diff --git a/CMakeLists.txt b/CMakeLists.txt index 7d95d3d..dce71ff 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -1,6 +1,7 @@ cmake_minimum_required (VERSION 3.5) project(whisper.cpp VERSION 1.5.4) +set(SOVERSION 1) # Add path to modules list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/") @@ -542,6 +543,12 @@ add_library(${TARGET} whisper.cpp ) +# Set the version numbers +set_target_properties(whisper PROPERTIES + VERSION ${PROJECT_VERSION} + SOVERSION ${SOVERSION} +) + include(DefaultTargetOptions) target_include_directories(${TARGET} PUBLIC diff --git a/README.md b/README.md index e470b1f..8773a0d 100644 --- a/README.md +++ b/README.md @@ -278,6 +278,7 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in - To ensure `coremltools` operates correctly, please confirm that [Xcode](https://developer.apple.com/xcode/) is installed and execute `xcode-select --install` to install the command-line tools. - Python 3.10 is recommended. + - MacOS Sonoma (version 14) or newer is recommended, as older versions of MacOS might experience issues with transcription hallucination. - [OPTIONAL] It is recommended to utilize a Python version management system, such as [Miniconda](https://docs.conda.io/en/latest/miniconda.html) for this step: - To create an environment, use: `conda create -n py310-whisper python=3.10 -y` - To activate the environment, use: `conda activate py310-whisper` diff --git a/bindings/go/whisper.go b/bindings/go/whisper.go index 9660662..87da83f 100644 --- a/bindings/go/whisper.go +++ b/bindings/go/whisper.go @@ -10,7 +10,7 @@ import ( /* #cgo LDFLAGS: -lwhisper -lm -lstdc++ -#cgo darwin LDFLAGS: -framework Accelerate +#cgo darwin LDFLAGS: -framework Accelerate -framework Metal -framework Foundation -framework CoreGraphics #include #include diff --git a/examples/bench/bench.cpp b/examples/bench/bench.cpp index 60d10a2..b77621a 100644 --- a/examples/bench/bench.cpp +++ b/examples/bench/bench.cpp @@ -8,7 +8,7 @@ // command-line parameters struct whisper_params { int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency()); - int32_t what = 0; // what to benchmark: 0 - whisper ecoder, 1 - memcpy, 2 - ggml_mul_mat + int32_t what = 0; // what to benchmark: 0 - whisper encoder, 1 - memcpy, 2 - ggml_mul_mat std::string model = "models/ggml-base.en.bin"; diff --git a/examples/main/main.cpp b/examples/main/main.cpp index cef9630..84db6a0 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -82,6 +82,14 @@ struct whisper_params { void whisper_print_usage(int argc, char ** argv, const whisper_params & params); +char* whisper_param_turn_lowercase(char* in){ + int string_len = strlen(in); + for(int i = 0; i < string_len; i++){ + *(in+i) = tolower((unsigned char)*(in+i)); + } + return in; +} + bool whisper_params_parse(int argc, char ** argv, whisper_params & params) { for (int i = 1; i < argc; i++) { std::string arg = argv[i]; @@ -135,7 +143,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) { else if (arg == "-pc" || arg == "--print-colors") { params.print_colors = true; } else if (arg == "-pp" || arg == "--print-progress") { params.print_progress = true; } else if (arg == "-nt" || arg == "--no-timestamps") { params.no_timestamps = true; } - else if (arg == "-l" || arg == "--language") { params.language = argv[++i]; } + else if (arg == "-l" || arg == "--language") { params.language = whisper_param_turn_lowercase(argv[++i]); } else if (arg == "-dl" || arg == "--detect-language") { params.detect_language = true; } else if ( arg == "--prompt") { params.prompt = argv[++i]; } else if (arg == "-m" || arg == "--model") { params.model = argv[++i]; } diff --git a/examples/talk-llama/llama.cpp b/examples/talk-llama/llama.cpp index 893bcdb..4225f95 100644 --- a/examples/talk-llama/llama.cpp +++ b/examples/talk-llama/llama.cpp @@ -104,6 +104,7 @@ #define LLAMA_MAX_NODES 8192 #define LLAMA_MAX_EXPERTS 8 + // // logging // @@ -211,10 +212,11 @@ enum llm_arch { LLM_ARCH_INTERNLM2, LLM_ARCH_MINICPM, LLM_ARCH_GEMMA, + LLM_ARCH_STARCODER2, LLM_ARCH_UNKNOWN, }; -static std::map LLM_ARCH_NAMES = { +static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_LLAMA, "llama" }, { LLM_ARCH_FALCON, "falcon" }, { LLM_ARCH_GPT2, "gpt2" }, @@ -238,6 +240,8 @@ static std::map LLM_ARCH_NAMES = { { LLM_ARCH_INTERNLM2, "internlm2" }, { LLM_ARCH_MINICPM, "minicpm" }, { LLM_ARCH_GEMMA, "gemma" }, + { LLM_ARCH_STARCODER2, "starcoder2" }, + { LLM_ARCH_UNKNOWN, "(unknown)" }, }; enum llm_kv { @@ -298,7 +302,7 @@ enum llm_kv { LLM_KV_TOKENIZER_RWKV, }; -static std::map LLM_KV_NAMES = { +static const std::map LLM_KV_NAMES = { { LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" }, { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" }, { LLM_KV_GENERAL_ALIGNMENT, "general.alignment" }, @@ -362,7 +366,7 @@ struct LLM_KV { llm_arch arch; std::string operator()(llm_kv kv) const { - return ::format(LLM_KV_NAMES[kv], LLM_ARCH_NAMES[arch]); + return ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch)); } }; @@ -397,7 +401,7 @@ enum llm_tensor { LLM_TENSOR_LAYER_OUT_NORM, }; -static std::map> LLM_TENSOR_NAMES = { +static const std::map> LLM_TENSOR_NAMES = { { LLM_ARCH_LLAMA, { @@ -779,6 +783,24 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_STARCODER2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_UNKNOWN, { @@ -812,38 +834,38 @@ struct LLM_TN { llm_arch arch; std::string operator()(llm_tensor tensor) const { - if (LLM_TENSOR_NAMES[arch].find(tensor) == LLM_TENSOR_NAMES[arch].end()) { + if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { return "__missing__"; } - return LLM_TENSOR_NAMES[arch].at(tensor); + return LLM_TENSOR_NAMES.at(arch).at(tensor); } std::string operator()(llm_tensor tensor, const std::string & suffix) const { - if (LLM_TENSOR_NAMES[arch].find(tensor) == LLM_TENSOR_NAMES[arch].end()) { + if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { return "__missing__"; } - return LLM_TENSOR_NAMES[arch].at(tensor) + "." + suffix; + return LLM_TENSOR_NAMES.at(arch).at(tensor) + "." + suffix; } std::string operator()(llm_tensor tensor, int bid) const { - if (LLM_TENSOR_NAMES[arch].find(tensor) == LLM_TENSOR_NAMES[arch].end()) { + if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { return "__missing__"; } - return ::format(LLM_TENSOR_NAMES[arch].at(tensor).c_str(), bid); + return ::format(LLM_TENSOR_NAMES.at(arch).at(tensor).c_str(), bid); } std::string operator()(llm_tensor tensor, const std::string & suffix, int bid) const { - if (LLM_TENSOR_NAMES[arch].find(tensor) == LLM_TENSOR_NAMES[arch].end()) { + if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { return "__missing__"; } - return ::format(LLM_TENSOR_NAMES[arch].at(tensor).c_str(), bid) + "." + suffix; + return ::format(LLM_TENSOR_NAMES.at(arch).at(tensor).c_str(), bid) + "." + suffix; } std::string operator()(llm_tensor tensor, const std::string & suffix, int bid, int xid) const { - if (LLM_TENSOR_NAMES[arch].find(tensor) == LLM_TENSOR_NAMES[arch].end()) { + if (LLM_TENSOR_NAMES.at(arch).find(tensor) == LLM_TENSOR_NAMES.at(arch).end()) { return "__missing__"; } - return ::format(LLM_TENSOR_NAMES[arch].at(tensor).c_str(), bid, xid) + "." + suffix; + return ::format(LLM_TENSOR_NAMES.at(arch).at(tensor).c_str(), bid, xid) + "." + suffix; } }; @@ -851,16 +873,16 @@ struct LLM_TN { // gguf helpers // -static std::map LLAMA_ROPE_SCALING_TYPES = { +static const std::map LLAMA_ROPE_SCALING_TYPES = { { LLAMA_ROPE_SCALING_TYPE_NONE, "none" }, { LLAMA_ROPE_SCALING_TYPE_LINEAR, "linear" }, { LLAMA_ROPE_SCALING_TYPE_YARN, "yarn" }, }; -static int32_t llama_rope_scaling_type_from_string(const std::string & name) { +static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::string & name) { for (const auto & kv : LLAMA_ROPE_SCALING_TYPES) { if (kv.second == name) { - return kv.first; + return (llama_rope_scaling_type) kv.first; } } @@ -1409,7 +1431,9 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(bool host_buffer buft = ggml_backend_cuda_host_buffer_type(); } #elif defined(GGML_USE_SYCL) - buft = ggml_backend_sycl_host_buffer_type(); + if (host_buffer) { + buft = ggml_backend_sycl_host_buffer_type(); + } #elif defined(GGML_USE_CPU_HBM) buft = ggml_backend_cpu_hbm_buffer_type(); #elif defined(GGML_USE_VULKAN) @@ -1463,6 +1487,12 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_split(int fallback_g } #endif +#ifdef GGML_USE_SYCL + if (ggml_backend_sycl_get_device_count() > 1) { + buft = ggml_backend_sycl_split_buffer_type(tensor_split); + } +#endif + if (buft == nullptr) { buft = llama_default_buffer_type_offload(fallback_gpu); } @@ -1474,6 +1504,8 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_split(int fallback_g static size_t llama_get_device_count() { #if defined(GGML_USE_CUBLAS) return ggml_backend_cuda_get_device_count(); +#elif defined(GGML_USE_SYCL) + return ggml_backend_sycl_get_device_count(); #elif defined(GGML_USE_VULKAN) return ggml_backend_vk_get_device_count(); #else @@ -1487,6 +1519,11 @@ static size_t llama_get_device_memory(int device) { size_t free; ggml_backend_cuda_get_device_memory(device, &total, &free); return free; +#elif defined(GGML_USE_SYCL) + size_t total; + size_t free; + ggml_backend_sycl_get_device_memory(device, &total, &free); + return free; #elif defined(GGML_USE_VULKAN) size_t total; size_t free; @@ -1575,7 +1612,6 @@ struct llama_hparams { float rope_freq_base_train; float rope_freq_scale_train; uint32_t n_yarn_orig_ctx; - int32_t rope_scaling_type_train; float f_clamp_kqv = 0.0f; float f_max_alibi_bias = 0.0f; @@ -1583,8 +1619,9 @@ struct llama_hparams { bool causal_attn = true; bool need_kq_pos = false; - enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_NONE; - enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE; + enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_NONE; + enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE; + enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE; bool operator!=(const llama_hparams & other) const { if (this->vocab_only != other.vocab_only) return true; @@ -1628,13 +1665,13 @@ struct llama_hparams { }; struct llama_cparams { - uint32_t n_ctx; // context size used during inference + uint32_t n_ctx; // context size used during inference uint32_t n_batch; uint32_t n_threads; // number of threads to use for generation uint32_t n_threads_batch; // number of threads to use for batch processing - float rope_freq_base; - float rope_freq_scale; + float rope_freq_base; + float rope_freq_scale; uint32_t n_yarn_orig_ctx; // These hyperparameters are not exposed in GGUF, because all @@ -1645,9 +1682,10 @@ struct llama_cparams { float yarn_beta_slow; float defrag_thold; - bool mul_mat_q; + bool embeddings; bool offload_kqv; - bool do_pooling; + + enum llama_pooling_type pooling_type; ggml_backend_sched_eval_callback cb_eval; void * cb_eval_user_data; @@ -1936,7 +1974,7 @@ struct llama_context { int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1) int32_t n_eval = 0; // number of eval calls - // decode output (2-dimensional array: [n_tokens][n_vocab]) + // logits output (2-dimensional array: [n_tokens][n_vocab]) std::vector logits; #ifndef NDEBUG // guard against access to unset logits @@ -1944,13 +1982,21 @@ struct llama_context { #endif bool logits_all = false; - // input embedding (1-dimensional array: [n_embd]) - std::vector embedding; + // embeddings output (2-dimensional array: [n_tokens][n_embd]) + // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE + std::vector embd; + + // sequence embeddings output (map of [n_embd] vectors) + // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE + std::map> embd_seq; // memory buffers used to evaluate the model std::vector buf_compute_meta; ggml_backend_sched_t sched = nullptr; + ggml_abort_callback abort_callback = nullptr; + void * abort_callback_data = nullptr; + // input tensors ggml_backend_buffer_t buf_input = nullptr; ggml_context * ctx_input = nullptr; @@ -2117,10 +2163,12 @@ static bool llama_kv_cache_find_slot( } // find how many cells are currently in use -static int32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache) { - for (uint32_t i = cache.size - 1; i > 0; --i) { - if (cache.cells[i].pos >= 0 && !cache.cells[i].is_empty()) { - return i + 1; +static uint32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache) { + for (uint32_t i = cache.size; i > 0; --i) { + const llama_kv_cell & cell = cache.cells[i - 1]; + + if (cell.pos >= 0 && !cell.is_empty()) { + return i; } } @@ -2892,7 +2940,11 @@ template<> bool llama_model_loader::get_key(const enum llm_kv kid, enum llama_pooling_type & result, const bool required) { uint32_t tmp; const bool found = get_key(kid, tmp, required); - result = (enum llama_pooling_type) tmp; + if (found) { + result = (enum llama_pooling_type) tmp; + } else { + result = LLAMA_POOLING_TYPE_UNSPECIFIED; + } return found; } @@ -3169,7 +3221,7 @@ static void llm_load_hparams( ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); ml.get_key(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn); ml.get_key(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, hparams.n_vocab_type); - ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type); + ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type, false); switch (hparams.n_layer) { case 3: @@ -3321,6 +3373,16 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_STARCODER2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + switch (hparams.n_layer) { + case 30: model.type = e_model::MODEL_3B; break; + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_15B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -4491,6 +4553,56 @@ static bool llm_load_tensors( layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); } } break; + case LLM_ARCH_STARCODER2: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); + + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false); + // if output is NULL, init from the input tok embed + if (model.output == NULL) { + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + ml.n_created--; // artificial tensor + ml.size_data += ggml_nbytes(model.output); + } + + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + + // optional bias tensors + layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}); + layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}); + layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}); + layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); + + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + + // optional bias tensors + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); + layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP , "bias", i), { n_ff}); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -4902,8 +5014,8 @@ static struct ggml_tensor * llm_build_kqv( ggml_mul_mat_set_prec(kq, GGML_PREC_F32); } -#if defined(GGML_USE_VULKAN) || defined(GGML_USE_KOMPUTE) -#pragma message("TODO: ALiBi support in ggml_soft_max_ext is not implemented for Vulkan, and Kompute") +#if defined(GGML_USE_KOMPUTE) +#pragma message("TODO: ALiBi support in ggml_soft_max_ext is not implemented for Kompute") #pragma message(" Falling back to ggml_alibi(). Will become an error in Mar 2024") #pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5488") if (hparams.f_max_alibi_bias > 0.0f) { @@ -4987,6 +5099,7 @@ static struct ggml_tensor * llm_build_kv( llm_build_kv_store(ctx, hparams, kv, graph, k_cur, v_cur, n_ctx, n_tokens, kv_head, cb, il); struct ggml_tensor * cur; + cur = llm_build_kqv(ctx, model, hparams, kv, graph, wo, wo_b, q_cur, kq_mask, kq_pos, n_ctx, n_tokens, n_kv, kq_scale, cb, il); cb(cur, "kqv_out", il); @@ -5074,7 +5187,7 @@ struct llm_build_context { n_kv (worst_case ? n_ctx : kv_self.n), kv_head (worst_case ? n_ctx - n_tokens : kv_self.head), n_orig_ctx (cparams.n_yarn_orig_ctx), - pooling_type (cparams.do_pooling ? hparams.pooling_type : LLAMA_POOLING_TYPE_NONE), + pooling_type (cparams.pooling_type), rope_type (hparams.rope_type), cb (cb), buf_compute_meta (lctx.buf_compute_meta) { @@ -5980,6 +6093,7 @@ struct llm_build_context { const int64_t n_embd_head = hparams.n_embd_head_v; const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); struct ggml_tensor * cur; @@ -5987,9 +6101,10 @@ struct llm_build_context { // get input vectors with right size const size_t stride1 = n_tokens * ggml_type_size(lctx.inp_tokens->type); - struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); + + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); struct ggml_tensor * inp_mean = ggml_view_2d(ctx0, lctx.inp_mean, n_tokens, n_tokens, stride1, 0); - struct ggml_tensor * inp_cls = ggml_view_1d(ctx0, lctx.inp_cls, n_tokens, 0); + struct ggml_tensor * inp_cls = ggml_view_1d(ctx0, lctx.inp_cls, n_tokens, 0); // construct input embeddings (token, type, position) inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); @@ -6007,39 +6122,38 @@ struct llm_build_context { cb(inpL, "inp_norm", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); - cb(KQ_mask, "KQ_mask", -1); // [n_kv, n_tokens] + struct ggml_tensor * KQ_mask = ggml_cont(ctx0, ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_tokens, n_tokens, n_tokens*ggml_type_size(lctx.inp_KQ_mask->type), 0)); + cb(KQ_mask, "KQ_mask", -1); // [n_tokens, n_tokens] // iterate layers for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * cur = inpL; + struct ggml_tensor * Qcur; + struct ggml_tensor * Kcur; + struct ggml_tensor * Vcur; + // self-attention if (model.arch == LLM_ARCH_BERT) { - struct ggml_tensor * Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), model.layers[il].bq); + Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), model.layers[il].bq); cb(Qcur, "Qcur", il); - struct ggml_tensor * Kcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), model.layers[il].bk); + Kcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), model.layers[il].bk); cb(Kcur, "Kcur", il); - struct ggml_tensor * Vcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), model.layers[il].bv); + Vcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), model.layers[il].bv); cb(Vcur, "Vcur", il); - // seems like we just need to do this for Q? - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - cb(cur, "kqv_out", il); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); } else { // compute Q and K and RoPE them cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); cb(cur, "wqkv", il); - struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); cb(Qcur, "Qcur", il); cb(Kcur, "Kcur", il); @@ -6058,13 +6172,41 @@ struct llm_build_context { ext_factor, attn_factor, beta_fast, beta_slow ); cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - cb(cur, "kqv_out", il); } + struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); + + struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + cb(kq, "kq", il); + + kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, nullptr, 1.0f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias); + cb(kq, "kq_soft_max_ext", il); + + struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens))); + cb(v, "v", il); + + struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq); + cb(kqv, "kqv", il); + + struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + cb(cur, "kqv_merged_cont", il); + + ggml_build_forward_expand(gf, cur); + + cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur); + if (model.layers[il].bo) { + cb(cur, "kqv_wo", il); + } + + if (model.layers[il].bo) { + cur = ggml_add(ctx0, cur, model.layers[il].bo); + } + cb(cur, "kqv_out", il); + // re-add the layer input cur = ggml_add(ctx0, cur, inpL); @@ -6104,16 +6246,29 @@ struct llm_build_context { // final output cur = inpL; + cb(cur, "result_embd", -1); // pooling layer - if (pooling_type == LLAMA_POOLING_TYPE_MEAN) { - cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, cur)), inp_mean); - } else if (pooling_type == LLAMA_POOLING_TYPE_CLS) { - cur = ggml_get_rows(ctx0, cur, inp_cls); - } else { - GGML_ASSERT(pooling_type == LLAMA_POOLING_TYPE_NONE && "Invalid pooling type"); + switch (pooling_type) { + case LLAMA_POOLING_TYPE_NONE: + { + // nop + } break; + case LLAMA_POOLING_TYPE_MEAN: + { + cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, cur)), inp_mean); + cb(cur, "result_embd_pooled", -1); + } break; + case LLAMA_POOLING_TYPE_CLS: + { + cur = ggml_get_rows(ctx0, cur, inp_cls); + cb(cur, "result_embd_pooled", -1); + } break; + case LLAMA_POOLING_TYPE_UNSPECIFIED: + { + GGML_ASSERT(false && "Invalid pooling type"); + } break; } - cb(cur, "result_embd", -1); ggml_build_forward_expand(gf, cur); @@ -7560,6 +7715,120 @@ struct llm_build_context { return gf; } + + struct ggml_cgraph * build_starcoder2() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); + cb(inp_pos, "inp_pos", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); + cb(KQ_mask, "KQ_mask", -1); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + cb(cur, "kqv_out", il); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } }; static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) { @@ -7706,6 +7975,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_gemma(); } break; + case LLM_ARCH_STARCODER2: + { + result = llm.build_starcoder2(); + } break; default: GGML_ASSERT(false); } @@ -7757,7 +8030,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos)); } - { + if (hparams.causal_attn) { const int64_t n_kv = kv_self.n; const int64_t n_tokens = batch.n_tokens; @@ -7772,16 +8045,40 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { for (int i = 0; i < n_kv; ++i) { float f; - if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || - (hparams.causal_attn && lctx.kv_self.cells[i].pos > pos)) { + if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) { f = -INFINITY; } else { - f = 0; + f = 0.0f; } data[h*(n_kv*n_tokens) + j*n_kv + i] = f; } } } + } else { + // non-causal attention attends only the tokens within the batch (i.e. the KV cache is not used) + const int64_t n_tokens = batch.n_tokens; + + assert(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer)); + + float * data = (float *) lctx.inp_KQ_mask->data; + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + const llama_seq_id seq_id = batch.seq_id[j][0]; + + for (int i = 0; i < n_tokens; ++i) { + float f = -INFINITY; + for (int s = 0; s < batch.n_seq_id[i]; ++s) { + if (batch.seq_id[i][s] == seq_id) { + f = 0.0f; + break; + } + } + + data[h*(n_tokens*n_tokens) + j*n_tokens + i] = f; + } + } + } } if (hparams.need_kq_pos) { @@ -7796,17 +8093,20 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } - if (cparams.do_pooling && hparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { + if (cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { const int64_t n_tokens = batch.n_tokens; GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_mean->buffer)); - float * data = (float *) lctx.inp_mean->data; + float * data = (float *) lctx.inp_mean->data; memset(lctx.inp_mean->data, 0, n_tokens * n_tokens * ggml_element_size(lctx.inp_mean)); std::vector sum(n_tokens, 0); for (int i = 0; i < n_tokens; ++i) { const llama_seq_id seq_id = batch.seq_id[i][0]; + + GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN"); + sum[seq_id] += 1; } @@ -7824,15 +8124,20 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } - if (cparams.do_pooling && hparams.pooling_type == LLAMA_POOLING_TYPE_CLS) { + if (cparams.pooling_type == LLAMA_POOLING_TYPE_CLS) { const int64_t n_tokens = batch.n_tokens; GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer)); + uint32_t * data = (uint32_t *) lctx.inp_cls->data; + memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls)); for (int i = 0; i < n_tokens; ++i) { const llama_seq_id seq_id = batch.seq_id[i][0]; - const llama_pos pos = batch.pos[i]; + const llama_pos pos = batch.pos[i]; + + GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS"); + if (pos == 0) { data[seq_id] = i; } @@ -7857,6 +8162,7 @@ static void llama_graph_compute( if (lctx.backend_cpu != nullptr) { ggml_backend_cpu_set_n_threads(lctx.backend_cpu, n_threads); + ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data); } ggml_backend_sched_graph_compute(lctx.sched, gf); @@ -7894,9 +8200,9 @@ static int llama_decode_internal( const auto n_batch = cparams.n_batch; GGML_ASSERT(n_tokens <= n_batch); + GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch; - GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT const int64_t t_start_us = ggml_time_us(); @@ -7945,24 +8251,27 @@ static int llama_decode_internal( batch.seq_id = seq_id_arr.data(); } - llama_kv_cache_update(&lctx); + // non-causal masks do not use the KV cache + if (hparams.causal_attn) { + llama_kv_cache_update(&lctx); - // if we have enough unused cells before the current head -> - // better to start searching from the beginning of the cache, hoping to fill it - if (kv_self.head > kv_self.used + 2*n_tokens) { - kv_self.head = 0; + // if we have enough unused cells before the current head -> + // better to start searching from the beginning of the cache, hoping to fill it + if (kv_self.head > kv_self.used + 2*n_tokens) { + kv_self.head = 0; + } + + if (!llama_kv_cache_find_slot(kv_self, batch)) { + return 1; + } + + // a heuristic, to avoid attending the full cache if it is not yet utilized + // after enough generations, the benefit from this heuristic disappears + // if we start defragmenting the cache, the benefit from this will be more important + kv_self.n = std::min(cparams.n_ctx, std::max(32u, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32))); + //kv_self.n = llama_kv_cache_cell_max(kv_self); } - if (!llama_kv_cache_find_slot(kv_self, batch)) { - return 1; - } - - // a heuristic, to avoid attending the full cache if it is not yet utilized - // after enough generations, the benefit from this heuristic disappears - // if we start defragmenting the cache, the benefit from this will be more important - kv_self.n = std::min((int32_t) cparams.n_ctx, std::max(32, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32))); - //kv_self.n = llama_kv_cache_cell_max(kv_self); - //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head); ggml_backend_sched_reset(lctx.sched); @@ -7971,20 +8280,26 @@ static int llama_decode_internal( ggml_cgraph * gf = llama_build_graph(lctx, batch, false); // the output is always the last tensor in the graph - struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1]; - struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2]; + struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1]; + struct ggml_tensor * embd = gf->nodes[gf->n_nodes - 2]; - if (strcmp(res->name, "result_output") == 0) { - // the embeddings could be the second to last tensor, or the third to last tensor - if (strcmp(embeddings->name, "result_norm") != 0) { - embeddings = gf->nodes[gf->n_nodes - 3]; - GGML_ASSERT(strcmp(embeddings->name, "result_norm") == 0); - } - } else if (strcmp(res->name, "result_embd") == 0) { - embeddings = res; - res = nullptr; + if (!hparams.causal_attn) { + res = nullptr; // do not extract logits for embedding models such as BERT + + // token or sequence embeddings + embd = gf->nodes[gf->n_nodes - 1]; + + GGML_ASSERT(strcmp(embd->name, "result_embd") == 0 || strcmp(embd->name, "result_embd_pooled") == 0); } else { - GGML_ASSERT(false); + if (strcmp(res->name, "result_output") == 0) { + // the token embeddings could be the second to last tensor, or the third to last tensor + if (strcmp(embd->name, "result_norm") != 0) { + embd = gf->nodes[gf->n_nodes - 3]; + GGML_ASSERT(strcmp(embd->name, "result_norm") == 0); + } + } else { + GGML_ASSERT(false && "missing result_output tensor"); + } } // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs); @@ -8051,46 +8366,82 @@ static int llama_decode_internal( logits_out.clear(); #endif - ggml_backend_t res_backend = ggml_backend_sched_get_node_backend(lctx.sched, res); - GGML_ASSERT(res_backend != nullptr); + ggml_backend_t backend_res = ggml_backend_sched_get_node_backend(lctx.sched, res); + GGML_ASSERT(backend_res != nullptr); + if (batch.logits) { logits_out.resize(n_vocab * n_tokens); for (uint32_t i = 0; i < n_tokens; i++) { if (batch.logits[i] == 0) { continue; } - ggml_backend_tensor_get_async(res_backend, res, logits_out.data() + (n_vocab*i), (n_vocab*i)*sizeof(float), n_vocab*sizeof(float)); + ggml_backend_tensor_get_async(backend_res, res, logits_out.data() + (n_vocab*i), (n_vocab*i)*sizeof(float), n_vocab*sizeof(float)); #ifndef NDEBUG logits_valid[i] = true; #endif } } else if (lctx.logits_all) { logits_out.resize(n_vocab * n_tokens); - ggml_backend_tensor_get_async(res_backend, res, logits_out.data(), 0, n_vocab*n_tokens*sizeof(float)); + ggml_backend_tensor_get_async(backend_res, res, logits_out.data(), 0, n_vocab*n_tokens*sizeof(float)); #ifndef NDEBUG std::fill(logits_valid.begin(), logits_valid.end(), true); #endif } else { logits_out.resize(n_vocab); - ggml_backend_tensor_get_async(res_backend, res, logits_out.data(), (n_vocab*(n_tokens - 1))*sizeof(float), n_vocab*sizeof(float)); + ggml_backend_tensor_get_async(backend_res, res, logits_out.data(), (n_vocab*(n_tokens - 1))*sizeof(float), n_vocab*sizeof(float)); #ifndef NDEBUG logits_valid[0] = true; #endif } - ggml_backend_synchronize(res_backend); + ggml_backend_synchronize(backend_res); } // extract embeddings - if (!lctx.embedding.empty()) { - auto & embedding_out = lctx.embedding; + if (cparams.embeddings && embd) { + ggml_backend_t backend_embd = ggml_backend_sched_get_node_backend(lctx.sched, embd); + GGML_ASSERT(backend_embd != nullptr); - const int64_t embd_pos = res ? n_embd * (n_tokens-1) : 0; - const int64_t embd_size = res ? n_embd : n_embd * n_tokens; + switch (cparams.pooling_type) { + case LLAMA_POOLING_TYPE_NONE: + { + // extract token embeddings + auto & embd_out = lctx.embd; - embedding_out.resize(embd_size); - ggml_backend_t embeddings_backend = ggml_backend_sched_get_node_backend(lctx.sched, embeddings); - ggml_backend_tensor_get_async(embeddings_backend, embeddings, embedding_out.data(), embd_pos*sizeof(float), embd_size*sizeof(float)); - ggml_backend_synchronize(embeddings_backend); + if (batch.logits) { + embd_out.resize(n_embd * n_tokens); + for (uint32_t i = 0; i < n_tokens; i++) { + if (batch.logits[i] == 0) { + continue; + } + + ggml_backend_tensor_get_async(backend_embd, embd, embd_out.data() + (n_embd*i), (n_embd*i)*sizeof(float), n_embd*sizeof(float)); + } + } + } break; + case LLAMA_POOLING_TYPE_CLS: + case LLAMA_POOLING_TYPE_MEAN: + { + GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0); + + // extract sequence embeddings + auto & embd_seq_out = lctx.embd_seq; + embd_seq_out.clear(); + + for (uint32_t i = 0; i < n_tokens; i++) { + const llama_seq_id seq_id = batch.seq_id[i][0]; + if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { + continue; + } + embd_seq_out[seq_id].resize(n_embd); + ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float)); + } + } break; + case LLAMA_POOLING_TYPE_UNSPECIFIED: + { + GGML_ASSERT(false && "unknown pooling type"); + } break; + } + ggml_backend_synchronize(backend_embd); } // measure the performance only for the single-token evals @@ -8384,19 +8735,19 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) { GGML_ASSERT(llama_is_byte_token(vocab, id)); const auto& token_data = vocab.id_to_token.at(id); switch (llama_vocab_get_type(vocab)) { - case LLAMA_VOCAB_TYPE_SPM: { - auto buf = token_data.text.substr(3, 2); - return strtol(buf.c_str(), NULL, 16); - } - case LLAMA_VOCAB_TYPE_BPE: { - GGML_ASSERT(false); - return unicode_to_bytes_bpe(token_data.text); - } - case LLAMA_VOCAB_TYPE_WPM: { - GGML_ASSERT(false); - } - default: - GGML_ASSERT(false); + case LLAMA_VOCAB_TYPE_SPM: { + auto buf = token_data.text.substr(3, 2); + return strtol(buf.c_str(), NULL, 16); + } + case LLAMA_VOCAB_TYPE_BPE: { + GGML_ASSERT(false); + return unicode_to_bytes_bpe(token_data.text); + } + case LLAMA_VOCAB_TYPE_WPM: { + GGML_ASSERT(false); + } + default: + GGML_ASSERT(false); } } @@ -8947,10 +9298,10 @@ struct llm_tokenizer_wpm { std::vector codepoints = codepoints_from_utf8(text); std::vector nfd_codepoints; for (uint32_t code : codepoints) { - auto it = nfd_map.find(code); - if (it != nfd_map.end()) { - for (uint32_t c : it->second) { - nfd_codepoints.push_back(c); + auto it = nfd_map.equal_range(code); + if (it.first != it.second) { + for (auto jt = it.first; jt != it.second; jt++) { + nfd_codepoints.push_back(jt->second); } } else { nfd_codepoints.push_back(code); @@ -9001,12 +9352,13 @@ struct llm_tokenizer_wpm { } uint32_t to_lower(uint32_t code) { + static const std::locale locale("en_US.UTF-8"); #if defined(_WIN32) if (code > 0xFFFF) { return code; } #endif - return std::tolower(wchar_t(code), std::locale("en_US.UTF-8")); + return std::tolower(wchar_t(code), locale); } bool is_ascii_punct(uint32_t code) { @@ -10062,10 +10414,6 @@ void llama_sample_temp(struct llama_context * ctx, llama_token_data_array * cand } } -void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) { - llama_sample_temp(ctx, candidates_p, temp); -} - void llama_sample_repetition_penalties( struct llama_context * ctx, llama_token_data_array * candidates, @@ -10192,38 +10540,6 @@ void llama_sample_apply_guidance( ctx->t_sample_us += ggml_time_us() - t_start_sample_us; } -void llama_sample_classifier_free_guidance( - struct llama_context * ctx, - llama_token_data_array * candidates, - struct llama_context * guidance_ctx, - float scale) { - GGML_ASSERT(ctx); - int64_t t_start_sample_us; - - t_start_sample_us = ggml_time_us(); - const size_t n_vocab = llama_n_vocab(llama_get_model(ctx)); - - GGML_ASSERT(n_vocab == candidates->size); - GGML_ASSERT(!candidates->sorted); - - std::vector logits_base(n_vocab); - for (size_t i = 0; i < n_vocab; ++i) { - logits_base[i] = candidates->data[i].logit; - } - - float * logits_guidance = llama_get_logits(guidance_ctx); - - ctx->t_sample_us += ggml_time_us() - t_start_sample_us; - llama_sample_apply_guidance(ctx, logits_base.data(), logits_guidance, scale); - t_start_sample_us = ggml_time_us(); - - for (size_t i = 0; i < n_vocab; ++i) { - candidates->data[i].logit = logits_base[i]; - } - - ctx->t_sample_us += ggml_time_us() - t_start_sample_us; -} - llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int32_t m, float * mu) { GGML_ASSERT(ctx); @@ -10657,7 +10973,7 @@ struct quantize_state_internal { {} }; -static void llama_convert_tensor_internal( +static void llama_tensor_dequantize_internal( struct ggml_tensor * tensor, std::vector> & output, std::vector & workers, const size_t nelements, const int nthread ) { @@ -10998,6 +11314,46 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty return new_type; } +static int32_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int chunk_size, int nrows, int n_per_row, int64_t * hist_cur, const float * imatrix, std::vector & workers, const int nthread) { + std::mutex mutex; + int counter = 0; + size_t new_size = 0; + if (nthread < 2) { + // single-thread + return ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, hist_cur, imatrix); + } + auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, chunk_size, + nrows, n_per_row, imatrix]() { + std::array local_hist = {}; + const int nrows_per_chunk = chunk_size / n_per_row; + size_t local_size = 0; + while (true) { + std::unique_lock lock(mutex); + int first_row = counter; counter += nrows_per_chunk; + if (first_row >= nrows) { + if (local_size > 0) { + for (int j=0; jftype; @@ -11110,7 +11466,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s std::vector workers; workers.reserve(nthread); - std::mutex mutex; int idx = 0; @@ -11224,7 +11579,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } else if (ggml_is_quantized(tensor->type) && !params->allow_requantize) { throw std::runtime_error(format("requantizing from type %s is disabled", ggml_type_name(tensor->type))); } else { - llama_convert_tensor_internal(tensor, f32_conv_buf, workers, nelements, nthread); + llama_tensor_dequantize_internal(tensor, f32_conv_buf, workers, nelements, nthread); f32_data = (float *) f32_conv_buf.data(); } @@ -11245,41 +11600,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s const int nchunk = (nelements + chunk_size - 1)/chunk_size; const int nthread_use = nthread > 1 ? std::max(1, std::min(nthread, nchunk)) : 1; - if (nthread_use < 2) { - new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, hist_cur.data(), imatrix); - } else { - int counter = 0; - new_size = 0; - auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, chunk_size, - nrows, n_per_row, imatrix]() { - std::array local_hist = {}; - const int nrows_per_chunk = chunk_size / n_per_row; - size_t local_size = 0; - while (true) { - std::unique_lock lock(mutex); - int first_row = counter; counter += nrows_per_chunk; - if (first_row >= nrows) { - if (local_size > 0) { - for (int j=0; j %8.2f MiB", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0); int64_t tot_count = 0; @@ -11656,6 +11977,7 @@ struct llama_context_params llama_context_default_params() { /*.n_threads =*/ GGML_DEFAULT_N_THREADS, // TODO: better default /*.n_threads_batch =*/ GGML_DEFAULT_N_THREADS, /*.rope_scaling_type =*/ LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED, + /*.pooling_type =*/ LLAMA_POOLING_TYPE_UNSPECIFIED, /*.rope_freq_base =*/ 0.0f, /*.rope_freq_scale =*/ 0.0f, /*.yarn_ext_factor =*/ -1.0f, @@ -11668,11 +11990,11 @@ struct llama_context_params llama_context_default_params() { /*.cb_eval_user_data =*/ nullptr, /*.type_k =*/ GGML_TYPE_F16, /*.type_v =*/ GGML_TYPE_F16, - /*.mul_mat_q =*/ true, /*.logits_all =*/ false, - /*.embedding =*/ false, + /*.embeddings =*/ false, /*.offload_kqv =*/ true, - /*.do_pooling =*/ true, + /*.abort_callback =*/ nullptr, + /*.abort_callback_data =*/ nullptr, }; return result; @@ -11724,15 +12046,6 @@ bool llama_supports_gpu_offload(void) { #endif } -// deprecated: -bool llama_mmap_supported(void) { - return llama_supports_mmap(); -} - -bool llama_mlock_supported(void) { - return llama_supports_mlock(); -} - void llama_backend_init(void) { ggml_time_init(); @@ -11829,9 +12142,9 @@ struct llama_context * llama_new_context_with_model( cparams.yarn_beta_fast = params.yarn_beta_fast; cparams.yarn_beta_slow = params.yarn_beta_slow; cparams.defrag_thold = params.defrag_thold; - cparams.mul_mat_q = params.mul_mat_q; + cparams.embeddings = params.embeddings; cparams.offload_kqv = params.offload_kqv; - cparams.do_pooling = params.do_pooling; + cparams.pooling_type = params.pooling_type; cparams.n_ctx = params.n_ctx == 0 ? hparams.n_ctx_train : params.n_ctx; cparams.rope_freq_base = params.rope_freq_base == 0.0f ? hparams.rope_freq_base_train : params.rope_freq_base; @@ -11857,6 +12170,14 @@ struct llama_context * llama_new_context_with_model( cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f; } + if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) { + if (hparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) { + cparams.pooling_type = LLAMA_POOLING_TYPE_NONE; + } else { + cparams.pooling_type = hparams.pooling_type; + } + } + if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); } @@ -11865,8 +12186,11 @@ struct llama_context * llama_new_context_with_model( LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base); LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale); - ctx->rng = std::mt19937(params.seed); - ctx->logits_all = params.logits_all; + ctx->abort_callback = params.abort_callback; + ctx->abort_callback_data = params.abort_callback_data; + + ctx->rng = std::mt19937(params.seed); + ctx->logits_all = params.logits_all; const ggml_type type_k = params.type_k; const ggml_type type_v = params.type_v; @@ -11924,13 +12248,31 @@ struct llama_context * llama_new_context_with_model( } #elif defined(GGML_USE_SYCL) if (model->n_gpu_layers > 0) { - ggml_backend_t backend = ggml_backend_sycl_init(model->main_gpu); - if (backend == nullptr) { - LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d backend\n", __func__, model->main_gpu); - llama_free(ctx); - return nullptr; + // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used + if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) { + int main_gpu_index = ggml_backend_sycl_get_device_index(model->main_gpu); + ggml_backend_t backend = ggml_backend_sycl_init(main_gpu_index); + if (backend == nullptr) { + LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d)backend\n", __func__, model->main_gpu, main_gpu_index); + llama_free(ctx); + return nullptr; + } + ctx->backends.push_back(backend); + } else { + // LLAMA_SPLIT_LAYER requires a backend for each GPU + int id_list[GGML_SYCL_MAX_DEVICES]; + ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES); + for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) { + int device_id = id_list[i]; + ggml_backend_t backend = ggml_backend_sycl_init(i); + if (backend == nullptr) { + LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d)backend\n", __func__, device_id, i); + llama_free(ctx); + return nullptr; + } + ctx->backends.push_back(backend); + } } - ctx->backends.push_back(backend); } #elif defined(GGML_USE_KOMPUTE) if (model->n_gpu_layers > 0) { @@ -11978,8 +12320,8 @@ struct llama_context * llama_new_context_with_model( // resized during inference, reserve maximum ctx->logits.reserve(hparams.n_vocab*cparams.n_batch); - if (params.embedding) { - ctx->embedding.resize(hparams.n_embd); + if (params.embeddings) { + ctx->embd.reserve(hparams.n_embd*cparams.n_batch); } // graph inputs @@ -12010,7 +12352,6 @@ struct llama_context * llama_new_context_with_model( ggml_set_name(ctx->inp_cls, "inp_cls"); ctx->buf_input = ggml_backend_alloc_ctx_tensors_from_buft(ctx->ctx_input, llama_default_buffer_type_cpu(true)); - LLAMA_LOG_INFO("%s: %10s input buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(ctx->buf_input), ggml_backend_buffer_get_size(ctx->buf_input) / 1024.0 / 1024.0); @@ -12131,6 +12472,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_QWEN2: case LLM_ARCH_PHI2: case LLM_ARCH_GEMMA: + case LLM_ARCH_STARCODER2: return LLAMA_ROPE_TYPE_NEOX; // all model arches should be listed explicitly here @@ -12244,15 +12586,6 @@ uint32_t llama_model_quantize( } } -int32_t llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lora, float scale, const char * path_base_model, int32_t n_threads) { - try { - return llama_apply_lora_from_file_internal(ctx->model, path_lora, scale, path_base_model, n_threads); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what()); - return 1; - } -} - int32_t llama_model_apply_lora_from_file(const struct llama_model * model, const char * path_lora, float scale, const char * path_base_model, int32_t n_threads) { try { return llama_apply_lora_from_file_internal(*model, path_lora, scale, path_base_model, n_threads); @@ -12423,10 +12756,15 @@ size_t llama_get_state_size(const struct llama_context * ctx) { // assume worst case for logits although only currently set ones are serialized const size_t s_logits = ctx->logits.capacity() * sizeof(float); const size_t s_embedding_size = sizeof(size_t); - const size_t s_embedding = ctx->embedding.size() * sizeof(float); - const size_t s_kv_size = sizeof(size_t); - const size_t s_kv_ntok = sizeof(int); + const size_t s_embedding = ctx->embd.capacity() * sizeof(float); + const size_t s_kv_buf_size = sizeof(size_t); + const size_t s_kv_head = sizeof(uint32_t); + const size_t s_kv_size = sizeof(uint32_t); + const size_t s_kv_used = sizeof(uint32_t); const size_t s_kv = ctx->kv_self.total_size(); + // TODO: assume the max is more than 1 seq_id per KV cell + const size_t s_kv_cell = sizeof(llama_pos) + sizeof(size_t) + sizeof(llama_seq_id); + const size_t s_kv_cells = ctx->kv_self.size * s_kv_cell; const size_t s_total = ( + s_rng_size @@ -12435,9 +12773,12 @@ size_t llama_get_state_size(const struct llama_context * ctx) { + s_logits + s_embedding_size + s_embedding + + s_kv_buf_size + + s_kv_head + s_kv_size - + s_kv_ntok + + s_kv_used + s_kv + + s_kv_cells ); return s_total; @@ -12524,12 +12865,12 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat // copy embeddings { - const size_t embedding_size = ctx->embedding.size(); + const size_t embeddings_size = ctx->embd.size(); - data_ctx->write(&embedding_size, sizeof(embedding_size)); + data_ctx->write(&embeddings_size, sizeof(embeddings_size)); - if (embedding_size) { - data_ctx->write(ctx->embedding.data(), embedding_size * sizeof(float)); + if (embeddings_size) { + data_ctx->write(ctx->embd.data(), embeddings_size * sizeof(float)); } } @@ -12537,15 +12878,13 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat { const auto & kv_self = ctx->kv_self; const auto & hparams = ctx->model.hparams; - const auto & cparams = ctx->cparams; const uint32_t n_layer = hparams.n_layer; const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(); const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(); - const uint32_t n_ctx = cparams.n_ctx; const size_t kv_buf_size = kv_self.total_size(); - const uint32_t kv_head = kv_self.head; + const uint32_t kv_head = llama_kv_cache_cell_max(kv_self); const uint32_t kv_size = kv_self.size; const uint32_t kv_used = kv_self.used; @@ -12565,7 +12904,7 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat // v is not contiguous, copy row by row const size_t v_row_size = ggml_row_size(kv_self.v_l[il]->type, kv_head); - const size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, n_ctx); + const size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, kv_size); tmp_buf.resize(v_row_size); for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) { @@ -12575,7 +12914,7 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat } } - for (uint32_t i = 0; i < kv_size; ++i) { + for (uint32_t i = 0; i < kv_head; ++i) { const auto & cell = kv_self.cells[i]; const llama_pos pos = cell.pos; @@ -12599,8 +12938,8 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { } // Sets the state reading from the specified source address -size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { - uint8_t * inp = src; +size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { + const uint8_t * inp = src; // set rng { @@ -12609,7 +12948,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { GGML_ASSERT(rng_size <= LLAMA_MAX_RNG_STATE); - std::string rng_str((char *)inp, rng_size); inp += rng_size; + std::string rng_str((const char *)inp, rng_size); inp += rng_size; std::istringstream rng_ss(rng_str); rng_ss >> ctx->rng; @@ -12635,15 +12974,17 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { // set embeddings { - size_t embedding_size; + size_t embeddings_size; - memcpy(&embedding_size, inp, sizeof(embedding_size)); inp += sizeof(embedding_size); + memcpy(&embeddings_size, inp, sizeof(embeddings_size)); inp += sizeof(embeddings_size); - GGML_ASSERT(ctx->embedding.capacity() == embedding_size); + GGML_ASSERT(ctx->embd.capacity() == embeddings_size); - if (embedding_size) { - memcpy(ctx->embedding.data(), inp, embedding_size * sizeof(float)); - inp += embedding_size * sizeof(float); + if (embeddings_size) { + ctx->embd.resize(embeddings_size); + + memcpy(ctx->embd.data(), inp, embeddings_size * sizeof(float)); + inp += embeddings_size * sizeof(float); } } @@ -12651,12 +12992,10 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { { const auto & kv_self = ctx->kv_self; const auto & hparams = ctx->model.hparams; - const auto & cparams = ctx->cparams; const uint32_t n_layer = hparams.n_layer; const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(); const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(); - const uint32_t n_ctx = cparams.n_ctx; size_t kv_buf_size; uint32_t kv_head; @@ -12679,7 +13018,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { // v is not contiguous, copy row by row const size_t v_row_size = ggml_row_size(kv_self.v_l[il]->type, kv_head); - const size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, n_ctx); + const size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, kv_size); for (int ir = 0; ir < (int) n_embd_v_gqa; ++ir) { ggml_backend_tensor_set(kv_self.v_l[il], inp, ir*v_row_stride, v_row_size); @@ -12688,13 +13027,15 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { } } + GGML_ASSERT(kv_self.size == kv_size); + ctx->kv_self.head = kv_head; ctx->kv_self.size = kv_size; ctx->kv_self.used = kv_used; ctx->kv_self.cells.resize(kv_size); - for (uint32_t i = 0; i < kv_size; ++i) { + for (uint32_t i = 0; i < kv_head; ++i) { llama_pos pos; size_t seq_id_size; @@ -12710,6 +13051,11 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { ctx->kv_self.cells[i].seq_id.insert(seq_id); } } + + for (uint32_t i = kv_head; i < kv_size; ++i) { + ctx->kv_self.cells[i].pos = -1; + ctx->kv_self.cells[i].seq_id.clear(); + } } const size_t nread = inp - src; @@ -12802,43 +13148,16 @@ bool llama_save_session_file(struct llama_context * ctx, const char * path_sessi return true; } -int llama_eval( - struct llama_context * ctx, - llama_token * tokens, - int32_t n_tokens, - int32_t n_past) { - llama_kv_cache_seq_rm(ctx->kv_self, -1, n_past, -1); - - const int ret = llama_decode_internal(*ctx, llama_batch_get_one(tokens, n_tokens, n_past, 0)); - if (ret < 0) { - LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret); - } - - return ret; -} - -int llama_eval_embd( - struct llama_context * ctx, - float * embd, - int32_t n_tokens, - int32_t n_past) { - llama_kv_cache_seq_rm(ctx->kv_self, -1, n_past, -1); - - llama_batch batch = { n_tokens, nullptr, embd, nullptr, nullptr, nullptr, nullptr, n_past, 1, 0, }; - - const int ret = llama_decode_internal(*ctx, batch); - if (ret < 0) { - LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret); - } - - return ret; -} - void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch) { ctx->cparams.n_threads = n_threads; ctx->cparams.n_threads_batch = n_threads_batch; } +void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) { + ctx->abort_callback = abort_callback; + ctx->abort_callback_data = abort_callback_data; +} + struct llama_batch llama_batch_get_one( llama_token * tokens, int32_t n_tokens, @@ -12915,11 +13234,20 @@ float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) { } float * llama_get_embeddings(struct llama_context * ctx) { - return ctx->embedding.data(); + return ctx->embd.data(); } float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) { - return ctx->embedding.data() + i*ctx->model.hparams.n_embd; + return ctx->embd.data() + i*ctx->model.hparams.n_embd; +} + +float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id) { + auto it = ctx->embd_seq.find(seq_id); + if (it == ctx->embd_seq.end()) { + return nullptr; + } + + return it->second.data(); } const char * llama_token_get_text(const struct llama_model * model, llama_token token) { @@ -13093,7 +13421,7 @@ static int32_t llama_chat_apply_template_internal( std::string & dest, bool add_ass) { // Taken from the research: https://github.com/ggerganov/llama.cpp/issues/5527 std::stringstream ss; - if (tmpl.find("<|im_start|>") != std::string::npos) { + if (tmpl == "chatml" || tmpl.find("<|im_start|>") != std::string::npos) { // chatml template for (auto message : chat) { ss << "<|im_start|>" << message->role << "\n" << message->content << "<|im_end|>\n"; @@ -13101,7 +13429,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|im_start|>assistant\n"; } - } else if (tmpl.find("[INST]") != std::string::npos) { + } else if (tmpl == "llama2" || tmpl.find("[INST]") != std::string::npos) { // llama2 template and its variants // [variant] support system message bool support_system_message = tmpl.find("<>") != std::string::npos; @@ -13136,7 +13464,7 @@ static int32_t llama_chat_apply_template_internal( } } // llama2 templates seem to not care about "add_generation_prompt" - } else if (tmpl.find("<|user|>") != std::string::npos) { + } else if (tmpl == "zephyr" || tmpl.find("<|user|>") != std::string::npos) { // zephyr template for (auto message : chat) { ss << "<|" << message->role << "|>" << "\n" << message->content << "<|endoftext|>\n"; @@ -13144,7 +13472,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|assistant|>\n"; } - } else if (tmpl.find("bos_token + message['role']") != std::string::npos) { + } else if (tmpl == "monarch" || tmpl.find("bos_token + message['role']") != std::string::npos) { // mlabonne/AlphaMonarch-7B template (the is included inside history) for (auto message : chat) { std::string bos = (message == chat.front()) ? "" : ""; // skip BOS for first message @@ -13153,7 +13481,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "assistant\n"; } - } else if (tmpl.find("") != std::string::npos) { + } else if (tmpl == "gemma" || tmpl.find("") != std::string::npos) { // google/gemma-7b-it std::string system_prompt = ""; for (auto message : chat) { @@ -13200,23 +13528,27 @@ LLAMA_API int32_t llama_chat_apply_template( int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size()); if (res < 0) { // worst case: there is no information about template, we will use chatml by default - curr_tmpl = "<|im_start|>"; // see llama_chat_apply_template_internal + curr_tmpl = "chatml"; // see llama_chat_apply_template_internal } else { curr_tmpl = std::string(model_template.data(), model_template.size()); } } + // format the chat to string std::vector chat_vec; chat_vec.resize(n_msg); for (size_t i = 0; i < n_msg; i++) { chat_vec[i] = &chat[i]; } + std::string formatted_chat; int32_t res = llama_chat_apply_template_internal(curr_tmpl, chat_vec, formatted_chat, add_ass); if (res < 0) { return res; } - strncpy(buf, formatted_chat.c_str(), length); + if (buf && length > 0) { + strncpy(buf, formatted_chat.c_str(), length); + } return res; } diff --git a/examples/talk-llama/llama.h b/examples/talk-llama/llama.h index 16e28e9..3dc162b 100644 --- a/examples/talk-llama/llama.h +++ b/examples/talk-llama/llama.h @@ -129,6 +129,7 @@ extern "C" { }; enum llama_pooling_type { + LLAMA_POOLING_TYPE_UNSPECIFIED = -1, LLAMA_POOLING_TYPE_NONE = 0, LLAMA_POOLING_TYPE_MEAN = 1, LLAMA_POOLING_TYPE_CLS = 2, @@ -162,7 +163,7 @@ extern "C" { // - embd : token embeddings (i.e. float vector of size n_embd) (used when token is NULL) // - pos : the positions of the respective token in the sequence // - seq_id : the sequence to which the respective token belongs - // - logits : if zero, the logits for the respective token will not be output + // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output // typedef struct llama_batch { int32_t n_tokens; @@ -172,7 +173,7 @@ extern "C" { llama_pos * pos; int32_t * n_seq_id; llama_seq_id ** seq_id; - int8_t * logits; + int8_t * logits; // TODO: rename this to "output" // NOTE: helpers for smooth API transition - can be deprecated in the future // for future-proof code, use the above fields instead and ignore everything below @@ -236,7 +237,10 @@ extern "C" { uint32_t n_batch; // prompt processing maximum batch size uint32_t n_threads; // number of threads to use for generation uint32_t n_threads_batch; // number of threads to use for batch processing - int32_t rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` + + enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` + enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id + // (ignored if no pooling layer) // ref: https://github.com/ggerganov/llama.cpp/pull/2054 float rope_freq_base; // RoPE base frequency, 0 = from model @@ -255,11 +259,15 @@ extern "C" { enum ggml_type type_v; // data type for V cache // Keep the booleans together to avoid misalignment during copy-by-value. - bool mul_mat_q; // if true, use experimental mul_mat_q kernels (DEPRECATED - always true) - bool logits_all; // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) - bool embedding; // embedding mode only + bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) + bool embeddings; // if true, extract embeddings (together with logits) bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU - bool do_pooling; // whether to pool (sum) embedding results by sequence id (ignored if no pooling layer) + + // Abort callback + // if it returns true, execution of llama_decode() will be aborted + // currently works only with CPU execution + ggml_abort_callback abort_callback; + void * abort_callback_data; }; // model quantization parameters @@ -364,9 +372,6 @@ extern "C" { LLAMA_API bool llama_supports_mlock (void); LLAMA_API bool llama_supports_gpu_offload(void); - LLAMA_API DEPRECATED(bool llama_mmap_supported (void), "use llama_supports_mmap() instead"); - LLAMA_API DEPRECATED(bool llama_mlock_supported(void), "use llama_supports_mlock() instead"); - LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); @@ -423,14 +428,6 @@ extern "C" { // The model needs to be reloaded before applying a new adapter, otherwise the adapter // will be applied on top of the previous one // Returns 0 on success - LLAMA_API DEPRECATED(int32_t llama_apply_lora_from_file( - struct llama_context * ctx, - const char * path_lora, - float scale, - const char * path_base_model, - int32_t n_threads), - "use llama_model_apply_lora_from_file instead"); - LLAMA_API int32_t llama_model_apply_lora_from_file( const struct llama_model * model, const char * path_lora, @@ -586,7 +583,7 @@ extern "C" { // Returns the number of bytes read LLAMA_API size_t llama_set_state_data( struct llama_context * ctx, - uint8_t * src); + const uint8_t * src); // Save/load session file LLAMA_API bool llama_load_session_file( @@ -606,27 +603,6 @@ extern "C" { // Decoding // - // Run the llama inference to obtain the logits and probabilities for the next token(s). - // tokens + n_tokens is the provided batch of new tokens to process - // n_past is the number of tokens to use from previous eval calls - // Returns 0 on success - // DEPRECATED: use llama_decode() instead - LLAMA_API DEPRECATED(int llama_eval( - struct llama_context * ctx, - llama_token * tokens, - int32_t n_tokens, - int32_t n_past), - "use llama_decode() instead"); - - // Same as llama_eval, but use float matrix input directly. - // DEPRECATED: use llama_decode() instead - LLAMA_API DEPRECATED(int llama_eval_embd( - struct llama_context * ctx, - float * embd, - int32_t n_tokens, - int32_t n_past), - "use llama_decode() instead"); - // Return batch for single sequence of tokens starting at pos_0 // // NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it @@ -665,7 +641,10 @@ extern "C" { // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens) LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch); - // Token logits obtained from the last call to llama_eval() + // Set abort callback + LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data); + + // Token logits obtained from the last call to llama_decode() // The logits for the last token are stored in the last row // Logits for which llama_batch.logits[i] == 0 are undefined // Rows: n_tokens provided with llama_batch @@ -676,14 +655,20 @@ extern "C" { // llama_get_logits(ctx) + i*n_vocab LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i); - // Get the embeddings for the input - // shape: [n_embd] (1-dimensional) + // Get all output token embeddings + // shape: [n_tokens*n_embd] (1-dimensional) LLAMA_API float * llama_get_embeddings(struct llama_context * ctx); - // Get the embeddings for the ith sequence + // Get the embeddings for the ith token // llama_get_embeddings(ctx) + i*n_embd + // shape: [n_embd] (1-dimensional) LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i); + // Get the embeddings for a sequence id + // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE + // shape: [n_embd] (1-dimensional) + LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id); + // // Vocab // @@ -800,13 +785,6 @@ extern "C" { float * logits_guidance, float scale); - LLAMA_API DEPRECATED(void llama_sample_classifier_free_guidance( - struct llama_context * ctx, - llama_token_data_array * candidates, - struct llama_context * guidance_ctx, - float scale), - "use llama_sample_apply_guidance() instead"); - /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. LLAMA_API void llama_sample_softmax( struct llama_context * ctx, @@ -860,12 +838,6 @@ extern "C" { llama_token_data_array * candidates, float temp); - LLAMA_API DEPRECATED(void llama_sample_temperature( - struct llama_context * ctx, - llama_token_data_array * candidates, - float temp), - "use llama_sample_temp instead"); - /// @details Apply constraints from grammar LLAMA_API void llama_sample_grammar( struct llama_context * ctx, diff --git a/examples/talk-llama/talk-llama.cpp b/examples/talk-llama/talk-llama.cpp index ddc9e76..4e1c175 100644 --- a/examples/talk-llama/talk-llama.cpp +++ b/examples/talk-llama/talk-llama.cpp @@ -391,6 +391,8 @@ int main(int argc, char ** argv) { prompt_llama = ::replace(prompt_llama, "{4}", chat_symb); + llama_batch batch = llama_batch_init(llama_n_ctx(ctx_llama), 0, 1); + // init session std::string path_session = params.path_session; std::vector session_tokens; @@ -426,8 +428,21 @@ int main(int argc, char ** argv) { printf("\n"); printf("%s : initializing - please wait ...\n", __func__); - if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0)) { - fprintf(stderr, "%s : failed to eval\n", __func__); + // prepare batch + { + batch.n_tokens = embd_inp.size(); + + for (int i = 0; i < batch.n_tokens; i++) { + batch.token[i] = embd_inp[i]; + batch.pos[i] = i; + batch.n_seq_id[i] = 1; + batch.seq_id[i][0] = 0; + batch.logits[i] = i == batch.n_tokens - 1; + } + } + + if (llama_decode(ctx_llama, batch)) { + fprintf(stderr, "%s : failed to decode\n", __func__); return 1; } @@ -647,8 +662,21 @@ int main(int argc, char ** argv) { n_session_consumed = session_tokens.size(); } - if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past)) { - fprintf(stderr, "%s : failed to eval\n", __func__); + // prepare batch + { + batch.n_tokens = embd.size(); + + for (int i = 0; i < batch.n_tokens; i++) { + batch.token[i] = embd[i]; + batch.pos[i] = n_past + i; + batch.n_seq_id[i] = 1; + batch.seq_id[i][0] = 0; + batch.logits[i] = i == batch.n_tokens - 1; + } + } + + if (llama_decode(ctx_llama, batch)) { + fprintf(stderr, "%s : failed to decode\n", __func__); return 1; } } diff --git a/examples/talk-llama/unicode.h b/examples/talk-llama/unicode.h index 620e2b5..f6be454 100644 --- a/examples/talk-llama/unicode.h +++ b/examples/talk-llama/unicode.h @@ -1,6 +1,7 @@ #pragma once #include +#include #include #include #include @@ -223,266 +224,311 @@ static const std::vector> control_ranges = { {0x2B81E, 0x2B81F}, {0x2CEA2, 0x2CEAF}, {0x2EBE1, 0x2F7FF}, {0x2FA1E, 0x2FFFF}, {0x3134B, 0xE00FF}, {0xE01F0, 0x10FFFF}, }; -static const std::unordered_map> nfd_map = { -{0xC0, {0x41, 0x300}}, {0xC1, {0x41, 0x301}}, {0xC2, {0x41, 0x302}}, {0xC3, {0x41, 0x303}}, {0xC4, {0x41, 0x308}}, {0xC5, {0x41, 0x30A}}, {0xC7, {0x43, 0x327}}, {0xC8, {0x45, 0x300}}, -{0xC9, {0x45, 0x301}}, {0xCA, {0x45, 0x302}}, {0xCB, {0x45, 0x308}}, {0xCC, {0x49, 0x300}}, {0xCD, {0x49, 0x301}}, {0xCE, {0x49, 0x302}}, {0xCF, {0x49, 0x308}}, {0xD1, {0x4E, 0x303}}, -{0xD2, {0x4F, 0x300}}, {0xD3, {0x4F, 0x301}}, {0xD4, {0x4F, 0x302}}, {0xD5, {0x4F, 0x303}}, {0xD6, {0x4F, 0x308}}, {0xD9, {0x55, 0x300}}, {0xDA, {0x55, 0x301}}, {0xDB, {0x55, 0x302}}, -{0xDC, {0x55, 0x308}}, {0xDD, {0x59, 0x301}}, {0xE0, {0x61, 0x300}}, {0xE1, {0x61, 0x301}}, {0xE2, {0x61, 0x302}}, {0xE3, {0x61, 0x303}}, {0xE4, {0x61, 0x308}}, {0xE5, {0x61, 0x30A}}, -{0xE7, {0x63, 0x327}}, {0xE8, {0x65, 0x300}}, {0xE9, {0x65, 0x301}}, {0xEA, {0x65, 0x302}}, {0xEB, {0x65, 0x308}}, {0xEC, {0x69, 0x300}}, {0xED, {0x69, 0x301}}, {0xEE, {0x69, 0x302}}, -{0xEF, {0x69, 0x308}}, {0xF1, {0x6E, 0x303}}, {0xF2, {0x6F, 0x300}}, {0xF3, {0x6F, 0x301}}, {0xF4, {0x6F, 0x302}}, {0xF5, {0x6F, 0x303}}, {0xF6, {0x6F, 0x308}}, {0xF9, {0x75, 0x300}}, -{0xFA, {0x75, 0x301}}, {0xFB, {0x75, 0x302}}, {0xFC, {0x75, 0x308}}, {0xFD, {0x79, 0x301}}, {0xFF, {0x79, 0x308}}, {0x100, {0x41, 0x304}}, {0x101, {0x61, 0x304}}, {0x102, {0x41, 0x306}}, -{0x103, {0x61, 0x306}}, {0x104, {0x41, 0x328}}, {0x105, {0x61, 0x328}}, {0x106, {0x43, 0x301}}, {0x107, {0x63, 0x301}}, {0x108, {0x43, 0x302}}, {0x109, {0x63, 0x302}}, {0x10A, {0x43, 0x307}}, -{0x10B, {0x63, 0x307}}, {0x10C, {0x43, 0x30C}}, {0x10D, {0x63, 0x30C}}, {0x10E, {0x44, 0x30C}}, {0x10F, {0x64, 0x30C}}, {0x112, {0x45, 0x304}}, {0x113, {0x65, 0x304}}, {0x114, {0x45, 0x306}}, -{0x115, {0x65, 0x306}}, {0x116, {0x45, 0x307}}, {0x117, {0x65, 0x307}}, {0x118, {0x45, 0x328}}, {0x119, {0x65, 0x328}}, {0x11A, {0x45, 0x30C}}, {0x11B, {0x65, 0x30C}}, {0x11C, {0x47, 0x302}}, -{0x11D, {0x67, 0x302}}, {0x11E, {0x47, 0x306}}, {0x11F, {0x67, 0x306}}, {0x120, {0x47, 0x307}}, {0x121, {0x67, 0x307}}, {0x122, {0x47, 0x327}}, {0x123, {0x67, 0x327}}, {0x124, {0x48, 0x302}}, -{0x125, {0x68, 0x302}}, {0x128, {0x49, 0x303}}, {0x129, {0x69, 0x303}}, {0x12A, {0x49, 0x304}}, {0x12B, {0x69, 0x304}}, {0x12C, {0x49, 0x306}}, {0x12D, {0x69, 0x306}}, {0x12E, {0x49, 0x328}}, -{0x12F, {0x69, 0x328}}, {0x130, {0x49, 0x307}}, {0x134, {0x4A, 0x302}}, {0x135, {0x6A, 0x302}}, {0x136, {0x4B, 0x327}}, {0x137, {0x6B, 0x327}}, {0x139, {0x4C, 0x301}}, {0x13A, {0x6C, 0x301}}, -{0x13B, {0x4C, 0x327}}, {0x13C, {0x6C, 0x327}}, {0x13D, {0x4C, 0x30C}}, {0x13E, {0x6C, 0x30C}}, {0x143, {0x4E, 0x301}}, {0x144, {0x6E, 0x301}}, {0x145, {0x4E, 0x327}}, {0x146, {0x6E, 0x327}}, -{0x147, {0x4E, 0x30C}}, {0x148, {0x6E, 0x30C}}, {0x14C, {0x4F, 0x304}}, {0x14D, {0x6F, 0x304}}, {0x14E, {0x4F, 0x306}}, {0x14F, {0x6F, 0x306}}, {0x150, {0x4F, 0x30B}}, {0x151, {0x6F, 0x30B}}, -{0x154, {0x52, 0x301}}, {0x155, {0x72, 0x301}}, {0x156, {0x52, 0x327}}, {0x157, {0x72, 0x327}}, {0x158, {0x52, 0x30C}}, {0x159, {0x72, 0x30C}}, {0x15A, {0x53, 0x301}}, {0x15B, {0x73, 0x301}}, -{0x15C, {0x53, 0x302}}, {0x15D, {0x73, 0x302}}, {0x15E, {0x53, 0x327}}, {0x15F, {0x73, 0x327}}, {0x160, {0x53, 0x30C}}, {0x161, {0x73, 0x30C}}, {0x162, {0x54, 0x327}}, {0x163, {0x74, 0x327}}, -{0x164, {0x54, 0x30C}}, {0x165, {0x74, 0x30C}}, {0x168, {0x55, 0x303}}, {0x169, {0x75, 0x303}}, {0x16A, {0x55, 0x304}}, {0x16B, {0x75, 0x304}}, {0x16C, {0x55, 0x306}}, {0x16D, {0x75, 0x306}}, -{0x16E, {0x55, 0x30A}}, {0x16F, {0x75, 0x30A}}, {0x170, {0x55, 0x30B}}, {0x171, {0x75, 0x30B}}, {0x172, {0x55, 0x328}}, {0x173, {0x75, 0x328}}, {0x174, {0x57, 0x302}}, {0x175, {0x77, 0x302}}, -{0x176, {0x59, 0x302}}, {0x177, {0x79, 0x302}}, {0x178, {0x59, 0x308}}, {0x179, {0x5A, 0x301}}, {0x17A, {0x7A, 0x301}}, {0x17B, {0x5A, 0x307}}, {0x17C, {0x7A, 0x307}}, {0x17D, {0x5A, 0x30C}}, -{0x17E, {0x7A, 0x30C}}, {0x1A0, {0x4F, 0x31B}}, {0x1A1, {0x6F, 0x31B}}, {0x1AF, {0x55, 0x31B}}, {0x1B0, {0x75, 0x31B}}, {0x1CD, {0x41, 0x30C}}, {0x1CE, {0x61, 0x30C}}, {0x1CF, {0x49, 0x30C}}, -{0x1D0, {0x69, 0x30C}}, {0x1D1, {0x4F, 0x30C}}, {0x1D2, {0x6F, 0x30C}}, {0x1D3, {0x55, 0x30C}}, {0x1D4, {0x75, 0x30C}}, {0x1D5, {0x55, 0x308, 0x304}}, {0x1D6, {0x75, 0x308, 0x304}}, -{0x1D7, {0x55, 0x308, 0x301}}, {0x1D8, {0x75, 0x308, 0x301}}, {0x1D9, {0x55, 0x308, 0x30C}}, {0x1DA, {0x75, 0x308, 0x30C}}, {0x1DB, {0x55, 0x308, 0x300}}, {0x1DC, {0x75, 0x308, 0x300}}, -{0x1DE, {0x41, 0x308, 0x304}}, {0x1DF, {0x61, 0x308, 0x304}}, {0x1E0, {0x41, 0x307, 0x304}}, {0x1E1, {0x61, 0x307, 0x304}}, {0x1E2, {0xC6, 0x304}}, {0x1E3, {0xE6, 0x304}}, {0x1E6, {0x47, 0x30C}}, -{0x1E7, {0x67, 0x30C}}, {0x1E8, {0x4B, 0x30C}}, {0x1E9, {0x6B, 0x30C}}, {0x1EA, {0x4F, 0x328}}, {0x1EB, {0x6F, 0x328}}, {0x1EC, {0x4F, 0x328, 0x304}}, {0x1ED, {0x6F, 0x328, 0x304}}, -{0x1EE, {0x1B7, 0x30C}}, {0x1EF, {0x292, 0x30C}}, {0x1F0, {0x6A, 0x30C}}, {0x1F4, {0x47, 0x301}}, {0x1F5, {0x67, 0x301}}, {0x1F8, {0x4E, 0x300}}, {0x1F9, {0x6E, 0x300}}, {0x1FA, {0x41, 0x30A, 0x301}}, -{0x1FB, {0x61, 0x30A, 0x301}}, {0x1FC, {0xC6, 0x301}}, {0x1FD, {0xE6, 0x301}}, {0x1FE, {0xD8, 0x301}}, {0x1FF, {0xF8, 0x301}}, {0x200, {0x41, 0x30F}}, {0x201, {0x61, 0x30F}}, {0x202, {0x41, 0x311}}, -{0x203, {0x61, 0x311}}, {0x204, {0x45, 0x30F}}, {0x205, {0x65, 0x30F}}, {0x206, {0x45, 0x311}}, {0x207, {0x65, 0x311}}, {0x208, {0x49, 0x30F}}, {0x209, {0x69, 0x30F}}, {0x20A, {0x49, 0x311}}, -{0x20B, {0x69, 0x311}}, {0x20C, {0x4F, 0x30F}}, {0x20D, {0x6F, 0x30F}}, {0x20E, {0x4F, 0x311}}, {0x20F, {0x6F, 0x311}}, {0x210, {0x52, 0x30F}}, {0x211, {0x72, 0x30F}}, {0x212, {0x52, 0x311}}, -{0x213, {0x72, 0x311}}, {0x214, {0x55, 0x30F}}, {0x215, {0x75, 0x30F}}, {0x216, {0x55, 0x311}}, {0x217, {0x75, 0x311}}, {0x218, {0x53, 0x326}}, {0x219, {0x73, 0x326}}, {0x21A, {0x54, 0x326}}, -{0x21B, {0x74, 0x326}}, {0x21E, {0x48, 0x30C}}, {0x21F, {0x68, 0x30C}}, {0x226, {0x41, 0x307}}, {0x227, {0x61, 0x307}}, {0x228, {0x45, 0x327}}, {0x229, {0x65, 0x327}}, {0x22A, {0x4F, 0x308, 0x304}}, -{0x22B, {0x6F, 0x308, 0x304}}, {0x22C, {0x4F, 0x303, 0x304}}, {0x22D, {0x6F, 0x303, 0x304}}, {0x22E, {0x4F, 0x307}}, {0x22F, {0x6F, 0x307}}, {0x230, {0x4F, 0x307, 0x304}}, -{0x231, {0x6F, 0x307, 0x304}}, {0x232, {0x59, 0x304}}, {0x233, {0x79, 0x304}}, {0x340, {0x300}}, {0x341, {0x301}}, {0x343, {0x313}}, {0x344, {0x308, 0x301}}, {0x374, {0x2B9}}, {0x37E, {0x3B}}, -{0x385, {0xA8, 0x301}}, {0x386, {0x391, 0x301}}, {0x387, {0xB7}}, {0x388, {0x395, 0x301}}, {0x389, {0x397, 0x301}}, {0x38A, {0x399, 0x301}}, {0x38C, {0x39F, 0x301}}, {0x38E, {0x3A5, 0x301}}, -{0x38F, {0x3A9, 0x301}}, {0x390, {0x3B9, 0x308, 0x301}}, {0x3AA, {0x399, 0x308}}, {0x3AB, {0x3A5, 0x308}}, {0x3AC, {0x3B1, 0x301}}, {0x3AD, {0x3B5, 0x301}}, {0x3AE, {0x3B7, 0x301}}, -{0x3AF, {0x3B9, 0x301}}, {0x3B0, {0x3C5, 0x308, 0x301}}, {0x3CA, {0x3B9, 0x308}}, {0x3CB, {0x3C5, 0x308}}, {0x3CC, {0x3BF, 0x301}}, {0x3CD, {0x3C5, 0x301}}, {0x3CE, {0x3C9, 0x301}}, -{0x3D3, {0x3D2, 0x301}}, {0x3D4, {0x3D2, 0x308}}, {0x400, {0x415, 0x300}}, {0x401, {0x415, 0x308}}, {0x403, {0x413, 0x301}}, {0x407, {0x406, 0x308}}, {0x40C, {0x41A, 0x301}}, {0x40D, {0x418, 0x300}}, -{0x40E, {0x423, 0x306}}, {0x419, {0x418, 0x306}}, {0x439, {0x438, 0x306}}, {0x450, {0x435, 0x300}}, {0x451, {0x435, 0x308}}, {0x453, {0x433, 0x301}}, {0x457, {0x456, 0x308}}, {0x45C, {0x43A, 0x301}}, -{0x45D, {0x438, 0x300}}, {0x45E, {0x443, 0x306}}, {0x476, {0x474, 0x30F}}, {0x477, {0x475, 0x30F}}, {0x4C1, {0x416, 0x306}}, {0x4C2, {0x436, 0x306}}, {0x4D0, {0x410, 0x306}}, {0x4D1, {0x430, 0x306}}, -{0x4D2, {0x410, 0x308}}, {0x4D3, {0x430, 0x308}}, {0x4D6, {0x415, 0x306}}, {0x4D7, {0x435, 0x306}}, {0x4DA, {0x4D8, 0x308}}, {0x4DB, {0x4D9, 0x308}}, {0x4DC, {0x416, 0x308}}, {0x4DD, {0x436, 0x308}}, -{0x4DE, {0x417, 0x308}}, {0x4DF, {0x437, 0x308}}, {0x4E2, {0x418, 0x304}}, {0x4E3, {0x438, 0x304}}, {0x4E4, {0x418, 0x308}}, {0x4E5, {0x438, 0x308}}, {0x4E6, {0x41E, 0x308}}, {0x4E7, {0x43E, 0x308}}, -{0x4EA, {0x4E8, 0x308}}, {0x4EB, {0x4E9, 0x308}}, {0x4EC, {0x42D, 0x308}}, {0x4ED, {0x44D, 0x308}}, {0x4EE, {0x423, 0x304}}, {0x4EF, {0x443, 0x304}}, {0x4F0, {0x423, 0x308}}, {0x4F1, {0x443, 0x308}}, -{0x4F2, {0x423, 0x30B}}, {0x4F3, {0x443, 0x30B}}, {0x4F4, {0x427, 0x308}}, {0x4F5, {0x447, 0x308}}, {0x4F8, {0x42B, 0x308}}, {0x4F9, {0x44B, 0x308}}, {0x622, {0x627, 0x653}}, {0x623, {0x627, 0x654}}, -{0x624, {0x648, 0x654}}, {0x625, {0x627, 0x655}}, {0x626, {0x64A, 0x654}}, {0x6C0, {0x6D5, 0x654}}, {0x6C2, {0x6C1, 0x654}}, {0x6D3, {0x6D2, 0x654}}, {0x929, {0x928, 0x93C}}, {0x931, {0x930, 0x93C}}, -{0x934, {0x933, 0x93C}}, {0x958, {0x915, 0x93C}}, {0x959, {0x916, 0x93C}}, {0x95A, {0x917, 0x93C}}, {0x95B, {0x91C, 0x93C}}, {0x95C, {0x921, 0x93C}}, {0x95D, {0x922, 0x93C}}, {0x95E, {0x92B, 0x93C}}, -{0x95F, {0x92F, 0x93C}}, {0x9CB, {0x9C7, 0x9BE}}, {0x9CC, {0x9C7, 0x9D7}}, {0x9DC, {0x9A1, 0x9BC}}, {0x9DD, {0x9A2, 0x9BC}}, {0x9DF, {0x9AF, 0x9BC}}, {0xA33, {0xA32, 0xA3C}}, {0xA36, {0xA38, 0xA3C}}, -{0xA59, {0xA16, 0xA3C}}, {0xA5A, {0xA17, 0xA3C}}, {0xA5B, {0xA1C, 0xA3C}}, {0xA5E, {0xA2B, 0xA3C}}, {0xB48, {0xB47, 0xB56}}, {0xB4B, {0xB47, 0xB3E}}, {0xB4C, {0xB47, 0xB57}}, {0xB5C, {0xB21, 0xB3C}}, -{0xB5D, {0xB22, 0xB3C}}, {0xB94, {0xB92, 0xBD7}}, {0xBCA, {0xBC6, 0xBBE}}, {0xBCB, {0xBC7, 0xBBE}}, {0xBCC, {0xBC6, 0xBD7}}, {0xC48, {0xC46, 0xC56}}, {0xCC0, {0xCBF, 0xCD5}}, {0xCC7, {0xCC6, 0xCD5}}, -{0xCC8, {0xCC6, 0xCD6}}, {0xCCA, {0xCC6, 0xCC2}}, {0xCCB, {0xCC6, 0xCC2, 0xCD5}}, {0xD4A, {0xD46, 0xD3E}}, {0xD4B, {0xD47, 0xD3E}}, {0xD4C, {0xD46, 0xD57}}, {0xDDA, {0xDD9, 0xDCA}}, -{0xDDC, {0xDD9, 0xDCF}}, {0xDDD, {0xDD9, 0xDCF, 0xDCA}}, {0xDDE, {0xDD9, 0xDDF}}, {0xF43, {0xF42, 0xFB7}}, {0xF4D, {0xF4C, 0xFB7}}, {0xF52, {0xF51, 0xFB7}}, {0xF57, {0xF56, 0xFB7}}, -{0xF5C, {0xF5B, 0xFB7}}, {0xF69, {0xF40, 0xFB5}}, {0xF73, {0xF71, 0xF72}}, {0xF75, {0xF71, 0xF74}}, {0xF76, {0xFB2, 0xF80}}, {0xF78, {0xFB3, 0xF80}}, {0xF81, {0xF71, 0xF80}}, {0xF93, {0xF92, 0xFB7}}, -{0xF9D, {0xF9C, 0xFB7}}, {0xFA2, {0xFA1, 0xFB7}}, {0xFA7, {0xFA6, 0xFB7}}, {0xFAC, {0xFAB, 0xFB7}}, {0xFB9, {0xF90, 0xFB5}}, {0x1026, {0x1025, 0x102E}}, {0x1B06, {0x1B05, 0x1B35}}, -{0x1B08, {0x1B07, 0x1B35}}, {0x1B0A, {0x1B09, 0x1B35}}, {0x1B0C, {0x1B0B, 0x1B35}}, {0x1B0E, {0x1B0D, 0x1B35}}, {0x1B12, {0x1B11, 0x1B35}}, {0x1B3B, {0x1B3A, 0x1B35}}, {0x1B3D, {0x1B3C, 0x1B35}}, -{0x1B40, {0x1B3E, 0x1B35}}, {0x1B41, {0x1B3F, 0x1B35}}, {0x1B43, {0x1B42, 0x1B35}}, {0x1E00, {0x41, 0x325}}, {0x1E01, {0x61, 0x325}}, {0x1E02, {0x42, 0x307}}, {0x1E03, {0x62, 0x307}}, -{0x1E04, {0x42, 0x323}}, {0x1E05, {0x62, 0x323}}, {0x1E06, {0x42, 0x331}}, {0x1E07, {0x62, 0x331}}, {0x1E08, {0x43, 0x327, 0x301}}, {0x1E09, {0x63, 0x327, 0x301}}, {0x1E0A, {0x44, 0x307}}, -{0x1E0B, {0x64, 0x307}}, {0x1E0C, {0x44, 0x323}}, {0x1E0D, {0x64, 0x323}}, {0x1E0E, {0x44, 0x331}}, {0x1E0F, {0x64, 0x331}}, {0x1E10, {0x44, 0x327}}, {0x1E11, {0x64, 0x327}}, {0x1E12, {0x44, 0x32D}}, -{0x1E13, {0x64, 0x32D}}, {0x1E14, {0x45, 0x304, 0x300}}, {0x1E15, {0x65, 0x304, 0x300}}, {0x1E16, {0x45, 0x304, 0x301}}, {0x1E17, {0x65, 0x304, 0x301}}, {0x1E18, {0x45, 0x32D}}, -{0x1E19, {0x65, 0x32D}}, {0x1E1A, {0x45, 0x330}}, {0x1E1B, {0x65, 0x330}}, {0x1E1C, {0x45, 0x327, 0x306}}, {0x1E1D, {0x65, 0x327, 0x306}}, {0x1E1E, {0x46, 0x307}}, {0x1E1F, {0x66, 0x307}}, -{0x1E20, {0x47, 0x304}}, {0x1E21, {0x67, 0x304}}, {0x1E22, {0x48, 0x307}}, {0x1E23, {0x68, 0x307}}, {0x1E24, {0x48, 0x323}}, {0x1E25, {0x68, 0x323}}, {0x1E26, {0x48, 0x308}}, {0x1E27, {0x68, 0x308}}, -{0x1E28, {0x48, 0x327}}, {0x1E29, {0x68, 0x327}}, {0x1E2A, {0x48, 0x32E}}, {0x1E2B, {0x68, 0x32E}}, {0x1E2C, {0x49, 0x330}}, {0x1E2D, {0x69, 0x330}}, {0x1E2E, {0x49, 0x308, 0x301}}, -{0x1E2F, {0x69, 0x308, 0x301}}, {0x1E30, {0x4B, 0x301}}, {0x1E31, {0x6B, 0x301}}, {0x1E32, {0x4B, 0x323}}, {0x1E33, {0x6B, 0x323}}, {0x1E34, {0x4B, 0x331}}, {0x1E35, {0x6B, 0x331}}, -{0x1E36, {0x4C, 0x323}}, {0x1E37, {0x6C, 0x323}}, {0x1E38, {0x4C, 0x323, 0x304}}, {0x1E39, {0x6C, 0x323, 0x304}}, {0x1E3A, {0x4C, 0x331}}, {0x1E3B, {0x6C, 0x331}}, {0x1E3C, {0x4C, 0x32D}}, -{0x1E3D, {0x6C, 0x32D}}, {0x1E3E, {0x4D, 0x301}}, {0x1E3F, {0x6D, 0x301}}, {0x1E40, {0x4D, 0x307}}, {0x1E41, {0x6D, 0x307}}, {0x1E42, {0x4D, 0x323}}, {0x1E43, {0x6D, 0x323}}, {0x1E44, {0x4E, 0x307}}, -{0x1E45, {0x6E, 0x307}}, {0x1E46, {0x4E, 0x323}}, {0x1E47, {0x6E, 0x323}}, {0x1E48, {0x4E, 0x331}}, {0x1E49, {0x6E, 0x331}}, {0x1E4A, {0x4E, 0x32D}}, {0x1E4B, {0x6E, 0x32D}}, -{0x1E4C, {0x4F, 0x303, 0x301}}, {0x1E4D, {0x6F, 0x303, 0x301}}, {0x1E4E, {0x4F, 0x303, 0x308}}, {0x1E4F, {0x6F, 0x303, 0x308}}, {0x1E50, {0x4F, 0x304, 0x300}}, {0x1E51, {0x6F, 0x304, 0x300}}, -{0x1E52, {0x4F, 0x304, 0x301}}, {0x1E53, {0x6F, 0x304, 0x301}}, {0x1E54, {0x50, 0x301}}, {0x1E55, {0x70, 0x301}}, {0x1E56, {0x50, 0x307}}, {0x1E57, {0x70, 0x307}}, {0x1E58, {0x52, 0x307}}, -{0x1E59, {0x72, 0x307}}, {0x1E5A, {0x52, 0x323}}, {0x1E5B, {0x72, 0x323}}, {0x1E5C, {0x52, 0x323, 0x304}}, {0x1E5D, {0x72, 0x323, 0x304}}, {0x1E5E, {0x52, 0x331}}, {0x1E5F, {0x72, 0x331}}, -{0x1E60, {0x53, 0x307}}, {0x1E61, {0x73, 0x307}}, {0x1E62, {0x53, 0x323}}, {0x1E63, {0x73, 0x323}}, {0x1E64, {0x53, 0x301, 0x307}}, {0x1E65, {0x73, 0x301, 0x307}}, {0x1E66, {0x53, 0x30C, 0x307}}, -{0x1E67, {0x73, 0x30C, 0x307}}, {0x1E68, {0x53, 0x323, 0x307}}, {0x1E69, {0x73, 0x323, 0x307}}, {0x1E6A, {0x54, 0x307}}, {0x1E6B, {0x74, 0x307}}, {0x1E6C, {0x54, 0x323}}, {0x1E6D, {0x74, 0x323}}, -{0x1E6E, {0x54, 0x331}}, {0x1E6F, {0x74, 0x331}}, {0x1E70, {0x54, 0x32D}}, {0x1E71, {0x74, 0x32D}}, {0x1E72, {0x55, 0x324}}, {0x1E73, {0x75, 0x324}}, {0x1E74, {0x55, 0x330}}, {0x1E75, {0x75, 0x330}}, -{0x1E76, {0x55, 0x32D}}, {0x1E77, {0x75, 0x32D}}, {0x1E78, {0x55, 0x303, 0x301}}, {0x1E79, {0x75, 0x303, 0x301}}, {0x1E7A, {0x55, 0x304, 0x308}}, {0x1E7B, {0x75, 0x304, 0x308}}, -{0x1E7C, {0x56, 0x303}}, {0x1E7D, {0x76, 0x303}}, {0x1E7E, {0x56, 0x323}}, {0x1E7F, {0x76, 0x323}}, {0x1E80, {0x57, 0x300}}, {0x1E81, {0x77, 0x300}}, {0x1E82, {0x57, 0x301}}, {0x1E83, {0x77, 0x301}}, -{0x1E84, {0x57, 0x308}}, {0x1E85, {0x77, 0x308}}, {0x1E86, {0x57, 0x307}}, {0x1E87, {0x77, 0x307}}, {0x1E88, {0x57, 0x323}}, {0x1E89, {0x77, 0x323}}, {0x1E8A, {0x58, 0x307}}, {0x1E8B, {0x78, 0x307}}, -{0x1E8C, {0x58, 0x308}}, {0x1E8D, {0x78, 0x308}}, {0x1E8E, {0x59, 0x307}}, {0x1E8F, {0x79, 0x307}}, {0x1E90, {0x5A, 0x302}}, {0x1E91, {0x7A, 0x302}}, {0x1E92, {0x5A, 0x323}}, {0x1E93, {0x7A, 0x323}}, -{0x1E94, {0x5A, 0x331}}, {0x1E95, {0x7A, 0x331}}, {0x1E96, {0x68, 0x331}}, {0x1E97, {0x74, 0x308}}, {0x1E98, {0x77, 0x30A}}, {0x1E99, {0x79, 0x30A}}, {0x1E9B, {0x17F, 0x307}}, {0x1EA0, {0x41, 0x323}}, -{0x1EA1, {0x61, 0x323}}, {0x1EA2, {0x41, 0x309}}, {0x1EA3, {0x61, 0x309}}, {0x1EA4, {0x41, 0x302, 0x301}}, {0x1EA5, {0x61, 0x302, 0x301}}, {0x1EA6, {0x41, 0x302, 0x300}}, -{0x1EA7, {0x61, 0x302, 0x300}}, {0x1EA8, {0x41, 0x302, 0x309}}, {0x1EA9, {0x61, 0x302, 0x309}}, {0x1EAA, {0x41, 0x302, 0x303}}, {0x1EAB, {0x61, 0x302, 0x303}}, {0x1EAC, {0x41, 0x323, 0x302}}, -{0x1EAD, {0x61, 0x323, 0x302}}, {0x1EAE, {0x41, 0x306, 0x301}}, {0x1EAF, {0x61, 0x306, 0x301}}, {0x1EB0, {0x41, 0x306, 0x300}}, {0x1EB1, {0x61, 0x306, 0x300}}, {0x1EB2, {0x41, 0x306, 0x309}}, -{0x1EB3, {0x61, 0x306, 0x309}}, {0x1EB4, {0x41, 0x306, 0x303}}, {0x1EB5, {0x61, 0x306, 0x303}}, {0x1EB6, {0x41, 0x323, 0x306}}, {0x1EB7, {0x61, 0x323, 0x306}}, {0x1EB8, {0x45, 0x323}}, -{0x1EB9, {0x65, 0x323}}, {0x1EBA, {0x45, 0x309}}, {0x1EBB, {0x65, 0x309}}, {0x1EBC, {0x45, 0x303}}, {0x1EBD, {0x65, 0x303}}, {0x1EBE, {0x45, 0x302, 0x301}}, {0x1EBF, {0x65, 0x302, 0x301}}, -{0x1EC0, {0x45, 0x302, 0x300}}, {0x1EC1, {0x65, 0x302, 0x300}}, {0x1EC2, {0x45, 0x302, 0x309}}, {0x1EC3, {0x65, 0x302, 0x309}}, {0x1EC4, {0x45, 0x302, 0x303}}, {0x1EC5, {0x65, 0x302, 0x303}}, -{0x1EC6, {0x45, 0x323, 0x302}}, {0x1EC7, {0x65, 0x323, 0x302}}, {0x1EC8, {0x49, 0x309}}, {0x1EC9, {0x69, 0x309}}, {0x1ECA, {0x49, 0x323}}, {0x1ECB, {0x69, 0x323}}, {0x1ECC, {0x4F, 0x323}}, -{0x1ECD, {0x6F, 0x323}}, {0x1ECE, {0x4F, 0x309}}, {0x1ECF, {0x6F, 0x309}}, {0x1ED0, {0x4F, 0x302, 0x301}}, {0x1ED1, {0x6F, 0x302, 0x301}}, {0x1ED2, {0x4F, 0x302, 0x300}}, -{0x1ED3, {0x6F, 0x302, 0x300}}, {0x1ED4, {0x4F, 0x302, 0x309}}, {0x1ED5, {0x6F, 0x302, 0x309}}, {0x1ED6, {0x4F, 0x302, 0x303}}, {0x1ED7, {0x6F, 0x302, 0x303}}, {0x1ED8, {0x4F, 0x323, 0x302}}, -{0x1ED9, {0x6F, 0x323, 0x302}}, {0x1EDA, {0x4F, 0x31B, 0x301}}, {0x1EDB, {0x6F, 0x31B, 0x301}}, {0x1EDC, {0x4F, 0x31B, 0x300}}, {0x1EDD, {0x6F, 0x31B, 0x300}}, {0x1EDE, {0x4F, 0x31B, 0x309}}, -{0x1EDF, {0x6F, 0x31B, 0x309}}, {0x1EE0, {0x4F, 0x31B, 0x303}}, {0x1EE1, {0x6F, 0x31B, 0x303}}, {0x1EE2, {0x4F, 0x31B, 0x323}}, {0x1EE3, {0x6F, 0x31B, 0x323}}, {0x1EE4, {0x55, 0x323}}, -{0x1EE5, {0x75, 0x323}}, {0x1EE6, {0x55, 0x309}}, {0x1EE7, {0x75, 0x309}}, {0x1EE8, {0x55, 0x31B, 0x301}}, {0x1EE9, {0x75, 0x31B, 0x301}}, {0x1EEA, {0x55, 0x31B, 0x300}}, -{0x1EEB, {0x75, 0x31B, 0x300}}, {0x1EEC, {0x55, 0x31B, 0x309}}, {0x1EED, {0x75, 0x31B, 0x309}}, {0x1EEE, {0x55, 0x31B, 0x303}}, {0x1EEF, {0x75, 0x31B, 0x303}}, {0x1EF0, {0x55, 0x31B, 0x323}}, -{0x1EF1, {0x75, 0x31B, 0x323}}, {0x1EF2, {0x59, 0x300}}, {0x1EF3, {0x79, 0x300}}, {0x1EF4, {0x59, 0x323}}, {0x1EF5, {0x79, 0x323}}, {0x1EF6, {0x59, 0x309}}, {0x1EF7, {0x79, 0x309}}, -{0x1EF8, {0x59, 0x303}}, {0x1EF9, {0x79, 0x303}}, {0x1F00, {0x3B1, 0x313}}, {0x1F01, {0x3B1, 0x314}}, {0x1F02, {0x3B1, 0x313, 0x300}}, {0x1F03, {0x3B1, 0x314, 0x300}}, {0x1F04, {0x3B1, 0x313, 0x301}}, -{0x1F05, {0x3B1, 0x314, 0x301}}, {0x1F06, {0x3B1, 0x313, 0x342}}, {0x1F07, {0x3B1, 0x314, 0x342}}, {0x1F08, {0x391, 0x313}}, {0x1F09, {0x391, 0x314}}, {0x1F0A, {0x391, 0x313, 0x300}}, -{0x1F0B, {0x391, 0x314, 0x300}}, {0x1F0C, {0x391, 0x313, 0x301}}, {0x1F0D, {0x391, 0x314, 0x301}}, {0x1F0E, {0x391, 0x313, 0x342}}, {0x1F0F, {0x391, 0x314, 0x342}}, {0x1F10, {0x3B5, 0x313}}, -{0x1F11, {0x3B5, 0x314}}, {0x1F12, {0x3B5, 0x313, 0x300}}, {0x1F13, {0x3B5, 0x314, 0x300}}, {0x1F14, {0x3B5, 0x313, 0x301}}, {0x1F15, {0x3B5, 0x314, 0x301}}, {0x1F18, {0x395, 0x313}}, -{0x1F19, {0x395, 0x314}}, {0x1F1A, {0x395, 0x313, 0x300}}, {0x1F1B, {0x395, 0x314, 0x300}}, {0x1F1C, {0x395, 0x313, 0x301}}, {0x1F1D, {0x395, 0x314, 0x301}}, {0x1F20, {0x3B7, 0x313}}, -{0x1F21, {0x3B7, 0x314}}, {0x1F22, {0x3B7, 0x313, 0x300}}, {0x1F23, {0x3B7, 0x314, 0x300}}, {0x1F24, {0x3B7, 0x313, 0x301}}, {0x1F25, {0x3B7, 0x314, 0x301}}, {0x1F26, {0x3B7, 0x313, 0x342}}, -{0x1F27, {0x3B7, 0x314, 0x342}}, {0x1F28, {0x397, 0x313}}, {0x1F29, {0x397, 0x314}}, {0x1F2A, {0x397, 0x313, 0x300}}, {0x1F2B, {0x397, 0x314, 0x300}}, {0x1F2C, {0x397, 0x313, 0x301}}, -{0x1F2D, {0x397, 0x314, 0x301}}, {0x1F2E, {0x397, 0x313, 0x342}}, {0x1F2F, {0x397, 0x314, 0x342}}, {0x1F30, {0x3B9, 0x313}}, {0x1F31, {0x3B9, 0x314}}, {0x1F32, {0x3B9, 0x313, 0x300}}, -{0x1F33, {0x3B9, 0x314, 0x300}}, {0x1F34, {0x3B9, 0x313, 0x301}}, {0x1F35, {0x3B9, 0x314, 0x301}}, {0x1F36, {0x3B9, 0x313, 0x342}}, {0x1F37, {0x3B9, 0x314, 0x342}}, {0x1F38, {0x399, 0x313}}, -{0x1F39, {0x399, 0x314}}, {0x1F3A, {0x399, 0x313, 0x300}}, {0x1F3B, {0x399, 0x314, 0x300}}, {0x1F3C, {0x399, 0x313, 0x301}}, {0x1F3D, {0x399, 0x314, 0x301}}, {0x1F3E, {0x399, 0x313, 0x342}}, -{0x1F3F, {0x399, 0x314, 0x342}}, {0x1F40, {0x3BF, 0x313}}, {0x1F41, {0x3BF, 0x314}}, {0x1F42, {0x3BF, 0x313, 0x300}}, {0x1F43, {0x3BF, 0x314, 0x300}}, {0x1F44, {0x3BF, 0x313, 0x301}}, -{0x1F45, {0x3BF, 0x314, 0x301}}, {0x1F48, {0x39F, 0x313}}, {0x1F49, {0x39F, 0x314}}, {0x1F4A, {0x39F, 0x313, 0x300}}, {0x1F4B, {0x39F, 0x314, 0x300}}, {0x1F4C, {0x39F, 0x313, 0x301}}, -{0x1F4D, {0x39F, 0x314, 0x301}}, {0x1F50, {0x3C5, 0x313}}, {0x1F51, {0x3C5, 0x314}}, {0x1F52, {0x3C5, 0x313, 0x300}}, {0x1F53, {0x3C5, 0x314, 0x300}}, {0x1F54, {0x3C5, 0x313, 0x301}}, -{0x1F55, {0x3C5, 0x314, 0x301}}, {0x1F56, {0x3C5, 0x313, 0x342}}, {0x1F57, {0x3C5, 0x314, 0x342}}, {0x1F59, {0x3A5, 0x314}}, {0x1F5B, {0x3A5, 0x314, 0x300}}, {0x1F5D, {0x3A5, 0x314, 0x301}}, -{0x1F5F, {0x3A5, 0x314, 0x342}}, {0x1F60, {0x3C9, 0x313}}, {0x1F61, {0x3C9, 0x314}}, {0x1F62, {0x3C9, 0x313, 0x300}}, {0x1F63, {0x3C9, 0x314, 0x300}}, {0x1F64, {0x3C9, 0x313, 0x301}}, -{0x1F65, {0x3C9, 0x314, 0x301}}, {0x1F66, {0x3C9, 0x313, 0x342}}, {0x1F67, {0x3C9, 0x314, 0x342}}, {0x1F68, {0x3A9, 0x313}}, {0x1F69, {0x3A9, 0x314}}, {0x1F6A, {0x3A9, 0x313, 0x300}}, -{0x1F6B, {0x3A9, 0x314, 0x300}}, {0x1F6C, {0x3A9, 0x313, 0x301}}, {0x1F6D, {0x3A9, 0x314, 0x301}}, {0x1F6E, {0x3A9, 0x313, 0x342}}, {0x1F6F, {0x3A9, 0x314, 0x342}}, {0x1F70, {0x3B1, 0x300}}, -{0x1F71, {0x3B1, 0x301}}, {0x1F72, {0x3B5, 0x300}}, {0x1F73, {0x3B5, 0x301}}, {0x1F74, {0x3B7, 0x300}}, {0x1F75, {0x3B7, 0x301}}, {0x1F76, {0x3B9, 0x300}}, {0x1F77, {0x3B9, 0x301}}, -{0x1F78, {0x3BF, 0x300}}, {0x1F79, {0x3BF, 0x301}}, {0x1F7A, {0x3C5, 0x300}}, {0x1F7B, {0x3C5, 0x301}}, {0x1F7C, {0x3C9, 0x300}}, {0x1F7D, {0x3C9, 0x301}}, {0x1F80, {0x3B1, 0x313, 0x345}}, -{0x1F81, {0x3B1, 0x314, 0x345}}, {0x1F82, {0x3B1, 0x313, 0x300, 0x345}}, {0x1F83, {0x3B1, 0x314, 0x300, 0x345}}, {0x1F84, {0x3B1, 0x313, 0x301, 0x345}}, {0x1F85, {0x3B1, 0x314, 0x301, 0x345}}, -{0x1F86, {0x3B1, 0x313, 0x342, 0x345}}, {0x1F87, {0x3B1, 0x314, 0x342, 0x345}}, {0x1F88, {0x391, 0x313, 0x345}}, {0x1F89, {0x391, 0x314, 0x345}}, {0x1F8A, {0x391, 0x313, 0x300, 0x345}}, -{0x1F8B, {0x391, 0x314, 0x300, 0x345}}, {0x1F8C, {0x391, 0x313, 0x301, 0x345}}, {0x1F8D, {0x391, 0x314, 0x301, 0x345}}, {0x1F8E, {0x391, 0x313, 0x342, 0x345}}, {0x1F8F, {0x391, 0x314, 0x342, 0x345}}, -{0x1F90, {0x3B7, 0x313, 0x345}}, {0x1F91, {0x3B7, 0x314, 0x345}}, {0x1F92, {0x3B7, 0x313, 0x300, 0x345}}, {0x1F93, {0x3B7, 0x314, 0x300, 0x345}}, {0x1F94, {0x3B7, 0x313, 0x301, 0x345}}, -{0x1F95, {0x3B7, 0x314, 0x301, 0x345}}, {0x1F96, {0x3B7, 0x313, 0x342, 0x345}}, {0x1F97, {0x3B7, 0x314, 0x342, 0x345}}, {0x1F98, {0x397, 0x313, 0x345}}, {0x1F99, {0x397, 0x314, 0x345}}, -{0x1F9A, {0x397, 0x313, 0x300, 0x345}}, {0x1F9B, {0x397, 0x314, 0x300, 0x345}}, {0x1F9C, {0x397, 0x313, 0x301, 0x345}}, {0x1F9D, {0x397, 0x314, 0x301, 0x345}}, {0x1F9E, {0x397, 0x313, 0x342, 0x345}}, -{0x1F9F, {0x397, 0x314, 0x342, 0x345}}, {0x1FA0, {0x3C9, 0x313, 0x345}}, {0x1FA1, {0x3C9, 0x314, 0x345}}, {0x1FA2, {0x3C9, 0x313, 0x300, 0x345}}, {0x1FA3, {0x3C9, 0x314, 0x300, 0x345}}, -{0x1FA4, {0x3C9, 0x313, 0x301, 0x345}}, {0x1FA5, {0x3C9, 0x314, 0x301, 0x345}}, {0x1FA6, {0x3C9, 0x313, 0x342, 0x345}}, {0x1FA7, {0x3C9, 0x314, 0x342, 0x345}}, {0x1FA8, {0x3A9, 0x313, 0x345}}, -{0x1FA9, {0x3A9, 0x314, 0x345}}, {0x1FAA, {0x3A9, 0x313, 0x300, 0x345}}, {0x1FAB, {0x3A9, 0x314, 0x300, 0x345}}, {0x1FAC, {0x3A9, 0x313, 0x301, 0x345}}, {0x1FAD, {0x3A9, 0x314, 0x301, 0x345}}, -{0x1FAE, {0x3A9, 0x313, 0x342, 0x345}}, {0x1FAF, {0x3A9, 0x314, 0x342, 0x345}}, {0x1FB0, {0x3B1, 0x306}}, {0x1FB1, {0x3B1, 0x304}}, {0x1FB2, {0x3B1, 0x300, 0x345}}, {0x1FB3, {0x3B1, 0x345}}, -{0x1FB4, {0x3B1, 0x301, 0x345}}, {0x1FB6, {0x3B1, 0x342}}, {0x1FB7, {0x3B1, 0x342, 0x345}}, {0x1FB8, {0x391, 0x306}}, {0x1FB9, {0x391, 0x304}}, {0x1FBA, {0x391, 0x300}}, {0x1FBB, {0x391, 0x301}}, -{0x1FBC, {0x391, 0x345}}, {0x1FBE, {0x3B9}}, {0x1FC1, {0xA8, 0x342}}, {0x1FC2, {0x3B7, 0x300, 0x345}}, {0x1FC3, {0x3B7, 0x345}}, {0x1FC4, {0x3B7, 0x301, 0x345}}, {0x1FC6, {0x3B7, 0x342}}, -{0x1FC7, {0x3B7, 0x342, 0x345}}, {0x1FC8, {0x395, 0x300}}, {0x1FC9, {0x395, 0x301}}, {0x1FCA, {0x397, 0x300}}, {0x1FCB, {0x397, 0x301}}, {0x1FCC, {0x397, 0x345}}, {0x1FCD, {0x1FBF, 0x300}}, -{0x1FCE, {0x1FBF, 0x301}}, {0x1FCF, {0x1FBF, 0x342}}, {0x1FD0, {0x3B9, 0x306}}, {0x1FD1, {0x3B9, 0x304}}, {0x1FD2, {0x3B9, 0x308, 0x300}}, {0x1FD3, {0x3B9, 0x308, 0x301}}, {0x1FD6, {0x3B9, 0x342}}, -{0x1FD7, {0x3B9, 0x308, 0x342}}, {0x1FD8, {0x399, 0x306}}, {0x1FD9, {0x399, 0x304}}, {0x1FDA, {0x399, 0x300}}, {0x1FDB, {0x399, 0x301}}, {0x1FDD, {0x1FFE, 0x300}}, {0x1FDE, {0x1FFE, 0x301}}, -{0x1FDF, {0x1FFE, 0x342}}, {0x1FE0, {0x3C5, 0x306}}, {0x1FE1, {0x3C5, 0x304}}, {0x1FE2, {0x3C5, 0x308, 0x300}}, {0x1FE3, {0x3C5, 0x308, 0x301}}, {0x1FE4, {0x3C1, 0x313}}, {0x1FE5, {0x3C1, 0x314}}, -{0x1FE6, {0x3C5, 0x342}}, {0x1FE7, {0x3C5, 0x308, 0x342}}, {0x1FE8, {0x3A5, 0x306}}, {0x1FE9, {0x3A5, 0x304}}, {0x1FEA, {0x3A5, 0x300}}, {0x1FEB, {0x3A5, 0x301}}, {0x1FEC, {0x3A1, 0x314}}, -{0x1FED, {0xA8, 0x300}}, {0x1FEE, {0xA8, 0x301}}, {0x1FEF, {0x60}}, {0x1FF2, {0x3C9, 0x300, 0x345}}, {0x1FF3, {0x3C9, 0x345}}, {0x1FF4, {0x3C9, 0x301, 0x345}}, {0x1FF6, {0x3C9, 0x342}}, -{0x1FF7, {0x3C9, 0x342, 0x345}}, {0x1FF8, {0x39F, 0x300}}, {0x1FF9, {0x39F, 0x301}}, {0x1FFA, {0x3A9, 0x300}}, {0x1FFB, {0x3A9, 0x301}}, {0x1FFC, {0x3A9, 0x345}}, {0x1FFD, {0xB4}}, {0x2000, {0x2002}}, -{0x2001, {0x2003}}, {0x2126, {0x3A9}}, {0x212A, {0x4B}}, {0x212B, {0x41, 0x30A}}, {0x219A, {0x2190, 0x338}}, {0x219B, {0x2192, 0x338}}, {0x21AE, {0x2194, 0x338}}, {0x21CD, {0x21D0, 0x338}}, -{0x21CE, {0x21D4, 0x338}}, {0x21CF, {0x21D2, 0x338}}, {0x2204, {0x2203, 0x338}}, {0x2209, {0x2208, 0x338}}, {0x220C, {0x220B, 0x338}}, {0x2224, {0x2223, 0x338}}, {0x2226, {0x2225, 0x338}}, -{0x2241, {0x223C, 0x338}}, {0x2244, {0x2243, 0x338}}, {0x2247, {0x2245, 0x338}}, {0x2249, {0x2248, 0x338}}, {0x2260, {0x3D, 0x338}}, {0x2262, {0x2261, 0x338}}, {0x226D, {0x224D, 0x338}}, -{0x226E, {0x3C, 0x338}}, {0x226F, {0x3E, 0x338}}, {0x2270, {0x2264, 0x338}}, {0x2271, {0x2265, 0x338}}, {0x2274, {0x2272, 0x338}}, {0x2275, {0x2273, 0x338}}, {0x2278, {0x2276, 0x338}}, -{0x2279, {0x2277, 0x338}}, {0x2280, {0x227A, 0x338}}, {0x2281, {0x227B, 0x338}}, {0x2284, {0x2282, 0x338}}, {0x2285, {0x2283, 0x338}}, {0x2288, {0x2286, 0x338}}, {0x2289, {0x2287, 0x338}}, -{0x22AC, {0x22A2, 0x338}}, {0x22AD, {0x22A8, 0x338}}, {0x22AE, {0x22A9, 0x338}}, {0x22AF, {0x22AB, 0x338}}, {0x22E0, {0x227C, 0x338}}, {0x22E1, {0x227D, 0x338}}, {0x22E2, {0x2291, 0x338}}, -{0x22E3, {0x2292, 0x338}}, {0x22EA, {0x22B2, 0x338}}, {0x22EB, {0x22B3, 0x338}}, {0x22EC, {0x22B4, 0x338}}, {0x22ED, {0x22B5, 0x338}}, {0x2329, {0x3008}}, {0x232A, {0x3009}}, -{0x2ADC, {0x2ADD, 0x338}}, {0x304C, {0x304B, 0x3099}}, {0x304E, {0x304D, 0x3099}}, {0x3050, {0x304F, 0x3099}}, {0x3052, {0x3051, 0x3099}}, {0x3054, {0x3053, 0x3099}}, {0x3056, {0x3055, 0x3099}}, -{0x3058, {0x3057, 0x3099}}, {0x305A, {0x3059, 0x3099}}, {0x305C, {0x305B, 0x3099}}, {0x305E, {0x305D, 0x3099}}, {0x3060, {0x305F, 0x3099}}, {0x3062, {0x3061, 0x3099}}, {0x3065, {0x3064, 0x3099}}, -{0x3067, {0x3066, 0x3099}}, {0x3069, {0x3068, 0x3099}}, {0x3070, {0x306F, 0x3099}}, {0x3071, {0x306F, 0x309A}}, {0x3073, {0x3072, 0x3099}}, {0x3074, {0x3072, 0x309A}}, {0x3076, {0x3075, 0x3099}}, -{0x3077, {0x3075, 0x309A}}, {0x3079, {0x3078, 0x3099}}, {0x307A, {0x3078, 0x309A}}, {0x307C, {0x307B, 0x3099}}, {0x307D, {0x307B, 0x309A}}, {0x3094, {0x3046, 0x3099}}, {0x309E, {0x309D, 0x3099}}, -{0x30AC, {0x30AB, 0x3099}}, {0x30AE, {0x30AD, 0x3099}}, {0x30B0, {0x30AF, 0x3099}}, {0x30B2, {0x30B1, 0x3099}}, {0x30B4, {0x30B3, 0x3099}}, {0x30B6, {0x30B5, 0x3099}}, {0x30B8, {0x30B7, 0x3099}}, -{0x30BA, {0x30B9, 0x3099}}, {0x30BC, {0x30BB, 0x3099}}, {0x30BE, {0x30BD, 0x3099}}, {0x30C0, {0x30BF, 0x3099}}, {0x30C2, {0x30C1, 0x3099}}, {0x30C5, {0x30C4, 0x3099}}, {0x30C7, {0x30C6, 0x3099}}, -{0x30C9, {0x30C8, 0x3099}}, {0x30D0, {0x30CF, 0x3099}}, {0x30D1, {0x30CF, 0x309A}}, {0x30D3, {0x30D2, 0x3099}}, {0x30D4, {0x30D2, 0x309A}}, {0x30D6, {0x30D5, 0x3099}}, {0x30D7, {0x30D5, 0x309A}}, -{0x30D9, {0x30D8, 0x3099}}, {0x30DA, {0x30D8, 0x309A}}, {0x30DC, {0x30DB, 0x3099}}, {0x30DD, {0x30DB, 0x309A}}, {0x30F4, {0x30A6, 0x3099}}, {0x30F7, {0x30EF, 0x3099}}, {0x30F8, {0x30F0, 0x3099}}, -{0x30F9, {0x30F1, 0x3099}}, {0x30FA, {0x30F2, 0x3099}}, {0x30FE, {0x30FD, 0x3099}}, {0xF900, {0x8C48}}, {0xF901, {0x66F4}}, {0xF902, {0x8ECA}}, {0xF903, {0x8CC8}}, {0xF904, {0x6ED1}}, -{0xF905, {0x4E32}}, {0xF906, {0x53E5}}, {0xF907, {0x9F9C}}, {0xF908, {0x9F9C}}, {0xF909, {0x5951}}, {0xF90A, {0x91D1}}, {0xF90B, {0x5587}}, {0xF90C, {0x5948}}, {0xF90D, {0x61F6}}, {0xF90E, {0x7669}}, -{0xF90F, {0x7F85}}, {0xF910, {0x863F}}, {0xF911, {0x87BA}}, {0xF912, {0x88F8}}, {0xF913, {0x908F}}, {0xF914, {0x6A02}}, {0xF915, {0x6D1B}}, {0xF916, {0x70D9}}, {0xF917, {0x73DE}}, {0xF918, {0x843D}}, -{0xF919, {0x916A}}, {0xF91A, {0x99F1}}, {0xF91B, {0x4E82}}, {0xF91C, {0x5375}}, {0xF91D, {0x6B04}}, {0xF91E, {0x721B}}, {0xF91F, {0x862D}}, {0xF920, {0x9E1E}}, {0xF921, {0x5D50}}, {0xF922, {0x6FEB}}, -{0xF923, {0x85CD}}, {0xF924, {0x8964}}, {0xF925, {0x62C9}}, {0xF926, {0x81D8}}, {0xF927, {0x881F}}, {0xF928, {0x5ECA}}, {0xF929, {0x6717}}, {0xF92A, {0x6D6A}}, {0xF92B, {0x72FC}}, {0xF92C, {0x90CE}}, -{0xF92D, {0x4F86}}, {0xF92E, {0x51B7}}, {0xF92F, {0x52DE}}, {0xF930, {0x64C4}}, {0xF931, {0x6AD3}}, {0xF932, {0x7210}}, {0xF933, {0x76E7}}, {0xF934, {0x8001}}, {0xF935, {0x8606}}, {0xF936, {0x865C}}, -{0xF937, {0x8DEF}}, {0xF938, {0x9732}}, {0xF939, {0x9B6F}}, {0xF93A, {0x9DFA}}, {0xF93B, {0x788C}}, {0xF93C, {0x797F}}, {0xF93D, {0x7DA0}}, {0xF93E, {0x83C9}}, {0xF93F, {0x9304}}, {0xF940, {0x9E7F}}, -{0xF941, {0x8AD6}}, {0xF942, {0x58DF}}, {0xF943, {0x5F04}}, {0xF944, {0x7C60}}, {0xF945, {0x807E}}, {0xF946, {0x7262}}, {0xF947, {0x78CA}}, {0xF948, {0x8CC2}}, {0xF949, {0x96F7}}, {0xF94A, {0x58D8}}, -{0xF94B, {0x5C62}}, {0xF94C, {0x6A13}}, {0xF94D, {0x6DDA}}, {0xF94E, {0x6F0F}}, {0xF94F, {0x7D2F}}, {0xF950, {0x7E37}}, {0xF951, {0x964B}}, {0xF952, {0x52D2}}, {0xF953, {0x808B}}, {0xF954, {0x51DC}}, -{0xF955, {0x51CC}}, {0xF956, {0x7A1C}}, {0xF957, {0x7DBE}}, {0xF958, {0x83F1}}, {0xF959, {0x9675}}, {0xF95A, {0x8B80}}, {0xF95B, {0x62CF}}, {0xF95C, {0x6A02}}, {0xF95D, {0x8AFE}}, {0xF95E, {0x4E39}}, -{0xF95F, {0x5BE7}}, {0xF960, {0x6012}}, {0xF961, {0x7387}}, {0xF962, {0x7570}}, {0xF963, {0x5317}}, {0xF964, {0x78FB}}, {0xF965, {0x4FBF}}, {0xF966, {0x5FA9}}, {0xF967, {0x4E0D}}, {0xF968, {0x6CCC}}, -{0xF969, {0x6578}}, {0xF96A, {0x7D22}}, {0xF96B, {0x53C3}}, {0xF96C, {0x585E}}, {0xF96D, {0x7701}}, {0xF96E, {0x8449}}, {0xF96F, {0x8AAA}}, {0xF970, {0x6BBA}}, {0xF971, {0x8FB0}}, {0xF972, {0x6C88}}, -{0xF973, {0x62FE}}, {0xF974, {0x82E5}}, {0xF975, {0x63A0}}, {0xF976, {0x7565}}, {0xF977, {0x4EAE}}, {0xF978, {0x5169}}, {0xF979, {0x51C9}}, {0xF97A, {0x6881}}, {0xF97B, {0x7CE7}}, {0xF97C, {0x826F}}, -{0xF97D, {0x8AD2}}, {0xF97E, {0x91CF}}, {0xF97F, {0x52F5}}, {0xF980, {0x5442}}, {0xF981, {0x5973}}, {0xF982, {0x5EEC}}, {0xF983, {0x65C5}}, {0xF984, {0x6FFE}}, {0xF985, {0x792A}}, {0xF986, {0x95AD}}, -{0xF987, {0x9A6A}}, {0xF988, {0x9E97}}, {0xF989, {0x9ECE}}, {0xF98A, {0x529B}}, {0xF98B, {0x66C6}}, {0xF98C, {0x6B77}}, {0xF98D, {0x8F62}}, {0xF98E, {0x5E74}}, {0xF98F, {0x6190}}, {0xF990, {0x6200}}, -{0xF991, {0x649A}}, {0xF992, {0x6F23}}, {0xF993, {0x7149}}, {0xF994, {0x7489}}, {0xF995, {0x79CA}}, {0xF996, {0x7DF4}}, {0xF997, {0x806F}}, {0xF998, {0x8F26}}, {0xF999, {0x84EE}}, {0xF99A, {0x9023}}, -{0xF99B, {0x934A}}, {0xF99C, {0x5217}}, {0xF99D, {0x52A3}}, {0xF99E, {0x54BD}}, {0xF99F, {0x70C8}}, {0xF9A0, {0x88C2}}, {0xF9A1, {0x8AAA}}, {0xF9A2, {0x5EC9}}, {0xF9A3, {0x5FF5}}, {0xF9A4, {0x637B}}, -{0xF9A5, {0x6BAE}}, {0xF9A6, {0x7C3E}}, {0xF9A7, {0x7375}}, {0xF9A8, {0x4EE4}}, {0xF9A9, {0x56F9}}, {0xF9AA, {0x5BE7}}, {0xF9AB, {0x5DBA}}, {0xF9AC, {0x601C}}, {0xF9AD, {0x73B2}}, {0xF9AE, {0x7469}}, -{0xF9AF, {0x7F9A}}, {0xF9B0, {0x8046}}, {0xF9B1, {0x9234}}, {0xF9B2, {0x96F6}}, {0xF9B3, {0x9748}}, {0xF9B4, {0x9818}}, {0xF9B5, {0x4F8B}}, {0xF9B6, {0x79AE}}, {0xF9B7, {0x91B4}}, {0xF9B8, {0x96B8}}, -{0xF9B9, {0x60E1}}, {0xF9BA, {0x4E86}}, {0xF9BB, {0x50DA}}, {0xF9BC, {0x5BEE}}, {0xF9BD, {0x5C3F}}, {0xF9BE, {0x6599}}, {0xF9BF, {0x6A02}}, {0xF9C0, {0x71CE}}, {0xF9C1, {0x7642}}, {0xF9C2, {0x84FC}}, -{0xF9C3, {0x907C}}, {0xF9C4, {0x9F8D}}, {0xF9C5, {0x6688}}, {0xF9C6, {0x962E}}, {0xF9C7, {0x5289}}, {0xF9C8, {0x677B}}, {0xF9C9, {0x67F3}}, {0xF9CA, {0x6D41}}, {0xF9CB, {0x6E9C}}, {0xF9CC, {0x7409}}, -{0xF9CD, {0x7559}}, {0xF9CE, {0x786B}}, {0xF9CF, {0x7D10}}, {0xF9D0, {0x985E}}, {0xF9D1, {0x516D}}, {0xF9D2, {0x622E}}, {0xF9D3, {0x9678}}, {0xF9D4, {0x502B}}, {0xF9D5, {0x5D19}}, {0xF9D6, {0x6DEA}}, -{0xF9D7, {0x8F2A}}, {0xF9D8, {0x5F8B}}, {0xF9D9, {0x6144}}, {0xF9DA, {0x6817}}, {0xF9DB, {0x7387}}, {0xF9DC, {0x9686}}, {0xF9DD, {0x5229}}, {0xF9DE, {0x540F}}, {0xF9DF, {0x5C65}}, {0xF9E0, {0x6613}}, -{0xF9E1, {0x674E}}, {0xF9E2, {0x68A8}}, {0xF9E3, {0x6CE5}}, {0xF9E4, {0x7406}}, {0xF9E5, {0x75E2}}, {0xF9E6, {0x7F79}}, {0xF9E7, {0x88CF}}, {0xF9E8, {0x88E1}}, {0xF9E9, {0x91CC}}, {0xF9EA, {0x96E2}}, -{0xF9EB, {0x533F}}, {0xF9EC, {0x6EBA}}, {0xF9ED, {0x541D}}, {0xF9EE, {0x71D0}}, {0xF9EF, {0x7498}}, {0xF9F0, {0x85FA}}, {0xF9F1, {0x96A3}}, {0xF9F2, {0x9C57}}, {0xF9F3, {0x9E9F}}, {0xF9F4, {0x6797}}, -{0xF9F5, {0x6DCB}}, {0xF9F6, {0x81E8}}, {0xF9F7, {0x7ACB}}, {0xF9F8, {0x7B20}}, {0xF9F9, {0x7C92}}, {0xF9FA, {0x72C0}}, {0xF9FB, {0x7099}}, {0xF9FC, {0x8B58}}, {0xF9FD, {0x4EC0}}, {0xF9FE, {0x8336}}, -{0xF9FF, {0x523A}}, {0xFA00, {0x5207}}, {0xFA01, {0x5EA6}}, {0xFA02, {0x62D3}}, {0xFA03, {0x7CD6}}, {0xFA04, {0x5B85}}, {0xFA05, {0x6D1E}}, {0xFA06, {0x66B4}}, {0xFA07, {0x8F3B}}, {0xFA08, {0x884C}}, -{0xFA09, {0x964D}}, {0xFA0A, {0x898B}}, {0xFA0B, {0x5ED3}}, {0xFA0C, {0x5140}}, {0xFA0D, {0x55C0}}, {0xFA10, {0x585A}}, {0xFA12, {0x6674}}, {0xFA15, {0x51DE}}, {0xFA16, {0x732A}}, {0xFA17, {0x76CA}}, -{0xFA18, {0x793C}}, {0xFA19, {0x795E}}, {0xFA1A, {0x7965}}, {0xFA1B, {0x798F}}, {0xFA1C, {0x9756}}, {0xFA1D, {0x7CBE}}, {0xFA1E, {0x7FBD}}, {0xFA20, {0x8612}}, {0xFA22, {0x8AF8}}, {0xFA25, {0x9038}}, -{0xFA26, {0x90FD}}, {0xFA2A, {0x98EF}}, {0xFA2B, {0x98FC}}, {0xFA2C, {0x9928}}, {0xFA2D, {0x9DB4}}, {0xFA2E, {0x90DE}}, {0xFA2F, {0x96B7}}, {0xFA30, {0x4FAE}}, {0xFA31, {0x50E7}}, {0xFA32, {0x514D}}, -{0xFA33, {0x52C9}}, {0xFA34, {0x52E4}}, {0xFA35, {0x5351}}, {0xFA36, {0x559D}}, {0xFA37, {0x5606}}, {0xFA38, {0x5668}}, {0xFA39, {0x5840}}, {0xFA3A, {0x58A8}}, {0xFA3B, {0x5C64}}, {0xFA3C, {0x5C6E}}, -{0xFA3D, {0x6094}}, {0xFA3E, {0x6168}}, {0xFA3F, {0x618E}}, {0xFA40, {0x61F2}}, {0xFA41, {0x654F}}, {0xFA42, {0x65E2}}, {0xFA43, {0x6691}}, {0xFA44, {0x6885}}, {0xFA45, {0x6D77}}, {0xFA46, {0x6E1A}}, -{0xFA47, {0x6F22}}, {0xFA48, {0x716E}}, {0xFA49, {0x722B}}, {0xFA4A, {0x7422}}, {0xFA4B, {0x7891}}, {0xFA4C, {0x793E}}, {0xFA4D, {0x7949}}, {0xFA4E, {0x7948}}, {0xFA4F, {0x7950}}, {0xFA50, {0x7956}}, -{0xFA51, {0x795D}}, {0xFA52, {0x798D}}, {0xFA53, {0x798E}}, {0xFA54, {0x7A40}}, {0xFA55, {0x7A81}}, {0xFA56, {0x7BC0}}, {0xFA57, {0x7DF4}}, {0xFA58, {0x7E09}}, {0xFA59, {0x7E41}}, {0xFA5A, {0x7F72}}, -{0xFA5B, {0x8005}}, {0xFA5C, {0x81ED}}, {0xFA5D, {0x8279}}, {0xFA5E, {0x8279}}, {0xFA5F, {0x8457}}, {0xFA60, {0x8910}}, {0xFA61, {0x8996}}, {0xFA62, {0x8B01}}, {0xFA63, {0x8B39}}, {0xFA64, {0x8CD3}}, -{0xFA65, {0x8D08}}, {0xFA66, {0x8FB6}}, {0xFA67, {0x9038}}, {0xFA68, {0x96E3}}, {0xFA69, {0x97FF}}, {0xFA6A, {0x983B}}, {0xFA6B, {0x6075}}, {0xFA6C, {0x242EE}}, {0xFA6D, {0x8218}}, {0xFA70, {0x4E26}}, -{0xFA71, {0x51B5}}, {0xFA72, {0x5168}}, {0xFA73, {0x4F80}}, {0xFA74, {0x5145}}, {0xFA75, {0x5180}}, {0xFA76, {0x52C7}}, {0xFA77, {0x52FA}}, {0xFA78, {0x559D}}, {0xFA79, {0x5555}}, {0xFA7A, {0x5599}}, -{0xFA7B, {0x55E2}}, {0xFA7C, {0x585A}}, {0xFA7D, {0x58B3}}, {0xFA7E, {0x5944}}, {0xFA7F, {0x5954}}, {0xFA80, {0x5A62}}, {0xFA81, {0x5B28}}, {0xFA82, {0x5ED2}}, {0xFA83, {0x5ED9}}, {0xFA84, {0x5F69}}, -{0xFA85, {0x5FAD}}, {0xFA86, {0x60D8}}, {0xFA87, {0x614E}}, {0xFA88, {0x6108}}, {0xFA89, {0x618E}}, {0xFA8A, {0x6160}}, {0xFA8B, {0x61F2}}, {0xFA8C, {0x6234}}, {0xFA8D, {0x63C4}}, {0xFA8E, {0x641C}}, -{0xFA8F, {0x6452}}, {0xFA90, {0x6556}}, {0xFA91, {0x6674}}, {0xFA92, {0x6717}}, {0xFA93, {0x671B}}, {0xFA94, {0x6756}}, {0xFA95, {0x6B79}}, {0xFA96, {0x6BBA}}, {0xFA97, {0x6D41}}, {0xFA98, {0x6EDB}}, -{0xFA99, {0x6ECB}}, {0xFA9A, {0x6F22}}, {0xFA9B, {0x701E}}, {0xFA9C, {0x716E}}, {0xFA9D, {0x77A7}}, {0xFA9E, {0x7235}}, {0xFA9F, {0x72AF}}, {0xFAA0, {0x732A}}, {0xFAA1, {0x7471}}, {0xFAA2, {0x7506}}, -{0xFAA3, {0x753B}}, {0xFAA4, {0x761D}}, {0xFAA5, {0x761F}}, {0xFAA6, {0x76CA}}, {0xFAA7, {0x76DB}}, {0xFAA8, {0x76F4}}, {0xFAA9, {0x774A}}, {0xFAAA, {0x7740}}, {0xFAAB, {0x78CC}}, {0xFAAC, {0x7AB1}}, -{0xFAAD, {0x7BC0}}, {0xFAAE, {0x7C7B}}, {0xFAAF, {0x7D5B}}, {0xFAB0, {0x7DF4}}, {0xFAB1, {0x7F3E}}, {0xFAB2, {0x8005}}, {0xFAB3, {0x8352}}, {0xFAB4, {0x83EF}}, {0xFAB5, {0x8779}}, {0xFAB6, {0x8941}}, -{0xFAB7, {0x8986}}, {0xFAB8, {0x8996}}, {0xFAB9, {0x8ABF}}, {0xFABA, {0x8AF8}}, {0xFABB, {0x8ACB}}, {0xFABC, {0x8B01}}, {0xFABD, {0x8AFE}}, {0xFABE, {0x8AED}}, {0xFABF, {0x8B39}}, {0xFAC0, {0x8B8A}}, -{0xFAC1, {0x8D08}}, {0xFAC2, {0x8F38}}, {0xFAC3, {0x9072}}, {0xFAC4, {0x9199}}, {0xFAC5, {0x9276}}, {0xFAC6, {0x967C}}, {0xFAC7, {0x96E3}}, {0xFAC8, {0x9756}}, {0xFAC9, {0x97DB}}, {0xFACA, {0x97FF}}, -{0xFACB, {0x980B}}, {0xFACC, {0x983B}}, {0xFACD, {0x9B12}}, {0xFACE, {0x9F9C}}, {0xFACF, {0x2284A}}, {0xFAD0, {0x22844}}, {0xFAD1, {0x233D5}}, {0xFAD2, {0x3B9D}}, {0xFAD3, {0x4018}}, -{0xFAD4, {0x4039}}, {0xFAD5, {0x25249}}, {0xFAD6, {0x25CD0}}, {0xFAD7, {0x27ED3}}, {0xFAD8, {0x9F43}}, {0xFAD9, {0x9F8E}}, {0xFB1D, {0x5D9, 0x5B4}}, {0xFB1F, {0x5F2, 0x5B7}}, {0xFB2A, {0x5E9, 0x5C1}}, -{0xFB2B, {0x5E9, 0x5C2}}, {0xFB2C, {0x5E9, 0x5BC, 0x5C1}}, {0xFB2D, {0x5E9, 0x5BC, 0x5C2}}, {0xFB2E, {0x5D0, 0x5B7}}, {0xFB2F, {0x5D0, 0x5B8}}, {0xFB30, {0x5D0, 0x5BC}}, {0xFB31, {0x5D1, 0x5BC}}, -{0xFB32, {0x5D2, 0x5BC}}, {0xFB33, {0x5D3, 0x5BC}}, {0xFB34, {0x5D4, 0x5BC}}, {0xFB35, {0x5D5, 0x5BC}}, {0xFB36, {0x5D6, 0x5BC}}, {0xFB38, {0x5D8, 0x5BC}}, {0xFB39, {0x5D9, 0x5BC}}, -{0xFB3A, {0x5DA, 0x5BC}}, {0xFB3B, {0x5DB, 0x5BC}}, {0xFB3C, {0x5DC, 0x5BC}}, {0xFB3E, {0x5DE, 0x5BC}}, {0xFB40, {0x5E0, 0x5BC}}, {0xFB41, {0x5E1, 0x5BC}}, {0xFB43, {0x5E3, 0x5BC}}, -{0xFB44, {0x5E4, 0x5BC}}, {0xFB46, {0x5E6, 0x5BC}}, {0xFB47, {0x5E7, 0x5BC}}, {0xFB48, {0x5E8, 0x5BC}}, {0xFB49, {0x5E9, 0x5BC}}, {0xFB4A, {0x5EA, 0x5BC}}, {0xFB4B, {0x5D5, 0x5B9}}, -{0xFB4C, {0x5D1, 0x5BF}}, {0xFB4D, {0x5DB, 0x5BF}}, {0xFB4E, {0x5E4, 0x5BF}}, {0x1109A, {0x11099, 0x110BA}}, {0x1109C, {0x1109B, 0x110BA}}, {0x110AB, {0x110A5, 0x110BA}}, -{0x1112E, {0x11131, 0x11127}}, {0x1112F, {0x11132, 0x11127}}, {0x1134B, {0x11347, 0x1133E}}, {0x1134C, {0x11347, 0x11357}}, {0x114BB, {0x114B9, 0x114BA}}, {0x114BC, {0x114B9, 0x114B0}}, -{0x114BE, {0x114B9, 0x114BD}}, {0x115BA, {0x115B8, 0x115AF}}, {0x115BB, {0x115B9, 0x115AF}}, {0x1D15E, {0x1D157, 0x1D165}}, {0x1D15F, {0x1D158, 0x1D165}}, {0x1D160, {0x1D158, 0x1D165, 0x1D16E}}, -{0x1D161, {0x1D158, 0x1D165, 0x1D16F}}, {0x1D162, {0x1D158, 0x1D165, 0x1D170}}, {0x1D163, {0x1D158, 0x1D165, 0x1D171}}, {0x1D164, {0x1D158, 0x1D165, 0x1D172}}, {0x1D1BB, {0x1D1B9, 0x1D165}}, -{0x1D1BC, {0x1D1BA, 0x1D165}}, {0x1D1BD, {0x1D1B9, 0x1D165, 0x1D16E}}, {0x1D1BE, {0x1D1BA, 0x1D165, 0x1D16E}}, {0x1D1BF, {0x1D1B9, 0x1D165, 0x1D16F}}, {0x1D1C0, {0x1D1BA, 0x1D165, 0x1D16F}}, -{0x2F800, {0x4E3D}}, {0x2F801, {0x4E38}}, {0x2F802, {0x4E41}}, {0x2F803, {0x20122}}, {0x2F804, {0x4F60}}, {0x2F805, {0x4FAE}}, {0x2F806, {0x4FBB}}, {0x2F807, {0x5002}}, {0x2F808, {0x507A}}, -{0x2F809, {0x5099}}, {0x2F80A, {0x50E7}}, {0x2F80B, {0x50CF}}, {0x2F80C, {0x349E}}, {0x2F80D, {0x2063A}}, {0x2F80E, {0x514D}}, {0x2F80F, {0x5154}}, {0x2F810, {0x5164}}, {0x2F811, {0x5177}}, -{0x2F812, {0x2051C}}, {0x2F813, {0x34B9}}, {0x2F814, {0x5167}}, {0x2F815, {0x518D}}, {0x2F816, {0x2054B}}, {0x2F817, {0x5197}}, {0x2F818, {0x51A4}}, {0x2F819, {0x4ECC}}, {0x2F81A, {0x51AC}}, -{0x2F81B, {0x51B5}}, {0x2F81C, {0x291DF}}, {0x2F81D, {0x51F5}}, {0x2F81E, {0x5203}}, {0x2F81F, {0x34DF}}, {0x2F820, {0x523B}}, {0x2F821, {0x5246}}, {0x2F822, {0x5272}}, {0x2F823, {0x5277}}, -{0x2F824, {0x3515}}, {0x2F825, {0x52C7}}, {0x2F826, {0x52C9}}, {0x2F827, {0x52E4}}, {0x2F828, {0x52FA}}, {0x2F829, {0x5305}}, {0x2F82A, {0x5306}}, {0x2F82B, {0x5317}}, {0x2F82C, {0x5349}}, -{0x2F82D, {0x5351}}, {0x2F82E, {0x535A}}, {0x2F82F, {0x5373}}, {0x2F830, {0x537D}}, {0x2F831, {0x537F}}, {0x2F832, {0x537F}}, {0x2F833, {0x537F}}, {0x2F834, {0x20A2C}}, {0x2F835, {0x7070}}, -{0x2F836, {0x53CA}}, {0x2F837, {0x53DF}}, {0x2F838, {0x20B63}}, {0x2F839, {0x53EB}}, {0x2F83A, {0x53F1}}, {0x2F83B, {0x5406}}, {0x2F83C, {0x549E}}, {0x2F83D, {0x5438}}, {0x2F83E, {0x5448}}, -{0x2F83F, {0x5468}}, {0x2F840, {0x54A2}}, {0x2F841, {0x54F6}}, {0x2F842, {0x5510}}, {0x2F843, {0x5553}}, {0x2F844, {0x5563}}, {0x2F845, {0x5584}}, {0x2F846, {0x5584}}, {0x2F847, {0x5599}}, -{0x2F848, {0x55AB}}, {0x2F849, {0x55B3}}, {0x2F84A, {0x55C2}}, {0x2F84B, {0x5716}}, {0x2F84C, {0x5606}}, {0x2F84D, {0x5717}}, {0x2F84E, {0x5651}}, {0x2F84F, {0x5674}}, {0x2F850, {0x5207}}, -{0x2F851, {0x58EE}}, {0x2F852, {0x57CE}}, {0x2F853, {0x57F4}}, {0x2F854, {0x580D}}, {0x2F855, {0x578B}}, {0x2F856, {0x5832}}, {0x2F857, {0x5831}}, {0x2F858, {0x58AC}}, {0x2F859, {0x214E4}}, -{0x2F85A, {0x58F2}}, {0x2F85B, {0x58F7}}, {0x2F85C, {0x5906}}, {0x2F85D, {0x591A}}, {0x2F85E, {0x5922}}, {0x2F85F, {0x5962}}, {0x2F860, {0x216A8}}, {0x2F861, {0x216EA}}, {0x2F862, {0x59EC}}, -{0x2F863, {0x5A1B}}, {0x2F864, {0x5A27}}, {0x2F865, {0x59D8}}, {0x2F866, {0x5A66}}, {0x2F867, {0x36EE}}, {0x2F868, {0x36FC}}, {0x2F869, {0x5B08}}, {0x2F86A, {0x5B3E}}, {0x2F86B, {0x5B3E}}, -{0x2F86C, {0x219C8}}, {0x2F86D, {0x5BC3}}, {0x2F86E, {0x5BD8}}, {0x2F86F, {0x5BE7}}, {0x2F870, {0x5BF3}}, {0x2F871, {0x21B18}}, {0x2F872, {0x5BFF}}, {0x2F873, {0x5C06}}, {0x2F874, {0x5F53}}, -{0x2F875, {0x5C22}}, {0x2F876, {0x3781}}, {0x2F877, {0x5C60}}, {0x2F878, {0x5C6E}}, {0x2F879, {0x5CC0}}, {0x2F87A, {0x5C8D}}, {0x2F87B, {0x21DE4}}, {0x2F87C, {0x5D43}}, {0x2F87D, {0x21DE6}}, -{0x2F87E, {0x5D6E}}, {0x2F87F, {0x5D6B}}, {0x2F880, {0x5D7C}}, {0x2F881, {0x5DE1}}, {0x2F882, {0x5DE2}}, {0x2F883, {0x382F}}, {0x2F884, {0x5DFD}}, {0x2F885, {0x5E28}}, {0x2F886, {0x5E3D}}, -{0x2F887, {0x5E69}}, {0x2F888, {0x3862}}, {0x2F889, {0x22183}}, {0x2F88A, {0x387C}}, {0x2F88B, {0x5EB0}}, {0x2F88C, {0x5EB3}}, {0x2F88D, {0x5EB6}}, {0x2F88E, {0x5ECA}}, {0x2F88F, {0x2A392}}, -{0x2F890, {0x5EFE}}, {0x2F891, {0x22331}}, {0x2F892, {0x22331}}, {0x2F893, {0x8201}}, {0x2F894, {0x5F22}}, {0x2F895, {0x5F22}}, {0x2F896, {0x38C7}}, {0x2F897, {0x232B8}}, {0x2F898, {0x261DA}}, -{0x2F899, {0x5F62}}, {0x2F89A, {0x5F6B}}, {0x2F89B, {0x38E3}}, {0x2F89C, {0x5F9A}}, {0x2F89D, {0x5FCD}}, {0x2F89E, {0x5FD7}}, {0x2F89F, {0x5FF9}}, {0x2F8A0, {0x6081}}, {0x2F8A1, {0x393A}}, -{0x2F8A2, {0x391C}}, {0x2F8A3, {0x6094}}, {0x2F8A4, {0x226D4}}, {0x2F8A5, {0x60C7}}, {0x2F8A6, {0x6148}}, {0x2F8A7, {0x614C}}, {0x2F8A8, {0x614E}}, {0x2F8A9, {0x614C}}, {0x2F8AA, {0x617A}}, -{0x2F8AB, {0x618E}}, {0x2F8AC, {0x61B2}}, {0x2F8AD, {0x61A4}}, {0x2F8AE, {0x61AF}}, {0x2F8AF, {0x61DE}}, {0x2F8B0, {0x61F2}}, {0x2F8B1, {0x61F6}}, {0x2F8B2, {0x6210}}, {0x2F8B3, {0x621B}}, -{0x2F8B4, {0x625D}}, {0x2F8B5, {0x62B1}}, {0x2F8B6, {0x62D4}}, {0x2F8B7, {0x6350}}, {0x2F8B8, {0x22B0C}}, {0x2F8B9, {0x633D}}, {0x2F8BA, {0x62FC}}, {0x2F8BB, {0x6368}}, {0x2F8BC, {0x6383}}, -{0x2F8BD, {0x63E4}}, {0x2F8BE, {0x22BF1}}, {0x2F8BF, {0x6422}}, {0x2F8C0, {0x63C5}}, {0x2F8C1, {0x63A9}}, {0x2F8C2, {0x3A2E}}, {0x2F8C3, {0x6469}}, {0x2F8C4, {0x647E}}, {0x2F8C5, {0x649D}}, -{0x2F8C6, {0x6477}}, {0x2F8C7, {0x3A6C}}, {0x2F8C8, {0x654F}}, {0x2F8C9, {0x656C}}, {0x2F8CA, {0x2300A}}, {0x2F8CB, {0x65E3}}, {0x2F8CC, {0x66F8}}, {0x2F8CD, {0x6649}}, {0x2F8CE, {0x3B19}}, -{0x2F8CF, {0x6691}}, {0x2F8D0, {0x3B08}}, {0x2F8D1, {0x3AE4}}, {0x2F8D2, {0x5192}}, {0x2F8D3, {0x5195}}, {0x2F8D4, {0x6700}}, {0x2F8D5, {0x669C}}, {0x2F8D6, {0x80AD}}, {0x2F8D7, {0x43D9}}, -{0x2F8D8, {0x6717}}, {0x2F8D9, {0x671B}}, {0x2F8DA, {0x6721}}, {0x2F8DB, {0x675E}}, {0x2F8DC, {0x6753}}, {0x2F8DD, {0x233C3}}, {0x2F8DE, {0x3B49}}, {0x2F8DF, {0x67FA}}, {0x2F8E0, {0x6785}}, -{0x2F8E1, {0x6852}}, {0x2F8E2, {0x6885}}, {0x2F8E3, {0x2346D}}, {0x2F8E4, {0x688E}}, {0x2F8E5, {0x681F}}, {0x2F8E6, {0x6914}}, {0x2F8E7, {0x3B9D}}, {0x2F8E8, {0x6942}}, {0x2F8E9, {0x69A3}}, -{0x2F8EA, {0x69EA}}, {0x2F8EB, {0x6AA8}}, {0x2F8EC, {0x236A3}}, {0x2F8ED, {0x6ADB}}, {0x2F8EE, {0x3C18}}, {0x2F8EF, {0x6B21}}, {0x2F8F0, {0x238A7}}, {0x2F8F1, {0x6B54}}, {0x2F8F2, {0x3C4E}}, -{0x2F8F3, {0x6B72}}, {0x2F8F4, {0x6B9F}}, {0x2F8F5, {0x6BBA}}, {0x2F8F6, {0x6BBB}}, {0x2F8F7, {0x23A8D}}, {0x2F8F8, {0x21D0B}}, {0x2F8F9, {0x23AFA}}, {0x2F8FA, {0x6C4E}}, {0x2F8FB, {0x23CBC}}, -{0x2F8FC, {0x6CBF}}, {0x2F8FD, {0x6CCD}}, {0x2F8FE, {0x6C67}}, {0x2F8FF, {0x6D16}}, {0x2F900, {0x6D3E}}, {0x2F901, {0x6D77}}, {0x2F902, {0x6D41}}, {0x2F903, {0x6D69}}, {0x2F904, {0x6D78}}, -{0x2F905, {0x6D85}}, {0x2F906, {0x23D1E}}, {0x2F907, {0x6D34}}, {0x2F908, {0x6E2F}}, {0x2F909, {0x6E6E}}, {0x2F90A, {0x3D33}}, {0x2F90B, {0x6ECB}}, {0x2F90C, {0x6EC7}}, {0x2F90D, {0x23ED1}}, -{0x2F90E, {0x6DF9}}, {0x2F90F, {0x6F6E}}, {0x2F910, {0x23F5E}}, {0x2F911, {0x23F8E}}, {0x2F912, {0x6FC6}}, {0x2F913, {0x7039}}, {0x2F914, {0x701E}}, {0x2F915, {0x701B}}, {0x2F916, {0x3D96}}, -{0x2F917, {0x704A}}, {0x2F918, {0x707D}}, {0x2F919, {0x7077}}, {0x2F91A, {0x70AD}}, {0x2F91B, {0x20525}}, {0x2F91C, {0x7145}}, {0x2F91D, {0x24263}}, {0x2F91E, {0x719C}}, {0x2F91F, {0x243AB}}, -{0x2F920, {0x7228}}, {0x2F921, {0x7235}}, {0x2F922, {0x7250}}, {0x2F923, {0x24608}}, {0x2F924, {0x7280}}, {0x2F925, {0x7295}}, {0x2F926, {0x24735}}, {0x2F927, {0x24814}}, {0x2F928, {0x737A}}, -{0x2F929, {0x738B}}, {0x2F92A, {0x3EAC}}, {0x2F92B, {0x73A5}}, {0x2F92C, {0x3EB8}}, {0x2F92D, {0x3EB8}}, {0x2F92E, {0x7447}}, {0x2F92F, {0x745C}}, {0x2F930, {0x7471}}, {0x2F931, {0x7485}}, -{0x2F932, {0x74CA}}, {0x2F933, {0x3F1B}}, {0x2F934, {0x7524}}, {0x2F935, {0x24C36}}, {0x2F936, {0x753E}}, {0x2F937, {0x24C92}}, {0x2F938, {0x7570}}, {0x2F939, {0x2219F}}, {0x2F93A, {0x7610}}, -{0x2F93B, {0x24FA1}}, {0x2F93C, {0x24FB8}}, {0x2F93D, {0x25044}}, {0x2F93E, {0x3FFC}}, {0x2F93F, {0x4008}}, {0x2F940, {0x76F4}}, {0x2F941, {0x250F3}}, {0x2F942, {0x250F2}}, {0x2F943, {0x25119}}, -{0x2F944, {0x25133}}, {0x2F945, {0x771E}}, {0x2F946, {0x771F}}, {0x2F947, {0x771F}}, {0x2F948, {0x774A}}, {0x2F949, {0x4039}}, {0x2F94A, {0x778B}}, {0x2F94B, {0x4046}}, {0x2F94C, {0x4096}}, -{0x2F94D, {0x2541D}}, {0x2F94E, {0x784E}}, {0x2F94F, {0x788C}}, {0x2F950, {0x78CC}}, {0x2F951, {0x40E3}}, {0x2F952, {0x25626}}, {0x2F953, {0x7956}}, {0x2F954, {0x2569A}}, {0x2F955, {0x256C5}}, -{0x2F956, {0x798F}}, {0x2F957, {0x79EB}}, {0x2F958, {0x412F}}, {0x2F959, {0x7A40}}, {0x2F95A, {0x7A4A}}, {0x2F95B, {0x7A4F}}, {0x2F95C, {0x2597C}}, {0x2F95D, {0x25AA7}}, {0x2F95E, {0x25AA7}}, -{0x2F95F, {0x7AEE}}, {0x2F960, {0x4202}}, {0x2F961, {0x25BAB}}, {0x2F962, {0x7BC6}}, {0x2F963, {0x7BC9}}, {0x2F964, {0x4227}}, {0x2F965, {0x25C80}}, {0x2F966, {0x7CD2}}, {0x2F967, {0x42A0}}, -{0x2F968, {0x7CE8}}, {0x2F969, {0x7CE3}}, {0x2F96A, {0x7D00}}, {0x2F96B, {0x25F86}}, {0x2F96C, {0x7D63}}, {0x2F96D, {0x4301}}, {0x2F96E, {0x7DC7}}, {0x2F96F, {0x7E02}}, {0x2F970, {0x7E45}}, -{0x2F971, {0x4334}}, {0x2F972, {0x26228}}, {0x2F973, {0x26247}}, {0x2F974, {0x4359}}, {0x2F975, {0x262D9}}, {0x2F976, {0x7F7A}}, {0x2F977, {0x2633E}}, {0x2F978, {0x7F95}}, {0x2F979, {0x7FFA}}, -{0x2F97A, {0x8005}}, {0x2F97B, {0x264DA}}, {0x2F97C, {0x26523}}, {0x2F97D, {0x8060}}, {0x2F97E, {0x265A8}}, {0x2F97F, {0x8070}}, {0x2F980, {0x2335F}}, {0x2F981, {0x43D5}}, {0x2F982, {0x80B2}}, -{0x2F983, {0x8103}}, {0x2F984, {0x440B}}, {0x2F985, {0x813E}}, {0x2F986, {0x5AB5}}, {0x2F987, {0x267A7}}, {0x2F988, {0x267B5}}, {0x2F989, {0x23393}}, {0x2F98A, {0x2339C}}, {0x2F98B, {0x8201}}, -{0x2F98C, {0x8204}}, {0x2F98D, {0x8F9E}}, {0x2F98E, {0x446B}}, {0x2F98F, {0x8291}}, {0x2F990, {0x828B}}, {0x2F991, {0x829D}}, {0x2F992, {0x52B3}}, {0x2F993, {0x82B1}}, {0x2F994, {0x82B3}}, -{0x2F995, {0x82BD}}, {0x2F996, {0x82E6}}, {0x2F997, {0x26B3C}}, {0x2F998, {0x82E5}}, {0x2F999, {0x831D}}, {0x2F99A, {0x8363}}, {0x2F99B, {0x83AD}}, {0x2F99C, {0x8323}}, {0x2F99D, {0x83BD}}, -{0x2F99E, {0x83E7}}, {0x2F99F, {0x8457}}, {0x2F9A0, {0x8353}}, {0x2F9A1, {0x83CA}}, {0x2F9A2, {0x83CC}}, {0x2F9A3, {0x83DC}}, {0x2F9A4, {0x26C36}}, {0x2F9A5, {0x26D6B}}, {0x2F9A6, {0x26CD5}}, -{0x2F9A7, {0x452B}}, {0x2F9A8, {0x84F1}}, {0x2F9A9, {0x84F3}}, {0x2F9AA, {0x8516}}, {0x2F9AB, {0x273CA}}, {0x2F9AC, {0x8564}}, {0x2F9AD, {0x26F2C}}, {0x2F9AE, {0x455D}}, {0x2F9AF, {0x4561}}, -{0x2F9B0, {0x26FB1}}, {0x2F9B1, {0x270D2}}, {0x2F9B2, {0x456B}}, {0x2F9B3, {0x8650}}, {0x2F9B4, {0x865C}}, {0x2F9B5, {0x8667}}, {0x2F9B6, {0x8669}}, {0x2F9B7, {0x86A9}}, {0x2F9B8, {0x8688}}, -{0x2F9B9, {0x870E}}, {0x2F9BA, {0x86E2}}, {0x2F9BB, {0x8779}}, {0x2F9BC, {0x8728}}, {0x2F9BD, {0x876B}}, {0x2F9BE, {0x8786}}, {0x2F9BF, {0x45D7}}, {0x2F9C0, {0x87E1}}, {0x2F9C1, {0x8801}}, -{0x2F9C2, {0x45F9}}, {0x2F9C3, {0x8860}}, {0x2F9C4, {0x8863}}, {0x2F9C5, {0x27667}}, {0x2F9C6, {0x88D7}}, {0x2F9C7, {0x88DE}}, {0x2F9C8, {0x4635}}, {0x2F9C9, {0x88FA}}, {0x2F9CA, {0x34BB}}, -{0x2F9CB, {0x278AE}}, {0x2F9CC, {0x27966}}, {0x2F9CD, {0x46BE}}, {0x2F9CE, {0x46C7}}, {0x2F9CF, {0x8AA0}}, {0x2F9D0, {0x8AED}}, {0x2F9D1, {0x8B8A}}, {0x2F9D2, {0x8C55}}, {0x2F9D3, {0x27CA8}}, -{0x2F9D4, {0x8CAB}}, {0x2F9D5, {0x8CC1}}, {0x2F9D6, {0x8D1B}}, {0x2F9D7, {0x8D77}}, {0x2F9D8, {0x27F2F}}, {0x2F9D9, {0x20804}}, {0x2F9DA, {0x8DCB}}, {0x2F9DB, {0x8DBC}}, {0x2F9DC, {0x8DF0}}, -{0x2F9DD, {0x208DE}}, {0x2F9DE, {0x8ED4}}, {0x2F9DF, {0x8F38}}, {0x2F9E0, {0x285D2}}, {0x2F9E1, {0x285ED}}, {0x2F9E2, {0x9094}}, {0x2F9E3, {0x90F1}}, {0x2F9E4, {0x9111}}, {0x2F9E5, {0x2872E}}, -{0x2F9E6, {0x911B}}, {0x2F9E7, {0x9238}}, {0x2F9E8, {0x92D7}}, {0x2F9E9, {0x92D8}}, {0x2F9EA, {0x927C}}, {0x2F9EB, {0x93F9}}, {0x2F9EC, {0x9415}}, {0x2F9ED, {0x28BFA}}, {0x2F9EE, {0x958B}}, -{0x2F9EF, {0x4995}}, {0x2F9F0, {0x95B7}}, {0x2F9F1, {0x28D77}}, {0x2F9F2, {0x49E6}}, {0x2F9F3, {0x96C3}}, {0x2F9F4, {0x5DB2}}, {0x2F9F5, {0x9723}}, {0x2F9F6, {0x29145}}, {0x2F9F7, {0x2921A}}, -{0x2F9F8, {0x4A6E}}, {0x2F9F9, {0x4A76}}, {0x2F9FA, {0x97E0}}, {0x2F9FB, {0x2940A}}, {0x2F9FC, {0x4AB2}}, {0x2F9FD, {0x29496}}, {0x2F9FE, {0x980B}}, {0x2F9FF, {0x980B}}, {0x2FA00, {0x9829}}, -{0x2FA01, {0x295B6}}, {0x2FA02, {0x98E2}}, {0x2FA03, {0x4B33}}, {0x2FA04, {0x9929}}, {0x2FA05, {0x99A7}}, {0x2FA06, {0x99C2}}, {0x2FA07, {0x99FE}}, {0x2FA08, {0x4BCE}}, {0x2FA09, {0x29B30}}, -{0x2FA0A, {0x9B12}}, {0x2FA0B, {0x9C40}}, {0x2FA0C, {0x9CFD}}, {0x2FA0D, {0x4CCE}}, {0x2FA0E, {0x4CED}}, {0x2FA0F, {0x9D67}}, {0x2FA10, {0x2A0CE}}, {0x2FA11, {0x4CF8}}, {0x2FA12, {0x2A105}}, -{0x2FA13, {0x2A20E}}, {0x2FA14, {0x2A291}}, {0x2FA15, {0x9EBB}}, {0x2FA16, {0x4D56}}, {0x2FA17, {0x9EF9}}, {0x2FA18, {0x9EFE}}, {0x2FA19, {0x9F05}}, {0x2FA1A, {0x9F0F}}, {0x2FA1B, {0x9F16}}, -{0x2FA1D, {0x2A600}}, +static const std::multimap nfd_map = { +{0xC0, 0x41}, {0xC0, 0x300}, {0xC1, 0x41}, {0xC1, 0x301}, {0xC2, 0x41}, {0xC2, 0x302}, {0xC3, 0x41}, {0xC3, 0x303}, {0xC4, 0x41}, {0xC4, 0x308}, {0xC5, 0x41}, {0xC5, 0x30A}, {0xC7, 0x43}, +{0xC7, 0x327}, {0xC8, 0x45}, {0xC8, 0x300}, {0xC9, 0x45}, {0xC9, 0x301}, {0xCA, 0x45}, {0xCA, 0x302}, {0xCB, 0x45}, {0xCB, 0x308}, {0xCC, 0x49}, {0xCC, 0x300}, {0xCD, 0x49}, {0xCD, 0x301}, +{0xCE, 0x49}, {0xCE, 0x302}, {0xCF, 0x49}, {0xCF, 0x308}, {0xD1, 0x4E}, {0xD1, 0x303}, {0xD2, 0x4F}, {0xD2, 0x300}, {0xD3, 0x4F}, {0xD3, 0x301}, {0xD4, 0x4F}, {0xD4, 0x302}, {0xD5, 0x4F}, +{0xD5, 0x303}, {0xD6, 0x4F}, {0xD6, 0x308}, {0xD9, 0x55}, {0xD9, 0x300}, {0xDA, 0x55}, {0xDA, 0x301}, {0xDB, 0x55}, {0xDB, 0x302}, {0xDC, 0x55}, {0xDC, 0x308}, {0xDD, 0x59}, {0xDD, 0x301}, +{0xE0, 0x61}, {0xE0, 0x300}, {0xE1, 0x61}, {0xE1, 0x301}, {0xE2, 0x61}, {0xE2, 0x302}, {0xE3, 0x61}, {0xE3, 0x303}, {0xE4, 0x61}, {0xE4, 0x308}, {0xE5, 0x61}, {0xE5, 0x30A}, {0xE7, 0x63}, +{0xE7, 0x327}, {0xE8, 0x65}, {0xE8, 0x300}, {0xE9, 0x65}, {0xE9, 0x301}, {0xEA, 0x65}, {0xEA, 0x302}, {0xEB, 0x65}, {0xEB, 0x308}, {0xEC, 0x69}, {0xEC, 0x300}, {0xED, 0x69}, {0xED, 0x301}, +{0xEE, 0x69}, {0xEE, 0x302}, {0xEF, 0x69}, {0xEF, 0x308}, {0xF1, 0x6E}, {0xF1, 0x303}, {0xF2, 0x6F}, {0xF2, 0x300}, {0xF3, 0x6F}, {0xF3, 0x301}, {0xF4, 0x6F}, {0xF4, 0x302}, {0xF5, 0x6F}, +{0xF5, 0x303}, {0xF6, 0x6F}, {0xF6, 0x308}, {0xF9, 0x75}, {0xF9, 0x300}, {0xFA, 0x75}, {0xFA, 0x301}, {0xFB, 0x75}, {0xFB, 0x302}, {0xFC, 0x75}, {0xFC, 0x308}, {0xFD, 0x79}, {0xFD, 0x301}, +{0xFF, 0x79}, {0xFF, 0x308}, {0x100, 0x41}, {0x100, 0x304}, {0x101, 0x61}, {0x101, 0x304}, {0x102, 0x41}, {0x102, 0x306}, {0x103, 0x61}, {0x103, 0x306}, {0x104, 0x41}, {0x104, 0x328}, {0x105, 0x61}, +{0x105, 0x328}, {0x106, 0x43}, {0x106, 0x301}, {0x107, 0x63}, {0x107, 0x301}, {0x108, 0x43}, {0x108, 0x302}, {0x109, 0x63}, {0x109, 0x302}, {0x10A, 0x43}, {0x10A, 0x307}, {0x10B, 0x63}, +{0x10B, 0x307}, {0x10C, 0x43}, {0x10C, 0x30C}, {0x10D, 0x63}, {0x10D, 0x30C}, {0x10E, 0x44}, {0x10E, 0x30C}, {0x10F, 0x64}, {0x10F, 0x30C}, {0x112, 0x45}, {0x112, 0x304}, {0x113, 0x65}, +{0x113, 0x304}, {0x114, 0x45}, {0x114, 0x306}, {0x115, 0x65}, {0x115, 0x306}, {0x116, 0x45}, {0x116, 0x307}, {0x117, 0x65}, {0x117, 0x307}, {0x118, 0x45}, {0x118, 0x328}, {0x119, 0x65}, +{0x119, 0x328}, {0x11A, 0x45}, {0x11A, 0x30C}, {0x11B, 0x65}, {0x11B, 0x30C}, {0x11C, 0x47}, {0x11C, 0x302}, {0x11D, 0x67}, {0x11D, 0x302}, {0x11E, 0x47}, {0x11E, 0x306}, {0x11F, 0x67}, +{0x11F, 0x306}, {0x120, 0x47}, {0x120, 0x307}, {0x121, 0x67}, {0x121, 0x307}, {0x122, 0x47}, {0x122, 0x327}, {0x123, 0x67}, {0x123, 0x327}, {0x124, 0x48}, {0x124, 0x302}, {0x125, 0x68}, +{0x125, 0x302}, {0x128, 0x49}, {0x128, 0x303}, {0x129, 0x69}, {0x129, 0x303}, {0x12A, 0x49}, {0x12A, 0x304}, {0x12B, 0x69}, {0x12B, 0x304}, {0x12C, 0x49}, {0x12C, 0x306}, {0x12D, 0x69}, +{0x12D, 0x306}, {0x12E, 0x49}, {0x12E, 0x328}, {0x12F, 0x69}, {0x12F, 0x328}, {0x130, 0x49}, {0x130, 0x307}, {0x134, 0x4A}, {0x134, 0x302}, {0x135, 0x6A}, {0x135, 0x302}, {0x136, 0x4B}, +{0x136, 0x327}, {0x137, 0x6B}, {0x137, 0x327}, {0x139, 0x4C}, {0x139, 0x301}, {0x13A, 0x6C}, {0x13A, 0x301}, {0x13B, 0x4C}, {0x13B, 0x327}, {0x13C, 0x6C}, {0x13C, 0x327}, {0x13D, 0x4C}, +{0x13D, 0x30C}, {0x13E, 0x6C}, {0x13E, 0x30C}, {0x143, 0x4E}, {0x143, 0x301}, {0x144, 0x6E}, {0x144, 0x301}, {0x145, 0x4E}, {0x145, 0x327}, {0x146, 0x6E}, {0x146, 0x327}, {0x147, 0x4E}, +{0x147, 0x30C}, {0x148, 0x6E}, {0x148, 0x30C}, {0x14C, 0x4F}, {0x14C, 0x304}, {0x14D, 0x6F}, {0x14D, 0x304}, {0x14E, 0x4F}, {0x14E, 0x306}, {0x14F, 0x6F}, {0x14F, 0x306}, {0x150, 0x4F}, +{0x150, 0x30B}, {0x151, 0x6F}, {0x151, 0x30B}, {0x154, 0x52}, {0x154, 0x301}, {0x155, 0x72}, {0x155, 0x301}, {0x156, 0x52}, {0x156, 0x327}, {0x157, 0x72}, {0x157, 0x327}, {0x158, 0x52}, +{0x158, 0x30C}, {0x159, 0x72}, {0x159, 0x30C}, {0x15A, 0x53}, {0x15A, 0x301}, {0x15B, 0x73}, {0x15B, 0x301}, {0x15C, 0x53}, {0x15C, 0x302}, {0x15D, 0x73}, {0x15D, 0x302}, {0x15E, 0x53}, +{0x15E, 0x327}, {0x15F, 0x73}, {0x15F, 0x327}, {0x160, 0x53}, {0x160, 0x30C}, {0x161, 0x73}, {0x161, 0x30C}, {0x162, 0x54}, {0x162, 0x327}, {0x163, 0x74}, {0x163, 0x327}, {0x164, 0x54}, +{0x164, 0x30C}, {0x165, 0x74}, {0x165, 0x30C}, {0x168, 0x55}, {0x168, 0x303}, {0x169, 0x75}, {0x169, 0x303}, {0x16A, 0x55}, {0x16A, 0x304}, {0x16B, 0x75}, {0x16B, 0x304}, {0x16C, 0x55}, +{0x16C, 0x306}, {0x16D, 0x75}, {0x16D, 0x306}, {0x16E, 0x55}, {0x16E, 0x30A}, {0x16F, 0x75}, {0x16F, 0x30A}, {0x170, 0x55}, {0x170, 0x30B}, {0x171, 0x75}, {0x171, 0x30B}, {0x172, 0x55}, +{0x172, 0x328}, {0x173, 0x75}, {0x173, 0x328}, {0x174, 0x57}, {0x174, 0x302}, {0x175, 0x77}, {0x175, 0x302}, {0x176, 0x59}, {0x176, 0x302}, {0x177, 0x79}, {0x177, 0x302}, {0x178, 0x59}, +{0x178, 0x308}, {0x179, 0x5A}, {0x179, 0x301}, {0x17A, 0x7A}, {0x17A, 0x301}, {0x17B, 0x5A}, {0x17B, 0x307}, {0x17C, 0x7A}, {0x17C, 0x307}, {0x17D, 0x5A}, {0x17D, 0x30C}, {0x17E, 0x7A}, +{0x17E, 0x30C}, {0x1A0, 0x4F}, {0x1A0, 0x31B}, {0x1A1, 0x6F}, {0x1A1, 0x31B}, {0x1AF, 0x55}, {0x1AF, 0x31B}, {0x1B0, 0x75}, {0x1B0, 0x31B}, {0x1CD, 0x41}, {0x1CD, 0x30C}, {0x1CE, 0x61}, +{0x1CE, 0x30C}, {0x1CF, 0x49}, {0x1CF, 0x30C}, {0x1D0, 0x69}, {0x1D0, 0x30C}, {0x1D1, 0x4F}, {0x1D1, 0x30C}, {0x1D2, 0x6F}, {0x1D2, 0x30C}, {0x1D3, 0x55}, {0x1D3, 0x30C}, {0x1D4, 0x75}, +{0x1D4, 0x30C}, {0x1D5, 0x55}, {0x1D5, 0x308}, {0x1D5, 0x304}, {0x1D6, 0x75}, {0x1D6, 0x308}, {0x1D6, 0x304}, {0x1D7, 0x55}, {0x1D7, 0x308}, {0x1D7, 0x301}, {0x1D8, 0x75}, {0x1D8, 0x308}, +{0x1D8, 0x301}, {0x1D9, 0x55}, {0x1D9, 0x308}, {0x1D9, 0x30C}, {0x1DA, 0x75}, {0x1DA, 0x308}, {0x1DA, 0x30C}, {0x1DB, 0x55}, {0x1DB, 0x308}, {0x1DB, 0x300}, {0x1DC, 0x75}, {0x1DC, 0x308}, +{0x1DC, 0x300}, {0x1DE, 0x41}, {0x1DE, 0x308}, {0x1DE, 0x304}, {0x1DF, 0x61}, {0x1DF, 0x308}, {0x1DF, 0x304}, {0x1E0, 0x41}, {0x1E0, 0x307}, {0x1E0, 0x304}, {0x1E1, 0x61}, {0x1E1, 0x307}, +{0x1E1, 0x304}, {0x1E2, 0xC6}, {0x1E2, 0x304}, {0x1E3, 0xE6}, {0x1E3, 0x304}, {0x1E6, 0x47}, {0x1E6, 0x30C}, {0x1E7, 0x67}, {0x1E7, 0x30C}, {0x1E8, 0x4B}, {0x1E8, 0x30C}, {0x1E9, 0x6B}, +{0x1E9, 0x30C}, {0x1EA, 0x4F}, {0x1EA, 0x328}, {0x1EB, 0x6F}, {0x1EB, 0x328}, {0x1EC, 0x4F}, {0x1EC, 0x328}, {0x1EC, 0x304}, {0x1ED, 0x6F}, {0x1ED, 0x328}, {0x1ED, 0x304}, {0x1EE, 0x1B7}, +{0x1EE, 0x30C}, {0x1EF, 0x292}, {0x1EF, 0x30C}, {0x1F0, 0x6A}, {0x1F0, 0x30C}, {0x1F4, 0x47}, {0x1F4, 0x301}, {0x1F5, 0x67}, {0x1F5, 0x301}, {0x1F8, 0x4E}, {0x1F8, 0x300}, {0x1F9, 0x6E}, +{0x1F9, 0x300}, {0x1FA, 0x41}, {0x1FA, 0x30A}, {0x1FA, 0x301}, {0x1FB, 0x61}, {0x1FB, 0x30A}, {0x1FB, 0x301}, {0x1FC, 0xC6}, {0x1FC, 0x301}, {0x1FD, 0xE6}, {0x1FD, 0x301}, {0x1FE, 0xD8}, +{0x1FE, 0x301}, {0x1FF, 0xF8}, {0x1FF, 0x301}, {0x200, 0x41}, {0x200, 0x30F}, {0x201, 0x61}, {0x201, 0x30F}, {0x202, 0x41}, {0x202, 0x311}, {0x203, 0x61}, {0x203, 0x311}, {0x204, 0x45}, +{0x204, 0x30F}, {0x205, 0x65}, {0x205, 0x30F}, {0x206, 0x45}, {0x206, 0x311}, {0x207, 0x65}, {0x207, 0x311}, {0x208, 0x49}, {0x208, 0x30F}, {0x209, 0x69}, {0x209, 0x30F}, {0x20A, 0x49}, +{0x20A, 0x311}, {0x20B, 0x69}, {0x20B, 0x311}, {0x20C, 0x4F}, {0x20C, 0x30F}, {0x20D, 0x6F}, {0x20D, 0x30F}, {0x20E, 0x4F}, {0x20E, 0x311}, {0x20F, 0x6F}, {0x20F, 0x311}, {0x210, 0x52}, +{0x210, 0x30F}, {0x211, 0x72}, {0x211, 0x30F}, {0x212, 0x52}, {0x212, 0x311}, {0x213, 0x72}, {0x213, 0x311}, {0x214, 0x55}, {0x214, 0x30F}, {0x215, 0x75}, {0x215, 0x30F}, {0x216, 0x55}, +{0x216, 0x311}, {0x217, 0x75}, {0x217, 0x311}, {0x218, 0x53}, {0x218, 0x326}, {0x219, 0x73}, {0x219, 0x326}, {0x21A, 0x54}, {0x21A, 0x326}, {0x21B, 0x74}, {0x21B, 0x326}, {0x21E, 0x48}, +{0x21E, 0x30C}, {0x21F, 0x68}, {0x21F, 0x30C}, {0x226, 0x41}, {0x226, 0x307}, {0x227, 0x61}, {0x227, 0x307}, {0x228, 0x45}, {0x228, 0x327}, {0x229, 0x65}, {0x229, 0x327}, {0x22A, 0x4F}, +{0x22A, 0x308}, {0x22A, 0x304}, {0x22B, 0x6F}, {0x22B, 0x308}, {0x22B, 0x304}, {0x22C, 0x4F}, {0x22C, 0x303}, {0x22C, 0x304}, {0x22D, 0x6F}, {0x22D, 0x303}, {0x22D, 0x304}, {0x22E, 0x4F}, +{0x22E, 0x307}, {0x22F, 0x6F}, {0x22F, 0x307}, {0x230, 0x4F}, {0x230, 0x307}, {0x230, 0x304}, {0x231, 0x6F}, {0x231, 0x307}, {0x231, 0x304}, {0x232, 0x59}, {0x232, 0x304}, {0x233, 0x79}, +{0x233, 0x304}, {0x340, 0x300}, {0x341, 0x301}, {0x343, 0x313}, {0x344, 0x308}, {0x344, 0x301}, {0x374, 0x2B9}, {0x37E, 0x3B}, {0x385, 0xA8}, {0x385, 0x301}, {0x386, 0x391}, {0x386, 0x301}, +{0x387, 0xB7}, {0x388, 0x395}, {0x388, 0x301}, {0x389, 0x397}, {0x389, 0x301}, {0x38A, 0x399}, {0x38A, 0x301}, {0x38C, 0x39F}, {0x38C, 0x301}, {0x38E, 0x3A5}, {0x38E, 0x301}, {0x38F, 0x3A9}, +{0x38F, 0x301}, {0x390, 0x3B9}, {0x390, 0x308}, {0x390, 0x301}, {0x3AA, 0x399}, {0x3AA, 0x308}, {0x3AB, 0x3A5}, {0x3AB, 0x308}, {0x3AC, 0x3B1}, {0x3AC, 0x301}, {0x3AD, 0x3B5}, {0x3AD, 0x301}, +{0x3AE, 0x3B7}, {0x3AE, 0x301}, {0x3AF, 0x3B9}, {0x3AF, 0x301}, {0x3B0, 0x3C5}, {0x3B0, 0x308}, {0x3B0, 0x301}, {0x3CA, 0x3B9}, {0x3CA, 0x308}, {0x3CB, 0x3C5}, {0x3CB, 0x308}, {0x3CC, 0x3BF}, +{0x3CC, 0x301}, {0x3CD, 0x3C5}, {0x3CD, 0x301}, {0x3CE, 0x3C9}, {0x3CE, 0x301}, {0x3D3, 0x3D2}, {0x3D3, 0x301}, {0x3D4, 0x3D2}, {0x3D4, 0x308}, {0x400, 0x415}, {0x400, 0x300}, {0x401, 0x415}, +{0x401, 0x308}, {0x403, 0x413}, {0x403, 0x301}, {0x407, 0x406}, {0x407, 0x308}, {0x40C, 0x41A}, {0x40C, 0x301}, {0x40D, 0x418}, {0x40D, 0x300}, {0x40E, 0x423}, {0x40E, 0x306}, {0x419, 0x418}, +{0x419, 0x306}, {0x439, 0x438}, {0x439, 0x306}, {0x450, 0x435}, {0x450, 0x300}, {0x451, 0x435}, {0x451, 0x308}, {0x453, 0x433}, {0x453, 0x301}, {0x457, 0x456}, {0x457, 0x308}, {0x45C, 0x43A}, +{0x45C, 0x301}, {0x45D, 0x438}, {0x45D, 0x300}, {0x45E, 0x443}, {0x45E, 0x306}, {0x476, 0x474}, {0x476, 0x30F}, {0x477, 0x475}, {0x477, 0x30F}, {0x4C1, 0x416}, {0x4C1, 0x306}, {0x4C2, 0x436}, +{0x4C2, 0x306}, {0x4D0, 0x410}, {0x4D0, 0x306}, {0x4D1, 0x430}, {0x4D1, 0x306}, {0x4D2, 0x410}, {0x4D2, 0x308}, {0x4D3, 0x430}, {0x4D3, 0x308}, {0x4D6, 0x415}, {0x4D6, 0x306}, {0x4D7, 0x435}, +{0x4D7, 0x306}, {0x4DA, 0x4D8}, {0x4DA, 0x308}, {0x4DB, 0x4D9}, {0x4DB, 0x308}, {0x4DC, 0x416}, {0x4DC, 0x308}, {0x4DD, 0x436}, {0x4DD, 0x308}, {0x4DE, 0x417}, {0x4DE, 0x308}, {0x4DF, 0x437}, +{0x4DF, 0x308}, {0x4E2, 0x418}, {0x4E2, 0x304}, {0x4E3, 0x438}, {0x4E3, 0x304}, {0x4E4, 0x418}, {0x4E4, 0x308}, {0x4E5, 0x438}, {0x4E5, 0x308}, {0x4E6, 0x41E}, {0x4E6, 0x308}, {0x4E7, 0x43E}, +{0x4E7, 0x308}, {0x4EA, 0x4E8}, {0x4EA, 0x308}, {0x4EB, 0x4E9}, {0x4EB, 0x308}, {0x4EC, 0x42D}, {0x4EC, 0x308}, {0x4ED, 0x44D}, {0x4ED, 0x308}, {0x4EE, 0x423}, {0x4EE, 0x304}, {0x4EF, 0x443}, +{0x4EF, 0x304}, {0x4F0, 0x423}, {0x4F0, 0x308}, {0x4F1, 0x443}, {0x4F1, 0x308}, {0x4F2, 0x423}, {0x4F2, 0x30B}, {0x4F3, 0x443}, {0x4F3, 0x30B}, {0x4F4, 0x427}, {0x4F4, 0x308}, {0x4F5, 0x447}, +{0x4F5, 0x308}, {0x4F8, 0x42B}, {0x4F8, 0x308}, {0x4F9, 0x44B}, {0x4F9, 0x308}, {0x622, 0x627}, {0x622, 0x653}, {0x623, 0x627}, {0x623, 0x654}, {0x624, 0x648}, {0x624, 0x654}, {0x625, 0x627}, +{0x625, 0x655}, {0x626, 0x64A}, {0x626, 0x654}, {0x6C0, 0x6D5}, {0x6C0, 0x654}, {0x6C2, 0x6C1}, {0x6C2, 0x654}, {0x6D3, 0x6D2}, {0x6D3, 0x654}, {0x929, 0x928}, {0x929, 0x93C}, {0x931, 0x930}, +{0x931, 0x93C}, {0x934, 0x933}, {0x934, 0x93C}, {0x958, 0x915}, {0x958, 0x93C}, {0x959, 0x916}, {0x959, 0x93C}, {0x95A, 0x917}, {0x95A, 0x93C}, {0x95B, 0x91C}, {0x95B, 0x93C}, {0x95C, 0x921}, +{0x95C, 0x93C}, {0x95D, 0x922}, {0x95D, 0x93C}, {0x95E, 0x92B}, {0x95E, 0x93C}, {0x95F, 0x92F}, {0x95F, 0x93C}, {0x9CB, 0x9C7}, {0x9CB, 0x9BE}, {0x9CC, 0x9C7}, {0x9CC, 0x9D7}, {0x9DC, 0x9A1}, +{0x9DC, 0x9BC}, {0x9DD, 0x9A2}, {0x9DD, 0x9BC}, {0x9DF, 0x9AF}, {0x9DF, 0x9BC}, {0xA33, 0xA32}, {0xA33, 0xA3C}, {0xA36, 0xA38}, {0xA36, 0xA3C}, {0xA59, 0xA16}, {0xA59, 0xA3C}, {0xA5A, 0xA17}, +{0xA5A, 0xA3C}, {0xA5B, 0xA1C}, {0xA5B, 0xA3C}, {0xA5E, 0xA2B}, {0xA5E, 0xA3C}, {0xB48, 0xB47}, {0xB48, 0xB56}, {0xB4B, 0xB47}, {0xB4B, 0xB3E}, {0xB4C, 0xB47}, {0xB4C, 0xB57}, {0xB5C, 0xB21}, +{0xB5C, 0xB3C}, {0xB5D, 0xB22}, {0xB5D, 0xB3C}, {0xB94, 0xB92}, {0xB94, 0xBD7}, {0xBCA, 0xBC6}, {0xBCA, 0xBBE}, {0xBCB, 0xBC7}, {0xBCB, 0xBBE}, {0xBCC, 0xBC6}, {0xBCC, 0xBD7}, {0xC48, 0xC46}, +{0xC48, 0xC56}, {0xCC0, 0xCBF}, {0xCC0, 0xCD5}, {0xCC7, 0xCC6}, {0xCC7, 0xCD5}, {0xCC8, 0xCC6}, {0xCC8, 0xCD6}, {0xCCA, 0xCC6}, {0xCCA, 0xCC2}, {0xCCB, 0xCC6}, {0xCCB, 0xCC2}, {0xCCB, 0xCD5}, +{0xD4A, 0xD46}, {0xD4A, 0xD3E}, {0xD4B, 0xD47}, {0xD4B, 0xD3E}, {0xD4C, 0xD46}, {0xD4C, 0xD57}, {0xDDA, 0xDD9}, {0xDDA, 0xDCA}, {0xDDC, 0xDD9}, {0xDDC, 0xDCF}, {0xDDD, 0xDD9}, {0xDDD, 0xDCF}, +{0xDDD, 0xDCA}, {0xDDE, 0xDD9}, {0xDDE, 0xDDF}, {0xF43, 0xF42}, {0xF43, 0xFB7}, {0xF4D, 0xF4C}, {0xF4D, 0xFB7}, {0xF52, 0xF51}, {0xF52, 0xFB7}, {0xF57, 0xF56}, {0xF57, 0xFB7}, {0xF5C, 0xF5B}, +{0xF5C, 0xFB7}, {0xF69, 0xF40}, {0xF69, 0xFB5}, {0xF73, 0xF71}, {0xF73, 0xF72}, {0xF75, 0xF71}, {0xF75, 0xF74}, {0xF76, 0xFB2}, {0xF76, 0xF80}, {0xF78, 0xFB3}, {0xF78, 0xF80}, {0xF81, 0xF71}, +{0xF81, 0xF80}, {0xF93, 0xF92}, {0xF93, 0xFB7}, {0xF9D, 0xF9C}, {0xF9D, 0xFB7}, {0xFA2, 0xFA1}, {0xFA2, 0xFB7}, {0xFA7, 0xFA6}, {0xFA7, 0xFB7}, {0xFAC, 0xFAB}, {0xFAC, 0xFB7}, {0xFB9, 0xF90}, +{0xFB9, 0xFB5}, {0x1026, 0x1025}, {0x1026, 0x102E}, {0x1B06, 0x1B05}, {0x1B06, 0x1B35}, {0x1B08, 0x1B07}, {0x1B08, 0x1B35}, {0x1B0A, 0x1B09}, {0x1B0A, 0x1B35}, {0x1B0C, 0x1B0B}, {0x1B0C, 0x1B35}, +{0x1B0E, 0x1B0D}, {0x1B0E, 0x1B35}, {0x1B12, 0x1B11}, {0x1B12, 0x1B35}, {0x1B3B, 0x1B3A}, {0x1B3B, 0x1B35}, {0x1B3D, 0x1B3C}, {0x1B3D, 0x1B35}, {0x1B40, 0x1B3E}, {0x1B40, 0x1B35}, {0x1B41, 0x1B3F}, +{0x1B41, 0x1B35}, {0x1B43, 0x1B42}, {0x1B43, 0x1B35}, {0x1E00, 0x41}, {0x1E00, 0x325}, {0x1E01, 0x61}, {0x1E01, 0x325}, {0x1E02, 0x42}, {0x1E02, 0x307}, {0x1E03, 0x62}, {0x1E03, 0x307}, +{0x1E04, 0x42}, {0x1E04, 0x323}, {0x1E05, 0x62}, {0x1E05, 0x323}, {0x1E06, 0x42}, {0x1E06, 0x331}, {0x1E07, 0x62}, {0x1E07, 0x331}, {0x1E08, 0x43}, {0x1E08, 0x327}, {0x1E08, 0x301}, {0x1E09, 0x63}, +{0x1E09, 0x327}, {0x1E09, 0x301}, {0x1E0A, 0x44}, {0x1E0A, 0x307}, {0x1E0B, 0x64}, {0x1E0B, 0x307}, {0x1E0C, 0x44}, {0x1E0C, 0x323}, {0x1E0D, 0x64}, {0x1E0D, 0x323}, {0x1E0E, 0x44}, {0x1E0E, 0x331}, +{0x1E0F, 0x64}, {0x1E0F, 0x331}, {0x1E10, 0x44}, {0x1E10, 0x327}, {0x1E11, 0x64}, {0x1E11, 0x327}, {0x1E12, 0x44}, {0x1E12, 0x32D}, {0x1E13, 0x64}, {0x1E13, 0x32D}, {0x1E14, 0x45}, {0x1E14, 0x304}, +{0x1E14, 0x300}, {0x1E15, 0x65}, {0x1E15, 0x304}, {0x1E15, 0x300}, {0x1E16, 0x45}, {0x1E16, 0x304}, {0x1E16, 0x301}, {0x1E17, 0x65}, {0x1E17, 0x304}, {0x1E17, 0x301}, {0x1E18, 0x45}, {0x1E18, 0x32D}, +{0x1E19, 0x65}, {0x1E19, 0x32D}, {0x1E1A, 0x45}, {0x1E1A, 0x330}, {0x1E1B, 0x65}, {0x1E1B, 0x330}, {0x1E1C, 0x45}, {0x1E1C, 0x327}, {0x1E1C, 0x306}, {0x1E1D, 0x65}, {0x1E1D, 0x327}, {0x1E1D, 0x306}, +{0x1E1E, 0x46}, {0x1E1E, 0x307}, {0x1E1F, 0x66}, {0x1E1F, 0x307}, {0x1E20, 0x47}, {0x1E20, 0x304}, {0x1E21, 0x67}, {0x1E21, 0x304}, {0x1E22, 0x48}, {0x1E22, 0x307}, {0x1E23, 0x68}, {0x1E23, 0x307}, +{0x1E24, 0x48}, {0x1E24, 0x323}, {0x1E25, 0x68}, {0x1E25, 0x323}, {0x1E26, 0x48}, {0x1E26, 0x308}, {0x1E27, 0x68}, {0x1E27, 0x308}, {0x1E28, 0x48}, {0x1E28, 0x327}, {0x1E29, 0x68}, {0x1E29, 0x327}, +{0x1E2A, 0x48}, {0x1E2A, 0x32E}, {0x1E2B, 0x68}, {0x1E2B, 0x32E}, {0x1E2C, 0x49}, {0x1E2C, 0x330}, {0x1E2D, 0x69}, {0x1E2D, 0x330}, {0x1E2E, 0x49}, {0x1E2E, 0x308}, {0x1E2E, 0x301}, {0x1E2F, 0x69}, +{0x1E2F, 0x308}, {0x1E2F, 0x301}, {0x1E30, 0x4B}, {0x1E30, 0x301}, {0x1E31, 0x6B}, {0x1E31, 0x301}, {0x1E32, 0x4B}, {0x1E32, 0x323}, {0x1E33, 0x6B}, {0x1E33, 0x323}, {0x1E34, 0x4B}, {0x1E34, 0x331}, +{0x1E35, 0x6B}, {0x1E35, 0x331}, {0x1E36, 0x4C}, {0x1E36, 0x323}, {0x1E37, 0x6C}, {0x1E37, 0x323}, {0x1E38, 0x4C}, {0x1E38, 0x323}, {0x1E38, 0x304}, {0x1E39, 0x6C}, {0x1E39, 0x323}, {0x1E39, 0x304}, +{0x1E3A, 0x4C}, {0x1E3A, 0x331}, {0x1E3B, 0x6C}, {0x1E3B, 0x331}, {0x1E3C, 0x4C}, {0x1E3C, 0x32D}, {0x1E3D, 0x6C}, {0x1E3D, 0x32D}, {0x1E3E, 0x4D}, {0x1E3E, 0x301}, {0x1E3F, 0x6D}, {0x1E3F, 0x301}, +{0x1E40, 0x4D}, {0x1E40, 0x307}, {0x1E41, 0x6D}, {0x1E41, 0x307}, {0x1E42, 0x4D}, {0x1E42, 0x323}, {0x1E43, 0x6D}, {0x1E43, 0x323}, {0x1E44, 0x4E}, {0x1E44, 0x307}, {0x1E45, 0x6E}, {0x1E45, 0x307}, +{0x1E46, 0x4E}, {0x1E46, 0x323}, {0x1E47, 0x6E}, {0x1E47, 0x323}, {0x1E48, 0x4E}, {0x1E48, 0x331}, {0x1E49, 0x6E}, {0x1E49, 0x331}, {0x1E4A, 0x4E}, {0x1E4A, 0x32D}, {0x1E4B, 0x6E}, {0x1E4B, 0x32D}, +{0x1E4C, 0x4F}, {0x1E4C, 0x303}, {0x1E4C, 0x301}, {0x1E4D, 0x6F}, {0x1E4D, 0x303}, {0x1E4D, 0x301}, {0x1E4E, 0x4F}, {0x1E4E, 0x303}, {0x1E4E, 0x308}, {0x1E4F, 0x6F}, {0x1E4F, 0x303}, {0x1E4F, 0x308}, +{0x1E50, 0x4F}, {0x1E50, 0x304}, {0x1E50, 0x300}, {0x1E51, 0x6F}, {0x1E51, 0x304}, {0x1E51, 0x300}, {0x1E52, 0x4F}, {0x1E52, 0x304}, {0x1E52, 0x301}, {0x1E53, 0x6F}, {0x1E53, 0x304}, {0x1E53, 0x301}, +{0x1E54, 0x50}, {0x1E54, 0x301}, {0x1E55, 0x70}, {0x1E55, 0x301}, {0x1E56, 0x50}, {0x1E56, 0x307}, {0x1E57, 0x70}, {0x1E57, 0x307}, {0x1E58, 0x52}, {0x1E58, 0x307}, {0x1E59, 0x72}, {0x1E59, 0x307}, +{0x1E5A, 0x52}, {0x1E5A, 0x323}, {0x1E5B, 0x72}, {0x1E5B, 0x323}, {0x1E5C, 0x52}, {0x1E5C, 0x323}, {0x1E5C, 0x304}, {0x1E5D, 0x72}, {0x1E5D, 0x323}, {0x1E5D, 0x304}, {0x1E5E, 0x52}, {0x1E5E, 0x331}, +{0x1E5F, 0x72}, {0x1E5F, 0x331}, {0x1E60, 0x53}, {0x1E60, 0x307}, {0x1E61, 0x73}, {0x1E61, 0x307}, {0x1E62, 0x53}, {0x1E62, 0x323}, {0x1E63, 0x73}, {0x1E63, 0x323}, {0x1E64, 0x53}, {0x1E64, 0x301}, +{0x1E64, 0x307}, {0x1E65, 0x73}, {0x1E65, 0x301}, {0x1E65, 0x307}, {0x1E66, 0x53}, {0x1E66, 0x30C}, {0x1E66, 0x307}, {0x1E67, 0x73}, {0x1E67, 0x30C}, {0x1E67, 0x307}, {0x1E68, 0x53}, {0x1E68, 0x323}, +{0x1E68, 0x307}, {0x1E69, 0x73}, {0x1E69, 0x323}, {0x1E69, 0x307}, {0x1E6A, 0x54}, {0x1E6A, 0x307}, {0x1E6B, 0x74}, {0x1E6B, 0x307}, {0x1E6C, 0x54}, {0x1E6C, 0x323}, {0x1E6D, 0x74}, {0x1E6D, 0x323}, +{0x1E6E, 0x54}, {0x1E6E, 0x331}, {0x1E6F, 0x74}, {0x1E6F, 0x331}, {0x1E70, 0x54}, {0x1E70, 0x32D}, {0x1E71, 0x74}, {0x1E71, 0x32D}, {0x1E72, 0x55}, {0x1E72, 0x324}, {0x1E73, 0x75}, {0x1E73, 0x324}, +{0x1E74, 0x55}, {0x1E74, 0x330}, {0x1E75, 0x75}, {0x1E75, 0x330}, {0x1E76, 0x55}, {0x1E76, 0x32D}, {0x1E77, 0x75}, {0x1E77, 0x32D}, {0x1E78, 0x55}, {0x1E78, 0x303}, {0x1E78, 0x301}, {0x1E79, 0x75}, +{0x1E79, 0x303}, {0x1E79, 0x301}, {0x1E7A, 0x55}, {0x1E7A, 0x304}, {0x1E7A, 0x308}, {0x1E7B, 0x75}, {0x1E7B, 0x304}, {0x1E7B, 0x308}, {0x1E7C, 0x56}, {0x1E7C, 0x303}, {0x1E7D, 0x76}, {0x1E7D, 0x303}, +{0x1E7E, 0x56}, {0x1E7E, 0x323}, {0x1E7F, 0x76}, {0x1E7F, 0x323}, {0x1E80, 0x57}, {0x1E80, 0x300}, {0x1E81, 0x77}, {0x1E81, 0x300}, {0x1E82, 0x57}, {0x1E82, 0x301}, {0x1E83, 0x77}, {0x1E83, 0x301}, +{0x1E84, 0x57}, {0x1E84, 0x308}, {0x1E85, 0x77}, {0x1E85, 0x308}, {0x1E86, 0x57}, {0x1E86, 0x307}, {0x1E87, 0x77}, {0x1E87, 0x307}, {0x1E88, 0x57}, {0x1E88, 0x323}, {0x1E89, 0x77}, {0x1E89, 0x323}, +{0x1E8A, 0x58}, {0x1E8A, 0x307}, {0x1E8B, 0x78}, {0x1E8B, 0x307}, {0x1E8C, 0x58}, {0x1E8C, 0x308}, {0x1E8D, 0x78}, {0x1E8D, 0x308}, {0x1E8E, 0x59}, {0x1E8E, 0x307}, {0x1E8F, 0x79}, {0x1E8F, 0x307}, +{0x1E90, 0x5A}, {0x1E90, 0x302}, {0x1E91, 0x7A}, {0x1E91, 0x302}, {0x1E92, 0x5A}, {0x1E92, 0x323}, {0x1E93, 0x7A}, {0x1E93, 0x323}, {0x1E94, 0x5A}, {0x1E94, 0x331}, {0x1E95, 0x7A}, {0x1E95, 0x331}, +{0x1E96, 0x68}, {0x1E96, 0x331}, {0x1E97, 0x74}, {0x1E97, 0x308}, {0x1E98, 0x77}, {0x1E98, 0x30A}, {0x1E99, 0x79}, {0x1E99, 0x30A}, {0x1E9B, 0x17F}, {0x1E9B, 0x307}, {0x1EA0, 0x41}, {0x1EA0, 0x323}, +{0x1EA1, 0x61}, {0x1EA1, 0x323}, {0x1EA2, 0x41}, {0x1EA2, 0x309}, {0x1EA3, 0x61}, {0x1EA3, 0x309}, {0x1EA4, 0x41}, {0x1EA4, 0x302}, {0x1EA4, 0x301}, {0x1EA5, 0x61}, {0x1EA5, 0x302}, {0x1EA5, 0x301}, +{0x1EA6, 0x41}, {0x1EA6, 0x302}, {0x1EA6, 0x300}, {0x1EA7, 0x61}, {0x1EA7, 0x302}, {0x1EA7, 0x300}, {0x1EA8, 0x41}, {0x1EA8, 0x302}, {0x1EA8, 0x309}, {0x1EA9, 0x61}, {0x1EA9, 0x302}, {0x1EA9, 0x309}, +{0x1EAA, 0x41}, {0x1EAA, 0x302}, {0x1EAA, 0x303}, {0x1EAB, 0x61}, {0x1EAB, 0x302}, {0x1EAB, 0x303}, {0x1EAC, 0x41}, {0x1EAC, 0x323}, {0x1EAC, 0x302}, {0x1EAD, 0x61}, {0x1EAD, 0x323}, {0x1EAD, 0x302}, +{0x1EAE, 0x41}, {0x1EAE, 0x306}, {0x1EAE, 0x301}, {0x1EAF, 0x61}, {0x1EAF, 0x306}, {0x1EAF, 0x301}, {0x1EB0, 0x41}, {0x1EB0, 0x306}, {0x1EB0, 0x300}, {0x1EB1, 0x61}, {0x1EB1, 0x306}, {0x1EB1, 0x300}, +{0x1EB2, 0x41}, {0x1EB2, 0x306}, {0x1EB2, 0x309}, {0x1EB3, 0x61}, {0x1EB3, 0x306}, {0x1EB3, 0x309}, {0x1EB4, 0x41}, {0x1EB4, 0x306}, {0x1EB4, 0x303}, {0x1EB5, 0x61}, {0x1EB5, 0x306}, {0x1EB5, 0x303}, +{0x1EB6, 0x41}, {0x1EB6, 0x323}, {0x1EB6, 0x306}, {0x1EB7, 0x61}, {0x1EB7, 0x323}, {0x1EB7, 0x306}, {0x1EB8, 0x45}, {0x1EB8, 0x323}, {0x1EB9, 0x65}, {0x1EB9, 0x323}, {0x1EBA, 0x45}, {0x1EBA, 0x309}, +{0x1EBB, 0x65}, {0x1EBB, 0x309}, {0x1EBC, 0x45}, {0x1EBC, 0x303}, {0x1EBD, 0x65}, {0x1EBD, 0x303}, {0x1EBE, 0x45}, {0x1EBE, 0x302}, {0x1EBE, 0x301}, {0x1EBF, 0x65}, {0x1EBF, 0x302}, {0x1EBF, 0x301}, +{0x1EC0, 0x45}, {0x1EC0, 0x302}, {0x1EC0, 0x300}, {0x1EC1, 0x65}, {0x1EC1, 0x302}, {0x1EC1, 0x300}, {0x1EC2, 0x45}, {0x1EC2, 0x302}, {0x1EC2, 0x309}, {0x1EC3, 0x65}, {0x1EC3, 0x302}, {0x1EC3, 0x309}, +{0x1EC4, 0x45}, {0x1EC4, 0x302}, {0x1EC4, 0x303}, {0x1EC5, 0x65}, {0x1EC5, 0x302}, {0x1EC5, 0x303}, {0x1EC6, 0x45}, {0x1EC6, 0x323}, {0x1EC6, 0x302}, {0x1EC7, 0x65}, {0x1EC7, 0x323}, {0x1EC7, 0x302}, +{0x1EC8, 0x49}, {0x1EC8, 0x309}, {0x1EC9, 0x69}, {0x1EC9, 0x309}, {0x1ECA, 0x49}, {0x1ECA, 0x323}, {0x1ECB, 0x69}, {0x1ECB, 0x323}, {0x1ECC, 0x4F}, {0x1ECC, 0x323}, {0x1ECD, 0x6F}, {0x1ECD, 0x323}, +{0x1ECE, 0x4F}, {0x1ECE, 0x309}, {0x1ECF, 0x6F}, {0x1ECF, 0x309}, {0x1ED0, 0x4F}, {0x1ED0, 0x302}, {0x1ED0, 0x301}, {0x1ED1, 0x6F}, {0x1ED1, 0x302}, {0x1ED1, 0x301}, {0x1ED2, 0x4F}, {0x1ED2, 0x302}, +{0x1ED2, 0x300}, {0x1ED3, 0x6F}, {0x1ED3, 0x302}, {0x1ED3, 0x300}, {0x1ED4, 0x4F}, {0x1ED4, 0x302}, {0x1ED4, 0x309}, {0x1ED5, 0x6F}, {0x1ED5, 0x302}, {0x1ED5, 0x309}, {0x1ED6, 0x4F}, {0x1ED6, 0x302}, +{0x1ED6, 0x303}, {0x1ED7, 0x6F}, {0x1ED7, 0x302}, {0x1ED7, 0x303}, {0x1ED8, 0x4F}, {0x1ED8, 0x323}, {0x1ED8, 0x302}, {0x1ED9, 0x6F}, {0x1ED9, 0x323}, {0x1ED9, 0x302}, {0x1EDA, 0x4F}, {0x1EDA, 0x31B}, +{0x1EDA, 0x301}, {0x1EDB, 0x6F}, {0x1EDB, 0x31B}, {0x1EDB, 0x301}, {0x1EDC, 0x4F}, {0x1EDC, 0x31B}, {0x1EDC, 0x300}, {0x1EDD, 0x6F}, {0x1EDD, 0x31B}, {0x1EDD, 0x300}, {0x1EDE, 0x4F}, {0x1EDE, 0x31B}, +{0x1EDE, 0x309}, {0x1EDF, 0x6F}, {0x1EDF, 0x31B}, {0x1EDF, 0x309}, {0x1EE0, 0x4F}, {0x1EE0, 0x31B}, {0x1EE0, 0x303}, {0x1EE1, 0x6F}, {0x1EE1, 0x31B}, {0x1EE1, 0x303}, {0x1EE2, 0x4F}, {0x1EE2, 0x31B}, +{0x1EE2, 0x323}, {0x1EE3, 0x6F}, {0x1EE3, 0x31B}, {0x1EE3, 0x323}, {0x1EE4, 0x55}, {0x1EE4, 0x323}, {0x1EE5, 0x75}, {0x1EE5, 0x323}, {0x1EE6, 0x55}, {0x1EE6, 0x309}, {0x1EE7, 0x75}, {0x1EE7, 0x309}, +{0x1EE8, 0x55}, {0x1EE8, 0x31B}, {0x1EE8, 0x301}, {0x1EE9, 0x75}, {0x1EE9, 0x31B}, {0x1EE9, 0x301}, {0x1EEA, 0x55}, {0x1EEA, 0x31B}, {0x1EEA, 0x300}, {0x1EEB, 0x75}, {0x1EEB, 0x31B}, {0x1EEB, 0x300}, +{0x1EEC, 0x55}, {0x1EEC, 0x31B}, {0x1EEC, 0x309}, {0x1EED, 0x75}, {0x1EED, 0x31B}, {0x1EED, 0x309}, {0x1EEE, 0x55}, {0x1EEE, 0x31B}, {0x1EEE, 0x303}, {0x1EEF, 0x75}, {0x1EEF, 0x31B}, {0x1EEF, 0x303}, +{0x1EF0, 0x55}, {0x1EF0, 0x31B}, {0x1EF0, 0x323}, {0x1EF1, 0x75}, {0x1EF1, 0x31B}, {0x1EF1, 0x323}, {0x1EF2, 0x59}, {0x1EF2, 0x300}, {0x1EF3, 0x79}, {0x1EF3, 0x300}, {0x1EF4, 0x59}, {0x1EF4, 0x323}, +{0x1EF5, 0x79}, {0x1EF5, 0x323}, {0x1EF6, 0x59}, {0x1EF6, 0x309}, {0x1EF7, 0x79}, {0x1EF7, 0x309}, {0x1EF8, 0x59}, {0x1EF8, 0x303}, {0x1EF9, 0x79}, {0x1EF9, 0x303}, {0x1F00, 0x3B1}, {0x1F00, 0x313}, +{0x1F01, 0x3B1}, {0x1F01, 0x314}, {0x1F02, 0x3B1}, {0x1F02, 0x313}, {0x1F02, 0x300}, {0x1F03, 0x3B1}, {0x1F03, 0x314}, {0x1F03, 0x300}, {0x1F04, 0x3B1}, {0x1F04, 0x313}, {0x1F04, 0x301}, +{0x1F05, 0x3B1}, {0x1F05, 0x314}, {0x1F05, 0x301}, {0x1F06, 0x3B1}, {0x1F06, 0x313}, {0x1F06, 0x342}, {0x1F07, 0x3B1}, {0x1F07, 0x314}, {0x1F07, 0x342}, {0x1F08, 0x391}, {0x1F08, 0x313}, +{0x1F09, 0x391}, {0x1F09, 0x314}, {0x1F0A, 0x391}, {0x1F0A, 0x313}, {0x1F0A, 0x300}, {0x1F0B, 0x391}, {0x1F0B, 0x314}, {0x1F0B, 0x300}, {0x1F0C, 0x391}, {0x1F0C, 0x313}, {0x1F0C, 0x301}, +{0x1F0D, 0x391}, {0x1F0D, 0x314}, {0x1F0D, 0x301}, {0x1F0E, 0x391}, {0x1F0E, 0x313}, {0x1F0E, 0x342}, {0x1F0F, 0x391}, {0x1F0F, 0x314}, {0x1F0F, 0x342}, {0x1F10, 0x3B5}, {0x1F10, 0x313}, +{0x1F11, 0x3B5}, {0x1F11, 0x314}, {0x1F12, 0x3B5}, {0x1F12, 0x313}, {0x1F12, 0x300}, {0x1F13, 0x3B5}, {0x1F13, 0x314}, {0x1F13, 0x300}, {0x1F14, 0x3B5}, {0x1F14, 0x313}, {0x1F14, 0x301}, +{0x1F15, 0x3B5}, {0x1F15, 0x314}, {0x1F15, 0x301}, {0x1F18, 0x395}, {0x1F18, 0x313}, {0x1F19, 0x395}, {0x1F19, 0x314}, {0x1F1A, 0x395}, {0x1F1A, 0x313}, {0x1F1A, 0x300}, {0x1F1B, 0x395}, +{0x1F1B, 0x314}, {0x1F1B, 0x300}, {0x1F1C, 0x395}, {0x1F1C, 0x313}, {0x1F1C, 0x301}, {0x1F1D, 0x395}, {0x1F1D, 0x314}, {0x1F1D, 0x301}, {0x1F20, 0x3B7}, {0x1F20, 0x313}, {0x1F21, 0x3B7}, +{0x1F21, 0x314}, {0x1F22, 0x3B7}, {0x1F22, 0x313}, {0x1F22, 0x300}, {0x1F23, 0x3B7}, {0x1F23, 0x314}, {0x1F23, 0x300}, {0x1F24, 0x3B7}, {0x1F24, 0x313}, {0x1F24, 0x301}, {0x1F25, 0x3B7}, +{0x1F25, 0x314}, {0x1F25, 0x301}, {0x1F26, 0x3B7}, {0x1F26, 0x313}, {0x1F26, 0x342}, {0x1F27, 0x3B7}, {0x1F27, 0x314}, {0x1F27, 0x342}, {0x1F28, 0x397}, {0x1F28, 0x313}, {0x1F29, 0x397}, +{0x1F29, 0x314}, {0x1F2A, 0x397}, {0x1F2A, 0x313}, {0x1F2A, 0x300}, {0x1F2B, 0x397}, {0x1F2B, 0x314}, {0x1F2B, 0x300}, {0x1F2C, 0x397}, {0x1F2C, 0x313}, {0x1F2C, 0x301}, {0x1F2D, 0x397}, +{0x1F2D, 0x314}, {0x1F2D, 0x301}, {0x1F2E, 0x397}, {0x1F2E, 0x313}, {0x1F2E, 0x342}, {0x1F2F, 0x397}, {0x1F2F, 0x314}, {0x1F2F, 0x342}, {0x1F30, 0x3B9}, {0x1F30, 0x313}, {0x1F31, 0x3B9}, +{0x1F31, 0x314}, {0x1F32, 0x3B9}, {0x1F32, 0x313}, {0x1F32, 0x300}, {0x1F33, 0x3B9}, {0x1F33, 0x314}, {0x1F33, 0x300}, {0x1F34, 0x3B9}, {0x1F34, 0x313}, {0x1F34, 0x301}, {0x1F35, 0x3B9}, +{0x1F35, 0x314}, {0x1F35, 0x301}, {0x1F36, 0x3B9}, {0x1F36, 0x313}, {0x1F36, 0x342}, {0x1F37, 0x3B9}, {0x1F37, 0x314}, {0x1F37, 0x342}, {0x1F38, 0x399}, {0x1F38, 0x313}, {0x1F39, 0x399}, +{0x1F39, 0x314}, {0x1F3A, 0x399}, {0x1F3A, 0x313}, {0x1F3A, 0x300}, {0x1F3B, 0x399}, {0x1F3B, 0x314}, {0x1F3B, 0x300}, {0x1F3C, 0x399}, {0x1F3C, 0x313}, {0x1F3C, 0x301}, {0x1F3D, 0x399}, +{0x1F3D, 0x314}, {0x1F3D, 0x301}, {0x1F3E, 0x399}, {0x1F3E, 0x313}, {0x1F3E, 0x342}, {0x1F3F, 0x399}, {0x1F3F, 0x314}, {0x1F3F, 0x342}, {0x1F40, 0x3BF}, {0x1F40, 0x313}, {0x1F41, 0x3BF}, +{0x1F41, 0x314}, {0x1F42, 0x3BF}, {0x1F42, 0x313}, {0x1F42, 0x300}, {0x1F43, 0x3BF}, {0x1F43, 0x314}, {0x1F43, 0x300}, {0x1F44, 0x3BF}, {0x1F44, 0x313}, {0x1F44, 0x301}, {0x1F45, 0x3BF}, +{0x1F45, 0x314}, {0x1F45, 0x301}, {0x1F48, 0x39F}, {0x1F48, 0x313}, {0x1F49, 0x39F}, {0x1F49, 0x314}, {0x1F4A, 0x39F}, {0x1F4A, 0x313}, {0x1F4A, 0x300}, {0x1F4B, 0x39F}, {0x1F4B, 0x314}, +{0x1F4B, 0x300}, {0x1F4C, 0x39F}, {0x1F4C, 0x313}, {0x1F4C, 0x301}, {0x1F4D, 0x39F}, {0x1F4D, 0x314}, {0x1F4D, 0x301}, {0x1F50, 0x3C5}, {0x1F50, 0x313}, {0x1F51, 0x3C5}, {0x1F51, 0x314}, +{0x1F52, 0x3C5}, {0x1F52, 0x313}, {0x1F52, 0x300}, {0x1F53, 0x3C5}, {0x1F53, 0x314}, {0x1F53, 0x300}, {0x1F54, 0x3C5}, {0x1F54, 0x313}, {0x1F54, 0x301}, {0x1F55, 0x3C5}, {0x1F55, 0x314}, +{0x1F55, 0x301}, {0x1F56, 0x3C5}, {0x1F56, 0x313}, {0x1F56, 0x342}, {0x1F57, 0x3C5}, {0x1F57, 0x314}, {0x1F57, 0x342}, {0x1F59, 0x3A5}, {0x1F59, 0x314}, {0x1F5B, 0x3A5}, {0x1F5B, 0x314}, +{0x1F5B, 0x300}, {0x1F5D, 0x3A5}, {0x1F5D, 0x314}, {0x1F5D, 0x301}, {0x1F5F, 0x3A5}, {0x1F5F, 0x314}, {0x1F5F, 0x342}, {0x1F60, 0x3C9}, {0x1F60, 0x313}, {0x1F61, 0x3C9}, {0x1F61, 0x314}, +{0x1F62, 0x3C9}, {0x1F62, 0x313}, {0x1F62, 0x300}, {0x1F63, 0x3C9}, {0x1F63, 0x314}, {0x1F63, 0x300}, {0x1F64, 0x3C9}, {0x1F64, 0x313}, {0x1F64, 0x301}, {0x1F65, 0x3C9}, {0x1F65, 0x314}, +{0x1F65, 0x301}, {0x1F66, 0x3C9}, {0x1F66, 0x313}, {0x1F66, 0x342}, {0x1F67, 0x3C9}, {0x1F67, 0x314}, {0x1F67, 0x342}, {0x1F68, 0x3A9}, {0x1F68, 0x313}, {0x1F69, 0x3A9}, {0x1F69, 0x314}, +{0x1F6A, 0x3A9}, {0x1F6A, 0x313}, {0x1F6A, 0x300}, {0x1F6B, 0x3A9}, {0x1F6B, 0x314}, {0x1F6B, 0x300}, {0x1F6C, 0x3A9}, {0x1F6C, 0x313}, {0x1F6C, 0x301}, {0x1F6D, 0x3A9}, {0x1F6D, 0x314}, +{0x1F6D, 0x301}, {0x1F6E, 0x3A9}, {0x1F6E, 0x313}, {0x1F6E, 0x342}, {0x1F6F, 0x3A9}, {0x1F6F, 0x314}, {0x1F6F, 0x342}, {0x1F70, 0x3B1}, {0x1F70, 0x300}, {0x1F71, 0x3B1}, {0x1F71, 0x301}, +{0x1F72, 0x3B5}, {0x1F72, 0x300}, {0x1F73, 0x3B5}, {0x1F73, 0x301}, {0x1F74, 0x3B7}, {0x1F74, 0x300}, {0x1F75, 0x3B7}, {0x1F75, 0x301}, {0x1F76, 0x3B9}, {0x1F76, 0x300}, {0x1F77, 0x3B9}, +{0x1F77, 0x301}, {0x1F78, 0x3BF}, {0x1F78, 0x300}, {0x1F79, 0x3BF}, {0x1F79, 0x301}, {0x1F7A, 0x3C5}, {0x1F7A, 0x300}, {0x1F7B, 0x3C5}, {0x1F7B, 0x301}, {0x1F7C, 0x3C9}, {0x1F7C, 0x300}, +{0x1F7D, 0x3C9}, {0x1F7D, 0x301}, {0x1F80, 0x3B1}, {0x1F80, 0x313}, {0x1F80, 0x345}, {0x1F81, 0x3B1}, {0x1F81, 0x314}, {0x1F81, 0x345}, {0x1F82, 0x3B1}, {0x1F82, 0x313}, {0x1F82, 0x300}, +{0x1F82, 0x345}, {0x1F83, 0x3B1}, {0x1F83, 0x314}, {0x1F83, 0x300}, {0x1F83, 0x345}, {0x1F84, 0x3B1}, {0x1F84, 0x313}, {0x1F84, 0x301}, {0x1F84, 0x345}, {0x1F85, 0x3B1}, {0x1F85, 0x314}, +{0x1F85, 0x301}, {0x1F85, 0x345}, {0x1F86, 0x3B1}, {0x1F86, 0x313}, {0x1F86, 0x342}, {0x1F86, 0x345}, {0x1F87, 0x3B1}, {0x1F87, 0x314}, {0x1F87, 0x342}, {0x1F87, 0x345}, {0x1F88, 0x391}, +{0x1F88, 0x313}, {0x1F88, 0x345}, {0x1F89, 0x391}, {0x1F89, 0x314}, {0x1F89, 0x345}, {0x1F8A, 0x391}, {0x1F8A, 0x313}, {0x1F8A, 0x300}, {0x1F8A, 0x345}, {0x1F8B, 0x391}, {0x1F8B, 0x314}, +{0x1F8B, 0x300}, {0x1F8B, 0x345}, {0x1F8C, 0x391}, {0x1F8C, 0x313}, {0x1F8C, 0x301}, {0x1F8C, 0x345}, {0x1F8D, 0x391}, {0x1F8D, 0x314}, {0x1F8D, 0x301}, {0x1F8D, 0x345}, {0x1F8E, 0x391}, +{0x1F8E, 0x313}, {0x1F8E, 0x342}, {0x1F8E, 0x345}, {0x1F8F, 0x391}, {0x1F8F, 0x314}, {0x1F8F, 0x342}, {0x1F8F, 0x345}, {0x1F90, 0x3B7}, {0x1F90, 0x313}, {0x1F90, 0x345}, {0x1F91, 0x3B7}, +{0x1F91, 0x314}, {0x1F91, 0x345}, {0x1F92, 0x3B7}, {0x1F92, 0x313}, {0x1F92, 0x300}, {0x1F92, 0x345}, {0x1F93, 0x3B7}, {0x1F93, 0x314}, {0x1F93, 0x300}, {0x1F93, 0x345}, {0x1F94, 0x3B7}, +{0x1F94, 0x313}, {0x1F94, 0x301}, {0x1F94, 0x345}, {0x1F95, 0x3B7}, {0x1F95, 0x314}, {0x1F95, 0x301}, {0x1F95, 0x345}, {0x1F96, 0x3B7}, {0x1F96, 0x313}, {0x1F96, 0x342}, {0x1F96, 0x345}, +{0x1F97, 0x3B7}, {0x1F97, 0x314}, {0x1F97, 0x342}, {0x1F97, 0x345}, {0x1F98, 0x397}, {0x1F98, 0x313}, {0x1F98, 0x345}, {0x1F99, 0x397}, {0x1F99, 0x314}, {0x1F99, 0x345}, {0x1F9A, 0x397}, +{0x1F9A, 0x313}, {0x1F9A, 0x300}, {0x1F9A, 0x345}, {0x1F9B, 0x397}, {0x1F9B, 0x314}, {0x1F9B, 0x300}, {0x1F9B, 0x345}, {0x1F9C, 0x397}, {0x1F9C, 0x313}, {0x1F9C, 0x301}, {0x1F9C, 0x345}, +{0x1F9D, 0x397}, {0x1F9D, 0x314}, {0x1F9D, 0x301}, {0x1F9D, 0x345}, {0x1F9E, 0x397}, {0x1F9E, 0x313}, {0x1F9E, 0x342}, {0x1F9E, 0x345}, {0x1F9F, 0x397}, {0x1F9F, 0x314}, {0x1F9F, 0x342}, +{0x1F9F, 0x345}, {0x1FA0, 0x3C9}, {0x1FA0, 0x313}, {0x1FA0, 0x345}, {0x1FA1, 0x3C9}, {0x1FA1, 0x314}, {0x1FA1, 0x345}, {0x1FA2, 0x3C9}, {0x1FA2, 0x313}, {0x1FA2, 0x300}, {0x1FA2, 0x345}, +{0x1FA3, 0x3C9}, {0x1FA3, 0x314}, {0x1FA3, 0x300}, {0x1FA3, 0x345}, {0x1FA4, 0x3C9}, {0x1FA4, 0x313}, {0x1FA4, 0x301}, {0x1FA4, 0x345}, {0x1FA5, 0x3C9}, {0x1FA5, 0x314}, {0x1FA5, 0x301}, +{0x1FA5, 0x345}, {0x1FA6, 0x3C9}, {0x1FA6, 0x313}, {0x1FA6, 0x342}, {0x1FA6, 0x345}, {0x1FA7, 0x3C9}, {0x1FA7, 0x314}, {0x1FA7, 0x342}, {0x1FA7, 0x345}, {0x1FA8, 0x3A9}, {0x1FA8, 0x313}, +{0x1FA8, 0x345}, {0x1FA9, 0x3A9}, {0x1FA9, 0x314}, {0x1FA9, 0x345}, {0x1FAA, 0x3A9}, {0x1FAA, 0x313}, {0x1FAA, 0x300}, {0x1FAA, 0x345}, {0x1FAB, 0x3A9}, {0x1FAB, 0x314}, {0x1FAB, 0x300}, +{0x1FAB, 0x345}, {0x1FAC, 0x3A9}, {0x1FAC, 0x313}, {0x1FAC, 0x301}, {0x1FAC, 0x345}, {0x1FAD, 0x3A9}, {0x1FAD, 0x314}, {0x1FAD, 0x301}, {0x1FAD, 0x345}, {0x1FAE, 0x3A9}, {0x1FAE, 0x313}, +{0x1FAE, 0x342}, {0x1FAE, 0x345}, {0x1FAF, 0x3A9}, {0x1FAF, 0x314}, {0x1FAF, 0x342}, {0x1FAF, 0x345}, {0x1FB0, 0x3B1}, {0x1FB0, 0x306}, {0x1FB1, 0x3B1}, {0x1FB1, 0x304}, {0x1FB2, 0x3B1}, +{0x1FB2, 0x300}, {0x1FB2, 0x345}, {0x1FB3, 0x3B1}, {0x1FB3, 0x345}, {0x1FB4, 0x3B1}, {0x1FB4, 0x301}, {0x1FB4, 0x345}, {0x1FB6, 0x3B1}, {0x1FB6, 0x342}, {0x1FB7, 0x3B1}, {0x1FB7, 0x342}, +{0x1FB7, 0x345}, {0x1FB8, 0x391}, {0x1FB8, 0x306}, {0x1FB9, 0x391}, {0x1FB9, 0x304}, {0x1FBA, 0x391}, {0x1FBA, 0x300}, {0x1FBB, 0x391}, {0x1FBB, 0x301}, {0x1FBC, 0x391}, {0x1FBC, 0x345}, +{0x1FBE, 0x3B9}, {0x1FC1, 0xA8}, {0x1FC1, 0x342}, {0x1FC2, 0x3B7}, {0x1FC2, 0x300}, {0x1FC2, 0x345}, {0x1FC3, 0x3B7}, {0x1FC3, 0x345}, {0x1FC4, 0x3B7}, {0x1FC4, 0x301}, {0x1FC4, 0x345}, +{0x1FC6, 0x3B7}, {0x1FC6, 0x342}, {0x1FC7, 0x3B7}, {0x1FC7, 0x342}, {0x1FC7, 0x345}, {0x1FC8, 0x395}, {0x1FC8, 0x300}, {0x1FC9, 0x395}, {0x1FC9, 0x301}, {0x1FCA, 0x397}, {0x1FCA, 0x300}, +{0x1FCB, 0x397}, {0x1FCB, 0x301}, {0x1FCC, 0x397}, {0x1FCC, 0x345}, {0x1FCD, 0x1FBF}, {0x1FCD, 0x300}, {0x1FCE, 0x1FBF}, {0x1FCE, 0x301}, {0x1FCF, 0x1FBF}, {0x1FCF, 0x342}, {0x1FD0, 0x3B9}, +{0x1FD0, 0x306}, {0x1FD1, 0x3B9}, {0x1FD1, 0x304}, {0x1FD2, 0x3B9}, {0x1FD2, 0x308}, {0x1FD2, 0x300}, {0x1FD3, 0x3B9}, {0x1FD3, 0x308}, {0x1FD3, 0x301}, {0x1FD6, 0x3B9}, {0x1FD6, 0x342}, +{0x1FD7, 0x3B9}, {0x1FD7, 0x308}, {0x1FD7, 0x342}, {0x1FD8, 0x399}, {0x1FD8, 0x306}, {0x1FD9, 0x399}, {0x1FD9, 0x304}, {0x1FDA, 0x399}, {0x1FDA, 0x300}, {0x1FDB, 0x399}, {0x1FDB, 0x301}, +{0x1FDD, 0x1FFE}, {0x1FDD, 0x300}, {0x1FDE, 0x1FFE}, {0x1FDE, 0x301}, {0x1FDF, 0x1FFE}, {0x1FDF, 0x342}, {0x1FE0, 0x3C5}, {0x1FE0, 0x306}, {0x1FE1, 0x3C5}, {0x1FE1, 0x304}, {0x1FE2, 0x3C5}, +{0x1FE2, 0x308}, {0x1FE2, 0x300}, {0x1FE3, 0x3C5}, {0x1FE3, 0x308}, {0x1FE3, 0x301}, {0x1FE4, 0x3C1}, {0x1FE4, 0x313}, {0x1FE5, 0x3C1}, {0x1FE5, 0x314}, {0x1FE6, 0x3C5}, {0x1FE6, 0x342}, +{0x1FE7, 0x3C5}, {0x1FE7, 0x308}, {0x1FE7, 0x342}, {0x1FE8, 0x3A5}, {0x1FE8, 0x306}, {0x1FE9, 0x3A5}, {0x1FE9, 0x304}, {0x1FEA, 0x3A5}, {0x1FEA, 0x300}, {0x1FEB, 0x3A5}, {0x1FEB, 0x301}, +{0x1FEC, 0x3A1}, {0x1FEC, 0x314}, {0x1FED, 0xA8}, {0x1FED, 0x300}, {0x1FEE, 0xA8}, {0x1FEE, 0x301}, {0x1FEF, 0x60}, {0x1FF2, 0x3C9}, {0x1FF2, 0x300}, {0x1FF2, 0x345}, {0x1FF3, 0x3C9}, {0x1FF3, 0x345}, +{0x1FF4, 0x3C9}, {0x1FF4, 0x301}, {0x1FF4, 0x345}, {0x1FF6, 0x3C9}, {0x1FF6, 0x342}, {0x1FF7, 0x3C9}, {0x1FF7, 0x342}, {0x1FF7, 0x345}, {0x1FF8, 0x39F}, {0x1FF8, 0x300}, {0x1FF9, 0x39F}, +{0x1FF9, 0x301}, {0x1FFA, 0x3A9}, {0x1FFA, 0x300}, {0x1FFB, 0x3A9}, {0x1FFB, 0x301}, {0x1FFC, 0x3A9}, {0x1FFC, 0x345}, {0x1FFD, 0xB4}, {0x2000, 0x2002}, {0x2001, 0x2003}, {0x2126, 0x3A9}, +{0x212A, 0x4B}, {0x212B, 0x41}, {0x212B, 0x30A}, {0x219A, 0x2190}, {0x219A, 0x338}, {0x219B, 0x2192}, {0x219B, 0x338}, {0x21AE, 0x2194}, {0x21AE, 0x338}, {0x21CD, 0x21D0}, {0x21CD, 0x338}, +{0x21CE, 0x21D4}, {0x21CE, 0x338}, {0x21CF, 0x21D2}, {0x21CF, 0x338}, {0x2204, 0x2203}, {0x2204, 0x338}, {0x2209, 0x2208}, {0x2209, 0x338}, {0x220C, 0x220B}, {0x220C, 0x338}, {0x2224, 0x2223}, +{0x2224, 0x338}, {0x2226, 0x2225}, {0x2226, 0x338}, {0x2241, 0x223C}, {0x2241, 0x338}, {0x2244, 0x2243}, {0x2244, 0x338}, {0x2247, 0x2245}, {0x2247, 0x338}, {0x2249, 0x2248}, {0x2249, 0x338}, +{0x2260, 0x3D}, {0x2260, 0x338}, {0x2262, 0x2261}, {0x2262, 0x338}, {0x226D, 0x224D}, {0x226D, 0x338}, {0x226E, 0x3C}, {0x226E, 0x338}, {0x226F, 0x3E}, {0x226F, 0x338}, {0x2270, 0x2264}, +{0x2270, 0x338}, {0x2271, 0x2265}, {0x2271, 0x338}, {0x2274, 0x2272}, {0x2274, 0x338}, {0x2275, 0x2273}, {0x2275, 0x338}, {0x2278, 0x2276}, {0x2278, 0x338}, {0x2279, 0x2277}, {0x2279, 0x338}, +{0x2280, 0x227A}, {0x2280, 0x338}, {0x2281, 0x227B}, {0x2281, 0x338}, {0x2284, 0x2282}, {0x2284, 0x338}, {0x2285, 0x2283}, {0x2285, 0x338}, {0x2288, 0x2286}, {0x2288, 0x338}, {0x2289, 0x2287}, +{0x2289, 0x338}, {0x22AC, 0x22A2}, {0x22AC, 0x338}, {0x22AD, 0x22A8}, {0x22AD, 0x338}, {0x22AE, 0x22A9}, {0x22AE, 0x338}, {0x22AF, 0x22AB}, {0x22AF, 0x338}, {0x22E0, 0x227C}, {0x22E0, 0x338}, +{0x22E1, 0x227D}, {0x22E1, 0x338}, {0x22E2, 0x2291}, {0x22E2, 0x338}, {0x22E3, 0x2292}, {0x22E3, 0x338}, {0x22EA, 0x22B2}, {0x22EA, 0x338}, {0x22EB, 0x22B3}, {0x22EB, 0x338}, {0x22EC, 0x22B4}, +{0x22EC, 0x338}, {0x22ED, 0x22B5}, {0x22ED, 0x338}, {0x2329, 0x3008}, {0x232A, 0x3009}, {0x2ADC, 0x2ADD}, {0x2ADC, 0x338}, {0x304C, 0x304B}, {0x304C, 0x3099}, {0x304E, 0x304D}, {0x304E, 0x3099}, +{0x3050, 0x304F}, {0x3050, 0x3099}, {0x3052, 0x3051}, {0x3052, 0x3099}, {0x3054, 0x3053}, {0x3054, 0x3099}, {0x3056, 0x3055}, {0x3056, 0x3099}, {0x3058, 0x3057}, {0x3058, 0x3099}, {0x305A, 0x3059}, +{0x305A, 0x3099}, {0x305C, 0x305B}, {0x305C, 0x3099}, {0x305E, 0x305D}, {0x305E, 0x3099}, {0x3060, 0x305F}, {0x3060, 0x3099}, {0x3062, 0x3061}, {0x3062, 0x3099}, {0x3065, 0x3064}, {0x3065, 0x3099}, +{0x3067, 0x3066}, {0x3067, 0x3099}, {0x3069, 0x3068}, {0x3069, 0x3099}, {0x3070, 0x306F}, {0x3070, 0x3099}, {0x3071, 0x306F}, {0x3071, 0x309A}, {0x3073, 0x3072}, {0x3073, 0x3099}, {0x3074, 0x3072}, +{0x3074, 0x309A}, {0x3076, 0x3075}, {0x3076, 0x3099}, {0x3077, 0x3075}, {0x3077, 0x309A}, {0x3079, 0x3078}, {0x3079, 0x3099}, {0x307A, 0x3078}, {0x307A, 0x309A}, {0x307C, 0x307B}, {0x307C, 0x3099}, +{0x307D, 0x307B}, {0x307D, 0x309A}, {0x3094, 0x3046}, {0x3094, 0x3099}, {0x309E, 0x309D}, {0x309E, 0x3099}, {0x30AC, 0x30AB}, {0x30AC, 0x3099}, {0x30AE, 0x30AD}, {0x30AE, 0x3099}, {0x30B0, 0x30AF}, +{0x30B0, 0x3099}, {0x30B2, 0x30B1}, {0x30B2, 0x3099}, {0x30B4, 0x30B3}, {0x30B4, 0x3099}, {0x30B6, 0x30B5}, {0x30B6, 0x3099}, {0x30B8, 0x30B7}, {0x30B8, 0x3099}, {0x30BA, 0x30B9}, {0x30BA, 0x3099}, +{0x30BC, 0x30BB}, {0x30BC, 0x3099}, {0x30BE, 0x30BD}, {0x30BE, 0x3099}, {0x30C0, 0x30BF}, {0x30C0, 0x3099}, {0x30C2, 0x30C1}, {0x30C2, 0x3099}, {0x30C5, 0x30C4}, {0x30C5, 0x3099}, {0x30C7, 0x30C6}, +{0x30C7, 0x3099}, {0x30C9, 0x30C8}, {0x30C9, 0x3099}, {0x30D0, 0x30CF}, {0x30D0, 0x3099}, {0x30D1, 0x30CF}, {0x30D1, 0x309A}, {0x30D3, 0x30D2}, {0x30D3, 0x3099}, {0x30D4, 0x30D2}, {0x30D4, 0x309A}, +{0x30D6, 0x30D5}, {0x30D6, 0x3099}, {0x30D7, 0x30D5}, {0x30D7, 0x309A}, {0x30D9, 0x30D8}, {0x30D9, 0x3099}, {0x30DA, 0x30D8}, {0x30DA, 0x309A}, {0x30DC, 0x30DB}, {0x30DC, 0x3099}, {0x30DD, 0x30DB}, +{0x30DD, 0x309A}, {0x30F4, 0x30A6}, {0x30F4, 0x3099}, {0x30F7, 0x30EF}, {0x30F7, 0x3099}, {0x30F8, 0x30F0}, {0x30F8, 0x3099}, {0x30F9, 0x30F1}, {0x30F9, 0x3099}, {0x30FA, 0x30F2}, {0x30FA, 0x3099}, +{0x30FE, 0x30FD}, {0x30FE, 0x3099}, {0xF900, 0x8C48}, {0xF901, 0x66F4}, {0xF902, 0x8ECA}, {0xF903, 0x8CC8}, {0xF904, 0x6ED1}, {0xF905, 0x4E32}, {0xF906, 0x53E5}, {0xF907, 0x9F9C}, {0xF908, 0x9F9C}, +{0xF909, 0x5951}, {0xF90A, 0x91D1}, {0xF90B, 0x5587}, {0xF90C, 0x5948}, {0xF90D, 0x61F6}, {0xF90E, 0x7669}, {0xF90F, 0x7F85}, {0xF910, 0x863F}, {0xF911, 0x87BA}, {0xF912, 0x88F8}, {0xF913, 0x908F}, +{0xF914, 0x6A02}, {0xF915, 0x6D1B}, {0xF916, 0x70D9}, {0xF917, 0x73DE}, {0xF918, 0x843D}, {0xF919, 0x916A}, {0xF91A, 0x99F1}, {0xF91B, 0x4E82}, {0xF91C, 0x5375}, {0xF91D, 0x6B04}, {0xF91E, 0x721B}, +{0xF91F, 0x862D}, {0xF920, 0x9E1E}, {0xF921, 0x5D50}, {0xF922, 0x6FEB}, {0xF923, 0x85CD}, {0xF924, 0x8964}, {0xF925, 0x62C9}, {0xF926, 0x81D8}, {0xF927, 0x881F}, {0xF928, 0x5ECA}, {0xF929, 0x6717}, +{0xF92A, 0x6D6A}, {0xF92B, 0x72FC}, {0xF92C, 0x90CE}, {0xF92D, 0x4F86}, {0xF92E, 0x51B7}, {0xF92F, 0x52DE}, {0xF930, 0x64C4}, {0xF931, 0x6AD3}, {0xF932, 0x7210}, {0xF933, 0x76E7}, {0xF934, 0x8001}, +{0xF935, 0x8606}, {0xF936, 0x865C}, {0xF937, 0x8DEF}, {0xF938, 0x9732}, {0xF939, 0x9B6F}, {0xF93A, 0x9DFA}, {0xF93B, 0x788C}, {0xF93C, 0x797F}, {0xF93D, 0x7DA0}, {0xF93E, 0x83C9}, {0xF93F, 0x9304}, +{0xF940, 0x9E7F}, {0xF941, 0x8AD6}, {0xF942, 0x58DF}, {0xF943, 0x5F04}, {0xF944, 0x7C60}, {0xF945, 0x807E}, {0xF946, 0x7262}, {0xF947, 0x78CA}, {0xF948, 0x8CC2}, {0xF949, 0x96F7}, {0xF94A, 0x58D8}, +{0xF94B, 0x5C62}, {0xF94C, 0x6A13}, {0xF94D, 0x6DDA}, {0xF94E, 0x6F0F}, {0xF94F, 0x7D2F}, {0xF950, 0x7E37}, {0xF951, 0x964B}, {0xF952, 0x52D2}, {0xF953, 0x808B}, {0xF954, 0x51DC}, {0xF955, 0x51CC}, +{0xF956, 0x7A1C}, {0xF957, 0x7DBE}, {0xF958, 0x83F1}, {0xF959, 0x9675}, {0xF95A, 0x8B80}, {0xF95B, 0x62CF}, {0xF95C, 0x6A02}, {0xF95D, 0x8AFE}, {0xF95E, 0x4E39}, {0xF95F, 0x5BE7}, {0xF960, 0x6012}, +{0xF961, 0x7387}, {0xF962, 0x7570}, {0xF963, 0x5317}, {0xF964, 0x78FB}, {0xF965, 0x4FBF}, {0xF966, 0x5FA9}, {0xF967, 0x4E0D}, {0xF968, 0x6CCC}, {0xF969, 0x6578}, {0xF96A, 0x7D22}, {0xF96B, 0x53C3}, +{0xF96C, 0x585E}, {0xF96D, 0x7701}, {0xF96E, 0x8449}, {0xF96F, 0x8AAA}, {0xF970, 0x6BBA}, {0xF971, 0x8FB0}, {0xF972, 0x6C88}, {0xF973, 0x62FE}, {0xF974, 0x82E5}, {0xF975, 0x63A0}, {0xF976, 0x7565}, +{0xF977, 0x4EAE}, {0xF978, 0x5169}, {0xF979, 0x51C9}, {0xF97A, 0x6881}, {0xF97B, 0x7CE7}, {0xF97C, 0x826F}, {0xF97D, 0x8AD2}, {0xF97E, 0x91CF}, {0xF97F, 0x52F5}, {0xF980, 0x5442}, {0xF981, 0x5973}, +{0xF982, 0x5EEC}, {0xF983, 0x65C5}, {0xF984, 0x6FFE}, {0xF985, 0x792A}, {0xF986, 0x95AD}, {0xF987, 0x9A6A}, {0xF988, 0x9E97}, {0xF989, 0x9ECE}, {0xF98A, 0x529B}, {0xF98B, 0x66C6}, {0xF98C, 0x6B77}, +{0xF98D, 0x8F62}, {0xF98E, 0x5E74}, {0xF98F, 0x6190}, {0xF990, 0x6200}, {0xF991, 0x649A}, {0xF992, 0x6F23}, {0xF993, 0x7149}, {0xF994, 0x7489}, {0xF995, 0x79CA}, {0xF996, 0x7DF4}, {0xF997, 0x806F}, +{0xF998, 0x8F26}, {0xF999, 0x84EE}, {0xF99A, 0x9023}, {0xF99B, 0x934A}, {0xF99C, 0x5217}, {0xF99D, 0x52A3}, {0xF99E, 0x54BD}, {0xF99F, 0x70C8}, {0xF9A0, 0x88C2}, {0xF9A1, 0x8AAA}, {0xF9A2, 0x5EC9}, +{0xF9A3, 0x5FF5}, {0xF9A4, 0x637B}, {0xF9A5, 0x6BAE}, {0xF9A6, 0x7C3E}, {0xF9A7, 0x7375}, {0xF9A8, 0x4EE4}, {0xF9A9, 0x56F9}, {0xF9AA, 0x5BE7}, {0xF9AB, 0x5DBA}, {0xF9AC, 0x601C}, {0xF9AD, 0x73B2}, +{0xF9AE, 0x7469}, {0xF9AF, 0x7F9A}, {0xF9B0, 0x8046}, {0xF9B1, 0x9234}, {0xF9B2, 0x96F6}, {0xF9B3, 0x9748}, {0xF9B4, 0x9818}, {0xF9B5, 0x4F8B}, {0xF9B6, 0x79AE}, {0xF9B7, 0x91B4}, {0xF9B8, 0x96B8}, +{0xF9B9, 0x60E1}, {0xF9BA, 0x4E86}, {0xF9BB, 0x50DA}, {0xF9BC, 0x5BEE}, {0xF9BD, 0x5C3F}, {0xF9BE, 0x6599}, {0xF9BF, 0x6A02}, {0xF9C0, 0x71CE}, {0xF9C1, 0x7642}, {0xF9C2, 0x84FC}, {0xF9C3, 0x907C}, +{0xF9C4, 0x9F8D}, {0xF9C5, 0x6688}, {0xF9C6, 0x962E}, {0xF9C7, 0x5289}, {0xF9C8, 0x677B}, {0xF9C9, 0x67F3}, {0xF9CA, 0x6D41}, {0xF9CB, 0x6E9C}, {0xF9CC, 0x7409}, {0xF9CD, 0x7559}, {0xF9CE, 0x786B}, +{0xF9CF, 0x7D10}, {0xF9D0, 0x985E}, {0xF9D1, 0x516D}, {0xF9D2, 0x622E}, {0xF9D3, 0x9678}, {0xF9D4, 0x502B}, {0xF9D5, 0x5D19}, {0xF9D6, 0x6DEA}, {0xF9D7, 0x8F2A}, {0xF9D8, 0x5F8B}, {0xF9D9, 0x6144}, +{0xF9DA, 0x6817}, {0xF9DB, 0x7387}, {0xF9DC, 0x9686}, {0xF9DD, 0x5229}, {0xF9DE, 0x540F}, {0xF9DF, 0x5C65}, {0xF9E0, 0x6613}, {0xF9E1, 0x674E}, {0xF9E2, 0x68A8}, {0xF9E3, 0x6CE5}, {0xF9E4, 0x7406}, +{0xF9E5, 0x75E2}, {0xF9E6, 0x7F79}, {0xF9E7, 0x88CF}, {0xF9E8, 0x88E1}, {0xF9E9, 0x91CC}, {0xF9EA, 0x96E2}, {0xF9EB, 0x533F}, {0xF9EC, 0x6EBA}, {0xF9ED, 0x541D}, {0xF9EE, 0x71D0}, {0xF9EF, 0x7498}, +{0xF9F0, 0x85FA}, {0xF9F1, 0x96A3}, {0xF9F2, 0x9C57}, {0xF9F3, 0x9E9F}, {0xF9F4, 0x6797}, {0xF9F5, 0x6DCB}, {0xF9F6, 0x81E8}, {0xF9F7, 0x7ACB}, {0xF9F8, 0x7B20}, {0xF9F9, 0x7C92}, {0xF9FA, 0x72C0}, +{0xF9FB, 0x7099}, {0xF9FC, 0x8B58}, {0xF9FD, 0x4EC0}, {0xF9FE, 0x8336}, {0xF9FF, 0x523A}, {0xFA00, 0x5207}, {0xFA01, 0x5EA6}, {0xFA02, 0x62D3}, {0xFA03, 0x7CD6}, {0xFA04, 0x5B85}, {0xFA05, 0x6D1E}, +{0xFA06, 0x66B4}, {0xFA07, 0x8F3B}, {0xFA08, 0x884C}, {0xFA09, 0x964D}, {0xFA0A, 0x898B}, {0xFA0B, 0x5ED3}, {0xFA0C, 0x5140}, {0xFA0D, 0x55C0}, {0xFA10, 0x585A}, {0xFA12, 0x6674}, {0xFA15, 0x51DE}, +{0xFA16, 0x732A}, {0xFA17, 0x76CA}, {0xFA18, 0x793C}, {0xFA19, 0x795E}, {0xFA1A, 0x7965}, {0xFA1B, 0x798F}, {0xFA1C, 0x9756}, {0xFA1D, 0x7CBE}, {0xFA1E, 0x7FBD}, {0xFA20, 0x8612}, {0xFA22, 0x8AF8}, +{0xFA25, 0x9038}, {0xFA26, 0x90FD}, {0xFA2A, 0x98EF}, {0xFA2B, 0x98FC}, {0xFA2C, 0x9928}, {0xFA2D, 0x9DB4}, {0xFA2E, 0x90DE}, {0xFA2F, 0x96B7}, {0xFA30, 0x4FAE}, {0xFA31, 0x50E7}, {0xFA32, 0x514D}, +{0xFA33, 0x52C9}, {0xFA34, 0x52E4}, {0xFA35, 0x5351}, {0xFA36, 0x559D}, {0xFA37, 0x5606}, {0xFA38, 0x5668}, {0xFA39, 0x5840}, {0xFA3A, 0x58A8}, {0xFA3B, 0x5C64}, {0xFA3C, 0x5C6E}, {0xFA3D, 0x6094}, +{0xFA3E, 0x6168}, {0xFA3F, 0x618E}, {0xFA40, 0x61F2}, {0xFA41, 0x654F}, {0xFA42, 0x65E2}, {0xFA43, 0x6691}, {0xFA44, 0x6885}, {0xFA45, 0x6D77}, {0xFA46, 0x6E1A}, {0xFA47, 0x6F22}, {0xFA48, 0x716E}, +{0xFA49, 0x722B}, {0xFA4A, 0x7422}, {0xFA4B, 0x7891}, {0xFA4C, 0x793E}, {0xFA4D, 0x7949}, {0xFA4E, 0x7948}, {0xFA4F, 0x7950}, {0xFA50, 0x7956}, {0xFA51, 0x795D}, {0xFA52, 0x798D}, {0xFA53, 0x798E}, +{0xFA54, 0x7A40}, {0xFA55, 0x7A81}, {0xFA56, 0x7BC0}, {0xFA57, 0x7DF4}, {0xFA58, 0x7E09}, {0xFA59, 0x7E41}, {0xFA5A, 0x7F72}, {0xFA5B, 0x8005}, {0xFA5C, 0x81ED}, {0xFA5D, 0x8279}, {0xFA5E, 0x8279}, +{0xFA5F, 0x8457}, {0xFA60, 0x8910}, {0xFA61, 0x8996}, {0xFA62, 0x8B01}, {0xFA63, 0x8B39}, {0xFA64, 0x8CD3}, {0xFA65, 0x8D08}, {0xFA66, 0x8FB6}, {0xFA67, 0x9038}, {0xFA68, 0x96E3}, {0xFA69, 0x97FF}, +{0xFA6A, 0x983B}, {0xFA6B, 0x6075}, {0xFA6C, 0x242EE}, {0xFA6D, 0x8218}, {0xFA70, 0x4E26}, {0xFA71, 0x51B5}, {0xFA72, 0x5168}, {0xFA73, 0x4F80}, {0xFA74, 0x5145}, {0xFA75, 0x5180}, {0xFA76, 0x52C7}, +{0xFA77, 0x52FA}, {0xFA78, 0x559D}, {0xFA79, 0x5555}, {0xFA7A, 0x5599}, {0xFA7B, 0x55E2}, {0xFA7C, 0x585A}, {0xFA7D, 0x58B3}, {0xFA7E, 0x5944}, {0xFA7F, 0x5954}, {0xFA80, 0x5A62}, {0xFA81, 0x5B28}, +{0xFA82, 0x5ED2}, {0xFA83, 0x5ED9}, {0xFA84, 0x5F69}, {0xFA85, 0x5FAD}, {0xFA86, 0x60D8}, {0xFA87, 0x614E}, {0xFA88, 0x6108}, {0xFA89, 0x618E}, {0xFA8A, 0x6160}, {0xFA8B, 0x61F2}, {0xFA8C, 0x6234}, +{0xFA8D, 0x63C4}, {0xFA8E, 0x641C}, {0xFA8F, 0x6452}, {0xFA90, 0x6556}, {0xFA91, 0x6674}, {0xFA92, 0x6717}, {0xFA93, 0x671B}, {0xFA94, 0x6756}, {0xFA95, 0x6B79}, {0xFA96, 0x6BBA}, {0xFA97, 0x6D41}, +{0xFA98, 0x6EDB}, {0xFA99, 0x6ECB}, {0xFA9A, 0x6F22}, {0xFA9B, 0x701E}, {0xFA9C, 0x716E}, {0xFA9D, 0x77A7}, {0xFA9E, 0x7235}, {0xFA9F, 0x72AF}, {0xFAA0, 0x732A}, {0xFAA1, 0x7471}, {0xFAA2, 0x7506}, +{0xFAA3, 0x753B}, {0xFAA4, 0x761D}, {0xFAA5, 0x761F}, {0xFAA6, 0x76CA}, {0xFAA7, 0x76DB}, {0xFAA8, 0x76F4}, {0xFAA9, 0x774A}, {0xFAAA, 0x7740}, {0xFAAB, 0x78CC}, {0xFAAC, 0x7AB1}, {0xFAAD, 0x7BC0}, +{0xFAAE, 0x7C7B}, {0xFAAF, 0x7D5B}, {0xFAB0, 0x7DF4}, {0xFAB1, 0x7F3E}, {0xFAB2, 0x8005}, {0xFAB3, 0x8352}, {0xFAB4, 0x83EF}, {0xFAB5, 0x8779}, {0xFAB6, 0x8941}, {0xFAB7, 0x8986}, {0xFAB8, 0x8996}, +{0xFAB9, 0x8ABF}, {0xFABA, 0x8AF8}, {0xFABB, 0x8ACB}, {0xFABC, 0x8B01}, {0xFABD, 0x8AFE}, {0xFABE, 0x8AED}, {0xFABF, 0x8B39}, {0xFAC0, 0x8B8A}, {0xFAC1, 0x8D08}, {0xFAC2, 0x8F38}, {0xFAC3, 0x9072}, +{0xFAC4, 0x9199}, {0xFAC5, 0x9276}, {0xFAC6, 0x967C}, {0xFAC7, 0x96E3}, {0xFAC8, 0x9756}, {0xFAC9, 0x97DB}, {0xFACA, 0x97FF}, {0xFACB, 0x980B}, {0xFACC, 0x983B}, {0xFACD, 0x9B12}, {0xFACE, 0x9F9C}, +{0xFACF, 0x2284A}, {0xFAD0, 0x22844}, {0xFAD1, 0x233D5}, {0xFAD2, 0x3B9D}, {0xFAD3, 0x4018}, {0xFAD4, 0x4039}, {0xFAD5, 0x25249}, {0xFAD6, 0x25CD0}, {0xFAD7, 0x27ED3}, {0xFAD8, 0x9F43}, +{0xFAD9, 0x9F8E}, {0xFB1D, 0x5D9}, {0xFB1D, 0x5B4}, {0xFB1F, 0x5F2}, {0xFB1F, 0x5B7}, {0xFB2A, 0x5E9}, {0xFB2A, 0x5C1}, {0xFB2B, 0x5E9}, {0xFB2B, 0x5C2}, {0xFB2C, 0x5E9}, {0xFB2C, 0x5BC}, +{0xFB2C, 0x5C1}, {0xFB2D, 0x5E9}, {0xFB2D, 0x5BC}, {0xFB2D, 0x5C2}, {0xFB2E, 0x5D0}, {0xFB2E, 0x5B7}, {0xFB2F, 0x5D0}, {0xFB2F, 0x5B8}, {0xFB30, 0x5D0}, {0xFB30, 0x5BC}, {0xFB31, 0x5D1}, +{0xFB31, 0x5BC}, {0xFB32, 0x5D2}, {0xFB32, 0x5BC}, {0xFB33, 0x5D3}, {0xFB33, 0x5BC}, {0xFB34, 0x5D4}, {0xFB34, 0x5BC}, {0xFB35, 0x5D5}, {0xFB35, 0x5BC}, {0xFB36, 0x5D6}, {0xFB36, 0x5BC}, +{0xFB38, 0x5D8}, {0xFB38, 0x5BC}, {0xFB39, 0x5D9}, {0xFB39, 0x5BC}, {0xFB3A, 0x5DA}, {0xFB3A, 0x5BC}, {0xFB3B, 0x5DB}, {0xFB3B, 0x5BC}, {0xFB3C, 0x5DC}, {0xFB3C, 0x5BC}, {0xFB3E, 0x5DE}, +{0xFB3E, 0x5BC}, {0xFB40, 0x5E0}, {0xFB40, 0x5BC}, {0xFB41, 0x5E1}, {0xFB41, 0x5BC}, {0xFB43, 0x5E3}, {0xFB43, 0x5BC}, {0xFB44, 0x5E4}, {0xFB44, 0x5BC}, {0xFB46, 0x5E6}, {0xFB46, 0x5BC}, +{0xFB47, 0x5E7}, {0xFB47, 0x5BC}, {0xFB48, 0x5E8}, {0xFB48, 0x5BC}, {0xFB49, 0x5E9}, {0xFB49, 0x5BC}, {0xFB4A, 0x5EA}, {0xFB4A, 0x5BC}, {0xFB4B, 0x5D5}, {0xFB4B, 0x5B9}, {0xFB4C, 0x5D1}, +{0xFB4C, 0x5BF}, {0xFB4D, 0x5DB}, {0xFB4D, 0x5BF}, {0xFB4E, 0x5E4}, {0xFB4E, 0x5BF}, {0x1109A, 0x11099}, {0x1109A, 0x110BA}, {0x1109C, 0x1109B}, {0x1109C, 0x110BA}, {0x110AB, 0x110A5}, +{0x110AB, 0x110BA}, {0x1112E, 0x11131}, {0x1112E, 0x11127}, {0x1112F, 0x11132}, {0x1112F, 0x11127}, {0x1134B, 0x11347}, {0x1134B, 0x1133E}, {0x1134C, 0x11347}, {0x1134C, 0x11357}, {0x114BB, 0x114B9}, +{0x114BB, 0x114BA}, {0x114BC, 0x114B9}, {0x114BC, 0x114B0}, {0x114BE, 0x114B9}, {0x114BE, 0x114BD}, {0x115BA, 0x115B8}, {0x115BA, 0x115AF}, {0x115BB, 0x115B9}, {0x115BB, 0x115AF}, {0x1D15E, 0x1D157}, +{0x1D15E, 0x1D165}, {0x1D15F, 0x1D158}, {0x1D15F, 0x1D165}, {0x1D160, 0x1D158}, {0x1D160, 0x1D165}, {0x1D160, 0x1D16E}, {0x1D161, 0x1D158}, {0x1D161, 0x1D165}, {0x1D161, 0x1D16F}, {0x1D162, 0x1D158}, +{0x1D162, 0x1D165}, {0x1D162, 0x1D170}, {0x1D163, 0x1D158}, {0x1D163, 0x1D165}, {0x1D163, 0x1D171}, {0x1D164, 0x1D158}, {0x1D164, 0x1D165}, {0x1D164, 0x1D172}, {0x1D1BB, 0x1D1B9}, {0x1D1BB, 0x1D165}, +{0x1D1BC, 0x1D1BA}, {0x1D1BC, 0x1D165}, {0x1D1BD, 0x1D1B9}, {0x1D1BD, 0x1D165}, {0x1D1BD, 0x1D16E}, {0x1D1BE, 0x1D1BA}, {0x1D1BE, 0x1D165}, {0x1D1BE, 0x1D16E}, {0x1D1BF, 0x1D1B9}, {0x1D1BF, 0x1D165}, +{0x1D1BF, 0x1D16F}, {0x1D1C0, 0x1D1BA}, {0x1D1C0, 0x1D165}, {0x1D1C0, 0x1D16F}, {0x2F800, 0x4E3D}, {0x2F801, 0x4E38}, {0x2F802, 0x4E41}, {0x2F803, 0x20122}, {0x2F804, 0x4F60}, {0x2F805, 0x4FAE}, +{0x2F806, 0x4FBB}, {0x2F807, 0x5002}, {0x2F808, 0x507A}, {0x2F809, 0x5099}, {0x2F80A, 0x50E7}, {0x2F80B, 0x50CF}, {0x2F80C, 0x349E}, {0x2F80D, 0x2063A}, {0x2F80E, 0x514D}, {0x2F80F, 0x5154}, +{0x2F810, 0x5164}, {0x2F811, 0x5177}, {0x2F812, 0x2051C}, {0x2F813, 0x34B9}, {0x2F814, 0x5167}, {0x2F815, 0x518D}, {0x2F816, 0x2054B}, {0x2F817, 0x5197}, {0x2F818, 0x51A4}, {0x2F819, 0x4ECC}, +{0x2F81A, 0x51AC}, {0x2F81B, 0x51B5}, {0x2F81C, 0x291DF}, {0x2F81D, 0x51F5}, {0x2F81E, 0x5203}, {0x2F81F, 0x34DF}, {0x2F820, 0x523B}, {0x2F821, 0x5246}, {0x2F822, 0x5272}, {0x2F823, 0x5277}, +{0x2F824, 0x3515}, {0x2F825, 0x52C7}, {0x2F826, 0x52C9}, {0x2F827, 0x52E4}, {0x2F828, 0x52FA}, {0x2F829, 0x5305}, {0x2F82A, 0x5306}, {0x2F82B, 0x5317}, {0x2F82C, 0x5349}, {0x2F82D, 0x5351}, +{0x2F82E, 0x535A}, {0x2F82F, 0x5373}, {0x2F830, 0x537D}, {0x2F831, 0x537F}, {0x2F832, 0x537F}, {0x2F833, 0x537F}, {0x2F834, 0x20A2C}, {0x2F835, 0x7070}, {0x2F836, 0x53CA}, {0x2F837, 0x53DF}, +{0x2F838, 0x20B63}, {0x2F839, 0x53EB}, {0x2F83A, 0x53F1}, {0x2F83B, 0x5406}, {0x2F83C, 0x549E}, {0x2F83D, 0x5438}, {0x2F83E, 0x5448}, {0x2F83F, 0x5468}, {0x2F840, 0x54A2}, {0x2F841, 0x54F6}, +{0x2F842, 0x5510}, {0x2F843, 0x5553}, {0x2F844, 0x5563}, {0x2F845, 0x5584}, {0x2F846, 0x5584}, {0x2F847, 0x5599}, {0x2F848, 0x55AB}, {0x2F849, 0x55B3}, {0x2F84A, 0x55C2}, {0x2F84B, 0x5716}, +{0x2F84C, 0x5606}, {0x2F84D, 0x5717}, {0x2F84E, 0x5651}, {0x2F84F, 0x5674}, {0x2F850, 0x5207}, {0x2F851, 0x58EE}, {0x2F852, 0x57CE}, {0x2F853, 0x57F4}, {0x2F854, 0x580D}, {0x2F855, 0x578B}, +{0x2F856, 0x5832}, {0x2F857, 0x5831}, {0x2F858, 0x58AC}, {0x2F859, 0x214E4}, {0x2F85A, 0x58F2}, {0x2F85B, 0x58F7}, {0x2F85C, 0x5906}, {0x2F85D, 0x591A}, {0x2F85E, 0x5922}, {0x2F85F, 0x5962}, +{0x2F860, 0x216A8}, {0x2F861, 0x216EA}, {0x2F862, 0x59EC}, {0x2F863, 0x5A1B}, {0x2F864, 0x5A27}, {0x2F865, 0x59D8}, {0x2F866, 0x5A66}, {0x2F867, 0x36EE}, {0x2F868, 0x36FC}, {0x2F869, 0x5B08}, +{0x2F86A, 0x5B3E}, {0x2F86B, 0x5B3E}, {0x2F86C, 0x219C8}, {0x2F86D, 0x5BC3}, {0x2F86E, 0x5BD8}, {0x2F86F, 0x5BE7}, {0x2F870, 0x5BF3}, {0x2F871, 0x21B18}, {0x2F872, 0x5BFF}, {0x2F873, 0x5C06}, +{0x2F874, 0x5F53}, {0x2F875, 0x5C22}, {0x2F876, 0x3781}, {0x2F877, 0x5C60}, {0x2F878, 0x5C6E}, {0x2F879, 0x5CC0}, {0x2F87A, 0x5C8D}, {0x2F87B, 0x21DE4}, {0x2F87C, 0x5D43}, {0x2F87D, 0x21DE6}, +{0x2F87E, 0x5D6E}, {0x2F87F, 0x5D6B}, {0x2F880, 0x5D7C}, {0x2F881, 0x5DE1}, {0x2F882, 0x5DE2}, {0x2F883, 0x382F}, {0x2F884, 0x5DFD}, {0x2F885, 0x5E28}, {0x2F886, 0x5E3D}, {0x2F887, 0x5E69}, +{0x2F888, 0x3862}, {0x2F889, 0x22183}, {0x2F88A, 0x387C}, {0x2F88B, 0x5EB0}, {0x2F88C, 0x5EB3}, {0x2F88D, 0x5EB6}, {0x2F88E, 0x5ECA}, {0x2F88F, 0x2A392}, {0x2F890, 0x5EFE}, {0x2F891, 0x22331}, +{0x2F892, 0x22331}, {0x2F893, 0x8201}, {0x2F894, 0x5F22}, {0x2F895, 0x5F22}, {0x2F896, 0x38C7}, {0x2F897, 0x232B8}, {0x2F898, 0x261DA}, {0x2F899, 0x5F62}, {0x2F89A, 0x5F6B}, {0x2F89B, 0x38E3}, +{0x2F89C, 0x5F9A}, {0x2F89D, 0x5FCD}, {0x2F89E, 0x5FD7}, {0x2F89F, 0x5FF9}, {0x2F8A0, 0x6081}, {0x2F8A1, 0x393A}, {0x2F8A2, 0x391C}, {0x2F8A3, 0x6094}, {0x2F8A4, 0x226D4}, {0x2F8A5, 0x60C7}, +{0x2F8A6, 0x6148}, {0x2F8A7, 0x614C}, {0x2F8A8, 0x614E}, {0x2F8A9, 0x614C}, {0x2F8AA, 0x617A}, {0x2F8AB, 0x618E}, {0x2F8AC, 0x61B2}, {0x2F8AD, 0x61A4}, {0x2F8AE, 0x61AF}, {0x2F8AF, 0x61DE}, +{0x2F8B0, 0x61F2}, {0x2F8B1, 0x61F6}, {0x2F8B2, 0x6210}, {0x2F8B3, 0x621B}, {0x2F8B4, 0x625D}, {0x2F8B5, 0x62B1}, {0x2F8B6, 0x62D4}, {0x2F8B7, 0x6350}, {0x2F8B8, 0x22B0C}, {0x2F8B9, 0x633D}, +{0x2F8BA, 0x62FC}, {0x2F8BB, 0x6368}, {0x2F8BC, 0x6383}, {0x2F8BD, 0x63E4}, {0x2F8BE, 0x22BF1}, {0x2F8BF, 0x6422}, {0x2F8C0, 0x63C5}, {0x2F8C1, 0x63A9}, {0x2F8C2, 0x3A2E}, {0x2F8C3, 0x6469}, +{0x2F8C4, 0x647E}, {0x2F8C5, 0x649D}, {0x2F8C6, 0x6477}, {0x2F8C7, 0x3A6C}, {0x2F8C8, 0x654F}, {0x2F8C9, 0x656C}, {0x2F8CA, 0x2300A}, {0x2F8CB, 0x65E3}, {0x2F8CC, 0x66F8}, {0x2F8CD, 0x6649}, +{0x2F8CE, 0x3B19}, {0x2F8CF, 0x6691}, {0x2F8D0, 0x3B08}, {0x2F8D1, 0x3AE4}, {0x2F8D2, 0x5192}, {0x2F8D3, 0x5195}, {0x2F8D4, 0x6700}, {0x2F8D5, 0x669C}, {0x2F8D6, 0x80AD}, {0x2F8D7, 0x43D9}, +{0x2F8D8, 0x6717}, {0x2F8D9, 0x671B}, {0x2F8DA, 0x6721}, {0x2F8DB, 0x675E}, {0x2F8DC, 0x6753}, {0x2F8DD, 0x233C3}, {0x2F8DE, 0x3B49}, {0x2F8DF, 0x67FA}, {0x2F8E0, 0x6785}, {0x2F8E1, 0x6852}, +{0x2F8E2, 0x6885}, {0x2F8E3, 0x2346D}, {0x2F8E4, 0x688E}, {0x2F8E5, 0x681F}, {0x2F8E6, 0x6914}, {0x2F8E7, 0x3B9D}, {0x2F8E8, 0x6942}, {0x2F8E9, 0x69A3}, {0x2F8EA, 0x69EA}, {0x2F8EB, 0x6AA8}, +{0x2F8EC, 0x236A3}, {0x2F8ED, 0x6ADB}, {0x2F8EE, 0x3C18}, {0x2F8EF, 0x6B21}, {0x2F8F0, 0x238A7}, {0x2F8F1, 0x6B54}, {0x2F8F2, 0x3C4E}, {0x2F8F3, 0x6B72}, {0x2F8F4, 0x6B9F}, {0x2F8F5, 0x6BBA}, +{0x2F8F6, 0x6BBB}, {0x2F8F7, 0x23A8D}, {0x2F8F8, 0x21D0B}, {0x2F8F9, 0x23AFA}, {0x2F8FA, 0x6C4E}, {0x2F8FB, 0x23CBC}, {0x2F8FC, 0x6CBF}, {0x2F8FD, 0x6CCD}, {0x2F8FE, 0x6C67}, {0x2F8FF, 0x6D16}, +{0x2F900, 0x6D3E}, {0x2F901, 0x6D77}, {0x2F902, 0x6D41}, {0x2F903, 0x6D69}, {0x2F904, 0x6D78}, {0x2F905, 0x6D85}, {0x2F906, 0x23D1E}, {0x2F907, 0x6D34}, {0x2F908, 0x6E2F}, {0x2F909, 0x6E6E}, +{0x2F90A, 0x3D33}, {0x2F90B, 0x6ECB}, {0x2F90C, 0x6EC7}, {0x2F90D, 0x23ED1}, {0x2F90E, 0x6DF9}, {0x2F90F, 0x6F6E}, {0x2F910, 0x23F5E}, {0x2F911, 0x23F8E}, {0x2F912, 0x6FC6}, {0x2F913, 0x7039}, +{0x2F914, 0x701E}, {0x2F915, 0x701B}, {0x2F916, 0x3D96}, {0x2F917, 0x704A}, {0x2F918, 0x707D}, {0x2F919, 0x7077}, {0x2F91A, 0x70AD}, {0x2F91B, 0x20525}, {0x2F91C, 0x7145}, {0x2F91D, 0x24263}, +{0x2F91E, 0x719C}, {0x2F91F, 0x243AB}, {0x2F920, 0x7228}, {0x2F921, 0x7235}, {0x2F922, 0x7250}, {0x2F923, 0x24608}, {0x2F924, 0x7280}, {0x2F925, 0x7295}, {0x2F926, 0x24735}, {0x2F927, 0x24814}, +{0x2F928, 0x737A}, {0x2F929, 0x738B}, {0x2F92A, 0x3EAC}, {0x2F92B, 0x73A5}, {0x2F92C, 0x3EB8}, {0x2F92D, 0x3EB8}, {0x2F92E, 0x7447}, {0x2F92F, 0x745C}, {0x2F930, 0x7471}, {0x2F931, 0x7485}, +{0x2F932, 0x74CA}, {0x2F933, 0x3F1B}, {0x2F934, 0x7524}, {0x2F935, 0x24C36}, {0x2F936, 0x753E}, {0x2F937, 0x24C92}, {0x2F938, 0x7570}, {0x2F939, 0x2219F}, {0x2F93A, 0x7610}, {0x2F93B, 0x24FA1}, +{0x2F93C, 0x24FB8}, {0x2F93D, 0x25044}, {0x2F93E, 0x3FFC}, {0x2F93F, 0x4008}, {0x2F940, 0x76F4}, {0x2F941, 0x250F3}, {0x2F942, 0x250F2}, {0x2F943, 0x25119}, {0x2F944, 0x25133}, {0x2F945, 0x771E}, +{0x2F946, 0x771F}, {0x2F947, 0x771F}, {0x2F948, 0x774A}, {0x2F949, 0x4039}, {0x2F94A, 0x778B}, {0x2F94B, 0x4046}, {0x2F94C, 0x4096}, {0x2F94D, 0x2541D}, {0x2F94E, 0x784E}, {0x2F94F, 0x788C}, +{0x2F950, 0x78CC}, {0x2F951, 0x40E3}, {0x2F952, 0x25626}, {0x2F953, 0x7956}, {0x2F954, 0x2569A}, {0x2F955, 0x256C5}, {0x2F956, 0x798F}, {0x2F957, 0x79EB}, {0x2F958, 0x412F}, {0x2F959, 0x7A40}, +{0x2F95A, 0x7A4A}, {0x2F95B, 0x7A4F}, {0x2F95C, 0x2597C}, {0x2F95D, 0x25AA7}, {0x2F95E, 0x25AA7}, {0x2F95F, 0x7AEE}, {0x2F960, 0x4202}, {0x2F961, 0x25BAB}, {0x2F962, 0x7BC6}, {0x2F963, 0x7BC9}, +{0x2F964, 0x4227}, {0x2F965, 0x25C80}, {0x2F966, 0x7CD2}, {0x2F967, 0x42A0}, {0x2F968, 0x7CE8}, {0x2F969, 0x7CE3}, {0x2F96A, 0x7D00}, {0x2F96B, 0x25F86}, {0x2F96C, 0x7D63}, {0x2F96D, 0x4301}, +{0x2F96E, 0x7DC7}, {0x2F96F, 0x7E02}, {0x2F970, 0x7E45}, {0x2F971, 0x4334}, {0x2F972, 0x26228}, {0x2F973, 0x26247}, {0x2F974, 0x4359}, {0x2F975, 0x262D9}, {0x2F976, 0x7F7A}, {0x2F977, 0x2633E}, +{0x2F978, 0x7F95}, {0x2F979, 0x7FFA}, {0x2F97A, 0x8005}, {0x2F97B, 0x264DA}, {0x2F97C, 0x26523}, {0x2F97D, 0x8060}, {0x2F97E, 0x265A8}, {0x2F97F, 0x8070}, {0x2F980, 0x2335F}, {0x2F981, 0x43D5}, +{0x2F982, 0x80B2}, {0x2F983, 0x8103}, {0x2F984, 0x440B}, {0x2F985, 0x813E}, {0x2F986, 0x5AB5}, {0x2F987, 0x267A7}, {0x2F988, 0x267B5}, {0x2F989, 0x23393}, {0x2F98A, 0x2339C}, {0x2F98B, 0x8201}, +{0x2F98C, 0x8204}, {0x2F98D, 0x8F9E}, {0x2F98E, 0x446B}, {0x2F98F, 0x8291}, {0x2F990, 0x828B}, {0x2F991, 0x829D}, {0x2F992, 0x52B3}, {0x2F993, 0x82B1}, {0x2F994, 0x82B3}, {0x2F995, 0x82BD}, +{0x2F996, 0x82E6}, {0x2F997, 0x26B3C}, {0x2F998, 0x82E5}, {0x2F999, 0x831D}, {0x2F99A, 0x8363}, {0x2F99B, 0x83AD}, {0x2F99C, 0x8323}, {0x2F99D, 0x83BD}, {0x2F99E, 0x83E7}, {0x2F99F, 0x8457}, +{0x2F9A0, 0x8353}, {0x2F9A1, 0x83CA}, {0x2F9A2, 0x83CC}, {0x2F9A3, 0x83DC}, {0x2F9A4, 0x26C36}, {0x2F9A5, 0x26D6B}, {0x2F9A6, 0x26CD5}, {0x2F9A7, 0x452B}, {0x2F9A8, 0x84F1}, {0x2F9A9, 0x84F3}, +{0x2F9AA, 0x8516}, {0x2F9AB, 0x273CA}, {0x2F9AC, 0x8564}, {0x2F9AD, 0x26F2C}, {0x2F9AE, 0x455D}, {0x2F9AF, 0x4561}, {0x2F9B0, 0x26FB1}, {0x2F9B1, 0x270D2}, {0x2F9B2, 0x456B}, {0x2F9B3, 0x8650}, +{0x2F9B4, 0x865C}, {0x2F9B5, 0x8667}, {0x2F9B6, 0x8669}, {0x2F9B7, 0x86A9}, {0x2F9B8, 0x8688}, {0x2F9B9, 0x870E}, {0x2F9BA, 0x86E2}, {0x2F9BB, 0x8779}, {0x2F9BC, 0x8728}, {0x2F9BD, 0x876B}, +{0x2F9BE, 0x8786}, {0x2F9BF, 0x45D7}, {0x2F9C0, 0x87E1}, {0x2F9C1, 0x8801}, {0x2F9C2, 0x45F9}, {0x2F9C3, 0x8860}, {0x2F9C4, 0x8863}, {0x2F9C5, 0x27667}, {0x2F9C6, 0x88D7}, {0x2F9C7, 0x88DE}, +{0x2F9C8, 0x4635}, {0x2F9C9, 0x88FA}, {0x2F9CA, 0x34BB}, {0x2F9CB, 0x278AE}, {0x2F9CC, 0x27966}, {0x2F9CD, 0x46BE}, {0x2F9CE, 0x46C7}, {0x2F9CF, 0x8AA0}, {0x2F9D0, 0x8AED}, {0x2F9D1, 0x8B8A}, +{0x2F9D2, 0x8C55}, {0x2F9D3, 0x27CA8}, {0x2F9D4, 0x8CAB}, {0x2F9D5, 0x8CC1}, {0x2F9D6, 0x8D1B}, {0x2F9D7, 0x8D77}, {0x2F9D8, 0x27F2F}, {0x2F9D9, 0x20804}, {0x2F9DA, 0x8DCB}, {0x2F9DB, 0x8DBC}, +{0x2F9DC, 0x8DF0}, {0x2F9DD, 0x208DE}, {0x2F9DE, 0x8ED4}, {0x2F9DF, 0x8F38}, {0x2F9E0, 0x285D2}, {0x2F9E1, 0x285ED}, {0x2F9E2, 0x9094}, {0x2F9E3, 0x90F1}, {0x2F9E4, 0x9111}, {0x2F9E5, 0x2872E}, +{0x2F9E6, 0x911B}, {0x2F9E7, 0x9238}, {0x2F9E8, 0x92D7}, {0x2F9E9, 0x92D8}, {0x2F9EA, 0x927C}, {0x2F9EB, 0x93F9}, {0x2F9EC, 0x9415}, {0x2F9ED, 0x28BFA}, {0x2F9EE, 0x958B}, {0x2F9EF, 0x4995}, +{0x2F9F0, 0x95B7}, {0x2F9F1, 0x28D77}, {0x2F9F2, 0x49E6}, {0x2F9F3, 0x96C3}, {0x2F9F4, 0x5DB2}, {0x2F9F5, 0x9723}, {0x2F9F6, 0x29145}, {0x2F9F7, 0x2921A}, {0x2F9F8, 0x4A6E}, {0x2F9F9, 0x4A76}, +{0x2F9FA, 0x97E0}, {0x2F9FB, 0x2940A}, {0x2F9FC, 0x4AB2}, {0x2F9FD, 0x29496}, {0x2F9FE, 0x980B}, {0x2F9FF, 0x980B}, {0x2FA00, 0x9829}, {0x2FA01, 0x295B6}, {0x2FA02, 0x98E2}, {0x2FA03, 0x4B33}, +{0x2FA04, 0x9929}, {0x2FA05, 0x99A7}, {0x2FA06, 0x99C2}, {0x2FA07, 0x99FE}, {0x2FA08, 0x4BCE}, {0x2FA09, 0x29B30}, {0x2FA0A, 0x9B12}, {0x2FA0B, 0x9C40}, {0x2FA0C, 0x9CFD}, {0x2FA0D, 0x4CCE}, +{0x2FA0E, 0x4CED}, {0x2FA0F, 0x9D67}, {0x2FA10, 0x2A0CE}, {0x2FA11, 0x4CF8}, {0x2FA12, 0x2A105}, {0x2FA13, 0x2A20E}, {0x2FA14, 0x2A291}, {0x2FA15, 0x9EBB}, {0x2FA16, 0x4D56}, {0x2FA17, 0x9EF9}, +{0x2FA18, 0x9EFE}, {0x2FA19, 0x9F05}, {0x2FA1A, 0x9F0F}, {0x2FA1B, 0x9F16}, {0x2FA1D, 0x2A600}, }; static std::string codepoint_to_utf8(uint32_t cp) { diff --git a/examples/whisper.android.java/app/src/main/jni/whisper/jni.c b/examples/whisper.android.java/app/src/main/jni/whisper/jni.c index f8e7eff..5f92810 100644 --- a/examples/whisper.android.java/app/src/main/jni/whisper/jni.c +++ b/examples/whisper.android.java/app/src/main/jni/whisper/jni.c @@ -245,6 +245,8 @@ Java_com_whispercpp_java_whisper_WhisperLib_benchMemcpy(JNIEnv *env, jobject thi UNUSED(thiz); const char *bench_ggml_memcpy = whisper_bench_memcpy_str(n_threads); jstring string = (*env)->NewStringUTF(env, bench_ggml_memcpy); + + return string; } JNIEXPORT jstring JNICALL @@ -253,5 +255,7 @@ Java_com_whispercpp_java_whisper_WhisperLib_benchGgmlMulMat(JNIEnv *env, jobject UNUSED(thiz); const char *bench_ggml_mul_mat = whisper_bench_ggml_mul_mat_str(n_threads); jstring string = (*env)->NewStringUTF(env, bench_ggml_mul_mat); + + return string; } diff --git a/extra/sync-ggml.last b/extra/sync-ggml.last index 5b775e8..1513555 100644 --- a/extra/sync-ggml.last +++ b/extra/sync-ggml.last @@ -1 +1 @@ -9fb1f17a51e3935d2ac3ab7969b6c2eb4ee6e5a3 +ac24ec57f1f18dc4c2eb3d3d3c22a75bc1d9b742 diff --git a/ggml-backend-impl.h b/ggml-backend-impl.h index 0e5bf0a..2e9ba58 100644 --- a/ggml-backend-impl.h +++ b/ggml-backend-impl.h @@ -91,13 +91,14 @@ extern "C" { // (optional) complete all pending operations void (*GGML_CALL synchronize)(ggml_backend_t backend); - // compute graph with a plan + // create a plan for ggml_cgraph and free it ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph); void (*GGML_CALL graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan); - void (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan); + // compute graph with a plan + enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan); // compute graph without a plan (async) - bool (*GGML_CALL graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph); + enum ggml_status (*GGML_CALL graph_compute) (ggml_backend_t backend, struct ggml_cgraph * cgraph); // check if the backend supports an operation bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op); diff --git a/ggml-backend.c b/ggml-backend.c index c86673b..d60d984 100644 --- a/ggml-backend.c +++ b/ggml-backend.c @@ -262,11 +262,11 @@ void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_pla backend->iface.graph_plan_free(backend, plan); } -void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { - backend->iface.graph_plan_compute(backend, plan); +enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { + return backend->iface.graph_plan_compute(backend, plan); } -bool ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { +enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { return backend->iface.graph_compute(backend, cgraph); } @@ -732,15 +732,15 @@ GGML_CALL static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, g GGML_UNUSED(backend); } -GGML_CALL static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { +GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan; - ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan); + return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan); GGML_UNUSED(backend); } -GGML_CALL static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { +GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context; struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads); @@ -755,8 +755,7 @@ GGML_CALL static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, str cplan.abort_callback = cpu_ctx->abort_callback; cplan.abort_callback_data = cpu_ctx->abort_callback_data; - ggml_graph_compute(cgraph, &cplan); - return true; + return ggml_graph_compute(cgraph, &cplan); } GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { @@ -1437,7 +1436,7 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) { return true; } -static bool ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) { +static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) { uint64_t copy_us[GGML_MAX_BACKENDS] = {0}; uint64_t compute_us[GGML_MAX_BACKENDS] = {0}; @@ -1472,8 +1471,9 @@ static bool ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) { uint64_t compute_start_us = ggml_time_us(); if (!sched->callback_eval) { - if (!ggml_backend_graph_compute(split_backend, &split->graph)) { - return false; + enum ggml_status ec = ggml_backend_graph_compute(split_backend, &split->graph); + if (ec != GGML_STATUS_SUCCESS) { + return ec; } //ggml_backend_synchronize(split_backend); // necessary to measure compute time } else { @@ -1494,8 +1494,9 @@ static bool ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) { struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1); - if (!ggml_backend_graph_compute(split_backend, &gv)) { - return false; + enum ggml_status ec = ggml_backend_graph_compute(split_backend, &gv); + if (ec != GGML_STATUS_SUCCESS) { + return ec; } if (need && !sched->callback_eval(t, false, sched->callback_eval_user_data)) { @@ -1519,7 +1520,7 @@ static bool ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) { } #endif - return true; + return GGML_STATUS_SUCCESS; } ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size) { @@ -1581,7 +1582,7 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * return true; } -bool ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { +enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS); if (!sched->is_reset) { @@ -1590,14 +1591,10 @@ bool ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cg ggml_backend_sched_split_graph(sched, graph); if (!ggml_backend_sched_alloc_splits(sched)) { - return false; + return GGML_STATUS_ALLOC_FAILED; } - if (!ggml_backend_sched_compute_splits(sched)) { - return false; - } - - return true; + return ggml_backend_sched_compute_splits(sched); } void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) { diff --git a/ggml-backend.h b/ggml-backend.h index 8fb54bd..8bed225 100644 --- a/ggml-backend.h +++ b/ggml-backend.h @@ -66,12 +66,13 @@ extern "C" { GGML_API void ggml_backend_synchronize(ggml_backend_t backend); - GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph); + GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph); + GGML_API void ggml_backend_graph_plan_free (ggml_backend_t backend, ggml_backend_graph_plan_t plan); - GGML_API void ggml_backend_graph_plan_free (ggml_backend_t backend, ggml_backend_graph_plan_t plan); - GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan); - GGML_API bool ggml_backend_graph_compute (ggml_backend_t backend, struct ggml_cgraph * cgraph); - GGML_API bool ggml_backend_supports_op (ggml_backend_t backend, const struct ggml_tensor * op); + GGML_API enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan); + GGML_API enum ggml_status ggml_backend_graph_compute (ggml_backend_t backend, struct ggml_cgraph * cgraph); + + GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op); // tensor copy between different backends GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst); @@ -157,26 +158,26 @@ extern "C" { typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data); // Initialize a backend scheduler - GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size); - GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched); + GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size); + GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched); // Initialize backend buffers from a measure graph - GGML_API bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph); + GGML_API bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph); // Get the number of splits of the last graph - GGML_API int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched); + GGML_API int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched); - GGML_API size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend); + GGML_API size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend); - GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend); - GGML_API ggml_backend_t ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node); + GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend); + GGML_API ggml_backend_t ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node); // Allocate and compute graph on the backend scheduler - GGML_API bool ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph); + GGML_API enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph); // Reset all assignments and allocators - must be called before changing the node backends - GGML_API void ggml_backend_sched_reset(ggml_backend_sched_t sched); + GGML_API void ggml_backend_sched_reset(ggml_backend_sched_t sched); // Set a callback to be called for each resulting node during graph compute - GGML_API void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data); + GGML_API void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data); // // Utils diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 0c6501e..72bcec8 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -616,6 +616,8 @@ static_assert(sizeof(block_iq4_xs) == sizeof(ggml_fp16_t) + sizeof(uint16_t) + Q #define CUDA_UPSCALE_BLOCK_SIZE 256 #define CUDA_CONCAT_BLOCK_SIZE 256 #define CUDA_PAD_BLOCK_SIZE 256 +#define CUDA_ARANGE_BLOCK_SIZE 256 +#define CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE 256 #define CUDA_ACC_BLOCK_SIZE 256 #define CUDA_IM2COL_BLOCK_SIZE 256 #define CUDA_POOL2D_BLOCK_SIZE 256 @@ -990,17 +992,21 @@ static __global__ void concat_f32(const float * x,const float * y, float * dst, nidx + blockIdx.y * ne0 + blockIdx.z * ne0 * gridDim.y; - dst[offset_dst] = x[offset_src]; + dst[offset_dst] = x[offset_src]; } else { int offset_src = nidx + blockIdx.y * ne0 + (blockIdx.z - ne02) * ne0 * gridDim.y; - dst[offset_dst] = y[offset_src]; + dst[offset_dst] = y[offset_src]; } } -static __global__ void upscale_f32(const float * x, float * dst, const int ne00, const int nb02, const int scale_factor) { +static __global__ void upscale_f32(const float * x, float * dst, const int ne00, const int ne00xne01, const int scale_factor) { + // blockIdx.z: idx of ne02*ne03 + // blockIdx.y: idx of ne01*scale_factor， aka ne1 + // blockIDx.x: idx of ne00*scale_factor / BLOCK_SIZE + // ne00xne01: ne00 * ne01 int ne0 = ne00 * scale_factor; int nidx = threadIdx.x + blockIdx.x * blockDim.x; if (nidx >= ne0) { @@ -1012,7 +1018,7 @@ static __global__ void upscale_f32(const float * x, float * dst, const int ne00, int offset_src = i00 + i01 * ne00 + - blockIdx.z * nb02; + blockIdx.z * ne00xne01; int offset_dst = nidx + blockIdx.y * ne0 + @@ -1020,7 +1026,10 @@ static __global__ void upscale_f32(const float * x, float * dst, const int ne00, dst[offset_dst] = x[offset_src]; } -static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02) { +static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) { + // blockIdx.z: idx of ne2*ne3, aka ne02*ne03 + // blockIdx.y: idx of ne1 + // blockIDx.x: idx of ne0 / BLOCK_SIZE int nidx = threadIdx.x + blockIdx.x * blockDim.x; if (nidx >= ne0) { return; @@ -1031,19 +1040,53 @@ static __global__ void pad_f32(const float * x, float * dst, const int ne0, cons nidx + blockIdx.y * ne0 + blockIdx.z * ne0 * gridDim.y; - if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02) { + if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) { int offset_src = nidx + blockIdx.y * ne00 + blockIdx.z * ne00 * ne01; - dst[offset_dst] = x[offset_src]; + dst[offset_dst] = x[offset_src]; } else { dst[offset_dst] = 0.0f; } } +static __global__ void arange_f32(float * dst, const int ne0, const float start, const float step) { + // blockIDx.x: idx of ne0 / BLOCK_SIZE + int nidx = threadIdx.x + blockIdx.x * blockDim.x; + if (nidx >= ne0) { + return; + } + dst[nidx] = start + step * nidx; +} + +static __global__ void timestep_embedding_f32(const float * timesteps, float * dst, const int nb1, const int dim, const int max_period) { + // blockIDx.y: idx of timesteps->ne[0] + // blockIDx.x: idx of ((dim + 1) / 2) / BLOCK_SIZE + int i = blockIdx.y; + int j = threadIdx.x + blockIdx.x * blockDim.x; + float * embed_data = (float *)((char *)dst + i*nb1); + + if (dim % 2 != 0 && j == ((dim + 1) / 2)) { + embed_data[dim] = 0.f; + } + + int half = dim / 2; + if (j >= half) { + return; + } + + float timestep = timesteps[i]; + float freq = (float)expf(-logf(max_period) * j / half); + float arg = timestep * freq; + embed_data[j] = cosf(arg); + embed_data[j + half] = sinf(arg); +} + template static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) { + // blockIdx.x: num_groups idx + // threadIdx.x: block_size idx int start = blockIdx.x * group_size; int end = start + group_size; @@ -2018,74 +2061,73 @@ static const __device__ uint32_t iq3xxs_grid[256] = { 0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04, }; -static const __device__ uint32_t iq3xs_grid[512] = { - 0x04040404, 0x0404040c, 0x04040414, 0x0404042c, 0x0404043e, 0x04040c04, 0x04040c0c, 0x04040c14, - 0x04040c24, 0x04040c34, 0x04041404, 0x0404140c, 0x0404142c, 0x04041c1c, 0x04042404, 0x04042414, - 0x0404242c, 0x0404243e, 0x04042c0c, 0x04042c1c, 0x04043404, 0x04043414, 0x04043e0c, 0x04043e24, - 0x04043e3e, 0x040c0404, 0x040c040c, 0x040c0414, 0x040c0424, 0x040c0c04, 0x040c0c0c, 0x040c0c2c, - 0x040c1404, 0x040c141c, 0x040c143e, 0x040c1c0c, 0x040c1c2c, 0x040c2424, 0x040c340c, 0x040c342c, - 0x040c3e14, 0x04140404, 0x0414040c, 0x0414042c, 0x0414043e, 0x04140c04, 0x04140c1c, 0x04140c34, - 0x0414140c, 0x0414142c, 0x04141c04, 0x04141c24, 0x04142414, 0x0414242c, 0x0414243e, 0x04142c0c, - 0x04142c1c, 0x04143e04, 0x04143e1c, 0x041c041c, 0x041c0c0c, 0x041c0c2c, 0x041c1404, 0x041c1414, - 0x041c1c0c, 0x041c1c1c, 0x041c1c34, 0x041c2424, 0x041c2c04, 0x041c2c14, 0x041c343e, 0x041c3e0c, - 0x041c3e2c, 0x04240404, 0x04240c1c, 0x04240c3e, 0x0424140c, 0x04241424, 0x04241c14, 0x04242404, - 0x0424241c, 0x04242c0c, 0x04243e04, 0x042c0414, 0x042c0424, 0x042c1404, 0x042c1414, 0x042c1434, - 0x042c1c1c, 0x042c240c, 0x042c242c, 0x042c243e, 0x042c3434, 0x042c3e1c, 0x04340434, 0x04340c0c, - 0x04340c1c, 0x04341c0c, 0x04342c14, 0x04343e0c, 0x043e0404, 0x043e0414, 0x043e0424, 0x043e1404, - 0x043e1414, 0x043e1434, 0x043e1c1c, 0x043e2c04, 0x043e2c24, 0x0c040404, 0x0c04040c, 0x0c040414, - 0x0c040424, 0x0c040c04, 0x0c040c0c, 0x0c040c1c, 0x0c040c2c, 0x0c040c3e, 0x0c041404, 0x0c041414, - 0x0c041c0c, 0x0c041c24, 0x0c041c34, 0x0c042c24, 0x0c042c34, 0x0c04340c, 0x0c043e14, 0x0c0c0404, - 0x0c0c040c, 0x0c0c041c, 0x0c0c0434, 0x0c0c0c04, 0x0c0c0c24, 0x0c0c140c, 0x0c0c1c04, 0x0c0c1c1c, - 0x0c0c240c, 0x0c0c2c04, 0x0c0c2c14, 0x0c0c3e04, 0x0c0c3e34, 0x0c140404, 0x0c140c14, 0x0c140c2c, - 0x0c140c3e, 0x0c141404, 0x0c141424, 0x0c141c14, 0x0c142404, 0x0c14241c, 0x0c142c2c, 0x0c143404, - 0x0c143e14, 0x0c1c040c, 0x0c1c0424, 0x0c1c043e, 0x0c1c0c04, 0x0c1c0c1c, 0x0c1c140c, 0x0c1c143e, - 0x0c1c1c04, 0x0c1c1c24, 0x0c1c240c, 0x0c1c3414, 0x0c1c3e04, 0x0c24041c, 0x0c24042c, 0x0c240c14, - 0x0c240c24, 0x0c241c0c, 0x0c241c1c, 0x0c242414, 0x0c242434, 0x0c242c04, 0x0c242c24, 0x0c2c040c, - 0x0c2c0c04, 0x0c2c0c1c, 0x0c2c140c, 0x0c2c1c04, 0x0c2c1c14, 0x0c2c2c0c, 0x0c341404, 0x0c341424, - 0x0c34143e, 0x0c342424, 0x0c342434, 0x0c3e040c, 0x0c3e041c, 0x0c3e0c04, 0x0c3e0c14, 0x0c3e140c, - 0x0c3e1c2c, 0x0c3e240c, 0x0c3e3414, 0x0c3e3e04, 0x14040404, 0x1404040c, 0x1404041c, 0x1404042c, - 0x1404043e, 0x14040c04, 0x14040c14, 0x14040c24, 0x14040c34, 0x1404140c, 0x1404141c, 0x1404143e, - 0x14041c04, 0x14041c14, 0x1404240c, 0x1404241c, 0x1404242c, 0x14042c04, 0x14042c14, 0x1404343e, - 0x14043e04, 0x14043e1c, 0x14043e2c, 0x140c0404, 0x140c0414, 0x140c0c04, 0x140c0c1c, 0x140c0c3e, - 0x140c1414, 0x140c142c, 0x140c1c0c, 0x140c1c24, 0x140c2414, 0x140c2c0c, 0x1414040c, 0x14140424, - 0x1414043e, 0x1414140c, 0x1414141c, 0x14141c04, 0x14141c3e, 0x1414240c, 0x14142c1c, 0x14142c3e, - 0x14143e0c, 0x14143e24, 0x141c0404, 0x141c0414, 0x141c042c, 0x141c0c0c, 0x141c1414, 0x141c1424, - 0x141c1c0c, 0x141c1c1c, 0x141c2414, 0x141c2c04, 0x141c3434, 0x1424040c, 0x1424043e, 0x14241404, - 0x1424141c, 0x14241c14, 0x14241c2c, 0x1424240c, 0x14243e14, 0x14243e2c, 0x142c0424, 0x142c0c0c, - 0x142c1414, 0x142c1c3e, 0x142c2404, 0x142c2c1c, 0x142c3e04, 0x14340404, 0x14340414, 0x1434043e, - 0x1434140c, 0x14342c2c, 0x1434340c, 0x143e042c, 0x143e0c0c, 0x143e1434, 0x143e1c04, 0x143e241c, - 0x143e2c04, 0x1c040414, 0x1c040c0c, 0x1c040c1c, 0x1c040c2c, 0x1c040c3e, 0x1c041414, 0x1c041c0c, - 0x1c041c1c, 0x1c041c2c, 0x1c042414, 0x1c042424, 0x1c04243e, 0x1c042c0c, 0x1c04341c, 0x1c043e0c, - 0x1c0c040c, 0x1c0c041c, 0x1c0c042c, 0x1c0c0c24, 0x1c0c140c, 0x1c0c141c, 0x1c0c2404, 0x1c0c3404, - 0x1c0c3e14, 0x1c0c3e34, 0x1c140404, 0x1c140c14, 0x1c141404, 0x1c141c14, 0x1c141c24, 0x1c142c04, - 0x1c1c040c, 0x1c1c0c04, 0x1c1c0c24, 0x1c1c140c, 0x1c1c141c, 0x1c1c143e, 0x1c1c1c04, 0x1c1c240c, - 0x1c1c241c, 0x1c1c243e, 0x1c1c2c2c, 0x1c1c3e1c, 0x1c24041c, 0x1c240c0c, 0x1c240c34, 0x1c241414, - 0x1c241c0c, 0x1c242c14, 0x1c243404, 0x1c243424, 0x1c2c040c, 0x1c2c0c04, 0x1c2c0c14, 0x1c2c142c, - 0x1c2c1c14, 0x1c2c2424, 0x1c2c2c34, 0x1c2c3e1c, 0x1c340c34, 0x1c34240c, 0x1c3e040c, 0x1c3e041c, - 0x1c3e1404, 0x1c3e1414, 0x1c3e1c2c, 0x24040404, 0x24040424, 0x24040c14, 0x24041404, 0x24041424, - 0x2404143e, 0x24041c14, 0x2404240c, 0x24042c04, 0x24043e04, 0x240c0414, 0x240c043e, 0x240c0c0c, - 0x240c0c1c, 0x240c1414, 0x240c1c04, 0x240c1c2c, 0x240c241c, 0x240c2c0c, 0x240c2c2c, 0x2414040c, - 0x2414041c, 0x24140c04, 0x24140c2c, 0x2414140c, 0x24141c1c, 0x24142404, 0x24142c3e, 0x24143414, - 0x24143e04, 0x241c0424, 0x241c0c0c, 0x241c0c1c, 0x241c1404, 0x241c1414, 0x241c1c0c, 0x241c1c2c, - 0x24240404, 0x24240414, 0x24241424, 0x24241c3e, 0x24242404, 0x24243e0c, 0x242c042c, 0x242c043e, - 0x242c140c, 0x242c3414, 0x24340c1c, 0x24341c24, 0x24343404, 0x243e0c04, 0x243e0c2c, 0x243e1c04, - 0x243e241c, 0x243e2c0c, 0x2c040414, 0x2c040c04, 0x2c040c24, 0x2c041414, 0x2c042404, 0x2c042424, - 0x2c04243e, 0x2c042c14, 0x2c043434, 0x2c043e24, 0x2c0c040c, 0x2c0c041c, 0x2c0c042c, 0x2c0c0c14, - 0x2c0c140c, 0x2c0c1c14, 0x2c0c3e14, 0x2c140404, 0x2c140c0c, 0x2c14141c, 0x2c141c04, 0x2c141c34, - 0x2c142c1c, 0x2c1c0414, 0x2c1c043e, 0x2c1c0c04, 0x2c1c143e, 0x2c1c2424, 0x2c1c2c0c, 0x2c1c342c, - 0x2c1c3e1c, 0x2c24040c, 0x2c240424, 0x2c241404, 0x2c241c14, 0x2c242434, 0x2c2c0c14, 0x2c2c1434, - 0x2c2c2c0c, 0x2c2c2c1c, 0x2c342414, 0x2c3e0414, 0x2c3e0424, 0x2c3e1414, 0x34040c0c, 0x34040c1c, - 0x34040c2c, 0x34041c0c, 0x34041c1c, 0x34043404, 0x340c0404, 0x340c1404, 0x340c143e, 0x340c3424, - 0x34140c14, 0x34141c24, 0x34142414, 0x34142c2c, 0x34143414, 0x34143e04, 0x341c0404, 0x341c0c24, - 0x341c140c, 0x341c2404, 0x3424142c, 0x3424241c, 0x34243414, 0x342c0404, 0x342c041c, 0x342c1c24, - 0x342c3404, 0x3434042c, 0x34342404, 0x343e0c0c, 0x343e0c1c, 0x3e040404, 0x3e040424, 0x3e04043e, - 0x3e041404, 0x3e041414, 0x3e041c34, 0x3e042404, 0x3e042c24, 0x3e043414, 0x3e0c0414, 0x3e0c0c0c, - 0x3e0c1424, 0x3e0c241c, 0x3e0c242c, 0x3e14040c, 0x3e140424, 0x3e140c04, 0x3e140c34, 0x3e14140c, - 0x3e141c04, 0x3e142c0c, 0x3e1c0414, 0x3e1c1c14, 0x3e1c1c2c, 0x3e1c2c1c, 0x3e24040c, 0x3e24042c, - 0x3e240c1c, 0x3e241404, 0x3e242c04, 0x3e2c1414, 0x3e2c2414, 0x3e340414, 0x3e341c0c, 0x3e3e0404, +static const __device__ uint32_t iq3s_grid[512] = { + 0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305, + 0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905, + 0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09, + 0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b, + 0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b, + 0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d, + 0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03, + 0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505, + 0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03, + 0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901, + 0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d, + 0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303, + 0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501, + 0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105, + 0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505, + 0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101, + 0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707, + 0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b, + 0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01, + 0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f, + 0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305, + 0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103, + 0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509, + 0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503, + 0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b, + 0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f, + 0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f, + 0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f, + 0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109, + 0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f, + 0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509, + 0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501, + 0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303, + 0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f, + 0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907, + 0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703, + 0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03, + 0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01, + 0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01, + 0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903, + 0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505, + 0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b, + 0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107, + 0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509, + 0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303, + 0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103, + 0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05, + 0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b, + 0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f, + 0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701, + 0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909, + 0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305, + 0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d, + 0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b, + 0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d, + 0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307, + 0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09, + 0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309, + 0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709, + 0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f, + 0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303, + 0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503, + 0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b, + 0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101, }; - static const __device__ uint64_t iq1s_grid[512] = { 0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000, 0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01, @@ -2392,9 +2434,9 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_ const int ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; const uint8_t * qs = x[i].qs + 8*ib; - const uint8_t * grid1 = (const uint8_t *)(iq3xs_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3xs_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256))); - const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf)) * 0.5f; + const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256))); + const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256))); + const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf)); const uint8_t signs = x[i].signs[4*ib + il]; for (int j = 0; j < 4; ++j) { y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); @@ -5211,8 +5253,8 @@ static __device__ __forceinline__ float vec_dot_iq3_s_q8_1( const int8_t * q8 = bq8_1[ib32].qs; int sumi = 0; for (int l = 0; l < 4; ++l) { - const uint32_t * grid1 = iq3xs_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256)); - const uint32_t * grid2 = iq3xs_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256)); + const uint32_t * grid1 = iq3s_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256)); + const uint32_t * grid2 = iq3s_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256)); uint32_t signs0 = __vcmpeq4(((bq2->signs[4*ib32+l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201); uint32_t signs1 = __vcmpeq4(((bq2->signs[4*ib32+l] >> 4) * 0x01010101) & 0x08040201, 0x08040201); const int grid_l = __vsub4(grid1[0] ^ signs0, signs0); @@ -5221,7 +5263,7 @@ static __device__ __forceinline__ float vec_dot_iq3_s_q8_1( sumi = __dp4a(grid_h, *((int *)q8+1), sumi); q8 += 8; } - const float d = (float)bq2->d * (0.5f + ((bq2->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * __low2float(bq8_1[ib32].ds) * 0.5f; + const float d = (float)bq2->d * (1 + 2*((bq2->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * __low2float(bq8_1[ib32].ds); return d * sumi; #else assert(false); @@ -6449,7 +6491,7 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; + const int64_t i = blockDim.x*blockIdx.x + threadIdx.x; if (i >= ne) { return; @@ -6457,17 +6499,17 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne, // determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor // then combine those indices with the corresponding byte offsets to get the total offsets - const int i03 = i/(ne00 * ne01 * ne02); - const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); - const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; - const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; - const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; + const int64_t i03 = i/(ne00 * ne01 * ne02); + const int64_t i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); + const int64_t i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; + const int64_t i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; + const int64_t x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; - const int i13 = i/(ne10 * ne11 * ne12); - const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); - const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; - const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; - const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13; + const int64_t i13 = i/(ne10 * ne11 * ne12); + const int64_t i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); + const int64_t i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; + const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; + const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13; cpy_1(cx + x_offset, cdst + dst_offset); } @@ -6905,6 +6947,7 @@ static __global__ void soft_max_f32(const float * x, const float * mask, const f // find the sum of exps in the block tmp = warp_reduce_sum(tmp); if (block_size > WARP_SIZE) { + __syncthreads(); if (warp_id == 0) { buf_iw[lane_id] = 0.0f; } @@ -6956,23 +6999,23 @@ static __global__ void clamp_f32(const float * x, float * dst, const float min, template static __global__ void im2col_kernel( - const float * x, T * dst, int batch_offset, - int offset_delta, int IC, int IW, int IH, int OH, int OW, int KW, int KH, int pelements, int CHW, + const float * x, T * dst, int64_t batch_offset, + int64_t offset_delta, int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH, int64_t pelements, int64_t CHW, int s0, int s1, int p0, int p1, int d0, int d1) { - const int i = threadIdx.x + blockIdx.x * blockDim.x; + const int64_t i = threadIdx.x + blockIdx.x * blockDim.x; if (i >= pelements) { return; } - const int ksize = OW * (KH > 1 ? KW : 1); - const int kx = i / ksize; - const int kd = kx * ksize; - const int ky = (i - kd) / OW; - const int ix = i % OW; + const int64_t ksize = OW * (KH > 1 ? KW : 1); + const int64_t kx = i / ksize; + const int64_t kd = kx * ksize; + const int64_t ky = (i - kd) / OW; + const int64_t ix = i % OW; - const int oh = blockIdx.y; - const int batch = blockIdx.z / IC; - const int ic = blockIdx.z % IC; + const int64_t oh = blockIdx.y; + const int64_t batch = blockIdx.z / IC; + const int64_t ic = blockIdx.z % IC; const int64_t iiw = ix * s0 + kx * d0 - p0; const int64_t iih = oh * s1 + ky * d1 - p1; @@ -7298,19 +7341,33 @@ static void concat_f32_cuda(const float * x, const float * y, float * dst, const concat_f32<<>>(x, y, dst, ne0, ne02); } -static void upscale_f32_cuda(const float * x, float * dst, const int ne00, const int ne01, const int ne02, const int scale_factor, cudaStream_t stream) { +static void upscale_f32_cuda(const float * x, float * dst, const int ne00, const int ne01, const int ne02, const int ne03, + const int scale_factor, cudaStream_t stream) { int ne0 = (ne00 * scale_factor); int num_blocks = (ne0 + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE; - dim3 gridDim(num_blocks, (ne01 * scale_factor), ne02); + dim3 gridDim(num_blocks, (ne01 * scale_factor), ne02*ne03); upscale_f32<<>>(x, dst, ne00, ne00 * ne01, scale_factor); } static void pad_f32_cuda(const float * x, float * dst, - const int ne00, const int ne01, const int ne02, - const int ne0, const int ne1, const int ne2, cudaStream_t stream) { + const int ne00, const int ne01, const int ne02, const int ne03, + const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) { int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE; - dim3 gridDim(num_blocks, ne1, ne2); - pad_f32<<>>(x, dst, ne0, ne00, ne01, ne02); + dim3 gridDim(num_blocks, ne1, ne2*ne3); + pad_f32<<>>(x, dst, ne0, ne00, ne01, ne02, ne03); +} + +static void arange_f32_cuda(float * dst, const int ne0, const float start, const float step, cudaStream_t stream) { + int num_blocks = (ne0 + CUDA_ARANGE_BLOCK_SIZE - 1) / CUDA_ARANGE_BLOCK_SIZE; + arange_f32<<>>(dst, ne0, start, step); +} + +static void timestep_embedding_f32_cuda(const float * x, float * dst, const int ne00, const int nb1, + const int dim, const int max_period, cudaStream_t stream) { + int half_ceil = (dim + 1) / 2; + int num_blocks = (half_ceil + CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE; + dim3 gridDim(num_blocks, ne00, 1); + timestep_embedding_f32<<>>(x, dst, nb1, dim, max_period); } static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) { @@ -8443,8 +8500,8 @@ static void soft_max_f32_cuda(const float * x, const float * mask, const float * template static void im2col_cuda(const float* x, T* dst, - int IW, int IH, int OW, int OH, int KW, int KH, int IC, - int batch, int batch_offset, int offset_delta, + int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC, + int64_t batch, int64_t batch_offset, int64_t offset_delta, int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) { const int parallel_elements = OW * KW * KH; const int num_blocks = (parallel_elements + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE; @@ -9123,7 +9180,7 @@ static void ggml_cuda_op_group_norm( int num_groups = dst->op_params[0]; int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups); - group_norm_f32_cuda(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream); + group_norm_f32_cuda(src0_dd, dst_dd, num_groups * src0->ne[3], group_size, ggml_nelements(src0), main_stream); (void) src1; (void) dst; @@ -9156,7 +9213,7 @@ static void ggml_cuda_op_upscale( const int scale_factor = dst->op_params[0]; - upscale_f32_cuda(src0_dd, dst_dd, src0->ne[0], src0->ne[1], src0->ne[2], scale_factor, main_stream); + upscale_f32_cuda(src0_dd, dst_dd, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], scale_factor, main_stream); (void) src1; (void) dst; @@ -9172,8 +9229,49 @@ static void ggml_cuda_op_pad( GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors pad_f32_cuda(src0_dd, dst_dd, - src0->ne[0], src0->ne[1], src0->ne[2], - dst->ne[0], dst->ne[1], dst->ne[2], main_stream); + src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], + dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +static void ggml_cuda_op_arange( + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { + + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + float start; + float stop; + float step; + memcpy(&start, (float *)dst->op_params + 0, sizeof(float)); + memcpy(&stop, (float *)dst->op_params + 1, sizeof(float)); + memcpy(&step, (float *)dst->op_params + 2, sizeof(float)); + + int64_t steps = (int64_t)ceil((stop - start) / step); + GGML_ASSERT(ggml_nelements(dst) == steps); + + arange_f32_cuda(dst_dd, dst->ne[0], start, step, main_stream); + + (void) src0; + (void) src1; + (void) src0_dd; + (void) src1_dd; +} + +static void ggml_cuda_op_timestep_embedding( + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + const int dim = dst->op_params[0]; + const int max_period = dst->op_params[1]; + + timestep_embedding_f32_cuda(src0_dd, dst_dd, src0->ne[0], dst->nb[1], dim, max_period, main_stream); (void) src1; (void) dst; @@ -10458,6 +10556,45 @@ static void ggml_cuda_pad(const ggml_tensor * src0, const ggml_tensor * src1, gg ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_pad); } +static void ggml_cuda_arange(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + + const bool dst_on_device = dst->backend == GGML_BACKEND_TYPE_GPU; + + // dd = data device + float * src0_ddf = nullptr; + float * src1_ddf = nullptr; + float * dst_ddf = nullptr; + + cuda_pool_alloc dst_f; + + ggml_cuda_set_device(g_main_device); + cudaStream_t main_stream = g_cudaStreams[g_main_device][0]; + + if (dst_on_device) { + dst_ddf = (float *) dst_extra->data_device[g_main_device]; + } else { + dst_ddf = dst_f.alloc(ggml_nelements(dst)); + } + + // do the computation + ggml_cuda_op_arange(src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream); + CUDA_CHECK(cudaGetLastError()); + + // copy dst to host if necessary + if (!dst_on_device) { + CUDA_CHECK(cudaMemcpyAsync(dst->data, dst_ddf, ggml_nbytes(dst), cudaMemcpyDeviceToHost, main_stream)); + } + + if (dst->backend == GGML_BACKEND_TYPE_CPU) { + CUDA_CHECK(cudaDeviceSynchronize()); + } +} + +static void ggml_cuda_timestep_embedding(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_timestep_embedding); +} + static void ggml_cuda_rms_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rms_norm); } @@ -11358,6 +11495,12 @@ GGML_CALL bool ggml_cuda_compute_forward(struct ggml_compute_params * params, st case GGML_OP_PAD: func = ggml_cuda_pad; break; + case GGML_OP_ARANGE: + func = ggml_cuda_arange; + break; + case GGML_OP_TIMESTEP_EMBEDDING: + func = ggml_cuda_timestep_embedding; + break; case GGML_OP_LEAKY_RELU: func = ggml_cuda_leaky_relu; break; @@ -12098,7 +12241,7 @@ GGML_CALL static void ggml_backend_cuda_synchronize(ggml_backend_t backend) { UNUSED(backend); } -GGML_CALL static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { +GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; ggml_cuda_set_main_device(cuda_ctx->device); @@ -12134,7 +12277,7 @@ GGML_CALL static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, gg GGML_ASSERT(ok); } - return true; + return GGML_STATUS_SUCCESS; } GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) { @@ -12253,6 +12396,8 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons case GGML_OP_GROUP_NORM: case GGML_OP_UPSCALE: case GGML_OP_PAD: + case GGML_OP_ARANGE: + case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_LEAKY_RELU: return true; default: diff --git a/ggml-kompute.cpp b/ggml-kompute.cpp index e740a76..83a7822 100644 --- a/ggml-kompute.cpp +++ b/ggml-kompute.cpp @@ -1927,10 +1927,10 @@ static ggml_backend_buffer_type_t ggml_backend_kompute_get_default_buffer_type(g return ggml_backend_kompute_buffer_type(ctx->device); } -static bool ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { +static ggml_status ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { auto * ctx = static_cast(backend->context); ggml_vk_graph_compute(ctx, cgraph); - return true; + return GGML_STATUS_SUCCESS; } static bool ggml_backend_kompute_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { diff --git a/ggml-metal.m b/ggml-metal.m index 71fcca5..00df228 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -163,6 +163,8 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_IM2COL_F32, GGML_METAL_KERNEL_TYPE_UPSCALE_F32, GGML_METAL_KERNEL_TYPE_PAD_F32, + GGML_METAL_KERNEL_TYPE_ARANGE_F32, + GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, @@ -569,6 +571,8 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32, im2col_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, timestep_embedding_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32, arange_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, argsort_f32_i32_desc, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, leaky_relu_f32, true); @@ -697,6 +701,8 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const return false; case GGML_OP_UPSCALE: case GGML_OP_PAD: + case GGML_OP_ARANGE: + case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_ARGSORT: case GGML_OP_LEAKY_RELU: return true; @@ -742,7 +748,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const } } -static bool ggml_metal_graph_compute( +static enum ggml_status ggml_metal_graph_compute( struct ggml_metal_context * ctx, struct ggml_cgraph * gf) { @@ -1091,7 +1097,8 @@ static bool ggml_metal_graph_compute( { GGML_ASSERT(ggml_is_contiguous(src0)); - const float scale = *(const float *) dst->op_params; + float scale; + memcpy(&scale, dst->op_params, sizeof(scale)); int64_t n = ggml_nelements(dst); @@ -1250,11 +1257,15 @@ static bool ggml_metal_graph_compute( pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX].pipeline; } - const float scale = ((float *) dst->op_params)[0]; - const float max_bias = ((float *) dst->op_params)[1]; + float scale; + float max_bias; + + memcpy(&scale, ((int32_t *) dst->op_params) + 0, sizeof(scale)); + memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias)); const int64_t nrows_x = ggml_nrows(src0); const int64_t nrows_y = src0->ne[1]; + const uint32_t n_head_kv = nrows_x/nrows_y; const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv)); @@ -2086,6 +2097,7 @@ static bool ggml_metal_graph_compute( //const int n_past = ((int32_t *) dst->op_params)[0]; const int n_head = ((int32_t *) dst->op_params)[1]; + float max_bias; memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float)); @@ -2300,6 +2312,50 @@ static bool ggml_metal_graph_compute( [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; } break; + case GGML_OP_ARANGE: + { + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + float start; + float step; + + memcpy(&start, ((int32_t *) dst->op_params) + 0, sizeof(float)); + memcpy(&step, ((int32_t *) dst->op_params) + 2, sizeof(float)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARANGE_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:0]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:1]; + [encoder setBytes:&start length:sizeof(start) atIndex:2]; + [encoder setBytes:&step length:sizeof(step) atIndex:3]; + + const int nth = MIN(1024, ne0); + + [encoder dispatchThreadgroups:MTLSizeMake(1, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_TIMESTEP_EMBEDDING: + { + GGML_ASSERT(src0->type == GGML_TYPE_F32); + + const int dim = dst->op_params[0]; + const int max_period = dst->op_params[1]; + + const int half = dim / 2; + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:2]; + [encoder setBytes:&dim length:sizeof(dim) atIndex:3]; + [encoder setBytes:&max_period length:sizeof(max_period) atIndex:4]; + + const int nth = MIN(1024, half); + + [encoder dispatchThreadgroups:MTLSizeMake(ne00, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; case GGML_OP_ARGSORT: { GGML_ASSERT(src0->type == GGML_TYPE_F32); @@ -2428,7 +2484,7 @@ static bool ggml_metal_graph_compute( MTLCommandBufferStatus status = [command_buffer status]; if (status != MTLCommandBufferStatusCompleted) { GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status); - return false; + return GGML_STATUS_FAILED; } } @@ -2437,7 +2493,7 @@ static bool ggml_metal_graph_compute( } } - return true; + return GGML_STATUS_SUCCESS; } //////////////////////////////////////////////////////////////////////////////// @@ -2739,7 +2795,7 @@ GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffe UNUSED(backend); } -GGML_CALL static bool ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { +GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context; return ggml_metal_graph_compute(metal_ctx, cgraph); diff --git a/ggml-metal.metal b/ggml-metal.metal index 74a5e0b..a65d126 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -1959,6 +1959,49 @@ kernel void kernel_pad_f32( } } +kernel void kernel_arange_f32( + device char * dst, + constant int64_t & ne0, + constant float & start, + constant float & step, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + + device float * dst_ptr = (device float *) dst; + + for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { + dst_ptr[i0] = start + step * i0; + } +} + +kernel void kernel_timestep_embedding_f32( + device const char * src0, + device char * dst, + constant uint64_t & nb1, + constant int & dim, + constant int & max_period, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + + int i = tgpig.x; + device float * embed_data = (device float *)(dst + i*nb1); + + int half_ = dim / 2; + for (int j = tpitg.x; j < half_; j += ntg.x) { + float timestep = ((device float *)src0)[i]; + float freq = (float)exp(-log((float)max_period) * j / half_); + float arg = timestep * freq; + embed_data[j ] = cos(arg); + embed_data[j + half_] = sin(arg); + } + + if (dim % 2 != 0 && tpitg.x == 0) { + embed_data[dim] = 0.f; + } +} + // bitonic sort implementation following the CUDA kernels as reference typedef void (argsort_t)( device const float * x, @@ -4087,71 +4130,71 @@ constexpr constant static uint32_t iq3xxs_grid[256] = { 0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04, }; -constexpr constant static uint32_t iq3xs_grid[512] = { - 0x04040404, 0x0404040c, 0x04040414, 0x0404042c, 0x0404043e, 0x04040c04, 0x04040c0c, 0x04040c14, - 0x04040c24, 0x04040c34, 0x04041404, 0x0404140c, 0x0404142c, 0x04041c1c, 0x04042404, 0x04042414, - 0x0404242c, 0x0404243e, 0x04042c0c, 0x04042c1c, 0x04043404, 0x04043414, 0x04043e0c, 0x04043e24, - 0x04043e3e, 0x040c0404, 0x040c040c, 0x040c0414, 0x040c0424, 0x040c0c04, 0x040c0c0c, 0x040c0c2c, - 0x040c1404, 0x040c141c, 0x040c143e, 0x040c1c0c, 0x040c1c2c, 0x040c2424, 0x040c340c, 0x040c342c, - 0x040c3e14, 0x04140404, 0x0414040c, 0x0414042c, 0x0414043e, 0x04140c04, 0x04140c1c, 0x04140c34, - 0x0414140c, 0x0414142c, 0x04141c04, 0x04141c24, 0x04142414, 0x0414242c, 0x0414243e, 0x04142c0c, - 0x04142c1c, 0x04143e04, 0x04143e1c, 0x041c041c, 0x041c0c0c, 0x041c0c2c, 0x041c1404, 0x041c1414, - 0x041c1c0c, 0x041c1c1c, 0x041c1c34, 0x041c2424, 0x041c2c04, 0x041c2c14, 0x041c343e, 0x041c3e0c, - 0x041c3e2c, 0x04240404, 0x04240c1c, 0x04240c3e, 0x0424140c, 0x04241424, 0x04241c14, 0x04242404, - 0x0424241c, 0x04242c0c, 0x04243e04, 0x042c0414, 0x042c0424, 0x042c1404, 0x042c1414, 0x042c1434, - 0x042c1c1c, 0x042c240c, 0x042c242c, 0x042c243e, 0x042c3434, 0x042c3e1c, 0x04340434, 0x04340c0c, - 0x04340c1c, 0x04341c0c, 0x04342c14, 0x04343e0c, 0x043e0404, 0x043e0414, 0x043e0424, 0x043e1404, - 0x043e1414, 0x043e1434, 0x043e1c1c, 0x043e2c04, 0x043e2c24, 0x0c040404, 0x0c04040c, 0x0c040414, - 0x0c040424, 0x0c040c04, 0x0c040c0c, 0x0c040c1c, 0x0c040c2c, 0x0c040c3e, 0x0c041404, 0x0c041414, - 0x0c041c0c, 0x0c041c24, 0x0c041c34, 0x0c042c24, 0x0c042c34, 0x0c04340c, 0x0c043e14, 0x0c0c0404, - 0x0c0c040c, 0x0c0c041c, 0x0c0c0434, 0x0c0c0c04, 0x0c0c0c24, 0x0c0c140c, 0x0c0c1c04, 0x0c0c1c1c, - 0x0c0c240c, 0x0c0c2c04, 0x0c0c2c14, 0x0c0c3e04, 0x0c0c3e34, 0x0c140404, 0x0c140c14, 0x0c140c2c, - 0x0c140c3e, 0x0c141404, 0x0c141424, 0x0c141c14, 0x0c142404, 0x0c14241c, 0x0c142c2c, 0x0c143404, - 0x0c143e14, 0x0c1c040c, 0x0c1c0424, 0x0c1c043e, 0x0c1c0c04, 0x0c1c0c1c, 0x0c1c140c, 0x0c1c143e, - 0x0c1c1c04, 0x0c1c1c24, 0x0c1c240c, 0x0c1c3414, 0x0c1c3e04, 0x0c24041c, 0x0c24042c, 0x0c240c14, - 0x0c240c24, 0x0c241c0c, 0x0c241c1c, 0x0c242414, 0x0c242434, 0x0c242c04, 0x0c242c24, 0x0c2c040c, - 0x0c2c0c04, 0x0c2c0c1c, 0x0c2c140c, 0x0c2c1c04, 0x0c2c1c14, 0x0c2c2c0c, 0x0c341404, 0x0c341424, - 0x0c34143e, 0x0c342424, 0x0c342434, 0x0c3e040c, 0x0c3e041c, 0x0c3e0c04, 0x0c3e0c14, 0x0c3e140c, - 0x0c3e1c2c, 0x0c3e240c, 0x0c3e3414, 0x0c3e3e04, 0x14040404, 0x1404040c, 0x1404041c, 0x1404042c, - 0x1404043e, 0x14040c04, 0x14040c14, 0x14040c24, 0x14040c34, 0x1404140c, 0x1404141c, 0x1404143e, - 0x14041c04, 0x14041c14, 0x1404240c, 0x1404241c, 0x1404242c, 0x14042c04, 0x14042c14, 0x1404343e, - 0x14043e04, 0x14043e1c, 0x14043e2c, 0x140c0404, 0x140c0414, 0x140c0c04, 0x140c0c1c, 0x140c0c3e, - 0x140c1414, 0x140c142c, 0x140c1c0c, 0x140c1c24, 0x140c2414, 0x140c2c0c, 0x1414040c, 0x14140424, - 0x1414043e, 0x1414140c, 0x1414141c, 0x14141c04, 0x14141c3e, 0x1414240c, 0x14142c1c, 0x14142c3e, - 0x14143e0c, 0x14143e24, 0x141c0404, 0x141c0414, 0x141c042c, 0x141c0c0c, 0x141c1414, 0x141c1424, - 0x141c1c0c, 0x141c1c1c, 0x141c2414, 0x141c2c04, 0x141c3434, 0x1424040c, 0x1424043e, 0x14241404, - 0x1424141c, 0x14241c14, 0x14241c2c, 0x1424240c, 0x14243e14, 0x14243e2c, 0x142c0424, 0x142c0c0c, - 0x142c1414, 0x142c1c3e, 0x142c2404, 0x142c2c1c, 0x142c3e04, 0x14340404, 0x14340414, 0x1434043e, - 0x1434140c, 0x14342c2c, 0x1434340c, 0x143e042c, 0x143e0c0c, 0x143e1434, 0x143e1c04, 0x143e241c, - 0x143e2c04, 0x1c040414, 0x1c040c0c, 0x1c040c1c, 0x1c040c2c, 0x1c040c3e, 0x1c041414, 0x1c041c0c, - 0x1c041c1c, 0x1c041c2c, 0x1c042414, 0x1c042424, 0x1c04243e, 0x1c042c0c, 0x1c04341c, 0x1c043e0c, - 0x1c0c040c, 0x1c0c041c, 0x1c0c042c, 0x1c0c0c24, 0x1c0c140c, 0x1c0c141c, 0x1c0c2404, 0x1c0c3404, - 0x1c0c3e14, 0x1c0c3e34, 0x1c140404, 0x1c140c14, 0x1c141404, 0x1c141c14, 0x1c141c24, 0x1c142c04, - 0x1c1c040c, 0x1c1c0c04, 0x1c1c0c24, 0x1c1c140c, 0x1c1c141c, 0x1c1c143e, 0x1c1c1c04, 0x1c1c240c, - 0x1c1c241c, 0x1c1c243e, 0x1c1c2c2c, 0x1c1c3e1c, 0x1c24041c, 0x1c240c0c, 0x1c240c34, 0x1c241414, - 0x1c241c0c, 0x1c242c14, 0x1c243404, 0x1c243424, 0x1c2c040c, 0x1c2c0c04, 0x1c2c0c14, 0x1c2c142c, - 0x1c2c1c14, 0x1c2c2424, 0x1c2c2c34, 0x1c2c3e1c, 0x1c340c34, 0x1c34240c, 0x1c3e040c, 0x1c3e041c, - 0x1c3e1404, 0x1c3e1414, 0x1c3e1c2c, 0x24040404, 0x24040424, 0x24040c14, 0x24041404, 0x24041424, - 0x2404143e, 0x24041c14, 0x2404240c, 0x24042c04, 0x24043e04, 0x240c0414, 0x240c043e, 0x240c0c0c, - 0x240c0c1c, 0x240c1414, 0x240c1c04, 0x240c1c2c, 0x240c241c, 0x240c2c0c, 0x240c2c2c, 0x2414040c, - 0x2414041c, 0x24140c04, 0x24140c2c, 0x2414140c, 0x24141c1c, 0x24142404, 0x24142c3e, 0x24143414, - 0x24143e04, 0x241c0424, 0x241c0c0c, 0x241c0c1c, 0x241c1404, 0x241c1414, 0x241c1c0c, 0x241c1c2c, - 0x24240404, 0x24240414, 0x24241424, 0x24241c3e, 0x24242404, 0x24243e0c, 0x242c042c, 0x242c043e, - 0x242c140c, 0x242c3414, 0x24340c1c, 0x24341c24, 0x24343404, 0x243e0c04, 0x243e0c2c, 0x243e1c04, - 0x243e241c, 0x243e2c0c, 0x2c040414, 0x2c040c04, 0x2c040c24, 0x2c041414, 0x2c042404, 0x2c042424, - 0x2c04243e, 0x2c042c14, 0x2c043434, 0x2c043e24, 0x2c0c040c, 0x2c0c041c, 0x2c0c042c, 0x2c0c0c14, - 0x2c0c140c, 0x2c0c1c14, 0x2c0c3e14, 0x2c140404, 0x2c140c0c, 0x2c14141c, 0x2c141c04, 0x2c141c34, - 0x2c142c1c, 0x2c1c0414, 0x2c1c043e, 0x2c1c0c04, 0x2c1c143e, 0x2c1c2424, 0x2c1c2c0c, 0x2c1c342c, - 0x2c1c3e1c, 0x2c24040c, 0x2c240424, 0x2c241404, 0x2c241c14, 0x2c242434, 0x2c2c0c14, 0x2c2c1434, - 0x2c2c2c0c, 0x2c2c2c1c, 0x2c342414, 0x2c3e0414, 0x2c3e0424, 0x2c3e1414, 0x34040c0c, 0x34040c1c, - 0x34040c2c, 0x34041c0c, 0x34041c1c, 0x34043404, 0x340c0404, 0x340c1404, 0x340c143e, 0x340c3424, - 0x34140c14, 0x34141c24, 0x34142414, 0x34142c2c, 0x34143414, 0x34143e04, 0x341c0404, 0x341c0c24, - 0x341c140c, 0x341c2404, 0x3424142c, 0x3424241c, 0x34243414, 0x342c0404, 0x342c041c, 0x342c1c24, - 0x342c3404, 0x3434042c, 0x34342404, 0x343e0c0c, 0x343e0c1c, 0x3e040404, 0x3e040424, 0x3e04043e, - 0x3e041404, 0x3e041414, 0x3e041c34, 0x3e042404, 0x3e042c24, 0x3e043414, 0x3e0c0414, 0x3e0c0c0c, - 0x3e0c1424, 0x3e0c241c, 0x3e0c242c, 0x3e14040c, 0x3e140424, 0x3e140c04, 0x3e140c34, 0x3e14140c, - 0x3e141c04, 0x3e142c0c, 0x3e1c0414, 0x3e1c1c14, 0x3e1c1c2c, 0x3e1c2c1c, 0x3e24040c, 0x3e24042c, - 0x3e240c1c, 0x3e241404, 0x3e242c04, 0x3e2c1414, 0x3e2c2414, 0x3e340414, 0x3e341c0c, 0x3e3e0404, +constexpr constant static uint32_t iq3s_grid[512] = { + 0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305, + 0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905, + 0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09, + 0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b, + 0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b, + 0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d, + 0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03, + 0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505, + 0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03, + 0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901, + 0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d, + 0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303, + 0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501, + 0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105, + 0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505, + 0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101, + 0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707, + 0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b, + 0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01, + 0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f, + 0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305, + 0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103, + 0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509, + 0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503, + 0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b, + 0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f, + 0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f, + 0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f, + 0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109, + 0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f, + 0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509, + 0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501, + 0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303, + 0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f, + 0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907, + 0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703, + 0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03, + 0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01, + 0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01, + 0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903, + 0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505, + 0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b, + 0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107, + 0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509, + 0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303, + 0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103, + 0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05, + 0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b, + 0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f, + 0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701, + 0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909, + 0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305, + 0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d, + 0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b, + 0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d, + 0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307, + 0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09, + 0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309, + 0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709, + 0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f, + 0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303, + 0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503, + 0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b, + 0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101, }; #define NGRID_IQ1S 512 @@ -4742,7 +4785,7 @@ void kernel_mul_mv_iq3_s_f32_impl( { int nval = 8; int pos = (32*sgitg + tiisg)*nval; - for (int i = 0; i < nval; ++i) values[pos + i] = iq3xs_grid[pos + i]; + for (int i = 0; i < nval; ++i) values[pos + i] = iq3s_grid[pos + i]; threadgroup_barrier(mem_flags::mem_threadgroup); } @@ -4769,12 +4812,14 @@ void kernel_mul_mv_iq3_s_f32_impl( for (int row = 0; row < N_DST; row++) { const float db = dh[0]; - const float d = db * (0.5f + ((sc[0] >> 4*(ib%2)) & 0xf)); + const float d = db * (1 + 2*((sc[0] >> 4*(ib%2)) & 0xf)); float2 sum = {0}; for (int l = 0; l < 4; ++l) { - const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256))); - const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256))); + const threadgroup uint32_t * table1 = qh[0] & kmask_iq2xs[2*l+0] ? values + 256 : values; + const threadgroup uint32_t * table2 = qh[0] & kmask_iq2xs[2*l+1] ? values + 256 : values; + const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(table1 + qs[2*l+0]); + const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(table2 + qs[2*l+1]); for (int j = 0; j < 4; ++j) { sum[0] += yl[8*l + j + 0] * grid1[j] * select(1, -1, signs[l] & kmask_iq2xs[j+0]); sum[1] += yl[8*l + j + 4] * grid2[j] * select(1, -1, signs[l] & kmask_iq2xs[j+4]); @@ -4795,7 +4840,7 @@ void kernel_mul_mv_iq3_s_f32_impl( for (int row = 0; row < N_DST; ++row) { all_sum = simd_sum(sumf[row]); if (tiisg == 0) { - dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.5f; + dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum; } } } @@ -5685,15 +5730,15 @@ void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & device const uint8_t * qs = xb->qs + 8*ib32; device const uint8_t * signs = xb->signs + 4*ib32 + 2*il; const uint8_t qh = xb->qh[ib32] >> 4*il; - const float dl = d * (0.5f + ((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * 0.5f; - constant uint8_t * grid1 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+0] | ((qh << 8) & 256))); - constant uint8_t * grid2 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+1] | ((qh << 7) & 256))); + const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf)); + constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256))); + constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256))); for (int i = 0; i < 4; ++i) { reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]); reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]); } - grid1 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+2] | ((qh << 6) & 256))); - grid2 = (constant uint8_t *)(iq3xs_grid + (qs[4*il+3] | ((qh << 5) & 256))); + grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256))); + grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256))); for (int i = 0; i < 4; ++i) { reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]); reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]); diff --git a/ggml-opencl.cpp b/ggml-opencl.cpp index df619a8..aa73d67 100644 --- a/ggml-opencl.cpp +++ b/ggml-opencl.cpp @@ -2231,7 +2231,7 @@ static ggml_backend_buffer_type_t ggml_backend_opencl_get_default_buffer_type(gg GGML_UNUSED(backend); } -static bool ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) { +static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) { for (int i = 0; i < graph->n_nodes; ++i) { ggml_tensor * node = graph->nodes[i]; switch (node->op) { @@ -2246,7 +2246,7 @@ static bool ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgrap } } - return true; + return GGML_STATUS_SUCCESS; GGML_UNUSED(backend); } diff --git a/ggml-quants.c b/ggml-quants.c index 371826f..dee33e0 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -51,6 +51,7 @@ #define UNUSED GGML_UNUSED +// some compilers don't provide _mm256_set_m128i, e.g. gcc 7 #define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1) #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) @@ -463,8 +464,8 @@ inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) { } // NOTE: not tested -inline static int8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) { - int8x16_t res; +inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) { + uint8x16_t res; res[ 0] = a[b[ 0]]; res[ 1] = a[b[ 1]]; @@ -3818,71 +3819,71 @@ static const uint32_t iq3xxs_grid[256] = { 0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04, }; -static const uint32_t iq3xs_grid[512] = { - 0x04040404, 0x0404040c, 0x04040414, 0x0404042c, 0x0404043e, 0x04040c04, 0x04040c0c, 0x04040c14, - 0x04040c24, 0x04040c34, 0x04041404, 0x0404140c, 0x0404142c, 0x04041c1c, 0x04042404, 0x04042414, - 0x0404242c, 0x0404243e, 0x04042c0c, 0x04042c1c, 0x04043404, 0x04043414, 0x04043e0c, 0x04043e24, - 0x04043e3e, 0x040c0404, 0x040c040c, 0x040c0414, 0x040c0424, 0x040c0c04, 0x040c0c0c, 0x040c0c2c, - 0x040c1404, 0x040c141c, 0x040c143e, 0x040c1c0c, 0x040c1c2c, 0x040c2424, 0x040c340c, 0x040c342c, - 0x040c3e14, 0x04140404, 0x0414040c, 0x0414042c, 0x0414043e, 0x04140c04, 0x04140c1c, 0x04140c34, - 0x0414140c, 0x0414142c, 0x04141c04, 0x04141c24, 0x04142414, 0x0414242c, 0x0414243e, 0x04142c0c, - 0x04142c1c, 0x04143e04, 0x04143e1c, 0x041c041c, 0x041c0c0c, 0x041c0c2c, 0x041c1404, 0x041c1414, - 0x041c1c0c, 0x041c1c1c, 0x041c1c34, 0x041c2424, 0x041c2c04, 0x041c2c14, 0x041c343e, 0x041c3e0c, - 0x041c3e2c, 0x04240404, 0x04240c1c, 0x04240c3e, 0x0424140c, 0x04241424, 0x04241c14, 0x04242404, - 0x0424241c, 0x04242c0c, 0x04243e04, 0x042c0414, 0x042c0424, 0x042c1404, 0x042c1414, 0x042c1434, - 0x042c1c1c, 0x042c240c, 0x042c242c, 0x042c243e, 0x042c3434, 0x042c3e1c, 0x04340434, 0x04340c0c, - 0x04340c1c, 0x04341c0c, 0x04342c14, 0x04343e0c, 0x043e0404, 0x043e0414, 0x043e0424, 0x043e1404, - 0x043e1414, 0x043e1434, 0x043e1c1c, 0x043e2c04, 0x043e2c24, 0x0c040404, 0x0c04040c, 0x0c040414, - 0x0c040424, 0x0c040c04, 0x0c040c0c, 0x0c040c1c, 0x0c040c2c, 0x0c040c3e, 0x0c041404, 0x0c041414, - 0x0c041c0c, 0x0c041c24, 0x0c041c34, 0x0c042c24, 0x0c042c34, 0x0c04340c, 0x0c043e14, 0x0c0c0404, - 0x0c0c040c, 0x0c0c041c, 0x0c0c0434, 0x0c0c0c04, 0x0c0c0c24, 0x0c0c140c, 0x0c0c1c04, 0x0c0c1c1c, - 0x0c0c240c, 0x0c0c2c04, 0x0c0c2c14, 0x0c0c3e04, 0x0c0c3e34, 0x0c140404, 0x0c140c14, 0x0c140c2c, - 0x0c140c3e, 0x0c141404, 0x0c141424, 0x0c141c14, 0x0c142404, 0x0c14241c, 0x0c142c2c, 0x0c143404, - 0x0c143e14, 0x0c1c040c, 0x0c1c0424, 0x0c1c043e, 0x0c1c0c04, 0x0c1c0c1c, 0x0c1c140c, 0x0c1c143e, - 0x0c1c1c04, 0x0c1c1c24, 0x0c1c240c, 0x0c1c3414, 0x0c1c3e04, 0x0c24041c, 0x0c24042c, 0x0c240c14, - 0x0c240c24, 0x0c241c0c, 0x0c241c1c, 0x0c242414, 0x0c242434, 0x0c242c04, 0x0c242c24, 0x0c2c040c, - 0x0c2c0c04, 0x0c2c0c1c, 0x0c2c140c, 0x0c2c1c04, 0x0c2c1c14, 0x0c2c2c0c, 0x0c341404, 0x0c341424, - 0x0c34143e, 0x0c342424, 0x0c342434, 0x0c3e040c, 0x0c3e041c, 0x0c3e0c04, 0x0c3e0c14, 0x0c3e140c, - 0x0c3e1c2c, 0x0c3e240c, 0x0c3e3414, 0x0c3e3e04, 0x14040404, 0x1404040c, 0x1404041c, 0x1404042c, - 0x1404043e, 0x14040c04, 0x14040c14, 0x14040c24, 0x14040c34, 0x1404140c, 0x1404141c, 0x1404143e, - 0x14041c04, 0x14041c14, 0x1404240c, 0x1404241c, 0x1404242c, 0x14042c04, 0x14042c14, 0x1404343e, - 0x14043e04, 0x14043e1c, 0x14043e2c, 0x140c0404, 0x140c0414, 0x140c0c04, 0x140c0c1c, 0x140c0c3e, - 0x140c1414, 0x140c142c, 0x140c1c0c, 0x140c1c24, 0x140c2414, 0x140c2c0c, 0x1414040c, 0x14140424, - 0x1414043e, 0x1414140c, 0x1414141c, 0x14141c04, 0x14141c3e, 0x1414240c, 0x14142c1c, 0x14142c3e, - 0x14143e0c, 0x14143e24, 0x141c0404, 0x141c0414, 0x141c042c, 0x141c0c0c, 0x141c1414, 0x141c1424, - 0x141c1c0c, 0x141c1c1c, 0x141c2414, 0x141c2c04, 0x141c3434, 0x1424040c, 0x1424043e, 0x14241404, - 0x1424141c, 0x14241c14, 0x14241c2c, 0x1424240c, 0x14243e14, 0x14243e2c, 0x142c0424, 0x142c0c0c, - 0x142c1414, 0x142c1c3e, 0x142c2404, 0x142c2c1c, 0x142c3e04, 0x14340404, 0x14340414, 0x1434043e, - 0x1434140c, 0x14342c2c, 0x1434340c, 0x143e042c, 0x143e0c0c, 0x143e1434, 0x143e1c04, 0x143e241c, - 0x143e2c04, 0x1c040414, 0x1c040c0c, 0x1c040c1c, 0x1c040c2c, 0x1c040c3e, 0x1c041414, 0x1c041c0c, - 0x1c041c1c, 0x1c041c2c, 0x1c042414, 0x1c042424, 0x1c04243e, 0x1c042c0c, 0x1c04341c, 0x1c043e0c, - 0x1c0c040c, 0x1c0c041c, 0x1c0c042c, 0x1c0c0c24, 0x1c0c140c, 0x1c0c141c, 0x1c0c2404, 0x1c0c3404, - 0x1c0c3e14, 0x1c0c3e34, 0x1c140404, 0x1c140c14, 0x1c141404, 0x1c141c14, 0x1c141c24, 0x1c142c04, - 0x1c1c040c, 0x1c1c0c04, 0x1c1c0c24, 0x1c1c140c, 0x1c1c141c, 0x1c1c143e, 0x1c1c1c04, 0x1c1c240c, - 0x1c1c241c, 0x1c1c243e, 0x1c1c2c2c, 0x1c1c3e1c, 0x1c24041c, 0x1c240c0c, 0x1c240c34, 0x1c241414, - 0x1c241c0c, 0x1c242c14, 0x1c243404, 0x1c243424, 0x1c2c040c, 0x1c2c0c04, 0x1c2c0c14, 0x1c2c142c, - 0x1c2c1c14, 0x1c2c2424, 0x1c2c2c34, 0x1c2c3e1c, 0x1c340c34, 0x1c34240c, 0x1c3e040c, 0x1c3e041c, - 0x1c3e1404, 0x1c3e1414, 0x1c3e1c2c, 0x24040404, 0x24040424, 0x24040c14, 0x24041404, 0x24041424, - 0x2404143e, 0x24041c14, 0x2404240c, 0x24042c04, 0x24043e04, 0x240c0414, 0x240c043e, 0x240c0c0c, - 0x240c0c1c, 0x240c1414, 0x240c1c04, 0x240c1c2c, 0x240c241c, 0x240c2c0c, 0x240c2c2c, 0x2414040c, - 0x2414041c, 0x24140c04, 0x24140c2c, 0x2414140c, 0x24141c1c, 0x24142404, 0x24142c3e, 0x24143414, - 0x24143e04, 0x241c0424, 0x241c0c0c, 0x241c0c1c, 0x241c1404, 0x241c1414, 0x241c1c0c, 0x241c1c2c, - 0x24240404, 0x24240414, 0x24241424, 0x24241c3e, 0x24242404, 0x24243e0c, 0x242c042c, 0x242c043e, - 0x242c140c, 0x242c3414, 0x24340c1c, 0x24341c24, 0x24343404, 0x243e0c04, 0x243e0c2c, 0x243e1c04, - 0x243e241c, 0x243e2c0c, 0x2c040414, 0x2c040c04, 0x2c040c24, 0x2c041414, 0x2c042404, 0x2c042424, - 0x2c04243e, 0x2c042c14, 0x2c043434, 0x2c043e24, 0x2c0c040c, 0x2c0c041c, 0x2c0c042c, 0x2c0c0c14, - 0x2c0c140c, 0x2c0c1c14, 0x2c0c3e14, 0x2c140404, 0x2c140c0c, 0x2c14141c, 0x2c141c04, 0x2c141c34, - 0x2c142c1c, 0x2c1c0414, 0x2c1c043e, 0x2c1c0c04, 0x2c1c143e, 0x2c1c2424, 0x2c1c2c0c, 0x2c1c342c, - 0x2c1c3e1c, 0x2c24040c, 0x2c240424, 0x2c241404, 0x2c241c14, 0x2c242434, 0x2c2c0c14, 0x2c2c1434, - 0x2c2c2c0c, 0x2c2c2c1c, 0x2c342414, 0x2c3e0414, 0x2c3e0424, 0x2c3e1414, 0x34040c0c, 0x34040c1c, - 0x34040c2c, 0x34041c0c, 0x34041c1c, 0x34043404, 0x340c0404, 0x340c1404, 0x340c143e, 0x340c3424, - 0x34140c14, 0x34141c24, 0x34142414, 0x34142c2c, 0x34143414, 0x34143e04, 0x341c0404, 0x341c0c24, - 0x341c140c, 0x341c2404, 0x3424142c, 0x3424241c, 0x34243414, 0x342c0404, 0x342c041c, 0x342c1c24, - 0x342c3404, 0x3434042c, 0x34342404, 0x343e0c0c, 0x343e0c1c, 0x3e040404, 0x3e040424, 0x3e04043e, - 0x3e041404, 0x3e041414, 0x3e041c34, 0x3e042404, 0x3e042c24, 0x3e043414, 0x3e0c0414, 0x3e0c0c0c, - 0x3e0c1424, 0x3e0c241c, 0x3e0c242c, 0x3e14040c, 0x3e140424, 0x3e140c04, 0x3e140c34, 0x3e14140c, - 0x3e141c04, 0x3e142c0c, 0x3e1c0414, 0x3e1c1c14, 0x3e1c1c2c, 0x3e1c2c1c, 0x3e24040c, 0x3e24042c, - 0x3e240c1c, 0x3e241404, 0x3e242c04, 0x3e2c1414, 0x3e2c2414, 0x3e340414, 0x3e341c0c, 0x3e3e0404, +static const uint32_t iq3s_grid[512] = { + 0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305, + 0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905, + 0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09, + 0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b, + 0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b, + 0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d, + 0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03, + 0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505, + 0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03, + 0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901, + 0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d, + 0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303, + 0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501, + 0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105, + 0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505, + 0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101, + 0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707, + 0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b, + 0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01, + 0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f, + 0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305, + 0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103, + 0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509, + 0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503, + 0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b, + 0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f, + 0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f, + 0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f, + 0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109, + 0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f, + 0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509, + 0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501, + 0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303, + 0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f, + 0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907, + 0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703, + 0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03, + 0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01, + 0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01, + 0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903, + 0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505, + 0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b, + 0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107, + 0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509, + 0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303, + 0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103, + 0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05, + 0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b, + 0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f, + 0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701, + 0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909, + 0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305, + 0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d, + 0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b, + 0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d, + 0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307, + 0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09, + 0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309, + 0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709, + 0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f, + 0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303, + 0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503, + 0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b, + 0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101, }; #define NGRID_IQ2XXS 512 @@ -4162,11 +4163,11 @@ void dequantize_row_iq3_s(const block_iq3_s * restrict x, float * restrict y, in const uint8_t * signs = x[i].signs; for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { - const float db1 = d * (0.5f + (x[i].scales[ib32/2] & 0xf)) * 0.5f; - const float db2 = d * (0.5f + (x[i].scales[ib32/2] >> 4)) * 0.5f; + const float db1 = d * (1 + 2*(x[i].scales[ib32/2] & 0xf)); + const float db2 = d * (1 + 2*(x[i].scales[ib32/2] >> 4)); for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3xs_grid + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3xs_grid + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256))); + const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256))); + const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256))); for (int j = 0; j < 4; ++j) { y[j+0] = db1 * grid1[j] * (signs[l] & kmask_iq2xs[j+0] ? -1.f : 1.f); y[j+4] = db1 * grid2[j] * (signs[l] & kmask_iq2xs[j+4] ? -1.f : 1.f); @@ -4176,8 +4177,8 @@ void dequantize_row_iq3_s(const block_iq3_s * restrict x, float * restrict y, in qs += 8; signs += 4; for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3xs_grid + (qs[2*l+0] | ((qh[1] << (8-2*l)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3xs_grid + (qs[2*l+1] | ((qh[1] << (7-2*l)) & 256))); + const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[1] << (8-2*l)) & 256))); + const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[1] << (7-2*l)) & 256))); for (int j = 0; j < 4; ++j) { y[j+0] = db2 * grid1[j] * (signs[l] & kmask_iq2xs[j+0] ? -1.f : 1.f); y[j+4] = db2 * grid2[j] * (signs[l] & kmask_iq2xs[j+4] ? -1.f : 1.f); @@ -9563,7 +9564,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * const __m128i odd_bits = _mm_shuffle_epi8(bit_helper, partial_sign_bits_for_counting); const __m128i full_sign_bits = _mm_or_si128(partial_sign_bits, odd_bits); - const __m256i full_signs = _mm256_set_m128i(full_sign_bits, full_sign_bits); + const __m256i full_signs = MM256_SET_M128I(full_sign_bits, full_sign_bits); const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)y[i].qs); const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)(y[i].qs+32)); @@ -9585,8 +9586,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1); const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2); - const __m256i sc1 = _mm256_set_m128i(_mm_set1_epi16(2*(x[i].scales[0] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[0] & 0xf)+1)); - const __m256i sc2 = _mm256_set_m128i(_mm_set1_epi16(2*(x[i].scales[1] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[1] & 0xf)+1)); + const __m256i sc1 = MM256_SET_M128I(_mm_set1_epi16(2*(x[i].scales[0] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[0] & 0xf)+1)); + const __m256i sc2 = MM256_SET_M128I(_mm_set1_epi16(2*(x[i].scales[1] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[1] & 0xf)+1)); const __m256i sum = _mm256_add_epi32(_mm256_madd_epi16(sc1, dot1), _mm256_madd_epi16(sc2, dot2)); @@ -9653,8 +9654,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * const __m128i full_signs_l = _mm256_castsi256_si128(full_sign_bits); const __m128i full_signs_h = _mm256_extractf128_si256(full_sign_bits, 1); - const __m256i full_signs_1 = _mm256_set_m128i(full_signs_l, full_signs_l); - const __m256i full_signs_2 = _mm256_set_m128i(full_signs_h, full_signs_h); + const __m256i full_signs_1 = MM256_SET_M128I(full_signs_l, full_signs_l); + const __m256i full_signs_2 = MM256_SET_M128I(full_signs_h, full_signs_h); __m256i signs; signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_1); @@ -9757,8 +9758,8 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,}; - const uint8x16x2_t mask1 = vld1q_u8_x2(k_mask1); - const uint8x16_t mask2 = vld1q_u8(k_mask2); + const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1); + const uint8x16_t mask2 = vld1q_u8(k_mask2); const uint8x16_t m1 = vdupq_n_u8(1); const int32x4_t vzero = vdupq_n_s32(0); @@ -10089,18 +10090,35 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v #if defined(__ARM_NEON) + typedef union { + uint16x8_t vec_index; + uint16_t index[8]; + } vec_index_t; + static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03 }; static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,}; - const uint8x16x2_t mask1 = vld1q_u8_x2(k_mask1); - const uint8x16_t mask2 = vld1q_u8(k_mask2); + static const int16_t k_shift[8] = {8, 7, 6, 5, 4, 3, 2, 1}; + + const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1); + const uint8x16_t mask2 = vld1q_u8(k_mask2); + + const int16x8_t hshift = vld1q_s16(k_shift); + const uint16x8_t m256 = vdupq_n_u16(256); + const uint8x16_t m1 = vdupq_n_u8(1); uint8x16x2_t vs; ggml_int8x16x4_t q3s; ggml_int8x16x4_t q8b; + vec_index_t idx; + +#if QK_K == 256 + uint32_t scales32[2]; + const uint8_t * scales8 = (const uint8_t *)scales32; +#endif float sumf = 0; for (int i = 0; i < nb; ++i) { @@ -10109,47 +10127,63 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v const uint8_t * restrict qh = x[i].qh; const uint16_t * restrict signs = (const uint16_t *)x[i].signs; const int8_t * restrict q8 = y[i].qs; + +#if QK_K == 256 + memcpy(scales32, x[i].scales, 4); + scales32[1] = (((scales32[0] >> 4) & 0x0f0f0f0f) << 1) | 0x01010101; + scales32[0] = ((scales32[0] & 0x0f0f0f0f) << 1) | 0x01010101; +#endif + int sumi1 = 0, sumi2 = 0; for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { q8b = ggml_vld1q_s8_x4(q8); q8 += 64; - const uint32x4_t aux32x4_0 = {iq3xs_grid[qs[ 0] | ((qh[ib32+0] << 8) & 256)], iq3xs_grid[qs[ 1] | ((qh[ib32+0] << 7) & 256)], - iq3xs_grid[qs[ 2] | ((qh[ib32+0] << 6) & 256)], iq3xs_grid[qs[ 3] | ((qh[ib32+0] << 5) & 256)]}; - const uint32x4_t aux32x4_1 = {iq3xs_grid[qs[ 4] | ((qh[ib32+0] << 4) & 256)], iq3xs_grid[qs[ 5] | ((qh[ib32+0] << 3) & 256)], - iq3xs_grid[qs[ 6] | ((qh[ib32+0] << 2) & 256)], iq3xs_grid[qs[ 7] | ((qh[ib32+0] << 1) & 256)]}; - const uint32x4_t aux32x4_2 = {iq3xs_grid[qs[ 8] | ((qh[ib32+1] << 8) & 256)], iq3xs_grid[qs[ 9] | ((qh[ib32+1] << 7) & 256)], - iq3xs_grid[qs[10] | ((qh[ib32+1] << 6) & 256)], iq3xs_grid[qs[11] | ((qh[ib32+1] << 5) & 256)]}; - const uint32x4_t aux32x4_3 = {iq3xs_grid[qs[12] | ((qh[ib32+1] << 4) & 256)], iq3xs_grid[qs[13] | ((qh[ib32+1] << 3) & 256)], - iq3xs_grid[qs[14] | ((qh[ib32+1] << 2) & 256)], iq3xs_grid[qs[15] | ((qh[ib32+1] << 1) & 256)]}; - qs += 16; + + const uint8x16_t idx_l = vld1q_u8(qs); qs += 16; + idx.vec_index = vorrq_u16(vmovl_u8(vget_low_u8 (idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+0]), hshift), m256)); + const uint32x4_t aux32x4_0 = {iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]], + iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]}; + const uint32x4_t aux32x4_1 = {iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]], + iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]}; + idx.vec_index = vorrq_u16(vmovl_u8(vget_high_u8(idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+1]), hshift), m256)); + const uint32x4_t aux32x4_2 = {iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]], + iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]}; + const uint32x4_t aux32x4_3 = {iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]], + iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]}; + vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | (signs[1] << 16))); vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2); vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2); - vs.val[0] = vceqq_u8(vs.val[0], mask2); - vs.val[1] = vceqq_u8(vs.val[1], mask2); + vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1); + vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1); - q3s.val[0] = vsubq_s8(vreinterpretq_s8_u8(veorq_u8(vs.val[0], vreinterpretq_u8_u32(aux32x4_0))), vreinterpretq_s8_u8(vs.val[0])); - q3s.val[1] = vsubq_s8(vreinterpretq_s8_u8(veorq_u8(vs.val[1], vreinterpretq_u8_u32(aux32x4_1))), vreinterpretq_s8_u8(vs.val[1])); + q3s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_0)); + q3s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_1)); vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | (signs[3] << 16))); vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2); vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2); - vs.val[0] = vceqq_u8(vs.val[0], mask2); - vs.val[1] = vceqq_u8(vs.val[1], mask2); + vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1); + vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1); signs += 4; - q3s.val[2] = vsubq_s8(vreinterpretq_s8_u8(veorq_u8(vs.val[0], vreinterpretq_u8_u32(aux32x4_2))), vreinterpretq_s8_u8(vs.val[0])); - q3s.val[3] = vsubq_s8(vreinterpretq_s8_u8(veorq_u8(vs.val[1], vreinterpretq_u8_u32(aux32x4_3))), vreinterpretq_s8_u8(vs.val[1])); + q3s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_2)); + q3s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_3)); const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]); const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]); +#if QK_K == 256 + sumi1 += vaddvq_s32(p1) * scales8[ib32/2+0]; + sumi2 += vaddvq_s32(p2) * scales8[ib32/2+4]; +#else sumi1 += vaddvq_s32(p1) * (1 + 2*(x[i].scales[ib32/2] & 0xf)); sumi2 += vaddvq_s32(p2) * (1 + 2*(x[i].scales[ib32/2] >> 4)); +#endif } sumf += d*(sumi1 + sumi2); } - *s = 0.25f * sumf; + *s = sumf; #elif defined(__AVX2__) @@ -10164,6 +10198,16 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1); const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2); + const __m256i idx_shift = _mm256_set_epi32(1, 2, 3, 4, 5, 6, 7, 8); + const __m256i idx_mask = _mm256_set1_epi32(256); + + typedef union { + __m256i vec[2]; + uint32_t index[16]; + } index_t; + + index_t idx; + __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; @@ -10176,24 +10220,25 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; - const __m256i q2_1 = _mm256_set_epi32(iq3xs_grid[qs[7] | ((qh[ib32+0] << 1) & 256)], - iq3xs_grid[qs[6] | ((qh[ib32+0] << 2) & 256)], - iq3xs_grid[qs[5] | ((qh[ib32+0] << 3) & 256)], - iq3xs_grid[qs[4] | ((qh[ib32+0] << 4) & 256)], - iq3xs_grid[qs[3] | ((qh[ib32+0] << 5) & 256)], - iq3xs_grid[qs[2] | ((qh[ib32+0] << 6) & 256)], - iq3xs_grid[qs[1] | ((qh[ib32+0] << 7) & 256)], - iq3xs_grid[qs[0] | ((qh[ib32+0] << 8) & 256)]); - qs += 8; - const __m256i q2_2 = _mm256_set_epi32(iq3xs_grid[qs[7] | ((qh[ib32+1] << 1) & 256)], - iq3xs_grid[qs[6] | ((qh[ib32+1] << 2) & 256)], - iq3xs_grid[qs[5] | ((qh[ib32+1] << 3) & 256)], - iq3xs_grid[qs[4] | ((qh[ib32+1] << 4) & 256)], - iq3xs_grid[qs[3] | ((qh[ib32+1] << 5) & 256)], - iq3xs_grid[qs[2] | ((qh[ib32+1] << 6) & 256)], - iq3xs_grid[qs[1] | ((qh[ib32+1] << 7) & 256)], - iq3xs_grid[qs[0] | ((qh[ib32+1] << 8) & 256)]); - qs += 8; + const __m256i idx_l = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)qs)); qs += 16; + idx.vec[0] = _mm256_set1_epi32(qh[ib32+0]); + idx.vec[1] = _mm256_set1_epi32(qh[ib32+1]); + idx.vec[0] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[0], idx_shift), idx_mask); + idx.vec[1] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[1], idx_shift), idx_mask); + idx.vec[0] = _mm256_or_si256(idx.vec[0], _mm256_cvtepi16_epi32(_mm256_castsi256_si128(idx_l))); + idx.vec[1] = _mm256_or_si256(idx.vec[1], _mm256_cvtepi16_epi32(_mm256_extractf128_si256(idx_l, 1))); + + // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange. + //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4); + //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4); + const __m256i q2_1 = _mm256_set_epi32( + iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]], + iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]] + ); + const __m256i q2_2 = _mm256_set_epi32( + iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]], + iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]] + ); __m256i aux256 = _mm256_set1_epi32(signs[0] | (signs[1] << 16)); aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2); @@ -10221,7 +10266,7 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v } - *s = 0.25f * hsum_float_8(accumf); + *s = hsum_float_8(accumf); #else @@ -10238,8 +10283,8 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v const uint32_t ls2 = 2*(x[i].scales[ib32/2] >> 4) + 1; int32_t sumi = 0; for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3xs_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3xs_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256))); + const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256))); + const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256))); for (int j = 0; j < 4; ++j) { sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); @@ -10251,8 +10296,8 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v bsum += sumi * ls1; sumi = 0; for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3xs_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3xs_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256))); + const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256))); + const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256))); for (int j = 0; j < 4; ++j) { sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); @@ -10265,7 +10310,7 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v } sumf += d * bsum; } - *s = 0.25f * sumf; + *s = sumf; #endif } @@ -10508,10 +10553,10 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void * const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[1].qs); const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[0].qs); const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[1].qs); - const __m256i q4b_1 = _mm256_set_m128i(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)), - _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b))); - const __m256i q4b_2 = _mm256_set_m128i(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)), - _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b))); + const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)), + _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b))); + const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)), + _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b))); const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1); const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2); const __m256i p_1 = _mm256_madd_epi16(p16_1, mone); @@ -10618,10 +10663,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)qs); qs += 16; const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32; - const __m256i q4b_1 = _mm256_set_m128i(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)), - _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b))); - const __m256i q4b_2 = _mm256_set_m128i(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)), - _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b))); + const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)), + _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b))); + const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)), + _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b))); const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1); const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2); const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32; @@ -11912,7 +11957,8 @@ static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, vo } float best = 0; float scale = max/(2*kMaxQ-1); - for (int is = -15; is <= 15; ++is) { + for (int k = 0; k < bs4; ++k) is_on_grid[k] = false; + for (int is = -9; is <= 9; ++is) { float id = (2*kMaxQ-1+is*0.2f)/max; float this_scale = 1/id; for (int k = 0; k < bs4; ++k) { @@ -11948,7 +11994,7 @@ static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, vo if (n_not_ongrid > 0 && scale > 0) { float id = 1/scale; for (int k = 0; k < bs4; ++k) { - if (is_on_grid[k]) continue; + //if (is_on_grid[k]) continue; uint16_t u = 0; for (int i = 0; i < 4; ++i) { int l = nearest_int(0.5f*(id*xval[4*k+i]-1)); @@ -12004,7 +12050,7 @@ static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, vo } float d = max_scale/31; - y[ibl].d = GGML_FP32_TO_FP16(d); + y[ibl].d = GGML_FP32_TO_FP16(d * 1.033f); float id = 1/d; for (int ib = 0; ib < QK_K/block_size; ib += 2) { int l1 = nearest_int(0.5f*(id*scales[ib+0]-1)); diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp index a054ec8..ddd951d 100644 --- a/ggml-sycl.cpp +++ b/ggml-sycl.cpp @@ -661,26 +661,29 @@ namespace dpct /// \param [out] total_memory The number of bytes of total memory on the SYCL device. void get_memory_info(size_t &free_memory, size_t &total_memory) { + total_memory = get_device_info().get_global_mem_size(); + const char *warning_info = "get_memory_info: [warning] ext_intel_free_memory is not " + "supported (export/set ZES_ENABLE_SYSMAN=1 to support), " + "use total memory as free memory"; #if (defined(__SYCL_COMPILER_VERSION) && __SYCL_COMPILER_VERSION >= 20221105) if (!has(sycl::aspect::ext_intel_free_memory)) { - std::cerr << "get_memory_info: ext_intel_free_memory is not supported." << std::endl; - free_memory = 0; + std::cerr << warning_info << std::endl; + free_memory = total_memory; } else { free_memory = get_info(); } #else - std::cerr << "get_memory_info: ext_intel_free_memory is not supported." << std::endl; - free_memory = 0; + std::cerr << warning_info << std::endl; + free_memory = total_memory; #if defined(_MSC_VER) && !defined(__clang__) #pragma message("Querying the number of bytes of free memory is not supported") #else #warning "Querying the number of bytes of free memory is not supported" #endif #endif - total_memory = get_device_info().get_global_mem_size(); } void get_device_info(device_info &out) const @@ -738,15 +741,25 @@ namespace dpct #endif // DPCT_USM_LEVEL_NONE } - sycl::queue *create_in_order_queue(bool enable_exception_handler = false) - { - std::lock_guard lock(m_mutex); - return create_queue_impl(enable_exception_handler, - sycl::property::queue::in_order()); + sycl::queue *create_queue(sycl::context context, sycl::device device, + bool enable_exception_handler = false) { + return create_in_order_queue(context, device, enable_exception_handler); } - sycl::queue *create_out_of_order_queue(bool enable_exception_handler = false) - { + sycl::queue *create_in_order_queue(bool enable_exception_handler = false) { + std::lock_guard lock(m_mutex); + return create_queue_impl(enable_exception_handler, + sycl::property::queue::in_order()); + } + + sycl::queue *create_in_order_queue(sycl::context context, sycl::device device, + bool enable_exception_handler = false) { + std::lock_guard lock(m_mutex); + return create_queue_impl(context, device, enable_exception_handler, + sycl::property::queue::in_order()); + } + + sycl::queue *create_out_of_order_queue(bool enable_exception_handler = false) { std::lock_guard lock(m_mutex); return create_queue_impl(enable_exception_handler); } @@ -809,6 +822,25 @@ namespace dpct return _queues.back().get(); } + template + sycl::queue *create_queue_impl(sycl::context context, sycl::device device, + bool enable_exception_handler, + Properties... properties) { + sycl::async_handler eh = {}; + if (enable_exception_handler) { + eh = exception_handler; + } + _queues.push_back(std::make_shared( + context, device, eh, + sycl::property_list( + #ifdef DPCT_PROFILING_ENABLED + sycl::property::queue::enable_profiling(), + #endif + properties...))); + + return _queues.back().get(); + } + void get_version(int &major, int &minor) const { detail::get_version(*this, major, minor); @@ -2862,6 +2894,254 @@ namespace dpct using err0 = detail::generic_error_type; using err1 = detail::generic_error_type; + static inline void dpct_free(void *ptr, sycl::queue &q = get_default_queue()) { + detail::dpct_free(ptr, q); + } + + /// dpct accessor used as device function parameter. + template class accessor; + template class accessor { + public: + using memory_t = detail::memory_traits; + using element_t = typename memory_t::element_t; + using pointer_t = typename memory_t::pointer_t; + using accessor_t = typename memory_t::template accessor_t<3>; + accessor(pointer_t data, const sycl::range<3> &in_range) + : _data(data), _range(in_range) {} + template + accessor(typename std::enable_if::type &acc) + : accessor(acc, acc.get_range()) {} + accessor(const accessor_t &acc, const sycl::range<3> &in_range) + : accessor(acc.get_pointer(), in_range) {} + accessor operator[](size_t index) const { + sycl::range<2> sub(_range.get(1), _range.get(2)); + return accessor(_data + index * sub.size(), sub); + } + + pointer_t get_ptr() const { return _data; } + + private: + pointer_t _data; + sycl::range<3> _range; + }; + template class accessor { + public: + using memory_t = detail::memory_traits; + using element_t = typename memory_t::element_t; + using pointer_t = typename memory_t::pointer_t; + using accessor_t = typename memory_t::template accessor_t<2>; + accessor(pointer_t data, const sycl::range<2> &in_range) + : _data(data), _range(in_range) {} + template + accessor(typename std::enable_if::type &acc) + : accessor(acc, acc.get_range()) {} + accessor(const accessor_t &acc, const sycl::range<2> &in_range) + : accessor(acc.get_pointer(), in_range) {} + + pointer_t operator[](size_t index) const { + return _data + _range.get(1) * index; + } + + pointer_t get_ptr() const { return _data; } + + private: + pointer_t _data; + sycl::range<2> _range; + }; + + namespace detail { + /// Device variable with address space of shared, global or constant. + template class device_memory { + public: + using accessor_t = + typename detail::memory_traits::template accessor_t; + using value_t = typename detail::memory_traits::value_t; + using dpct_accessor_t = dpct::accessor; + + device_memory() : device_memory(sycl::range(1)) {} + + /// Constructor of 1-D array with initializer list + device_memory(const sycl::range &in_range, + std::initializer_list &&init_list) + : device_memory(in_range) { + assert(init_list.size() <= in_range.size()); + _host_ptr = (value_t *)std::malloc(_size); + std::memset(_host_ptr, 0, _size); + std::memcpy(_host_ptr, init_list.begin(), init_list.size() * sizeof(T)); + } + + /// Constructor of 2-D array with initializer list + template + device_memory( + const typename std::enable_if>::type &in_range, + std::initializer_list> &&init_list) + : device_memory(in_range) { + assert(init_list.size() <= in_range[0]); + _host_ptr = (value_t *)std::malloc(_size); + std::memset(_host_ptr, 0, _size); + auto tmp_data = _host_ptr; + for (auto sub_list : init_list) { + assert(sub_list.size() <= in_range[1]); + std::memcpy(tmp_data, sub_list.begin(), + sub_list.size() * sizeof(T)); + tmp_data += in_range[1]; + } + } + + /// Constructor with range + device_memory(const sycl::range &range_in) + : _size(range_in.size() * sizeof(T)), _range(range_in), + _reference(false), _host_ptr(nullptr), _device_ptr(nullptr) { + static_assert( + (Memory == global) || (Memory == constant) || (Memory == shared), + "device memory region should be global, constant or shared"); + // Make sure that singleton class mem_mgr and dev_mgr will destruct + // later than this. + detail::mem_mgr::instance(); + dev_mgr::instance(); + } + + /// Constructor with range + template + device_memory(Args... Arguments) + : device_memory(sycl::range(Arguments...)) {} + + ~device_memory() { + if (_device_ptr && !_reference) + dpct::dpct_free(_device_ptr); + if (_host_ptr) + std::free(_host_ptr); + } + + /// Allocate memory with default queue, and init memory if has initial + /// value. + void init() { init(dpct::get_default_queue()); } + /// Allocate memory with specified queue, and init memory if has initial + /// value. + void init(sycl::queue &q) { + if (_device_ptr) + return; + if (!_size) + return; + allocate_device(q); + if (_host_ptr) + detail::dpct_memcpy(q, _device_ptr, _host_ptr, _size, + host_to_device); + } + + /// The variable is assigned to a device pointer. + void assign(value_t *src, size_t size) { + this->~device_memory(); + new (this) device_memory(src, size); + } + + /// Get memory pointer of the memory object, which is virtual pointer when + /// usm is not used, and device pointer when usm is used. + value_t *get_ptr() { return get_ptr(get_default_queue()); } + /// Get memory pointer of the memory object, which is virtual pointer when + /// usm is not used, and device pointer when usm is used. + value_t *get_ptr(sycl::queue &q) { + init(q); + return _device_ptr; + } + + /// Get the device memory object size in bytes. + size_t get_size() { return _size; } + + template + typename std::enable_if::type &operator[](size_t index) { + init(); + #ifdef DPCT_USM_LEVEL_NONE + return dpct::get_buffer::type>( + _device_ptr) + .template get_access()[index]; + #else + return _device_ptr[index]; + #endif // DPCT_USM_LEVEL_NONE + } + + #ifdef DPCT_USM_LEVEL_NONE + /// Get sycl::accessor for the device memory object when usm is not used. + accessor_t get_access(sycl::handler &cgh) { + return get_buffer(_device_ptr) + .template reinterpret(_range) + .template get_access::mode, + detail::memory_traits::target>(cgh); + } + #else + /// Get dpct::accessor with dimension info for the device memory object + /// when usm is used and dimension is greater than 1. + template + typename std::enable_if::type + get_access(sycl::handler &cgh) { + return dpct_accessor_t((T *)_device_ptr, _range); + } + #endif // DPCT_USM_LEVEL_NONE + + private: + device_memory(value_t *memory_ptr, size_t size) + : _size(size), _range(size / sizeof(T)), _reference(true), + _device_ptr(memory_ptr) {} + + void allocate_device(sycl::queue &q) { + #ifndef DPCT_USM_LEVEL_NONE + if (Memory == shared) { + _device_ptr = (value_t *)sycl::malloc_shared(_size, q.get_device(), + q.get_context()); + return; + } + #ifdef SYCL_EXT_ONEAPI_USM_DEVICE_READ_ONLY + if (Memory == constant) { + _device_ptr = (value_t *)sycl::malloc_device( + _size, q.get_device(), q.get_context(), + sycl::ext::oneapi::property::usm::device_read_only()); + return; + } + #endif + #endif + _device_ptr = (value_t *)detail::dpct_malloc(_size, q); + } + + size_t _size; + sycl::range _range; + bool _reference; + value_t *_host_ptr; + value_t *_device_ptr; + }; + template + class device_memory : public device_memory { + public: + using base = device_memory; + using value_t = typename base::value_t; + using accessor_t = + typename detail::memory_traits::template accessor_t<0>; + + /// Constructor with initial value. + device_memory(const value_t &val) : base(sycl::range<1>(1), {val}) {} + + /// Default constructor + device_memory() : base(1) {} + + #ifdef DPCT_USM_LEVEL_NONE + /// Get sycl::accessor for the device memory object when usm is not used. + accessor_t get_access(sycl::handler &cgh) { + auto buf = get_buffer(base::get_ptr()) + .template reinterpret(sycl::range<1>(1)); + return accessor_t(buf, cgh); + } + #endif // DPCT_USM_LEVEL_NONE + }; + } // namespace detail + + template + using global_memory = detail::device_memory; + template + using constant_memory = detail::device_memory; + template + using shared_memory = detail::device_memory; + + } // COPY from DPCT head files @@ -2906,6 +3186,15 @@ static int g_work_group_size = 0; #pragma warning(disable: 4244 4267) // possible loss of data #endif +// dmmv = dequantize_mul_mat_vec +#ifndef GGML_SYCL_DMMV_X +#define GGML_SYCL_DMMV_X 32 +#endif +#ifndef GGML_SYCL_MMV_Y +#define GGML_SYCL_MMV_Y 1 +#endif + + static_assert(sizeof(sycl::half) == sizeof(ggml_fp16_t), "wrong fp16 size"); static void crash(){ @@ -2943,14 +3232,11 @@ bool ggml_sycl_loaded(void); void * ggml_sycl_host_malloc(size_t size); void ggml_sycl_host_free(void * ptr); bool ggml_sycl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); -void ggml_sycl_set_tensor_split(const float * tensor_split); -void ggml_sycl_transform_tensor(void * data, struct ggml_tensor * tensor); void ggml_sycl_free_data(struct ggml_tensor * tensor); void ggml_sycl_assign_buffers(struct ggml_tensor * tensor); void ggml_sycl_assign_buffers_no_scratch(struct ggml_tensor * tensor); void ggml_sycl_assign_buffers_force_inplace(struct ggml_tensor * tensor); void ggml_sycl_assign_buffers_no_alloc(struct ggml_tensor * tensor); -void ggml_sycl_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset); void ggml_sycl_copy_to_device(struct ggml_tensor * tensor); void ggml_sycl_set_main_device(int main_device); void ggml_sycl_set_mul_mat_q(bool mul_mat_q); @@ -2963,6 +3249,14 @@ int get_main_device(); void print_ggml_tensor(const char*name, struct ggml_tensor *src); void log_tensor_with_cnt(const char* name, struct ggml_tensor * src, int stop_cnt); +void dev2dev_memcpy(sycl::queue &q_dst, sycl::queue &q_src, void *ptr_dst, + const void *ptr_src, size_t size) { + char *host_buf = (char *)malloc(size); + q_src.memcpy(host_buf, (const char *)ptr_src, size).wait(); + q_dst.memcpy((char *)ptr_dst, host_buf, size).wait(); + free(host_buf); +} + static __dpct_inline__ int get_int_from_int8(const int8_t *x8, const int &i32) { const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment @@ -3023,7 +3317,7 @@ typedef void (*ggml_sycl_op_flatten_t)(const ggml_tensor *src0, #define QK4_0 32 #define QR4_0 2 #define QI4_0 (QK4_0 / (4 * QR4_0)) -typedef struct dpct_type_471834 { +typedef struct dpct_type_block_q4_0 { sycl::half d; // delta uint8_t qs[QK4_0 / 2]; // nibbles / quants } block_q4_0; @@ -3032,7 +3326,7 @@ static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 #define QK4_1 32 #define QR4_1 2 #define QI4_1 (QK4_1 / (4 * QR4_1)) -typedef struct dpct_type_143705 { +typedef struct dpct_type_block_q4_1 { sycl::half2 dm; // dm.x = delta, dm.y = min uint8_t qs[QK4_1 / 2]; // nibbles / quants } block_q4_1; @@ -3041,7 +3335,7 @@ static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong #define QK5_0 32 #define QR5_0 2 #define QI5_0 (QK5_0 / (4 * QR5_0)) -typedef struct dpct_type_673649 { +typedef struct dpct_type_block_q5_0 { sycl::half d; // delta uint8_t qh[4]; // 5-th bit of quants uint8_t qs[QK5_0 / 2]; // nibbles / quants @@ -3051,7 +3345,7 @@ static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5 #define QK5_1 32 #define QR5_1 2 #define QI5_1 (QK5_1 / (4 * QR5_1)) -typedef struct dpct_type_135589 { +typedef struct dpct_type_block_q5_1 { sycl::half2 dm; // dm.x = delta, dm.y = min uint8_t qh[4]; // 5-th bit of quants uint8_t qs[QK5_1 / 2]; // nibbles / quants @@ -3061,7 +3355,7 @@ static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + #define QK8_0 32 #define QR8_0 1 #define QI8_0 (QK8_0 / (4 * QR8_0)) -typedef struct dpct_type_122878 { +typedef struct dpct_type_block_q8_0 { sycl::half d; // delta int8_t qs[QK8_0]; // quants } block_q8_0; @@ -3070,7 +3364,7 @@ static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 blo #define QK8_1 32 #define QR8_1 1 #define QI8_1 (QK8_1 / (4 * QR8_1)) -typedef struct dpct_type_143721 { +typedef struct dpct_type_block_q8_1 { sycl::half2 ds; // ds.x = delta, ds.y = sum int8_t qs[QK8_0]; // quants } block_q8_1; @@ -3104,7 +3398,7 @@ typedef float (*vec_dot_q_mul_mat_sycl_t)( #define QR2_K 4 #define QI2_K (QK_K / (4*QR2_K)) -typedef struct dpct_type_619598 { +typedef struct dpct_type_block_q2_K { uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits uint8_t qs[QK_K/4]; // quants sycl::half2 dm; // super-block scale for quantized scales/mins @@ -3113,7 +3407,7 @@ static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "w #define QR3_K 4 #define QI3_K (QK_K / (4*QR3_K)) -typedef struct dpct_type_138576 { +typedef struct dpct_type_block_q3_K { uint8_t hmask[QK_K/8]; // quants - high bit uint8_t qs[QK_K/4]; // quants - low 2 bits #ifdef GGML_QKK_64 @@ -3129,13 +3423,13 @@ typedef struct dpct_type_138576 { #define QI4_K (QK_K / (4*QR4_K)) #ifdef GGML_QKK_64 typedef struct { - half dm[2]; // super-block scales/mins + sycl::half dm[2]; // super-block scales/mins uint8_t scales[2]; // 4-bit block scales/mins uint8_t qs[QK_K/2]; // 4--bit quants } block_q4_K; -static_assert(sizeof(block_q4_K) == sizeof(half2) + QK_K/2 + 2, "wrong q4_K block size/padding"); +static_assert(sizeof(block_q4_K) == sizeof(sycl::half2) + QK_K/2 + 2, "wrong q4_K block size/padding"); #else -typedef struct dpct_type_154943 { +typedef struct dpct_type_block_q4_K { sycl::half2 dm; // super-block scale for quantized scales/mins uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits uint8_t qs[QK_K/2]; // 4--bit quants @@ -3147,14 +3441,14 @@ static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, #define QI5_K (QK_K / (4*QR5_K)) #ifdef GGML_QKK_64 typedef struct { - half d; // super-block scale + sycl::half d; // super-block scale int8_t scales[QK_K/16]; // block scales uint8_t qh[QK_K/8]; // quants, high bit uint8_t qs[QK_K/2]; // quants, low 4 bits } block_q5_K; static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding"); #else -typedef struct dpct_type_866817 { +typedef struct dpct_type_block_q5_K { sycl::half2 dm; // super-block scale for quantized scales/mins uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits uint8_t qh[QK_K/8]; // quants, high bit @@ -3165,7 +3459,7 @@ static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/ #define QR6_K 2 #define QI6_K (QK_K / (4*QR6_K)) -typedef struct dpct_type_107281 { +typedef struct dpct_type_block_q6_K { uint8_t ql[QK_K/2]; // quants, lower 4 bits uint8_t qh[QK_K/4]; // quants, upper 2 bits int8_t scales[QK_K/16]; // scales @@ -3173,6 +3467,31 @@ typedef struct dpct_type_107281 { } block_q6_K; static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding"); +#define QR2_XXS 8 +#define QI2_XXS (QK_K / (4*QR2_XXS)) +typedef struct dpct_type_block_iq2_xxs { + sycl::half d; + uint16_t qs[QK_K/8]; +} block_iq2_xxs; +static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding"); + +#define QR2_XS 8 +#define QI2_XS (QK_K / (4*QR2_XS)) +typedef struct dpct_type_block_iq2_xs { + sycl::half d; + uint16_t qs[QK_K/8]; + uint8_t scales[QK_K/32]; +} block_iq2_xs; +static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding"); + +#define QR3_XXS 8 +#define QI3_XXS (QK_K / (4*QR3_XXS)) +typedef struct dpct_type_block_iq3_xxs { + sycl::half d; + uint8_t qs[3*(QK_K/8)]; +} block_iq3_xxs; +static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding"); + #define WARP_SIZE 32 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses @@ -3180,6 +3499,8 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_ #define SYCL_SILU_BLOCK_SIZE 256 #define SYCL_TANH_BLOCK_SIZE 256 #define SYCL_RELU_BLOCK_SIZE 256 +#define SYCL_HARDSIGMOID_BLOCK_SIZE 256 +#define SYCL_HARDSWISH_BLOCK_SIZE 256 #define SYCL_SQR_BLOCK_SIZE 256 #define SYCL_CPY_BLOCK_SIZE 32 #define SYCL_SCALE_BLOCK_SIZE 256 @@ -3196,6 +3517,7 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_ #define SYCL_PAD_BLOCK_SIZE 256 #define SYCL_ACC_BLOCK_SIZE 256 #define SYCL_IM2COL_BLOCK_SIZE 256 +#define SYCL_POOL2D_BLOCK_SIZE 256 // dmmv = dequantize_mul_mat_vec #ifndef GGML_SYCL_DMMV_X @@ -3218,8 +3540,7 @@ static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUA #define MUL_MAT_SRC1_COL_STRIDE 128 #define MAX_STREAMS 8 -static dpct::queue_ptr g_syclStreams[GGML_SYCL_MAX_DEVICES][MAX_STREAMS] = { - {0}}; +static dpct::queue_ptr g_syclStreams[GGML_SYCL_MAX_DEVICES][MAX_STREAMS] = {{0}}; struct ggml_tensor_extra_gpu { void * data_device[GGML_SYCL_MAX_DEVICES]; // 1 pointer for each device for split tensors @@ -3228,30 +3549,108 @@ struct ggml_tensor_extra_gpu { [MAX_STREAMS]; // events for synchronizing multiple GPUs }; -inline dpct::err0 ggml_sycl_set_device(const int device) try { - int current_device; +class sycl_gpu_mgr { + public: + std::vector gpus; + std::vector devices; + sycl::queue *first_queue; + sycl::context co_ctx; + int max_compute_units = 0; + int work_group_size = 0; + std::string gpus_list = ""; - SYCL_CHECK(CHECK_TRY_ERROR( - current_device = dpct::dev_mgr::instance().current_device_id())); + sycl_gpu_mgr() { + detect_sycl_gpu_list_with_max_cu(); + get_allow_gpus(); + create_context_with_gpus(); + } - // GGML_SYCL_DEBUG("ggml_sycl_set_device device=%d, current_device=%d\n", device, current_device); - if (device == current_device) { - return 0; - } + void create_context_with_gpus() { + sycl::context ctx = sycl::context(devices); + assert(gpus.size() > 0); + first_queue = dpct::get_current_device().create_queue(ctx, devices[0]); + co_ctx = first_queue->get_context(); + } - return CHECK_TRY_ERROR(dpct::select_device(device)); -} -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - crash(); - std::exit(1); -} + sycl::context &get_co_ctx() { return co_ctx; } + void get_allow_gpus() { + gpus_list = ""; + for (size_t i = 0; i < gpus.size(); ++i) { + gpus_list += std::to_string(gpus[i]); + gpus_list += ","; + } + if (gpus_list.length() > 2) { + gpus_list.pop_back(); + } + } + + bool is_allowed_gpu(int device_id) { + return std::find(gpus.begin(), gpus.end(), device_id) != gpus.end(); + } + + void detect_sycl_gpu_list_with_max_cu() try { + int device_count = dpct::dev_mgr::instance().device_count(); + + for (int id = 0; id < device_count; id++) { + sycl::device device = dpct::dev_mgr::instance().get_device(id); + if (!device.is_gpu()) + continue; + dpct::device_info prop; + dpct::get_device_info(prop, device); + if (max_compute_units < prop.get_max_compute_units()) + max_compute_units = prop.get_max_compute_units(); + } + + for (int id = 0; id < device_count; id++) { + sycl::device device = dpct::dev_mgr::instance().get_device(id); + if (!device.is_gpu()) + continue; + dpct::device_info prop; + dpct::get_device_info(prop, device); + if (max_compute_units == prop.get_max_compute_units() && + prop.get_major_version() == 1) { + gpus.push_back(id); + devices.push_back(device); + work_group_size = prop.get_max_work_group_size(); + } + } + return; + } catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); + } + + int get_gpu_count() { return (int)gpus.size(); } + + int get_index(int id) { + for (int i = 0; i < (int)gpus.size(); i++) { + if (gpus[i] == id) + return i; + } + assert(false); + return -1; + } + + int get_next_index(int id) { + int cur_index = get_index(id); + for (int i = cur_index + 1; i < (int)gpus.size(); i++) { + if (gpus[i] == id) + return i; + } + assert(false); + return -1; + } +}; + +static sycl_gpu_mgr *g_sycl_gpu_mgr = NULL; static int g_device_count = -1; static int g_all_sycl_device_count = -1; static int g_main_device = -1; -static int g_main_device_index = -1; +static int g_main_device_id = -1; + +static std::array g_default_tensor_split = {}; static float g_tensor_split[GGML_SYCL_MAX_DEVICES] = {0}; @@ -3268,8 +3667,6 @@ struct sycl_device_id2index { int index; }; -static sycl_device_id2index g_sycl_device_id2index[GGML_SYCL_MAX_DEVICES] = { {-1} }; - static void * g_scratch_buffer = nullptr; static size_t g_scratch_size = 0; // disabled by default static size_t g_scratch_offset = 0; @@ -3290,6 +3687,63 @@ static void bad_arch(const sycl::stream &stream_ct1) { (void) bad_arch; // suppress unused function warning } +/* +device_index: device index from 0 to n (continue numbers). + It is used for device select/set in SYCL backend internal data structure. +*/ +void check_allow_gpu_index(const int device_index) { + if (device_index >= g_device_count) { + char error_buf[256]; + snprintf(error_buf, sizeof(error_buf), + "%s error: device_index:%d is out of range: [0-%d]", __func__, + device_index, g_device_count - 1); + fprintf(stderr, "%s\n", error_buf); + assert(false); + } +} + +/* +device_id: device ID is shown by ggml_backend_sycl_print_sycl_devices(). + It is only used to set current working device. +*/ +void check_allow_gpu_id(const int device_id) { + if (!g_sycl_gpu_mgr->is_allowed_gpu(device_id)) { + char error_buf[256]; + snprintf(error_buf, sizeof(error_buf), + "error: cannot set device=%d, which is not allowed. Please " + "set GPU ID in: [%s]", + device_id, g_sycl_gpu_mgr->gpus_list.c_str()); + fprintf(stderr, "%s\n", error_buf); + throw std::invalid_argument(error_buf); + } +} + +int get_current_device_id() { + return dpct::dev_mgr::instance().current_device_id(); +} + +inline dpct::err0 ggml_sycl_set_device(const int device) try { + + int device_id = g_sycl_gpu_mgr->gpus[device]; + check_allow_gpu_id(device_id); + + int current_device_id; + SYCL_CHECK(CHECK_TRY_ERROR(current_device_id = get_current_device_id())); + + // GGML_SYCL_DEBUG("ggml_sycl_set_device device_id=%d, + // current_device_id=%d\n", device, current_device); + if (device_id == current_device_id) { + return 0; + } + + return CHECK_TRY_ERROR(dpct::select_device(device_id)); +} catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + crash(); + std::exit(1); +} + void log_ggml_var_device(const char*name, float *src, size_t total_elements, bool src_on_device){ if(!g_ggml_sycl_debug) return; if(!src){ @@ -3302,25 +3756,55 @@ void log_ggml_var_device(const char*name, float *src, size_t total_elements, boo size_t total_size = total_elements*sizeof(float); float *local_buf = NULL; - // printf("total_size %d2, src_on_device %d\n", total_size, src_on_device); if(src_on_device) { local_buf = (float *) ggml_sycl_host_malloc(total_size); - // printf("local buf %p size %d bytes\n", local_buf, total_size); ggml_sycl_set_device(g_main_device); - dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0]; - main_stream->memcpy(local_buf, src, total_size); + dpct::queue_ptr main_stream = g_syclStreams[g_main_device][0]; + main_stream->memcpy(local_buf, src, total_size).wait(); } else { local_buf = (float *)src; - // printf("local buf from src-> data %p\n", local_buf); } std::ofstream logfile; logfile.open(filename); - // printf("local buf element %d\n", total_elements); for(size_t i=0; imemcpy(local_buf, src, total_size).wait(); + } + else { + local_buf = (sycl::half *)src; + } + + std::ofstream logfile; + logfile.open(filename); + for(size_t i=0; iextra; - src_data = (float*)src_extra->data_device[g_main_device_index]; + src_data = (float*)src_extra->data_device[g_main_device]; } else { src_data = (float *)src->data; @@ -3359,10 +3843,6 @@ void log_tensor_with_cnt(const char* name, struct ggml_tensor * src, int stop_cn sprintf(filename, "%s_%07d", name, log_file_name_idx); log_file_name_idx++; print_ggml_tensor(filename, src); - // print_ggml_tensor("ggml_sycl_rms_norm_src0", (ggml_tensor *)src0); - // print_ggml_tensor("ggml_sycl_rms_norm_src1", (ggml_tensor *)src1); - // int *ptr = NULL; - // *ptr = 0; } static __dpct_inline__ float warp_reduce_sum(float x, @@ -3583,6 +4063,28 @@ static void relu_f32(const float * x, float * dst, const int k, dst[i] = sycl::fmax((float)(x[i]), (float)0); } +static void hardsigmoid_f32(const float * x, float * dst, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i >= k) { + return; + } + dst[i] = sycl::fmin(1.0f, sycl::fmax(0.0f, (x[i] + 3.0f) / 6.0f)); +} + +static void hardswish_f32(const float * x, float * dst, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i >= k) { + return; + } + dst[i] = x[i] * sycl::fmin(1.0f, sycl::fmax(0.0f, (x[i] + 3.0f) / 6.0f)); +} + static void leaky_relu_f32(const float *x, float *dst, const int k, const float negative_slope, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + @@ -3943,6 +4445,66 @@ static __dpct_inline__ void dequantize_q8_0(const void *vx, const int ib, #endif // GGML_SYCL_F16 } +template +static void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32, + const sycl::nd_item<3> &item_ct1) { + + const int i = item_ct1.get_group(2); + + // assume 32 threads + const int tid = item_ct1.get_local_id(2); + const int il = tid/8; + const int ir = tid%8; + const int ib = 8*i + ir; + if (ib >= nb32) { + return; + } + + dst_t * y = yy + 256*i + 32*ir + 4*il; + + const block_q4_0 * x = (const block_q4_0 *)vx + ib; + const float d = sycl::vec(x->d) + .convert()[0]; + const float dm = -8*d; + + const uint8_t * q = x->qs + 4*il; + + for (int l = 0; l < 4; ++l) { + y[l+ 0] = d * (q[l] & 0xF) + dm; + y[l+16] = d * (q[l] >> 4) + dm; + } +} + +template +static void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32, + const sycl::nd_item<3> &item_ct1) { + + const int i = item_ct1.get_group(2); + + // assume 32 threads + const int tid = item_ct1.get_local_id(2); + const int il = tid/8; + const int ir = tid%8; + const int ib = 8*i + ir; + if (ib >= nb32) { + return; + } + + dst_t * y = yy + 256*i + 32*ir + 4*il; + + const block_q4_1 * x = (const block_q4_1 *)vx + ib; + const sycl::float2 d = + x->dm.convert(); + + const uint8_t * q = x->qs + 4*il; + + for (int l = 0; l < 4; ++l) { + y[l + 0] = d.x() * (q[l] & 0xF) + d.y(); + y[l + 16] = d.x() * (q[l] >> 4) + d.y(); + } +} + + //================================== k-quants template @@ -3972,8 +4534,9 @@ static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restri const int il = tid%16; // 0...15 const uint8_t q = x[i].qs[il] >> (2*is); dst_t * y = yy + i*QK_K + 16*is + il; - float dall = __low2half(x[i].dm); - float dmin = __high2half(x[i].dm); + + float dall = x[i].dm[0]; + float dmin = x[i].dm[1]; y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4); #endif @@ -4012,7 +4575,7 @@ static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restri for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)); #else - const int tid = threadIdx.x; + const int tid = item_ct1.get_local_id(2); const int is = tid/16; // 0 or 1 const int il = tid%16; // 0...15 const int im = il/8; // 0...1 @@ -4078,7 +4641,7 @@ static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restri y[l +32] = d2 * (q[l] >> 4) - m2; } #else - const int tid = threadIdx.x; + const int tid = item_ct1.get_local_id(2); const uint8_t * q = x[i].qs; dst_t * y = yy + i*QK_K; const float d = (float)x[i].dm[0]; @@ -4123,7 +4686,7 @@ static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restri y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2; y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2; #else - const int tid = threadIdx.x; + const int tid = item_ct1.get_local_id(2); const uint8_t q = x[i].qs[tid]; const int im = tid/8; // 0...3 const int in = tid%8; // 0...7 @@ -4165,7 +4728,7 @@ static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restri #else // assume 32 threads - const int tid = threadIdx.x; + const int tid = item_ct1.get_local_id(2); const int ip = tid/16; // 0 or 1 const int il = tid - 16*ip; // 0...15 @@ -4182,6 +4745,474 @@ static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restri #endif } +static dpct::global_memory + iq2xxs_grid(sycl::range<1>(256), + { + 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, + 0x0808080808082b08, 0x0808080808082b2b, 0x0808080808190819, + 0x0808080808191908, 0x08080808082b0808, 0x08080808082b082b, + 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819, + 0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, + 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808, + 0x080808082b08082b, 0x080808082b082b2b, 0x080808082b2b082b, + 0x0808081908080819, 0x0808081908081908, 0x0808081908190808, + 0x0808081908191919, 0x0808081919080808, 0x080808192b081908, + 0x080808192b192b08, 0x0808082b08080808, 0x0808082b0808082b, + 0x0808082b082b082b, 0x0808082b2b08082b, 0x0808190808080819, + 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819, + 0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, + 0x0808190819082b08, 0x08081908192b0808, 0x080819082b080819, + 0x080819082b081908, 0x080819082b190808, 0x080819082b2b1908, + 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08, + 0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, + 0x080819192b080808, 0x080819192b190819, 0x0808192b08082b19, + 0x0808192b08190808, 0x0808192b19080808, 0x0808192b2b081908, + 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919, + 0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, + 0x08082b0819080819, 0x08082b0819081908, 0x08082b0819190808, + 0x08082b081919082b, 0x08082b082b082b08, 0x08082b1908081908, + 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908, + 0x0819080808080819, 0x0819080808081908, 0x0819080808190808, + 0x08190808082b0819, 0x0819080819080808, 0x08190808192b0808, + 0x081908082b081908, 0x081908082b190808, 0x081908082b191919, + 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808, + 0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, + 0x0819082b082b1908, 0x0819082b19081919, 0x0819190808080808, + 0x0819190808082b08, 0x08191908082b0808, 0x08191908082b1919, + 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08, + 0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, + 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808, + 0x08192b0819080808, 0x08192b082b080819, 0x08192b1908080808, + 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819, + 0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, + 0x082b080819081908, 0x082b0808192b0819, 0x082b08082b080808, + 0x082b08082b08082b, 0x082b0819082b2b19, 0x082b081919082b08, + 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819, + 0x082b190808081908, 0x082b190808190808, 0x082b190819080808, + 0x082b19081919192b, 0x082b191908080808, 0x082b191919080819, + 0x082b1919192b1908, 0x082b192b2b190808, 0x082b2b0808082b08, + 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908, + 0x1908080808080819, 0x1908080808081908, 0x1908080808190808, + 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908, + 0x1908080819080808, 0x1908080819082b08, 0x190808081919192b, + 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908, + 0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, + 0x19080819192b0819, 0x190808192b080808, 0x190808192b081919, + 0x1908082b08080819, 0x1908082b08190808, 0x1908082b19082b08, + 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808, + 0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, + 0x190819082b192b19, 0x190819190819082b, 0x19081919082b1908, + 0x1908192b08080808, 0x19082b0808080819, 0x19082b0808081908, + 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, + 0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, + 0x19082b192b08082b, 0x19082b2b19081919, 0x19082b2b2b190808, + 0x1919080808080808, 0x1919080808082b08, 0x1919080808190819, + 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808, + 0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, + 0x191908192b2b1908, 0x1919082b2b190819, 0x191919082b190808, + 0x191919082b19082b, 0x1919191908082b2b, 0x1919192b08080819, + 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819, + 0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, + 0x19192b2b08082b08, 0x192b080808081908, 0x192b080808190808, + 0x192b080819080808, 0x192b0808192b2b08, 0x192b081908080808, + 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808, + 0x192b190808080808, 0x192b190808081919, 0x192b191908190808, + 0x192b19190819082b, 0x192b19192b081908, 0x192b2b081908082b, + 0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808082b2b, + 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908, + 0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, + 0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, + 0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, + 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908, + 0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, + 0x2b082b080808082b, 0x2b082b1908081908, 0x2b082b2b08190819, + 0x2b19080808081908, 0x2b19080808190808, 0x2b190808082b1908, + 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b, + 0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, + 0x2b191908082b082b, 0x2b19190819081908, 0x2b19191919190819, + 0x2b192b082b080819, 0x2b192b19082b0808, 0x2b2b08080808082b, + 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19, + 0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, + 0x2b2b2b1908081908, + }); + +static dpct::global_memory + iq2xs_grid(sycl::range<1>(512), + { + 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, + 0x0808080808082b08, 0x0808080808082b2b, 0x0808080808190819, + 0x0808080808191908, 0x080808080819192b, 0x0808080808192b19, + 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, + 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, + 0x080808081908192b, 0x0808080819082b19, 0x0808080819190808, + 0x080808081919082b, 0x0808080819191919, 0x0808080819192b08, + 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808, + 0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, + 0x080808082b190819, 0x080808082b191908, 0x080808082b192b19, + 0x080808082b2b0808, 0x0808081908080819, 0x0808081908081908, + 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, + 0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, + 0x0808081908192b2b, 0x08080819082b0819, 0x08080819082b1908, + 0x0808081919080808, 0x080808191908082b, 0x0808081919081919, + 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908, + 0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, + 0x080808192b081908, 0x080808192b190808, 0x0808082b08080808, + 0x0808082b0808082b, 0x0808082b08081919, 0x0808082b08082b08, + 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808, + 0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, + 0x0808082b19191919, 0x0808082b2b080808, 0x0808082b2b082b2b, + 0x0808190808080819, 0x0808190808081908, 0x080819080808192b, + 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, + 0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, + 0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, + 0x0808190819081919, 0x0808190819082b08, 0x0808190819190819, + 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808, + 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, + 0x0808191908080808, 0x080819190808082b, 0x0808191908081919, + 0x0808191908082b08, 0x0808191908190819, 0x0808191908191908, + 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908, + 0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, + 0x0808192b08080819, 0x0808192b08081908, 0x0808192b08190808, + 0x0808192b082b192b, 0x0808192b19080808, 0x0808192b1908082b, + 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b, + 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, + 0x08082b0808190819, 0x08082b0808191908, 0x08082b08082b0808, + 0x08082b08082b1919, 0x08082b0819080819, 0x08082b0819081908, + 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808, + 0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, + 0x08082b1908081908, 0x08082b1908190808, 0x08082b1919080808, + 0x08082b192b080819, 0x08082b192b082b19, 0x08082b2b08080808, + 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b, + 0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, + 0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, + 0x081908080819082b, 0x0819080808191919, 0x0819080808192b08, + 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808, + 0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, + 0x0819080819190819, 0x0819080819191908, 0x08190808192b0808, + 0x08190808192b2b2b, 0x081908082b080819, 0x081908082b081908, + 0x081908082b190808, 0x0819081908080808, 0x081908190808082b, + 0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, + 0x0819081908191908, 0x08190819082b0808, 0x0819081919080819, + 0x0819081919081908, 0x0819081919190808, 0x081908192b080808, + 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819, + 0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, + 0x0819082b19080808, 0x0819082b192b0808, 0x0819190808080808, + 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08, + 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808, + 0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, + 0x0819190819190808, 0x08191908192b1908, 0x081919082b080808, + 0x0819191908080819, 0x0819191908081908, 0x0819191908190808, + 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908, + 0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, + 0x08192b0808190808, 0x08192b080819082b, 0x08192b0819080808, + 0x08192b0819191908, 0x08192b082b08192b, 0x08192b1908080808, + 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819, + 0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, + 0x082b080808081919, 0x082b080808082b08, 0x082b080808082b2b, + 0x082b080808190819, 0x082b080808191908, 0x082b0808082b0808, + 0x082b080819080819, 0x082b080819081908, 0x082b080819190808, + 0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, + 0x082b081908081908, 0x082b081908190808, 0x082b081919080808, + 0x082b081919082b08, 0x082b0819192b1919, 0x082b082b08080808, + 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08, + 0x082b190808080819, 0x082b190808081908, 0x082b190808190808, + 0x082b1908082b2b19, 0x082b190819080808, 0x082b191908080808, + 0x082b191919080819, 0x082b19191919082b, 0x082b19192b192b19, + 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b, + 0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, + 0x082b2b08082b0808, 0x082b2b0819191919, 0x082b2b082b082b08, + 0x082b2b082b2b082b, 0x082b2b19192b2b08, 0x082b2b192b190808, + 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b, + 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, + 0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, + 0x1908080808190808, 0x190808080819082b, 0x1908080808191919, + 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908, + 0x1908080819080808, 0x190808081908082b, 0x1908080819081919, + 0x1908080819082b08, 0x1908080819082b2b, 0x1908080819190819, + 0x1908080819191908, 0x19080808192b0808, 0x19080808192b1919, + 0x190808082b080819, 0x190808082b081908, 0x190808082b190808, + 0x1908081908080808, 0x190808190808082b, 0x1908081908081919, + 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908, + 0x19080819082b0808, 0x1908081919080819, 0x1908081919081908, + 0x1908081919190808, 0x190808192b080808, 0x190808192b081919, + 0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, + 0x1908082b08190808, 0x1908082b0819082b, 0x1908082b082b2b19, + 0x1908082b19080808, 0x1908190808080808, 0x190819080808082b, + 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819, + 0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, + 0x1908190819080819, 0x1908190819081908, 0x1908190819190808, + 0x190819082b080808, 0x190819082b191908, 0x1908191908080819, + 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908, + 0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, + 0x1908192b08082b2b, 0x1908192b19081908, 0x1908192b19190808, + 0x19082b0808080819, 0x19082b0808081908, 0x19082b0808190808, + 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908, + 0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, + 0x19082b1919081908, 0x19082b1919190808, 0x19082b19192b2b19, + 0x19082b2b08081908, 0x1919080808080808, 0x191908080808082b, + 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, + 0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, + 0x1919080819080819, 0x1919080819081908, 0x1919080819190808, + 0x191908082b080808, 0x1919081908080819, 0x1919081908081908, + 0x1919081908190808, 0x1919081908191919, 0x1919081919080808, + 0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, + 0x1919082b2b2b2b2b, 0x1919190808080819, 0x1919190808081908, + 0x1919190808190808, 0x19191908082b0819, 0x1919190819080808, + 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819, + 0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, + 0x191919192b082b08, 0x1919192b082b0819, 0x1919192b192b2b08, + 0x1919192b2b2b0819, 0x19192b0808080808, 0x19192b0808191908, + 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19, + 0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, + 0x19192b2b2b081919, 0x192b080808080819, 0x192b080808081908, + 0x192b080808190808, 0x192b080819080808, 0x192b080819191908, + 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19, + 0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, + 0x192b082b2b19082b, 0x192b190808080808, 0x192b19080819192b, + 0x192b191908190808, 0x192b191919080808, 0x192b191919081919, + 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b, + 0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, + 0x192b2b2b192b082b, 0x2b08080808080808, 0x2b0808080808082b, + 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819, + 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b, + 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, + 0x2b0808082b080808, 0x2b0808082b08082b, 0x2b0808082b2b2b08, + 0x2b0808082b2b2b2b, 0x2b08081908080819, 0x2b08081908081908, + 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808, + 0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, + 0x2b08082b082b0808, 0x2b08082b2b080808, 0x2b08082b2b08082b, + 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08, 0x2b08190808080819, + 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b, + 0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, + 0x2b0819082b082b19, 0x2b08191908080808, 0x2b08191919081908, + 0x2b0819192b2b1919, 0x2b08192b08192b08, 0x2b08192b192b2b2b, + 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919, + 0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, + 0x2b082b082b2b2b08, 0x2b082b190808192b, 0x2b082b2b082b082b, + 0x2b082b2b2b080808, 0x2b082b2b2b082b08, 0x2b082b2b2b19192b, + 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908, + 0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, + 0x2b1908082b081908, 0x2b19081908080808, 0x2b190819082b082b, + 0x2b190819192b1908, 0x2b19082b1919192b, 0x2b19082b2b082b19, + 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908, + 0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, + 0x2b1919192b190808, 0x2b1919192b19082b, 0x2b19192b19080819, + 0x2b192b0819190819, 0x2b192b082b2b192b, 0x2b192b1919082b19, + 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808, + 0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, + 0x2b2b0808082b0808, 0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, + 0x2b2b081919190819, 0x2b2b081919192b19, 0x2b2b08192b2b192b, + 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08, + 0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, + 0x2b2b190819080808, 0x2b2b19082b191919, 0x2b2b192b192b1919, + 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b, 0x2b2b2b08082b0808, + 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808, + 0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, + 0x2b2b2b192b08192b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, + 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b, + }); + +static dpct::global_memory iq3xxs_grid( + sycl::range<1>(256), + { + 0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, + 0x04041404, 0x04041414, 0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, + 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14, 0x040c140c, 0x040c142c, + 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404, + 0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, + 0x04141c1c, 0x04141c3e, 0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, + 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c, 0x041c3e04, 0x04240c1c, + 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c, + 0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, + 0x043e0c24, 0x043e0c34, 0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, + 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c, 0x0c041c04, 0x0c041c14, + 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c, + 0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, + 0x0c14140c, 0x0c141c04, 0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, + 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c, 0x0c24042c, 0x0c242c04, + 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414, + 0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, + 0x14041414, 0x14041434, 0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, + 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c, 0x140c1c04, 0x140c341c, + 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e, + 0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, + 0x141c0c04, 0x141c0c24, 0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, + 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24, 0x143e040c, 0x143e041c, + 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c, + 0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, + 0x1c0c1404, 0x1c0c1c0c, 0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, + 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14, 0x1c1c0c0c, 0x1c1c1c1c, + 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414, + 0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, + 0x24040424, 0x24040c3e, 0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, + 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404, 0x24143404, 0x24143434, + 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c, + 0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, + 0x2c040c14, 0x2c04240c, 0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, + 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14, 0x2c1c0414, 0x2c1c2c1c, + 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c, + 0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, + 0x34043424, 0x340c140c, 0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, + 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14, 0x34341c1c, 0x343e041c, + 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14, + 0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, + 0x3e1c0404, 0x3e1c0c2c, 0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, + 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04, + }); + +static dpct::global_memory ksigns_iq2xs( + sycl::range<1>(128), + { + 0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, + 141, 142, 15, 144, 17, 18, 147, 20, 149, 150, 23, 24, 153, + 154, 27, 156, 29, 30, 159, 160, 33, 34, 163, 36, 165, 166, + 39, 40, 169, 170, 43, 172, 45, 46, 175, 48, 177, 178, 51, + 180, 53, 54, 183, 184, 57, 58, 187, 60, 189, 190, 63, 192, + 65, 66, 195, 68, 197, 198, 71, 72, 201, 202, 75, 204, 77, + 78, 207, 80, 209, 210, 83, 212, 85, 86, 215, 216, 89, 90, + 219, 92, 221, 222, 95, 96, 225, 226, 99, 228, 101, 102, 231, + 232, 105, 106, 235, 108, 237, 238, 111, 240, 113, 114, 243, 116, + 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255, + }); + +static dpct::global_memory + ksigns64(sycl::range<1>(128), + { + 0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, + 0x000000000000ffff, 0xff00000000ff0000, 0x0000000000ff00ff, + 0x0000000000ffff00, 0xff00000000ffffff, 0xff000000ff000000, + 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff, + 0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, + 0x00000000ffffffff, 0xff0000ff00000000, 0x000000ff000000ff, + 0x000000ff0000ff00, 0xff0000ff0000ffff, 0x000000ff00ff0000, + 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff, + 0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, + 0x000000ffff00ffff, 0xff0000ffffff0000, 0x000000ffffff00ff, + 0x000000ffffffff00, 0xff0000ffffffffff, 0xff00ff0000000000, + 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff, + 0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, + 0x0000ff0000ffffff, 0x0000ff00ff000000, 0xff00ff00ff0000ff, + 0xff00ff00ff00ff00, 0x0000ff00ff00ffff, 0xff00ff00ffff0000, + 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff, + 0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, + 0x0000ffff0000ffff, 0xff00ffff00ff0000, 0x0000ffff00ff00ff, + 0x0000ffff00ffff00, 0xff00ffff00ffffff, 0xff00ffffff000000, + 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff, + 0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, + 0x0000ffffffffffff, 0xffff000000000000, 0x00ff0000000000ff, + 0x00ff00000000ff00, 0xffff00000000ffff, 0x00ff000000ff0000, + 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff, + 0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, + 0x00ff0000ff00ffff, 0xffff0000ffff0000, 0x00ff0000ffff00ff, + 0x00ff0000ffffff00, 0xffff0000ffffffff, 0x00ff00ff00000000, + 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff, + 0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, + 0xffff00ff00ffffff, 0xffff00ffff000000, 0x00ff00ffff0000ff, + 0x00ff00ffff00ff00, 0xffff00ffff00ffff, 0x00ff00ffffff0000, + 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff, + 0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, + 0x00ffff000000ffff, 0xffffff0000ff0000, 0x00ffff0000ff00ff, + 0x00ffff0000ffff00, 0xffffff0000ffffff, 0xffffff00ff000000, + 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff, + 0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, + 0x00ffff00ffffffff, 0xffffffff00000000, 0x00ffffff000000ff, + 0x00ffffff0000ff00, 0xffffffff0000ffff, 0x00ffffff00ff0000, + 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff, + 0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, + 0x00ffffffff00ffff, 0xffffffffffff0000, 0x00ffffffffff00ff, + 0x00ffffffffffff00, 0xffffffffffffffff, + }); +//#endif + +static dpct::global_memory + kmask_iq2xs(sycl::range<1>(8), {1, 2, 4, 8, 16, 32, 64, 128}); + +template +static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1, + const uint64_t *iq2xxs_grid_ptr, + const uint8_t *ksigns_iq2xs_ptr, + const uint8_t *kmask_iq2xs_ptr) { + + const int i = item_ct1.get_group(2); + const block_iq2_xxs * x = (const block_iq2_xxs *) vx; + + const int tid = item_ct1.get_local_id(2); +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint16_t * q2 = x[i].qs + 4*ib; + const uint8_t * aux8 = (const uint8_t *)q2; + const uint8_t * grid = (const uint8_t *)(iq2xxs_grid_ptr + aux8[il]); + const uint32_t aux32 = q2[2] | (q2[3] << 16); + const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f; + const uint8_t signs = ksigns_iq2xs_ptr[(aux32 >> 7*il) & 127]; + for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs_ptr[j] ? -1.f : 1.f); +#else + assert(false); +#endif + +} + +template +static void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1, + const uint64_t *iq2xs_grid, + const uint8_t *ksigns_iq2xs, + const uint8_t *kmask_iq2xs) { + + const int i = item_ct1.get_group(2); + const block_iq2_xs * x = (const block_iq2_xs *) vx; + + const int tid = item_ct1.get_local_id(2); +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint16_t * q2 = x[i].qs + 4*ib; + const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511)); + const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; + const uint8_t signs = ksigns_iq2xs[q2[il] >> 9]; + for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); +#else + assert(false); +#endif + +} + +template +static void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1, + const uint32_t *iq3xxs_grid, + const uint8_t *ksigns_iq2xs, + const uint8_t *kmask_iq2xs) { + + const int i = item_ct1.get_group(2); + const block_iq3_xxs * x = (const block_iq3_xxs *) vx; + + const int tid = item_ct1.get_local_id(2); +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint8_t * q3 = x[i].qs + 8*ib; + const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib; + const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]); + const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]); + const uint32_t aux32 = gas[0] | (gas[1] << 16); + const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f; + const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127]; + for (int j = 0; j < 4; ++j) { + y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); + y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); + } +#else + assert(false); +#endif + +} + /* DPCT1110:4: The total declared local variable size in device function dequantize_mul_mat_vec_q2_k exceeds 128 bytes and may cause high register @@ -4260,13 +5291,16 @@ static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx, } #else - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 + const int tid = item_ct1.get_local_id(2) / + (2 * K_QUANTS_PER_ITERATION); // 0...15 or 0...7 + const int ix = item_ct1.get_local_id(2) % + (2 * K_QUANTS_PER_ITERATION); // 0....1 or 0...3 const int offset = tid * K_QUANTS_PER_ITERATION; uint32_t uaux[2]; const uint8_t * d = (const uint8_t *)uaux; + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { const float * y = yy + i * QK_K + offset; @@ -4276,7 +5310,8 @@ static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx, uaux[0] = s[0] & 0x0f0f0f0f; uaux[1] = (s[0] >> 4) & 0x0f0f0f0f; - const float2 dall = __half22float2(x[i].dm); + const sycl::float2 dall = + x[i].dm.convert(); float sum1 = 0, sum2 = 0; for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { @@ -4287,8 +5322,9 @@ static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx, + y[l+48] * d[3] * ((ql >> 6) & 3); sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7]; } - tmp += dall.x * sum1 - dall.y * sum2; + tmp += dall.x() * sum1 - dall.y() * sum2; } + #endif // sum up partial sums and write back result @@ -4383,8 +5419,8 @@ static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx, } #else - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 + const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 + const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 const int offset = tid * K_QUANTS_PER_ITERATION; // 0...15 or 0...14 const int in = offset/8; // 0 or 1 const int im = offset%8; // 0...7 @@ -4533,8 +5569,8 @@ static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx, } #else - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); + const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...15 + const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); const int step = tid * K_QUANTS_PER_ITERATION; @@ -4674,8 +5710,8 @@ static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx, } #else - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); + const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...15 + const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); const int step = tid * K_QUANTS_PER_ITERATION; const int im = step/8; const int in = step%8; @@ -4783,8 +5819,8 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa #else - const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...7 - const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0...3 + const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...7 + const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); // 0...3 const int step = tid * K_QUANTS_PER_ITERATION; @@ -4964,8 +6000,8 @@ static void k_get_rows_float( template static void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, const sycl::nd_item<3> &item_ct1) { - const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - 2 * item_ct1.get_local_id(2); + const int i = 2 * (item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2)); if (i >= k) { return; @@ -4984,6 +6020,21 @@ static void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y[iybs + iqs + y_offset] = v.y(); } +template +static void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i >= k) { + return; + } + + const src_t * x = (src_t *) vx; + + y[i] = x[i]; +} + // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q @@ -6402,8 +7453,8 @@ vec_dot_q4_K_q8_1(const void *__restrict__ vbq, const float dall = bq4_K->dm[0]; const float dmin = bq4_K->dm[1]; - const float d8_1 = __low2float(bq8_1[0].ds); - const float d8_2 = __low2float(bq8_1[1].ds); + const float d8_1 = bq8_1[0].ds[0]; + const float d8_2 = bq8_1[1].ds[1]; const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); @@ -6414,10 +7465,10 @@ vec_dot_q4_K_q8_1(const void *__restrict__ vbq, const int v1 = q4[0]; const int v2 = q4[4]; - const int dot1 = __dp4a(ui2, v2 & 0x0f0f0f0f, __dp4a(ui1, v1 & 0x0f0f0f0f, 0)); - const int dot2 = __dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, __dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0)); - const int dot3 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0)); - const int dot4 = __dp4a(0x01010101, ui4, __dp4a(0x01010101, ui3, 0)); + const int dot1 = dpct::dp4a(ui2, v2 & 0x0f0f0f0f, dpct::dp4a(ui1, v1 & 0x0f0f0f0f, 0)); + const int dot2 = dpct::dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, dpct::dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0)); + const int dot3 = dpct::dp4a(0x01010101, ui2, dpct::dp4a(0x01010101, ui1, 0)); + const int dot4 = dpct::dp4a(0x01010101, ui4, dpct::dp4a(0x01010101, ui3, 0)); sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]); sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]); @@ -6586,8 +7637,8 @@ vec_dot_q5_K_q8_1(const void *__restrict__ vbq, const float d = bq5_K->d; - const float d8_1 = __low2half(bq8_1[0].ds); - const float d8_2 = __low2half(bq8_1[1].ds); + const float d8_1 = bq8_1[0].ds[0]; + const float d8_2 = bq8_1[1].ds[1]; const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); @@ -6608,8 +7659,8 @@ vec_dot_q5_K_q8_1(const void *__restrict__ vbq, const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f); const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f); - const float sumf_d = d8_1 * (__dp4a(ui1, v1, 0) * s[0] + __dp4a(ui2, v2, 0) * s[1]) - + d8_2 * (__dp4a(ui3, v3, 0) * s[2] + __dp4a(ui4, v4, 0) * s[3]); + const float sumf_d = d8_1 * (dpct::dp4a(ui1, v1, 0) * s[0] + dpct::dp4a(ui2, v2, 0) * s[1]) + + d8_2 * (dpct::dp4a(ui3, v3, 0) * s[2] + dpct::dp4a(ui4, v4, 0) * s[3]); return d * sumf_d; @@ -6865,6 +7916,150 @@ static __dpct_inline__ float vec_dot_q6_K_q8_1_mul_mat( return vec_dot_q6_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, x_dmf[i * (WARP_SIZE/QI6_K) + i/QI6_K], &y_df[index_y/QI8_1]); } + +static __dpct_inline__ float +vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs, + const uint64_t *iq2xxs_grid, const uint8_t *ksigns_iq2xs, + const uint8_t *kmask_iq2xs) { +#if QK_K == 256 + const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq; + +#if QR2_XXS == 8 + const int ib32 = iqs; + const uint16_t * q2 = bq2->qs + 4*ib32; + const uint8_t * aux8 = (const uint8_t *)q2; + const int8_t * q8 = bq8_1[ib32].qs; + uint32_t aux32 = q2[2] | (q2[3] << 16); + int sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); + const uint8_t signs = ksigns_iq2xs[aux32 & 127]; + for (int j = 0; j < 8; ++j) { + sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + aux32 >>= 7; + } + const float d = (float)bq2->d * (0.5f + aux32) * bq8_1[ib32].ds[0] * 0.25f; + return d * sumi; +#else + // iqs is 0...15 + const int ib32 = iqs/2; + const int il = iqs%2; + const uint16_t * q2 = bq2->qs + 4*ib32; + const uint8_t * aux8 = (const uint8_t *)q2; + const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]); + const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]); + const uint32_t aux32 = q2[2] | (q2[3] << 16); + const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * bq8_1[ib32].ds[0] * 0.25f; + const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127]; + const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127]; + const int8_t * q8 = bq8_1[ib32].qs + 16*il; + int sumi1 = 0, sumi2 = 0; + for (int j = 0; j < 8; ++j) { + sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1); + sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1); + } + return d * (sumi1 + sumi2); +#endif +#else + assert(false); + return 0.f; +#endif +} + +static __dpct_inline__ float +vec_dot_iq2_xs_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs, + const uint64_t *iq2xs_grid, const uint64_t *ksigns64) { +#if DPCT_COMPATIBILITY_TEMP >= \ + MIN_CC_DP4A // lowest compute capability for integer intrinsics +#if QK_K == 256 + const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq; + + const int ib32 = iqs; + const uint16_t * q2 = bq2->qs + 4*ib32; + const int8_t * q8 = bq8_1[ib32].qs; + const uint8_t ls1 = bq2->scales[ib32] & 0xf; + const uint8_t ls2 = bq2->scales[ib32] >> 4; + int sumi1 = 0; + for (int l = 0; l < 2; ++l) { + const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); + const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); + const int grid_l = dpct::vectorized_binary( + grid[0] ^ signs[0], signs[0], std::minus<>()); + const int grid_h = dpct::vectorized_binary( + grid[1] ^ signs[1], signs[1], std::minus<>()); + sumi1 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi1); + sumi1 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi1); + q8 += 8; + } + int sumi2 = 0; + for (int l = 2; l < 4; ++l) { + const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); + const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); + const int grid_l = dpct::vectorized_binary( + grid[0] ^ signs[0], signs[0], std::minus<>()); + const int grid_h = dpct::vectorized_binary( + grid[1] ^ signs[1], signs[1], std::minus<>()); + sumi2 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi2); + sumi2 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi2); + q8 += 8; + } + const float d = (float)bq2->d * bq8_1[ib32].ds[0] * 0.25f; + return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); +#else + assert(false); + return 0.f; +#endif +#else + assert(false); + return 0.f; +#endif +} + +static __dpct_inline__ float +vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs, + const uint32_t *iq3xxs_grid, const uint64_t *ksigns64) { +#if DPCT_COMPATIBILITY_TEMP >= \ + MIN_CC_DP4A // lowest compute capability for integer intrinsics +#if QK_K == 256 + const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq; + + const int ib32 = iqs; + const uint8_t * q3 = bq2->qs + 8*ib32; + const uint16_t * gas = (const uint16_t *)(bq2->qs + QK_K/4) + 2*ib32; + const int8_t * q8 = bq8_1[ib32].qs; + uint32_t aux32 = gas[0] | (gas[1] << 16); + int sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint32_t * grid1 = iq3xxs_grid + q3[2*l+0]; + const uint32_t * grid2 = iq3xxs_grid + q3[2*l+1]; + const uint32_t * signs = (const uint32_t *)(ksigns64 + (aux32 & 127)); + const int grid_l = dpct::vectorized_binary( + grid1[0] ^ signs[0], signs[0], std::minus<>()); + const int grid_h = dpct::vectorized_binary( + grid2[0] ^ signs[1], signs[1], std::minus<>()); + sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi); + sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi); + q8 += 8; + aux32 >>= 7; + } + const float d = (float)bq2->d * (0.5f + aux32) * bq8_1[ib32].ds[0] * 0.5f; + return d * sumi; +#else + assert(false); + return 0.f; +#endif +#else + assert(false); + return 0.f; +#endif +} + + template @@ -7446,7 +8641,8 @@ template static void template static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows, - const sycl::nd_item<3> &item_ct1) { + const sycl::nd_item<3> &item_ct1, + const uint32_t *iq3xxs_grid_ptr, const uint64_t *ksigns64_ptr) { const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1); @@ -7463,12 +8659,11 @@ static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict_ const block_q_t * x = (const block_q_t *) vx; const block_q8_1 * y = (const block_q8_1 *) vy; - for (int i = 0; i < blocks_per_row; i += blocks_per_warp) { - const int ibx = row * blocks_per_row + i + - item_ct1.get_local_id(2) / (qi / vdr); // x block index + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index - const int iby = (i + item_ct1.get_local_id(2) / (qi / vdr)) * - (qk / QK8_1); // y block index that aligns with ibx + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx const int iqs = vdr * @@ -7490,6 +8685,145 @@ static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict_ } } +template +static void mul_mat_vec_q_iq2_xxs_q8_1(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows, + const sycl::nd_item<3> &item_ct1, + const uint64_t *iq2xxs_grid_ptr, const uint8_t *ksigns_iq2xs_ptr, + const uint8_t *kmask_iq2xs_ptr ) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_iq2_xxs_q8_1(&x[ibx], &y[iby], iqs, iq2xxs_grid_ptr, ksigns_iq2xs_ptr, kmask_iq2xs_ptr); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +template +static void mul_mat_vec_q_iq2_xs_q8_1(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows, + const sycl::nd_item<3> &item_ct1, + const uint64_t *iq2xs_grid_ptr, const uint64_t *ksigns64_ptr ) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_iq2_xs_q8_1(&x[ibx], &y[iby], iqs, iq2xs_grid_ptr, ksigns64_ptr); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +template +static void mul_mat_vec_q_iq3_xxs_q8_1(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows, + const sycl::nd_item<3> &item_ct1, + const uint32_t *iq3xxs_grid_ptr, const uint64_t *ksigns64_ptr ) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_iq3_xxs_q8_1(&x[ibx], &y[iby], iqs, iq3xxs_grid_ptr, ksigns64_ptr); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + template static void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows, const sycl::nd_item<3> &item_ct1) { @@ -7695,7 +9029,7 @@ static void cpy_1_f16_f16(const char * cxi, char * cdsti) { static void cpy_1_f16_f32(const char * cxi, char * cdsti) { const sycl::half *xi = (const sycl::half *)cxi; - float *dsti = (float *)cdsti; + float * dsti = (float *) cdsti; *dsti = *xi; } @@ -8297,6 +9631,62 @@ static void im2col_kernel(const float *x, T *dst, int offset_delta, } } +template +static void pool2d_nchw_kernel( + const int ih, const int iw, const int oh, const int ow, + const int kh, const int kw, const int sh, const int sw, + const int ph, const int pw, const int parallel_elements, + const Ti* src, To* dst, const enum ggml_op_pool op, + const sycl::nd_item<3> &item_ct1) { + int idx = item_ct1.get_local_id(2) + + item_ct1.get_group(2) * item_ct1.get_local_range(2); + if (idx >= parallel_elements) { + return; + } + + const int I_HW = ih * iw; + const int O_HW = oh * ow; + const int nc = idx / O_HW; + const int cur_oh = idx % O_HW / ow; + const int cur_ow = idx % O_HW % ow; + const Ti* i_ptr = src + nc * I_HW; + To* o_ptr = dst + nc * O_HW; + const int start_h = cur_oh * sh - ph; + const int bh = sycl::max(0, start_h); + const int eh = sycl::min(ih, start_h + kh); + const int start_w = cur_ow * sw - pw; + const int bw = sycl::max(0, start_w); + const int ew = sycl::min(iw, start_w + kw); + + To res = 0; + + switch (op) { + case GGML_OP_POOL_AVG: res = 0; break; + case GGML_OP_POOL_MAX: res = -FLT_MAX; break; + } + + for (int i = bh; i < eh; i += 1) { + for (int j = bw; j < ew; j += 1) { +#if DPCT_COMPATIBILITY_TEMP >= 350 + /* + DPCT1098:106: The '*' expression is used instead of the __ldg + call. These two expressions do not provide the exact same + functionality. Check the generated code for potential precision + and/or performance issues. + */ + Ti cur = *(i_ptr + i * iw + j); +#else + Ti cur = i_ptr[i * iw + j]; +#endif + switch (op) { + case GGML_OP_POOL_AVG: res += (cur / (kh * kw)); break; + case GGML_OP_POOL_MAX: res = sycl::max(res, (To)cur); break; + } + } + } + o_ptr[cur_oh * ow + cur_ow] = res; +} + template static void get_rows_sycl(const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, const void *src0_dd, @@ -8585,6 +9975,30 @@ static void relu_f32_sycl(const float *x, float *dst, const int k, }); } +static void hardsigmoid_f32_sycl(const float *x, float *dst, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_HARDSIGMOID_BLOCK_SIZE - 1) / SYCL_HARDSIGMOID_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + hardsigmoid_f32(x, dst, k, item_ct1); + }); +} + +static void hardswish_f32_sycl(const float *x, float *dst, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_HARDSWISH_BLOCK_SIZE - 1) / SYCL_HARDSWISH_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + hardswish_f32(x, dst, k, item_ct1); + }); +} + static void leaky_relu_f32_sycl(const float *x, float *dst, const int k, const float negative_slope, dpct::queue_ptr stream) { @@ -8811,11 +10225,10 @@ template static void dequantize_block_sycl(const void *__restrict__ vx, dst_t *__restrict__ y, const int k, dpct::queue_ptr stream) { - const int num_blocks = (k + SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / SYCL_DEQUANTIZE_BLOCK_SIZE; + const int num_blocks = (k + 2*SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / (2*SYCL_DEQUANTIZE_BLOCK_SIZE); { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( sycl::nd_range<3>( sycl::range<3>(1, 1, num_blocks) * @@ -8844,7 +10257,18 @@ static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int k, }); } #else - dequantize_block_q2_K<<>>(vx, y); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 32), + sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q2_K(vx, y, item_ct1); + }); + } + #endif } @@ -8865,10 +10289,57 @@ static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int k, }); } #else - dequantize_block_q3_K<<>>(vx, y); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 32), + sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q3_K(vx, y, item_ct1); + }); + } #endif } +template +static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb32 = k / 32; + const int nb = (k + 255) / 256; + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 32), + sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q4_0(vx, y, nb32, item_ct1); + }); + } +} + +template +static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb32 = k / 32; + const int nb = (k + 255) / 256; + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 32), + sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q4_1(vx, y, nb32, item_ct1); + }); + } +} + + template static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int k, dpct::queue_ptr stream) { @@ -8903,7 +10374,18 @@ static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int k, }); } #else - dequantize_block_q5_K<<>>(vx, y); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 32), + sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q5_K(vx, y, item_ct1); + }); + } + #endif } @@ -8924,11 +10406,132 @@ static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int k, }); } #else - dequantize_block_q6_K<<>>(vx, y); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 32), + sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q6_K(vx, y, item_ct1); + }); + } + #endif } -static to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type) { + +template +static void dequantize_row_iq2_xxs_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_K; + { + iq2xxs_grid.init(*stream); + ksigns_iq2xs.init(*stream); + kmask_iq2xs.init(*stream); + + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + auto iq2xxs_grid_ptr_ct1 = iq2xxs_grid.get_ptr(); + auto ksigns_iq2xs_ptr_ct1 = ksigns_iq2xs.get_ptr(); + auto kmask_iq2xs_ptr_ct1 = kmask_iq2xs.get_ptr(); + + cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 32), + sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_iq2_xxs( + vx, y, item_ct1, iq2xxs_grid_ptr_ct1, + ksigns_iq2xs_ptr_ct1, kmask_iq2xs_ptr_ct1); + }); + }); + } +} + +template +static void dequantize_row_iq2_xs_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_K; + { + iq2xs_grid.init(*stream); + ksigns_iq2xs.init(*stream); + kmask_iq2xs.init(*stream); + + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + auto iq2xs_grid_ptr_ct1 = iq2xs_grid.get_ptr(); + auto ksigns_iq2xs_ptr_ct1 = ksigns_iq2xs.get_ptr(); + auto kmask_iq2xs_ptr_ct1 = kmask_iq2xs.get_ptr(); + + cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 32), + sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_iq2_xs( + vx, y, item_ct1, iq2xs_grid_ptr_ct1, + ksigns_iq2xs_ptr_ct1, kmask_iq2xs_ptr_ct1); + }); + }); + } +} + +template +static void dequantize_row_iq3_xxs_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_K; + { + iq3xxs_grid.init(*stream); + ksigns_iq2xs.init(*stream); + kmask_iq2xs.init(*stream); + + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr(); + auto ksigns_iq2xs_ptr_ct1 = ksigns_iq2xs.get_ptr(); + auto kmask_iq2xs_ptr_ct1 = kmask_iq2xs.get_ptr(); + + cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 32), + sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_iq3_xxs( + vx, y, item_ct1, iq3xxs_grid_ptr_ct1, + ksigns_iq2xs_ptr_ct1, kmask_iq2xs_ptr_ct1); + }); + }); + } +} + +template +static void convert_unary_sycl(const void *__restrict__ vx, + dst_t *__restrict__ y, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / SYCL_DEQUANTIZE_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>( + sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + convert_unary(vx, y, k, item_ct1); + }); + } +} + + +static to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type) try { + int id; switch (type) { case GGML_TYPE_Q4_0: return dequantize_block_sycl; @@ -8950,19 +10553,30 @@ static to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type) { return dequantize_row_q5_K_sycl; case GGML_TYPE_Q6_K: return dequantize_row_q6_K_sycl; + case GGML_TYPE_IQ2_XXS: + return dequantize_row_iq2_xxs_sycl; + case GGML_TYPE_IQ2_XS: + return dequantize_row_iq2_xs_sycl; + case GGML_TYPE_IQ3_XXS: + return dequantize_row_iq3_xxs_sycl; case GGML_TYPE_F32: - return dequantize_block_sycl<1, 1, convert_f32>; + return convert_unary_sycl; default: return nullptr; } } +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} static to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type) { switch (type) { case GGML_TYPE_Q4_0: - return dequantize_block_sycl; + return dequantize_row_q4_0_sycl; case GGML_TYPE_Q4_1: - return dequantize_block_sycl; + return dequantize_row_q4_1_sycl; case GGML_TYPE_Q5_0: return dequantize_block_sycl; case GGML_TYPE_Q5_1: @@ -8979,8 +10593,14 @@ static to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type) { return dequantize_row_q5_K_sycl; case GGML_TYPE_Q6_K: return dequantize_row_q6_K_sycl; + case GGML_TYPE_IQ2_XXS: + return dequantize_row_iq2_xxs_sycl; + case GGML_TYPE_IQ2_XS: + return dequantize_row_iq2_xs_sycl; + case GGML_TYPE_IQ3_XXS: + return dequantize_row_iq3_xxs_sycl; case GGML_TYPE_F16: - return dequantize_block_sycl<1, 1, convert_f16>; + return convert_unary_sycl; default: return nullptr; } @@ -9190,41 +10810,384 @@ static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y, } } -template -static void mul_mat_vec_q_sycl_submitter(const void *vx, const void *vy, + +static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK4_0 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq3xxs_grid.init(*stream); + ksigns64.init(*stream); + + stream->submit([&](sycl::handler &cgh) { + auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr(); + auto ksigns64_ptr_ct1 = ksigns64.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1, + iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1); + }); + }); + } +} + +static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK4_1 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq3xxs_grid.init(*stream); + ksigns64.init(*stream); + + stream->submit([&](sycl::handler &cgh) { + auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr(); + auto ksigns64_ptr_ct1 = ksigns64.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1, + iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1); + }); + }); + } +} + +static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK5_0 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq3xxs_grid.init(*stream); + ksigns64.init(*stream); + + stream->submit([&](sycl::handler &cgh) { + auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr(); + auto ksigns64_ptr_ct1 = ksigns64.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1, + iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1); + }); + }); + } +} + +static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK5_1 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq3xxs_grid.init(*stream); + ksigns64.init(*stream); + + stream->submit([&](sycl::handler &cgh) { + auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr(); + auto ksigns64_ptr_ct1 = ksigns64.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1, + iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1); + }); + }); + } +} + +static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK8_0 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq3xxs_grid.init(*stream); + ksigns64.init(*stream); + + stream->submit([&](sycl::handler &cgh) { + auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr(); + auto ksigns64_ptr_ct1 = ksigns64.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1, + iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1); + }); + }); + } +} + +static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq3xxs_grid.init(*stream); + ksigns64.init(*stream); + + stream->submit([&](sycl::handler &cgh) { + auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr(); + auto ksigns64_ptr_ct1 = ksigns64.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1, + iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1); + }); + }); + } +} + +static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq3xxs_grid.init(*stream); + ksigns64.init(*stream); + + stream->submit([&](sycl::handler &cgh) { + auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr(); + auto ksigns64_ptr_ct1 = ksigns64.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1, + iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1); + }); + }); + } +} + +static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq3xxs_grid.init(*stream); + ksigns64.init(*stream); + + stream->submit([&](sycl::handler &cgh) { + auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr(); + auto ksigns64_ptr_ct1 = ksigns64.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1, + iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1); + }); + }); + } +} + +static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq3xxs_grid.init(*stream); + ksigns64.init(*stream); + + stream->submit([&](sycl::handler &cgh) { + auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr(); + auto ksigns64_ptr_ct1 = ksigns64.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1, + iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1); + }); + }); + } +} + +static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq3xxs_grid.init(*stream); + ksigns64.init(*stream); + + stream->submit([&](sycl::handler &cgh) { + auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr(); + auto ksigns64_ptr_ct1 = ksigns64.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1, + iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1); + }); + }); + } +} + +static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq2xxs_grid.init(*stream); + ksigns_iq2xs.init(*stream); + kmask_iq2xs.init(*stream); + + + stream->submit([&](sycl::handler &cgh) { + auto iq2xxs_grid_ptr_ct1 = iq2xxs_grid.get_ptr(); + auto ksigns_iq2xs_ptr_ct1 = ksigns_iq2xs.get_ptr(); + auto kmask_iq2xs_ptr_ct1 = kmask_iq2xs.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q_iq2_xxs_q8_1( + vx, vy, dst, ncols, nrows, item_ct1, + iq2xxs_grid_ptr_ct1, ksigns_iq2xs_ptr_ct1, kmask_iq2xs_ptr_ct1); + }); + }); + } +} + +static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy, float *dst, const int ncols, const int nrows, dpct::queue_ptr stream) { - GGML_ASSERT(ncols % QK4_0 == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), [= - ](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq2xs_grid.init(*stream); + ksigns64.init(*stream); + + stream->submit([&](sycl::handler &cgh) { + auto iq2xs_grid_ptr_ct1 = iq2xs_grid.get_ptr(); + auto ksigns64_ptr_ct1 = ksigns64.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q_iq2_xs_q8_1( + vx, vy, dst, ncols, nrows, item_ct1, + iq2xs_grid_ptr_ct1, ksigns64_ptr_ct1); + }); + }); + } } -int get_device_index_by_id(int id){ - int res = g_sycl_device_id2index[id].index; - // GGML_SYCL_DEBUG("get_device_index_by_id id=%d device_index=%d\n", id, res); - GGML_ASSERT(res>=0); - return res; +static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + iq3xxs_grid.init(*stream); + ksigns64.init(*stream); + + stream->submit([&](sycl::handler &cgh) { + auto iq3xxs_grid_ptr_ct1 = iq3xxs_grid.get_ptr(); + auto ksigns64_ptr_ct1 = ksigns64.get_ptr(); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q_iq3_xxs_q8_1( + vx, vy, dst, ncols, nrows, item_ct1, + iq3xxs_grid_ptr_ct1, ksigns64_ptr_ct1); + }); + }); + } } -int get_device_id_by_index(int index){ - int res = g_device_caps[index].device_id; - GGML_ASSERT(res>=0); - return res; -} - - -int get_current_device_index(){ - return get_device_index_by_id(dpct::dev_mgr::instance().current_device_id()); -} static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, float *dst, const int ncols_x, @@ -9234,7 +11197,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, int id; SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); + CHECK_TRY_ERROR(id = get_current_device_id())); const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; @@ -9349,7 +11312,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, int id; SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); + CHECK_TRY_ERROR(id = get_current_device_id())); const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; @@ -9464,7 +11427,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, int id; SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); + CHECK_TRY_ERROR(id = get_current_device_id())); const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; @@ -9579,7 +11542,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, int id; SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); + CHECK_TRY_ERROR(id = get_current_device_id())); const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; @@ -9694,7 +11657,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, int id; SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); + CHECK_TRY_ERROR(id = get_current_device_id())); const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; @@ -9809,7 +11772,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, int id; SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); + CHECK_TRY_ERROR(id = get_current_device_id())); const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; @@ -9932,7 +11895,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, int id; SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); + CHECK_TRY_ERROR(id = get_current_device_id())); const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; @@ -10060,7 +12023,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, int id; SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); + CHECK_TRY_ERROR(id = get_current_device_id())); const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; @@ -10181,7 +12144,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, int id; SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); + CHECK_TRY_ERROR(id = get_current_device_id())); const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; @@ -10302,7 +12265,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, int id; SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); + CHECK_TRY_ERROR(id = get_current_device_id())); const int compute_capability = g_device_caps[id].cc; int mmq_x, mmq_y, nwarps; @@ -10458,6 +12421,31 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl( } } +static void +ggml_cpy_f16_f32_sycl(const char *cx, char *cdst, const int ne, const int ne00, + const int ne01, const int ne02, const int nb00, + const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, + const int nb10, const int nb11, const int nb12, + const int nb13, dpct::queue_ptr stream) { + + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, + nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + static void ggml_cpy_f32_f32_sycl(const char *cx, char *cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, @@ -11014,12 +13002,9 @@ struct sycl_buffer { static sycl_buffer g_sycl_buffer_pool[GGML_SYCL_MAX_DEVICES][MAX_SYCL_BUFFERS]; static size_t g_sycl_pool_size[GGML_SYCL_MAX_DEVICES] = {0}; -static void *ggml_sycl_pool_malloc_leg(size_t size, size_t *actual_size) try { +static void *ggml_sycl_pool_malloc_leg(int device_index, size_t size, size_t *actual_size) try { scoped_spin_lock lock(g_sycl_pool_lock); - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); - // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg index %d\n", id); + // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg device_index %d size=%lu\n", device_index, size); #ifdef DEBUG_SYCL_MALLOC int nnz = 0; size_t max_size = 0; @@ -11027,7 +13012,7 @@ static void *ggml_sycl_pool_malloc_leg(size_t size, size_t *actual_size) try { size_t best_diff = 1ull << 36; int ibest = -1; for (int i = 0; i < MAX_SYCL_BUFFERS; ++i) { - sycl_buffer& b = g_sycl_buffer_pool[id][i]; + sycl_buffer& b = g_sycl_buffer_pool[device_index][i]; if (b.ptr != nullptr) { #ifdef DEBUG_SYCL_MALLOC ++nnz; @@ -11043,7 +13028,7 @@ static void *ggml_sycl_pool_malloc_leg(size_t size, size_t *actual_size) try { *actual_size = b.size; b.ptr = nullptr; b.size = 0; - // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg return 1 %p\n", ptr); + // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg return 1 %p and rm in pool\n", ptr); return ptr; } } @@ -11051,30 +13036,30 @@ static void *ggml_sycl_pool_malloc_leg(size_t size, size_t *actual_size) try { } } if (ibest >= 0) { - sycl_buffer& b = g_sycl_buffer_pool[id][ibest]; + sycl_buffer& b = g_sycl_buffer_pool[device_index][ibest]; void * ptr = b.ptr; *actual_size = b.size; b.ptr = nullptr; b.size = 0; - // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg return 2 %p\n", ptr); + // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg return 2 %p and rm in pool\n", ptr); return ptr; } void * ptr; size_t look_ahead_size = (size_t) (1.05 * size); look_ahead_size = 256 * ((look_ahead_size + 255)/256); - const dpct::queue_ptr stream = g_syclStreams[id][0]; + const dpct::queue_ptr stream = g_syclStreams[device_index][0]; SYCL_CHECK( CHECK_TRY_ERROR(ptr = (void *)sycl::malloc_device( look_ahead_size, *stream))); *actual_size = look_ahead_size; - g_sycl_pool_size[id] += look_ahead_size; + g_sycl_pool_size[device_index] += look_ahead_size; #ifdef DEBUG_SYCL_MALLOC fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, id, nnz, (uint32_t)(max_size/1024/1024), (uint32_t)(g_sycl_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024)); #endif - // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg return %p\n", ptr); + // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%lu, return %p\n", look_ahead_size, ptr); return ptr; } catch (sycl::exception const &exc) { @@ -11083,15 +13068,11 @@ catch (sycl::exception const &exc) { std::exit(1); } -static void ggml_sycl_pool_free_leg(void *ptr, size_t size) try { +static void ggml_sycl_pool_free_leg(int device_index, void *ptr, size_t size) try { scoped_spin_lock lock(g_sycl_pool_lock); - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); - - const dpct::queue_ptr stream = g_syclStreams[id][0]; + const dpct::queue_ptr stream = g_syclStreams[device_index][0]; for (int i = 0; i < MAX_SYCL_BUFFERS; ++i) { - sycl_buffer& b = g_sycl_buffer_pool[id][i]; + sycl_buffer& b = g_sycl_buffer_pool[device_index][i]; if (b.ptr == nullptr) { b.ptr = ptr; b.size = size; @@ -11100,7 +13081,7 @@ static void ggml_sycl_pool_free_leg(void *ptr, size_t size) try { } fprintf(stderr, "WARNING: sycl buffer pool full, increase MAX_SYCL_BUFFERS\n"); SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, *stream))); - g_sycl_pool_size[id] -= size; + g_sycl_pool_size[device_index] -= size; } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ @@ -11117,7 +13098,8 @@ DPCT1082:64: Migration of CUmemGenericAllocationHandle type is not supported. static dpct::device_ptr g_sycl_pool_addr[GGML_SYCL_MAX_DEVICES] = {0}; static size_t g_sycl_pool_used[GGML_SYCL_MAX_DEVICES] = {0}; -static void *ggml_sycl_pool_malloc_vmm(size_t size, size_t *actual_size) try { +static void *ggml_sycl_pool_malloc_vmm(int device_index, size_t size, size_t *actual_size) try { + GGML_UNUSED(device_index); GGML_UNUSED(size); GGML_UNUSED(actual_size); return NULL; @@ -11128,20 +13110,16 @@ catch (sycl::exception const &exc) { std::exit(1); } -static void ggml_sycl_pool_free_vmm(void *ptr, size_t size) try { +static void ggml_sycl_pool_free_vmm(int device_index, void *ptr, size_t size) try { scoped_spin_lock lock(g_sycl_pool_lock); - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = dpct::dev_mgr::instance().current_device_id())); - #ifdef DEBUG_SYCL_MALLOC - printf("sycl pool[%d]: freed %llu bytes at %llx\n", id, (unsigned long long) size, ptr); + printf("sycl pool[%d]: freed %llu bytes at %llx\n", device_index, (unsigned long long) size, ptr); #endif - g_sycl_pool_used[id] -= size; + g_sycl_pool_used[device_index] -= size; // all deallocations must be in reverse order of the allocations - GGML_ASSERT(ptr == (void *) (g_sycl_pool_addr[id] + g_sycl_pool_used[id])); + GGML_ASSERT(ptr == (void *) (g_sycl_pool_addr[device_index] + g_sycl_pool_used[device_index])); } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ @@ -11149,14 +13127,11 @@ catch (sycl::exception const &exc) { std::exit(1); } -static void *ggml_sycl_pool_malloc(size_t size, size_t *actual_size) try { - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); - if (g_device_caps[id].vmm) { - return ggml_sycl_pool_malloc_vmm(size, actual_size); +static void *ggml_sycl_pool_malloc(int device_index, size_t size, size_t *actual_size) try { + if (g_device_caps[device_index].vmm) { + return ggml_sycl_pool_malloc_vmm(device_index, size, actual_size); } else { - return ggml_sycl_pool_malloc_leg(size, actual_size); + return ggml_sycl_pool_malloc_leg(device_index, size, actual_size); } } catch (sycl::exception const &exc) { @@ -11165,14 +13140,11 @@ catch (sycl::exception const &exc) { std::exit(1); } -static void ggml_sycl_pool_free(void *ptr, size_t size) try { - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); - if (g_device_caps[id].vmm) { - ggml_sycl_pool_free_vmm(ptr, size); +static void ggml_sycl_pool_free(int device_index, void *ptr, size_t size) try { + if (g_device_caps[device_index].vmm) { + ggml_sycl_pool_free_vmm(device_index, ptr, size); } else { - ggml_sycl_pool_free_leg(ptr, size); + ggml_sycl_pool_free_leg(device_index, ptr, size); } } catch (sycl::exception const &exc) { @@ -11184,14 +13156,18 @@ catch (sycl::exception const &exc) { template struct sycl_pool_alloc { + int device_index = -1; + int device_id = -1; T * ptr = nullptr; size_t actual_size = 0; // size is in number of elements T * alloc(size_t size) { GGML_ASSERT(ptr == nullptr); - ptr = (T *) ggml_sycl_pool_malloc(size * sizeof(T), &this->actual_size); - // GGML_SYCL_DEBUG("alloc %lu return %p actual size=%lu\n", size * sizeof(T), ptr, this->actual_size); + device_id = get_current_device_id(); + device_index = g_sycl_gpu_mgr->get_index(device_id); + ptr = (T *) ggml_sycl_pool_malloc(device_index, size * sizeof(T), &this->actual_size); + // GGML_SYCL_DEBUG("sycl_pool_alloc %lu return %p actual size=%lu\n", size * sizeof(T), ptr, this->actual_size); return ptr; } @@ -11201,7 +13177,7 @@ struct sycl_pool_alloc { ~sycl_pool_alloc() { if (ptr != nullptr) { - ggml_sycl_pool_free(ptr, actual_size); + ggml_sycl_pool_free(device_index, ptr, actual_size); } } @@ -11222,44 +13198,57 @@ bool ggml_sycl_loaded(void) { return g_sycl_loaded; } -void ggml_backend_sycl_print_sycl_devices(){ - int device_count = dpct::dev_mgr::instance().device_count(); - fprintf(stderr, "found %d SYCL devices:\n", device_count); - for (int id = 0; id < device_count; ++id) { - dpct::device_info prop; - SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( - prop, dpct::dev_mgr::instance().get_device(id)))); - sycl::device cur_device = dpct::dev_mgr::instance().get_device(id); - fprintf(stderr, " Device %d: %s,\tcompute capability %d.%d,\n\tmax compute_units %d,\tmax work group size %d,\tmax sub group size %d,\tglobal mem size %lu\n", id, - prop.get_name(), prop.get_major_version(), - prop.get_minor_version(), - prop.get_max_compute_units(), - prop.get_max_work_group_size(), - prop.get_max_sub_group_size(), - prop.get_global_mem_size() - ); - } - // fprintf(stderr, "\n"); +void print_device_detail(int id) { + dpct::device_info prop; + SYCL_CHECK(CHECK_TRY_ERROR( + dpct::get_device_info(prop, dpct::dev_mgr::instance().get_device(id)))); + sycl::device cur_device = dpct::dev_mgr::instance().get_device(id); + std::string version; + version += std::to_string(prop.get_major_version()); + version += "."; + version += std::to_string(prop.get_minor_version()); + + fprintf(stderr, "|%2d|%45s|%18s|%17d|%14d|%13d|%15lu|\n", id, + prop.get_name(), version.c_str(), prop.get_max_compute_units(), + prop.get_max_work_group_size(), prop.get_max_sub_group_size(), + prop.get_global_mem_size()); } -int get_sycl_env(const char* env_name, int default_val){ - char * user_device_string = getenv(env_name); +void ggml_backend_sycl_print_sycl_devices() { + int device_count = dpct::dev_mgr::instance().device_count(); + fprintf(stderr, "found %d SYCL devices:\n", device_count); + fprintf(stderr, "|ID| Name |compute capability|Max compute units|Max work group|Max sub group|Global mem size|\n"); + fprintf(stderr, "|--|---------------------------------------------|------------------|-----------------|--------------|-------------|---------------|\n"); + for (int id = 0; id < device_count; ++id) { + print_device_detail(id); + } +} + +void print_gpu_device_list() { + fprintf(stderr, "detect %d SYCL GPUs: [%s] with Max compute units:%d\n", + g_sycl_gpu_mgr->get_gpu_count(), + g_sycl_gpu_mgr->gpus_list.c_str(), + g_sycl_gpu_mgr->max_compute_units); +} + +int get_sycl_env(const char *env_name, int default_val) { + char *user_device_string = getenv(env_name); int user_number = default_val; unsigned n; - if (user_device_string != NULL && sscanf(user_device_string, " %u", &n) == 1) { - user_number = (int)n; - } else { - user_number=default_val; - } + if (user_device_string != NULL && + sscanf(user_device_string, " %u", &n) == 1) { + user_number = (int)n; + } else { + user_number = default_val; + } return user_number; } -int get_work_group_size(int user_device_id){ +int get_work_group_size(int user_device_id) { dpct::device_info prop; - dpct::get_device_info( - prop, - dpct::dev_mgr::instance().get_device(user_device_id)); + dpct::get_device_info(prop, + dpct::dev_mgr::instance().get_device(user_device_id)); return prop.get_max_work_group_size(); } @@ -11268,113 +13257,81 @@ void ggml_init_sycl() try { if (!initialized) { g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0); + fprintf(stderr, "%s: GGML_SYCL_DEBUG: %d\n", __func__, g_ggml_sycl_debug); - printf("GGML_SYCL_DEBUG=%d\n", g_ggml_sycl_debug); - - int user_device_id = get_sycl_env("GGML_SYCL_DEVICE", 0); - +#if defined(GGML_SYCL_F16) + fprintf(stderr, "%s: GGML_SYCL_F16: yes\n", __func__); +#else + fprintf(stderr, "%s: GGML_SYCL_F16: no\n", __func__); +#endif if (CHECK_TRY_ERROR(g_all_sycl_device_count = - dpct::dev_mgr::instance().device_count()) != - 0) { + dpct::dev_mgr::instance().device_count()) != 0) { initialized = true; g_sycl_loaded = false; return; } GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES); + ggml_backend_sycl_print_sycl_devices(); + + if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr(); + + g_device_count = g_sycl_gpu_mgr->get_gpu_count(); + g_work_group_size = g_sycl_gpu_mgr->work_group_size; + + print_gpu_device_list(); + int64_t total_vram = 0; -#if defined(GGML_SYCL_F16) - fprintf(stderr, "%s: GGML_SYCL_F16: yes\n", __func__); -#else - fprintf(stderr, "%s: GGML_SYCL_F16: no\n", __func__); -#endif - - +/* NOT REMOVE, keep it for next optimize for XMX. #if defined(SYCL_USE_XMX) fprintf(stderr, "%s: SYCL_USE_XMX: yes\n", __func__); #else fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__); #endif - ggml_backend_sycl_print_sycl_devices(); +*/ for (int id = 0; id < GGML_SYCL_MAX_DEVICES; ++id) { - g_sycl_device_id2index[id].index = -1; g_device_caps[id].vmm = 0; g_device_caps[id].device_id = -1; g_device_caps[id].cc = 0; g_tensor_split[id] = 0; + g_default_tensor_split[id] = 0; } - int device_inx = -1; - for (int id = 0; id < g_all_sycl_device_count; ++id) { - if(id!=user_device_id) continue; - - device_inx++; - - g_device_caps[device_inx].vmm = 0; - g_device_caps[device_inx].device_id = id; - g_sycl_device_id2index[id].index = device_inx; + for (int i = 0; i < g_device_count; ++i) { + int device_id = g_sycl_gpu_mgr->gpus[i]; + g_device_caps[i].vmm = 0; dpct::device_info prop; SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( - prop, dpct::dev_mgr::instance().get_device(id)))); + prop, dpct::dev_mgr::instance().get_device(device_id)))); - g_tensor_split[device_inx] = total_vram; + g_default_tensor_split[i] = total_vram; total_vram += prop.get_global_mem_size(); - g_device_caps[device_inx].cc = + g_device_caps[i].cc = 100 * prop.get_major_version() + 10 * prop.get_minor_version(); - - } - device_inx = -1; - for (int id = 0; id < g_all_sycl_device_count; ++id) { - if(id!=user_device_id) continue; - device_inx++; - g_tensor_split[device_inx] /= total_vram; } - device_inx = -1; - for (int id = 0; id < g_all_sycl_device_count; ++id) { - if(id!=user_device_id) continue; - device_inx++; - SYCL_CHECK(ggml_sycl_set_device(id)); + for (int i = 0; i < g_device_count; ++i) { + g_default_tensor_split[i] /= total_vram; + } + + for (int i = 0; i < g_device_count; ++i) { + SYCL_CHECK(ggml_sycl_set_device(i)); // create sycl streams for (int is = 0; is < MAX_STREAMS; ++is) { - /* - DPCT1025:88: The SYCL queue is created ignoring the flag and - priority options. - */ SYCL_CHECK(CHECK_TRY_ERROR( - g_syclStreams[device_inx][is] = - dpct::get_current_device().create_queue())); + g_syclStreams[i][is] = + dpct::get_current_device().create_queue( + g_sycl_gpu_mgr->get_co_ctx(), dpct::get_current_device()))); } - const dpct::queue_ptr stream = g_syclStreams[device_inx][0]; + const dpct::queue_ptr stream = g_syclStreams[i][0]; // create sycl handle - SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[device_inx] = - stream)); - /* - DPCT1027:89: The call to syclSetMathMode was replaced with 0 - because this functionality is redundant in SYCL. - */ - SYCL_CHECK(0); + SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[i] = stream)); } - // configure logging to stdout - // SYCL_CHECK(syclLoggerConfigure(1, 1, 0, nullptr)); - - //hardcode, force set to 1 device - g_device_count = 1; - ggml_sycl_set_main_device(user_device_id); - ggml_sycl_set_device(user_device_id); - g_work_group_size = get_work_group_size(user_device_id); - // fprintf(stderr, "Using Device %d\n", user_device_id); - - // for (int id = 0; id < g_all_sycl_device_count; ++id) { - // GGML_SYCL_DEBUG("id=%d g_device_caps[%d].device_id=%d g_sycl_device_id2index[%d].index=%d ", id, id, - // g_device_caps[id].device_id, id, g_sycl_device_id2index[id].index); - // } - initialized = true; g_sycl_loaded = true; } @@ -11385,31 +13342,6 @@ catch (sycl::exception const &exc) { std::exit(1); } - -void ggml_sycl_set_tensor_split(const float * tensor_split) { - if (tensor_split == nullptr) { - return; - } - bool all_zero = true; - for (int i = 0; i < g_device_count; ++i) { - if (tensor_split[i] != 0.0f) { - all_zero = false; - break; - } - } - if (all_zero) { - return; - } - float split_sum = 0.0f; - for (int i = 0; i < g_device_count; ++i) { - g_tensor_split[i] = split_sum; - split_sum += tensor_split[i]; - } - for (int i = 0; i < g_device_count; ++i) { - g_tensor_split[i] /= split_sum; - } -} - void *ggml_sycl_host_malloc(size_t size) try { if (getenv("GGML_SYCL_NO_PINNED") != nullptr) { return nullptr; @@ -11419,28 +13351,14 @@ void *ggml_sycl_host_malloc(size_t size) try { //allow to use dpct::get_in_order_queue() for host malloc dpct::err0 err = CHECK_TRY_ERROR( ptr = (void *)sycl::malloc_host(size, dpct::get_in_order_queue())); - /* - DPCT1000:82: Error handling if-stmt was detected but could not be rewritten. - */ + if (err != 0) { // clear the error - /* - DPCT1026:83: The call to syclGetLastError was removed because this - functionality is redundant in SYCL. - */ - /* - DPCT1001:81: The statement could not be removed. - */ fprintf( stderr, "WARNING: failed to allocate %.2f MB of pinned memory: %s\n", - /* - DPCT1009:84: SYCL uses exceptions to report errors and does not use - the error codes. The original code was commented out and a warning - string was inserted. You need to rewrite this code. - */ size / 1024.0 / 1024.0, - "syclGetErrorString is not supported" /*syclGetErrorString(err)*/); + "syclGetErrorString is not supported"); return nullptr; } @@ -11480,7 +13398,7 @@ static dpct::err0 ggml_sycl_cpy_tensor_2d(void *dst, ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra; int id; SYCL_CHECK(CHECK_TRY_ERROR( - id = get_current_device_index())); + id = get_current_device_id())); // GGML_SYCL_DEBUG("current device index %d\n", id); src_ptr = (char *) extra->data_device[id]; } else { @@ -11714,7 +13632,6 @@ inline void ggml_sycl_op_tanh(const ggml_tensor *src0, const ggml_tensor *src1, GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); - tanh_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); (void) src1; @@ -11737,6 +13654,37 @@ inline void ggml_sycl_op_relu(const ggml_tensor *src0, const ggml_tensor *src1, (void) src1_dd; } +static void ggml_sycl_op_hardsigmoid(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + hardsigmoid_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +static void ggml_sycl_op_hardswish(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + hardswish_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + inline void ggml_sycl_op_leaky_relu(const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, const float *src0_dd, const float *src1_dd, @@ -11905,7 +13853,7 @@ inline void ggml_sycl_op_mul_mat_q( int device_id; SYCL_CHECK( - CHECK_TRY_ERROR(device_id = dpct::dev_mgr::instance().current_device_id())); + CHECK_TRY_ERROR(device_id = get_current_device_id())); // the main device has a larger memory buffer to hold the results from all GPUs // nrows_dst == nrows of the matrix that the dequantize_mul_mat kernel writes into @@ -11957,16 +13905,16 @@ catch (sycl::exception const &exc) { std::exit(1); } -static int64_t get_row_rounding(ggml_type type) { +static int64_t get_row_rounding(ggml_type type, const std::array & tensor_split) { int64_t min_compute_capability = INT_MAX; int64_t max_compute_capability = INT_MIN; - for (int64_t id = 0; id < g_device_count; ++id) { - if (g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { - if (min_compute_capability > g_device_caps[id].cc) { - min_compute_capability = g_device_caps[id].cc; + for (int i = 0; i < g_device_count; ++i) { + if (tensor_split[i] < (i + 1 < g_device_count ? tensor_split[i + 1] : 1.0f)) { + if (min_compute_capability > g_device_caps[i].cc) { + min_compute_capability = g_device_caps[i].cc; } - if (max_compute_capability < g_device_caps[id].cc) { - max_compute_capability = g_device_caps[id].cc; + if (max_compute_capability < g_device_caps[i].cc) { + max_compute_capability = g_device_caps[i].cc; } } } @@ -11986,12 +13934,16 @@ static int64_t get_row_rounding(ggml_type type) { case GGML_TYPE_Q3_K: case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: return max_compute_capability >= VER_GEN9 ? 128 : 64; case GGML_TYPE_Q6_K: return 64; default: GGML_ASSERT(false); } + } inline void ggml_sycl_op_mul_mat_vec_q( @@ -12006,63 +13958,46 @@ inline void ggml_sycl_op_mul_mat_vec_q( const int64_t ne00 = src0->ne[0]; const int64_t row_diff = row_high - row_low; - // TODO: support these quantization types - GGML_ASSERT(!(src0->type == GGML_TYPE_IQ2_XXS || - src0->type == GGML_TYPE_IQ2_XS || - src0->type == GGML_TYPE_IQ3_XXS || - src0->type == GGML_TYPE_IQ1_S)); - switch (src0->type) { case GGML_TYPE_Q4_0: - mul_mat_vec_q_sycl_submitter( - src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; + mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; case GGML_TYPE_Q4_1: - mul_mat_vec_q_sycl_submitter( - src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; + mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; case GGML_TYPE_Q5_0: - mul_mat_vec_q_sycl_submitter( - src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; + mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; case GGML_TYPE_Q5_1: - mul_mat_vec_q_sycl_submitter( - src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; + mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; case GGML_TYPE_Q8_0: - mul_mat_vec_q_sycl_submitter( - src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; + mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; case GGML_TYPE_Q2_K: - mul_mat_vec_q_sycl_submitter( - src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; + mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; case GGML_TYPE_Q3_K: - mul_mat_vec_q_sycl_submitter( - src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; + mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; case GGML_TYPE_Q4_K: - mul_mat_vec_q_sycl_submitter( - src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; + mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; case GGML_TYPE_Q5_K: - mul_mat_vec_q_sycl_submitter( - src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; + mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; case GGML_TYPE_Q6_K: - mul_mat_vec_q_sycl_submitter( - src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; + mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ2_XXS: + mul_mat_vec_iq2_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ2_XS: + mul_mat_vec_iq2_xs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ3_XXS: + mul_mat_vec_iq3_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; default: GGML_ASSERT(false); break; @@ -12075,6 +14010,7 @@ inline void ggml_sycl_op_mul_mat_vec_q( (void) src1_padded_row_size; } + inline void ggml_sycl_op_dequantize_mul_mat_vec( const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, @@ -12082,10 +14018,11 @@ inline void ggml_sycl_op_dequantize_mul_mat_vec( const int64_t src1_ncols, const int64_t src1_padded_row_size, const dpct::queue_ptr &stream) { - GGML_TENSOR_BINARY_OP_LOCALS; - + const int64_t ne00 = src0->ne[0]; const int64_t row_diff = row_high - row_low; + GGML_ASSERT(src1->type == GGML_TYPE_F32); + // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics #ifdef GGML_SYCL_F16 sycl_pool_alloc src1_dfloat_a; @@ -12097,15 +14034,10 @@ inline void ggml_sycl_op_dequantize_mul_mat_vec( src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16; if (src1_convert_f16) { - if (src1->type == GGML_TYPE_F16) { - src1_dfloat = (sycl::half *)src1->data + src1_padded_row_size; - } else { - src1_dfloat = src1_dfloat_a.alloc(ne00); - ggml_cpy_f32_f16_sycl((const char *)src1_ddf_i, (char *)src1_dfloat, - ne00, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, - nb13, stream); - } + src1_dfloat = src1_dfloat_a.alloc(ne00); + const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type); + GGML_ASSERT(to_fp16_sycl != nullptr); + to_fp16_sycl(src1_ddf_i, src1_dfloat, ne00, stream); } #else const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion @@ -12176,27 +14108,22 @@ inline void ggml_sycl_op_mul_mat_sycl( const int64_t row_diff = row_high - row_low; int id; - int device_id = dpct::dev_mgr::instance().current_device_id(); SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_index())); + CHECK_TRY_ERROR(id = get_current_device_id())); // the main device has a larger memory buffer to hold the results from all GPUs // ldc == nrows of the matrix that cuBLAS writes into - int ldc = dst->backend == GGML_BACKEND_TYPE_GPU && device_id == g_main_device ? ne0 : row_diff; + int ldc = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device ? ne0 : row_diff; #ifdef GGML_SYCL_F16 bool use_fp16 = true; // TODO(Yu) SYCL capability check #else bool use_fp16 = false; #endif - // if (compute_capability >= VER_GEN9 && (src0->type == GGML_TYPE_F16 || - // ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == - // src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) { if ((src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && use_fp16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) { - // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32 // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat_sycl - fp16 path\n"); sycl_pool_alloc src0_as_f16; if (src0->type != GGML_TYPE_F16) { @@ -12225,7 +14152,6 @@ inline void ggml_sycl_op_mul_mat_sycl( const sycl::half alpha_f16 = 1.0f; const sycl::half beta_f16 = 0.0f; - SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[id] = stream)); SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm( *g_sycl_handles[id], oneapi::mkl::transpose::trans, @@ -12234,21 +14160,28 @@ inline void ggml_sycl_op_mul_mat_sycl( src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16, dst_f16.get(), dpct::library_data_t::real_half, ldc, dpct::library_data_t::real_half))); - + g_sycl_handles[id]->wait(); const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16); to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream); } else { // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat_sycl - fp32 path\n"); sycl_pool_alloc src0_ddq_as_f32; - + sycl_pool_alloc src1_ddq_as_f32; if (src0->type != GGML_TYPE_F32) { const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src0->type); GGML_ASSERT(to_fp32_sycl != nullptr); src0_ddq_as_f32.alloc(row_diff*ne00); to_fp32_sycl(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream); } + if (src1->type != GGML_TYPE_F32) { + const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src1->type); + GGML_ASSERT(to_fp32_sycl != nullptr); + src1_ddq_as_f32.alloc(src1_ncols*ne10); + to_fp32_sycl(src1_ddf_i, src1_ddq_as_f32.get(), src1_ncols*ne10, stream); + } const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get(); + const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get(); const float alpha = 1.0f; const float beta = 0.0f; @@ -12258,10 +14191,10 @@ inline void ggml_sycl_op_mul_mat_sycl( *g_sycl_handles[id], oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, dpct::get_value(&alpha, *g_sycl_handles[id]), src0_ddf_i, ne00, - src1_ddf_i, ne10, dpct::get_value(&beta, *g_sycl_handles[id]), + src1_ddf1_i, ne10, dpct::get_value(&beta, *g_sycl_handles[id]), dst_dd_i, ldc))); + g_sycl_handles[id]->wait(); } - (void) dst; (void) src1_ddq_i; (void) src1_padded_row_size; @@ -12382,6 +14315,48 @@ inline void ggml_sycl_op_alibi(const ggml_tensor *src0, const ggml_tensor *src1, (void) src1_dd; } +static void ggml_sycl_op_pool2d(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int32_t * opts = (const int32_t *)dst->op_params; + enum ggml_op_pool op = static_cast(opts[0]); + const int k0 = opts[1]; + const int k1 = opts[2]; + const int s0 = opts[3]; + const int s1 = opts[4]; + const int p0 = opts[5]; + const int p1 = opts[6]; + + const int64_t IH = src0->ne[1]; + const int64_t IW = src0->ne[0]; + + const int64_t N = dst->ne[3]; + const int64_t OC = dst->ne[2]; + const int64_t OH = dst->ne[1]; + const int64_t OW = dst->ne[0]; + + const int parallel_elements = N * OC * OH * OW; + const int num_blocks = (parallel_elements + SYCL_POOL2D_BLOCK_SIZE - 1) / SYCL_POOL2D_BLOCK_SIZE; + sycl::range<3> block_nums(1, 1, num_blocks); + main_stream->parallel_for( + sycl::nd_range<3>(block_nums * + sycl::range<3>(1, 1, SYCL_IM2COL_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_IM2COL_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + pool2d_nchw_kernel(IH, IW, OH, OW, k1, k0, s1, s0, p1, p0, + parallel_elements, src0_dd, dst_dd, op, + item_ct1); + }); + + (void) src1; + (void) src1_dd; +} + inline void ggml_sycl_op_im2col(const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, const float *src0_dd, const float *src1_dd, @@ -12606,12 +14581,12 @@ static void ggml_sycl_op_flatten(const ggml_tensor *src0, sycl_pool_alloc dst_f; ggml_sycl_set_device(g_main_device); - dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0]; - // GGML_SYCL_DEBUG("g_main_device_index=%d, main_stream=%p src0_on_device=%d, src1_on_device=%d, dst_on_device=%d\n", - // g_main_device_index, main_stream, src0_on_device, src1_on_device, dst_on_device); + dpct::queue_ptr main_stream = g_syclStreams[g_main_device][0]; + // GGML_SYCL_DEBUG("g_main_device=%d, main_stream=%p src0_on_device=%d, src1_on_device=%d, dst_on_device=%d\n", + // g_main_device, main_stream, src0_on_device, src1_on_device, dst_on_device); if (src0_on_device) { - src0_ddf = (float *) src0_extra->data_device[g_main_device_index]; + src0_ddf = (float *) src0_extra->data_device[g_main_device]; } else { src0_ddf = src0_f.alloc(ggml_nelements(src0)); // GGML_SYCL_DEBUG("before ggml_sycl_cpy_tensor_2d src0_ddf=%p, src0=%p\n", src0_ddf, src0); @@ -12620,15 +14595,14 @@ static void ggml_sycl_op_flatten(const ggml_tensor *src0, if (use_src1) { if (src1_on_device) { - src1_ddf = (float *) src1_extra->data_device[g_main_device_index]; + src1_ddf = (float *) src1_extra->data_device[g_main_device]; } else { src1_ddf = src1_f.alloc(ggml_nelements(src1)); SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src1_ddf, src1, 0, 0, 0, nrows1, main_stream)); } } if (dst_on_device) { - dst_ddf = (float *) dst_extra->data_device[g_main_device_index]; - // printf("zjy dst_ddf=%p main_stream=%p g_main_device_index=%d\n", dst_ddf, main_stream, g_main_device_index); + dst_ddf = (float *) dst_extra->data_device[g_main_device]; } else { dst_ddf = dst_f.alloc(ggml_nelements(dst)); } @@ -12646,7 +14620,7 @@ static void ggml_sycl_op_flatten(const ggml_tensor *src0, // copy dst to host if necessary if (!dst_on_device) { SYCL_CHECK(CHECK_TRY_ERROR( - main_stream->memcpy(dst->data, dst_ddf, ggml_nbytes(dst)))); + main_stream->memcpy(dst->data, dst_ddf, ggml_nbytes(dst)).wait())); } if (dst->backend == GGML_BACKEND_TYPE_CPU) { @@ -12672,21 +14646,19 @@ static void ggml_sycl_set_peer_access(const int n_tokens) { } #ifdef NDEBUG - for (int id = 0; id < g_device_count; ++id) { - SYCL_CHECK(ggml_sycl_set_device(get_device_id_by_index(id))); + for (int i = 0; i < g_device_count; ++i) { + SYCL_CHECK(ggml_sycl_set_device(i)); // SYCL_CHECK(syclDeviceSynchronize()); } - for (int id = 0; id < g_device_count; ++id) { - SYCL_CHECK(ggml_sycl_set_device(get_device_id_by_index(id))); - int device_id = g_device_caps[id].device_id; + for (int i = 0; i < g_device_count; ++i) { + SYCL_CHECK(ggml_sycl_set_device(i)); for (int id_other = 0; id_other < g_device_count; ++id_other) { - int device_id_other = g_device_caps[id_other].device_id; - if (device_id == id_other) { + if (i == id_other) { continue; } - if (device_id != g_main_device && device_id_other != g_main_device) { + if (i != g_main_device && id_other != g_main_device) { continue; } @@ -12706,6 +14678,10 @@ static void ggml_sycl_set_peer_access(const int n_tokens) { peer_access_enabled = enable_peer_access; } +struct ggml_backend_sycl_split_buffer_type_context { + std::array tensor_split; +}; + static void ggml_sycl_op_mul_mat(const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, ggml_sycl_op_mul_mat_t op, @@ -12726,6 +14702,7 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0, GGML_ASSERT(dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT); GGML_ASSERT(src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT); + GGML_ASSERT(src1->type == GGML_TYPE_F32 || (src1->ne[2] == 1 && src1->ne[3] == 1)); GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0); @@ -12751,80 +14728,90 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0, GGML_ASSERT(!(split && ne03 > 1)); GGML_ASSERT(!(split && ne02 < ne12)); - // dd = data device - char * src0_dd[GGML_SYCL_MAX_DEVICES] = {nullptr}; - float * src1_ddf[GGML_SYCL_MAX_DEVICES] = {nullptr}; // float - char * src1_ddq[GGML_SYCL_MAX_DEVICES] = {nullptr}; // q8_1 - float * dst_dd[GGML_SYCL_MAX_DEVICES] = {nullptr}; + std::array tensor_split; + if (split) { + // TODO: check that src0->buffer->buft is a split buffer type, replace GGML_BACKEND_TYPE_GPU_SPLIT check + // GGML_ASSERT(src0->buffer != nullptr && src0->buffer->buft == ...); + ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *) src0->buffer->buft->context; + tensor_split = buft_ctx->tensor_split; + } - // as = actual size - size_t src0_as[GGML_SYCL_MAX_DEVICES] = {0}; - size_t src1_asf[GGML_SYCL_MAX_DEVICES] = {0}; - size_t src1_asq[GGML_SYCL_MAX_DEVICES] = {0}; - size_t dst_as[GGML_SYCL_MAX_DEVICES] = {0}; + struct dev_data { + sycl_pool_alloc src0_dd_alloc; + sycl_pool_alloc src1_ddf_alloc; + sycl_pool_alloc src1_ddq_alloc; + sycl_pool_alloc dst_dd_alloc; - int64_t row_low[GGML_SYCL_MAX_DEVICES]; - int64_t row_high[GGML_SYCL_MAX_DEVICES]; + char *src0_dd = nullptr; + float *src1_ddf = nullptr; // float + char *src1_ddq = nullptr; // q8_1 + float *dst_dd = nullptr; + + int64_t row_low; + int64_t row_high; + }; + + dev_data dev[GGML_SYCL_MAX_DEVICES]; int used_devices = 0; + dpct::queue_ptr main_stream = g_syclStreams[g_main_device][0]; - for (int64_t id = 0; id < g_device_count; ++id) { + for (int i = 0; i < g_device_count; ++i) { // by default, use all rows - row_low[id] = 0; - row_high[id] = ne01; + dev[i].row_low = 0; + dev[i].row_high = ne01; // for multi GPU, get the row boundaries from tensor split // and round to mul_mat_q tile sizes if (split) { - const int64_t rounding = get_row_rounding(src0->type); + const int64_t rounding = get_row_rounding(src0->type, tensor_split); - if (id != 0) { - row_low[id] = ne01*g_tensor_split[id]; - if (row_low[id] < ne01) { - row_low[id] -= row_low[id] % rounding; + if (i != 0) { + dev[i].row_low = ne01*tensor_split[i]; + if (dev[i].row_low < ne01) { + dev[i].row_low -= dev[i].row_low % rounding; } } - if (id != g_device_count - 1) { - row_high[id] = ne01*g_tensor_split[id + 1]; - if (row_high[id] < ne01) { - row_high[id] -= row_high[id] % rounding; + if (i != g_device_count - 1) { + dev[i].row_high = ne01*tensor_split[i + 1]; + if (dev[i].row_high < ne01) { + dev[i].row_high -= dev[i].row_high % rounding; } } } } - for (int64_t id = 0; id < g_device_count; ++id) { - if ((!split && id != g_main_device_index) || row_low[id] == row_high[id]) { + for (int i = 0; i < g_device_count; ++i) { + if ((!split && i != g_main_device) || dev[i].row_low == dev[i].row_high) { continue; } used_devices++; - const bool src1_on_device = src1->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index; - const bool dst_on_device = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index; + const bool src1_on_device = src1->backend == GGML_BACKEND_TYPE_GPU && i == g_main_device; + const bool dst_on_device = dst->backend == GGML_BACKEND_TYPE_GPU && i == g_main_device; - ggml_sycl_set_device(get_device_id_by_index(id)); - const dpct::queue_ptr stream = g_syclStreams[id][0]; + ggml_sycl_set_device(i); + dpct::queue_ptr stream = g_syclStreams[i][0]; if (src0_on_device && src0_is_contiguous) { - src0_dd[id] = (char *) src0_extra->data_device[id]; + dev[i].src0_dd = (char *) src0_extra->data_device[i]; } else { - // const size_t size_src0_ddq = split ? (row_high[id]-row_low[id])*ne00 * src0_ts/src0_bs : ggml_nbytes(src0); - src0_dd[id] = (char *) ggml_sycl_pool_malloc(ggml_nbytes(src0), &src0_as[id]); + dev[i].src0_dd = dev[i].src0_dd_alloc.alloc(ggml_nbytes(src0)); } if (src1_on_device && src1_is_contiguous) { - src1_ddf[id] = (float *) src1_extra->data_device[id]; + dev[i].src1_ddf = (float *) src1_extra->data_device[i]; } else { - src1_ddf[id] = (float *) ggml_sycl_pool_malloc(ggml_nbytes(src1), &src1_asf[id]); + dev[i].src1_ddf = dev[i].src1_ddf_alloc.alloc(ggml_nelements(src1)); } if (convert_src1_to_q8_1) { - src1_ddq[id] = (char *) ggml_sycl_pool_malloc(nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs, &src1_asq[id]); + dev[i].src1_ddq = dev[i].src1_ddq_alloc.alloc(nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs); if (src1_on_device && src1_is_contiguous) { - quantize_row_q8_1_sycl(src1_ddf[id], src1_ddq[id], ne10, nrows1, src1_padded_col_size, stream); + quantize_row_q8_1_sycl(dev[i].src1_ddf, dev[i].src1_ddq, ne10, nrows1, src1_padded_col_size, stream); /* DPCT1010:90: SYCL uses exceptions to report errors and does not use the error codes. The call was replaced with 0. You need to @@ -12835,25 +14822,25 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0, } if (dst_on_device) { - dst_dd[id] = (float *) dst_extra->data_device[id]; + dev[i].dst_dd = (float *) dst_extra->data_device[i]; } else { - const size_t size_dst_ddf = split ? (row_high[id]-row_low[id])*ne1*sizeof(float) : ggml_nbytes(dst); - dst_dd[id] = (float *) ggml_sycl_pool_malloc(size_dst_ddf, &dst_as[id]); + const size_t size_dst_ddf = split ? (dev[i].row_high - dev[i].row_low)*ne1 : ggml_nelements(dst); + dev[i].dst_dd = dev[i].dst_dd_alloc.alloc(size_dst_ddf); } } // if multiple devices are used they need to wait for the main device // here an event is recorded that signals that the main device has finished calculating the input data if (split && used_devices > 1) { - SYCL_CHECK(ggml_sycl_set_device(g_main_device)); + ggml_sycl_set_device(g_main_device); /* DPCT1024:91: The original code returned the error code that was further consumed by the program logic. This original code was replaced with 0. You may need to rewrite the program logic consuming the error code. */ SYCL_CHECK(CHECK_TRY_ERROR( - *src0_extra->events[g_main_device_index][0] = - g_syclStreams[g_main_device_index][0]->ext_oneapi_submit_barrier())); + *src0_extra->events[g_main_device][0] = + g_syclStreams[g_main_device][0]->ext_oneapi_submit_barrier())); } const int64_t src1_col_stride = split && used_devices > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11; @@ -12861,22 +14848,27 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0, const int64_t is = split ? (src1_col_0/src1_col_stride) % MAX_STREAMS : 0; const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride; - for (int64_t id = 0; id < g_device_count; ++id) { - if ((!split && id != g_main_device_index) || row_low[id] == row_high[id]) { + for (int i = 0; i < g_device_count; ++i) { + if ((!split && i != g_main_device) || dev[i].row_low == dev[i].row_high) { continue; } - const bool src1_on_device = src1->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index; - const bool dst_on_device = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index; - const int64_t row_diff = row_high[id] - row_low[id]; + const bool src1_on_device = src1->backend == GGML_BACKEND_TYPE_GPU && i == g_main_device; + const bool dst_on_device = dst->backend == GGML_BACKEND_TYPE_GPU && i == g_main_device; + const int64_t row_diff = dev[i].row_high - dev[i].row_low; - ggml_sycl_set_device(get_device_id_by_index(id)); - const dpct::queue_ptr stream = g_syclStreams[id][is]; + ggml_sycl_set_device(i); + dpct::queue_ptr stream = g_syclStreams[i][is]; // wait for main GPU data if necessary - if (split && (id != g_main_device_index || is != 0)) { + if (split && (i != g_main_device || is != 0)) { + /* + DPCT1009:163: SYCL uses exceptions to report errors and does not + use the error codes. The original code was commented out and a + warning string was inserted. You need to rewrite this code. + */ SYCL_CHECK(CHECK_TRY_ERROR(stream->ext_oneapi_submit_barrier( - {*src0_extra->events[g_main_device_index][0]}))); + {*src0_extra->events[g_main_device][0]}))); } for (int64_t i0 = 0; i0 < ne13*ne12; ++i0) { @@ -12886,30 +14878,32 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0, const size_t src1_ddq_i_offset = (i0*ne11 + src1_col_0) * src1_padded_col_size*q8_1_ts/q8_1_bs; // for split tensors the data begins at i0 == i0_offset_low - char * src0_dd_i = src0_dd[id] + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs; - float * src1_ddf_i = src1_ddf[id] + (i0*ne11 + src1_col_0) * ne10; - char * src1_ddq_i = src1_ddq[id] + src1_ddq_i_offset; - float * dst_dd_i = dst_dd[id] + (i0*ne1 + src1_col_0) * (dst_on_device ? ne0 : row_diff); + char * src0_dd_i = dev[i].src0_dd + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs; + float * src1_ddf_i = dev[i].src1_ddf + (i0*ne11 + src1_col_0) * ne10; + char * src1_ddq_i = dev[i].src1_ddq + src1_ddq_i_offset; + float * dst_dd_i = dev[i].dst_dd + (i0*ne1 + src1_col_0) * (dst_on_device ? ne0 : row_diff); // the main device memory buffer can be on VRAM scratch, with space for all partial results // in that case an offset on dst_ddf_i is needed - if (dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index) { - dst_dd_i += row_low[id]; // offset is 0 if no tensor split + if (dst->backend == GGML_BACKEND_TYPE_GPU && i == g_main_device) { + dst_dd_i += dev[i].row_low; // offset is 0 if no tensor split } // copy src0, src1 to device if necessary if (src1->backend == GGML_BACKEND_TYPE_GPU && src1_is_contiguous) { - if (id != g_main_device_index) { + if (i != g_main_device) { if (convert_src1_to_q8_1) { - char * src1_ddq_i_source = src1_ddq[g_main_device_index] + src1_ddq_i_offset; - SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy( + char * src1_ddq_i_source = dev[g_main_device].src1_ddq + src1_ddq_i_offset; + SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy( src1_ddq_i, src1_ddq_i_source, src1_ncols * src1_padded_col_size * q8_1_ts / - q8_1_bs))); + q8_1_bs).wait())); } else { - float * src1_ddf_i_source = (float *) src1_extra->data_device[g_main_device_index]; + + float * src1_ddf_i_source = (float *) src1_extra->data_device[g_main_device]; src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10; - SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy( + + SYCL_CHECK(CHECK_TRY_ERROR(dev2dev_memcpy(*stream, *main_stream, src1_ddf_i, src1_ddf_i_source, src1_ncols * ne10 * sizeof(float)))); } @@ -12932,14 +14926,14 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0, } if (src1_col_0 == 0 && (!src0_on_device || !src0_is_contiguous) && i02 % i02_divisor == 0) { - SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src0_dd_i, src0, i03, i02/i02_divisor, row_low[id], row_high[id], stream)); + SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src0_dd_i, src0, i03, i02/i02_divisor, dev[i].row_low, dev[i].row_high, stream)); } if (src1->type == GGML_TYPE_F16) { src1_padded_col_size = (i0 * ne11 + src1_col_0) * ne10; } // do the computation op(src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i, - row_low[id], row_high[id], src1_ncols, src1_padded_col_size, stream); + dev[i].row_low, dev[i].row_high, src1_ncols, src1_padded_col_size, stream); /* DPCT1010:93: SYCL uses exceptions to report errors and does not use the error codes. The call was replaced with 0. You need to @@ -12955,7 +14949,7 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0, dst_off_device = dst->data; kind = dpct::device_to_host; } else if (dst->backend == GGML_BACKEND_TYPE_GPU) { - dst_off_device = dst_extra->data_device[g_main_device_index]; + dst_off_device = dst_extra->data_device[g_main_device]; kind = dpct::device_to_device; } else { GGML_ASSERT(false); @@ -12968,23 +14962,41 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0, // If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results. float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3); GGML_ASSERT(dst->nb[1] == ne0*sizeof(float)); - dhf_dst_i += src1_col_0*ne0 + row_low[id]; - SYCL_CHECK(CHECK_TRY_ERROR(dpct::async_dpct_memcpy( - dhf_dst_i, ne0 * sizeof(float), dst_dd_i, - row_diff * sizeof(float), row_diff * sizeof(float), - src1_ncols, kind, *stream))); + dhf_dst_i += src1_col_0*ne0 + dev[i].row_low; + + //todo, dirty solution. Need be updated when device2device memcpy() is supported. + if (kind == dpct::device_to_device) { + size_t dst_size = ggml_nbytes_pad(dst); + float *host_buf = (float *)malloc(dst_size); + SYCL_CHECK(CHECK_TRY_ERROR(dpct::async_dpct_memcpy( + host_buf, ne0 * sizeof(float), dst_dd_i, + row_diff * sizeof(float), row_diff * sizeof(float), + src1_ncols, dpct::device_to_host, *stream))); + dpct::dev_mgr::instance().get_device(g_sycl_gpu_mgr->gpus[i]).queues_wait_and_throw(); + SYCL_CHECK(CHECK_TRY_ERROR(dpct::async_dpct_memcpy( + dhf_dst_i, ne0 * sizeof(float), host_buf, + row_diff * sizeof(float), row_diff * sizeof(float), + src1_ncols, dpct::host_to_device, *main_stream))); + dpct::dev_mgr::instance().get_device(g_sycl_gpu_mgr->gpus[g_main_device]).queues_wait_and_throw(); + free(host_buf); + } else { + SYCL_CHECK(CHECK_TRY_ERROR(dpct::async_dpct_memcpy( + dhf_dst_i, ne0 * sizeof(float), dst_dd_i, + row_diff * sizeof(float), row_diff * sizeof(float), + src1_ncols, kind, *stream))); + } } else { float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3); GGML_ASSERT(dst->nb[1] == ne0*sizeof(float)); dhf_dst_i += src1_col_0*ne0; SYCL_CHECK(CHECK_TRY_ERROR( stream->memcpy(dhf_dst_i, dst_dd_i, - src1_ncols * ne0 * sizeof(float)))); + src1_ncols * ne0 * sizeof(float)).wait())); } } // add event for the main device to wait on until other device is done - if (split && (id != g_main_device_index || is != 0)) { + if (split && (i != g_main_device || is != 0)) { /* DPCT1024:94: The original code returned the error code that was further consumed by the program logic. This original @@ -12992,48 +15004,27 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0, program logic consuming the error code. */ SYCL_CHECK(CHECK_TRY_ERROR( - *src0_extra->events[id][is] = + *src0_extra->events[i][is] = stream->ext_oneapi_submit_barrier())); } } } } - for (int64_t id = 0; id < g_device_count; ++id) { - if ((!split && id != g_main_device_index) || row_low[id] == row_high[id]) { - continue; - } - SYCL_CHECK(ggml_sycl_set_device(get_device_id_by_index(id))); - - // free buffers again when done - if (dst_as[id] > 0) { - ggml_sycl_pool_free(dst_dd[id], dst_as[id]); - } - if (src1_asq[id] > 0) { - ggml_sycl_pool_free(src1_ddq[id], src1_asq[id]); - } - if (src1_asf[id] > 0) { - ggml_sycl_pool_free(src1_ddf[id], src1_asf[id]); - } - if (src0_as[id] > 0) { - ggml_sycl_pool_free(src0_dd[id], src0_as[id]); - } - } - // main device waits for all other devices to be finished if (split && g_device_count > 1) { int64_t is_max = (ne11 + MUL_MAT_SRC1_COL_STRIDE - 1) / MUL_MAT_SRC1_COL_STRIDE; is_max = is_max <= MAX_STREAMS ? is_max : MAX_STREAMS; - SYCL_CHECK(ggml_sycl_set_device(g_main_device)); - for (int64_t id = 0; id < g_device_count; ++id) { - if (row_low[id] == row_high[id]) { + ggml_sycl_set_device(g_main_device); + for (int i = 0; i < g_device_count; ++i) { + if (dev[i].row_low == dev[i].row_high) { continue; } for (int64_t is = 0; is < is_max; ++is) { SYCL_CHECK(CHECK_TRY_ERROR( - g_syclStreams[g_main_device_index][0]->ext_oneapi_submit_barrier( - {*src0_extra->events[id][is]}))); + g_syclStreams[g_main_device][0]->ext_oneapi_submit_barrier( + {*src0_extra->events[i][is]}))); } } } @@ -13050,110 +15041,132 @@ catch (sycl::exception const &exc) { std::exit(1); } + static void ggml_sycl_repeat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_repeat); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_get_rows(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_get_rows); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_add); - // log_tensor_with_cnt("log_ggml_sycl_add_src0", (struct ggml_tensor *) src0, 6); - // log_tensor_with_cnt("log_ggml_sycl_add_src1", (struct ggml_tensor *)src1, 6); - // log_tensor_with_cnt("log_ggml_sycl_add_dst", dst, 6); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_acc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_acc); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_mul(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_mul); - // log_tensor_with_cnt("log_ggml_sycl_mul_src0", (struct ggml_tensor *)src0, 6); - // log_tensor_with_cnt("log_ggml_sycl_mul_src1", (struct ggml_tensor *)src1, 6); - // log_tensor_with_cnt("log_ggml_sycl_mul_dst", dst, 6); - + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_div(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_div); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_gelu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_gelu); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_silu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_silu); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_gelu_quick(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_gelu_quick); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_tanh(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_tanh); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_relu); + GGML_SYCL_DEBUG("call %s done\n", __func__); +} + +static void ggml_sycl_hardsigmoid(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_hardsigmoid); + GGML_SYCL_DEBUG("call %s done\n", __func__); +} + +static void ggml_sycl_hardswish(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_hardswish); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_leaky_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_leaky_relu); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_sqr(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_sqr); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_norm); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_group_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_group_norm); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_concat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_concat); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_upscale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_upscale); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_pad(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_pad); + GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_rms_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_rms_norm); - // log_tensor_with_cnt("log_ggml_sycl_rms_norm_src0", (struct ggml_tensor *)src0, 6); - // log_tensor_with_cnt("log_ggml_sycl_rms_norm_src1", (struct ggml_tensor *)src1, 6); - // log_tensor_with_cnt("log_ggml_sycl_rms_norm_dst", dst, 6); + GGML_SYCL_DEBUG("call %s done\n", __func__); } bool ggml_sycl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { @@ -13188,16 +15201,16 @@ static void ggml_sycl_mul_mat_vec_p021(const ggml_tensor *src0, const int64_t ne12 = src1->ne[2]; SYCL_CHECK(ggml_sycl_set_device(g_main_device)); - dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0]; + dpct::queue_ptr main_stream = g_syclStreams[g_main_device][0]; ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; - void * src0_ddq = src0_extra->data_device[g_main_device_index]; + void * src0_ddq = src0_extra->data_device[g_main_device]; ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; - float * src1_ddf = (float *) src1_extra->data_device[g_main_device_index]; + float * src1_ddf = (float *) src1_extra->data_device[g_main_device]; ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; - float * dst_ddf = (float *) dst_extra->data_device[g_main_device_index]; + float * dst_ddf = (float *) dst_extra->data_device[g_main_device]; ggml_mul_mat_p021_f16_f32_sycl(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream); } @@ -13227,16 +15240,16 @@ static void ggml_sycl_mul_mat_vec_nc(const ggml_tensor *src0, const int64_t ne12 = src1->ne[2]; SYCL_CHECK(ggml_sycl_set_device(g_main_device)); - dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0]; + dpct::queue_ptr main_stream = g_syclStreams[g_main_device][0]; ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; - void * src0_ddq = src0_extra->data_device[g_main_device_index]; + void * src0_ddq = src0_extra->data_device[g_main_device]; ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; - float * src1_ddf = (float *) src1_extra->data_device[g_main_device_index]; + float * src1_ddf = (float *) src1_extra->data_device[g_main_device]; ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; - float * dst_ddf = (float *) dst_extra->data_device[g_main_device_index]; + float * dst_ddf = (float *) dst_extra->data_device[g_main_device]; const int64_t row_stride_x = nb01 / sizeof(sycl::half); const int64_t channel_stride_x = nb02 / sizeof(sycl::half); @@ -13269,60 +15282,56 @@ static void k_compute_batched_ptrs(const sycl::half *src0_as_f16, int64_t i03 = i13 / r3; int64_t i02 = i12 / r2; - ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03; - ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12/2 + i13*nb13/2; - ptrs_dst[0*ne23 + i12 + i13*ne12] = ( char *) dst + i12*nbd2 + i13*nbd3; + ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03; + ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13; + ptrs_dst[0*ne23 + i12 + i13*ne12] = ( char *) dst + i12*nbd2 + i13*nbd3; } -static void ggml_sycl_mul_mat_mat_batched_sycl(const ggml_tensor *src0, - const ggml_tensor *src1, - ggml_tensor *dst) try { +static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0, + const ggml_tensor *src1, + ggml_tensor *dst) try { GGML_ASSERT(!ggml_is_transposed(src0)); GGML_ASSERT(!ggml_is_transposed(src1)); - GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT); GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne); + GGML_TENSOR_BINARY_OP_LOCALS - GGML_TENSOR_LOCALS(int64_t, nb0, src0, nb); - - GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne); - - GGML_TENSOR_LOCALS(int64_t, nb1, src1, nb); - - const int64_t ne1 = ggml_nelements(src1); - const int64_t ne = ggml_nelements(dst); + const int64_t ne_dst = ggml_nelements(dst); SYCL_CHECK(ggml_sycl_set_device(g_main_device)); - dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0]; + dpct::queue_ptr main_stream = g_syclStreams[g_main_device][0]; SYCL_CHECK( - CHECK_TRY_ERROR(g_sycl_handles[g_main_device_index] = main_stream)); + CHECK_TRY_ERROR(g_sycl_handles[g_main_device] = main_stream)); ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; - void * src0_ddq = src0_extra->data_device[g_main_device_index]; + void * src0_ddq = src0_extra->data_device[g_main_device]; sycl::half *src0_as_f16 = (sycl::half *)src0_ddq; ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; - float * src1_ddf = (float *) src1_extra->data_device[g_main_device_index]; + float * src1_ddf = (float *) src1_extra->data_device[g_main_device]; ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; - float * dst_ddf = (float *) dst_extra->data_device[g_main_device_index]; + float * dst_ddf = (float *) dst_extra->data_device[g_main_device]; // convert src1 to fp16 - const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type); - GGML_ASSERT(to_fp16_sycl != nullptr); - - sycl_pool_alloc src1_as_f16(ne1); - to_fp16_sycl(src1_ddf, src1_as_f16.get(), ne1, main_stream); + sycl_pool_alloc src1_f16_alloc; + if (src1->type != GGML_TYPE_F16) { + const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type); + const int64_t ne_src1 = ggml_nelements(src1); + src1_f16_alloc.alloc(ne_src1); + GGML_ASSERT(to_fp16_sycl != nullptr); + to_fp16_sycl(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream); + } + sycl::half *src1_f16 = src1->type == GGML_TYPE_F16 ? (sycl::half *)src1_ddf + : src1_f16_alloc.get(); sycl_pool_alloc dst_f16; char * dst_t; - dpct::library_data_t cu_compute_type = dpct::library_data_t::real_half; - dpct::library_data_t cu_data_type = dpct::library_data_t::real_half; + dpct::library_data_t cu_compute_type = dpct::library_data_t::real_float; + dpct::library_data_t cu_data_type = dpct::library_data_t::real_float; // dst strides size_t nbd2 = dst->nb[2]; @@ -13334,23 +15343,13 @@ static void ggml_sycl_mul_mat_mat_batched_sycl(const ggml_tensor *src0, const float alpha_f32 = 1.0f; const float beta_f32 = 0.0f; - const void * alpha = &alpha_f16; - const void * beta = &beta_f16; + const void * alpha = &alpha_f32; + const void * beta = &beta_f32; - if (dst->op_params[0] == GGML_PREC_DEFAULT) { - dst_t = (char *) dst_f16.alloc(ne); + // TODO: Renable (dst->op_params[0] =! GGML_PREC_DEFAULT) pathway + // oneMKL open source supports half, half, float, float: datatypes - nbd2 /= sizeof(float) / sizeof(sycl::half); - nbd3 /= sizeof(float) / sizeof(sycl::half); - } else { - dst_t = (char *) dst_ddf; - - cu_compute_type = dpct::library_data_t::real_float; - cu_data_type = dpct::library_data_t::real_float; - - alpha = &alpha_f32; - beta = &beta_f32; - } + dst_t = (char *) dst_ddf; GGML_ASSERT(ne12 % ne02 == 0); GGML_ASSERT(ne13 % ne03 == 0); @@ -13368,7 +15367,7 @@ static void ggml_sycl_mul_mat_mat_batched_sycl(const ggml_tensor *src0, int i02 = i12 / r2; SYCL_CHECK( - syclGemmEx(g_sycl_handles[g_main_device_index], CUBLAS_OP_T, CUBLAS_OP_N, + syclGemmEx(g_sycl_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N, ne01, ne11, ne10, alpha, (const char *) src0_as_f16 + i02*src0->nb[2] + i03*src0->nb[3] , SYCL_R_16F, nb01/sizeof(half), (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, SYCL_R_16F, nb11/sizeof(float), @@ -13381,18 +15380,17 @@ static void ggml_sycl_mul_mat_mat_batched_sycl(const ggml_tensor *src0, #else if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) { // there is no broadcast and src0, src1 are contiguous across dims 2, 3 - // use syclGemmStridedBatchedEx SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch( - *g_sycl_handles[g_main_device_index], oneapi::mkl::transpose::trans, + *g_sycl_handles[g_main_device], oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha, (const char *)src0_as_f16, dpct::library_data_t::real_half, - nb01 / sizeof(sycl::half), src0->nb[2] / sizeof(sycl::half), - (const char *)src1_as_f16.get(), dpct::library_data_t::real_half, - nb11 / sizeof(float), src1->nb[2] / sizeof(float), beta, - (char *)dst_t, cu_data_type, ne01, dst->nb[2] / sizeof(float), + nb01 / nb00, nb02 / nb00, + (const char *)src1_f16, dpct::library_data_t::real_half, + nb11 / nb10, nb12 / nb10, beta, + (char *)dst_t, cu_data_type, ne01, nb2 / nb0, ne12 * ne13, cu_compute_type))); + g_sycl_handles[g_main_device]->wait(); } else { - // use syclGemmBatchedEx const int ne23 = ne12*ne13; sycl_pool_alloc ptrs_src(2*ne23); @@ -13409,44 +15407,34 @@ static void ggml_sycl_mul_mat_mat_batched_sycl(const ggml_tensor *src0, {sycl::aspect::fp16}); main_stream->submit([&](sycl::handler &cgh) { - const sycl::half *src1_as_f16_get_ct1 = src1_as_f16.get(); - const void **ptrs_src_get_ct3 = ptrs_src.get(); - void **ptrs_dst_get_ct4 = ptrs_dst.get(); - + const void **ptrs_src_get = ptrs_src.get(); + void **ptrs_dst_get = ptrs_dst.get(); + size_t nb12_scaled = src1->type == GGML_TYPE_F16 ? nb12 : nb12 / 2; + size_t nb13_scaled = src1->type == GGML_TYPE_F16 ? nb13 : nb13 / 2; cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_compute_batched_ptrs( - src0_as_f16, src1_as_f16_get_ct1, - dst_t, ptrs_src_get_ct3, - ptrs_dst_get_ct4, ne12, ne13, ne23, - nb02, nb03, nb12, nb13, nbd2, nbd3, r2, - r3, item_ct1); + src0_as_f16, src1_f16, + dst_t, ptrs_src_get, + ptrs_dst_get, ne12, ne13, ne23, + nb02, nb03, nb12_scaled, nb13_scaled, + nbd2, nbd3, r2, r3, item_ct1); }); - }); + }).wait(); } - /* - DPCT1010:95: SYCL uses exceptions to report errors and does not use the - error codes. The call was replaced with 0. You need to rewrite this - code. - */ - SYCL_CHECK(0); - SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch( - *g_sycl_handles[g_main_device_index], oneapi::mkl::transpose::trans, + *g_sycl_handles[g_main_device], oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha, (const void **)(ptrs_src.get() + 0 * ne23), - dpct::library_data_t::real_half, nb01 / sizeof(sycl::half), + dpct::library_data_t::real_half, nb01 / nb00, (const void **)(ptrs_src.get() + 1 * ne23), - dpct::library_data_t::real_half, nb11 / sizeof(float), beta, + dpct::library_data_t::real_half, nb11 / nb10, beta, (void **)(ptrs_dst.get() + 0 * ne23), cu_data_type, ne01, ne23, cu_compute_type))); + g_sycl_handles[g_main_device]->wait(); } #endif - if (dst->op_params[0] == GGML_PREC_DEFAULT) { - const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16); - to_fp32_sycl(dst_f16.get(), dst_ddf, ne, main_stream); - } } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ @@ -13454,6 +15442,7 @@ catch (sycl::exception const &exc) { std::exit(1); } + static void ggml_sycl_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { const bool all_on_device = (src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT) && @@ -13463,9 +15452,9 @@ static void ggml_sycl_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1 const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT; int64_t min_compute_capability = INT_MAX; - for (int64_t id = 0; id < g_device_count; ++id) { - if (min_compute_capability > g_device_caps[id].cc && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { - min_compute_capability = g_device_caps[id].cc; + for (int i = 0; i < g_device_count; ++i) { + if (min_compute_capability > g_device_caps[i].cc && g_tensor_split[i] < (i + 1 < g_device_count ? g_tensor_split[i + 1] : 1.0f)) { + min_compute_capability = g_device_caps[i].cc; } } @@ -13491,10 +15480,10 @@ static void ggml_sycl_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1 // KQV single-batch // GGML_SYCL_DEBUG("ggml_sycl_mul_mat_vec_nc\n"); ggml_sycl_mul_mat_vec_nc(src0, src1, dst); - } else if (!split && all_on_device && use_xmx && src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) { + } else if (!split && all_on_device && use_xmx && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) { // KQ + KQV multi-batch - // GGML_SYCL_DEBUG("ggml_sycl_mul_mat_mat_batched_sycl\n"); - ggml_sycl_mul_mat_mat_batched_sycl(src0, src1, dst); + // GGML_SYCL_DEBUG("ggml_sycl_mul_mat_batched_sycl\n"); + ggml_sycl_mul_mat_batched_sycl(src0, src1, dst); } else if (src0->type == GGML_TYPE_F32) { // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat\n"); ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false); @@ -13558,13 +15547,13 @@ static __global__ void k_compute_batched_ptrs_id( src0_f16 = (half *) srcs_ar[i]; } else { src0_f16 = src0_as_f16; - if (threadIdx.x == 0 && threadIdx.y == 0) { + if (item_ct1.get_local_id(2) == 0 && threadIdx.y == 0) { const to_fp16_sycl_t to_fp16 = ggml_get_to_fp16_sycl(src0_type); to_fp16(srcs_ar[i], src0_f16, src0_ne, syclStreamFireAndForget); } } - int i13 = blockIdx.x * blockDim.x + threadIdx.x; + int i13 = blockIdx.x * blockDim.x + item_ct1.get_local_id(2); int i12 = blockIdx.y * blockDim.y + threadIdx.y; if (i13 >= ne13 || i12 >= ne12) { @@ -13606,30 +15595,30 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) { const int64_t ne = ggml_nelements(dst); SYCL_CHECK(ggml_sycl_set_device(g_main_device)); - syclStream_t main_stream = g_syclStreams[g_main_device_index][0]; + syclStream_t main_stream = g_syclStreams[g_main_device][0]; - SYCL_CHECK(syclSetStream(g_sycl_handles[g_main_device_index], main_stream)); + SYCL_CHECK(syclSetStream(g_sycl_handles[g_main_device], main_stream)); //ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; - //void * src0_ddq = src0_extra->data_device[g_main_device_index]; + //void * src0_ddq = src0_extra->data_device[g_main_device]; //half * src0_as_f16 = (half *) src0_ddq; ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; - float * src1_ddf = (float *) src1_extra->data_device[g_main_device_index]; + float * src1_ddf = (float *) src1_extra->data_device[g_main_device]; ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; - float * dst_ddf = (float *) dst_extra->data_device[g_main_device_index]; + float * dst_ddf = (float *) dst_extra->data_device[g_main_device]; // convert src1 to fp16 const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type); GGML_ASSERT(to_fp16_sycl != nullptr); size_t src1_as = 0; - half * src1_as_f16 = (half *) ggml_sycl_pool_malloc(ne1 * sizeof(half), &src1_as); + half * src1_as_f16 = (half *) ggml_sycl_pool_malloc(g_main_device, ne1 * sizeof(half), &src1_as); to_fp16_sycl(src1_ddf, src1_as_f16, ne1, main_stream); size_t dst_as = 0; - half * dst_f16 = (half *) ggml_sycl_pool_malloc(ne * sizeof(half), &dst_as); + half * dst_f16 = (half *) ggml_sycl_pool_malloc(g_main_device, ne * sizeof(half), &dst_as); GGML_ASSERT(ne12 % ne02 == 0); GGML_ASSERT(ne13 % ne03 == 0); @@ -13650,14 +15639,14 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) { size_t ptrs_src_s = 0; size_t ptrs_dst_s = 0; - ptrs_src = (const void **) ggml_sycl_pool_malloc(2*ne23*sizeof(void *), &ptrs_src_s); - ptrs_dst = ( void **) ggml_sycl_pool_malloc(1*ne23*sizeof(void *), &ptrs_dst_s); + ptrs_src = (const void **) ggml_sycl_pool_malloc(g_main_device, 2*ne23*sizeof(void *), &ptrs_src_s); + ptrs_dst = ( void **) ggml_sycl_pool_malloc(g_main_device, 1*ne23*sizeof(void *), &ptrs_dst_s); int64_t src0_ne = ggml_nelements(src00); half * src0_as_f16 = nullptr; size_t src0_as = 0; if (src00->type != GGML_TYPE_F16) { - src0_as_f16 = (half *) ggml_sycl_pool_malloc(src0_ne * sizeof(half), &src0_as); + src0_as_f16 = (half *) ggml_sycl_pool_malloc(g_main_device, src0_ne * sizeof(half), &src0_as); } static_assert(GGML_MAX_SRC == 6, "GGML_MAX_SRC == 6"); @@ -13672,16 +15661,16 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) { r2, r3, src00->type, src0_as_f16, src0_ne, src1_as_f16, dst_f16, - (const int *)((ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device_index], id, - dst->src[2] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[2]->extra)->data_device[g_main_device_index] : nullptr, - dst->src[3] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[3]->extra)->data_device[g_main_device_index] : nullptr, - dst->src[4] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[4]->extra)->data_device[g_main_device_index] : nullptr, - dst->src[5] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[5]->extra)->data_device[g_main_device_index] : nullptr + (const int *)((ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device], id, + dst->src[2] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[2]->extra)->data_device[g_main_device] : nullptr, + dst->src[3] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[3]->extra)->data_device[g_main_device] : nullptr, + dst->src[4] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[4]->extra)->data_device[g_main_device] : nullptr, + dst->src[5] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[5]->extra)->data_device[g_main_device] : nullptr ); SYCL_CHECK(syclGetLastError()); SYCL_CHECK( - syclGemmBatchedEx(g_sycl_handles[g_main_device_index], CUBLAS_OP_T, CUBLAS_OP_N, + syclGemmBatchedEx(g_sycl_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N, ne01, ne11, ne10, &alpha_f16, (const void **) (ptrs_src + 0*ne23), SYCL_R_16F, ne00, (const void **) (ptrs_src + 1*ne23), SYCL_R_16F, ne10, @@ -13691,20 +15680,20 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) { CUBLAS_GEMM_DEFAULT_TENSOR_OP)); if (src0_as != 0) { - ggml_sycl_pool_free(src0_as_f16, src0_as); + ggml_sycl_pool_free(g_main_device, src0_as_f16, src0_as); } if (ptrs_src_s != 0) { - ggml_sycl_pool_free(ptrs_src, ptrs_src_s); + ggml_sycl_pool_free(g_main_device, ptrs_src, ptrs_src_s); } if (ptrs_dst_s != 0) { - ggml_sycl_pool_free(ptrs_dst, ptrs_dst_s); + ggml_sycl_pool_free(g_main_device, ptrs_dst, ptrs_dst_s); } const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16); to_fp32_sycl(dst_f16, dst_ddf, ne, main_stream); - ggml_sycl_pool_free(src1_as_f16, src1_as); - ggml_sycl_pool_free(dst_f16, dst_as); + ggml_sycl_pool_free(g_main_device, src1_as_f16, src1_as); + ggml_sycl_pool_free(g_main_device, dst_f16, dst_as); } #endif @@ -13725,13 +15714,13 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0, std::vector ids_host(ggml_nbytes(ids)); - const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0]; + const dpct::queue_ptr stream = g_syclStreams[g_main_device][0]; if (ids->backend == GGML_BACKEND_TYPE_GPU) { - const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device_index]; + const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device]; SYCL_CHECK(CHECK_TRY_ERROR( - stream->memcpy(ids_host.data(), ids_dev, ggml_nbytes(ids)))); - SYCL_CHECK(CHECK_TRY_ERROR(stream->wait())); + stream->memcpy(ids_host.data(), ids_dev, ggml_nbytes(ids)).wait())); + // SYCL_CHECK(CHECK_TRY_ERROR(stream->wait())); } else { memcpy(ids_host.data(), ids->data, ggml_nbytes(ids)); } @@ -13752,9 +15741,9 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0, dst_row.extra = &dst_row_extra; char * src1_original = src1->backend == GGML_BACKEND_TYPE_CPU ? - (char *) src1->data : (char *) src1_extra->data_device[g_main_device_index]; + (char *) src1->data : (char *) src1_extra->data_device[g_main_device]; char * dst_original = dst->backend == GGML_BACKEND_TYPE_CPU ? - (char *) dst->data : (char *) dst_extra->data_device[g_main_device_index]; + (char *) dst->data : (char *) dst_extra->data_device[g_main_device]; if (src1->ne[1] == 1) { GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU); @@ -13771,10 +15760,10 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0, const struct ggml_tensor * src0_row = dst->src[row_id + 2]; - src1_row_extra.data_device[g_main_device_index] = src1_original + i01*src1->nb[1]; + src1_row_extra.data_device[g_main_device] = src1_original + i01*src1->nb[1]; src1_row.data = (char *) src1->data + i01*src1->nb[1]; // TODO why is this set? - dst_row_extra.data_device[g_main_device_index] = dst_original + i01*dst->nb[1]; + dst_row_extra.data_device[g_main_device] = dst_original + i01*dst->nb[1]; dst_row.data = (char *) dst->data + i01*dst->nb[1]; // TODO why is this set? ggml_sycl_mul_mat(src0_row, &src1_row, &dst_row); @@ -13783,8 +15772,8 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0, sycl_pool_alloc src1_contiguous(sizeof(float)*ggml_nelements(src1)); sycl_pool_alloc dst_contiguous(sizeof(float)*ggml_nelements(dst)); - src1_row_extra.data_device[g_main_device_index] = src1_contiguous.get(); - dst_row_extra.data_device[g_main_device_index] = dst_contiguous.get(); + src1_row_extra.data_device[g_main_device] = src1_contiguous.get(); + dst_row_extra.data_device[g_main_device] = dst_contiguous.get(); for (int32_t row_id = 0; row_id < n_as; ++row_id) { const struct ggml_tensor * src0_row = dst->src[row_id + 2]; @@ -13801,7 +15790,7 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0, SYCL_CHECK(CHECK_TRY_ERROR( stream->memcpy(src1_contiguous.get() + num_src1_rows * nb11, - src1_original + i01 * nb11, nb11))); + src1_original + i01 * nb11, nb11).wait())); num_src1_rows++; } @@ -13834,7 +15823,7 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0, SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy( dst_original + i01 * nb1, - dst_contiguous.get() + num_src1_rows * nb1, nb1))); + dst_contiguous.get() + num_src1_rows * nb1, nb1).wait())); num_src1_rows++; } } @@ -13872,13 +15861,13 @@ static void ggml_sycl_cpy(const ggml_tensor *src0, const ggml_tensor *src1, GGML_TENSOR_BINARY_OP_LOCALS; SYCL_CHECK(ggml_sycl_set_device(g_main_device)); - dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0]; + dpct::queue_ptr main_stream = g_syclStreams[g_main_device][0]; const ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; const ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; - char * src0_ddc = (char *) src0_extra->data_device[g_main_device_index]; - char * src1_ddc = (char *) src1_extra->data_device[g_main_device_index]; + char * src0_ddc = (char *) src0_extra->data_device[g_main_device]; + char * src1_ddc = (char *) src1_extra->data_device[g_main_device]; if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) { ggml_cpy_f32_f32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); @@ -13890,6 +15879,8 @@ static void ggml_sycl_cpy(const ggml_tensor *src0, const ggml_tensor *src1, ggml_cpy_f32_q4_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) { ggml_cpy_f32_q4_1_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) { + ggml_cpy_f16_f32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) { ggml_cpy_f16_f16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); } else if (src0->type == GGML_TYPE_I16 && src1->type == GGML_TYPE_I16) { @@ -13933,6 +15924,10 @@ static void ggml_sycl_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_alibi); } +static void ggml_sycl_pool2d(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_pool2d); +} + static void ggml_sycl_im2col(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_im2col); } @@ -13959,93 +15954,6 @@ static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_spl return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]); } -void ggml_sycl_transform_tensor(void *data, struct ggml_tensor *tensor) try { - const int64_t nrows = ggml_nrows(tensor); - - const int64_t ne0 = tensor->ne[0]; - - const size_t nb1 = tensor->nb[1]; - - ggml_backend_type backend = tensor->backend; - ggml_tensor_extra_gpu * extra = new struct ggml_tensor_extra_gpu; - memset(extra, 0, sizeof(*extra)); - - for (int64_t id = 0; id < g_device_count; ++id) { - if (backend == GGML_BACKEND_TYPE_GPU && id != g_main_device_index) { - continue; - } - ggml_sycl_set_device(get_device_id_by_index(id)); - const dpct::queue_ptr stream = g_syclStreams[id][0]; - - int64_t row_low, row_high; - if (backend == GGML_BACKEND_TYPE_GPU) { - row_low = 0; - row_high = nrows; - } else if (backend == GGML_BACKEND_TYPE_GPU_SPLIT) { - const int64_t rounding = get_row_rounding(tensor->type); - - row_low = id == 0 ? 0 : nrows*g_tensor_split[id]; - row_low -= row_low % rounding; - - if (id == g_device_count - 1) { - row_high = nrows; - } else { - row_high = nrows*g_tensor_split[id + 1]; - row_high -= row_high % rounding; - } - } else { - GGML_ASSERT(false); - } - if (row_low == row_high) { - continue; - } - - int64_t nrows_split = row_high - row_low; - - const size_t offset_split = row_low*nb1; - size_t size = ggml_nbytes_split(tensor, nrows_split); - const size_t original_size = size; - - // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses - if (ne0 % MATRIX_ROW_PADDING != 0) { - size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); - } - - char * buf; - SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)sycl::malloc_device( - size, *stream))); - char * buf_host = (char *)data + offset_split; - - // set padding to 0 to avoid possible NaN values - if (size > original_size) { - SYCL_CHECK(CHECK_TRY_ERROR( - (*stream) - .memset(buf + original_size, 0, size - original_size) - .wait())); - } - - SYCL_CHECK(CHECK_TRY_ERROR((*stream) - .memcpy(buf, buf_host, original_size) - .wait())); - - extra->data_device[id] = buf; - - if (backend == GGML_BACKEND_TYPE_GPU_SPLIT) { - for (int64_t is = 0; is < MAX_STREAMS; ++is) { - SYCL_CHECK(CHECK_TRY_ERROR(extra->events[id][is] = - new sycl::event())); - } - } - } - - tensor->extra = extra; -} -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} - void ggml_sycl_free_data(struct ggml_tensor *tensor) try { if (!tensor || !tensor->extra || (tensor->backend != GGML_BACKEND_TYPE_GPU && tensor->backend != GGML_BACKEND_TYPE_GPU_SPLIT) ) { return; @@ -14053,18 +15961,18 @@ void ggml_sycl_free_data(struct ggml_tensor *tensor) try { ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; - for (int64_t id = 0; id < g_device_count; ++id) { - const dpct::queue_ptr stream = g_syclStreams[id][0]; - if (extra->data_device[id] != nullptr) { - SYCL_CHECK(ggml_sycl_set_device(get_device_id_by_index(id))); - SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(extra->data_device[id], *stream))); + for (int i = 0; i < g_device_count; ++i) { + const dpct::queue_ptr stream = g_syclStreams[i][0]; + if (extra->data_device[i] != nullptr) { + SYCL_CHECK(ggml_sycl_set_device(i)); + SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(extra->data_device[i], *stream))); } for (int64_t is = 0; is < MAX_STREAMS; ++is) { - if (extra->events[id][is] != nullptr) { - SYCL_CHECK(ggml_sycl_set_device(get_device_id_by_index(id))); + if (extra->events[i][is] != nullptr) { + SYCL_CHECK(ggml_sycl_set_device(i)); SYCL_CHECK(CHECK_TRY_ERROR( - dpct::destroy_event(extra->events[id][is]))); + dpct::destroy_event(extra->events[i][is]))); } } } @@ -14124,22 +16032,22 @@ static void ggml_sycl_assign_buffers_impl(struct ggml_tensor *tensor, const size_t size = ggml_nbytes(tensor); SYCL_CHECK(ggml_sycl_set_device(g_main_device)); - const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0]; + const dpct::queue_ptr stream = g_syclStreams[g_main_device][0]; if (inplace && (tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU || tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT)) { ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra; - char * src0_ddc = (char *) src0_extra->data_device[g_main_device_index]; + char * src0_ddc = (char *) src0_extra->data_device[g_main_device]; size_t offset = 0; if (tensor->op == GGML_OP_VIEW) { memcpy(&offset, tensor->op_params, sizeof(size_t)); } extra = ggml_sycl_alloc_temp_tensor_extra(); - extra->data_device[g_main_device_index] = src0_ddc + offset; + extra->data_device[g_main_device] = src0_ddc + offset; } else if (tensor->op == GGML_OP_CPY) { ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu * ) tensor->src[1]->extra; - void * src1_ddv = src1_extra->data_device[g_main_device_index]; + void * src1_ddv = src1_extra->data_device[g_main_device]; extra = ggml_sycl_alloc_temp_tensor_extra(); - extra->data_device[g_main_device_index] = src1_ddv; + extra->data_device[g_main_device] = src1_ddv; } else if (scratch) { GGML_ASSERT(size <= g_scratch_size); if (g_scratch_offset + size > g_scratch_size) { @@ -14154,7 +16062,7 @@ static void ggml_sycl_assign_buffers_impl(struct ggml_tensor *tensor, g_scratch_buffer = data; } extra = ggml_sycl_alloc_temp_tensor_extra(); - extra->data_device[g_main_device_index] = data + g_scratch_offset; + extra->data_device[g_main_device] = data + g_scratch_offset; g_scratch_offset += size; @@ -14167,44 +16075,7 @@ static void ggml_sycl_assign_buffers_impl(struct ggml_tensor *tensor, (*stream).memset(data, 0, size).wait())); extra = new ggml_tensor_extra_gpu; memset(extra, 0, sizeof(*extra)); - extra->data_device[g_main_device_index] = data; - } - - tensor->extra = extra; -} -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} - -void ggml_sycl_assign_scratch_offset(struct ggml_tensor *tensor, - size_t offset) try { - if (g_scratch_size == 0) { - return; - } - if (g_scratch_buffer == nullptr) { - ggml_sycl_set_device(g_main_device); - const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0]; - SYCL_CHECK( - CHECK_TRY_ERROR(g_scratch_buffer = (void *)sycl::malloc_device( - g_scratch_size, *stream))); - } - - ggml_tensor_extra_gpu * extra = ggml_sycl_alloc_temp_tensor_extra(); - - const bool inplace = tensor->view_src != nullptr; - - if (inplace && (tensor->view_src->backend == GGML_BACKEND_TYPE_GPU || tensor->view_src->backend == GGML_BACKEND_TYPE_GPU_SPLIT)) { - ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->view_src->extra; - char * src0_ddc = (char *) src0_extra->data_device[g_main_device_index]; - size_t view_offset = 0; - if (tensor->op == GGML_OP_VIEW) { - memcpy(&view_offset, tensor->op_params, sizeof(size_t)); - } - extra->data_device[g_main_device_index] = src0_ddc + view_offset; - } else { - extra->data_device[g_main_device_index] = (char *) g_scratch_buffer + offset; + extra->data_device[g_main_device] = data; } tensor->extra = extra; @@ -14221,9 +16092,9 @@ void ggml_sycl_copy_to_device(struct ggml_tensor *tensor) try { ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; SYCL_CHECK(ggml_sycl_set_device(g_main_device)); - const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0]; + const dpct::queue_ptr stream = g_syclStreams[g_main_device][0]; SYCL_CHECK(CHECK_TRY_ERROR((*stream) - .memcpy(extra->data_device[g_main_device_index], + .memcpy(extra->data_device[g_main_device], tensor->data, ggml_nbytes(tensor)) .wait())); } @@ -14250,21 +16121,17 @@ void ggml_sycl_assign_buffers_force_inplace(struct ggml_tensor * tensor) { } void ggml_sycl_set_main_device(const int main_device) try { + if (g_main_device == main_device) return; + check_allow_gpu_index(main_device); + g_main_device = main_device; + g_main_device_id = g_sycl_gpu_mgr->gpus[main_device]; - if (main_device >= g_all_sycl_device_count) { - fprintf(stderr, "warning: cannot set main_device=%d because there are only %d devices. Using device %d instead.\n", - main_device, g_all_sycl_device_count, g_main_device); - return; - } - - if (g_main_device != main_device && g_device_count >= 1) { - g_main_device = main_device; - g_main_device_index = get_device_index_by_id(g_main_device); + if (g_ggml_sycl_debug) { dpct::device_info prop; SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( - prop, dpct::dev_mgr::instance().get_device(g_main_device)))); + prop, dpct::dev_mgr::instance().get_device(g_main_device_id)))); fprintf(stderr, "Using device %d (%s) as main device\n", - g_main_device, prop.get_name()); + g_main_device_id, prop.get_name()); } } catch (sycl::exception const &exc) { @@ -14287,7 +16154,7 @@ void ggml_sycl_free_scratch() try { return; } ggml_sycl_set_device(g_main_device); - const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0]; + const dpct::queue_ptr stream = g_syclStreams[g_main_device][0]; SYCL_CHECK(CHECK_TRY_ERROR( sycl::free(g_scratch_buffer, *stream))); @@ -14359,6 +16226,12 @@ bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_ case GGML_UNARY_OP_RELU: func = ggml_sycl_relu; break; + case GGML_UNARY_OP_HARDSIGMOID: + func = ggml_sycl_hardsigmoid; + break; + case GGML_UNARY_OP_HARDSWISH: + func = ggml_sycl_hardswish; + break; default: return false; } @@ -14433,6 +16306,9 @@ bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_ case GGML_OP_IM2COL: func = ggml_sycl_im2col; break; + case GGML_OP_POOL_2D: + func = ggml_sycl_pool2d; + break; case GGML_OP_SUM_ROWS: func = ggml_sycl_sum_rows; break; @@ -14458,27 +16334,15 @@ bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_ } GGML_API GGML_CALL void ggml_sycl_get_gpu_list(int *id_list, int max_len) try { - int max_compute_units = -1; - for(int i=0;igpus.size();i++){ + if (i>=max_len) break; + id_list[i] = g_sycl_gpu_mgr->gpus[i]; } return; } @@ -14505,8 +16369,9 @@ catch (sycl::exception const &exc) { GGML_API GGML_CALL void ggml_sycl_get_device_description(int device, char *description, size_t description_size) try { dpct::device_info prop; + int device_id = g_sycl_gpu_mgr->gpus[device]; SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( - prop, dpct::dev_mgr::instance().get_device(device)))); + prop, dpct::dev_mgr::instance().get_device(device_id)))); snprintf(description, description_size, "%s", prop.get_name()); } catch (sycl::exception const &exc) { @@ -14515,17 +16380,36 @@ catch (sycl::exception const &exc) { std::exit(1); } +GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, + size_t *total) try { + ggml_sycl_set_device(device); + + /* + DPCT1009:218: SYCL uses exceptions to report errors and does not use the + error codes. The original code was commented out and a warning string was + inserted. You need to rewrite this code. + */ + /* + DPCT1106:217: 'cudaMemGetInfo' was migrated with the Intel extensions for + device information which may not be supported by all compilers or runtimes. + You may need to adjust the code. + */ + int device_id = g_sycl_gpu_mgr->gpus[device]; + SYCL_CHECK(CHECK_TRY_ERROR( + dpct::dev_mgr::instance().get_device(device_id).get_memory_info(*free, *total))); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + //////////////////////////////////////////////////////////////////////////////// // backend interface #define UNUSED GGML_UNUSED -struct ggml_backend_sycl_context { - int device; - std::string name; -}; - // sycl buffer struct ggml_backend_sycl_buffer_context { @@ -14535,7 +16419,12 @@ struct ggml_backend_sycl_buffer_context { size_t temp_tensor_extra_index = 0; std::string name; - ggml_backend_sycl_buffer_context(int device, void * dev_ptr) : device(device), dev_ptr(dev_ptr) {} + ggml_backend_sycl_buffer_context(int device, void * dev_ptr) : + device(device), dev_ptr(dev_ptr) { + check_allow_gpu_index(device); + int id = g_sycl_gpu_mgr->gpus[device]; + name = (GGML_SYCL_NAME + std::to_string(id)); + } ~ ggml_backend_sycl_buffer_context() { delete[] temp_tensor_extras; @@ -14566,10 +16455,9 @@ GGML_CALL static bool ggml_backend_buffer_is_sycl(ggml_backend_buffer_t buffer) static void ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try { - ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; + ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; ggml_sycl_set_device(ctx->device); - int device_index = get_device_index_by_id(ctx->device); - const dpct::queue_ptr stream = g_syclStreams[device_index][0]; + const dpct::queue_ptr stream = g_syclStreams[ctx->device][0]; SYCL_CHECK( CHECK_TRY_ERROR(sycl::free(ctx->dev_ptr, *stream))); @@ -14582,13 +16470,14 @@ catch (sycl::exception const &exc) { } static void * ggml_backend_sycl_buffer_get_base(ggml_backend_buffer_t buffer) { - ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; + ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; return ctx->dev_ptr; } -static void ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer, - ggml_tensor *tensor) try { - ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; +GGML_CALL static void +ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer, + ggml_tensor *tensor) try { + ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context; if (tensor->view_src != NULL && tensor->view_offs == 0) { assert(tensor->view_src->buffer->buft == buffer->buft); @@ -14600,27 +16489,20 @@ static void ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor_extra_gpu * extra = ctx->ggml_sycl_alloc_temp_tensor_extra(); extra->data_device[ctx->device] = tensor->data; - tensor->backend = GGML_BACKEND_TYPE_GPU; tensor->extra = extra; if (ggml_is_quantized(tensor->type)) { // initialize padding to 0 to avoid possible NaN values - int64_t row_low = 0; - int64_t row_high = ggml_nrows(tensor); - int64_t nrows_split = row_high - row_low; - - size_t original_size = ggml_nbytes_split(tensor, nrows_split); + size_t original_size = ggml_nbytes(tensor); size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor); if (padded_size > original_size && tensor->view_src == nullptr) { SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[ctx->device][0]->memset( (char *)tensor->data + original_size, 0, - padded_size - original_size))); + padded_size - original_size).wait())); } } - - UNUSED(buffer); } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ @@ -14634,13 +16516,12 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer, size_t size) try { GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU); - ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; + ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; ggml_sycl_set_device(ctx->device); - int device_index = get_device_index_by_id(ctx->device); - const dpct::queue_ptr stream = g_syclStreams[device_index][0]; + const dpct::queue_ptr stream = g_syclStreams[ctx->device][0]; SYCL_CHECK( - CHECK_TRY_ERROR(dpct::get_current_device().queues_wait_and_throw())); + CHECK_TRY_ERROR(dpct::dev_mgr::instance().get_device(ctx->device).queues_wait_and_throw())); SYCL_CHECK( CHECK_TRY_ERROR((*stream) @@ -14659,14 +16540,13 @@ static void ggml_backend_sycl_buffer_get_tensor(ggml_backend_buffer_t buffer, size_t size) try { GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU); - ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; + ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; ggml_sycl_set_device(ctx->device); - int device_index = get_device_index_by_id(ctx->device); - const dpct::queue_ptr stream = g_syclStreams[device_index][0]; + const dpct::queue_ptr stream = g_syclStreams[ctx->device][0]; SYCL_CHECK( - CHECK_TRY_ERROR(dpct::get_current_device().queues_wait_and_throw())); + CHECK_TRY_ERROR(dpct::dev_mgr::instance().get_device(ctx->device).queues_wait_and_throw())); SYCL_CHECK(CHECK_TRY_ERROR( (*stream) @@ -14679,13 +16559,73 @@ catch (sycl::exception const &exc) { std::exit(1); } +GGML_CALL static bool +ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer, + const ggml_tensor *src, + ggml_tensor *dst) try { + if (ggml_backend_buffer_is_sycl(src->buffer)) { + ggml_backend_sycl_buffer_context * src_ctx = (ggml_backend_sycl_buffer_context *)src->buffer->context; + ggml_backend_sycl_buffer_context * dst_ctx = (ggml_backend_sycl_buffer_context *)buffer->context; + + ggml_sycl_set_device(src_ctx->device); + /* + DPCT1009:198: SYCL uses exceptions to report errors and does not use the + error codes. The original code was commented out and a warning string + was inserted. You need to rewrite this code. + */ + SYCL_CHECK(CHECK_TRY_ERROR( + dpct::dev_mgr::instance().get_device(src_ctx->device).queues_wait_and_throw())); + ggml_sycl_set_device(dst_ctx->device); + /* + DPCT1009:199: SYCL uses exceptions to report errors and does not use the + error codes. The original code was commented out and a warning string + was inserted. You need to rewrite this code. + */ + SYCL_CHECK(CHECK_TRY_ERROR( + dpct::dev_mgr::instance().get_device(dst_ctx->device).queues_wait_and_throw())); + /* + DPCT1009:200: SYCL uses exceptions to report errors and does not use the + error codes. The original code was commented out and a warning string + was inserted. You need to rewrite this code. + */ + + dpct::queue_ptr stream_dst = g_syclStreams[dst_ctx->device][0]; + dpct::queue_ptr stream_src = g_syclStreams[src_ctx->device][0]; + size_t size = ggml_nbytes(src); + + //todo. it's dirty solutino to walkaroud known issue:device2device cross GPUs. + dev2dev_memcpy(*stream_dst, *stream_src, dst->data, src->data, size); + +//todo, it's known issue：error in device2device cross GPUs. reused when the issue is fixed. DON"T remove +#if 0 + SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy( + (char *)dst->data, (const char *)src->data, size).wait())); + + /* + DPCT1009:201: SYCL uses exceptions to report errors and does not use the + error codes. The original code was commented out and a warning string + was inserted. You need to rewrite this code. + */ + SYCL_CHECK(CHECK_TRY_ERROR( + dpct::dev_mgr::instance().get_device(dst_ctx->device).queues_wait_and_throw())); +#endif + return true; + } + return false; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + + static void ggml_backend_sycl_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) try { ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; ggml_sycl_set_device(ctx->device); - int device_index = get_device_index_by_id(ctx->device); - const dpct::queue_ptr stream = g_syclStreams[device_index][0]; + const dpct::queue_ptr stream = g_syclStreams[ctx->device][0]; SYCL_CHECK( CHECK_TRY_ERROR(dpct::get_current_device().queues_wait_and_throw())); @@ -14706,7 +16646,7 @@ static struct ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = { /* .init_tensor = */ ggml_backend_sycl_buffer_init_tensor, /* .set_tensor = */ ggml_backend_sycl_buffer_set_tensor, /* .get_tensor = */ ggml_backend_sycl_buffer_get_tensor, - /* .cpy_tensor = */ NULL, + /* .cpy_tensor = */ ggml_backend_sycl_buffer_cpy_tensor, /* .clear = */ ggml_backend_sycl_buffer_clear, /* .reset = */ NULL, }; @@ -14717,29 +16657,28 @@ struct ggml_backend_sycl_buffer_type_context { std::string name; }; +struct ggml_backend_sycl_context { + int device; + std::string name; +}; + GGML_CALL static const char * ggml_backend_sycl_buffer_type_name(ggml_backend_buffer_type_t buft) { ggml_backend_sycl_buffer_type_context * ctx = (ggml_backend_sycl_buffer_type_context *)buft->context; return ctx->name.c_str(); } - -static ggml_backend_buffer_t +GGML_CALL static ggml_backend_buffer_t ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) try { ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context; - int device = (int) buft_ctx->device; - - ggml_sycl_set_device(device); - int device_index = get_device_index_by_id(device); - const dpct::queue_ptr stream = g_syclStreams[device_index][0]; + ggml_sycl_set_device(buft_ctx->device); + const dpct::queue_ptr stream = g_syclStreams[buft_ctx->device][0]; size = std::max(size, (size_t)1); // syclMalloc returns null for size 0 void * dev_ptr; SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)sycl::malloc_device( size, *stream))); - - ggml_backend_sycl_buffer_context * ctx = new ggml_backend_sycl_buffer_context(device, dev_ptr); - + ggml_backend_sycl_buffer_context * ctx = new ggml_backend_sycl_buffer_context(buft_ctx->device, dev_ptr); return ggml_backend_buffer_init(buft, ggml_backend_sycl_buffer_interface, ctx, size); } catch (sycl::exception const &exc) { @@ -14748,9 +16687,8 @@ catch (sycl::exception const &exc) { std::exit(1); } -static size_t ggml_backend_sycl_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { +GGML_CALL static size_t ggml_backend_sycl_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { return 128; - UNUSED(buft); } @@ -14760,13 +16698,8 @@ static size_t ggml_backend_sycl_buffer_type_get_max_size(ggml_backend_buffer_typ UNUSED(buft); } -static size_t ggml_backend_sycl_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { - int64_t row_low = 0; - int64_t row_high = ggml_nrows(tensor); - int64_t nrows_split = row_high - row_low; - - size_t size = ggml_nbytes_split(tensor, nrows_split); - +GGML_CALL static size_t ggml_backend_sycl_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { + size_t size = ggml_nbytes(tensor); int64_t ne0 = tensor->ne[0]; if (ggml_is_quantized(tensor->type)) { @@ -14780,10 +16713,13 @@ static size_t ggml_backend_sycl_buffer_type_get_alloc_size(ggml_backend_buffer_t UNUSED(buft); } -static bool ggml_backend_sycl_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) { - return ggml_backend_is_sycl(backend); - - UNUSED(buft); +GGML_CALL static bool ggml_backend_sycl_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) { + if (!ggml_backend_is_sycl(backend)) { + return false; + } + ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context; + ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; + return buft_ctx->device == sycl_ctx->device; } static ggml_backend_buffer_type_i ggml_backend_sycl_buffer_type_interface = { @@ -14802,10 +16738,10 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) { static bool ggml_backend_sycl_buffer_type_initialized = false; if (!ggml_backend_sycl_buffer_type_initialized) { - for (int i = 0; i < GGML_SYCL_MAX_DEVICES; i++) { + for (int i = 0; i < g_device_count; i++) { ggml_backend_sycl_buffer_types[i] = { /* .iface = */ ggml_backend_sycl_buffer_type_interface, - /* .context = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(i)}, + /* .context = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(g_sycl_gpu_mgr->gpus[i])}, }; } ggml_backend_sycl_buffer_type_initialized = true; @@ -14814,6 +16750,391 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) { return &ggml_backend_sycl_buffer_types[device]; } +// sycl split buffer type +static void get_row_split(int64_t * row_low, int64_t * row_high, const ggml_tensor * tensor, const std::array & tensor_split, int id) { + const int64_t nrows = ggml_nrows(tensor); + const int64_t rounding = get_row_rounding(tensor->type, tensor_split); + + *row_low = id == 0 ? 0 : nrows*tensor_split[id]; + *row_low -= *row_low % rounding; + if (id == g_device_count - 1) { + *row_high = nrows; + } else { + *row_high = nrows*tensor_split[id + 1]; + *row_high -= *row_high % rounding; + } +} + +struct ggml_backend_sycl_split_buffer_context { + ~ggml_backend_sycl_split_buffer_context() try { + for (ggml_tensor_extra_gpu * extra : tensor_extras) { + for (int i = 0; i < g_device_count; ++i) { + // int id = g_sycl_gpu_mgr->gpus[i]; + for (int64_t is = 0; is < MAX_STREAMS; ++is) { + if (extra->events[i][is] != nullptr) { + /* + DPCT1009:206: SYCL uses exceptions to report errors and + does not use the error codes. The original code was + commented out and a warning string was inserted. You + need to rewrite this code. + */ + SYCL_CHECK(CHECK_TRY_ERROR( + dpct::destroy_event(extra->events[i][is]))); + } + } + if (extra->data_device[i] != nullptr) { + /* + DPCT1009:207: SYCL uses exceptions to report errors and does + not use the error codes. The original code was commented out + and a warning string was inserted. You need to rewrite this + code. + */ + ggml_sycl_set_device(i); + SYCL_CHECK(CHECK_TRY_ERROR(sycl::free( + extra->data_device[i], *g_syclStreams[i][0]))); + } + } + delete extra; + } + } + catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); + } + + std::vector tensor_extras; +}; + +GGML_CALL static const char * ggml_backend_sycl_split_buffer_get_name(ggml_backend_buffer_t buffer) { + return GGML_SYCL_NAME "_Split"; + + UNUSED(buffer); +} + +// unused at the moment +//static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer) { +// return buffer->iface.get_name == ggml_backend_sycl_split_buffer_get_name; +//} + +GGML_CALL static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) { + ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context; + delete ctx; +} + +GGML_CALL static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) { + // the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced + return (void *)0x1000; + + UNUSED(buffer); +} + +GGML_CALL static void +ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer, + ggml_tensor *tensor) try { + GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported + + ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context; + ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *)buffer->buft->context; + + const int64_t ne0 = tensor->ne[0]; + + ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{}; + + ctx->tensor_extras.push_back(extra); + + for (int i = 0; i < g_device_count; ++i) { + // int id = g_sycl_gpu_mgr->gpus[i]; + int64_t row_low, row_high; + get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, i); + + int64_t nrows_split = row_high - row_low; + if (nrows_split == 0) { + continue; + } + + size_t size = ggml_nbytes_split(tensor, nrows_split); + const size_t original_size = size; + + // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses + if (ne0 % MATRIX_ROW_PADDING != 0) { + size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); + } + + // FIXME: do not crash if cudaMalloc fails + // currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first + ggml_sycl_set_device(i); + char * buf; + /* + DPCT1009:208: SYCL uses exceptions to report errors and does not use the + error codes. The original code was commented out and a warning string + was inserted. You need to rewrite this code. + */ + SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)sycl::malloc_device( + size, *g_syclStreams[i][0]))); + + // set padding to 0 to avoid possible NaN values + if (size > original_size) { + /* + DPCT1009:209: SYCL uses exceptions to report errors and does not use + the error codes. The original code was commented out and a warning + string was inserted. You need to rewrite this code. + */ + SYCL_CHECK(CHECK_TRY_ERROR( + (*g_syclStreams[i][0]) + .memset(buf + original_size, 0, size - original_size) + .wait())); + } + + extra->data_device[i] = buf; + + for (int64_t is = 0; is < MAX_STREAMS; ++is) { + /* + DPCT1009:210: SYCL uses exceptions to report errors and does not use + the error codes. The original code was commented out and a warning + string was inserted. You need to rewrite this code. + */ + SYCL_CHECK( + CHECK_TRY_ERROR(extra->events[i][is] = new sycl::event())); + } + } + tensor->backend = GGML_BACKEND_TYPE_GPU_SPLIT; + tensor->extra = extra; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +GGML_CALL static void +ggml_backend_sycl_split_buffer_set_tensor(ggml_backend_buffer_t buffer, + ggml_tensor *tensor, const void *data, + size_t offset, size_t size) try { + // split tensors must always be set in their entirety at once + GGML_ASSERT(offset == 0); + GGML_ASSERT(size == ggml_nbytes(tensor)); + + ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *)buffer->buft->context; + + const int64_t ne0 = tensor->ne[0]; + const size_t nb1 = tensor->nb[1]; + ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra; + + for (int i = 0; i < g_device_count; ++i) { + // int id = g_sycl_gpu_mgr->gpus[i]; + int64_t row_low, row_high; + get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, i); + + int64_t nrows_split = row_high - row_low; + if (nrows_split == 0) { + continue; + } + + const size_t offset_split = row_low*nb1; + size_t size = ggml_nbytes_split(tensor, nrows_split); + const size_t original_size = size; + + // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses + if (ne0 % MATRIX_ROW_PADDING != 0) { + size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); + } + + const char * buf_host = (const char *)data + offset_split; + /* + DPCT1009:211: SYCL uses exceptions to report errors and does not use the + error codes. The original code was commented out and a warning string + was inserted. You need to rewrite this code. + */ + ggml_sycl_set_device(i); + SYCL_CHECK(CHECK_TRY_ERROR( + (*g_syclStreams[i][0]) + .memcpy(extra->data_device[i], buf_host, original_size) + .wait())); + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +GGML_CALL static void +ggml_backend_sycl_split_buffer_get_tensor(ggml_backend_buffer_t buffer, + const ggml_tensor *tensor, void *data, + size_t offset, size_t size) try { + // split tensors must always be set in their entirety at once + GGML_ASSERT(offset == 0); + GGML_ASSERT(size == ggml_nbytes(tensor)); + + ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *)buffer->buft->context; + + const int64_t ne0 = tensor->ne[0]; + const size_t nb1 = tensor->nb[1]; + ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra; + + for (int i = 0; i < g_device_count; ++i) { + // int id = g_sycl_gpu_mgr->gpus[i]; + int64_t row_low, row_high; + get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, i); + + int64_t nrows_split = row_high - row_low; + if (nrows_split == 0) { + continue; + } + + const size_t offset_split = row_low*nb1; + size_t size = ggml_nbytes_split(tensor, nrows_split); + const size_t original_size = size; + + // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses + if (ne0 % MATRIX_ROW_PADDING != 0) { + size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); + } + + char * buf_host = (char *)data + offset_split; + /* + DPCT1009:212: SYCL uses exceptions to report errors and does not use the + error codes. The original code was commented out and a warning string + was inserted. You need to rewrite this code. + */ + ggml_sycl_set_device(i); + SYCL_CHECK(CHECK_TRY_ERROR( + (*g_syclStreams[i][0]) + .memcpy(buf_host, extra->data_device[i], original_size) + .wait())); + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +GGML_CALL static void ggml_backend_sycl_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { + UNUSED(buffer); + UNUSED(value); +} + +static struct ggml_backend_buffer_i ggml_backend_sycl_split_buffer_interface = { + /* .get_name = */ ggml_backend_sycl_split_buffer_get_name, + /* .free_buffer = */ ggml_backend_sycl_split_buffer_free_buffer, + /* .get_base = */ ggml_backend_sycl_split_buffer_get_base, + /* .init_tensor = */ ggml_backend_sycl_split_buffer_init_tensor, + /* .set_tensor = */ ggml_backend_sycl_split_buffer_set_tensor, + /* .get_tensor = */ ggml_backend_sycl_split_buffer_get_tensor, + /* .cpy_tensor = */ NULL, + /* .clear = */ ggml_backend_sycl_split_buffer_clear, + /* .reset = */ NULL, +}; + +GGML_CALL static const char * ggml_backend_sycl_split_buffer_type_name(ggml_backend_buffer_type_t buft) { + return GGML_SYCL_NAME "_Split"; + + UNUSED(buft); +} + +GGML_CALL static ggml_backend_buffer_t ggml_backend_sycl_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point + // instead, we allocate them for each tensor separately in init_tensor + // however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated, + // as returned by get_alloc_size. this limit is enforced during tensor allocation by ggml-alloc, so it must be correct. + ggml_backend_sycl_split_buffer_context * ctx = new ggml_backend_sycl_split_buffer_context(); + + return ggml_backend_buffer_init(buft, ggml_backend_sycl_split_buffer_interface, ctx, size); +} + +GGML_CALL static size_t ggml_backend_sycl_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { + return 128; + UNUSED(buft); +} + +GGML_CALL static size_t ggml_backend_sycl_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { + ggml_backend_sycl_split_buffer_type_context * ctx = (ggml_backend_sycl_split_buffer_type_context *)buft->context; + + size_t total_size = 0; + + const int64_t ne0 = tensor->ne[0]; + + for (int i = 0; i < g_device_count; ++i) { + // int id = g_sycl_gpu_mgr->gpus[i]; + int64_t row_low, row_high; + get_row_split(&row_low, &row_high, tensor, ctx->tensor_split, i); + + int64_t nrows_split = row_high - row_low; + if (nrows_split == 0) { + continue; + } + + total_size += ggml_nbytes_split(tensor, nrows_split); + + // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses + if (ne0 % MATRIX_ROW_PADDING != 0) { + total_size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); + } + } + + return total_size; +} + +GGML_CALL static bool ggml_backend_sycl_split_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) { + return ggml_backend_is_sycl(backend); + + UNUSED(buft); +} + +GGML_CALL static bool ggml_backend_sycl_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) { + return false; + + UNUSED(buft); +} + +static ggml_backend_buffer_type_i ggml_backend_sycl_split_buffer_type_interface = { + /* .get_name = */ ggml_backend_sycl_split_buffer_type_name, + /* .alloc_buffer = */ ggml_backend_sycl_split_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_sycl_split_buffer_type_get_alignment, + /* .get_max_size = */ NULL, // defaults to SIZE_MAX + /* .get_alloc_size = */ ggml_backend_sycl_split_buffer_type_get_alloc_size, + /* .supports_backend = */ ggml_backend_sycl_split_buffer_type_supports_backend, + /* .is_host = */ ggml_backend_sycl_split_buffer_type_is_host, +}; + +GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split) { + // FIXME: this is not thread safe + static std::map, struct ggml_backend_buffer_type> buft_map; + + std::array tensor_split_arr = {}; + + bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + GGML_SYCL_MAX_DEVICES, [](float x) { return x == 0.0f; }); + if (all_zero) { + tensor_split_arr = g_default_tensor_split; + } else { + float split_sum = 0.0f; + for (int i = 0; i < g_device_count; ++i) { + // int id = g_sycl_gpu_mgr->gpus[i]; + tensor_split_arr[i] = split_sum; + split_sum += tensor_split[i]; + } + for (int i = 0; i < g_device_count; ++i) { + // int id = g_sycl_gpu_mgr->gpus[i]; + tensor_split_arr[i] /= split_sum; + } + } + + auto it = buft_map.find(tensor_split_arr); + if (it != buft_map.end()) { + return &it->second; + } + + struct ggml_backend_buffer_type buft { + /* .iface = */ ggml_backend_sycl_split_buffer_type_interface, + /* .context = */ new ggml_backend_sycl_split_buffer_type_context{tensor_split_arr}, + }; + + auto result = buft_map.emplace(tensor_split_arr, buft); + return &result.first->second; +} + // host buffer type GGML_CALL static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_type_t buft) { @@ -14843,6 +17164,7 @@ static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggm // FIXME: this is a hack to avoid having to implement a new buffer type ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size); buffer->buft = buft; + buffer->iface.get_name = ggml_backend_sycl_host_buffer_name; buffer->iface.free_buffer = ggml_backend_sycl_host_buffer_free_buffer; return buffer; @@ -14867,36 +17189,35 @@ ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() { // backend -static const char * ggml_backend_sycl_name(ggml_backend_t backend) { - return GGML_SYCL_NAME; +GGML_CALL static const char * ggml_backend_sycl_name(ggml_backend_t backend) { - UNUSED(backend); + ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; + + return sycl_ctx->name.c_str(); } -static void ggml_backend_sycl_free(ggml_backend_t backend) { +GGML_CALL static void ggml_backend_sycl_free(ggml_backend_t backend) { ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; delete sycl_ctx; delete backend; } -static ggml_backend_buffer_type_t ggml_backend_sycl_get_default_buffer_type(ggml_backend_t backend) { - ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; +GGML_CALL static ggml_backend_buffer_type_t ggml_backend_sycl_get_default_buffer_type(ggml_backend_t backend) { + ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; return ggml_backend_sycl_buffer_type(sycl_ctx->device); } -static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend, +GGML_CALL static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend, ggml_tensor *tensor, const void *data, size_t offset, size_t size) try { ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; - GGML_ASSERT(tensor->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type"); GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU); - SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->memcpy( - (char *)tensor->data + offset, data, size))); + (char *)tensor->data + offset, data, size).wait())); } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ @@ -14904,17 +17225,38 @@ catch (sycl::exception const &exc) { std::exit(1); } -static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend, +GGML_CALL static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend, const ggml_tensor *tensor, void *data, size_t offset, size_t size) try { ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; - GGML_ASSERT(tensor->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type"); GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU); - SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->memcpy( - data, (const char *)tensor->data + offset, size))); + data, (const char *)tensor->data + offset, size).wait())); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +GGML_CALL static bool ggml_backend_sycl_cpy_tensor_async(ggml_backend_t backend, + const ggml_tensor *src, + ggml_tensor *dst) try { + ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; + if (dst->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && ggml_backend_buffer_is_sycl(src->buffer)) { + /* + DPCT1009:215: SYCL uses exceptions to report errors and does not use the + error codes. The original code was commented out and a warning string + was inserted. You need to rewrite this code. + */ + SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->memcpy( + dst->data, src->data, ggml_nbytes(dst)).wait())); + return true; + } + + return false; } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ @@ -14924,7 +17266,6 @@ catch (sycl::exception const &exc) { static void ggml_backend_sycl_synchronize(ggml_backend_t backend) try { ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; - SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->wait())); UNUSED(backend); @@ -14935,32 +17276,8 @@ catch (sycl::exception const &exc) { std::exit(1); } -static ggml_backend_graph_plan_t ggml_backend_sycl_graph_plan_create(ggml_backend_t backend, const ggml_cgraph * cgraph) { - GGML_ASSERT(!"not implemented"); - - return nullptr; - - UNUSED(backend); - UNUSED(cgraph); -} - -static void ggml_backend_sycl_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { - GGML_ASSERT(!"not implemented"); - - UNUSED(backend); - UNUSED(plan); -} - -static void ggml_backend_sycl_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { - GGML_ASSERT(!"not implemented"); - - UNUSED(backend); - UNUSED(plan); -} - -static bool ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { +GGML_CALL static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; - ggml_sycl_set_main_device(sycl_ctx->device); ggml_compute_params params = {}; @@ -14968,63 +17285,41 @@ static bool ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph params.ith = 0; for (int i = 0; i < cgraph->n_nodes; i++) { ggml_tensor * node = cgraph->nodes[i]; - - if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE) + if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) { continue; - - assert(node->backend == GGML_BACKEND_TYPE_GPU); + } +#ifndef NDEBUG + assert(node->backend == GGML_BACKEND_TYPE_GPU || node->backend == GGML_BACKEND_TYPE_GPU_SPLIT); assert(node->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device)); assert(node->extra != nullptr); for (int j = 0; j < GGML_MAX_SRC; j++) { if (node->src[j] != nullptr) { - assert(node->src[j]->backend == GGML_BACKEND_TYPE_GPU); + assert(node->src[j]->backend == GGML_BACKEND_TYPE_GPU || node->src[j]->backend == GGML_BACKEND_TYPE_GPU_SPLIT); assert(node->src[j]->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device)); assert(node->src[j]->extra != nullptr); } } - +#endif bool ok = ggml_sycl_compute_forward(¶ms, node); if (!ok) { fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op)); } GGML_ASSERT(ok); - -#if 0 - if (node->type == GGML_TYPE_F32) { - syclDeviceSynchronize(); - std::vector tmp(ggml_nelements(node), 0.0f); - syclMemcpy(tmp.data(), node->data, ggml_nelements(node)*sizeof(float), syclMemcpyDeviceToHost); - printf("\n%s (%s) (%s %s) (%s %s): ", node->name, ggml_op_name(node->op), - ggml_type_name(node->src[0]->type), - node->src[1] ? ggml_type_name(node->src[1]->type) : "none", - node->src[0]->name, - node->src[1] ? node->src[1]->name : "none"); - double sum = 0.0; - double sq_sum = 0.0; - for (int i = 0; i < ggml_nelements(node); i++) { - printf("%f ", tmp[i]); - sum += tmp[i]; - sq_sum += tmp[i]*tmp[i]; - } - printf("\n"); - printf("sum: %f, ", sum); - printf("sq_sum: %f\n", sq_sum); - } -#endif } - UNUSED(backend); - return true; + return GGML_STATUS_SUCCESS; } -static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_tensor * op) { +GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_tensor * op) { switch (op->op) { case GGML_OP_UNARY: switch (ggml_get_unary_op(op)) { case GGML_UNARY_OP_GELU: case GGML_UNARY_OP_SILU: case GGML_UNARY_OP_RELU: + case GGML_UNARY_OP_HARDSIGMOID: + case GGML_UNARY_OP_HARDSWISH: case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_TANH: return true; @@ -15047,20 +17342,12 @@ static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_ten if (a->ne[3] != b->ne[3]) { return false; } - - if (a->type == GGML_TYPE_IQ1_S) { + ggml_type a_type = a->type; + if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS || + a_type == GGML_TYPE_IQ1_S || a_type == GGML_TYPE_IQ4_NL || a_type == GGML_TYPE_IQ3_S || + a_type == GGML_TYPE_IQ2_S || a_type == GGML_TYPE_IQ4_XS) { return false; } - if (a->type == GGML_TYPE_IQ3_XXS) { - return false; - } - if (a->type == GGML_TYPE_IQ2_XXS) { - return false; - } - if (a->type == GGML_TYPE_IQ2_XS) { - return false; - } - return true; } break; case GGML_OP_GET_ROWS: @@ -15100,25 +17387,24 @@ static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_ten if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { return true; } + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { + return true; + } return false; } break; case GGML_OP_CONCAT: { ggml_type src0_type = op->src[0]->type; - if (src0_type == GGML_TYPE_F32) { - return true; - } else { - return false; - } + return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16; } break; + case GGML_OP_DUP: case GGML_OP_NONE: case GGML_OP_RESHAPE: + case GGML_OP_REPEAT: case GGML_OP_VIEW: case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: case GGML_OP_NORM: - case GGML_OP_REPEAT: - case GGML_OP_DUP: case GGML_OP_ADD: case GGML_OP_MUL: case GGML_OP_DIV: @@ -15132,6 +17418,7 @@ static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_ten case GGML_OP_ROPE: case GGML_OP_ALIBI: case GGML_OP_IM2COL: + case GGML_OP_POOL_2D: case GGML_OP_SUM_ROWS: case GGML_OP_ARGSORT: case GGML_OP_ACC: @@ -15153,11 +17440,11 @@ static ggml_backend_i ggml_backend_sycl_interface = { /* .get_default_buffer_type = */ ggml_backend_sycl_get_default_buffer_type, /* .set_tensor_async = */ ggml_backend_sycl_set_tensor_async, /* .get_tensor_async = */ ggml_backend_sycl_get_tensor_async, - /* .cpy_tensor_async = */ NULL, + /* .cpy_tensor_async = */ ggml_backend_sycl_cpy_tensor_async, /* .synchronize = */ ggml_backend_sycl_synchronize, - /* .graph_plan_create = */ ggml_backend_sycl_graph_plan_create, - /* .graph_plan_free = */ ggml_backend_sycl_graph_plan_free, - /* .graph_plan_compute = */ ggml_backend_sycl_graph_plan_compute, + /* .graph_plan_create = */ NULL, + /* .graph_plan_free = */ NULL, + /* .graph_plan_compute = */ NULL, /* .graph_compute = */ ggml_backend_sycl_graph_compute, /* .supports_op = */ ggml_backend_sycl_supports_op, }; @@ -15167,20 +17454,17 @@ static ggml_guid_t ggml_backend_sycl_guid() { return &guid; } -ggml_backend_t ggml_backend_sycl_init(int device) { +GGML_CALL ggml_backend_t ggml_backend_sycl_init(int device) { ggml_init_sycl(); // TODO: remove from ggml.c - if (device < 0 || device >= ggml_sycl_get_device_count()) { - fprintf(stderr, "%s: error: invalid device %d\n", __func__, device); - return nullptr; - } + check_allow_gpu_index(device); // not strictly necessary, but it may reduce the overhead of the first graph_compute ggml_sycl_set_main_device(device); - + int id = g_sycl_gpu_mgr->gpus[device]; ggml_backend_sycl_context * ctx = new ggml_backend_sycl_context { /* .device = */ device, - /* .name = */ GGML_SYCL_NAME + std::to_string(device), + /* .name = */ GGML_SYCL_NAME + std::to_string(id), }; ggml_backend_t sycl_backend = new ggml_backend { @@ -15196,22 +17480,33 @@ bool ggml_backend_is_sycl(ggml_backend_t backend) { return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_sycl_guid()); } -static ggml_backend_t ggml_backend_reg_sycl_init(const char * params, void * user_data) { +GGML_CALL int ggml_backend_sycl_get_device_count() { + if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr(); + return g_sycl_gpu_mgr->get_gpu_count(); +} + +GGML_CALL static ggml_backend_t ggml_backend_reg_sycl_init(const char * params, void * user_data) { ggml_backend_t sycl_backend = ggml_backend_sycl_init((int) (intptr_t) user_data); return sycl_backend; UNUSED(params); } +GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id) { + return g_sycl_gpu_mgr->get_index(device_id); +} + extern "C" int ggml_backend_sycl_reg_devices(); int ggml_backend_sycl_reg_devices() { - int device_count = ggml_sycl_get_device_count(); - - for (int i = 0; i < device_count; i++) { + if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr(); + g_device_count = g_sycl_gpu_mgr->get_gpu_count(); + assert(g_device_count>0); + for (int i = 0; i < g_device_count; i++) { + int id = g_sycl_gpu_mgr->gpus[i]; char name[128]; - snprintf(name, sizeof(name), "%s%d", GGML_SYCL_NAME, i); + snprintf(name, sizeof(name), "%s%d", GGML_SYCL_NAME, id); ggml_backend_register(name, ggml_backend_reg_sycl_init, ggml_backend_sycl_buffer_type(i), (void *) (intptr_t) i); } - return device_count; + return g_device_count; } diff --git a/ggml-sycl.h b/ggml-sycl.h index 891f2d0..bf5b11b 100644 --- a/ggml-sycl.h +++ b/ggml-sycl.h @@ -24,6 +24,11 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void); GGML_API void ggml_backend_sycl_print_sycl_devices(void); GGML_API GGML_CALL void ggml_sycl_get_gpu_list(int *id_list, int max_len); GGML_API GGML_CALL void ggml_sycl_get_device_description(int device, char *description, size_t description_size); +GGML_API GGML_CALL int ggml_backend_sycl_get_device_count(); +GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split); +GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total); +GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id); + #ifdef __cplusplus } #endif diff --git a/ggml-vulkan.cpp b/ggml-vulkan.cpp index 314e3d7..5a1b3f4 100644 --- a/ggml-vulkan.cpp +++ b/ggml-vulkan.cpp @@ -69,6 +69,33 @@ struct vk_queue { vk::PipelineStageFlags stage_flags; }; +struct vk_pipeline_struct { + std::string name; + vk::ShaderModule shader_module; + vk::DescriptorSetLayout dsl; + std::vector descriptor_pools; + std::vector descriptor_sets; + uint32_t descriptor_set_idx; + vk::PipelineLayout layout; + vk::Pipeline pipeline; + uint32_t push_constant_size; + uint32_t parameter_count; + std::array wg_denoms; + uint32_t align; +}; + +typedef std::shared_ptr vk_pipeline; +typedef std::weak_ptr vk_pipeline_ref; + +static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline); + +struct vk_matmul_pipeline_struct { + vk_pipeline l, m, s; + vk_pipeline a_l, a_m, a_s; +}; + +typedef std::shared_ptr vk_matmul_pipeline; + struct vk_device { vk::PhysicalDevice physical_device; vk::PhysicalDeviceProperties properties; @@ -84,10 +111,61 @@ struct vk_device { uint32_t subgroup_size; bool uma; + bool initialized; + size_t idx; + + vk_matmul_pipeline pipeline_matmul_f32; + vk_matmul_pipeline pipeline_matmul_f16; + vk_matmul_pipeline pipeline_matmul_f16_f32; + vk_pipeline pipeline_matmul_split_k_reduce; + + vk_matmul_pipeline pipeline_dequant_mul_mat_mat[VK_NUM_TYPES]; + + vk_pipeline pipeline_dequant[VK_NUM_TYPES]; + vk_pipeline pipeline_dequant_mul_mat_vec_f32[VK_NUM_TYPES]; + + vk_pipeline pipeline_mul_mat_vec_p021_f16_f32; + vk_pipeline pipeline_mul_mat_vec_nc_f16_f32; + vk_pipeline pipeline_get_rows[VK_NUM_TYPES]; + vk_pipeline pipeline_get_rows_f32[VK_NUM_TYPES]; + vk_pipeline pipeline_mul_f32; + vk_pipeline pipeline_add_f32; + vk_pipeline pipeline_scale_f32; + vk_pipeline pipeline_sqr_f32; + vk_pipeline pipeline_clamp_f32; + vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16; + vk_pipeline pipeline_norm_f32; + vk_pipeline pipeline_rms_norm_f32; + vk_pipeline pipeline_gelu_f32; + vk_pipeline pipeline_silu_f32; + vk_pipeline pipeline_relu_f32; + vk_pipeline pipeline_diag_mask_inf_f32; + vk_pipeline pipeline_soft_max_f32; + vk_pipeline pipeline_rope_f32, pipeline_rope_f16; + vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16; + vk_pipeline pipeline_argsort_f32; + + std::vector pipelines; + ~vk_device() { #ifdef GGML_VULKAN_DEBUG std::cerr << "destroy device " << name << std::endl; #endif + device.destroyCommandPool(compute_queue.pool); + if (!single_queue) { + device.destroyCommandPool(transfer_queue.pool); + } + + for (auto& pipeline : pipelines) { + if (pipeline.expired()) { + continue; + } + + vk_pipeline pl = pipeline.lock(); + ggml_vk_destroy_pipeline(device, pl); + } + pipelines.clear(); + device.destroy(); } }; @@ -125,21 +203,6 @@ struct vk_subbuffer { uint64_t size; }; -struct vk_pipeline { - std::string name; - vk::ShaderModule shader_module; - vk::DescriptorSetLayout dsl; - std::vector descriptor_pools; - std::vector descriptor_sets; - uint32_t descriptor_set_idx; - vk::PipelineLayout layout; - vk::Pipeline pipeline; - uint32_t push_constant_size; - uint32_t parameter_count; - std::array wg_denoms; - uint32_t align; -}; - struct vk_semaphore { vk::Semaphore s; uint64_t value; @@ -160,11 +223,21 @@ struct vk_op_push_constants { float param2; }; -struct vk_op_cpy_push_constants { +struct vk_op_unary_push_constants { uint32_t ne; - uint32_t ne00; uint32_t ne01; uint32_t nb00; uint32_t nb01; uint32_t nb02; - uint32_t ne10; uint32_t ne11; uint32_t nb10; uint32_t nb11; uint32_t nb12; + uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03; + uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13; uint32_t d_offset; + float param1; float param2; +}; + +struct vk_op_binary_push_constants { + uint32_t ne; + uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03; + uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13; + uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23; uint32_t nb20; uint32_t nb21; uint32_t nb22; uint32_t nb23; + uint32_t d_offset; + float param1; float param2; }; struct vk_op_diag_mask_push_constants { @@ -196,6 +269,22 @@ struct vk_op_rope_neox_push_constants { float inv_ndims; }; +struct vk_op_soft_max_push_constants { + uint32_t KX; + uint32_t KY; + uint32_t KZ; + float scale; + float max_bias; + float m0; + float m1; + uint32_t n_head_log2; +}; + +struct vk_op_argsort_push_constants { + uint32_t ncols; + bool ascending; +}; + // Allow pre-recording command buffers struct vk_staging_memcpy { vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {} @@ -236,7 +325,6 @@ struct ggml_tensor_extra_gpu { }; struct ggml_vk_garbage_collector { - std::vector pipelines; std::vector tl_semaphores; std::vector semaphores; std::vector events; @@ -247,35 +335,7 @@ struct ggml_vk_garbage_collector { struct ggml_backend_vk_context { std::string name; - std::weak_ptr device; - vk_pipeline pipeline_matmul_f32_l, pipeline_matmul_f32_m, pipeline_matmul_f32_s; - vk_pipeline pipeline_matmul_f32_aligned_l, pipeline_matmul_f32_aligned_m, pipeline_matmul_f32_aligned_s; - vk_pipeline pipeline_matmul_f16_l, pipeline_matmul_f16_m, pipeline_matmul_f16_s; - vk_pipeline pipeline_matmul_f16_aligned_l, pipeline_matmul_f16_aligned_m, pipeline_matmul_f16_aligned_s; - vk_pipeline pipeline_matmul_f16_f32_l, pipeline_matmul_f16_f32_m, pipeline_matmul_f16_f32_s; - vk_pipeline pipeline_matmul_f16_f32_aligned_l, pipeline_matmul_f16_f32_aligned_m, pipeline_matmul_f16_f32_aligned_s; - vk_pipeline pipeline_matmul_split_k_reduce; - vk_pipeline pipeline_dequant[VK_NUM_TYPES]; - vk_pipeline pipeline_dequant_mul_mat_vec_f32[VK_NUM_TYPES]; - vk_pipeline pipeline_mul_mat_vec_p021_f16_f32; - vk_pipeline pipeline_mul_mat_vec_nc_f16_f32; - vk_pipeline pipeline_get_rows[VK_NUM_TYPES]; - vk_pipeline pipeline_get_rows_f32[VK_NUM_TYPES]; - vk_pipeline pipeline_mul_f32; - vk_pipeline pipeline_add_f32; - vk_pipeline pipeline_scale_f32; - vk_pipeline pipeline_sqr_f32; - vk_pipeline pipeline_clamp_f32; - vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16; - vk_pipeline pipeline_norm_f32; - vk_pipeline pipeline_rms_norm_f32; - vk_pipeline pipeline_gelu_f32; - vk_pipeline pipeline_silu_f32; - vk_pipeline pipeline_relu_f32; - vk_pipeline pipeline_diag_mask_inf_f32; - vk_pipeline pipeline_soft_max_f32; - vk_pipeline pipeline_rope_f32, pipeline_rope_f16; - vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16; + std::shared_ptr device; size_t semaphore_idx, event_idx; ggml_vk_garbage_collector gc; @@ -304,13 +364,31 @@ struct vk_instance { std::vector device_indices; - std::shared_ptr devices[GGML_VK_MAX_DEVICES]; ggml_backend_t backends[GGML_VK_MAX_DEVICES]; ggml_backend_vk_context contexts[GGML_VK_MAX_DEVICES]; ggml_backend_buffer_type buffer_types[GGML_VK_MAX_DEVICES]; bool initialized[GGML_VK_MAX_DEVICES]; }; +static std::shared_ptr ggml_vk_get_device(size_t idx) { +#ifdef GGML_VULKAN_DEBUG + std::cerr << "ggml_vk_get_device(" << idx << ")" << std::endl; +#endif + static std::weak_ptr devices[GGML_VK_MAX_DEVICES]; + + if (devices[idx].expired()) { +#ifdef GGML_VULKAN_DEBUG + std::cerr << "Initializing new vk_device" << std::endl; +#endif + std::shared_ptr device = std::make_shared(); + device->initialized = false; + devices[idx] = device; + return device; + } + + return devices[idx].lock(); +} + #ifdef GGML_VULKAN_CHECK_RESULTS static size_t vk_skip_checks; static size_t vk_output_tensor; @@ -334,14 +412,15 @@ static void ggml_vk_create_pipeline(ggml_backend_vk_context * ctx, vk_pipeline& GGML_ASSERT(parameter_count > 0); GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT - pipeline.name = name; - pipeline.parameter_count = parameter_count; - pipeline.push_constant_size = push_constant_size; - pipeline.wg_denoms = wg_denoms; - pipeline.align = align; + pipeline = std::make_shared(); + pipeline->name = name; + pipeline->parameter_count = parameter_count; + pipeline->push_constant_size = push_constant_size; + pipeline->wg_denoms = wg_denoms; + pipeline->align = align; vk::ShaderModuleCreateInfo shader_module_create_info({}, spv_size, reinterpret_cast(spv_data)); - pipeline.shader_module = ctx->device.lock()->device.createShaderModule(shader_module_create_info); + pipeline->shader_module = ctx->device->device.createShaderModule(shader_module_create_info); std::vector dsl_binding; std::vector dsl_binding_flags; @@ -355,49 +434,49 @@ static void ggml_vk_create_pipeline(ggml_backend_vk_context * ctx, vk_pipeline& vk::PushConstantRange pcr( vk::ShaderStageFlagBits::eCompute, 0, - pipeline.push_constant_size + pipeline->push_constant_size ); vk::DescriptorSetLayoutCreateInfo descriptor_set_layout_create_info( {}, dsl_binding); descriptor_set_layout_create_info.setPNext(&dslbfci); - pipeline.dsl = ctx->device.lock()->device.createDescriptorSetLayout(descriptor_set_layout_create_info); + pipeline->dsl = ctx->device->device.createDescriptorSetLayout(descriptor_set_layout_create_info); // Check if device supports multiple descriptors per pool - if (ctx->device.lock()->descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_UNKNOWN) { + if (ctx->device->descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_UNKNOWN) { const uint32_t alloc_count = 2; // Try allocating multiple sets from one pool // This fails on AMD for some reason, so add a fall back to allocating one pool per set - vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline.parameter_count); + vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count); vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, alloc_count, descriptor_pool_size); - vk::DescriptorPool pool = ctx->device.lock()->device.createDescriptorPool(descriptor_pool_create_info); + vk::DescriptorPool pool = ctx->device->device.createDescriptorPool(descriptor_pool_create_info); std::vector layouts(alloc_count); for (uint32_t i = 0; i < alloc_count; i++) { - layouts[i] = pipeline.dsl; + layouts[i] = pipeline->dsl; } try { vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pool, alloc_count, layouts.data()); - std::vector sets = ctx->device.lock()->device.allocateDescriptorSets(descriptor_set_alloc_info); + std::vector sets = ctx->device->device.allocateDescriptorSets(descriptor_set_alloc_info); } catch(vk::OutOfPoolMemoryError const&) { - ctx->device.lock()->descriptor_set_mode = VK_DEVICE_DESCRIPTOR_POOL_MODE_SINGLE; + ctx->device->descriptor_set_mode = VK_DEVICE_DESCRIPTOR_POOL_MODE_SINGLE; } - ctx->device.lock()->device.destroyDescriptorPool(pool); + ctx->device->device.destroyDescriptorPool(pool); } - if (ctx->device.lock()->descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI) { - vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline.parameter_count); + if (ctx->device->descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI) { + vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count); vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, 128, descriptor_pool_size); - pipeline.descriptor_pools.push_back(ctx->device.lock()->device.createDescriptorPool(descriptor_pool_create_info)); + pipeline->descriptor_pools.push_back(ctx->device->device.createDescriptorPool(descriptor_pool_create_info)); } - pipeline.descriptor_set_idx = 0; + pipeline->descriptor_set_idx = 0; - vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), pipeline.dsl, pcr); - pipeline.layout = ctx->device.lock()->device.createPipelineLayout(pipeline_layout_create_info); + vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), pipeline->dsl, pcr); + pipeline->layout = ctx->device->device.createPipelineLayout(pipeline_layout_create_info); std::vector specialization_entries(specialization_constants.size()); @@ -417,72 +496,75 @@ static void ggml_vk_create_pipeline(ggml_backend_vk_context * ctx, vk_pipeline& vk::PipelineShaderStageCreateInfo pipeline_shader_create_info( vk::PipelineShaderStageCreateFlags(), vk::ShaderStageFlagBits::eCompute, - pipeline.shader_module, + pipeline->shader_module, entrypoint.c_str(), &specialization_info); vk::ComputePipelineCreateInfo compute_pipeline_create_info( vk::PipelineCreateFlags(), pipeline_shader_create_info, - pipeline.layout); - pipeline.pipeline = ctx->device.lock()->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value; + pipeline->layout); + pipeline->pipeline = ctx->device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value; - ctx->gc.pipelines.push_back(&pipeline); + ctx->device->pipelines.push_back(pipeline); } -static void ggml_vk_destroy_pipeline(ggml_backend_vk_context * ctx, vk_pipeline * pipeline) { +static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline) { +#ifdef GGML_VULKAN_DEBUG + std::cerr << "ggml_pipeline_destroy_pipeline(" << pipeline->name << ")" << std::endl; +#endif for (auto& pool : pipeline->descriptor_pools) { - ctx->device.lock()->device.destroyDescriptorPool(pool); + device.destroyDescriptorPool(pool); } pipeline->descriptor_pools.clear(); pipeline->descriptor_sets.clear(); pipeline->descriptor_set_idx = 0; - ctx->device.lock()->device.destroyDescriptorSetLayout(pipeline->dsl); + device.destroyDescriptorSetLayout(pipeline->dsl); - ctx->device.lock()->device.destroyPipelineLayout(pipeline->layout); + device.destroyPipelineLayout(pipeline->layout); - ctx->device.lock()->device.destroyShaderModule(pipeline->shader_module); + device.destroyShaderModule(pipeline->shader_module); - ctx->device.lock()->device.destroyPipeline(pipeline->pipeline); + device.destroyPipeline(pipeline->pipeline); } static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx, vk_pipeline& pipeline, uint32_t n) { #ifdef GGML_VULKAN_DEBUG - std::cerr << "ggml_pipeline_allocate_descriptor_sets(" << pipeline.name << ", " << n << ")" << std::endl; + std::cerr << "ggml_pipeline_allocate_descriptor_sets(" << pipeline->name << ", " << n << ")" << std::endl; #endif - if (pipeline.descriptor_sets.size() >= pipeline.descriptor_set_idx + n) { + if (pipeline->descriptor_sets.size() >= pipeline->descriptor_set_idx + n) { // Enough descriptors are available return; } - if (ctx->device.lock()->descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI) { - const uint32_t alloc_count = pipeline.descriptor_set_idx + n - pipeline.descriptor_sets.size(); + if (ctx->device->descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI) { + const uint32_t alloc_count = pipeline->descriptor_set_idx + n - pipeline->descriptor_sets.size(); std::vector layouts(alloc_count); for (uint32_t i = 0; i < alloc_count; i++) { - layouts[i] = pipeline.dsl; + layouts[i] = pipeline->dsl; } - vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline.descriptor_pools[0], alloc_count, layouts.data()); - std::vector sets = ctx->device.lock()->device.allocateDescriptorSets(descriptor_set_alloc_info); - pipeline.descriptor_sets.insert(pipeline.descriptor_sets.end(), sets.begin(), sets.end()); + vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline->descriptor_pools[0], alloc_count, layouts.data()); + std::vector sets = ctx->device->device.allocateDescriptorSets(descriptor_set_alloc_info); + pipeline->descriptor_sets.insert(pipeline->descriptor_sets.end(), sets.begin(), sets.end()); } else { - for (uint32_t i = pipeline.descriptor_sets.size(); i < pipeline.descriptor_set_idx + n; i++) { - vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline.parameter_count); + for (uint32_t i = pipeline->descriptor_sets.size(); i < pipeline->descriptor_set_idx + n; i++) { + vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count); vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, 1, descriptor_pool_size); - pipeline.descriptor_pools.push_back(ctx->device.lock()->device.createDescriptorPool(descriptor_pool_create_info)); + pipeline->descriptor_pools.push_back(ctx->device->device.createDescriptorPool(descriptor_pool_create_info)); - vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline.descriptor_pools[i], 1, &pipeline.dsl); - std::vector sets = ctx->device.lock()->device.allocateDescriptorSets(descriptor_set_alloc_info); - pipeline.descriptor_sets.push_back(sets[0]); + vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline->descriptor_pools[i], 1, &pipeline->dsl); + std::vector sets = ctx->device->device.allocateDescriptorSets(descriptor_set_alloc_info); + pipeline->descriptor_sets.push_back(sets[0]); } } } static void ggml_pipeline_cleanup(vk_pipeline& pipeline) { #ifdef GGML_VULKAN_DEBUG - std::cerr << "ggml_pipeline_cleanup(" << pipeline.name << ")" << std::endl; + std::cerr << "ggml_pipeline_cleanup(" << pipeline->name << ")" << std::endl; #endif - pipeline.descriptor_set_idx = 0; + pipeline->descriptor_set_idx = 0; } static vk::CommandBuffer ggml_vk_create_cmd_buffer(ggml_backend_vk_context * ctx, vk_queue& q) { @@ -498,7 +580,7 @@ static vk::CommandBuffer ggml_vk_create_cmd_buffer(ggml_backend_vk_context * ctx q.pool, vk::CommandBufferLevel::ePrimary, 1); - const std::vector cmd_buffers = ctx->device.lock()->device.allocateCommandBuffers(command_buffer_alloc_info); + const std::vector cmd_buffers = ctx->device->device.allocateCommandBuffers(command_buffer_alloc_info); auto buf = cmd_buffers.front(); q.cmd_buffers.push_back(buf); @@ -643,11 +725,11 @@ static void ggml_vk_create_queue(ggml_backend_vk_context * ctx, vk_queue& q, uin q.queue_family_index = queue_family_index; vk::CommandPoolCreateInfo command_pool_create_info_compute(vk::CommandPoolCreateFlags(VK_COMMAND_POOL_CREATE_TRANSIENT_BIT), queue_family_index); - q.pool = ctx->device.lock()->device.createCommandPool(command_pool_create_info_compute); + q.pool = ctx->device->device.createCommandPool(command_pool_create_info_compute); q.cmd_buffer_idx = 0; - q.queue = ctx->device.lock()->device.getQueue(queue_family_index, queue_index); + q.queue = ctx->device->device.getQueue(queue_family_index, queue_index); q.stage_flags = stage_flags; } @@ -671,7 +753,7 @@ static vk_semaphore * ggml_vk_create_binary_semaphore(ggml_backend_vk_context * vk::SemaphoreTypeCreateInfo tci{ vk::SemaphoreType::eBinary, 0 }; vk::SemaphoreCreateInfo ci{}; ci.setPNext(&tci); - vk::Semaphore semaphore = ctx->device.lock()->device.createSemaphore(ci); + vk::Semaphore semaphore = ctx->device->device.createSemaphore(ci); ctx->gc.semaphores.push_back({ semaphore, 0 }); return &ctx->gc.semaphores[ctx->gc.semaphores.size() - 1]; } @@ -684,7 +766,7 @@ static vk_semaphore * ggml_vk_create_timeline_semaphore(ggml_backend_vk_context vk::SemaphoreTypeCreateInfo tci{ vk::SemaphoreType::eTimeline, 0 }; vk::SemaphoreCreateInfo ci{}; ci.setPNext(&tci); - vk::Semaphore semaphore = ctx->device.lock()->device.createSemaphore(ci); + vk::Semaphore semaphore = ctx->device->device.createSemaphore(ci); ctx->gc.tl_semaphores.push_back({ semaphore, 0 }); } return &ctx->gc.tl_semaphores[ctx->semaphore_idx++]; @@ -692,7 +774,7 @@ static vk_semaphore * ggml_vk_create_timeline_semaphore(ggml_backend_vk_context static vk::Event ggml_vk_create_event(ggml_backend_vk_context * ctx) { if (ctx->event_idx >= ctx->gc.events.size()) { - ctx->gc.events.push_back(ctx->device.lock()->device.createEvent({})); + ctx->gc.events.push_back(ctx->device->device.createEvent({})); } return ctx->gc.events[ctx->event_idx++]; } @@ -703,7 +785,7 @@ static void ggml_vk_queue_cleanup(ggml_backend_vk_context * ctx, vk_queue& q) { #endif // Requires command buffers to be done - ctx->device.lock()->device.resetCommandPool(q.pool); + ctx->device->device.resetCommandPool(q.pool); q.cmd_buffer_idx = 0; } @@ -740,11 +822,11 @@ static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t siz nullptr, }; - buf->buffer = ctx->device.lock()->device.createBuffer(buffer_create_info); + buf->buffer = ctx->device->device.createBuffer(buffer_create_info); - vk::MemoryRequirements mem_req = ctx->device.lock()->device.getBufferMemoryRequirements(buf->buffer); + vk::MemoryRequirements mem_req = ctx->device->device.getBufferMemoryRequirements(buf->buffer); - vk::PhysicalDeviceMemoryProperties mem_props = ctx->device.lock()->physical_device.getMemoryProperties(); + vk::PhysicalDeviceMemoryProperties mem_props = ctx->device->physical_device.getMemoryProperties(); uint32_t memory_type_index = UINT32_MAX; @@ -757,30 +839,30 @@ static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t siz } if (memory_type_index == UINT32_MAX) { - ctx->device.lock()->device.destroyBuffer(buf->buffer); + ctx->device->device.destroyBuffer(buf->buffer); buf->size = 0; throw vk::OutOfDeviceMemoryError("No suitable memory type found"); } try { - buf->device_memory = ctx->device.lock()->device.allocateMemory({ mem_req.size, memory_type_index }); + buf->device_memory = ctx->device->device.allocateMemory({ mem_req.size, memory_type_index }); } catch (const vk::SystemError& e) { // Out of Host/Device memory, clean up buffer - ctx->device.lock()->device.destroyBuffer(buf->buffer); + ctx->device->device.destroyBuffer(buf->buffer); buf->size = 0; throw e; } buf->ptr = nullptr; if (buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) { - buf->ptr = ctx->device.lock()->device.mapMemory(buf->device_memory, 0, VK_WHOLE_SIZE); + buf->ptr = ctx->device->device.mapMemory(buf->device_memory, 0, VK_WHOLE_SIZE); } - ctx->device.lock()->device.bindBufferMemory(buf->buffer, buf->device_memory, 0); + ctx->device->device.bindBufferMemory(buf->buffer, buf->device_memory, 0); buf->ctx = ctx; - buf->device = ctx->device.lock(); + buf->device = ctx->device; #ifdef GGML_VULKAN_DEBUG std::cerr << "Created buffer " << buf->buffer << std::endl; @@ -802,7 +884,7 @@ static vk_buffer ggml_vk_create_buffer_check(ggml_backend_vk_context * ctx, size static vk_buffer ggml_vk_create_buffer_device(ggml_backend_vk_context * ctx, size_t size) { vk_buffer buf; try { - if (ctx->device.lock()->uma) { + if (ctx->device->uma) { // Fall back to host memory type buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent); } else { @@ -883,10 +965,16 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) { std::cerr << "ggml_vk_load_shaders(" << ctx->name << ")" << std::endl; #endif + const std::shared_ptr device = ctx->device; + // mulmat - std::initializer_list warptile_l = { 128, 128, 128, 16, ctx->device.lock()->subgroup_size * 2, 64, 2, 4, 4, ctx->device.lock()->subgroup_size }; - std::initializer_list warptile_m = { 128, 64, 64, 16, ctx->device.lock()->subgroup_size, 32, 2, 4, 2, ctx->device.lock()->subgroup_size }; - std::initializer_list warptile_s = { ctx->device.lock()->subgroup_size, 32, 32, 16, 32, 32, 2, 2, 2, ctx->device.lock()->subgroup_size }; + std::initializer_list warptile_l = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size }; + std::initializer_list warptile_m = { 128, 64, 64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size }; + std::initializer_list warptile_s = { device->subgroup_size, 32, 32, 16, 32, 32, 2, 2, 2, device->subgroup_size }; + + std::initializer_list warptile_mmq_l = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size }; + std::initializer_list warptile_mmq_m = { 128, 64, 64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size }; + std::initializer_list warptile_mmq_s = { device->subgroup_size, 32, 32, 32, 32, 32, 2, 2, 2, device->subgroup_size }; std::array l_wg_denoms = {128, 128, 1 }; std::array m_wg_denoms = { 64, 64, 1 }; @@ -896,126 +984,206 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) { uint32_t m_align = 64; uint32_t s_align = 32; - if (ctx->device.lock()->fp16) { - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f32_l, "matmul_f32_l", matmul_f32_l_len, matmul_f32_l_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f32_m, "matmul_f32_m", matmul_f32_m_len, matmul_f32_m_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f32_s, "matmul_f32_s", matmul_f32_s_len, matmul_f32_s_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f32_aligned_l, "matmul_f32_aligned_l", matmul_f32_aligned_l_len, matmul_f32_aligned_l_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f32_aligned_m, "matmul_f32_aligned_m", matmul_f32_aligned_m_len, matmul_f32_aligned_m_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f32_aligned_s, "matmul_f32_aligned_s", matmul_f32_aligned_s_len, matmul_f32_aligned_s_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align); + ctx->device->pipeline_matmul_f32 = std::make_shared(); + ctx->device->pipeline_matmul_f16_f32 = std::make_shared(); + ctx->device->pipeline_matmul_f16 = std::make_shared(); + ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0] = std::make_shared(); + ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1] = std::make_shared(); + ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0] = std::make_shared(); + ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1] = std::make_shared(); + ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0] = std::make_shared(); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_l, "matmul_f16_l", matmul_f16_l_len, matmul_f16_l_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_m, "matmul_f16_m", matmul_f16_m_len, matmul_f16_m_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_s, "matmul_f16_s", matmul_f16_s_len, matmul_f16_s_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_aligned_l, "matmul_f16_aligned_l", matmul_f16_aligned_l_len, matmul_f16_aligned_l_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_aligned_m, "matmul_f16_aligned_m", matmul_f16_aligned_m_len, matmul_f16_aligned_m_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_aligned_s, "matmul_f16_aligned_s", matmul_f16_aligned_s_len, matmul_f16_aligned_s_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align); + if (device->fp16) { + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_len, matmul_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_len, matmul_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_len, matmul_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_aligned_len, matmul_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_aligned_len, matmul_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_aligned_len, matmul_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_f32_l, "matmul_f16_f32_l", matmul_f16_f32_l_len, matmul_f16_f32_l_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_f32_m, "matmul_f16_f32_m", matmul_f16_f32_m_len, matmul_f16_f32_m_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_f32_s, "matmul_f16_f32_s", matmul_f16_f32_s_len, matmul_f16_f32_s_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_f32_aligned_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_l_len, matmul_f16_f32_aligned_l_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_f32_aligned_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_m_len, matmul_f16_f32_aligned_m_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_f32_aligned_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_s_len, matmul_f16_f32_aligned_s_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_len, matmul_f16_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_len, matmul_f16_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_len, matmul_f16_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align); + + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align); + + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_0_f32_l", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_0_f32_m", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_0_f32_s", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); } else { - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f32_l, "matmul_f32_l", matmul_f32_l_fp32_len, matmul_f32_l_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f32_m, "matmul_f32_m", matmul_f32_m_fp32_len, matmul_f32_m_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f32_s, "matmul_f32_s", matmul_f32_s_fp32_len, matmul_f32_s_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f32_aligned_l, "matmul_f32_aligned_l", matmul_f32_aligned_l_fp32_len, matmul_f32_aligned_l_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f32_aligned_m, "matmul_f32_aligned_m", matmul_f32_aligned_m_fp32_len, matmul_f32_aligned_m_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f32_aligned_s, "matmul_f32_aligned_s", matmul_f32_aligned_s_fp32_len, matmul_f32_aligned_s_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_fp32_len, matmul_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_fp32_len, matmul_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_fp32_len, matmul_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_aligned_fp32_len, matmul_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_aligned_fp32_len, matmul_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_aligned_fp32_len, matmul_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_l, "matmul_f16_l", matmul_f16_l_fp32_len, matmul_f16_l_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_m, "matmul_f16_m", matmul_f16_m_fp32_len, matmul_f16_m_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_s, "matmul_f16_s", matmul_f16_s_fp32_len, matmul_f16_s_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_aligned_l, "matmul_f16_aligned_l", matmul_f16_aligned_l_fp32_len, matmul_f16_aligned_l_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_aligned_m, "matmul_f16_aligned_m", matmul_f16_aligned_m_fp32_len, matmul_f16_aligned_m_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_aligned_s, "matmul_f16_aligned_s", matmul_f16_aligned_s_fp32_len, matmul_f16_aligned_s_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_f32_l, "matmul_f16_f32_l", matmul_f16_f32_l_fp32_len, matmul_f16_f32_l_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_f32_m, "matmul_f16_f32_m", matmul_f16_f32_m_fp32_len, matmul_f16_f32_m_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_f32_s, "matmul_f16_f32_s", matmul_f16_f32_s_fp32_len, matmul_f16_f32_s_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_f32_aligned_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_l_fp32_len, matmul_f16_f32_aligned_l_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_f32_aligned_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_m_fp32_len, matmul_f16_f32_aligned_m_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_f16_f32_aligned_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_s_fp32_len, matmul_f16_f32_aligned_s_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align); + + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); } - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32", mul_mat_vec_f16_f32_len, mul_mat_vec_f16_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32", mul_mat_vec_q4_0_f32_len, mul_mat_vec_q4_0_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32", mul_mat_vec_q4_1_f32_len, mul_mat_vec_q4_1_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32", mul_mat_vec_q5_0_f32_len, mul_mat_vec_q5_0_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32", mul_mat_vec_q5_1_f32_len, mul_mat_vec_q5_1_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32", mul_mat_vec_q8_0_f32_len, mul_mat_vec_q8_0_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_K_f32", mul_mat_vec_q2_K_f32_len, mul_mat_vec_q2_K_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_K_f32", mul_mat_vec_q3_K_f32_len, mul_mat_vec_q3_K_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_K_f32", mul_mat_vec_q4_K_f32_len, mul_mat_vec_q4_K_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_K_f32", mul_mat_vec_q5_K_f32_len, mul_mat_vec_q5_K_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_K_f32", mul_mat_vec_q6_K_f32_len, mul_mat_vec_q6_K_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32", mul_mat_vec_f16_f32_len, mul_mat_vec_f16_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32", mul_mat_vec_q4_0_f32_len, mul_mat_vec_q4_0_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32", mul_mat_vec_q4_1_f32_len, mul_mat_vec_q4_1_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32", mul_mat_vec_q5_0_f32_len, mul_mat_vec_q5_0_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32", mul_mat_vec_q5_1_f32_len, mul_mat_vec_q5_1_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32", mul_mat_vec_q8_0_f32_len, mul_mat_vec_q8_0_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_K_f32", mul_mat_vec_q2_K_f32_len, mul_mat_vec_q2_K_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_K_f32", mul_mat_vec_q3_K_f32_len, mul_mat_vec_q3_K_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_K_f32", mul_mat_vec_q4_K_f32_len, mul_mat_vec_q4_K_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_K_f32", mul_mat_vec_q5_K_f32_len, mul_mat_vec_q5_K_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_K_f32", mul_mat_vec_q6_K_f32_len, mul_mat_vec_q6_K_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1); // dequant shaders - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16", f32_to_f16_len, f32_to_f16_data, "main", 2, 4 * sizeof(int), { 64, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant[GGML_TYPE_F16 ], "dequant_f16", dequant_f16_len, dequant_f16_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant[GGML_TYPE_Q4_0], "dequant_q4_0", dequant_q4_0_len, dequant_q4_0_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant[GGML_TYPE_Q4_1], "dequant_q4_1", dequant_q4_1_len, dequant_q4_1_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant[GGML_TYPE_Q5_0], "dequant_q5_0", dequant_q5_0_len, dequant_q5_0_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant[GGML_TYPE_Q5_1], "dequant_q5_1", dequant_q5_1_len, dequant_q5_1_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant[GGML_TYPE_Q8_0], "dequant_q8_0", dequant_q8_0_len, dequant_q8_0_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant[GGML_TYPE_Q2_K], "dequant_q2_K", dequant_q2_K_len, dequant_q2_K_data, "main", 2, 4 * sizeof(int), {256 * 64, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant[GGML_TYPE_Q3_K], "dequant_q3_K", dequant_q3_K_len, dequant_q3_K_data, "main", 2, 4 * sizeof(int), {256 * 64, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant[GGML_TYPE_Q4_K], "dequant_q4_K", dequant_q4_K_len, dequant_q4_K_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant[GGML_TYPE_Q5_K], "dequant_q5_K", dequant_q5_K_len, dequant_q5_K_data, "main", 2, 4 * sizeof(int), {256 * 64, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_dequant[GGML_TYPE_Q6_K], "dequant_q6_K", dequant_q6_K_len, dequant_q6_K_data, "main", 2, 4 * sizeof(int), {256 * 64, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16", dequant_f32_len, dequant_f32_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q4_0], "dequant_q4_0", dequant_q4_0_len, dequant_q4_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q4_1], "dequant_q4_1", dequant_q4_1_len, dequant_q4_1_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q5_0], "dequant_q5_0", dequant_q5_0_len, dequant_q5_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q5_1], "dequant_q5_1", dequant_q5_1_len, dequant_q5_1_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q8_0], "dequant_q8_0", dequant_q8_0_len, dequant_q8_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q2_K], "dequant_q2_K", dequant_q2_K_len, dequant_q2_K_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q3_K], "dequant_q3_K", dequant_q3_K_len, dequant_q3_K_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q4_K], "dequant_q4_K", dequant_q4_K_len, dequant_q4_K_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q5_K], "dequant_q5_K", dequant_q5_K_len, dequant_q5_K_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant[GGML_TYPE_Q6_K], "dequant_q6_K", dequant_q6_K_len, dequant_q6_K_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1); // get_rows - ggml_vk_create_pipeline(ctx, ctx->pipeline_get_rows[GGML_TYPE_F16 ], "get_rows_f16", get_rows_f16_len, get_rows_f16_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_get_rows[GGML_TYPE_Q4_0], "get_rows_q4_0", get_rows_q4_0_len, get_rows_q4_0_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_get_rows[GGML_TYPE_Q4_1], "get_rows_q4_1", get_rows_q4_1_len, get_rows_q4_1_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_get_rows[GGML_TYPE_Q5_0], "get_rows_q5_0", get_rows_q5_0_len, get_rows_q5_0_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_get_rows[GGML_TYPE_Q5_1], "get_rows_q5_1", get_rows_q5_1_len, get_rows_q5_1_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_get_rows[GGML_TYPE_Q8_0], "get_rows_q8_0", get_rows_q8_0_len, get_rows_q8_0_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_get_rows[GGML_TYPE_F16 ], "get_rows_f16", get_rows_f16_len, get_rows_f16_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_get_rows[GGML_TYPE_Q4_0], "get_rows_q4_0", get_rows_q4_0_len, get_rows_q4_0_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_get_rows[GGML_TYPE_Q4_1], "get_rows_q4_1", get_rows_q4_1_len, get_rows_q4_1_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_get_rows[GGML_TYPE_Q5_0], "get_rows_q5_0", get_rows_q5_0_len, get_rows_q5_0_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_get_rows[GGML_TYPE_Q5_1], "get_rows_q5_1", get_rows_q5_1_len, get_rows_q5_1_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_get_rows[GGML_TYPE_Q8_0], "get_rows_q8_0", get_rows_q8_0_len, get_rows_q8_0_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_get_rows_f32[GGML_TYPE_F32 ], "get_rows_f16_f32", get_rows_f16_f32_len, get_rows_f16_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_get_rows_f32[GGML_TYPE_Q4_0], "get_rows_q4_0_f32", get_rows_q4_0_f32_len, get_rows_q4_0_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_get_rows_f32[GGML_TYPE_Q4_1], "get_rows_q4_1_f32", get_rows_q4_1_f32_len, get_rows_q4_1_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_get_rows_f32[GGML_TYPE_Q5_0], "get_rows_q5_0_f32", get_rows_q5_0_f32_len, get_rows_q5_0_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_get_rows_f32[GGML_TYPE_Q5_1], "get_rows_q5_1_f32", get_rows_q5_1_f32_len, get_rows_q5_1_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_get_rows_f32[GGML_TYPE_Q8_0], "get_rows_q8_0_f32", get_rows_q8_0_f32_len, get_rows_q8_0_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_get_rows_f32[GGML_TYPE_F32 ], "get_rows_f16_f32", get_rows_f16_f32_len, get_rows_f16_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_get_rows_f32[GGML_TYPE_Q4_0], "get_rows_q4_0_f32", get_rows_q4_0_f32_len, get_rows_q4_0_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_get_rows_f32[GGML_TYPE_Q4_1], "get_rows_q4_1_f32", get_rows_q4_1_f32_len, get_rows_q4_1_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_get_rows_f32[GGML_TYPE_Q5_0], "get_rows_q5_0_f32", get_rows_q5_0_f32_len, get_rows_q5_0_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_get_rows_f32[GGML_TYPE_Q5_1], "get_rows_q5_1_f32", get_rows_q5_1_f32_len, get_rows_q5_1_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_get_rows_f32[GGML_TYPE_Q8_0], "get_rows_q8_0_f32", get_rows_q8_0_f32_len, get_rows_q8_0_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_mul_mat_vec_p021_f16_f32, "mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, "mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_cpy_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_cpy_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_cpy_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_gelu_f32, "gelu_f32", gelu_f32_len, gelu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_silu_f32, "silu_f32", silu_f32_len, silu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_relu_f32, "relu_f32", relu_f32_len, relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_gelu_f32, "gelu_f32", gelu_f32_len, gelu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_silu_f32, "silu_f32", silu_f32_len, silu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_relu_f32, "relu_f32", relu_f32_len, relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_rope_f32, "rope_f32", rope_f32_len, rope_f32_data, "main", 3, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_rope_f16, "rope_f16", rope_f16_len, rope_f16_data, "main", 3, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_f32, "rope_f32", rope_f32_len, rope_f32_data, "main", 3, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_f16, "rope_f16", rope_f16_len, rope_f16_data, "main", 3, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 3, sizeof(vk_op_rope_neox_push_constants), {1, 512, 1}, {}, 1); - ggml_vk_create_pipeline(ctx, ctx->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 3, sizeof(vk_op_rope_neox_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 3, sizeof(vk_op_rope_neox_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 3, sizeof(vk_op_rope_neox_push_constants), {1, 512, 1}, {}, 1); + + ggml_vk_create_pipeline(ctx, ctx->device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1); } static void ggml_vk_print_gpu_info(size_t idx) { @@ -1057,8 +1225,8 @@ static void ggml_vk_print_gpu_info(size_t idx) { } } - const char* GGML_VULKAN_DISABLE_F16 = getenv("GGML_VULKAN_DISABLE_F16"); - bool force_disable_f16 = GGML_VULKAN_DISABLE_F16 != nullptr; + const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16"); + bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr; bool fp16 = !force_disable_f16 && fp16_storage && fp16_compute; @@ -1106,7 +1274,9 @@ void ggml_vk_instance_init() { const std::vector instance_extensions = vk::enumerateInstanceExtensionProperties(); const bool validation_ext = ggml_vk_instance_validation_ext_available(instance_extensions); +#ifdef __APPLE__ const bool portability_enumeration_ext = ggml_vk_instance_portability_enumeration_ext_available(instance_extensions); +#endif std::vector layers; @@ -1117,13 +1287,17 @@ void ggml_vk_instance_init() { if (validation_ext) { extensions.push_back("VK_EXT_validation_features"); } +#ifdef __APPLE__ if (portability_enumeration_ext) { extensions.push_back("VK_KHR_portability_enumeration"); } +#endif vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags{}, &app_info, layers, extensions); +#ifdef __APPLE__ if (portability_enumeration_ext) { instance_create_info.flags |= vk::InstanceCreateFlagBits::eEnumeratePortabilityKHR; } +#endif std::vector features_enable; vk::ValidationFeaturesEXT validation_features; @@ -1182,140 +1356,152 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) { throw std::runtime_error("Device not found"); } - vk_instance.devices[idx] = std::make_shared(); - ctx->device = vk_instance.devices[idx]; - ctx->device.lock()->physical_device = devices[dev_num]; - const std::vector ext_props = ctx->device.lock()->physical_device.enumerateDeviceExtensionProperties(); + ctx->device = ggml_vk_get_device(idx); + if (!ctx->device->initialized) { + ctx->device->physical_device = devices[dev_num]; + const std::vector ext_props = ctx->device->physical_device.enumerateDeviceExtensionProperties(); - bool maintenance4_support = false; + bool maintenance4_support = false; - // Check if maintenance4 is supported - for (const auto& properties : ext_props) { - if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) { - maintenance4_support = true; + // Check if maintenance4 is supported + for (const auto& properties : ext_props) { + if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) { + maintenance4_support = true; + } } - } - vk::PhysicalDeviceProperties2 props2; - vk::PhysicalDeviceMaintenance3Properties props3; - vk::PhysicalDeviceMaintenance4Properties props4; - vk::PhysicalDeviceSubgroupProperties subgroup_props; - props2.pNext = &props3; - props3.pNext = &subgroup_props; - if (maintenance4_support) { - subgroup_props.pNext = &props4; - } - ctx->device.lock()->physical_device.getProperties2(&props2); - ctx->device.lock()->properties = props2.properties; - - if (maintenance4_support) { - ctx->device.lock()->max_memory_allocation_size = std::min(props3.maxMemoryAllocationSize, props4.maxBufferSize); - } else { - ctx->device.lock()->max_memory_allocation_size = props3.maxMemoryAllocationSize; - } - - ctx->device.lock()->vendor_id = ctx->device.lock()->properties.vendorID; - ctx->device.lock()->subgroup_size = subgroup_props.subgroupSize; - ctx->device.lock()->uma = ctx->device.lock()->properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu; - - bool fp16_storage = false; - bool fp16_compute = false; - - for (const auto& properties : ext_props) { - if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) { - fp16_storage = true; - } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) { - fp16_compute = true; + vk::PhysicalDeviceProperties2 props2; + vk::PhysicalDeviceMaintenance3Properties props3; + vk::PhysicalDeviceMaintenance4Properties props4; + vk::PhysicalDeviceSubgroupProperties subgroup_props; + props2.pNext = &props3; + props3.pNext = &subgroup_props; + if (maintenance4_support) { + subgroup_props.pNext = &props4; } - } + ctx->device->physical_device.getProperties2(&props2); + ctx->device->properties = props2.properties; - const char* GGML_VULKAN_DISABLE_F16 = getenv("GGML_VULKAN_DISABLE_F16"); - bool force_disable_f16 = GGML_VULKAN_DISABLE_F16 != nullptr; + const char* GGML_VK_FORCE_MAX_ALLOCATION_SIZE = getenv("GGML_VK_FORCE_MAX_ALLOCATION_SIZE"); - ctx->device.lock()->fp16 = !force_disable_f16 && fp16_storage && fp16_compute; + if (GGML_VK_FORCE_MAX_ALLOCATION_SIZE != nullptr) { + ctx->device->max_memory_allocation_size = std::stoi(GGML_VK_FORCE_MAX_ALLOCATION_SIZE); + } else if (maintenance4_support) { + ctx->device->max_memory_allocation_size = std::min(props3.maxMemoryAllocationSize, props4.maxBufferSize); + } else { + ctx->device->max_memory_allocation_size = props3.maxMemoryAllocationSize; + } - std::vector queue_family_props = ctx->device.lock()->physical_device.getQueueFamilyProperties(); + ctx->device->vendor_id = ctx->device->properties.vendorID; + ctx->device->subgroup_size = subgroup_props.subgroupSize; + ctx->device->uma = ctx->device->properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu; - // Try to find a non-graphics compute queue and transfer-focused queues - const uint32_t compute_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eCompute, vk::QueueFlagBits::eGraphics, -1, 1); - const uint32_t transfer_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eTransfer, vk::QueueFlagBits::eCompute | vk::QueueFlagBits::eGraphics, compute_queue_family_index, 1); + bool fp16_storage = false; + bool fp16_compute = false; - const float priorities[] = { 1.0f, 1.0f }; - ctx->device.lock()->single_queue = compute_queue_family_index == transfer_queue_family_index && queue_family_props[compute_queue_family_index].queueCount == 1; + for (const auto& properties : ext_props) { + if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) { + fp16_storage = true; + } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) { + fp16_compute = true; + } + } - std::vector device_queue_create_infos; - if (compute_queue_family_index != transfer_queue_family_index) { - device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities}); - device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), transfer_queue_family_index, 1, priorities + 1}); - } else if(!ctx->device.lock()->single_queue) { - device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 2, priorities}); - } else { - device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities}); - } - vk::DeviceCreateInfo device_create_info; - std::vector device_extensions; - vk::PhysicalDeviceFeatures device_features = ctx->device.lock()->physical_device.getFeatures(); + const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16"); + const bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr; - VkPhysicalDeviceFeatures2 device_features2; - device_features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2; - device_features2.pNext = nullptr; - device_features2.features = (VkPhysicalDeviceFeatures)device_features; + ctx->device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute; - VkPhysicalDeviceVulkan11Features vk11_features; - vk11_features.pNext = nullptr; - vk11_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES; - device_features2.pNext = &vk11_features; + std::vector queue_family_props = ctx->device->physical_device.getQueueFamilyProperties(); - VkPhysicalDeviceVulkan12Features vk12_features; - vk12_features.pNext = nullptr; - vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES; - vk11_features.pNext = &vk12_features; + // Try to find a non-graphics compute queue and transfer-focused queues + const uint32_t compute_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eCompute, vk::QueueFlagBits::eGraphics, -1, 1); + const uint32_t transfer_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eTransfer, vk::QueueFlagBits::eCompute | vk::QueueFlagBits::eGraphics, compute_queue_family_index, 1); - vkGetPhysicalDeviceFeatures2(ctx->device.lock()->physical_device, &device_features2); + const float priorities[] = { 1.0f, 1.0f }; + ctx->device->single_queue = compute_queue_family_index == transfer_queue_family_index && queue_family_props[compute_queue_family_index].queueCount == 1; - ctx->device.lock()->fp16 = ctx->device.lock()->fp16 && vk12_features.shaderFloat16; + std::vector device_queue_create_infos; + if (compute_queue_family_index != transfer_queue_family_index) { + device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities}); + device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), transfer_queue_family_index, 1, priorities + 1}); + } else if(!ctx->device->single_queue) { + device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 2, priorities}); + } else { + device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities}); + } + vk::DeviceCreateInfo device_create_info; + std::vector device_extensions; + vk::PhysicalDeviceFeatures device_features = ctx->device->physical_device.getFeatures(); - if (!vk11_features.storageBuffer16BitAccess) { - std::cerr << "ggml_vulkan: device " << GGML_VK_NAME << idx << " does not support 16-bit storage." << std::endl; - throw std::runtime_error("Unsupported device"); - } + VkPhysicalDeviceFeatures2 device_features2; + device_features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2; + device_features2.pNext = nullptr; + device_features2.features = (VkPhysicalDeviceFeatures)device_features; - device_extensions.push_back("VK_KHR_16bit_storage"); + VkPhysicalDeviceVulkan11Features vk11_features; + vk11_features.pNext = nullptr; + vk11_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES; + device_features2.pNext = &vk11_features; + + VkPhysicalDeviceVulkan12Features vk12_features; + vk12_features.pNext = nullptr; + vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES; + vk11_features.pNext = &vk12_features; + + vkGetPhysicalDeviceFeatures2(ctx->device->physical_device, &device_features2); + + ctx->device->fp16 = ctx->device->fp16 && vk12_features.shaderFloat16; + + if (!vk11_features.storageBuffer16BitAccess) { + std::cerr << "ggml_vulkan: device " << GGML_VK_NAME << idx << " does not support 16-bit storage." << std::endl; + throw std::runtime_error("Unsupported device"); + } + + device_extensions.push_back("VK_KHR_16bit_storage"); #ifdef GGML_VULKAN_VALIDATE - device_extensions.push_back("VK_KHR_shader_non_semantic_info"); + device_extensions.push_back("VK_KHR_shader_non_semantic_info"); #endif - if (ctx->device.lock()->fp16) { - device_extensions.push_back("VK_KHR_shader_float16_int8"); - } - ctx->device.lock()->name = ctx->device.lock()->properties.deviceName.data(); + if (ctx->device->fp16) { + device_extensions.push_back("VK_KHR_shader_float16_int8"); + } + ctx->device->name = ctx->device->properties.deviceName.data(); - device_create_info = { - vk::DeviceCreateFlags(), - device_queue_create_infos, - {}, - device_extensions - }; - device_create_info.setPNext(&device_features2); - ctx->device.lock()->device = ctx->device.lock()->physical_device.createDevice(device_create_info); + device_create_info = { + vk::DeviceCreateFlags(), + device_queue_create_infos, + {}, + device_extensions + }; + device_create_info.setPNext(&device_features2); + ctx->device->device = ctx->device->physical_device.createDevice(device_create_info); - ctx->device.lock()->descriptor_set_mode = VK_DEVICE_DESCRIPTOR_POOL_MODE_UNKNOWN; + ctx->device->descriptor_set_mode = VK_DEVICE_DESCRIPTOR_POOL_MODE_UNKNOWN; - // Shaders - ggml_vk_load_shaders(ctx); + // Queues + ggml_vk_create_queue(ctx, ctx->device->compute_queue, compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer }); - // Queues - ggml_vk_create_queue(ctx, ctx->device.lock()->compute_queue, compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer }); - if (!ctx->device.lock()->single_queue) { - const uint32_t transfer_queue_index = compute_queue_family_index == transfer_queue_family_index ? 1 : 0; - ggml_vk_create_queue(ctx, ctx->device.lock()->transfer_queue, transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer }); - } else { - // TODO: Use pointer or reference to avoid copy - ctx->device.lock()->transfer_queue = ctx->device.lock()->compute_queue; + // Shaders + ggml_vk_load_shaders(ctx); + + if (!ctx->device->single_queue) { + const uint32_t transfer_queue_index = compute_queue_family_index == transfer_queue_family_index ? 1 : 0; + ggml_vk_create_queue(ctx, ctx->device->transfer_queue, transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer }); + } else { + // TODO: Use pointer or reference to avoid copy + ctx->device->transfer_queue = ctx->device->compute_queue; + } + + ctx->device->idx = dev_num; + ctx->device->initialized = true; + } else if (ctx->device->idx != dev_num) { + std::cerr << "ggml_vulkan: Device " << ctx->device->name << " already initialized with index " << ctx->device->idx << ", but trying to reinitialize with index " << dev_num << std::endl; + throw std::runtime_error("Device already initialized"); } - ctx->fence = ctx->device.lock()->device.createFence({}); + ctx->fence = ctx->device->device.createFence({}); ctx->compute_ctx = nullptr; ctx->transfer_ctx = nullptr; @@ -1333,7 +1519,7 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) { #endif } -static vk_pipeline* ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type type) { +static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type type) { #ifdef GGML_VULKAN_DEBUG std::cerr << "ggml_vk_get_to_fp16()" << std::endl; #endif @@ -1354,10 +1540,36 @@ static vk_pipeline* ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type return nullptr; } - return &ctx->pipeline_dequant[type]; + return ctx->device->pipeline_dequant[type]; } -static vk_pipeline* ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type type) { +static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type) { +#ifdef GGML_VULKAN_DEBUG + std::cerr << "ggml_vk_get_mul_mat_mat_pipeline()" << std::endl; +#endif + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) { + return ctx->device->pipeline_matmul_f32; + } + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { + return ctx->device->pipeline_matmul_f16_f32; + } + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { + return ctx->device->pipeline_matmul_f16; + } + + GGML_ASSERT(src1_type == GGML_TYPE_F32); + + switch (src0_type) { + case GGML_TYPE_Q4_0: + break; + default: + return nullptr; + } + + return ctx->device->pipeline_dequant_mul_mat_mat[src0_type]; +} + +static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type type) { #ifdef GGML_VULKAN_DEBUG std::cerr << "ggml_vk_get_dequantize_mul_mat_vec()" << std::endl; #endif @@ -1378,7 +1590,7 @@ static vk_pipeline* ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * return nullptr; } - return &ctx->pipeline_dequant_mul_mat_vec_f32[type]; + return ctx->device->pipeline_dequant_mul_mat_vec_f32[type]; } static vk_buffer ggml_vk_pool_malloc(ggml_backend_vk_context * ctx, size_t size) { @@ -1457,8 +1669,8 @@ static void * ggml_vk_host_malloc(ggml_backend_vk_context * ctx, size_t size) { if(!(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible)) { fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory\n", size/1024.0/1024.0); - ctx->device.lock()->device.freeMemory(buf->device_memory); - ctx->device.lock()->device.destroyBuffer(buf->buffer); + ctx->device->device.freeMemory(buf->device_memory); + ctx->device->device.destroyBuffer(buf->buffer); return nullptr; } @@ -1522,30 +1734,30 @@ static vk_submission ggml_vk_begin_submission(ggml_backend_vk_context * ctx, vk_ } static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline& pipeline, std::vector&& buffers, size_t push_constant_size, const void* push_constants, std::array elements) { - const uint32_t wg0 = CEIL_DIV(elements[0], pipeline.wg_denoms[0]); - const uint32_t wg1 = CEIL_DIV(elements[1], pipeline.wg_denoms[1]); - const uint32_t wg2 = CEIL_DIV(elements[2], pipeline.wg_denoms[2]); + const uint32_t wg0 = CEIL_DIV(elements[0], pipeline->wg_denoms[0]); + const uint32_t wg1 = CEIL_DIV(elements[1], pipeline->wg_denoms[1]); + const uint32_t wg2 = CEIL_DIV(elements[2], pipeline->wg_denoms[2]); #ifdef GGML_VULKAN_DEBUG - std::cerr << "ggml_vk_dispatch_pipeline(" << pipeline.name << ", (" << wg0 << "," << wg1 << "," << wg2 << "))" << std::endl; + std::cerr << "ggml_vk_dispatch_pipeline(" << pipeline->name << ", (" << wg0 << "," << wg1 << "," << wg2 << "))" << std::endl; #endif std::vector descriptor_buffer_infos; std::vector write_descriptor_sets; - GGML_ASSERT(pipeline.descriptor_set_idx < pipeline.descriptor_sets.size()); - GGML_ASSERT(buffers.size() == pipeline.parameter_count); - vk::DescriptorSet& descriptor_set = pipeline.descriptor_sets[pipeline.descriptor_set_idx++]; - for (uint32_t i = 0; i < pipeline.parameter_count; i++) { + GGML_ASSERT(pipeline->descriptor_set_idx < pipeline->descriptor_sets.size()); + GGML_ASSERT(buffers.size() == pipeline->parameter_count); + vk::DescriptorSet& descriptor_set = pipeline->descriptor_sets[pipeline->descriptor_set_idx++]; + for (uint32_t i = 0; i < pipeline->parameter_count; i++) { descriptor_buffer_infos.push_back({buffers[i].buffer->buffer, buffers[i].offset, buffers[i].size}); } - for (uint32_t i = 0; i < pipeline.parameter_count; i++) { + for (uint32_t i = 0; i < pipeline->parameter_count; i++) { write_descriptor_sets.push_back({descriptor_set, i, 0, 1, vk::DescriptorType::eStorageBuffer, nullptr, &descriptor_buffer_infos[i]}); } - ctx->device.lock()->device.updateDescriptorSets(write_descriptor_sets, {}); + ctx->device->device.updateDescriptorSets(write_descriptor_sets, {}); - subctx->s->buffer.pushConstants(pipeline.layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size, push_constants); - subctx->s->buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline.pipeline); + subctx->s->buffer.pushConstants(pipeline->layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size, push_constants); + subctx->s->buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline->pipeline); subctx->s->buffer.bindDescriptorSets(vk::PipelineBindPoint::eCompute, - pipeline.layout, + pipeline->layout, 0, { descriptor_set }, {}); @@ -1804,7 +2016,7 @@ static void ggml_vk_buffer_write_2d(ggml_backend_vk_context * ctx, vk_buffer& ds memcpy((uint8_t *)dst->ptr + offset + i * width, (const uint8_t *) src + i * spitch, width); } } else { - vk_context * subctx = ggml_vk_create_context(ctx, ctx->device.lock()->transfer_queue); + vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); ggml_vk_ctx_begin(ctx, subctx); ggml_vk_buffer_write_2d_async(ctx, subctx, dst, offset, src, spitch, width, height, true); ggml_vk_ctx_end(subctx); @@ -1814,8 +2026,9 @@ static void ggml_vk_buffer_write_2d(ggml_backend_vk_context * ctx, vk_buffer& ds } ggml_vk_submit(subctx, ctx->fence); - VK_CHECK(ctx->device.lock()->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences"); - ctx->device.lock()->device.resetFences({ ctx->fence }); + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences"); + ctx->device->device.resetFences({ ctx->fence }); + ggml_vk_queue_cleanup(ctx, ctx->device->transfer_queue); } } @@ -1900,18 +2113,19 @@ static void ggml_vk_buffer_read(ggml_backend_vk_context * ctx, vk_buffer& src, s memcpy(dst, (uint8_t *) src->ptr + offset, size); } else { - vk_context * subctx = ggml_vk_create_context(ctx, ctx->device.lock()->transfer_queue); + vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); ggml_vk_ctx_begin(ctx, subctx); ggml_vk_buffer_read_async(ctx, subctx, src, offset, dst, size, true); ggml_vk_ctx_end(subctx); ggml_vk_submit(subctx, ctx->fence); - VK_CHECK(ctx->device.lock()->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "vk_buffer_read waitForFences"); - ctx->device.lock()->device.resetFences({ ctx->fence }); + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "vk_buffer_read waitForFences"); + ctx->device->device.resetFences({ ctx->fence }); for (auto& cpy : subctx->out_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); } + ggml_vk_queue_cleanup(ctx, ctx->device->transfer_queue); } } @@ -1935,15 +2149,13 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr // Copy within the device ggml_backend_vk_context * ctx = src->ctx; - VkBufferCopy bc{ src_offset, dst_offset, size }; - - vk_context * subctx = ggml_vk_create_context(ctx, ctx->device.lock()->transfer_queue); + vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); ggml_vk_ctx_begin(ctx, subctx); ggml_vk_buffer_copy_async(subctx, dst, dst_offset, src, src_offset, size); ggml_vk_ctx_end(subctx); ggml_vk_submit(subctx, ctx->fence); - VK_CHECK(ctx->device.lock()->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences"); - ctx->device.lock()->device.resetFences({ ctx->fence }); + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences"); + ctx->device->device.resetFences({ ctx->fence }); } else { #ifdef GGML_VULKAN_DEBUG std::cerr << "ggml_vk_buffer_copy(MULTI_DEVICE, " << size << ")" << std::endl; @@ -1971,14 +2183,14 @@ static void ggml_vk_buffer_memset(ggml_backend_vk_context * ctx, vk_buffer& dst, // Make sure ctx owns the buffer GGML_ASSERT(dst->ctx == ctx); - vk_context * subctx = ggml_vk_create_context(ctx, ctx->device.lock()->transfer_queue); + vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); ggml_vk_ctx_begin(ctx, subctx); subctx->s->buffer.fillBuffer(dst->buffer, offset, size, c); ggml_vk_ctx_end(subctx); ggml_vk_submit(subctx, ctx->fence); - VK_CHECK(ctx->device.lock()->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "vk_memset waitForFences"); - ctx->device.lock()->device.resetFences({ ctx->fence }); + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "vk_memset waitForFences"); + ctx->device->device.resetFences({ ctx->fence }); } static void ggml_vk_h2d_tensor_2d(ggml_backend_vk_context * ctx, vk_context * subctx, vk_buffer& dst, size_t offset, const ggml_tensor * src, uint64_t i3, uint64_t i2, uint64_t i1) { @@ -2039,176 +2251,63 @@ static void ggml_vk_d2h_tensor_2d(ggml_backend_vk_context * ctx, vk_context * su static uint32_t ggml_vk_guess_split_k(int m, int n, int k) { #ifdef GGML_VULKAN_DEBUG - std::cerr << "ggml_vk_guess_split_k(" << m << ", " << n << ", " << k << ")"; + std::cerr << "ggml_vk_guess_split_k(" << m << ", " << n << ", " << k << ")" << std::endl; #endif if (k > 128 && (m < 128 || n < 128) && m > 2 && n > 2) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << " = 4" << std::endl; -#endif return 4; } -#ifdef GGML_VULKAN_DEBUG - std::cerr << " = 1" << std::endl; -#endif return 1; } -static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ctx, int m, int n) { +static vk_pipeline ggml_vk_guess_matmul_pipeline_amd(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) { + if (m <= 32 || n <= 32) { + return aligned ? mmp->a_s : mmp->s; + } + return aligned ? mmp->a_m : mmp->m; + + GGML_UNUSED(ctx); +} + +static vk_pipeline ggml_vk_guess_matmul_pipeline_apple(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, bool aligned) { + return aligned ? mmp->a_m : mmp->m; + + GGML_UNUSED(ctx); +} + +static vk_pipeline ggml_vk_guess_matmul_pipeline_intel(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, bool aligned) { + return aligned ? mmp->a_s : mmp->s; + + GGML_UNUSED(ctx); +} + +static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) { +#ifdef GGML_VULKAN_DEBUG + std::cerr << "ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ")" << std::endl; +#endif + switch (ctx->device->vendor_id) { + case VK_VENDOR_ID_AMD: + return ggml_vk_guess_matmul_pipeline_amd(ctx, mmp, m, n, aligned); + case VK_VENDOR_ID_APPLE: + return ggml_vk_guess_matmul_pipeline_apple(ctx, mmp, aligned); + case VK_VENDOR_ID_INTEL: + return ggml_vk_guess_matmul_pipeline_intel(ctx, mmp, aligned); + } + + if (m <= 32 || n <= 32) { + return aligned ? mmp->a_s : mmp->s; + } + if (m <= 64 || n <= 64) { + return aligned ? mmp->a_m : mmp->m; + } + return aligned ? mmp->a_l : mmp->l; +} + +static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n) { #ifdef GGML_VULKAN_DEBUG std::cerr << "ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ")" << std::endl; #endif - if (m <= 32 || n <= 32) { - return ctx->pipeline_matmul_f32_aligned_s.align; - } - if (ctx->device.lock()->subgroup_size == 64 || m <= 64 || n <= 64) { - return ctx->pipeline_matmul_f32_aligned_m.align; - } - return ctx->pipeline_matmul_f32_aligned_l.align; -} - -static vk_pipeline* ggml_vk_guess_matmul_pipeline_amd(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, int m, int n, bool aligned) { - if (bit16_x && bit16_y) { - if (m <= 32 || n <= 32) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << " S" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s; - } -#ifdef GGML_VULKAN_DEBUG - std::cerr << " M" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f16_aligned_m : &ctx->pipeline_matmul_f16_m; - } - if (bit16_x && !bit16_y) { - if (m <= 32 || n <= 32) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << " S" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s; - } -#ifdef GGML_VULKAN_DEBUG - std::cerr << " M" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_m : &ctx->pipeline_matmul_f16_f32_m; - } - if (!bit16_x && bit16_y) { - GGML_ASSERT(false); - } - - if (m <= 32 || n <= 32) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << " S" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s; - } -#ifdef GGML_VULKAN_DEBUG - std::cerr << " M" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f32_aligned_m : &ctx->pipeline_matmul_f32_m; -} - -static vk_pipeline* ggml_vk_guess_matmul_pipeline_apple(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, bool aligned) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << " M" << std::endl; -#endif - if (bit16_x && bit16_y) { - return aligned ? &ctx->pipeline_matmul_f16_aligned_m : &ctx->pipeline_matmul_f16_m; - } - if (bit16_x && !bit16_y) { - return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_m : &ctx->pipeline_matmul_f16_f32_m; - } - if (!bit16_x && bit16_y) { - GGML_ASSERT(false); - } - return aligned ? &ctx->pipeline_matmul_f32_aligned_m : &ctx->pipeline_matmul_f32_m; -} - -static vk_pipeline* ggml_vk_guess_matmul_pipeline_intel(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, bool aligned) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << " S" << std::endl; -#endif - if (bit16_x && bit16_y) { - return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s; - } - if (bit16_x && !bit16_y) { - return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s; - } - if (!bit16_x && bit16_y) { - GGML_ASSERT(false); - } - return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s; -} - -static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, int m, int n, bool aligned) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << "ggml_vk_guess_matmul_pipeline(" << bit16_x << ", " << bit16_y << ", " << m << ", " << n << ", " << aligned << ")"; -#endif - switch (ctx->device.lock()->vendor_id) { - case VK_VENDOR_ID_AMD: - return ggml_vk_guess_matmul_pipeline_amd(ctx, bit16_x, bit16_y, m, n, aligned); - case VK_VENDOR_ID_APPLE: - return ggml_vk_guess_matmul_pipeline_apple(ctx, bit16_x, bit16_y, aligned); - case VK_VENDOR_ID_INTEL: - return ggml_vk_guess_matmul_pipeline_intel(ctx, bit16_x, bit16_y, aligned); - } - - if (bit16_x && bit16_y) { - if (m <= 32 || n <= 32) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << " S" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s; - } - if (m <= 64 || n <= 64) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << " M" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f16_aligned_m : &ctx->pipeline_matmul_f16_m; - } -#ifdef GGML_VULKAN_DEBUG - std::cerr << " L" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f16_aligned_l : &ctx->pipeline_matmul_f16_l; - } - if (bit16_x && !bit16_y) { - if (m <= 32 || n <= 32) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << " S" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s; - } - if (m <= 64 || n <= 64) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << " M" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_m : &ctx->pipeline_matmul_f16_f32_m; - } -#ifdef GGML_VULKAN_DEBUG - std::cerr << " L" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_l : &ctx->pipeline_matmul_f16_f32_l; - } - if (!bit16_x && bit16_y) { - GGML_ASSERT(false); - } - - if (m <= 32 || n <= 32) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << " S" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s; - } - if (m <= 64 || n <= 64) { -#ifdef GGML_VULKAN_DEBUG - std::cerr << " M" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f32_aligned_m : &ctx->pipeline_matmul_f32_m; - } -#ifdef GGML_VULKAN_DEBUG - std::cerr << " L" << std::endl; -#endif - return aligned ? &ctx->pipeline_matmul_f32_aligned_l : &ctx->pipeline_matmul_f32_l; + return ggml_vk_guess_matmul_pipeline(ctx, mmp, m, n, false)->align; } static void ggml_vk_matmul(ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline& pipeline, vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& split_k_buffer, uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d, uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3, uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d) { @@ -2226,10 +2325,10 @@ static void ggml_vk_matmul(ggml_backend_vk_context * ctx, vk_context * subctx, v const std::array pc1 = { m, n, k, stride_a, stride_b, stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, batch_stride_a, batch_stride_b, batch_stride_d }; // Make sure enough workgroups get assigned for split k to work - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1.size() * sizeof(uint32_t), pc1.data(), { (CEIL_DIV(m, pipeline.wg_denoms[0]) * pipeline.wg_denoms[0]) * split_k, n, batch }); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1.size() * sizeof(uint32_t), pc1.data(), { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch }); ggml_vk_sync_buffers(subctx); const std::array pc2 = { (uint32_t)(m * n * batch), split_k }; - ggml_vk_dispatch_pipeline(ctx, subctx, ctx->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2.size() * sizeof(uint32_t), pc2.data(), { m * n * batch, 1, 1 }); + ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2.size() * sizeof(uint32_t), pc2.data(), { m * n * batch, 1, 1 }); } static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) { @@ -2239,41 +2338,39 @@ static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) { tensor->nb[3] == tensor->nb[2]*tensor->ne[2]; } -static vk_pipeline * ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, ggml_type from, ggml_type to) { +static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, ggml_type from, ggml_type to) { if (from == GGML_TYPE_F32 && to == GGML_TYPE_F32) { - return &ctx->pipeline_cpy_f32_f32; + return ctx->device->pipeline_cpy_f32_f32; } if (from == GGML_TYPE_F32 && to == GGML_TYPE_F16) { - return &ctx->pipeline_cpy_f32_f16; + return ctx->device->pipeline_cpy_f32_f16; } if (from == GGML_TYPE_F16 && to == GGML_TYPE_F16) { - return &ctx->pipeline_cpy_f16_f16; + return ctx->device->pipeline_cpy_f16_f16; } std::cerr << "Missing CPY op for types: " << ggml_type_name(from) << " " << ggml_type_name(to) << std::endl; GGML_ASSERT(false); } -static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline * pipeline, const ggml_tensor * tensor, vk_subbuffer&& in, vk_subbuffer&& out, ggml_type buffer_type, bool aligned=true) { +static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context * subctx, vk_pipeline pipeline, const ggml_tensor * tensor, vk_subbuffer&& in, vk_subbuffer&& out) { #ifdef GGML_VULKAN_DEBUG std::cerr << "ggml_vk_cpy_to_contiguous((" << tensor << ", type=" << tensor->type << ", backend=" << tensor->backend << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << "), "; std::cerr << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ")" << std::endl; #endif const int tensor_type_size = ggml_type_size(tensor->type); - const int dst_type_size = ggml_type_size(buffer_type); - const uint32_t ne = tensor->ne[0] * tensor->ne[1] * tensor->ne[2]; + const uint32_t ne = ggml_nelements(tensor); - const uint32_t nb2 = aligned ? ggml_vk_align_size(dst_type_size * tensor->ne[0] * tensor->ne[1], ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size : tensor->ne[0] * tensor->ne[1]; - - const vk_op_cpy_push_constants pc = { + const vk_op_unary_push_constants pc = { (uint32_t)ne, - (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->nb[0] / tensor_type_size, (uint32_t)tensor->nb[1] / tensor_type_size, (uint32_t)tensor->nb[2] / tensor_type_size, - (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], 1 , (uint32_t)tensor->ne[0] , nb2, + (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], (uint32_t)tensor->nb[0] / tensor_type_size, (uint32_t)tensor->nb[1] / tensor_type_size, (uint32_t)tensor->nb[2] / tensor_type_size, (uint32_t)tensor->nb[3] / tensor_type_size, + (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], 1 , (uint32_t)tensor->ne[0] , (uint32_t)(tensor->ne[0] * tensor->ne[1]) , (uint32_t)(tensor->ne[0] * tensor->ne[1] * tensor->ne[2]), 0, + 0.0f, 0.0f, }; ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, *pipeline, { in, out }, sizeof(vk_op_cpy_push_constants), &pc, { ne, 1, 1 }); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, { ne, 1, 1 }); } static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -2313,7 +2410,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su bool src0_uma = false; bool src1_uma = false; - if (ctx->device.lock()->uma) { + if (ctx->device->uma) { ggml_vk_host_get(ctx, src0->data, d_Qx, qx_buf_offset); ggml_vk_host_get(ctx, src1->data, d_Qy, qy_buf_offset); src0_uma = d_Qx != nullptr; @@ -2326,10 +2423,17 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su const bool x_non_contig = !load_x && !ggml_vk_dim01_contiguous(src0); const bool y_non_contig = !load_y && !ggml_vk_dim01_contiguous(src1); - const bool f16_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig; + const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig; - const bool qx_needs_dequant = src0->type != GGML_TYPE_F16 || x_non_contig; - const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig; + vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type); + + const bool qx_needs_dequant = mmp == nullptr || x_non_contig; + const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig; + + if (mmp == nullptr) { + // Fall back to dequant + f16 mulmat + mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16); + } // Not implemented GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT @@ -2338,17 +2442,17 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su const int y_ne = ne11 * ne10; const int d_ne = ne11 * ne01; - const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, ne01, ne11)); + const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11)); const bool aligned = ne10 == kpad; const uint32_t split_k = ggml_vk_guess_split_k(ne01, ne11, ne10); - vk_pipeline * pipeline = ggml_vk_guess_matmul_pipeline(ctx, true, !f16_f32_kernel, ne01, ne11, aligned); + vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned); const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type); const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); - const uint64_t x_sz = sizeof(ggml_fp16_t) * x_ne; - const uint64_t y_sz = f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne; + const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne; + const uint64_t y_sz = y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne; const uint64_t d_sz = sizeof(float) * d_ne; vk_buffer d_D = extra->buffer_gpu.lock(); @@ -2379,7 +2483,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su } else { d_X = d_Qx; x_buf_offset = qx_buf_offset; - GGML_ASSERT(qx_sz == x_sz); // NOLINT + GGML_ASSERT(qx_sz == x_sz); } if (qy_needs_dequant) { d_Y = ctx->prealloc_y; @@ -2390,8 +2494,8 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su GGML_ASSERT(qy_sz == y_sz); } - vk_pipeline * to_fp16_vk_0 = nullptr; - vk_pipeline * to_fp16_vk_1 = nullptr; + vk_pipeline to_fp16_vk_0 = nullptr; + vk_pipeline to_fp16_vk_1 = nullptr; if (x_non_contig) { to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, GGML_TYPE_F16); @@ -2407,19 +2511,19 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT // Allocate descriptor sets - ggml_pipeline_allocate_descriptor_sets(ctx, *pipeline, ne12 * ne13); + ggml_pipeline_allocate_descriptor_sets(ctx, pipeline, 1); if (qx_needs_dequant) { - ggml_pipeline_allocate_descriptor_sets(ctx, *to_fp16_vk_0, x_non_contig ? 1 : ne12 * ne13); + ggml_pipeline_allocate_descriptor_sets(ctx, to_fp16_vk_0, 1); } if (qy_needs_dequant) { - ggml_pipeline_allocate_descriptor_sets(ctx, *to_fp16_vk_1, y_non_contig ? 1 : ne12 * ne13); + ggml_pipeline_allocate_descriptor_sets(ctx, to_fp16_vk_1, 1); } if (split_k > 1) { - ggml_pipeline_allocate_descriptor_sets(ctx, ctx->pipeline_matmul_split_k_reduce, ne12 * ne13); + ggml_pipeline_allocate_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, 1); } if (x_non_contig) { - ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE }, dst->type, false); + ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE }); } else if (load_x || qx_needs_dequant) { if (load_x) { // copy data to device @@ -2428,13 +2532,13 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su } if (qx_needs_dequant) { - const std::vector pc = { (int)ne01, (int)ne10, (int)ne10, (int)ne10 }; + const std::vector pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) }; ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, *to_fp16_vk_0, { { d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, { d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1}); + ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { { d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, { d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1}); } } if (y_non_contig) { - ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, dst->type); + ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }); } else if (load_y) { ggml_vk_h2d_tensor_2d(ctx, subctx, d_Qy, 0, src1, 0, 0, ggml_nrows(src1)); } @@ -2451,7 +2555,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su } // compute - ggml_vk_matmul(ctx, subctx, *pipeline, { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 }, { d_D, d_buf_offset, d_sz * ne12 * ne13 }, { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k }, ne01, ne11, ne10, ne10, ne10, ne01, split_k, ne12*ne13, ne02, ne12, r2, r3, stride_batch_x, stride_batch_y, ne20*ne21); // NOLINT + ggml_vk_matmul(ctx, subctx, pipeline, { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 }, { d_D, d_buf_offset, d_sz * ne12 * ne13 }, { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k }, ne01, ne11, ne10, ne10, ne10, ne01, split_k, ne12*ne13, ne02, ne12, r2, r3, stride_batch_x, stride_batch_y, ne20*ne21); // NOLINT if (dst->backend == GGML_BACKEND_TYPE_CPU) { // copy dst to host @@ -2499,7 +2603,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context bool src0_uma = false; bool src1_uma = false; - if (ctx->device.lock()->uma) { + if (ctx->device->uma) { ggml_vk_host_get(ctx, src0->data, d_Qx, qx_buf_offset); ggml_vk_host_get(ctx, src1->data, d_Qy, qy_buf_offset); src0_uma = d_Qx != nullptr; @@ -2521,9 +2625,9 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context const uint64_t y_ne = ne11 * ne10; const uint64_t d_ne = ne11 * ne01; - const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment); + const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment); const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); - const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) : qx_sz; + const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : qx_sz; const uint64_t y_sz = f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne; const uint64_t d_sz = sizeof(float) * d_ne; @@ -2563,8 +2667,8 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context GGML_ASSERT(qy_sz == y_sz); } - vk_pipeline * to_fp16_vk_0 = nullptr; - vk_pipeline* to_fp16_vk_1 = nullptr; + vk_pipeline to_fp16_vk_0 = nullptr; + vk_pipeline to_fp16_vk_1 = nullptr; if (x_non_contig) { to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, src0->type); } @@ -2573,30 +2677,30 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context } else { to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); } - vk_pipeline* dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type); + vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type); GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT GGML_ASSERT(dmmv != nullptr); // Allocate descriptor sets if (qx_needs_dequant) { - ggml_pipeline_allocate_descriptor_sets(ctx, *to_fp16_vk_0, 1); + ggml_pipeline_allocate_descriptor_sets(ctx, to_fp16_vk_0, 1); } if (qy_needs_dequant) { - ggml_pipeline_allocate_descriptor_sets(ctx, *to_fp16_vk_1, y_non_contig ? 1 : ne12 * ne13); + ggml_pipeline_allocate_descriptor_sets(ctx, to_fp16_vk_1, y_non_contig ? 1 : ne12 * ne13); } - ggml_pipeline_allocate_descriptor_sets(ctx, *dmmv, ne12 * ne13); + ggml_pipeline_allocate_descriptor_sets(ctx, dmmv, ne12 * ne13); if (x_non_contig) { - GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment)); - ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE }, src0->type); + GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment)); + ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE }); } else if (load_x) { // copy data to device ggml_vk_h2d_tensor_2d(ctx, subctx, d_Qx, 0, src0, 0, 0, ggml_nrows(src0)); } if (y_non_contig) { GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne); - ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, src1->type); + ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }); } else if (load_y) { ggml_vk_h2d_tensor_2d(ctx, subctx, d_Qy, 0, src1, 0, 0, ggml_nrows(src1)); } @@ -2613,22 +2717,22 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context const uint64_t y_offset = y_buf_offset + y_sz * it_idx1; const uint64_t d_offset = d_buf_offset + d_sz * it_idx1; - const uint64_t y_buffer_offset = (y_offset / ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) * ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment; + const uint64_t y_buffer_offset = (y_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment; const uint64_t y_shader_offset = y_offset - y_buffer_offset; - const uint64_t d_buffer_offset = (d_offset / ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) * ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment; + const uint64_t d_buffer_offset = (d_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment; const uint64_t d_shader_offset = d_offset - d_buffer_offset; if (!y_non_contig && qy_needs_dequant) { - const std::vector pc = { (int)ne11, (int)ne10, (int)ne10, (int)ne10 }; + const std::vector pc = { (uint32_t)ne11, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(y_ne / 32) }; ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, *to_fp16_vk_1, { { d_Qy, qy_offset, qy_sz }, { d_Y, y_offset, y_sz } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)y_ne, 1, 1}); + ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_1, { { d_Qy, qy_offset, qy_sz }, { d_Y, y_offset, y_sz } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)y_ne, 1, 1}); } // compute - const std::array pc = { (int)ne00, (int)(y_shader_offset / ggml_type_size(src1->type)), (int)(d_shader_offset / ggml_type_size(dst->type))}; + const std::array pc = { (uint32_t)ne00, (uint32_t)(y_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type))}; ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, *dmmv, { { d_X, x_offset, x_sz }, { d_Y, y_buffer_offset, y_sz + y_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 3 * sizeof(int), &pc, { (uint32_t)ne01, 1, 1}); + ggml_vk_dispatch_pipeline(ctx, subctx, dmmv, { { d_X, x_offset, x_sz }, { d_Y, y_buffer_offset, y_sz + y_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 3 * sizeof(int), &pc, { (uint32_t)ne01, 1, 1}); if (dst->backend == GGML_BACKEND_TYPE_CPU) { // copy dst to host @@ -2674,7 +2778,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c bool src1_uma = false; - if (ctx->device.lock()->uma) { + if (ctx->device->uma) { ggml_vk_host_get(ctx, src1->data, d_Qy, qy_buf_offset); src1_uma = d_Qy != nullptr; } @@ -2685,7 +2789,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c const uint64_t y_ne = ne10 * ne11 * ne12; const uint64_t d_ne = ne01 * ne11 * ne12; - const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment); + const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment); const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); const uint64_t d_sz = sizeof(float) * d_ne; @@ -2704,12 +2808,12 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c } // Allocate descriptor sets - ggml_pipeline_allocate_descriptor_sets(ctx, ctx->pipeline_mul_mat_vec_p021_f16_f32, 1); + ggml_pipeline_allocate_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, 1); - const uint64_t qy_buffer_offset = (qy_buf_offset / ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) * ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment; + const uint64_t qy_buffer_offset = (qy_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment; const uint64_t qy_shader_offset = qy_buf_offset - qy_buffer_offset; - const uint64_t d_buffer_offset = (d_buf_offset / ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) * ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment; + const uint64_t d_buffer_offset = (d_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment; const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset; if (load_y) { @@ -2719,7 +2823,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c // compute const std::array pc = { (uint32_t)ne00, (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) }; ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, ctx->pipeline_mul_mat_vec_p021_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 }); + ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 }); if (dst->backend == GGML_BACKEND_TYPE_CPU) { // copy dst to host @@ -2766,7 +2870,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con bool src1_uma = false; - if (ctx->device.lock()->uma) { + if (ctx->device->uma) { ggml_vk_host_get(ctx, src1->data, d_Qy, qy_buf_offset); src1_uma = d_Qy != nullptr; } @@ -2797,12 +2901,12 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con } // Allocate descriptor sets - ggml_pipeline_allocate_descriptor_sets(ctx, ctx->pipeline_mul_mat_vec_nc_f16_f32, 1); + ggml_pipeline_allocate_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, 1); - const uint64_t qy_buffer_offset = (qy_buf_offset / ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) * ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment; + const uint64_t qy_buffer_offset = (qy_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment; const uint64_t qy_shader_offset = qy_buf_offset - qy_buffer_offset; - const uint64_t d_buffer_offset = (d_buf_offset / ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) * ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment; + const uint64_t d_buffer_offset = (d_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment; const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset; if (load_y) { @@ -2812,7 +2916,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con // compute const std::array pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) }; ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, ctx->pipeline_mul_mat_vec_nc_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 }); + ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 }); if (dst->backend == GGML_BACKEND_TYPE_CPU) { // copy dst to host @@ -2850,6 +2954,10 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx, } } +// static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context * subctx, const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { +// +// } + static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { // guaranteed to be an integer due to the check in ggml_can_repeat const uint64_t ne0 = dst->ne[0]; @@ -2921,40 +3029,40 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx } -static vk_pipeline* ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_op op) { +static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op) { switch (op) { case GGML_OP_ADD: if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_add_f32; + return ctx->device->pipeline_add_f32; } return nullptr; case GGML_OP_GET_ROWS: GGML_ASSERT(src1->type == GGML_TYPE_I32); if (dst->type == GGML_TYPE_F16) { - return &ctx->pipeline_get_rows[src0->type]; + return ctx->device->pipeline_get_rows[src0->type]; } if (dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_get_rows_f32[src0->type]; + return ctx->device->pipeline_get_rows_f32[src0->type]; } return nullptr; case GGML_OP_MUL: if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_mul_f32; + return ctx->device->pipeline_mul_f32; } return nullptr; case GGML_OP_SCALE: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_scale_f32; + return ctx->device->pipeline_scale_f32; } return nullptr; case GGML_OP_SQR: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_sqr_f32; + return ctx->device->pipeline_sqr_f32; } return nullptr; case GGML_OP_CLAMP: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_clamp_f32; + return ctx->device->pipeline_clamp_f32; } return nullptr; case GGML_OP_CPY: @@ -2963,29 +3071,29 @@ static vk_pipeline* ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ggml_vk_get_cpy_pipeline(ctx, src0->type, dst->type); case GGML_OP_NORM: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_norm_f32; + return ctx->device->pipeline_norm_f32; } return nullptr; case GGML_OP_RMS_NORM: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_rms_norm_f32; + return ctx->device->pipeline_rms_norm_f32; } return nullptr; case GGML_OP_UNARY: switch (ggml_get_unary_op(dst)) { case GGML_UNARY_OP_SILU: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_silu_f32; + return ctx->device->pipeline_silu_f32; } break; case GGML_UNARY_OP_GELU: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_gelu_f32; + return ctx->device->pipeline_gelu_f32; } break; case GGML_UNARY_OP_RELU: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_relu_f32; + return ctx->device->pipeline_relu_f32; } break; default: @@ -2994,12 +3102,12 @@ static vk_pipeline* ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return nullptr; case GGML_OP_DIAG_MASK_INF: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_diag_mask_inf_f32; + return ctx->device->pipeline_diag_mask_inf_f32; } return nullptr; case GGML_OP_SOFT_MAX: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_soft_max_f32; + if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && (src2 == nullptr || src2->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_soft_max_f32; } return nullptr; case GGML_OP_ROPE: @@ -3014,21 +3122,26 @@ static vk_pipeline* ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const if (is_neox) { if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_rope_neox_f32; + return ctx->device->pipeline_rope_neox_f32; } if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { - return &ctx->pipeline_rope_neox_f16; + return ctx->device->pipeline_rope_neox_f16; } } else { if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return &ctx->pipeline_rope_f32; + return ctx->device->pipeline_rope_f32; } if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { - return &ctx->pipeline_rope_f16; + return ctx->device->pipeline_rope_f16; } } return nullptr; } + case GGML_OP_ARGSORT: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_I32) { + return ctx->device->pipeline_argsort_f32; + } + return nullptr; default: return nullptr; } @@ -3044,17 +3157,19 @@ static ggml_vk_func_t ggml_vk_op_get_func(ggml_op op) { } template -static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_op op, const PC&& pc) { +static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, const PC&& pc) { #ifdef GGML_VULKAN_DEBUG std::cerr << "ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3]; if (src1 != nullptr) { std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3]; } + if (src2 != nullptr) { + std::cerr << "), (" << src2 << ", name=" << src2->name << ", type=" << src2->type << ", backend=" << src2->backend << ", ne0=" << src2->ne[0] << ", ne1=" << src2->ne[1] << ", ne2=" << src2->ne[2] << ", ne3=" << src2->ne[3] << ", nb0=" << src2->nb[0] << ", nb1=" << src2->nb[1] << ", nb2=" << src2->nb[2] << ", nb3=" << src2->nb[3]; + } std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "), " << ggml_op_name(op) << ")" << std::endl; #endif GGML_ASSERT(!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))); // NOLINT GGML_ASSERT(op == GGML_OP_CPY || ggml_vk_dim01_contiguous(src0)); // NOLINT - GGML_ASSERT(src1 == nullptr || ggml_vk_dim01_contiguous(src1)); // NOLINT GGML_ASSERT(dst->extra != nullptr); const uint64_t ne00 = src0->ne[0]; const uint64_t ne01 = src0->ne[1]; @@ -3071,7 +3186,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c const uint64_t nb2 = dst->nb[2]; const uint64_t nb3 = dst->nb[3]; - vk_pipeline * pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, dst, op); + const bool use_src2 = src2 != nullptr; + const uint64_t ne2 = use_src2 ? src2->ne[0] * src2->ne[1] : 0; + + vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, dst, op); ggml_vk_func_t op_func; if (pipeline == nullptr) { @@ -3092,29 +3210,39 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra; ggml_tensor_extra_gpu * extra_src1 = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr; + ggml_tensor_extra_gpu * extra_src2 = use_src2 ? (ggml_tensor_extra_gpu *) src2->extra : nullptr; vk_buffer d_X = nullptr; size_t x_buf_offset = 0; vk_buffer d_Y = nullptr; size_t y_buf_offset = 0; + vk_buffer d_Z = nullptr; + size_t z_buf_offset = 0; bool src0_uma = false; bool src1_uma = false; + bool src2_uma = false; - if (ctx->device.lock()->uma) { + if (ctx->device->uma) { ggml_vk_host_get(ctx, src0->data, d_X, x_buf_offset); src0_uma = d_X != nullptr; if (use_src1) { ggml_vk_host_get(ctx, src1->data, d_Y, y_buf_offset); src1_uma = d_Y != nullptr; } + if (use_src2) { + ggml_vk_host_get(ctx, src1->data, d_Z, z_buf_offset); + src2_uma = d_Z != nullptr; + } } const bool transfer_src0 = src0->backend != GGML_BACKEND_TYPE_GPU && !src0_uma; const bool transfer_src1 = use_src1 && src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma; + const bool transfer_src2 = use_src2 && src2->backend != GGML_BACKEND_TYPE_GPU && !src2_uma; - uint64_t x_sz = ggml_vk_align_size(ggml_type_size(src0->type) * ne0, ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment); - uint64_t y_sz = use_src1 ? ggml_vk_align_size(ggml_type_size(src1->type) * ne1, ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) : 0; + uint64_t x_sz = ggml_vk_align_size(ggml_type_size(src0->type) * ne0, ctx->device->properties.limits.minStorageBufferOffsetAlignment); + uint64_t y_sz = use_src1 ? ggml_vk_align_size(ggml_type_size(src1->type) * ne1, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : 0; + uint64_t z_sz = use_src2 ? ggml_vk_align_size(ggml_type_size(src2->type) * ne2, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : 0; uint64_t d_sz = ggml_type_size(dst->type) * ne0; vk_buffer d_D = extra->buffer_gpu.lock(); @@ -3125,7 +3253,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c } GGML_ASSERT(d_D != nullptr); - uint64_t d_buf_offset = (extra->offset / ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) * ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment; + uint64_t d_buf_offset = (extra->offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment; GGML_ASSERT(d_buf_offset == extra->offset || op == GGML_OP_CPY); // NOLINT if (transfer_src0) { d_X = ctx->prealloc_qx; @@ -3142,6 +3270,13 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c GGML_ASSERT(d_Y != nullptr); } + GGML_ASSERT(!transfer_src2); + if (use_src2 && !src2_uma) { + d_Z = extra_src2->buffer_gpu.lock(); + z_buf_offset = extra_src2->offset; + GGML_ASSERT(d_Z != nullptr); + } + if (op == GGML_OP_CPY) { GGML_ASSERT(!transfer_src0); GGML_ASSERT(!transfer_src1); @@ -3169,7 +3304,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c // Single call if dimension 2 is contiguous if (op == GGML_OP_CPY || (ggml_is_contiguous(src0) && (src1 == nullptr || ggml_is_contiguous(src1)))) { - ggml_pipeline_allocate_descriptor_sets(ctx, *pipeline, 1); + ggml_pipeline_allocate_descriptor_sets(ctx, pipeline, 1); switch (dst->op) { case GGML_OP_NORM: @@ -3198,16 +3333,30 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c } } - if (!use_src1 && op == GGML_OP_SOFT_MAX) { - // Empty src1 is possible on soft_max, but the shader needs a buffer + if (op == GGML_OP_SOFT_MAX) { + // Empty src1 and src2 are possible on soft_max, but the shader needs buffers + vk_subbuffer subbuf_y; + if (use_src1) { + subbuf_y = { d_Y, y_buf_offset, y_sz }; + } else { + subbuf_y = { ctx->prealloc_y, 0, ctx->prealloc_y->size }; + } + + vk_subbuffer subbuf_z; + if (use_src2) { + subbuf_z = { d_Z, z_buf_offset, z_sz }; + } else { + subbuf_z = { ctx->prealloc_y, 0, ctx->prealloc_y->size }; + } + ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, *pipeline, { { d_X, x_buf_offset, x_sz }, { ctx->prealloc_y, 0, ctx->prealloc_y->size }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, subbuf_y, subbuf_z, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); } else if (use_src1) { ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, *pipeline, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); } else { ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, *pipeline, { { d_X, x_buf_offset, x_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); } if (dst->backend == GGML_BACKEND_TYPE_CPU && op == GGML_OP_CPY) { ggml_vk_d2h_tensor_2d(ctx, subctx, d_D, 0, dst); @@ -3217,7 +3366,9 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c ggml_vk_buffer_read_async(ctx, subctx, d_D, 0, d, d_sz); } } else { - ggml_pipeline_allocate_descriptor_sets(ctx, *pipeline, ne02 * ne03); + GGML_ASSERT(op != GGML_OP_SOFT_MAX); + + ggml_pipeline_allocate_descriptor_sets(ctx, pipeline, ne02 * ne03); switch (dst->op) { case GGML_OP_NORM: @@ -3242,16 +3393,12 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c const uint32_t y_offset = y_sz * it_idx1; const uint32_t d_offset = d_sz * it_idx0; - if (!use_src1 && op == GGML_OP_SOFT_MAX) { - // Empty src1 is possible on soft_max, but the shader needs a buffer + if (use_src1) { ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, *pipeline, { { d_X, x_buf_offset, x_sz }, { ctx->prealloc_y, 0, ctx->prealloc_y->size }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); - } else if (use_src1) { - ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, *pipeline, { { d_X, x_buf_offset + x_offset, x_sz }, { d_Y, y_buf_offset + y_offset, y_sz }, { d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset + x_offset, x_sz }, { d_Y, y_buf_offset + y_offset, y_sz }, { d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements); } else { ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, *pipeline, { { d_X, x_buf_offset + x_offset, x_sz }, { d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset + x_offset, x_sz }, { d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements); } if (dst->backend == GGML_BACKEND_TYPE_CPU) { // copy dst to host @@ -3263,69 +3410,141 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c } static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_vk_op_f32(ctx, subctx, src0, src1, dst, GGML_OP_REPEAT, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f }); + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_REPEAT, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f }); } static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_vk_op_f32(ctx, subctx, src0, src1, dst, GGML_OP_GET_ROWS, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f }); + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_GET_ROWS, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f }); } static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_vk_op_f32(ctx, subctx, src0, src1, dst, GGML_OP_ADD, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f }); + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t src1_type_size = ggml_type_size(src1->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ADD, { + (uint32_t)ggml_nelements(src0), + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, + 0, + 0.0f, 0.0f, + }); } static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { - ggml_vk_op_f32(ctx, subctx, src0, src1, dst, GGML_OP_MUL, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f }); + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t src1_type_size = ggml_type_size(src1->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_MUL, { + (uint32_t)ggml_nelements(src0), + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, + 0, + 0.0f, 0.0f, + }); } static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { float * op_params = (float *)dst->op_params; - ggml_vk_op_f32(ctx, subctx, src0, nullptr, dst, GGML_OP_SCALE, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }); + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SCALE, { + (uint32_t)ggml_nelements(src0), + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, + 0, + op_params[0], 0.0f + }); } static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { - ggml_vk_op_f32(ctx, subctx, src0, nullptr, dst, GGML_OP_SQR, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }); + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQR, { + (uint32_t)ggml_nelements(src0), + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, + 0, + 0.0f, 0.0f, + }); } static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { float * op_params = (float *)dst->op_params; - ggml_vk_op_f32(ctx, subctx, src0, nullptr, dst, GGML_OP_CLAMP, { (uint32_t)ggml_nelements(src0), 0, op_params[0], op_params[1] }); + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CLAMP, { + (uint32_t)ggml_nelements(src0), + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, + 0, + op_params[0], op_params[1], + }); } static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra; - const int src0_type_size = ggml_type_size(src0->type); - const int dst_type_size = ggml_type_size(dst->type); - const uint32_t d_offset = (extra->offset % ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size; - ggml_vk_op_f32(ctx, subctx, src0, nullptr, dst, GGML_OP_CPY, { + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + const uint32_t d_offset = (extra->offset % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size; + + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, { (uint32_t)ggml_nelements(src0), - (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, - (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, d_offset, + 0.0f, 0.0f, }); } static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { - ggml_vk_op_f32(ctx, subctx, src0, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f }); + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f }); } static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { float * op_params = (float *)dst->op_params; - ggml_vk_op_f32(ctx, subctx, src0, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }); + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }); } static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { - ggml_vk_op_f32(ctx, subctx, src0, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }); + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }); } static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { int32_t * op_params = (int32_t *)dst->op_params; - ggml_vk_op_f32(ctx, subctx, src0, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }); + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }); } -static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { +static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) { float * op_params = (float *)dst->op_params; - ggml_vk_op_f32(ctx, subctx, src0, src1, dst, GGML_OP_SOFT_MAX, { (uint32_t)src0->ne[0], (uint32_t)(src1 != nullptr ? ggml_nrows(src1) : 0), op_params[0], 0.0f }); + + float scale = op_params[0]; + float max_bias = op_params[1]; + + const uint32_t ncols = (uint32_t)src0->ne[0]; + const uint32_t nrows_x = (uint32_t)ggml_nrows(src0); + const uint32_t nrows_y = (uint32_t)src0->ne[1]; + + const uint32_t n_head_kv = nrows_x/nrows_y; + const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv)); + + const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + + ggml_vk_op_f32(ctx, subctx, src0, src1, src2, dst, GGML_OP_SOFT_MAX, { + ncols, + nrows_y, + src2 != nullptr ? (uint32_t)1 : (uint32_t)0, + scale, max_bias, + m0, m1, + n_head_log2, + }); } static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -3351,12 +3570,17 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, con if (is_neox) { const float theta_scale = powf(freq_base, -2.0f/n_dims); const float inv_ndims = -1.0f / n_dims; - ggml_vk_op_f32(ctx, subctx, src0, src1, dst, GGML_OP_ROPE, { (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1], freq_base, ext_factor, attn_factor, corr_dims[0], corr_dims[1], 0.0f, 0.0f, theta_scale, inv_ndims }); + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ROPE, { (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1], freq_base, ext_factor, attn_factor, corr_dims[0], corr_dims[1], 0.0f, 0.0f, theta_scale, inv_ndims }); } else { - ggml_vk_op_f32(ctx, subctx, src0, src1, dst, GGML_OP_ROPE, { (uint32_t)src0->ne[0], freq_scale, (uint32_t)src0->ne[1], freq_base, ext_factor, attn_factor, corr_dims[0], corr_dims[1], 0.0f, 0.0f }); + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ROPE, { (uint32_t)src0->ne[0], freq_scale, (uint32_t)src0->ne[1], freq_base, ext_factor, attn_factor, corr_dims[0], corr_dims[1], 0.0f, 0.0f }); } } +static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { + int32_t * op_params = (int32_t *)dst->op_params; + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGSORT, { (uint32_t)src0->ne[0], ((ggml_sort_order) op_params[0]) == GGML_SORT_ORDER_ASC }); +} + static void ggml_vk_nop(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) { // If backend is CPU, data from src0 has to be copied off the device if (dst->backend == GGML_BACKEND_TYPE_CPU) { @@ -3408,43 +3632,43 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t const size_t y_ne = k * n * batch; const size_t d_ne = m * n * batch; - vk_pipeline * p; + vk_pipeline p; std::string shname; if (shader_size == 0) { if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f32_aligned_s; + p = ctx->device->pipeline_matmul_f32->a_s; shname = "F32_ALIGNED_S"; } else if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f16_f32_aligned_s; + p = ctx->device->pipeline_matmul_f16_f32->a_s; shname = "F16_F32_ALIGNED_S"; } else if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f16_aligned_s; + p = ctx->device->pipeline_matmul_f16->a_s; shname = "F16_ALIGNED_S"; } else { GGML_ASSERT(false); } } else if (shader_size == 1) { if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f32_aligned_m; + p = ctx->device->pipeline_matmul_f32->a_m; shname = "F32_ALIGNED_M"; } else if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f16_f32_aligned_m; + p = ctx->device->pipeline_matmul_f16_f32->a_m; shname = "F16_F32_ALIGNED_M"; } else if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f16_aligned_m; + p = ctx->device->pipeline_matmul_f16->a_m; shname = "F16_ALIGNED_M"; } else { GGML_ASSERT(false); } } else if (shader_size == 2) { if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f32_aligned_l; + p = ctx->device->pipeline_matmul_f32->a_l; shname = "F32_ALIGNED_L"; } else if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f16_f32_aligned_l; + p = ctx->device->pipeline_matmul_f16_f32->a_l; shname = "F16_F32_ALIGNED_L"; } else if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f16_aligned_l; + p = ctx->device->pipeline_matmul_f16->a_l; shname = "F16_ALIGNED_L"; } else { GGML_ASSERT(false); @@ -3458,43 +3682,43 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t if (k != kpad) { if (shader_size == 0) { if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f32_s; + p = ctx->device->pipeline_matmul_f32->s; shname = "F32_S"; } else if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f16_f32_s; + p = ctx->device->pipeline_matmul_f16_f32->s; shname = "F16_F32_S"; } else if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f16_s; + p = ctx->device->pipeline_matmul_f16->s; shname = "F16_S"; } } else if (shader_size == 1) { if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f32_m; + p = ctx->device->pipeline_matmul_f32->m; shname = "F32_M"; } else if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f16_f32_m; + p = ctx->device->pipeline_matmul_f16_f32->m; shname = "F16_F32_M"; } else if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f16_m; + p = ctx->device->pipeline_matmul_f16->m; shname = "F16_M"; } } else if (shader_size == 2) { if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f32_l; + p = ctx->device->pipeline_matmul_f32->l; shname = "F32_L"; } else if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f16_f32_l; + p = ctx->device->pipeline_matmul_f16_f32->l; shname = "F16_F32_L"; } else if (std::is_same() && std::is_same()) { - p = &ctx->pipeline_matmul_f16_l; + p = ctx->device->pipeline_matmul_f16->l; shname = "F16_L"; } } } - ggml_pipeline_allocate_descriptor_sets(ctx, *p, num_it); + ggml_pipeline_allocate_descriptor_sets(ctx, p, num_it); if (split_k > 1) { - ggml_pipeline_allocate_descriptor_sets(ctx, ctx->pipeline_matmul_split_k_reduce, num_it); + ggml_pipeline_allocate_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, num_it); if (ctx->prealloc_split_k == nullptr || ctx->prealloc_split_k->size < sizeof(float) * d_ne * split_k) { // Resize buffer @@ -3524,9 +3748,11 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t } for (size_t i = 0; i < y_ne; i++) { if (std::is_same()) { - y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; + // y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; + y[i] = (i % k == i / k) ? 1.0f : 0.0f; } else if (std::is_same()) { - y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f); + // y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f); + y[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f); } else { GGML_ASSERT(false); } @@ -3535,17 +3761,17 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t ggml_vk_buffer_write(ctx, d_X, 0, x, sizeof(X_TYPE) * k * m * batch); ggml_vk_buffer_write(ctx, d_Y, 0, y, sizeof(Y_TYPE) * k * n * batch); - vk_context * subctx = ggml_vk_create_context(ctx, ctx->device.lock()->compute_queue); + vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); for (size_t i = 0; i < num_it; i++) { ggml_vk_ctx_begin(ctx, subctx); - ggml_vk_matmul(ctx, subctx, *p, ggml_vk_subbuffer(d_X), ggml_vk_subbuffer(d_Y), ggml_vk_subbuffer(d_D), ggml_vk_subbuffer(ctx->prealloc_split_k), m, n, k, k, k, m, split_k, batch, batch, batch, 1, 1, k*m, k*n, m*n); + ggml_vk_matmul(ctx, subctx, p, ggml_vk_subbuffer(d_X), ggml_vk_subbuffer(d_Y), ggml_vk_subbuffer(d_D), ggml_vk_subbuffer(ctx->prealloc_split_k), m, n, k, k, k, m, split_k, batch, batch, batch, 1, 1, k*m, k*n, m*n); ggml_vk_ctx_end(subctx); } auto begin = std::chrono::high_resolution_clock::now(); ggml_vk_submit(subctx, ctx->fence); - VK_CHECK(ctx->device.lock()->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_matmul waitForFences"); - ctx->device.lock()->device.resetFences({ ctx->fence }); + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_matmul waitForFences"); + ctx->device->device.resetFences({ ctx->fence }); auto end = std::chrono::high_resolution_clock::now(); double time = std::chrono::duration_cast(end-begin).count() / 1000.0; @@ -3624,6 +3850,8 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl; std::cerr << "Actual result: " << std::endl << std::endl; ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); + std::cerr << std::endl; + ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n + 15, first_err_b); std::cerr << "Expected result: " << std::endl << std::endl; ggml_vk_print_matrix_area(d_chk, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); @@ -3649,15 +3877,15 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t free(d_chk); - ggml_vk_queue_cleanup(ctx, ctx->device.lock()->transfer_queue); - ggml_vk_queue_cleanup(ctx, ctx->device.lock()->compute_queue); + ggml_vk_queue_cleanup(ctx, ctx->device->transfer_queue); + ggml_vk_queue_cleanup(ctx, ctx->device->compute_queue); ggml_vk_destroy_buffer(d_X); ggml_vk_destroy_buffer(d_Y); ggml_vk_destroy_buffer(d_D); - ggml_pipeline_cleanup(*p); - ggml_pipeline_cleanup(ctx->pipeline_matmul_split_k_reduce); + ggml_pipeline_cleanup(p); + ggml_pipeline_cleanup(ctx->device->pipeline_matmul_split_k_reduce); free(x); free(y); @@ -3730,7 +3958,7 @@ static void ggml_vk_test_h2d_nc(ggml_backend_vk_context * ctx, size_t ne0, size_ data[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; } - vk_context * subctx = ggml_vk_create_context(ctx, ctx->device.lock()->compute_queue); + vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); ggml_vk_ctx_begin(ctx, subctx); vk_buffer buffer = ggml_vk_create_buffer_check(ctx, ggml_nbytes(tensor), vk::MemoryPropertyFlagBits::eDeviceLocal); @@ -3739,8 +3967,8 @@ static void ggml_vk_test_h2d_nc(ggml_backend_vk_context * ctx, size_t ne0, size_ ggml_vk_ctx_end(subctx); ggml_vk_submit(subctx, ctx->fence); - VK_CHECK(ctx->device.lock()->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_h2d_nc waitForFences"); - ctx->device.lock()->device.resetFences({ ctx->fence }); + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_h2d_nc waitForFences"); + ctx->device->device.resetFences({ ctx->fence }); ggml_vk_buffer_read(ctx, buffer, 0, result_data, ggml_nbytes(tensor)); @@ -3812,7 +4040,7 @@ static void ggml_vk_test_transfer(ggml_backend_vk_context * ctx, size_t ne, bool x[i] = rand() / (float)RAND_MAX; } - vk_context * subctx = ggml_vk_create_context(ctx, ctx->device.lock()->compute_queue); + vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); ggml_vk_ctx_begin(ctx, subctx); auto begin = std::chrono::high_resolution_clock::now(); @@ -3826,8 +4054,8 @@ static void ggml_vk_test_transfer(ggml_backend_vk_context * ctx, size_t ne, bool ggml_vk_ctx_end(subctx); ggml_vk_submit(subctx, ctx->fence); - VK_CHECK(ctx->device.lock()->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_transfer waitForFences"); - ctx->device.lock()->device.resetFences({ ctx->fence }); + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_transfer waitForFences"); + ctx->device->device.resetFences({ ctx->fence }); auto end = std::chrono::high_resolution_clock::now(); @@ -3841,8 +4069,8 @@ static void ggml_vk_test_transfer(ggml_backend_vk_context * ctx, size_t ne, bool ggml_vk_ctx_end(subctx); ggml_vk_submit(subctx, ctx->fence); - VK_CHECK(ctx->device.lock()->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_transfer waitForFences"); - ctx->device.lock()->device.resetFences({ ctx->fence }); + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_transfer waitForFences"); + ctx->device->device.resetFences({ ctx->fence }); for (auto& cpy : subctx->out_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); @@ -3873,6 +4101,48 @@ static void ggml_vk_test_transfer(ggml_backend_vk_context * ctx, size_t ne, bool } } +static void ggml_vk_quantize_data(const float * from, void * to, size_t ne, ggml_type quant) { + std::vector hist_cur(1 << 4, 0); + + switch(quant) { + case GGML_TYPE_F32: + memcpy(to, from, sizeof(float) * ne); + break; + case GGML_TYPE_Q4_0: + ggml_quantize_q4_0(from, to, ne, ne, hist_cur.data()); + break; + case GGML_TYPE_Q4_1: + ggml_quantize_q4_1(from, to, ne, ne, hist_cur.data()); + break; + case GGML_TYPE_Q5_0: + ggml_quantize_q5_0(from, to, ne, ne, hist_cur.data()); + break; + case GGML_TYPE_Q5_1: + ggml_quantize_q5_1(from, to, ne, ne, hist_cur.data()); + break; + case GGML_TYPE_Q8_0: + ggml_quantize_q8_0(from, to, ne, ne, hist_cur.data()); + break; + case GGML_TYPE_Q2_K: + ggml_quantize_q2_K(from, to, ne, ne, hist_cur.data()); + break; + case GGML_TYPE_Q3_K: + ggml_quantize_q3_K(from, to, ne, ne, hist_cur.data()); + break; + case GGML_TYPE_Q4_K: + ggml_quantize_q4_K(from, to, ne, ne, hist_cur.data()); + break; + case GGML_TYPE_Q5_K: + ggml_quantize_q5_K(from, to, ne, ne, hist_cur.data()); + break; + case GGML_TYPE_Q6_K: + ggml_quantize_q6_K(from, to, ne, ne, hist_cur.data()); + break; + default: + GGML_ASSERT(false); + } +} + static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_type quant) { #ifdef GGML_VULKAN_DEBUG std::cerr << "ggml_vk_test_dequant(" << ne << ")" << std::endl; @@ -3890,72 +4160,59 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_ x[i] = rand() / (float)RAND_MAX; } - std::vector hist_cur(1 << 4, 0); + vk_pipeline p = ctx->device->pipeline_dequant[quant]; - vk_pipeline& p = ctx->pipeline_dequant[quant]; - - switch(quant) { - case GGML_TYPE_Q4_0: - ggml_quantize_q4_0(x, qx, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q4_1: - ggml_quantize_q4_1(x, qx, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q5_0: - ggml_quantize_q5_0(x, qx, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q5_1: - ggml_quantize_q4_1(x, qx, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q8_0: - ggml_quantize_q8_0(x, qx, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q2_K: - ggml_quantize_q2_K(x, qx, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q3_K: - ggml_quantize_q3_K(x, qx, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q4_K: - ggml_quantize_q4_K(x, qx, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q5_K: - ggml_quantize_q5_K(x, qx, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q6_K: - ggml_quantize_q6_K(x, qx, ne, ne, hist_cur.data()); - break; - default: - GGML_ASSERT(false); - } + ggml_vk_quantize_data(x, qx, ne, quant); ggml_pipeline_allocate_descriptor_sets(ctx, p, 1); ggml_vk_buffer_write(ctx, qx_buf, 0, qx, qx_sz); - vk_context * subctx = ggml_vk_create_context(ctx, ctx->device.lock()->compute_queue); + vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); ggml_vk_ctx_begin(ctx, subctx); - const std::vector pc = { 1, (int)ne, (int)ne, (int)ne }; + const std::vector pc = { 1, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne }; ggml_vk_dispatch_pipeline(ctx, subctx, p, { { qx_buf, 0, qx_sz }, { x_buf, 0, x_sz_f16 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)ne, 1, 1}); ggml_vk_ctx_end(subctx); auto begin = std::chrono::high_resolution_clock::now(); ggml_vk_submit(subctx, ctx->fence); - VK_CHECK(ctx->device.lock()->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences"); - ctx->device.lock()->device.resetFences({ ctx->fence }); + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences"); + ctx->device->device.resetFences({ ctx->fence }); auto end = std::chrono::high_resolution_clock::now(); double ms_dequant = std::chrono::duration_cast(end-begin).count() / 1000.0; ggml_vk_buffer_read(ctx, x_buf, 0, x_chk, x_sz_f16); + int first_err = -1; + double avg_err = 0.0; for (size_t i = 0; i < ne; i++) { - avg_err += std::fabs(x[i] - ggml_fp16_to_fp32(x_chk[i])); + double error = std::fabs(x[i] - ggml_fp16_to_fp32(x_chk[i])); + avg_err += error; + + if (first_err < 0 && error > 0.05) { + first_err = i; + } } - std::cerr << "TEST DEQUANT " << ggml_type_name(quant) << " time=" << ms_dequant << "ms avg_err=" << avg_err / ne << std::endl; + avg_err /= ne; + + std::cerr << "TEST DEQUANT " << ggml_type_name(quant) << " time=" << ms_dequant << "ms avg_err=" << avg_err << std::endl; + + if (avg_err > 0.1) { + std::cerr << "first_error = " << first_err << std::endl; + std::cerr << "Actual result: " << std::endl << std::endl; + for (int i = std::max(0, first_err - 5); i < std::min((int)ne, first_err + 5); i++) { + std::cerr << ggml_fp16_to_fp32(x_chk[i]) << ", "; + } + std::cerr << std::endl << "Expected result: " << std::endl << std::endl; + for (int i = std::max(0, first_err - 5); i < std::min((int)ne, first_err + 5); i++) { + std::cerr << x[i] << ", "; + } + std::cerr << std::endl; + } ggml_vk_destroy_buffer(x_buf); ggml_vk_destroy_buffer(qx_buf); @@ -3964,6 +4221,190 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_ free(qx); free(x_chk); } + +static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, size_t n, size_t k, size_t batch, size_t num_it, size_t split_k, size_t shader_size, ggml_type quant) { +#ifdef GGML_VULKAN_DEBUG + std::cerr << "ggml_vk_test_dequant_matmul(" << m << ", " << n << ", " << k << ", " << batch << ", " << num_it << ", " << split_k << ", " << ggml_type_name(quant) << ")" << std::endl; +#endif + const size_t x_ne = m * k * batch; + const size_t y_ne = k * n * batch; + const size_t d_ne = m * n * batch; + + vk_pipeline p; + std::string shname; + if (shader_size == 0) { + p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_s; + shname = std::string(ggml_type_name(quant)) + "_ALIGNED_S"; + } else if (shader_size == 1) { + p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_m; + shname = std::string(ggml_type_name(quant)) + "_ALIGNED_M"; + } else if (shader_size == 2) { + p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_l; + shname = std::string(ggml_type_name(quant)) + "_ALIGNED_L"; + } else { + GGML_ASSERT(0); + } + + const size_t kpad = ggml_vk_align_size(k, p->align); + + if (k != kpad) { + if (shader_size == 0) { + p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->s; + shname = std::string(ggml_type_name(quant)) + "_S"; + } else if (shader_size == 1) { + p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->m; + shname = std::string(ggml_type_name(quant)) + "_M"; + } else if (shader_size == 2) { + p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->l; + shname = std::string(ggml_type_name(quant)) + "_L"; + } else { + GGML_ASSERT(0); + } + } + + const size_t x_sz = sizeof(float) * x_ne; + const size_t y_sz = sizeof(float) * y_ne; + const size_t qx_sz = x_ne * ggml_type_size(quant)/ggml_blck_size(quant); + const size_t d_sz = sizeof(float) * d_ne; + float * x = (float *) malloc(x_sz); + float * y = (float *) malloc(y_sz); + void * qx = malloc(qx_sz); + vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx, qx_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); + vk_buffer y_buf = ggml_vk_create_buffer_check(ctx, y_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); + vk_buffer d_buf = ggml_vk_create_buffer_check(ctx, d_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); + float * d = (float *) malloc(d_sz); + float * d_chk = (float *) malloc(d_sz); + + for (size_t i = 0; i < x_ne; i++) { + x[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; + } + + ggml_vk_quantize_data(x, qx, x_ne, quant); + + for (size_t i = 0; i < y_ne; i++) { + // y[i] = rand() / (float)RAND_MAX; + y[i] = (i % k == i / k) ? 1.0f : 0.0f; + } + + ggml_pipeline_allocate_descriptor_sets(ctx, p, num_it); + if (split_k > 1) { + ggml_pipeline_allocate_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, num_it); + + if (ctx->prealloc_split_k == nullptr || ctx->prealloc_split_k->size < sizeof(float) * d_ne * split_k) { + // Resize buffer + if (ctx->prealloc_split_k != nullptr) { + ggml_vk_destroy_buffer(ctx->prealloc_split_k); + } + ctx->prealloc_split_k = ggml_vk_create_buffer_check(ctx, sizeof(float) * d_ne * split_k, vk::MemoryPropertyFlagBits::eDeviceLocal); + } + } + + ggml_vk_buffer_write(ctx, qx_buf, 0, qx, qx_sz); + ggml_vk_buffer_write(ctx, y_buf, 0, y, y_sz); + + vk_context * subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); + for (size_t i = 0; i < num_it; i++) { + ggml_vk_ctx_begin(ctx, subctx); + ggml_vk_matmul(ctx, subctx, p, ggml_vk_subbuffer(qx_buf), ggml_vk_subbuffer(y_buf), ggml_vk_subbuffer(d_buf), ggml_vk_subbuffer(ctx->prealloc_split_k), m, n, k, k, k, m, split_k, batch, batch, batch, 1, 1, k*m, k*n, m*n); + ggml_vk_ctx_end(subctx); + } + + auto begin = std::chrono::high_resolution_clock::now(); + + ggml_vk_submit(subctx, ctx->fence); + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences"); + ctx->device->device.resetFences({ ctx->fence }); + + auto end = std::chrono::high_resolution_clock::now(); + + double time_ms = std::chrono::duration_cast(end-begin).count() / 1000.0; + ggml_vk_buffer_read(ctx, d_buf, 0, d, d_sz); + + ggml_init_params iparams = { + /*.mem_size =*/ 1024*1024*1024, + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + + ggml_context * ggml_ctx = ggml_init(iparams); + + ggml_tensor * src0_ggml = ggml_new_tensor_3d(ggml_ctx, quant, k, m, batch); + ggml_tensor * src1_ggml = ggml_new_tensor_3d(ggml_ctx, GGML_TYPE_F32, k, n, batch); + ggml_tensor * tensor_ggml = ggml_mul_mat(ggml_ctx, src0_ggml, src1_ggml); + + src0_ggml->data = qx; + src1_ggml->data = y; + tensor_ggml->data = d_chk; + + ctx->disable = true; + + ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx); + ggml_build_forward_expand(cgraph, tensor_ggml); + + ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 1); + + ctx->disable = false; + + ggml_free(ggml_ctx); + + double avg_err = 0.0; + int first_err_n = -1; + int first_err_m = -1; + int first_err_b = -1; + + for (size_t i = 0; i < m*n*batch; i++) { + double err = std::fabs(d[i] - d_chk[i]); + avg_err += err; + + if ((err > 0.05f || std::isnan(err)) && first_err_n == -1) { + first_err_b = i / (m * n); + first_err_n = (i % (m * n)) / m; + first_err_m = (i % (m * n)) % m; + } + } + + avg_err /= m * n; + + std::cerr << "TEST MMQ " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time_ms / num_it << "ms avg_err=" << avg_err << std::endl; + + if (avg_err > 0.1 || std::isnan(avg_err)) { + std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl; + std::cerr << "Actual result: " << std::endl << std::endl; + ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); + std::cerr << std::endl; + std::cerr << "Expected result: " << std::endl << std::endl; + ggml_vk_print_matrix_area(d_chk, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); + + if (split_k > 1) { + float * split_k_buf = (float *) malloc(sizeof(float) * d_ne * split_k); + ggml_vk_buffer_read(ctx, ctx->prealloc_split_k, 0, split_k_buf, sizeof(float) * d_ne * split_k); + + std::cerr << "d_buf0: " << std::endl << std::endl; + ggml_vk_print_matrix_area(split_k_buf, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); + + std::cerr << "d_buf1: " << std::endl << std::endl; + ggml_vk_print_matrix_area(split_k_buf + d_ne, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); + + std::cerr << "d_buf2: " << std::endl << std::endl; + ggml_vk_print_matrix_area(split_k_buf + 2 * d_ne, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); + + std::cerr << "d_buf3: " << std::endl << std::endl; + ggml_vk_print_matrix_area(split_k_buf + 3 * d_ne, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); + + free(split_k_buf); + } + } + + ggml_vk_destroy_buffer(qx_buf); + ggml_vk_destroy_buffer(y_buf); + ggml_vk_destroy_buffer(d_buf); + + free(x); + free(qx); + free(y); + free(d); + free(d_chk); +} #endif static ggml_tensor_extra_gpu * ggml_vk_tensor_create_extra(ggml_tensor * tensor) { @@ -3976,18 +4417,8 @@ static ggml_tensor_extra_gpu * ggml_vk_tensor_create_extra(ggml_tensor * tensor) return extra; } -static ggml_tensor * ggml_vk_find_last_use(const ggml_tensor * node, ggml_cgraph * graph) { - GGML_ASSERT(node != nullptr); - - for (int i = graph->n_nodes - 1; i >= 0; i--) { - for (int j = 0; j < GGML_MAX_SRC; j++) { - if (graph->nodes[i]->src[j] == node) { - return graph->nodes[i]; - } - } - } - - return nullptr; +static bool ggml_vk_cpu_assist_op(const ggml_tensor * node) { + return node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID; } static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggml_tensor * node){ @@ -3998,7 +4429,7 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm || (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_TYPE_GPU || node->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT)) || (node->src[1] != nullptr && (node->src[1]->backend == GGML_BACKEND_TYPE_GPU)); - if (ctx->disable || (!any_on_device && node->op != GGML_OP_MUL_MAT)) { + if (ctx->disable || (!any_on_device && !ggml_vk_cpu_assist_op(node))) { return; } @@ -4029,7 +4460,7 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm const bool f16_f32_kernel = use_src1 && src1->type == GGML_TYPE_F32; int split_k; - if (node->op == GGML_OP_MUL_MAT) { + if (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID) { split_k = ggml_vk_guess_split_k(ne01, ne11, ne10); } else { split_k = 1; @@ -4038,11 +4469,11 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm const uint32_t y_ne = ne10 * ne11; const uint32_t d_ne = ne20 * ne21; - const uint64_t qx_sz = use_src0 ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) * ne02 * ne03 : 0; - const uint64_t qy_sz = use_src1 ? ggml_vk_align_size(ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type), ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) * ne12 * ne13 : 0; - const uint64_t x_sz = use_src0 ? ggml_vk_align_size(sizeof(ggml_fp16_t) * x_ne, ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) * ne02 * ne03 : 0; - const uint64_t y_sz = use_src1 ? ggml_vk_align_size(f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne, ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) * ne12 * ne13 : 0; - uint64_t d_sz = ggml_vk_align_size(ggml_type_size(node->type) * d_ne, ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) * ne22 * ne23; + const uint64_t qx_sz = use_src0 ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ne02 * ne03 : 0; + const uint64_t qy_sz = use_src1 ? ggml_vk_align_size(ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ne12 * ne13 : 0; + const uint64_t x_sz = use_src0 ? ggml_vk_align_size(sizeof(ggml_fp16_t) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ne02 * ne03 : 0; + const uint64_t y_sz = use_src1 ? ggml_vk_align_size(f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ne12 * ne13 : 0; + uint64_t d_sz = ggml_vk_align_size(ggml_type_size(node->type) * d_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ne22 * ne23; const uint64_t split_k_size = split_k > 1 ? d_sz * 4 : 0; if (extra->buffer_gpu.expired()) { @@ -4070,6 +4501,7 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: case GGML_OP_ROPE: + case GGML_OP_ARGSORT: break; case GGML_OP_UNARY: switch (ggml_get_unary_op(node)) { @@ -4082,6 +4514,7 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm } break; case GGML_OP_MUL_MAT: + case GGML_OP_MUL_MAT_ID: if (ctx->prealloc_size_qx < qx_sz) { ctx->prealloc_size_qx = qx_sz; } @@ -4115,21 +4548,66 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { #endif #if defined(GGML_VULKAN_RUN_TESTS) ctx->staging = ggml_vk_create_buffer_check(ctx, 100ul * 1024ul * 1024ul, - vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached + vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent); ggml_vk_test_transfer(ctx, 8192 * 1000, false); ggml_vk_test_transfer(ctx, 8192 * 1000, true); - ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q4_0); - ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q4_1); - ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q5_0); - ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q5_1); - ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q8_0); - ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q2_K); - ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q3_K); - ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q4_K); - ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q5_K); - ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q6_K); + ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_F32); + ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_0); + ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_1); + ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_0); + ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_1); + ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q8_0); + ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q2_K); + ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q3_K); + ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_K); + ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_K); + ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q6_K); + + ggml_vk_test_matmul(ctx, 128, 512, 512, 2, 100, 1, 0); + ggml_vk_test_matmul(ctx, 128, 512, 512, 2, 100, 1, 1); + ggml_vk_test_matmul(ctx, 128, 512, 512, 2, 100, 1, 2); + ggml_vk_test_matmul(ctx, 128, 512, 512, 2, 100, 4, 0); + ggml_vk_test_matmul(ctx, 128, 512, 512, 2, 100, 4, 1); + ggml_vk_test_matmul(ctx, 128, 512, 512, 2, 100, 4, 2); + + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_0); + + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_1); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_1); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_1); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_1); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_1); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_1); + + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_0); + + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_1); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_1); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_1); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_1); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_1); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_1); + + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q8_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q8_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q8_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q8_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q8_0); + ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q8_0); + + std::cerr << std::endl; const std::vector vals { 8, 8, 8, @@ -4219,7 +4697,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod || (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_TYPE_GPU || node->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT)) || (node->src[1] != nullptr && node->src[1]->backend == GGML_BACKEND_TYPE_GPU); - if (ctx->disable || (!any_on_device && node->op != GGML_OP_MUL_MAT) || (node->op == GGML_OP_MUL_MAT && !any_on_device && !ggml_vk_can_mul_mat(node->src[0], node->src[1], node))) { + if (ctx->disable || (!any_on_device && !ggml_vk_cpu_assist_op(node)) || (ggml_vk_cpu_assist_op(node) && !any_on_device && !ggml_vk_can_mul_mat(node->src[0], node->src[1], node))) { return; } @@ -4231,6 +4709,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod const ggml_tensor * src0 = node->src[0]; const ggml_tensor * src1 = node->src[1]; + const ggml_tensor * src2 = node->src[2]; ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra; @@ -4265,7 +4744,9 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_SOFT_MAX: case GGML_OP_ROPE: case GGML_OP_MUL_MAT: + case GGML_OP_MUL_MAT_ID: case GGML_OP_NONE: + case GGML_OP_ARGSORT: break; default: if (any_on_device) { @@ -4276,7 +4757,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod } if (ctx->compute_ctx == nullptr) { - ctx->compute_ctx = ggml_vk_create_context(ctx, ctx->device.lock()->compute_queue); + ctx->compute_ctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); ggml_vk_ctx_begin(ctx, ctx->compute_ctx); } @@ -4347,16 +4828,25 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod break; case GGML_OP_SOFT_MAX: - ggml_vk_soft_max(ctx, ctx->compute_ctx, src0, src1, node); + ggml_vk_soft_max(ctx, ctx->compute_ctx, src0, src1, src2, node); break; case GGML_OP_ROPE: ggml_vk_rope(ctx, ctx->compute_ctx, src0, src1, node); + break; + case GGML_OP_ARGSORT: + ggml_vk_argsort(ctx, ctx->compute_ctx, src0, node); break; case GGML_OP_MUL_MAT: ggml_vk_mul_mat(ctx, ctx->compute_ctx, src0, src1, node); + break; + case GGML_OP_MUL_MAT_ID: + //ggml_vk_mul_mat_id(ctx, ctx->compute_ctx, src0, src1, node); + std::cerr << "ggml_vulkan: GGML_OP_MUL_MAT_ID not implemented yet." << std::endl; + GGML_ASSERT(false); + break; default: return; @@ -4383,7 +4873,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_ || (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU || tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT)) || (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_TYPE_GPU); - if (ctx->disable || (!any_on_device && tensor->op != GGML_OP_MUL_MAT)) { + if (ctx->disable || (!any_on_device && !ggml_vk_cpu_assist_op(tensor))) { return false; } @@ -4409,6 +4899,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_ case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: case GGML_OP_NONE: + case GGML_OP_ARGSORT: extra = (ggml_tensor_extra_gpu *) tensor->extra; break; @@ -4424,6 +4915,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_ } break; case GGML_OP_MUL_MAT: + case GGML_OP_MUL_MAT_ID: if (!any_on_device && !ggml_vk_can_mul_mat(tensor->src[0], tensor->src[1], tensor)) { return false; } @@ -4469,8 +4961,8 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_ } if (tensor == subctx.exit_tensor) { - VK_CHECK(ctx->device.lock()->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences"); - ctx->device.lock()->device.resetFences({ ctx->fence }); + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences"); + ctx->device->device.resetFences({ ctx->fence }); // Do staging buffer copies for (auto& cpy : subctx.out_memcpys) { @@ -4498,20 +4990,25 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) { } ctx->gc.temp_buffers.clear(); - for (auto * pipeline : ctx->gc.pipelines) { - ggml_pipeline_cleanup(*pipeline); + for (auto& pipeline : ctx->device->pipelines) { + if (pipeline.expired()) { + continue; + } + + vk_pipeline pl = pipeline.lock(); + ggml_pipeline_cleanup(pl); } - ggml_vk_queue_cleanup(ctx, ctx->device.lock()->compute_queue); - ggml_vk_queue_cleanup(ctx, ctx->device.lock()->transfer_queue); + ggml_vk_queue_cleanup(ctx, ctx->device->compute_queue); + ggml_vk_queue_cleanup(ctx, ctx->device->transfer_queue); for (size_t i = 0; i < ctx->gc.semaphores.size(); i++) { - ctx->device.lock()->device.destroySemaphore({ ctx->gc.semaphores[i].s }); + ctx->device->device.destroySemaphore({ ctx->gc.semaphores[i].s }); } ctx->gc.semaphores.clear(); for (size_t i = 0; i < ctx->gc.tl_semaphores.size(); i++) { - ctx->device.lock()->device.destroySemaphore({ ctx->gc.tl_semaphores[i].s }); + ctx->device->device.destroySemaphore({ ctx->gc.tl_semaphores[i].s }); } ctx->gc.tl_semaphores.clear(); ctx->semaphore_idx = 0; @@ -4519,7 +5016,7 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) { ctx->event_idx = 0; for (auto& event : ctx->gc.events) { - ctx->device.lock()->device.resetEvent(event); + ctx->device->device.resetEvent(event); } ctx->staging_offset = 0; @@ -4556,21 +5053,11 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) { ctx->staging_size = 0; for (auto& event : ctx->gc.events) { - ctx->device.lock()->device.destroyEvent(event); + ctx->device->device.destroyEvent(event); } ctx->gc.events.clear(); - for (auto* pipeline : ctx->gc.pipelines) { - ggml_vk_destroy_pipeline(ctx, pipeline); - } - ctx->gc.pipelines.clear(); - - ctx->device.lock()->device.destroyFence(ctx->fence); - - ctx->device.lock()->device.destroyCommandPool(ctx->device.lock()->compute_queue.pool); - if (!ctx->device.lock()->single_queue) { - ctx->device.lock()->device.destroyCommandPool(ctx->device.lock()->transfer_queue.pool); - } + ctx->device->device.destroyFence(ctx->fence); } GGML_CALL static int ggml_vk_get_device_count() { @@ -4781,7 +5268,6 @@ GGML_CALL static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t bu GGML_CALL static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) { if (ggml_backend_buffer_is_vk(src->buffer)) { - ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context; ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *) src->extra; ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; @@ -4793,6 +5279,8 @@ GGML_CALL static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t bu return true; } return false; + + UNUSED(buffer); } GGML_CALL static void ggml_backend_vk_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { @@ -4839,12 +5327,12 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer( GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *) buft->context; - return ctx->ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment; + return ctx->ctx->device->properties.limits.minStorageBufferOffsetAlignment; } GGML_CALL static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) { ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *) buft->context; - return ctx->ctx->device.lock()->max_memory_allocation_size; + return ctx->ctx->device->max_memory_allocation_size; } GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { @@ -4930,7 +5418,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_bu } GGML_CALL static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { - return vk_instance.contexts[0].device.lock()->properties.limits.minMemoryMapAlignment; + return vk_instance.contexts[0].device->properties.limits.minMemoryMapAlignment; UNUSED(buft); } @@ -4975,8 +5463,7 @@ GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend) { ggml_vk_cleanup(ctx); - // Release device - vk_instance.devices[ctx->idx].reset(); + ctx->device.reset(); ctx->initialized = false; vk_instance.initialized[idx] = false; @@ -5005,7 +5492,7 @@ GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, g if (ctx->transfer_ctx == nullptr) { // Initialize new transfer context - ctx->transfer_ctx = ggml_vk_create_context(ctx, ctx->device.lock()->transfer_queue); + ctx->transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); ggml_vk_ctx_begin(ctx, ctx->transfer_ctx); } @@ -5026,7 +5513,7 @@ GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, c if (ctx->transfer_ctx == nullptr) { // Initialize new transfer context - ctx->transfer_ctx = ggml_vk_create_context(ctx, ctx->device.lock()->transfer_queue); + ctx->transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); ggml_vk_ctx_begin(ctx, ctx->transfer_ctx); } @@ -5046,7 +5533,7 @@ GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, c if (ctx->transfer_ctx == nullptr) { // Initialize new transfer context - ctx->transfer_ctx = ggml_vk_create_context(ctx, ctx->device.lock()->transfer_queue); + ctx->transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue); ggml_vk_ctx_begin(ctx, ctx->transfer_ctx); } @@ -5076,8 +5563,8 @@ GGML_CALL static void ggml_backend_vk_synchronize(ggml_backend_t backend) { } ggml_vk_submit(ctx->transfer_ctx, ctx->fence); - VK_CHECK(ctx->device.lock()->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_backend_vk_synchronize waitForFences"); - ctx->device.lock()->device.resetFences({ ctx->fence }); + VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_backend_vk_synchronize waitForFences"); + ctx->device->device.resetFences({ ctx->fence }); for (auto& cpy : ctx->transfer_ctx->out_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); @@ -5086,7 +5573,7 @@ GGML_CALL static void ggml_backend_vk_synchronize(ggml_backend_t backend) { ctx->transfer_ctx = nullptr; } -GGML_CALL static bool ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { +GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; for (int i = 0; i < cgraph->n_nodes; i++) { @@ -5129,7 +5616,7 @@ GGML_CALL static bool ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml ggml_vk_graph_cleanup(ctx); - return true; + return GGML_STATUS_SUCCESS; UNUSED(backend); } @@ -5147,6 +5634,7 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const } break; case GGML_OP_MUL_MAT: + case GGML_OP_MUL_MAT_ID: { struct ggml_tensor * a; struct ggml_tensor * b; @@ -5220,6 +5708,7 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const case GGML_OP_CONT: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: + case GGML_OP_ARGSORT: return true; default: return false; @@ -5422,7 +5911,8 @@ static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tenso ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; - ggml_vk_buffer_read(ctx, extra->buffer_gpu, extra->offset, tensor_data, tensor_size); + vk_buffer buffer_gpu = extra->buffer_gpu.lock(); + ggml_vk_buffer_read(ctx, buffer_gpu, extra->offset, tensor_data, tensor_size); } std::cerr << "TENSOR CHECK " << name << " (" << tensor->name << "): " << ggml_op_name(tensor->op) << std::endl; @@ -5498,6 +5988,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_ ggml_tensor * src0 = tensor->src[0]; ggml_tensor * src1 = tensor->src[1]; + ggml_tensor * src2 = tensor->src[2]; struct ggml_init_params iparams = { /*.mem_size =*/ 1024*1024*1024, @@ -5509,13 +6000,16 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_ struct ggml_tensor * src0_clone = nullptr; struct ggml_tensor * src1_clone = nullptr; + struct ggml_tensor * src2_clone = nullptr; struct ggml_tensor * tensor_clone = nullptr; size_t src0_size; size_t src1_size; + size_t src2_size; void * src0_buffer; void * src1_buffer; + void * src2_buffer; if (src0 != nullptr) { src0_clone = ggml_dup_tensor(ggml_ctx, src0); @@ -5529,12 +6023,13 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_ memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS); } else if (src0->backend == GGML_BACKEND_TYPE_GPU) { ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src0->extra; + vk_buffer buffer_gpu = extra->buffer_gpu.lock(); uint64_t offset = extra->offset; if (!ggml_is_contiguous(src0) && ggml_vk_dim01_contiguous(src0)) { for (int i3 = 0; i3 < src0->ne[3]; i3++) { for (int i2 = 0; i2 < src0->ne[2]; i2++) { const int idx = i3*src0->ne[2] + i2; - ggml_vk_buffer_read(ctx, extra->buffer_gpu, offset + idx * src0->nb[2], ((char *)src0_clone->data + idx * src0_clone->nb[2]), src0->ne[1] * src0->nb[1]); + ggml_vk_buffer_read(ctx, buffer_gpu, offset + idx * src0->nb[2], ((char *)src0_clone->data + idx * src0_clone->nb[2]), src0->ne[1] * src0->nb[1]); } } @@ -5544,10 +6039,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_ src0_clone->nb[i] = src0_clone->nb[i - 1]*src0_clone->ne[i - 1]; } } else { - if (offset + src0_size >= extra->buffer_gpu->size) { - src0_size = extra->buffer_gpu->size - offset; + if (offset + src0_size >= buffer_gpu->size) { + src0_size = buffer_gpu->size - offset; } - ggml_vk_buffer_read(ctx, extra->buffer_gpu, offset, src0_clone->data, src0_size); + ggml_vk_buffer_read(ctx, buffer_gpu, offset, src0_clone->data, src0_size); memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS); } } else { @@ -5572,12 +6067,13 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_ memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS); } else if (src1->backend == GGML_BACKEND_TYPE_GPU) { ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src1->extra; + vk_buffer buffer_gpu = extra->buffer_gpu.lock(); uint64_t offset = extra->offset; if (!ggml_is_contiguous(src1) && ggml_vk_dim01_contiguous(src1)) { for (int i3 = 0; i3 < src1->ne[3]; i3++) { for (int i2 = 0; i2 < src1->ne[2]; i2++) { const int idx = i3*src1->ne[2] + i2; - ggml_vk_buffer_read(ctx, extra->buffer_gpu, offset + idx * src1->nb[2], ((char *)src1_clone->data + idx * src1_clone->nb[2]), src1->ne[1] * src1->nb[1]); + ggml_vk_buffer_read(ctx, buffer_gpu, offset + idx * src1->nb[2], ((char *)src1_clone->data + idx * src1_clone->nb[2]), src1->ne[1] * src1->nb[1]); } } @@ -5587,10 +6083,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_ src1_clone->nb[i] = src1_clone->nb[i - 1]*src1_clone->ne[i - 1]; } } else { - if (offset + src1_size >= extra->buffer_gpu->size) { - src1_size = extra->buffer_gpu->size - offset; + if (offset + src1_size >= buffer_gpu->size) { + src1_size = buffer_gpu->size - offset; } - ggml_vk_buffer_read(ctx, extra->buffer_gpu, offset, src1_clone->data, src1_size); + ggml_vk_buffer_read(ctx, buffer_gpu, offset, src1_clone->data, src1_size); memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS); } } else { @@ -5619,6 +6115,66 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_ ggml_vk_check_tensor(std::string(ggml_op_name(tensor->op)) + "->src1", src1_clone); } + if (src2 != nullptr) { + src2_clone = ggml_dup_tensor(ggml_ctx, src2); + + src2_size = ggml_nbytes(src2); + + src2_buffer = malloc(src2_size); + src2_clone->data = src2_buffer; + if (src2->backend == GGML_BACKEND_TYPE_CPU) { + memcpy(src2_clone->data, src2->data, src2_size); + memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS); + } else if (src2->backend == GGML_BACKEND_TYPE_GPU) { + ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src2->extra; + vk_buffer buf = extra->buffer_gpu.lock(); + uint64_t offset = extra->offset; + if (!ggml_is_contiguous(src2) && ggml_vk_dim01_contiguous(src2)) { + for (int i3 = 0; i3 < src2->ne[3]; i3++) { + for (int i2 = 0; i2 < src2->ne[2]; i2++) { + const int idx = i3*src2->ne[2] + i2; + ggml_vk_buffer_read(ctx, buf, offset + idx * src2->nb[2], ((char *)src2_clone->data + idx * src2_clone->nb[2]), src2->ne[1] * src2->nb[1]); + } + } + + src2_clone->nb[0] = src2->nb[0]; + src2_clone->nb[1] = src2->nb[1]; + for (int i = 2; i < GGML_MAX_DIMS; i++) { + src2_clone->nb[i] = src2_clone->nb[i - 1]*src2_clone->ne[i - 1]; + } + } else { + if (offset + src2_size >= buf->size) { + src2_size = buf->size - offset; + } + ggml_vk_buffer_read(ctx, buf, offset, src2_clone->data, src2_size); + memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS); + } + } else { + GGML_ASSERT(false); + } + + if (vk_output_tensor > 0 && vk_output_tensor == check_counter) { + ggml_vk_print_tensor(ctx, src2, "src2"); + std::cerr << "TENSOR CHECK: " << ggml_op_name(src2_clone->op) << " (check " << check_counter << ")" << std::endl; + std::cerr << "src2_clone=" << tensor << " src2_clone->backend: " << src2_clone->backend << " src2_clone->type: " << ggml_type_name(src2_clone->type) << " ne0=" << src2_clone->ne[0] << " nb0=" << src2_clone->nb[0] << " ne1=" << src2_clone->ne[1] << " nb1=" << src2_clone->nb[1] << " ne2=" << src2_clone->ne[2] << " nb2=" << src2_clone->nb[2] << " ne3=" << src2_clone->ne[3] << " nb3=" << src2_clone->nb[3] << std::endl; + if (src2->src[0] != nullptr) { + std::cerr << "src2->src[0]=" << src2->src[0] << " op=" << ggml_op_name(src2->src[0]->op) << " type=" << ggml_type_name(src2->src[0]->type) << " backend=" << src2->src[0]->backend << " ne0=" << src2->src[0]->ne[0] << " nb0=" << src2->src[0]->nb[0] << " ne1=" << src2->src[0]->ne[1] << " nb1=" << src2->src[0]->nb[1] << " ne2=" << src2->src[0]->ne[2] << " nb2=" << src2->src[0]->nb[2] << " ne3=" << src2->src[0]->ne[3] << " nb3=" << src2->src[0]->nb[3] << std::endl; + } + if (src2->src[1] != nullptr) { + std::cerr << "src2->src[1]=" << src2->src[1] << " op=" << ggml_op_name(src2->src[1]->op) << " type=" << ggml_type_name(src2->src[1]->type) << " backend=" << src2->src[1]->backend << " ne0=" << src2->src[1]->ne[0] << " nb0=" << src2->src[1]->nb[0] << " ne1=" << src2->src[1]->ne[1] << " nb1=" << src2->src[1]->nb[1] << " ne2=" << src2->src[1]->ne[2] << " nb2=" << src2->src[1]->nb[2] << " ne3=" << src2->src[1]->ne[3] << " nb3=" << src2->src[1]->nb[3] << std::endl; + } + std::cerr << std::endl << "Result:" << std::endl; + ggml_vk_print_tensor_area(src2_clone, src2_clone->data, 5, 5, 0, 0); + std::cerr << std::endl; + std::cerr << std::endl << "Result:" << std::endl; + ggml_vk_print_tensor_area(src2_clone, src2_clone->data, 5, 5, 1, 0); + std::cerr << std::endl; + std::vector done; + ggml_vk_print_graph_origin(src2_clone, done); + } + + ggml_vk_check_tensor(std::string(ggml_op_name(tensor->op)) + "->src2", src2_clone); + } if (tensor->op == GGML_OP_MUL_MAT) { tensor_clone = ggml_mul_mat(ggml_ctx, src0_clone, src1_clone); @@ -5638,7 +6194,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_ tensor_clone = ggml_rms_norm(ggml_ctx, src0_clone, *(float *)tensor->op_params); } else if (tensor->op == GGML_OP_SOFT_MAX) { if (src1 != nullptr) { - tensor_clone = ggml_soft_max_ext(ggml_ctx, src0_clone, src1_clone, *(float *)tensor->op_params); + tensor_clone = ggml_soft_max_ext(ggml_ctx, src0_clone, src1_clone, src2_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]); } else { tensor_clone = ggml_soft_max(ggml_ctx, src0_clone); } @@ -5721,6 +6277,9 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_ if (src1 != nullptr) { free(src1_buffer); } + if (src2 != nullptr) { + free(src1_buffer); + } ggml_free(ggml_ctx); } @@ -5747,11 +6306,12 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_ ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; - if (extra->offset + tensor_size >= extra->buffer_gpu->size) { - tensor_size = extra->buffer_gpu->size - (extra->offset); + vk_buffer buffer_gpu = extra->buffer_gpu.lock(); + if (extra->offset + tensor_size >= buffer_gpu->size) { + tensor_size = buffer_gpu->size - (extra->offset); } - ggml_vk_buffer_read(ctx, extra->buffer_gpu, extra->offset, tensor_data, tensor_size); + ggml_vk_buffer_read(ctx, buffer_gpu, extra->offset, tensor_data, tensor_size); } float first_error_result = -1.0f; diff --git a/ggml-vulkan.h b/ggml-vulkan.h index 9645126..e4317c3 100644 --- a/ggml-vulkan.h +++ b/ggml-vulkan.h @@ -10,6 +10,7 @@ extern "C" { #define GGML_VK_NAME "Vulkan" #define GGML_VK_MAX_DEVICES 16 +GGML_API void ggml_vk_instance_init(void); GGML_API void ggml_vk_init_cpu_assist(void); GGML_API void ggml_vk_preallocate_buffers_graph_cpu_assist(struct ggml_tensor * node); diff --git a/ggml.c b/ggml.c index f29b9f1..92b17ee 100644 --- a/ggml.c +++ b/ggml.c @@ -320,6 +320,17 @@ static ggml_fp16_t ggml_table_exp_f16[1 << 16]; // precomputed f32 table for f16 (256 KB) (ggml-impl.h) float ggml_table_f32_f16[1 << 16]; +const char * ggml_status_to_string(enum ggml_status status) { + switch (status) { + case GGML_STATUS_ALLOC_FAILED: return "GGML status: error (failed to allocate memory)"; + case GGML_STATUS_FAILED: return "GGML status: error (operation failed)"; + case GGML_STATUS_SUCCESS: return "GGML status: success"; + case GGML_STATUS_ABORTED: return "GGML status: warning (operation aborted)"; + } + + return "GGML status: unknown"; +} + // note: do not use these inside ggml.c // these are meant to be used via the ggml.h API float ggml_fp16_to_fp32(ggml_fp16_t x) { @@ -1822,6 +1833,8 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "POOL_2D", "UPSCALE", "PAD", + "ARANGE", + "TIMESTEP_EMBEDDING", "ARGSORT", "LEAKY_RELU", @@ -1850,7 +1863,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "CROSS_ENTROPY_LOSS_BACK", }; -static_assert(GGML_OP_COUNT == 72, "GGML_OP_COUNT != 72"); +static_assert(GGML_OP_COUNT == 74, "GGML_OP_COUNT != 74"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -1908,6 +1921,8 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "pool_2d(x)", "upscale(x)", "pad(x)", + "arange(start, stop, step)", + "timestep_embedding(timesteps, dim, max_period)", "argsort(x)", "leaky_relu(x)", @@ -1936,7 +1951,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "cross_entropy_loss_back(x,y)", }; -static_assert(GGML_OP_COUNT == 72, "GGML_OP_COUNT != 72"); +static_assert(GGML_OP_COUNT == 74, "GGML_OP_COUNT != 74"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -2139,7 +2154,10 @@ void ggml_numa_init(enum ggml_numa_strategy numa_flag) { getcpu_ret = getcpu(¤t_cpu, &g_state.numa.current_node); #else // old glibc doesn't have a wrapper for this call. Fall back on direct syscall - getcpu_ret = syscall(SYS_getcpu,¤t_cpu,&g_state.numa.current_node); +# if !defined(SYS_getcpu) && defined(SYS_get_cpu) +# define SYS_getcpu SYS_get_cpu // some older glibc versions use this name +# endif + getcpu_ret = syscall(SYS_getcpu, ¤t_cpu, &g_state.numa.current_node); #endif if (g_state.numa.n_nodes < 1 || g_state.numa.total_cpus < 1 || getcpu_ret != 0) { @@ -2895,11 +2913,21 @@ static int32_t ggml_get_op_params_i32(const struct ggml_tensor * tensor, uint32_ return ((const int32_t *)(tensor->op_params))[i]; } +static float ggml_get_op_params_f32(const struct ggml_tensor * tensor, uint32_t i) { + assert(i < GGML_MAX_OP_PARAMS / sizeof(float)); + return ((const float *)(tensor->op_params))[i]; +} + static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) { assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t)); ((int32_t *)(tensor->op_params))[i] = value; } +static void ggml_set_op_params_f32(struct ggml_tensor * tensor, uint32_t i, float value) { + assert(i < GGML_MAX_OP_PARAMS / sizeof(float)); + ((float *)(tensor->op_params))[i] = value; +} + struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor) { memset(tensor->data, 0, ggml_nbytes(tensor)); return tensor; @@ -5898,6 +5926,55 @@ struct ggml_tensor * ggml_upscale( return ggml_upscale_impl(ctx, a, scale_factor); } +struct ggml_tensor * ggml_arange( + struct ggml_context * ctx, + float start, + float stop, + float step) { + + GGML_ASSERT(stop > start); + + const int64_t steps = (int64_t) ceilf((stop - start) / step); + + struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps); + + result->op = GGML_OP_ARANGE; + ggml_set_op_params_f32(result, 0, start); + ggml_set_op_params_f32(result, 1, stop); + ggml_set_op_params_f32(result, 2, step); + + return result; +} + +struct ggml_tensor * ggml_timestep_embedding( + struct ggml_context * ctx, + struct ggml_tensor * timesteps, + int dim, + int max_period) { + bool is_node = false; + + if (timesteps->grad) { + GGML_ASSERT(false); // TODO: implement backward + is_node = true; + } + + int actual_dim = dim; + if (dim % 2 != 0) { + actual_dim = dim + 1; + } + + struct ggml_tensor * result = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, actual_dim, timesteps->ne[0]); + + result->op = GGML_OP_TIMESTEP_EMBEDDING; + ggml_set_op_params_i32(result, 0, dim); + ggml_set_op_params_i32(result, 1, max_period); + + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = timesteps; + + return result; +} + // ggml_argsort struct ggml_tensor * ggml_argsort( @@ -10231,7 +10308,7 @@ static void ggml_compute_forward_group_norm_f32( int n_channels = src0->ne[2]; int n_groups = dst->op_params[0]; int n_channels_per_group = (n_channels + n_groups - 1) / n_groups; - for (int i = ith; i < n_groups; i+=nth) { + for (int i = ith; i < n_groups; i += nth) { int start = i * n_channels_per_group; int end = start + n_channels_per_group; if (end > n_channels) { @@ -10245,28 +10322,32 @@ static void ggml_compute_forward_group_norm_f32( for (int64_t i01 = 0; i01 < ne01; i01++) { const float * x = (float *)((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03); + ggml_float sumr = 0.0; for (int64_t i00 = 0; i00 < ne00; i00++) { - sum += (ggml_float)x[i00]; + sumr += (ggml_float)x[i00]; } + sum += sumr; } } - float mean = sum / (ne00 * ne01 * step); - ggml_float sum2 = 0.0; + const float mean = sum / (ne00 * ne01 * step); + ggml_float sum2 = 0.0; for (int64_t i02 = start; i02 < end; i02++) { for (int64_t i01 = 0; i01 < ne01; i01++) { const float * x = (float *)((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03); float * y = (float *)((char *) dst->data + i01 * nb1 + i02 * nb2 + i03 * nb3); + ggml_float sumr = 0.0; for (int64_t i00 = 0; i00 < ne00; i00++) { float v = x[i00] - mean; y[i00] = v; - sum2 += (ggml_float)(v * v); + sumr += (ggml_float)(v * v); } + sum2 += sumr; } } - float variance = sum2 / (ne00 * ne01 * step); + const float variance = sum2 / (ne00 * ne01 * step); const float scale = 1.0f / sqrtf(variance + eps); for (int64_t i02 = start; i02 < end; i02++) { @@ -13547,6 +13628,106 @@ static void ggml_compute_forward_pad( } } + +// ggml_compute_forward_arange + +static void ggml_compute_forward_arange_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) { + return; + } + + GGML_ASSERT(dst->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const float start = ggml_get_op_params_f32(dst, 0); + const float stop = ggml_get_op_params_f32(dst, 1); + const float step = ggml_get_op_params_f32(dst, 2); + + const int64_t steps = (int64_t) ceilf((stop - start) / step); + + GGML_ASSERT(ggml_nelements(dst) == steps); + + for (int64_t i = ith; i < steps; i+= nth) { + float value = start + step * i; + ((float *)dst->data)[i] = value; + } +} + +static void ggml_compute_forward_arange( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + switch (dst->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_arange_f32(params, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + +static void ggml_compute_forward_timestep_embedding_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) { + return; + } + + const struct ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + const int dim = ggml_get_op_params_i32(dst, 0); + const int max_period = ggml_get_op_params_i32(dst, 1); + + int half = dim / 2; + + for (int64_t i = 0; i < ne00; i++) { + float * embed_data = (float *)((char *) dst->data + i*nb1); + for (int64_t j = ith; j < half; j += nth) { + float timestep = ((float *)src0->data)[i]; + float freq = (float)expf(-logf(max_period) * j / half); + float arg = timestep * freq; + embed_data[j] = cosf(arg); + embed_data[j + half] = sinf(arg); + } + if (dim % 2 != 0 && ith == 0) { + embed_data[dim] = 0.f; + } + } +} + +static void ggml_compute_forward_timestep_embedding( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_timestep_embedding_f32(params, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_argsort static void ggml_compute_forward_argsort_f32( @@ -15615,6 +15796,14 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_pad(params, tensor); } break; + case GGML_OP_ARANGE: + { + ggml_compute_forward_arange(params, tensor); + } break; + case GGML_OP_TIMESTEP_EMBEDDING: + { + ggml_compute_forward_timestep_embedding(params, tensor); + } break; case GGML_OP_ARGSORT: { ggml_compute_forward_argsort(params, tensor); @@ -16617,6 +16806,14 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { GGML_ASSERT(false); // TODO: not implemented } break; + case GGML_OP_ARANGE: + { + GGML_ASSERT(false); // TODO: not implemented + } break; + case GGML_OP_TIMESTEP_EMBEDDING: + { + GGML_ASSERT(false); // TODO: not implemented + } break; case GGML_OP_ARGSORT: { GGML_ASSERT(false); // TODO: not implemented @@ -17217,6 +17414,7 @@ struct ggml_compute_state { ggml_thread_t thrd; int ith; struct ggml_compute_state_shared * shared; + enum ggml_status ec; }; static void ggml_graph_compute_perf_stats_node(struct ggml_tensor * node, const struct ggml_compute_state_shared * st) { @@ -17368,6 +17566,14 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { { n_tasks = n_threads; } break; + case GGML_OP_ARANGE: + { + n_tasks = n_threads; + } break; + case GGML_OP_TIMESTEP_EMBEDDING: + { + n_tasks = n_threads; + } break; case GGML_OP_ARGSORT: { n_tasks = n_threads; @@ -17502,7 +17708,8 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { while (true) { if (cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) { state->shared->node_n += 1; - return (thread_ret_t) GGML_EXIT_ABORTED; + state->ec = GGML_STATUS_ABORTED; + return 0; } if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) { @@ -17624,7 +17831,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { } } - return GGML_EXIT_SUCCESS; + return 0; } struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threads) { @@ -17820,7 +18027,7 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa return cplan; } -int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { +enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { { GGML_ASSERT(cplan); GGML_ASSERT(cplan->n_threads > 0); @@ -17864,6 +18071,7 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { .thrd = 0, .ith = j, .shared = &state_shared, + .ec = GGML_STATUS_SUCCESS, }; const int rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_thread, &workers[j]); @@ -17874,12 +18082,14 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { workers[0].ith = 0; workers[0].shared = &state_shared; + workers[0].ec = GGML_STATUS_SUCCESS; const int64_t perf_start_cycles = ggml_perf_cycles(); const int64_t perf_start_time_us = ggml_perf_time_us(); // this is a work thread too - int compute_status = (size_t) ggml_graph_compute_thread(&workers[0]); + ggml_graph_compute_thread(&workers[0]); + enum ggml_status compute_status = workers[0].ec; // don't leave affinity set on the main thread clear_numa_thread_affinity(); @@ -17889,6 +18099,8 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { for (int j = 1; j < n_threads; j++) { const int rc = ggml_thread_join(workers[j].thrd, NULL); GGML_ASSERT(rc == 0); + if (workers[j].ec != GGML_STATUS_SUCCESS) + compute_status = workers[j].ec; } } @@ -17916,14 +18128,14 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { return compute_status; } -void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads) { +enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads) { struct ggml_cplan cplan = ggml_graph_plan(cgraph, n_threads); struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_WORK_BUFFER, cplan.work_size); cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs; - ggml_graph_compute(cgraph, &cplan); + return ggml_graph_compute(cgraph, &cplan); } struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name) { diff --git a/ggml.h b/ggml.h index 0a6d3c0..0ea4f88 100644 --- a/ggml.h +++ b/ggml.h @@ -315,6 +315,16 @@ extern "C" { #endif + enum ggml_status { + GGML_STATUS_ALLOC_FAILED = -2, + GGML_STATUS_FAILED = -1, + GGML_STATUS_SUCCESS = 0, + GGML_STATUS_ABORTED = 1, + }; + + // get ggml_status name string + GGML_API GGML_CALL const char * ggml_status_to_string(enum ggml_status status); + typedef uint16_t ggml_fp16_t; // convert FP16 <-> FP32 @@ -454,6 +464,8 @@ extern "C" { GGML_OP_POOL_2D, GGML_OP_UPSCALE, // nearest interpolate GGML_OP_PAD, + GGML_OP_ARANGE, + GGML_OP_TIMESTEP_EMBEDDING, GGML_OP_ARGSORT, GGML_OP_LEAKY_RELU, @@ -1661,6 +1673,15 @@ extern "C" { int p2, int p3); + // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151 + // timesteps: [N,] + // return: [N, dim] + GGML_API struct ggml_tensor * ggml_timestep_embedding( + struct ggml_context * ctx, + struct ggml_tensor * timesteps, + int dim, + int max_period); + // sort rows enum ggml_sort_order { GGML_SORT_ORDER_ASC, @@ -1672,6 +1693,12 @@ extern "C" { struct ggml_tensor * a, enum ggml_sort_order order); + GGML_API struct ggml_tensor * ggml_arange( + struct ggml_context * ctx, + float start, + float stop, + float step); + // top k elements per row GGML_API struct ggml_tensor * ggml_top_k( struct ggml_context * ctx, @@ -1923,12 +1950,11 @@ extern "C" { // ggml_graph_plan() has to be called before ggml_graph_compute() // when plan.work_size > 0, caller must allocate memory for plan.work_data - GGML_API struct ggml_cplan ggml_graph_plan (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/); - GGML_API int ggml_graph_compute( struct ggml_cgraph * cgraph, struct ggml_cplan * cplan); - + GGML_API struct ggml_cplan ggml_graph_plan (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/); + GGML_API enum ggml_status ggml_graph_compute ( struct ggml_cgraph * cgraph, struct ggml_cplan * cplan); // same as ggml_graph_compute() but the work data is allocated as a part of the context // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data - GGML_API void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads); + GGML_API enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads); GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name); diff --git a/whisper.cpp b/whisper.cpp index 2e0a6e2..289d703 100644 --- a/whisper.cpp +++ b/whisper.cpp @@ -187,7 +187,7 @@ static bool ggml_graph_compute_helper( ggml_backend_metal_set_n_cb(backend, n_threads); } #endif - return ggml_backend_graph_compute(backend, graph); + return ggml_backend_graph_compute(backend, graph) == GGML_STATUS_SUCCESS; } // faster matrix multiplications for tensors that do not have dimension 0 divisible by "pad" @@ -413,7 +413,7 @@ struct whisper_batch { whisper_token * token; whisper_pos * pos; - int32_t * n_seq_id; + int32_t * n_seq_id; // always 1, here for consistency with llama.cpp whisper_seq_id ** seq_id; // null terminated int8_t * logits; }; @@ -3852,7 +3852,7 @@ const char * whisper_print_system_info(void) { s += "VSX = " + std::to_string(ggml_cpu_has_vsx()) + " | "; s += "CUDA = " + std::to_string(ggml_cpu_has_cublas()) + " | "; s += "COREML = " + std::to_string(whisper_has_coreml()) + " | "; - s += "OPENVINO = " + std::to_string(whisper_has_openvino()) + " | "; + s += "OPENVINO = " + std::to_string(whisper_has_openvino()) ; return s.c_str(); } @@ -4759,6 +4759,19 @@ static void whisper_process_logits( #endif } +static bool whisper_sequence_tokens_equal(const whisper_sequence & a, const whisper_sequence & b) { + if (a.tokens.size() != b.tokens.size()) { + return false; + } + // sequences are more likely to diverge at the end + for (int i = a.tokens.size() - 1; i >= 0; i--) { + if (a.tokens[i].id != b.tokens[i].id) { + return false; + } + } + return true; +} + static whisper_token_data whisper_sample_token( whisper_context & ctx, const whisper_decoder & decoder, @@ -5357,7 +5370,10 @@ int whisper_full_with_state( beam_candidates.begin(), beam_candidates.end(), [](const beam_candidate & a, const beam_candidate & b) { - return a.sequence.sum_logprobs_all > b.sequence.sum_logprobs_all; + if (a.sequence.sum_logprobs_all != b.sequence.sum_logprobs_all) { + return a.sequence.sum_logprobs_all > b.sequence.sum_logprobs_all; + } + return a.decoder_idx < b.decoder_idx; }); uint32_t cur_c = 0; @@ -5375,7 +5391,7 @@ int whisper_full_with_state( auto & cur = beam_candidates[cur_c++]; - while (beam_candidates.size() > cur_c && beam_candidates[cur_c].sequence.sum_logprobs_all == cur.sequence.sum_logprobs_all && i > 0) { + while (beam_candidates.size() > cur_c && whisper_sequence_tokens_equal(beam_candidates[cur_c].sequence, cur.sequence) && i > 0) { ++cur_c; }