diff --git a/examples/common-ggml.cpp b/examples/common-ggml.cpp index 141c6a2..9215dbe 100644 --- a/examples/common-ggml.cpp +++ b/examples/common-ggml.cpp @@ -6,7 +6,6 @@ static const std::map GGML_FTYPE_MAP = { {"q4_0", GGML_FTYPE_MOSTLY_Q4_0}, {"q4_1", GGML_FTYPE_MOSTLY_Q4_1}, - {"q4_2", GGML_FTYPE_MOSTLY_Q4_2}, {"q5_0", GGML_FTYPE_MOSTLY_Q5_0}, {"q5_1", GGML_FTYPE_MOSTLY_Q5_1}, {"q8_0", GGML_FTYPE_MOSTLY_Q8_0}, @@ -46,7 +45,6 @@ bool ggml_common_quantize_0( switch (ftype) { case GGML_FTYPE_MOSTLY_Q4_0: qtype = GGML_TYPE_Q4_0; break; case GGML_FTYPE_MOSTLY_Q4_1: qtype = GGML_TYPE_Q4_1; break; - case GGML_FTYPE_MOSTLY_Q4_2: qtype = GGML_TYPE_Q4_2; break; case GGML_FTYPE_MOSTLY_Q5_0: qtype = GGML_TYPE_Q5_0; break; case GGML_FTYPE_MOSTLY_Q5_1: qtype = GGML_TYPE_Q5_1; break; case GGML_FTYPE_MOSTLY_Q8_0: qtype = GGML_TYPE_Q8_0; break; @@ -171,10 +169,6 @@ bool ggml_common_quantize_0( { cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data()); } break; - case GGML_TYPE_Q4_2: - { - cur_size = ggml_quantize_q4_2(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data()); - } break; case GGML_TYPE_Q5_0: { cur_size = ggml_quantize_q5_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data()); diff --git a/examples/common.cpp b/examples/common.cpp index 019a8ef..a8461fb 100644 --- a/examples/common.cpp +++ b/examples/common.cpp @@ -38,6 +38,20 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { } else if (arg == "-h" || arg == "--help") { gpt_print_usage(argc, argv, params); exit(0); + } else if (arg == "-f" || arg == "--file") { + if (++i > argc) { + fprintf(stderr, "Invalid file param"); + break; + } + std::ifstream file(argv[i]); + if (!file) { + fprintf(stderr, "error: failed to open file '%s'\n", argv[i]); + break; + } + std::copy(std::istreambuf_iterator(file), std::istreambuf_iterator(), back_inserter(params.prompt)); + if (params.prompt.back() == '\n') { + params.prompt.pop_back(); + } } else { fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); gpt_print_usage(argc, argv, params); @@ -57,6 +71,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads); fprintf(stderr, " -p PROMPT, --prompt PROMPT\n"); fprintf(stderr, " prompt to start generation with (default: random)\n"); + fprintf(stderr, " -f FNAME, --file FNAME\n"); + fprintf(stderr, " load prompt from a file\n"); fprintf(stderr, " -n N, --n_predict N number of tokens to predict (default: %d)\n", params.n_predict); fprintf(stderr, " --top_k N top-k sampling (default: %d)\n", params.top_k); fprintf(stderr, " --top_p N top-p sampling (default: %.1f)\n", params.top_p); @@ -192,6 +208,10 @@ std::map json_parse(const std::string & fname) { return result; } +void gpt_vocab::add_special_token(const std::string & token) { + special_tokens.push_back(token); +} + std::vector gpt_tokenize(const gpt_vocab & vocab, const std::string & text) { std::vector words; @@ -200,6 +220,20 @@ std::vector gpt_tokenize(const gpt_vocab & vocab, const std::stri std::string str = text; std::string pat = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)"; + // Generate the subpattern from the special_tokens vector if it's not empty + if (!vocab.special_tokens.empty()) { + std::string special_tokens_subpattern; + for (const auto & token : vocab.special_tokens) { + if (!special_tokens_subpattern.empty()) { + special_tokens_subpattern += "|"; + } + special_tokens_subpattern += token; + } + + // Modify the regex pattern with the generated special tokens subpattern + pat = special_tokens_subpattern + "|" + pat; + } + std::regex re(pat); std::smatch m; diff --git a/examples/common.h b/examples/common.h index 48252ce..29d0792 100644 --- a/examples/common.h +++ b/examples/common.h @@ -53,6 +53,9 @@ struct gpt_vocab { std::map token_to_id; std::map id_to_token; + std::vector special_tokens; + + void add_special_token(const std::string & token); }; // poor-man's JSON parsing diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index dc20e4e..f2fdb0f 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -25,7 +25,7 @@ struct whisper_hparams { int32_t n_text_head = 6; int32_t n_text_layer = 4; int32_t n_mels = 80; - int32_t f16 = 1; + int32_t ftype = 1; }; struct whisper_filters { @@ -79,7 +79,10 @@ bool whisper_model_quantize(const std::string & fname_inp, const std::string & f finp.read((char *) &hparams.n_text_head, sizeof(hparams.n_text_head)); finp.read((char *) &hparams.n_text_layer, sizeof(hparams.n_text_layer)); finp.read((char *) &hparams.n_mels, sizeof(hparams.n_mels)); - finp.read((char *) &hparams.f16, sizeof(hparams.f16)); + finp.read((char *) &hparams.ftype, sizeof(hparams.ftype)); + + const int32_t qntvr_src = hparams.ftype / GGML_QNT_VERSION_FACTOR; + const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype; fprintf(stderr, "%s: n_vocab = %d\n", __func__, hparams.n_vocab); fprintf(stderr, "%s: n_audio_ctx = %d\n", __func__, hparams.n_audio_ctx); @@ -91,7 +94,10 @@ bool whisper_model_quantize(const std::string & fname_inp, const std::string & f fprintf(stderr, "%s: n_text_head = %d\n", __func__, hparams.n_text_head); fprintf(stderr, "%s: n_text_layer = %d\n", __func__, hparams.n_text_layer); fprintf(stderr, "%s: n_mels = %d\n", __func__, hparams.n_mels); - fprintf(stderr, "%s: f16 = %d\n", __func__, hparams.f16); + fprintf(stderr, "%s: ftype (src) = %d\n", __func__, hparams.ftype); + fprintf(stderr, "%s: qntvr (src) = %d\n", __func__, qntvr_src); + fprintf(stderr, "%s: ftype (dst) = %d\n", __func__, ftype_dst); + fprintf(stderr, "%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION); fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab)); fout.write((char *) &hparams.n_audio_ctx, sizeof(hparams.n_audio_ctx)); @@ -103,7 +109,7 @@ bool whisper_model_quantize(const std::string & fname_inp, const std::string & f fout.write((char *) &hparams.n_text_head, sizeof(hparams.n_text_head)); fout.write((char *) &hparams.n_text_layer, sizeof(hparams.n_text_layer)); fout.write((char *) &hparams.n_mels, sizeof(hparams.n_mels)); - fout.write((char *) &ftype, sizeof(hparams.f16)); + fout.write((char *) &ftype_dst, sizeof(hparams.ftype)); } // load mel filters diff --git a/ggml-cuda.cu b/ggml-cuda.cu index e8a1e77..eb9f0df 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -32,9 +32,15 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size"); } \ } while (0) +typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, float & v0, float & v1); typedef void (*to_fp32_cuda_t)(const void * x, float * y, int k, cudaStream_t stream); +typedef void (*dequantize_mul_mat_vec_cuda_t)(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream); + +// QK = number of values after dequantization +// QR = QK / number of values before dequantization #define QK4_0 32 +#define QR4_0 2 typedef struct { float d; // delta uint8_t qs[QK4_0 / 2]; // nibbles / quants @@ -42,6 +48,7 @@ typedef struct { static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding"); #define QK4_1 32 +#define QR4_1 2 typedef struct { float d; // delta float m; // min @@ -49,14 +56,8 @@ typedef struct { } block_q4_1; static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding"); -#define QK4_2 16 -typedef struct { - half d; // delta - uint8_t qs[QK4_2 / 2]; // nibbles / quants -} block_q4_2; -static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding"); - #define QK5_0 32 +#define QR5_0 2 typedef struct { half d; // delta uint8_t qh[4]; // 5-th bit of quants @@ -65,6 +66,7 @@ typedef struct { static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding"); #define QK5_1 32 +#define QR5_1 2 typedef struct { half d; // delta half m; // min @@ -74,36 +76,121 @@ typedef struct { static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding"); #define QK8_0 32 +#define QR8_0 1 typedef struct { float d; // delta int8_t qs[QK8_0]; // quants } block_q8_0; static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding"); +#define CUDA_DMMV_BLOCK_SIZE 32 + +static __device__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){ + const block_q4_0 * x = (const block_q4_0 *) vx; + + const float d = x[ib].d; + + const uint8_t vui = x[ib].qs[iqs]; + + const int8_t vi0 = vui & 0xF; + const int8_t vi1 = vui >> 4; + + v0 = (vi0 - 8)*d; + v1 = (vi1 - 8)*d; +} + +static __device__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, float & v0, float & v1){ + const block_q4_1 * x = (const block_q4_1 *) vx; + + const float d = x[ib].d; + const float m = x[ib].m; + + const uint8_t vui = x[ib].qs[iqs]; + + const int8_t vi0 = vui & 0xF; + const int8_t vi1 = vui >> 4; + + v0 = vi0*d + m; + v1 = vi1*d + m; +} + +static __device__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){ + const block_q5_0 * x = (const block_q5_0 *) vx; + + const float d = x[ib].d; + + uint32_t qh; + memcpy(&qh, x[ib].qh, sizeof(qh)); + + const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10; + + const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0) - 16; + const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1) - 16; + + v0 = x0*d; + v1 = x1*d; +} + +static __device__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, float & v0, float & v1){ + const block_q5_1 * x = (const block_q5_1 *) vx; + + const float d = x[ib].d; + const float m = x[ib].m; + + uint32_t qh; + memcpy(&qh, x[ib].qh, sizeof(qh)); + + const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10; + + const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0); + const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1); + + v0 = x0*d + m; + v1 = x1*d + m; +} + +static __device__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){ + const block_q8_0 * x = (const block_q8_0 *) vx; + + const float d = x[ib].d; + + const int8_t vi0 = x[ib].qs[iqs + 0]; + const int8_t vi1 = x[ib].qs[iqs + 1]; + + v0 = vi0*d; + v1 = vi1*d; +} + +static __device__ void convert_f16(const void * vx, const int ib, const int iqs, float & v0, float & v1){ + const half * x = (const half *) vx; + + v0 = __half2float(x[ib + 0]); + v1 = __half2float(x[ib + 1]); +} + static __global__ void dequantize_block_q4_0(const void * vx, float * y) { + static const int qk = QK4_0; + const block_q4_0 * x = (const block_q4_0 *) vx; const int i = blockIdx.x; const float d = x[i].d; - const uint8_t * pp = x[i].qs; + for (int j = 0; j < qk/2; ++j) { + const int x0 = (x[i].qs[j] & 0xf) - 8; + const int x1 = (x[i].qs[j] >> 4) - 8; - for (int l = 0; l < QK4_0; l += 2) { - const uint8_t vi = pp[l/2]; - - const int8_t vi0 = vi & 0xf; - const int8_t vi1 = vi >> 4; - - const float v0 = (vi0 - 8)*d; - const float v1 = (vi1 - 8)*d; - - y[i*QK4_0 + l + 0] = v0; - y[i*QK4_0 + l + 1] = v1; + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; } } static __global__ void dequantize_block_q4_1(const void * vx, float * y) { + static const int qk = QK4_1; + const block_q4_1 * x = (const block_q4_1 *) vx; const int i = blockIdx.x; @@ -111,75 +198,42 @@ static __global__ void dequantize_block_q4_1(const void * vx, float * y) { const float d = x[i].d; const float m = x[i].m; - const uint8_t * pp = x[i].qs; + for (int j = 0; j < qk/2; ++j) { + const int x0 = (x[i].qs[j] & 0xf); + const int x1 = (x[i].qs[j] >> 4); - for (int l = 0; l < QK4_1; l += 2) { - const uint8_t vi = pp[l/2]; - - const int8_t vi0 = vi & 0xf; - const int8_t vi1 = vi >> 4; - - const float v0 = vi0*d + m; - const float v1 = vi1*d + m; - - y[i*QK4_1 + l + 0] = v0; - y[i*QK4_1 + l + 1] = v1; - } -} - -static __global__ void dequantize_block_q4_2(const void * vx, float * y) { - const block_q4_2 * x = (const block_q4_2 *) vx; - - const int i = blockIdx.x; - - const float d = x[i].d; - - const uint8_t * pp = x[i].qs; - - for (int l = 0; l < QK4_2; l += 2) { - const uint8_t vi = pp[l/2]; - - const int8_t vi0 = vi & 0xf; - const int8_t vi1 = vi >> 4; - - const float v0 = (vi0 - 8)*d; - const float v1 = (vi1 - 8)*d; - - y[i*QK4_2 + l + 0] = v0; - y[i*QK4_2 + l + 1] = v1; + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; } } static __global__ void dequantize_block_q5_0(const void * vx, float * y) { + static const int qk = QK5_0; + const block_q5_0 * x = (const block_q5_0 *) vx; const int i = blockIdx.x; const float d = x[i].d; - const uint8_t * pp = x[i].qs; - uint32_t qh; memcpy(&qh, x[i].qh, sizeof(qh)); - for (int l = 0; l < QK5_0; l += 2) { - const uint8_t vi = pp[l/2]; + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - const int8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4; - const int8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4; + const int32_t x0 = ((x[i].qs[j] & 0xf) | xh_0) - 16; + const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; - const int8_t vi0 = ((vi & 0xf) | vh0); - const int8_t vi1 = ((vi >> 4) | vh1); - - const float v0 = (vi0 - 16)*d; - const float v1 = (vi1 - 16)*d; - - y[i*QK5_0 + l + 0] = v0; - y[i*QK5_0 + l + 1] = v1; + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; } } static __global__ void dequantize_block_q5_1(const void * vx, float * y) { + static const int qk = QK5_1; + const block_q5_1 * x = (const block_q5_1 *) vx; const int i = blockIdx.x; @@ -187,41 +241,70 @@ static __global__ void dequantize_block_q5_1(const void * vx, float * y) { const float d = x[i].d; const float m = x[i].m; - const uint8_t * pp = x[i].qs; - uint32_t qh; memcpy(&qh, x[i].qh, sizeof(qh)); - for (int l = 0; l < QK5_1; l += 2) { - const uint8_t vi = pp[l/2]; + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - const int8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4; - const int8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4; + const int x0 = (x[i].qs[j] & 0xf) | xh_0; + const int x1 = (x[i].qs[j] >> 4) | xh_1; - const int8_t vi0 = (vi & 0xf) | vh0; - const int8_t vi1 = (vi >> 4) | vh1; - - const float v0 = vi0*d + m; - const float v1 = vi1*d + m; - - y[i*QK5_1 + l + 0] = v0; - y[i*QK5_1 + l + 1] = v1; + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; } } static __global__ void dequantize_block_q8_0(const void * vx, float * y) { + static const int qk = QK8_0; + const block_q8_0 * x = (const block_q8_0 *) vx; const int i = blockIdx.x; const float d = x[i].d; - const int8_t * pp = x[i].qs; + for (int j = 0; j < qk; ++j) { + y[i*qk + j] = x[i].qs[j]*d; + } +} - for (int l = 0; l < QK8_0; l++) { - const int8_t vi = pp[l]; +template +static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y, float * dst, const int ncols) { + const int row = blockIdx.x; + const int tid = threadIdx.x; - y[i*QK8_0 + l] = vi*d; + const int y_offset = qr == 1 ? 1 : qk/2; + + __shared__ float tmp[block_size]; // separate sum for each thread + tmp[tid] = 0; + + for (int i = 0; i < ncols/block_size; i += 2) { + const int col = i*block_size + 2*tid; + const int ib = (row*ncols + col)/qk; // block index + const int iqs = (col%qk)/qr; // quant index + const int iybs = col - col%qk; // y block start index + + // dequantize + float v0, v1; + dequantize_kernel(vx, ib, iqs, v0, v1); + + // matrix multiplication + tmp[tid] += v0 * y[iybs + iqs + 0]; + tmp[tid] += v1 * y[iybs + iqs + y_offset]; + } + + // sum up partial sums and write back result + __syncthreads(); + for (int s=block_size/2; s>0; s>>=1) { + if (tid < s) { + tmp[tid] += tmp[tid + s]; + } + __syncthreads(); + } + if (tid == 0) { + dst[row] = tmp[0]; } } @@ -235,11 +318,6 @@ static void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStre dequantize_block_q4_1<<>>(vx, y); } -static void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream) { - const int nb = k / QK4_2; - dequantize_block_q4_2<<>>(vx, y); -} - static void dequantize_row_q5_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) { const int nb = k / QK5_0; dequantize_block_q5_0<<>>(vx, y); @@ -255,6 +333,36 @@ static void dequantize_row_q8_0_cuda(const void * vx, float * y, int k, cudaStre dequantize_block_q8_0<<>>(vx, y); } +static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols); +} + +static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols); +} + +static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols); +} + +static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols); +} + +static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols); +} + // TODO: optimize static __global__ void convert_fp16_to_fp32(const void * vx, float * y) { const half * x = (const half *) vx; @@ -268,14 +376,18 @@ static void convert_fp16_to_fp32_cuda(const void * x, float * y, int k, cudaStre convert_fp16_to_fp32<<>>(x, y); } +static void convert_mul_mat_vec_f16_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { + GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0); + dequantize_mul_mat_vec + <<>>(vx, y, dst, ncols); +} + static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) { switch (type) { case GGML_TYPE_Q4_0: return dequantize_row_q4_0_cuda; case GGML_TYPE_Q4_1: return dequantize_row_q4_1_cuda; - case GGML_TYPE_Q4_2: - return dequantize_row_q4_2_cuda; case GGML_TYPE_Q5_0: return dequantize_row_q5_0_cuda; case GGML_TYPE_Q5_1: @@ -289,8 +401,27 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) { } } +static dequantize_mul_mat_vec_cuda_t ggml_get_dequantize_mul_mat_vec_cuda(ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + return dequantize_mul_mat_vec_q4_0_cuda; + case GGML_TYPE_Q4_1: + return dequantize_mul_mat_vec_q4_1_cuda; + case GGML_TYPE_Q5_0: + return dequantize_mul_mat_vec_q5_0_cuda; + case GGML_TYPE_Q5_1: + return dequantize_mul_mat_vec_q5_1_cuda; + case GGML_TYPE_Q8_0: + return dequantize_mul_mat_vec_q8_0_cuda; + case GGML_TYPE_F16: + return convert_mul_mat_vec_f16_cuda; + default: + return nullptr; + } +} + // buffer pool for cuda -#define MAX_CUDA_BUFFERS 16 +#define MAX_CUDA_BUFFERS 256 struct scoped_spin_lock { std::atomic_flag& lock; @@ -348,7 +479,7 @@ static void ggml_cuda_pool_free(void * ptr, size_t size) { CUDA_CHECK(cudaFree(ptr)); } -#define GGML_CUDA_MAX_STREAMS 8 +#define GGML_CUDA_MAX_STREAMS 8 // Set this to 1 for reproducible matrix multiplication. #define GGML_CUDA_MAX_EVENTS 64 static cublasHandle_t g_cublasH = nullptr; static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_STREAMS] = { nullptr }; @@ -587,6 +718,7 @@ static void ggml_cuda_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor const int nb2 = dst->nb[2]; const int nb3 = dst->nb[3]; const ggml_type type = src0->type; + const bool mul_mat_vec = ne11 == 1; const float alpha = 1.0f; const float beta = 0.0f; @@ -597,12 +729,16 @@ static void ggml_cuda_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor const size_t q_sz = ggml_type_size(type) * x_ne / ggml_blck_size(type); size_t x_size, y_size, d_size, q_size; - float * d_X = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * x_ne, &x_size); + float * d_X = nullptr; + if (!mul_mat_vec) { + d_X = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * x_ne, &x_size); + } float * d_Y = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * y_ne, &y_size); float * d_D = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * d_ne, &d_size); char * d_Q = (char *) ggml_cuda_pool_malloc(n_mm * q_sz, &q_size); const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(type); + dequantize_mul_mat_vec_cuda_t dmmv = ggml_get_dequantize_mul_mat_vec_cuda(type); GGML_ASSERT(to_fp32_cuda != nullptr); for (int64_t i03 = 0; i03 < ne03; i03++) { @@ -612,31 +748,54 @@ static void ggml_cuda_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor cudaStream_t cudaStream2 = g_cudaStreams2[i % GGML_CUDA_MAX_STREAMS]; cudaEvent_t cudaEvent = g_cudaEvents[i % GGML_CUDA_MAX_EVENTS]; - float * c_X = d_X + i * x_ne; float * c_Y = d_Y + i * y_ne; float * c_D = d_D + i * d_ne; char * c_Q = d_Q + i * q_sz; - // copy src0 and convert to fp32 on device - CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Q, src0, i03, i02, cudaStream2)); - to_fp32_cuda(c_Q, c_X, x_ne, cudaStream2); - CUDA_CHECK(cudaGetLastError()); - CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2)); + // copy src0 to device if necessary + if (src0->backend == GGML_BACKEND_CPU) { + CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Q, src0, i03, i02, cudaStream2)); + } else if (src0->backend == GGML_BACKEND_CUDA) { + c_Q = ((char *) src0->data) + i * q_sz; + } else { + GGML_ASSERT(false); + } + if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel + CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2)); - // copy src1 to device - CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream)); + // copy src1 to device + CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream)); - // wait for conversion - CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0)); + // wait for data + CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0)); - // compute - CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream)); - CUBLAS_CHECK( - cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N, - ne01, ne11, ne10, - &alpha, c_X, ne00, - c_Y, ne10, - &beta, c_D, ne01)); + // compute + dmmv(c_Q, c_Y, c_D, ne00, ne01, cudaStream); + CUDA_CHECK(cudaGetLastError()); + + } else { // general dequantization kernel + cuBLAS matrix matrix multiplication + float * c_X = d_X + i * x_ne; + + // convert src0 to fp32 on device + to_fp32_cuda(c_Q, c_X, x_ne, cudaStream2); + CUDA_CHECK(cudaGetLastError()); + CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2)); + + // copy src1 to device + CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream)); + + // wait for conversion + CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0)); + + // compute + CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream)); + CUBLAS_CHECK( + cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + &alpha, c_X, ne00, + c_Y, ne10, + &beta, c_D, ne01)); + } // copy dst to host float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); @@ -645,7 +804,9 @@ static void ggml_cuda_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor } CUDA_CHECK(cudaDeviceSynchronize()); - ggml_cuda_pool_free(d_X, x_size); + if (!mul_mat_vec) { + ggml_cuda_pool_free(d_X, x_size); + } ggml_cuda_pool_free(d_Y, y_size); ggml_cuda_pool_free(d_D, d_size); ggml_cuda_pool_free(d_Q, q_size); @@ -661,8 +822,7 @@ bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_te if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && - (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) { - + ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_CUDA)) { return true; } @@ -714,3 +874,25 @@ size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct return 0; } } + +void ggml_cuda_transform_tensor(ggml_tensor * tensor) { + const int64_t ne0 = tensor->ne[0]; + const int64_t ne1 = tensor->ne[1]; + const int64_t ne2 = tensor->ne[2]; + const int64_t ne3 = tensor->ne[3]; + + const ggml_type type = tensor->type; + const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type); + + size_t q_size; + char * d_Q = (char *) ggml_cuda_pool_malloc(q_sz, &q_size); + + cudaStream_t cudaStream2 = g_cudaStreams2[0]; + + // copy tensor to device + CUDA_CHECK(ggml_cuda_h2d_tensor_2d(d_Q, tensor, 0, 0, cudaStream2)); + CUDA_CHECK(cudaDeviceSynchronize()); + + tensor->data = d_Q; + tensor->backend = GGML_BACKEND_CUDA; +} diff --git a/ggml-cuda.h b/ggml-cuda.h index f7d6a8b..4e2c242 100644 --- a/ggml-cuda.h +++ b/ggml-cuda.h @@ -14,6 +14,8 @@ void ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens void * ggml_cuda_host_malloc(size_t size); void ggml_cuda_host_free(void * ptr); +void ggml_cuda_transform_tensor(struct ggml_tensor * tensor); + #ifdef __cplusplus } #endif diff --git a/ggml-opencl.c b/ggml-opencl.c index 4389eca..31ab13b 100644 --- a/ggml-opencl.c +++ b/ggml-opencl.c @@ -12,129 +12,129 @@ #define MULTILINE_QUOTE(...) #__VA_ARGS__ const char * clblast_dequant = MULTILINE_QUOTE( +typedef uchar uint8_t; +typedef int int32_t; +typedef uint uint32_t; + +constant uint QK4_0 = 32; struct block_q4_0 { float d; - uchar qs[16]; + uint8_t qs[QK4_0 / 2]; }; -__kernel void dequantize_row_q4_0(__global struct block_q4_0* blocks, __global float* result) { - const uint i = get_global_id(0) / 32; - const uint l = get_local_id(0); - - const float d = blocks[i].d; - - const uchar vi = blocks[i].qs[l]; - - const uint index = i*32 + l*2; - result[index + 0] = ((vi & 0xf) - 8)*d; - result[index + 1] = ((vi >> 4) - 8)*d; -} - +constant uint QK4_1 = 32; struct block_q4_1 { float d; float m; - uchar qs[16]; + uint8_t qs[QK4_1 / 2]; }; -__kernel void dequantize_row_q4_1(__global struct block_q4_1* blocks, __global float* result) { - const uint i = get_global_id(0) / 32; - const uint l = get_local_id(0); - - const float d = blocks[i].d; - const float m = blocks[i].m; - - const uchar vi = blocks[i].qs[l]; - - const uint index = i*32 + l*2; - result[index + 0] = (vi & 0xf) * d + m; - result[index + 1] = (vi >> 4) * d + m; -} - -struct block_q4_2 +constant uint QK5_0 = 32; +struct __attribute__ ((packed)) block_q5_0 { - ushort d; - uchar qs[8]; + half d; + uint32_t qh; + uint8_t qs[QK5_0 / 2]; }; -__kernel void dequantize_row_q4_2(__global struct block_q4_2* blocks, __global float* result) { - const uint i = get_global_id(0) / 16; - const uint l = get_local_id(0); - - const float d = vload_half(0, (__global half*) &blocks[i].d); - - const uchar vi = blocks[i].qs[l]; - - const uint index = i*16 + l*2; - result[index + 0] = ((vi & 0xf) - 8)*d; - result[index + 1] = ((vi >> 4) - 8)*d; -} - - -struct block_q5_0 -{ - float d; - uint qh; - uchar qs[16]; -}; - -__kernel void dequantize_row_q5_0(__global struct block_q5_0* blocks, __global float* result) { - const uint i = get_global_id(0) / 32; - const uint l = get_local_id(0); - - const float d = blocks[i].d; - - const uchar vi = blocks[i].qs[l]; - - const uint l2 = l * 2; - - const uchar vh0 = ((blocks[i].qh & (1 << (l2 + 0))) >> (l2 + 0)) << 4; - const uchar vh1 = ((blocks[i].qh & (1 << (l2 + 1))) >> (l2 + 1)) << 4; - - const uint index = i*32 + l2; - result[index + 0] = (((vi & 0xf) | vh0) - 16)*d; - result[index + 1] = (((vi >> 4) | vh1) - 16)*d; -} - +constant uint QK5_1 = 32; struct block_q5_1 { - ushort d; - ushort m; - uint qh; - uchar qs[16]; + half d; + half m; + uint32_t qh; + uint8_t qs[QK5_1 / 2]; }; -__kernel void dequantize_row_q5_1(__global struct block_q5_1* blocks, __global float* result) { - const uint i = get_global_id(0) / 32; - const uint l = get_local_id(0); - - const float d = vload_half(0, (__global half*) &blocks[i].d); - const float m = vload_half(0, (__global half*) &blocks[i].m); - - const uchar vi = blocks[i].qs[l]; - - const uint l2 = l * 2; - - const uchar vh0 = ((blocks[i].qh & (1 << (l2 + 0))) >> (l2 + 0)) << 4; - const uchar vh1 = ((blocks[i].qh & (1 << (l2 + 1))) >> (l2 + 1)) << 4; - - const uint index = i*32 + l2; - result[index + 0] = ((vi & 0xf) | vh0)*d + m; - result[index + 1] = ((vi >> 4) | vh1)*d + m; -} - +constant uint QK8_0 = 32; struct block_q8_0 { float d; - char qs[32]; + uint8_t qs[QK8_0]; }; -__kernel void dequantize_row_q8_0(__global struct block_q8_0* blocks, __global float* result) { - const uint i = get_global_id(0) / 32; - const uint l = get_local_id(0); - result[i*32 + l] = blocks[i].qs[l] * blocks[i].d; +__kernel void dequantize_row_q4_0(__global struct block_q4_0* x, __global float* y) { + constant uint qk = QK4_0; + + const uint i = get_global_id(0) / qk; + const uint j = get_local_id(0); + + const float d = x[i].d; + + const int x0 = (x[i].qs[j] & 0xf) - 8; + const int x1 = (x[i].qs[j] >> 4) - 8; + + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; +} + +__kernel void dequantize_row_q4_1(__global struct block_q4_1* x, __global float* y) { + constant uint qk = QK4_1; + + const uint i = get_global_id(0) / qk; + const uint j = get_local_id(0); + + const float d = x[i].d; + const float m = x[i].m; + + const int x0 = (x[i].qs[j] & 0xf); + const int x1 = (x[i].qs[j] >> 4); + + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; +} + +__kernel void dequantize_row_q5_0(__global struct block_q5_0* x, __global float* y) { + constant uint qk = QK5_0; + + const uint i = get_global_id(0) / qk; + const uint j = get_local_id(0); + + const float d = vload_half(0, (__global half*) &x[i].d); + + uint32_t qh = x[i].qh; + + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; + + const int32_t x0 = ((x[i].qs[j] & 0xf) | xh_0) - 16; + const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; + + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; +} + +__kernel void dequantize_row_q5_1(__global struct block_q5_1* x, __global float* y) { + constant uint qk = QK5_1; + + const uint i = get_global_id(0) / qk; + const uint j = get_local_id(0); + + const float d = vload_half(0, (__global half*) &x[i].d); + const float m = vload_half(0, (__global half*) &x[i].m); + + uint32_t qh = x[i].qh; + + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; + + const int x0 = (x[i].qs[j] & 0xf) | xh_0; + const int x1 = (x[i].qs[j] >> 4) | xh_1; + + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; +} + +__kernel void dequantize_row_q8_0(__global struct block_q8_0* x, __global float* y) { + constant uint qk = QK8_0; + const uint i = get_global_id(0) / qk; + const uint j = get_local_id(0); + + const float d = x[i].d; + y[i*qk + j] = x[i].qs[j]*d; } ); @@ -148,26 +148,12 @@ __kernel void dequantize_row_q8_0(__global struct block_q8_0* blocks, __global f } \ } while (0) -#define QK5_0 32 -typedef struct { - ggml_fp16_t d; // delta - uint8_t qh[4]; // 5-th bit of quants - uint8_t qs[QK5_0 / 2]; // nibbles / quants -} block_q5_0; - - -typedef struct { - float d; // delta - uint32_t qh; // 5-th bit of quants - uint8_t qs[QK5_0 / 2]; // nibbles / quants -} cl_block_q5_0; - static cl_platform_id platform; static cl_device_id device; static cl_context context; static cl_command_queue queue; static cl_program program; -static cl_kernel kernel_q4_0, kernel_q4_1, kernel_q4_2, kernel_q5_0, kernel_q5_1, kernel_q8_0; +static cl_kernel kernel_q4_0, kernel_q4_1, kernel_q5_0, kernel_q5_1, kernel_q8_0; static cl_mem cl_buffer_a, cl_buffer_qb, cl_buffer_b, cl_buffer_c; static size_t cl_size_a = 0, cl_size_qb = 0, cl_size_b = 0, cl_size_c = 0; @@ -238,8 +224,6 @@ void ggml_cl_init(void) { CL_CHECK(err, "clCreateKernel"); kernel_q4_1 = clCreateKernel(program, "dequantize_row_q4_1", &err); CL_CHECK(err, "clCreateKernel"); - kernel_q4_2 = clCreateKernel(program, "dequantize_row_q4_2", &err); - CL_CHECK(err, "clCreateKernel"); kernel_q5_0 = clCreateKernel(program, "dequantize_row_q5_0", &err); CL_CHECK(err, "clCreateKernel"); kernel_q5_1 = clCreateKernel(program, "dequantize_row_q5_1", &err); @@ -274,7 +258,6 @@ void ggml_cl_sgemm_wrapper( cl_kernel kernel; size_t global = n * k, local, size_qb; bool dequant; - cl_block_q5_0* cl_host_b; switch (btype) { case GGML_TYPE_F32: @@ -292,28 +275,11 @@ void ggml_cl_sgemm_wrapper( local = 16; size_qb = global * (sizeof(float) * 2 + local) / 32; break; - case GGML_TYPE_Q4_2: - dequant = true; - kernel = kernel_q4_2; - local = 8; - size_qb = global * (sizeof(ggml_fp16_t) + local) / 16; - break; case GGML_TYPE_Q5_0: dequant = true; kernel = kernel_q5_0; local = 16; - // For some reason OpenCL seems to be incapable of working with structs of size 22. - // 20 and 24 bytes are fine. Workaround to do the fp16 to fp32 step on CPU... - // TODO Find the reason, fix and remove workaround. - const block_q5_0* b = (const block_q5_0*) host_b; - cl_host_b = (cl_block_q5_0*) malloc(sizeof(cl_block_q5_0) * global / 32); - for (size_t i = 0; i < global / 32; i++) { - cl_host_b[i].d = ggml_fp16_to_fp32(b[i].d); - memcpy(&cl_host_b[i].qh, b[i].qh, sizeof(uint32_t)); - memcpy(&cl_host_b[i].qs, b[i].qs, QK5_0 / 2); - } - host_b = (const float*) cl_host_b; - size_qb = global * (sizeof(float) + sizeof(uint32_t) + local) / 32; + size_qb = global * (sizeof(ggml_fp16_t) + sizeof(uint32_t) + local) / 32; break; case GGML_TYPE_Q5_1: dequant = true; @@ -392,7 +358,4 @@ void ggml_cl_sgemm_wrapper( clWaitForEvents(1, &ev_c); clReleaseEvent(ev_sgemm); clReleaseEvent(ev_c); - if (btype == GGML_TYPE_Q5_0) { - free((void*) cl_host_b); - } } diff --git a/ggml.c b/ggml.c index addf0c3..58727bb 100644 --- a/ggml.c +++ b/ggml.c @@ -137,6 +137,9 @@ inline static void* ggml_aligned_malloc(size_t size) { #if defined(GGML_USE_ACCELERATE) #include +#if defined(GGML_USE_CLBLAST) // allow usage of CLBlast alongside Accelerate functions +#include "ggml-opencl.h" +#endif #elif defined(GGML_USE_OPENBLAS) #include #elif defined(GGML_USE_CUBLAS) @@ -180,9 +183,13 @@ typedef double ggml_float; #undef bool #define bool _Bool #else +#if defined(_MSC_VER) || defined(__MINGW32__) +#include +#else #include #endif #endif +#endif #ifdef __F16C__ @@ -332,8 +339,9 @@ static float table_f32_f16[1 << 16]; #define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s) #define B8(c,s ) B7(c,s, c), B7(c,s, s) -// precomputed tables for expanding 8bits to 8 bytes (shl 4) -static const uint64_t table_b2b_u[1 << 8] = { B8(00, 10) }; +// precomputed tables for expanding 8bits to 8 bytes: +static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4 +static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4 #endif // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32, @@ -464,26 +472,20 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float); // quantization // -#if __AVX__ || __AVX2__ || __AVX512F__ -// Unpack 16 4-bit fields into 16 bytes -// The output vector contains 16 bytes, each one in [ 0 .. 15 ] interval -static inline __m128i bytes_from_nibbles_16(const uint8_t * rsi) -{ - // Load 8 bytes from memory - __m128i tmp = _mm_loadl_epi64( ( const __m128i* )rsi ); - - // Expand bytes into uint16_t values - __m128i bytes = _mm_cvtepu8_epi16( tmp ); - - // Unpack values into individual bytes - const __m128i lowMask = _mm_set1_epi8( 0xF ); - __m128i high = _mm_andnot_si128( lowMask, bytes ); - __m128i low = _mm_and_si128( lowMask, bytes ); - high = _mm_slli_epi16( high, 4 ); - bytes = _mm_or_si128( low, high ); - return bytes; +#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) +// multiply int8_t, add results pairwise twice +static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) { + // Get absolute values of x vectors + const __m128i ax = _mm_sign_epi8(x, x); + // Sign the values of the y vectors + const __m128i sy = _mm_sign_epi8(y, x); + // Perform multiplication and create 16-bit values + const __m128i dot = _mm_maddubs_epi16(ax, sy); + const __m128i ones = _mm_set1_epi16(1); + return _mm_madd_epi16(ones, dot); } +#if __AVX__ || __AVX2__ || __AVX512F__ // horizontally add 8 floats static inline float hsum_float_8(const __m256 x) { __m128 res = _mm256_extractf128_ps(x, 1); @@ -516,8 +518,8 @@ static inline __m256i bytes_from_bits_32(const uint8_t * x) { uint32_t x32; memcpy(&x32, x, sizeof(uint32_t)); const __m256i shuf_mask = _mm256_set_epi64x( - 0x0303030303030303, 0x0202020202020202, - 0x0101010101010101, 0x0000000000000000); + 0x0303030303030303, 0x0202020202020202, + 0x0101010101010101, 0x0000000000000000); __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask); const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe); bytes = _mm256_or_si256(bytes, bit_mask); @@ -528,19 +530,10 @@ static inline __m256i bytes_from_bits_32(const uint8_t * x) { // The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) { - // Load 16 bytes from memory - __m128i tmp = _mm_loadu_si128( ( const __m128i* )rsi ); - - // Expand bytes into uint16_t values - __m256i bytes = _mm256_cvtepu8_epi16( tmp ); - - // Unpack values into individual bytes + const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi); + const __m256i bytes = _mm256_set_m128i(_mm_srli_epi16(tmp, 4), tmp); const __m256i lowMask = _mm256_set1_epi8( 0xF ); - __m256i high = _mm256_andnot_si256( lowMask, bytes ); - __m256i low = _mm256_and_si256( lowMask, bytes ); - high = _mm256_slli_epi16( high, 4 ); - bytes = _mm256_or_si256( low, high ); - return bytes; + return _mm256_and_si256(lowMask, bytes); } // add int16_t pairwise and return as float vector @@ -604,7 +597,19 @@ static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 ) return _mm_packus_epi16( bytes1, bytes2); } #endif +#elif defined(__SSSE3__) +// horizontally add 4x4 floats +static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) { + __m128 res_0 =_mm_hadd_ps(a, b); + __m128 res_1 =_mm_hadd_ps(c, d); + __m128 res =_mm_hadd_ps(res_0, res_1); + res =_mm_hadd_ps(res, res); + res =_mm_hadd_ps(res, res); + + return _mm_cvtss_f32(res); +} #endif // __AVX__ || __AVX2__ || __AVX512F__ +#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) #if __ARM_NEON @@ -670,94 +675,6 @@ float vmaxvq_f32(float32x4_t v) { MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3))); } -int8x8_t vzip1_s8(int8x8_t a, int8x8_t b) { - int8x8_t res; - - res[0] = a[0]; res[1] = b[0]; - res[2] = a[1]; res[3] = b[1]; - res[4] = a[2]; res[5] = b[2]; - res[6] = a[3]; res[7] = b[3]; - - return res; -} - -int8x8_t vzip2_s8(int8x8_t a, int8x8_t b) { - int8x8_t res; - - res[0] = a[4]; res[1] = b[4]; - res[2] = a[5]; res[3] = b[5]; - res[4] = a[6]; res[5] = b[6]; - res[6] = a[7]; res[7] = b[7]; - - return res; -} - -uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) { - uint8x8_t res; - - res[0] = a[0]; res[1] = b[0]; - res[2] = a[1]; res[3] = b[1]; - res[4] = a[2]; res[5] = b[2]; - res[6] = a[3]; res[7] = b[3]; - - return res; -} - -uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) { - uint8x8_t res; - - res[0] = a[4]; res[1] = b[4]; - res[2] = a[5]; res[3] = b[5]; - res[4] = a[6]; res[5] = b[6]; - res[6] = a[7]; res[7] = b[7]; - - return res; -} - -int8x16_t vzip1q_s8(int8x16_t a, int8x16_t b) { - int8x16_t res; - - res[0] = a[0]; res[1] = b[0]; res[2] = a[1]; res[3] = b[1]; - res[4] = a[2]; res[5] = b[2]; res[6] = a[3]; res[7] = b[3]; - res[8] = a[4]; res[9] = b[4]; res[10] = a[5]; res[11] = b[5]; - res[12] = a[6]; res[13] = b[6]; res[14] = a[7]; res[15] = b[7]; - - return res; -} - -int8x16_t vzip2q_s8(int8x16_t a, int8x16_t b) { - int8x16_t res; - - res[0] = a[8]; res[1] = b[8]; res[2] = a[9]; res[3] = b[9]; - res[4] = a[10]; res[5] = b[10]; res[6] = a[11]; res[7] = b[11]; - res[8] = a[12]; res[9] = b[12]; res[10] = a[13]; res[11] = b[13]; - res[12] = a[14]; res[13] = b[14]; res[14] = a[15]; res[15] = b[15]; - - return res; -} - -uint8x16_t vzip1q_u8(uint8x16_t a, uint8x16_t b) { - uint8x16_t res; - - res[0] = a[0]; res[1] = b[0]; res[2] = a[1]; res[3] = b[1]; - res[4] = a[2]; res[5] = b[2]; res[6] = a[3]; res[7] = b[3]; - res[8] = a[4]; res[9] = b[4]; res[10] = a[5]; res[11] = b[5]; - res[12] = a[6]; res[13] = b[6]; res[14] = a[7]; res[15] = b[7]; - - return res; -} - -uint8x16_t vzip2q_u8(uint8x16_t a, uint8x16_t b) { - uint8x16_t res; - - res[0] = a[8]; res[1] = b[8]; res[2] = a[9]; res[3] = b[9]; - res[4] = a[10]; res[5] = b[10]; res[6] = a[11]; res[7] = b[11]; - res[8] = a[12]; res[9] = b[12]; res[10] = a[13]; res[11] = b[13]; - res[12] = a[14]; res[13] = b[14]; res[14] = a[15]; res[15] = b[15]; - - return res; -} - int32x4_t vcvtnq_s32_f32(float32x4_t v) { int32x4_t res; @@ -788,13 +705,6 @@ typedef struct { } block_q4_1; static_assert(sizeof(block_q4_1) == 2 * sizeof(float) + QK4_1 / 2, "wrong q4_1 block size/padding"); -#define QK4_2 16 -typedef struct { - ggml_fp16_t d; // delta - uint8_t qs[QK4_2 / 2]; // nibbles / quants -} block_q4_2; -static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding"); - #define QK5_0 32 typedef struct { ggml_fp16_t d; // delta @@ -821,150 +731,255 @@ static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block siz #define QK8_1 32 typedef struct { - float d; // delta - float s0; // d * sum(qs[i]) low - float s1; // d * sum(qs[i]) high - int8_t qs[QK8_1]; // quants + float d; // delta + float s; // d * sum(qs[i]) + int8_t qs[QK8_1]; // quants } block_q8_1; -static_assert(sizeof(block_q8_1) == 3*sizeof(float) + QK8_1, "wrong q8_1 block size/padding"); +static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding"); // reference implementation for deterministic creation of model files static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) { - assert(k % QK4_0 == 0); - const int nb = k / QK4_0; + static const int qk = QK4_0; - uint8_t pp[QK4_0/2]; + assert(k % qk == 0); + + const int nb = k / qk; for (int i = 0; i < nb; i++) { float amax = 0.0f; // absolute max - float max = 0.0f; + float max = 0.0f; - for (int l = 0; l < QK4_0; l++) { - const float v = x[i*QK4_0 + l]; + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; if (amax < fabsf(v)) { amax = fabsf(v); - max = v; + max = v; } } - const float d = max / -8; + const float d = max / -8; const float id = d ? 1.0f/d : 0.0f; y[i].d = d; - for (int l = 0; l < QK4_0; l += 2) { - const float v0 = x[i*QK4_0 + l + 0]*id; - const float v1 = x[i*QK4_0 + l + 1]*id; + for (int j = 0; j < qk/2; ++j) { + const float x0 = x[i*qk + 0 + j]*id; + const float x1 = x[i*qk + qk/2 + j]*id; - const uint8_t vi0 = MIN(15, (int8_t)roundf(v0) + 8); - const uint8_t vi1 = MIN(15, (int8_t)roundf(v1) + 8); + const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f)); + const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f)); - assert(vi0 < 16); - assert(vi1 < 16); - - pp[l/2] = vi0 | (vi1 << 4); + y[i].qs[j] = xi0; + y[i].qs[j] |= xi1 << 4; } - - memcpy(y[i].qs, pp, sizeof(pp)); } } -static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_0 == 0); - const int nb = k / QK4_0; +static void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) { + quantize_row_q4_0_reference(x, y, k); +} - block_q4_0 * restrict y = vy; +static void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k) { + const int qk = QK4_1; + + assert(k % qk == 0); + + const int nb = k / qk; -#if defined(__POWER9_VECTOR__) - const vector float v85 = vec_splats(8.5f); - const vector signed int v15 = vec_splats(15); for (int i = 0; i < nb; i++) { - float max = 0.0f; - float min = 0.0f; + float min = FLT_MAX; + float max = -FLT_MAX; - vector float asrcv [8]; - vector float srcv [8]; - vector float maxv[8]; - vector float minv[8]; + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; - for (int l = 0; l < 8; l++) srcv[l] = *(vector float *)(x + i*32 + 4*l); - //for (int l = 0; l < 8; l++) asrcv[l] = vec_abs(srcv[l]); + if (v < min) min = v; + if (v > max) max = v; + } - for (int l = 0; l < 4; l++) maxv[2*l] = vec_max(asrcv[2*l], asrcv[2*l+1]); - //for (int l = 0; l < 2; l++) maxv[4*l] = vec_max(maxv[4*l], maxv[4*l+2]); - maxv[0] = vec_max(maxv[0], maxv[2]); - maxv[4] = vec_max(maxv[4], maxv[6]); - //for (int l = 0; l < 1; l++) maxv[8*l] = vec_max(maxv[8*l], maxv[8*l+4]); - maxv[0] = vec_max(maxv[0], maxv[4]); - - for (int l = 0; l < 4; l++) minv[2*l] = vec_min(asrcv[2*l], asrcv[2*l+1]); - //for (int l = 0; l < 2; l++) minv[4*l] = vec_min(minv[4*l], minv[4*l+2]); - minv[0] = vec_min(minv[0], minv[2]); - minv[4] = vec_min(minv[4], minv[6]); - //for (int l = 0; l < 1; l++) minv[8*l] = vec_min(minv[8*l], minv[8*l+4]); - minv[0] = vec_min(minv[0], minv[4]); - - - max = MAX( - MAX(vec_extract(maxv[0], 0), vec_extract(maxv[0], 1)), - MAX(vec_extract(maxv[0], 2), vec_extract(maxv[0], 3))); - min = MIN( - MIN(vec_extract(minv[0], 0), vec_extract(minv[0], 1)), - MIN(vec_extract(minv[0], 2), vec_extract(minv[0], 3))); - - const float magnitude = max >= fabsf(min) ? max : min; - const float d = magnitude / -8; - const float id = d ? 1.0/d : 0.0; + const float d = (max - min) / ((1 << 4) - 1); + const float id = d ? 1.0f/d : 0.0f; y[i].d = d; + y[i].m = min; - const vector float vid = vec_splats(id); - uint8_t * restrict pb = y[i].qs; - for (int l = 0; l < 8; l++) { - const vector float vf = vec_madd(srcv[l], vid, v85); - const vector signed int vi = vec_signed(vf); - const vector signed int vc = vec_min(vi, v15); + for (int j = 0; j < qk/2; ++j) { + const float x0 = (x[i*qk + 0 + j] - min)*id; + const float x1 = (x[i*qk + qk/2 + j] - min)*id; - pb[2*l + 0] = vec_extract(vc, 0) | (vec_extract(vc, 1) << 4); - pb[2*l + 1] = vec_extract(vc, 2) | (vec_extract(vc, 3) << 4); + const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f)); + const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f)); + + y[i].qs[j] = xi0; + y[i].qs[j] |= xi1 << 4; } } -#elif __ARM_NEON +} + +static void quantize_row_q4_1(const float * restrict x, void * restrict y, int k) { + quantize_row_q4_1_reference(x, y, k); +} + +static void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k) { + static const int qk = QK5_0; + + assert(k % qk == 0); + + const int nb = k / qk; + for (int i = 0; i < nb; i++) { - float32x4_t srcv [8]; - float32x4_t maxv[8]; - float32x4_t minv[8]; + float amax = 0.0f; // absolute max + float max = 0.0f; - for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l); + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; + if (amax < fabsf(v)) { + amax = fabsf(v); + max = v; + } + } - for (int l = 0; l < 4; l++) maxv[2*l] = vmaxq_f32(srcv[2*l], srcv[2*l+1]); - for (int l = 0; l < 2; l++) maxv[4*l] = vmaxq_f32(maxv[4*l], maxv[4*l+2]); - for (int l = 0; l < 1; l++) maxv[8*l] = vmaxq_f32(maxv[8*l], maxv[8*l+4]); + const float d = max / -16; + const float id = d ? 1.0f/d : 0.0f; - for (int l = 0; l < 4; l++) minv[2*l] = vminq_f32(srcv[2*l], srcv[2*l+1]); - for (int l = 0; l < 2; l++) minv[4*l] = vminq_f32(minv[4*l], minv[4*l+2]); - for (int l = 0; l < 1; l++) minv[8*l] = vminq_f32(minv[8*l], minv[8*l+4]); + y[i].d = GGML_FP32_TO_FP16(d); - const float max = vmaxvq_f32(maxv[0]); - const float min = vminvq_f32(minv[0]); + uint32_t qh = 0; - const float magnitude = max >= fabsf(min) ? max : min; - const float d = magnitude / -8; + for (int j = 0; j < qk/2; ++j) { + const float x0 = x[i*qk + 0 + j]*id; + const float x1 = x[i*qk + qk/2 + j]*id; + + const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f)); + const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f)); + + y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4); + + // get the 5-th bit and store it in qh at the right position + qh |= ((xi0 & 0x10) >> 4) << (j + 0); + qh |= ((xi1 & 0x10) >> 4) << (j + qk/2); + } + + memcpy(&y[i].qh, &qh, sizeof(qh)); + } +} + +static void quantize_row_q5_0(const float * restrict x, void * restrict y, int k) { + quantize_row_q5_0_reference(x, y, k); +} + +static void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k) { + const int qk = QK5_1; + + assert(k % qk == 0); + + const int nb = k / qk; + + for (int i = 0; i < nb; i++) { + float min = FLT_MAX; + float max = -FLT_MAX; + + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; + + if (v < min) min = v; + if (v > max) max = v; + } + + const float d = (max - min) / ((1 << 5) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + y[i].m = GGML_FP32_TO_FP16(min); + + uint32_t qh = 0; + + for (int j = 0; j < qk/2; ++j) { + const float x0 = (x[i*qk + 0 + j] - min)*id; + const float x1 = (x[i*qk + qk/2 + j] - min)*id; + + const uint8_t xi0 = (uint8_t)(x0 + 0.5f); + const uint8_t xi1 = (uint8_t)(x1 + 0.5f); + + y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4); + + // get the 5-th bit and store it in qh at the right position + qh |= ((xi0 & 0x10) >> 4) << (j + 0); + qh |= ((xi1 & 0x10) >> 4) << (j + qk/2); + } + + memcpy(&y[i].qh, &qh, sizeof(y[i].qh)); + } +} + +static void quantize_row_q5_1(const float * restrict x, void * restrict y, int k) { + quantize_row_q5_1_reference(x, y, k); +} + +// reference implementation for deterministic creation of model files +static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k) { + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_0; j++) { + const float v = x[i*QK8_0 + j]; + amax = MAX(amax, fabsf(v)); + } + + const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f/d : 0.0f; y[i].d = d; - for (int l = 0; l < 8; l++) { - const float32x4_t v = vmulq_n_f32(srcv[l], id); - const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(8.5f)); - const int32x4_t vi = vcvtq_s32_f32(vf); - const int32x4_t vc = vminq_s32(vi, vdupq_n_s32(15)); + for (int j = 0; j < QK8_0; ++j) { + const float x0 = x[i*QK8_0 + j]*id; - y[i].qs[2*l + 0] = vgetq_lane_s32(vc, 0) | (vgetq_lane_s32(vc, 1) << 4); - y[i].qs[2*l + 1] = vgetq_lane_s32(vc, 2) | (vgetq_lane_s32(vc, 3) << 4); + y[i].qs[j] = roundf(x0); } } -#elif defined(__AVX2__) +} + +static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) { + assert(QK8_0 == 32); + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + block_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + for (int i = 0; i < nb; i++) { + float32x4_t srcv [8]; + float32x4_t asrcv[8]; + float32x4_t amaxv[8]; + + for (int j = 0; j < 8; j++) srcv[j] = vld1q_f32(x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]); + + for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]); + + const float amax = vmaxvq_f32(amaxv[0]); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + for (int j = 0; j < 8; j++) { + const float32x4_t v = vmulq_n_f32(srcv[j], id); + const int32x4_t vi = vcvtnq_s32_f32(v); + + y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3); + } + } +#elif defined(__AVX2__) || defined(__AVX__) for (int i = 0; i < nb; i++) { // Load elements into 4 AVX vectors __m256 v0 = _mm256_loadu_ps( x ); @@ -973,31 +988,22 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int __m256 v3 = _mm256_loadu_ps( x + 24 ); x += 32; - // Compute max for the block - __m256 max = _mm256_max_ps( v0, v1 ); - __m256 maxTmp = _mm256_max_ps( v2, v3 ); - max = _mm256_max_ps( max, maxTmp ); + // Compute max(abs(e)) for the block + const __m256 signBit = _mm256_set1_ps( -0.0f ); + __m256 maxAbs = _mm256_andnot_ps( signBit, v0 ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) ); - __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( max, 1 ), _mm256_castps256_ps128( max ) ); + __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) ); max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); const float maxScalar = _mm_cvtss_f32( max4 ); - // Compute min for the block - __m256 min = _mm256_min_ps( v0, v1 ); - __m256 minTmp = _mm256_min_ps( v2, v3 ); - min = _mm256_min_ps( min, minTmp ); - - __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( min, 1 ), _mm256_castps256_ps128( min ) ); - min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) ); - min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) ); - const float minScalar = _mm_cvtss_f32( min4 ); - // Quantize these floats - const float magnitude = maxScalar >= fabsf(minScalar) ? maxScalar : minScalar; - const float d = magnitude / -8.0f; + const float d = maxScalar / 127.f; y[i].d = d; - const float id = ( magnitude != 0.0f ) ? -8.0f / magnitude : 0.0f; + const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; const __m256 mul = _mm256_set1_ps( id ); // Apply the multiplier @@ -1018,6 +1024,7 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int __m256i i2 = _mm256_cvtps_epi32( v2 ); __m256i i3 = _mm256_cvtps_epi32( v3 ); +#if defined(__AVX2__) // Convert int32 to int16 i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 @@ -1030,70 +1037,8 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); i0 = _mm256_permutevar8x32_epi32( i0, perm ); - // Apply offset and clamp to translate the range from [ -8 .. +8 ] into [ +0 .. +15 ] - const __m256i off = _mm256_set1_epi8( 8 ); - i0 = _mm256_add_epi8( i0, off ); - const __m256i maxNibble = _mm256_set1_epi8( 15 ); - i0 = _mm256_min_epi8( i0, maxNibble ); - - // Compress the vector into 4 bit/value, and store - __m128i res = packNibbles( i0 ); - _mm_storeu_si128( ( __m128i* )y[i].qs, res ); - } -#elif defined(__AVX__) - for (int i = 0; i < nb; i++) { - // Load elements into 4 AVX vectors - __m256 v0 = _mm256_loadu_ps( x ); - __m256 v1 = _mm256_loadu_ps( x + 8 ); - __m256 v2 = _mm256_loadu_ps( x + 16 ); - __m256 v3 = _mm256_loadu_ps( x + 24 ); - x += 32; - - // Compute max for the block - __m256 max = _mm256_max_ps( v0, v1 ); - __m256 maxTmp = _mm256_max_ps( v2, v3 ); - max = _mm256_max_ps( max, maxTmp ); - - __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( max, 1 ), _mm256_castps256_ps128( max ) ); - max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); - max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); - const float maxScalar = _mm_cvtss_f32( max4 ); - - // Compute min for the block - __m256 min = _mm256_min_ps( v0, v1 ); - __m256 minTmp = _mm256_min_ps( v2, v3 ); - min = _mm256_min_ps( min, minTmp ); - - __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( min, 1 ), _mm256_castps256_ps128( min ) ); - min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) ); - min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) ); - const float minScalar = _mm_cvtss_f32( min4 ); - - // Quantize these floats - const float magnitude = maxScalar >= fabsf(minScalar) ? maxScalar : minScalar; - const float d = magnitude / -8.0f; - y[i].d = d; - const float id = ( magnitude != 0.0f ) ? -8.0f / magnitude : 0.0f; - const __m256 mul = _mm256_set1_ps( id ); - - // Apply the multiplier - v0 = _mm256_mul_ps( v0, mul ); - v1 = _mm256_mul_ps( v1, mul ); - v2 = _mm256_mul_ps( v2, mul ); - v3 = _mm256_mul_ps( v3, mul ); - - // Round to nearest integer - v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); - v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); - v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); - v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); - - // Convert floats to integers - __m256i i0 = _mm256_cvtps_epi32( v0 ); - __m256i i1 = _mm256_cvtps_epi32( v1 ); - __m256i i2 = _mm256_cvtps_epi32( v2 ); - __m256i i3 = _mm256_cvtps_epi32( v3 ); - + _mm256_storeu_si256((__m256i *)y[i].qs, i0); +#else // Since we don't have in AVX some necessary functions, // we split the registers in half and call AVX2 analogs from SSE __m128i ni0 = _mm256_castsi256_si128( i0 ); @@ -1114,418 +1059,27 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int ni0 = _mm_packs_epi16( ni0, ni2 ); ni4 = _mm_packs_epi16( ni4, ni6 ); - // Apply offset and clamp to translate the range from [ -8 .. +8 ] into [ +0 .. +15 ] - const __m128i off = _mm_set1_epi8( 8 ); - ni0 = _mm_add_epi8( ni0, off ); - ni4 = _mm_add_epi8( ni4, off ); - const __m128i maxNibble = _mm_set1_epi8( 15 ); - ni0 = _mm_min_epi8( ni0, maxNibble ); - ni4 = _mm_min_epi8( ni4, maxNibble ); - - // Compress the vector into 4 bit/value, and store - __m128i res = packNibbles( ni0, ni4 ); - _mm_storeu_si128( ( __m128i* )y[i].qs, res ); - } -#elif defined(__wasm_simd128__) - for (int i = 0; i < nb; i++) { - float max = 0.0f; - float min = 0.0f; - - v128_t srcv [8]; - v128_t maxv[8]; - v128_t minv[8]; - - for (int l = 0; l < 8; l++) srcv[l] = wasm_v128_load(x + i*32 + 4*l); - - for (int l = 0; l < 4; l++) maxv[2*l] = wasm_f32x4_max(srcv[2*l], srcv[2*l+1]); - for (int l = 0; l < 2; l++) maxv[4*l] = wasm_f32x4_max(maxv[4*l], maxv[4*l+2]); - for (int l = 0; l < 1; l++) maxv[8*l] = wasm_f32x4_max(maxv[8*l], maxv[8*l+4]); - - for (int l = 0; l < 4; l++) minv[2*l] = wasm_f32x4_min(srcv[2*l], srcv[2*l+1]); - for (int l = 0; l < 2; l++) minv[4*l] = wasm_f32x4_min(minv[4*l], minv[4*l+2]); - for (int l = 0; l < 1; l++) minv[8*l] = wasm_f32x4_min(minv[8*l], minv[8*l+4]); - - max = MAX( - MAX(wasm_f32x4_extract_lane(maxv[0], 0), wasm_f32x4_extract_lane(maxv[0], 1)), - MAX(wasm_f32x4_extract_lane(maxv[0], 2), wasm_f32x4_extract_lane(maxv[0], 3))); - min = MIN( - MIN(wasm_f32x4_extract_lane(minv[0], 0), wasm_f32x4_extract_lane(minv[0], 1)), - MIN(wasm_f32x4_extract_lane(minv[0], 2), wasm_f32x4_extract_lane(minv[0], 3))); - - const float magnitude = max >= fabsf(min) ? max : min; - const float d = magnitude / -8; - const float id = d ? 1.0/d : 0.0; - - y[i].d = d; - - for (int l = 0; l < 8; l++) { - const v128_t v = wasm_f32x4_mul(srcv[l], wasm_f32x4_splat(id)); - const v128_t vf = wasm_f32x4_add(v, wasm_f32x4_splat(8.5f)); - const v128_t vi = wasm_i32x4_trunc_sat_f32x4(vf); - const v128_t vc = wasm_i32x4_min(vi, wasm_i32x4_splat(15)); - - y[i].qs[2*l + 0] = wasm_i32x4_extract_lane(vc, 0) | (wasm_i32x4_extract_lane(vc, 1) << 4); - y[i].qs[2*l + 1] = wasm_i32x4_extract_lane(vc, 2) | (wasm_i32x4_extract_lane(vc, 3) << 4); - } + _mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0); + _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4); +#endif } #else // scalar - quantize_row_q4_0_reference(x, y, k); -#endif -} - -static void quantize_row_q4_1_reference(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_1 == 0); - const int nb = k / QK4_1; - - block_q4_1 * restrict y = vy; - - uint8_t pp[QK4_1/2]; - - for (int i = 0; i < nb; i++) { - float min = FLT_MAX; - float max = -FLT_MAX; - - for (int l = 0; l < QK4_1; l++) { - const float v = x[i*QK4_1 + l]; - if (v < min) min = v; - if (v > max) max = v; - } - - const float d = (max - min) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = d; - y[i].m = min; - - for (int l = 0; l < QK4_1; l += 2) { - const float v0 = (x[i*QK4_1 + l + 0] - min)*id; - const float v1 = (x[i*QK4_1 + l + 1] - min)*id; - - const uint8_t vi0 = roundf(v0); - const uint8_t vi1 = roundf(v1); - - assert(vi0 < 16); - assert(vi1 < 16); - - pp[l/2] = vi0 | (vi1 << 4); - } - - memcpy(y[i].qs, pp, sizeof(pp)); - } -} - -static void quantize_row_q4_1(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_1 == 0); - - const int nb = k / QK4_1; - - block_q4_1 * restrict y = vy; - -#if defined(__AVX2__) - for (int i = 0; i < nb; i++) { - // Load elements into 4 AVX vectors - __m256 v0 = _mm256_loadu_ps( x ); - __m256 v1 = _mm256_loadu_ps( x + 8 ); - __m256 v2 = _mm256_loadu_ps( x + 16 ); - __m256 v3 = _mm256_loadu_ps( x + 24 ); - x += 32; - - // Compute max for the block - __m256 vmax; - vmax = _mm256_max_ps( v0, v1 ); - vmax = _mm256_max_ps( vmax, v2 ); - vmax = _mm256_max_ps( vmax, v3 ); - - __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( vmax, 1 ), _mm256_castps256_ps128( vmax ) ); - max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); - max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); - const float maxScalar = _mm_cvtss_f32( max4 ); - - // Compute min for the block - __m256 vmin; - vmin = _mm256_min_ps( v0, v1 ); - vmin = _mm256_min_ps( vmin, v2 ); - vmin = _mm256_min_ps( vmin, v3 ); - - __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( vmin, 1 ), _mm256_castps256_ps128( vmin ) ); - min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) ); - min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) ); - const float minScalar = _mm_cvtss_f32( min4 ); - - // Quantize these floats - const float d = (maxScalar - minScalar) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].m = minScalar; - y[i].d = d; - - // x = (x-min)*id - const __m256 mul = _mm256_set1_ps( id ); - const __m256 off = _mm256_set1_ps( minScalar ); - v0 = _mm256_mul_ps( _mm256_sub_ps( v0, off ), mul ); - v1 = _mm256_mul_ps( _mm256_sub_ps( v1, off ), mul ); - v2 = _mm256_mul_ps( _mm256_sub_ps( v2, off ), mul ); - v3 = _mm256_mul_ps( _mm256_sub_ps( v3, off ), mul ); - - // Round to nearest integer - v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); - v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); - v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); - v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); - - // Convert floats to integers - __m256i i0 = _mm256_cvtps_epi32( v0 ); - __m256i i1 = _mm256_cvtps_epi32( v1 ); - __m256i i2 = _mm256_cvtps_epi32( v2 ); - __m256i i3 = _mm256_cvtps_epi32( v3 ); - - // Convert int32 to int16 - i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 - i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 - // Convert int16 to int8 - i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 - - // We got our precious signed bytes, but the order is now wrong - // These AVX2 pack instructions process 16-byte pieces independently - // The following instruction is fixing the order - const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); - i0 = _mm256_permutevar8x32_epi32( i0, perm ); - - // Compress the vector into 4 bit/value, and store - __m128i res = packNibbles( i0 ); - _mm_storeu_si128( ( __m128i* )y[i].qs, res ); - } -#elif __ARM_NEON - for (int i = 0; i < nb; i++) { - float32x4_t srcv[8]; - float32x4_t minv[8]; - float32x4_t maxv[8]; - - for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*QK4_1 + 4*l); - - for (int l = 0; l < 4; l++) minv[2*l] = vminq_f32(srcv[2*l], srcv[2*l + 1]); - for (int l = 0; l < 2; l++) minv[4*l] = vminq_f32(minv[4*l], minv[4*l + 2]); - for (int l = 0; l < 1; l++) minv[8*l] = vminq_f32(minv[8*l], minv[8*l + 4]); - - for (int l = 0; l < 4; l++) maxv[2*l] = vmaxq_f32(srcv[2*l], srcv[2*l + 1]); - for (int l = 0; l < 2; l++) maxv[4*l] = vmaxq_f32(maxv[4*l], maxv[4*l + 2]); - for (int l = 0; l < 1; l++) maxv[8*l] = vmaxq_f32(maxv[8*l], maxv[8*l + 4]); - - const float min = vminvq_f32(minv[0]); - const float max = vmaxvq_f32(maxv[0]); - - const float d = (max - min) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = d; - y[i].m = min; - - const float32x4_t minv0 = vdupq_n_f32(min); - - for (int l = 0; l < 8; l++) { - const float32x4_t v = vmulq_n_f32(vsubq_f32(srcv[l], minv0), id); - const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(0.5f)); // needed to round to nearest - const int32x4_t vi = vcvtq_s32_f32(vf); - - y[i].qs[2*l + 0] = vgetq_lane_s32(vi, 0) | (vgetq_lane_s32(vi, 1) << 4); - y[i].qs[2*l + 1] = vgetq_lane_s32(vi, 2) | (vgetq_lane_s32(vi, 3) << 4); - } - } -#else - // scalar - quantize_row_q4_1_reference(x, vy, k); -#endif -} - -// reference implementation for deterministic creation of model files -static void quantize_row_q4_2_reference(const float * restrict x, block_q4_2 * restrict y, int k) { - assert(k % QK4_2 == 0); - - const int nb = k / QK4_2; - - for (int i = 0; i < nb; i++) { - float amax = 0.0f; // absolute max - float max = 0.0f; - - for (int l = 0; l < QK4_2; l++) { - const float v = x[i*QK4_2 + l]; - if (amax < fabsf(v)) { - amax = fabsf(v); - max = v; - } - } - - const float d = max / -8; - - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = GGML_FP32_TO_FP16(d); - - for (int l = 0; l < QK4_2; l += 2) { - const float v0 = x[i*QK4_2 + l + 0]*id; - const float v1 = x[i*QK4_2 + l + 1]*id; - - const uint8_t vi0 = MIN(15, (uint8_t)(v0 + 8.5f)); - const uint8_t vi1 = MIN(15, (uint8_t)(v1 + 8.5f)); - - assert(vi0 < 16); - assert(vi1 < 16); - - y[i].qs[l/2] = vi0 | (vi1 << 4); - } - } -} - -static void quantize_row_q4_2(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_2 == 0); - - block_q4_2 * restrict y = vy; - - quantize_row_q4_2_reference(x, y, k); -} - -static void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k) { - assert(k % QK5_0 == 0); - const int nb = k / QK5_0; - - for (int i = 0; i < nb; i++) { - float amax = 0.0f; // absolute max - float max = 0.0f; - - for (int l = 0; l < QK5_0; l++) { - const float v = x[i*QK5_0 + l]; - if (amax < fabsf(v)) { - amax = fabsf(v); - max = v; - } - } - - const float d = max / -16; - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = GGML_FP32_TO_FP16(d); - - uint32_t qh = 0; - - for (int l = 0; l < QK5_0; l += 2) { - const float v0 = x[i*QK5_0 + l + 0]*id; - const float v1 = x[i*QK5_0 + l + 1]*id; - - const uint32_t vi0 = MIN(31, (int) (v0 + 16.5f)); - const uint32_t vi1 = MIN(31, (int) (v1 + 16.5f)); - - y[i].qs[l/2] = (vi0 & 0x0F) | ((vi1 & 0x0F) << 4); - - // get the 5-th bit and store it in qh at the right position - qh |= ((vi0 & 0x10) >> 4) << (l + 0); - qh |= ((vi1 & 0x10) >> 4) << (l + 1); - } - - memcpy(&y[i].qh, &qh, sizeof(y[i].qh)); - } -} - -static void quantize_row_q5_0(const float * restrict x, void * restrict vy, int k) { - assert(k % QK5_0 == 0); - - block_q5_0 * restrict y = vy; - - quantize_row_q5_0_reference(x, y, k); -} - -static void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k) { - assert(k % QK5_1 == 0); - const int nb = k / QK5_1; - - for (int i = 0; i < nb; i++) { - float min = FLT_MAX; - float max = -FLT_MAX; - - for (int l = 0; l < QK5_1; l++) { - const float v = x[i*QK5_1 + l]; - if (v < min) min = v; - if (v > max) max = v; - } - - const float d = (max - min) / ((1 << 5) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = GGML_FP32_TO_FP16(d); - y[i].m = GGML_FP32_TO_FP16(min); - - uint32_t qh = 0; - - for (int l = 0; l < QK5_1; l += 2) { - const float v0 = (x[i*QK5_1 + l + 0] - min)*id; - const float v1 = (x[i*QK5_1 + l + 1] - min)*id; - - const uint32_t vi0 = (int) (v0 + 0.5f); - const uint32_t vi1 = (int) (v1 + 0.5f); - - y[i].qs[l/2] = (vi0 & 0x0F) | ((vi1 & 0x0F) << 4); - - // get the 5-th bit and store it in qh at the right position - qh |= ((vi0 & 0x10) >> 4) << (l + 0); - qh |= ((vi1 & 0x10) >> 4) << (l + 1); - } - - memcpy(&y[i].qh, &qh, sizeof(y[i].qh)); - } -} - -static void quantize_row_q5_1(const float * restrict x, void * restrict vy, int k) { - assert(k % QK5_1 == 0); - - block_q5_1 * restrict y = vy; - - quantize_row_q5_1_reference(x, y, k); -} - -// reference implementation for deterministic creation of model files -static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k) { - assert(k % QK8_0 == 0); - const int nb = k / QK8_0; - - for (int i = 0; i < nb; i++) { - float amax = 0.0f; // absolute max - - for (int l = 0; l < QK8_0; l++) { - const float v = x[i*QK8_0 + l]; - amax = MAX(amax, fabsf(v)); - } - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = d; - - for (int l = 0; l < QK8_0; ++l) { - const float v0 = x[i*QK8_0 + l]*id; - - y[i].qs[l] = roundf(v0); - } - } -} - -static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) { - assert(k % QK8_0 == 0); - - block_q8_0 * restrict y = vy; - quantize_row_q8_0_reference(x, y, k); +#endif } // reference implementation for deterministic creation of model files static void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k) { + assert(QK8_1 == 32); assert(k % QK8_1 == 0); const int nb = k / QK8_1; for (int i = 0; i < nb; i++) { float amax = 0.0f; // absolute max - for (int l = 0; l < QK8_1; l++) { - const float v = x[i*QK8_1 + l]; + for (int j = 0; j < QK8_1; j++) { + const float v = x[i*QK8_1 + j]; amax = MAX(amax, fabsf(v)); } @@ -1534,22 +1088,20 @@ static void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * r y[i].d = d; - int sum0 = 0; - int sum1 = 0; + int sum = 0; - for (int l = 0; l < QK8_1/2; ++l) { - const float v0 = x[i*QK8_1 + l]*id; - const float v1 = x[i*QK8_1 + QK8_1/2 + l]*id; + for (int j = 0; j < QK8_1/2; ++j) { + const float v0 = x[i*QK8_1 + j]*id; + const float v1 = x[i*QK8_1 + QK8_1/2 + j]*id; - y[i].qs[ l] = roundf(v0); - y[i].qs[QK8_1/2 + l] = roundf(v1); + y[i].qs[ j] = roundf(v0); + y[i].qs[QK8_1/2 + j] = roundf(v1); - sum0 += y[i].qs[ l]; - sum1 += y[i].qs[QK8_1/2 + l]; + sum += y[i].qs[ j]; + sum += y[i].qs[QK8_1/2 + j]; } - y[i].s0 = d * sum0; - y[i].s1 = d * sum1; + y[i].s = d * sum; } } @@ -1565,12 +1117,12 @@ static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int float32x4_t asrcv[8]; float32x4_t amaxv[8]; - for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l); - for (int l = 0; l < 8; l++) asrcv[l] = vabsq_f32(srcv[l]); + for (int j = 0; j < 8; j++) srcv[j] = vld1q_f32(x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]); - for (int l = 0; l < 4; l++) amaxv[2*l] = vmaxq_f32(asrcv[2*l], asrcv[2*l+1]); - for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]); - for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]); + for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]); const float amax = vmaxvq_f32(amaxv[0]); @@ -1579,40 +1131,21 @@ static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int y[i].d = d; - int32x4_t accv0 = vdupq_n_s32(0); - int32x4_t accv1 = vdupq_n_s32(0); + int32x4_t accv = vdupq_n_s32(0); - // low half - for (int l = 0; l < 4; l++) { - const float32x4_t v = vmulq_n_f32(srcv[l], id); + for (int j = 0; j < 8; j++) { + const float32x4_t v = vmulq_n_f32(srcv[j], id); const int32x4_t vi = vcvtnq_s32_f32(v); - y[i].qs[4*l + 0] = vgetq_lane_s32(vi, 0); - y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1); - y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2); - y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3); + y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3); - accv0 = vaddq_s32(accv0, vi); + accv = vaddq_s32(accv, vi); } - // high half - for (int l = 4; l < 8; l++) { - const float32x4_t v = vmulq_n_f32(srcv[l], id); - const int32x4_t vi = vcvtnq_s32_f32(v); - - y[i].qs[4*l + 0] = vgetq_lane_s32(vi, 0); - y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1); - y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2); - y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3); - - accv1 = vaddq_s32(accv1, vi); - } - - const int32_t sum0 = vaddvq_s32(accv0); - const int32_t sum1 = vaddvq_s32(accv1); - - y[i].s0 = d * sum0; - y[i].s1 = d * sum1; + y[i].s = d * vaddvq_s32(accv); } #elif defined(__AVX2__) || defined(__AVX__) for (int i = 0; i < nb; i++) { @@ -1661,9 +1194,7 @@ static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int #if defined(__AVX2__) // Compute the sum of the quants and set y[i].s - //y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))); - y[i].s0 = d * hsum_i32_8(_mm256_add_epi32(i0, i1)); - y[i].s1 = d * hsum_i32_8(_mm256_add_epi32(i2, i3)); + y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))); // Convert int32 to int16 i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 @@ -1693,8 +1224,7 @@ static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int // Compute the sum of the quants and set y[i].s const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3)); const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7)); - y[i].s0 = d * hsum_i32_4(s0); - y[i].s1 = d * hsum_i32_4(s1); + y[i].s = d * hsum_i32_4(_mm_add_epi32(s0, s1)); // Convert int32 to int16 ni0 = _mm_packs_epi32( ni0, ni1 ); @@ -1715,359 +1245,127 @@ static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int #endif } -static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, int k) { - assert(k % QK4_0 == 0); - const int nb = k / QK4_0; +static void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k) { + static const int qk = QK4_0; - const block_q4_0 * restrict x = vx; + assert(k % qk == 0); -#if defined(__AVX2__) - for (int i = 0; i < nb; i++) { - // scale factor - const __m256 d_v = _mm256_broadcast_ss(&x[i].d); + const int nb = k / qk; - const uint8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK4_0; l += 32) { - // Load 32x4-bit integers into 32x8-bit integers - __m256i vx8 = bytes_from_nibbles_32(pp+l/2); - - // Subtract 8 from the integers - vx8 = _mm256_sub_epi8(vx8, _mm256_set1_epi8(8)); - - // Convert to 16-bit int - const __m256i vx16_lo = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 0)); - const __m256i vx16_hi = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 1)); - - // Convert to 32-bit int -> float 32 - const __m256 vf[4] = { - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 0))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 1))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 0))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 1))) - }; - - // Scale and store - for (int j = 0; j < 4; j++) { - const __m256 result = _mm256_mul_ps(vf[j], d_v); - _mm256_storeu_ps(y + i * QK4_0 + l + j*8, result); - } - } - } -#elif defined(__ARM_NEON) - for (int i = 0; i < nb; i++) { - const float32x4_t vd = vdupq_n_f32(x[i].d); - - const uint8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK4_0; l += 16) { - // Load 16x4-bit integers into 8x8-bit integers - const uint8x8_t v8 = vld1_u8(pp + l/2); - - // Expand 4-bit qs to 8-bit bytes - const uint8x8_t v0 = vand_u8(v8, vdup_n_u8(0x0F)); - const uint8x8_t v1 = vshr_n_u8(v8, 4); - - // Convert to signed 8-bit integers - const int8x8_t vs_0 = vreinterpret_s8_u8(v0); - const int8x8_t vs_1 = vreinterpret_s8_u8(v1); - - // Subtract 8 from each byte - const int8x8_t vb_0 = vsub_s8(vs_0, vdup_n_s8(8)); - const int8x8_t vb_1 = vsub_s8(vs_1, vdup_n_s8(8)); - - // Interleave and combine - const int8x8_t vx_0 = vzip1_s8(vb_0, vb_1); - const int8x8_t vx_1 = vzip2_s8(vb_0, vb_1); - - const int8x16_t vq = vcombine_s8(vx_0, vx_1); - - // convert to 2x int16x8_t - const int16x8_t vi_0 = vmovl_s8(vget_low_s8 (vq)); - const int16x8_t vi_1 = vmovl_s8(vget_high_s8(vq)); - - // convert to 4x float32x4_t - const float32x4_t vf_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16 (vi_0))); - const float32x4_t vf_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vi_0))); - const float32x4_t vf_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16 (vi_1))); - const float32x4_t vf_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vi_1))); - - // Multiply by d - const float32x4_t r0 = vmulq_f32(vf_0, vd); - const float32x4_t r1 = vmulq_f32(vf_1, vd); - const float32x4_t r2 = vmulq_f32(vf_2, vd); - const float32x4_t r3 = vmulq_f32(vf_3, vd); - - // Store - vst1q_f32(y + i*QK4_0 + l + 0, r0); - vst1q_f32(y + i*QK4_0 + l + 4, r1); - vst1q_f32(y + i*QK4_0 + l + 8, r2); - vst1q_f32(y + i*QK4_0 + l + 12, r3); - } - } -#else - // scalar for (int i = 0; i < nb; i++) { const float d = x[i].d; - const uint8_t * restrict pp = x[i].qs; + for (int j = 0; j < qk/2; ++j) { + const int x0 = (x[i].qs[j] & 0x0F) - 8; + const int x1 = (x[i].qs[j] >> 4) - 8; - for (int l = 0; l < QK4_0; l += 2) { - const uint8_t vi = pp[l/2]; - - const int8_t vi0 = vi & 0x0F; - const int8_t vi1 = vi >> 4; - - const float v0 = (vi0 - 8)*d; - const float v1 = (vi1 - 8)*d; - - //printf("d = %f, vi = %d, vi0 = %d, vi1 = %d, v0 = %f, v1 = %f\n", d, vi, vi0, vi1, v0, v1); - - y[i*QK4_0 + l + 0] = v0; - y[i*QK4_0 + l + 1] = v1; - - assert(!isnan(y[i*QK4_0 + l + 0])); - assert(!isnan(y[i*QK4_0 + l + 1])); + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; } } -#endif } -static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, int k) { - assert(k % QK4_1 == 0); - const int nb = k / QK4_1; +static void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k) { + static const int qk = QK4_1; - const block_q4_1 * restrict x = vx; + assert(k % qk == 0); -#if defined(__AVX2__) - for (int i = 0; i < nb; i++) { - const __m256 d_v = _mm256_broadcast_ss(&x[i].d); - const __m256 d_m = _mm256_broadcast_ss(&x[i].m); + const int nb = k / qk; - const uint8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK4_1; l += 32) { - // Load 32x4-bit integers into 32x8-bit integers - __m256i vx8 = bytes_from_nibbles_32(pp+l/2); - - // Convert to 16-bit int - const __m256i vx16_lo = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 0)); - const __m256i vx16_hi = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 1)); - - // Convert to 32-bit int -> float 32 - const __m256 vf[4] = { - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 0))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 1))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 0))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 1))) - }; - - // Scale, add m and store - for (int j = 0; j < 4; j++) { - const __m256 result = _mm256_add_ps(_mm256_mul_ps(vf[j], d_v), d_m); - _mm256_storeu_ps(y + i * QK4_1 + l + j*8, result); - } - } - } -#elif defined(__ARM_NEON) - for (int i = 0; i < nb; i++) { - const float32x4_t vd = vdupq_n_f32(x[i].d); - const float32x4_t vm = vdupq_n_f32(x[i].m); - - const uint8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK4_1; l += 16) { - // Load 16x4-bit integers into 8x8-bit integers - const uint8x8_t v8 = vld1_u8(pp + l/2); - - // Expand 4-bit qs to 8-bit bytes - const uint8x8_t v0 = vand_u8(v8, vdup_n_u8(0x0F)); - const uint8x8_t v1 = vshr_n_u8(v8, 4); - - // Interleave and combine - const uint8x8_t vx_0 = vzip1_u8(v0, v1); - const uint8x8_t vx_1 = vzip2_u8(v0, v1); - - const uint8x16_t vq = vcombine_u8(vx_0, vx_1); - - // convert to 2x uint16x8_t - const uint16x8_t vi_0 = vmovl_u8(vget_low_u8 (vq)); - const uint16x8_t vi_1 = vmovl_u8(vget_high_u8(vq)); - - // convert to 4x float32x4_t - const float32x4_t vf_0 = vcvtq_f32_u32(vmovl_u16(vget_low_u16 (vi_0))); - const float32x4_t vf_1 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vi_0))); - const float32x4_t vf_2 = vcvtq_f32_u32(vmovl_u16(vget_low_u16 (vi_1))); - const float32x4_t vf_3 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vi_1))); - - // multiply by d and add m - const float32x4_t r0 = vmlaq_f32(vm, vf_0, vd); - const float32x4_t r1 = vmlaq_f32(vm, vf_1, vd); - const float32x4_t r2 = vmlaq_f32(vm, vf_2, vd); - const float32x4_t r3 = vmlaq_f32(vm, vf_3, vd); - - // Store - vst1q_f32(y + i*QK4_1 + l + 0, r0); - vst1q_f32(y + i*QK4_1 + l + 4, r1); - vst1q_f32(y + i*QK4_1 + l + 8, r2); - vst1q_f32(y + i*QK4_1 + l + 12, r3); - } - } -#else for (int i = 0; i < nb; i++) { const float d = x[i].d; const float m = x[i].m; - const uint8_t * restrict pp = x[i].qs; + for (int j = 0; j < qk/2; ++j) { + const int x0 = (x[i].qs[j] & 0x0F); + const int x1 = (x[i].qs[j] >> 4); - for (int l = 0; l < QK4_1; l += 2) { - const uint8_t vi = pp[l/2]; - - const int8_t vi0 = vi & 0x0F; - const int8_t vi1 = vi >> 4; - - const float v0 = vi0*d + m; - const float v1 = vi1*d + m; - - y[i*QK4_1 + l + 0] = v0; - y[i*QK4_1 + l + 1] = v1; - - assert(!isnan(y[i*QK4_1 + l + 0])); - assert(!isnan(y[i*QK4_1 + l + 1])); - } - } -#endif -} - -static void dequantize_row_q4_2(const void * restrict vx, float * restrict y, int k) { - assert(k % QK4_2 == 0); - const int nb = k / QK4_2; - - const block_q4_2 * restrict x = vx; - - for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d); - - const uint8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK4_2; l += 2) { - const uint8_t vi = pp[l/2]; - - const int8_t vi0 = vi & 0x0F; - const int8_t vi1 = vi >> 4; - - const float v0 = (vi0 - 8)*d; - const float v1 = (vi1 - 8)*d; - - y[i*QK4_2 + l + 0] = v0; - y[i*QK4_2 + l + 1] = v1; - - assert(!isnan(y[i*QK4_2 + l + 0])); - assert(!isnan(y[i*QK4_2 + l + 1])); + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; } } } -static void dequantize_row_q5_0(const void * restrict vx, float * restrict y, int k) { - assert(k % QK5_0 == 0); - const int nb = k / QK5_0; +static void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k) { + static const int qk = QK5_0; - const block_q5_0 * restrict x = vx; + assert(k % qk == 0); + + const int nb = k / qk; for (int i = 0; i < nb; i++) { const float d = GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict pp = x[i].qs; - uint32_t qh; memcpy(&qh, x[i].qh, sizeof(qh)); - for (int l = 0; l < QK5_0; l += 2) { - const uint8_t vi = pp[l/2]; + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - // extract the 5-th bit from qh - const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4; - const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4; + const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16; + const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; - const int8_t vi0 = (vi & 0x0F) | vh0; - const int8_t vi1 = (vi >> 4) | vh1; - - const float v0 = (vi0 - 16)*d; - const float v1 = (vi1 - 16)*d; - - y[i*QK5_0 + l + 0] = v0; - y[i*QK5_0 + l + 1] = v1; - - assert(!isnan(y[i*QK5_0 + l + 0])); - assert(!isnan(y[i*QK5_0 + l + 1])); + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; } } } -static void dequantize_row_q5_1(const void * restrict vx, float * restrict y, int k) { - assert(k % QK5_1 == 0); - const int nb = k / QK5_1; +static void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k) { + static const int qk = QK5_1; - const block_q5_1 * restrict x = vx; + assert(k % qk == 0); + + const int nb = k / qk; for (int i = 0; i < nb; i++) { const float d = GGML_FP16_TO_FP32(x[i].d); const float m = GGML_FP16_TO_FP32(x[i].m); - const uint8_t * restrict pp = x[i].qs; - uint32_t qh; memcpy(&qh, x[i].qh, sizeof(qh)); - for (int l = 0; l < QK5_1; l += 2) { - const uint8_t vi = pp[l/2]; + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - // extract the 5-th bit from qh - const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4; - const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4; + const int x0 = (x[i].qs[j] & 0x0F) | xh_0; + const int x1 = (x[i].qs[j] >> 4) | xh_1; - const uint8_t vi0 = (vi & 0x0F) | vh0; - const uint8_t vi1 = (vi >> 4) | vh1; - - const float v0 = vi0*d + m; - const float v1 = vi1*d + m; - - y[i*QK5_1 + l + 0] = v0; - y[i*QK5_1 + l + 1] = v1; - - assert(!isnan(y[i*QK5_1 + l + 0])); - assert(!isnan(y[i*QK5_1 + l + 1])); + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; } } } static void dequantize_row_q8_0(const void * restrict vx, float * restrict y, int k) { - assert(k % QK8_0 == 0); - const int nb = k / QK8_0; + static const int qk = QK8_0; + + assert(k % qk == 0); + + const int nb = k / qk; const block_q8_0 * restrict x = vx; for (int i = 0; i < nb; i++) { const float d = x[i].d; - const int8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK8_0; ++l) { - y[i*QK8_0 + l] = pp[l]*d; + for (int j = 0; j < qk; ++j) { + y[i*qk + j] = x[i].qs[j]*d; } } } static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); -static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = { [GGML_TYPE_Q4_0] = { - .dequantize_row_q = dequantize_row_q4_0, + .dequantize_row_q = (dequantize_row_q_t) dequantize_row_q4_0, .quantize_row_q = quantize_row_q4_0, .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_0_reference, .quantize_row_q_dot = quantize_row_q8_0, @@ -2075,23 +1373,15 @@ static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = { .vec_dot_type = GGML_TYPE_Q8_0, }, [GGML_TYPE_Q4_1] = { - .dequantize_row_q = dequantize_row_q4_1, + .dequantize_row_q = (dequantize_row_q_t)dequantize_row_q4_1, .quantize_row_q = quantize_row_q4_1, .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_1_reference, .quantize_row_q_dot = quantize_row_q8_1, .vec_dot_q = ggml_vec_dot_q4_1_q8_1, .vec_dot_type = GGML_TYPE_Q8_1, }, - [GGML_TYPE_Q4_2] = { - .dequantize_row_q = dequantize_row_q4_2, - .quantize_row_q = quantize_row_q4_2, - .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_2_reference, - .quantize_row_q_dot = quantize_row_q8_0, - .vec_dot_q = ggml_vec_dot_q4_2_q8_0, - .vec_dot_type = GGML_TYPE_Q8_0, - }, [GGML_TYPE_Q5_0] = { - .dequantize_row_q = dequantize_row_q5_0, + .dequantize_row_q = (dequantize_row_q_t) dequantize_row_q5_0, .quantize_row_q = quantize_row_q5_0, .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q5_0_reference, .quantize_row_q_dot = quantize_row_q8_0, @@ -2099,7 +1389,7 @@ static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = { .vec_dot_type = GGML_TYPE_Q8_0, }, [GGML_TYPE_Q5_1] = { - .dequantize_row_q = dequantize_row_q5_1, + .dequantize_row_q = (dequantize_row_q_t) dequantize_row_q5_1, .quantize_row_q = quantize_row_q5_1, .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q5_1_reference, .quantize_row_q_dot = quantize_row_q8_1, @@ -2641,6 +1931,7 @@ inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t v) { inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; } inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] + y[i]; } +inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float v) { for (int i = 0; i < n; ++i) z[i] = x[i] + v; } inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] += x[i]; } inline static void ggml_vec_acc1_f32(const int n, float * y, const float v) { for (int i = 0; i < n; ++i) y[i] += v; } inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] - y[i]; } @@ -2724,9 +2015,10 @@ inline static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t } static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_0; + const int qk = QK8_0; + const int nb = n / qk; - assert(n % QK8_0 == 0); + assert(n % qk == 0); assert(nb % 2 == 0); const block_q4_0 * restrict x = vx; @@ -2760,12 +2052,6 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b); const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b); - // interleave - const int8x16_t v0_0lz = vzip1q_s8(v0_0ls, v0_0hs); - const int8x16_t v0_0hz = vzip2q_s8(v0_0ls, v0_0hs); - const int8x16_t v0_1lz = vzip1q_s8(v0_1ls, v0_1hs); - const int8x16_t v0_1hz = vzip2q_s8(v0_1ls, v0_1hs); - // load y const int8x16_t v1_0l = vld1q_s8(y0->qs); const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); @@ -2774,21 +2060,21 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * #if defined(__ARM_FEATURE_DOTPROD) // dot product into int32x4_t - const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l), v0_0hz, v1_0h); - const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l), v0_1hz, v1_1h); + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h); sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); #else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0h)); - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1h)); const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); @@ -2834,74 +2120,181 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * // Compute combined scale for the block const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); - __m128i i32[2]; - for (int j = 0; j < 2; ++j) { - // Load 8 bytes, and unpack 4 bit fields into bytes, making 16 bytes - __m128i bx = bytes_from_nibbles_16(x[i].qs + 8*j); - __m128i by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16*j)); + const __m128i lowMask = _mm_set1_epi8(0xF); + const __m128i off = _mm_set1_epi8(8); - // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. - const __m128i off = _mm_set1_epi8( 8 ); - bx = _mm_sub_epi8( bx, off ); + const __m128i tmp = _mm_loadu_si128((const __m128i *)x[i].qs); - // Get absolute values of x vectors - const __m128i ax = _mm_sign_epi8(bx, bx); + __m128i bx = _mm_and_si128(lowMask, tmp); + __m128i by = _mm_loadu_si128((const __m128i *)y[i].qs); + bx = _mm_sub_epi8(bx, off); + const __m128i i32_0 = mul_sum_i8_pairs(bx, by); - // Sign the values of the y vectors - const __m128i sy = _mm_sign_epi8(by, bx); - - // Perform multiplication and create 16-bit values - const __m128i dot = _mm_maddubs_epi16(ax, sy); - - const __m128i ones = _mm_set1_epi16(1); - i32[j] = _mm_madd_epi16(ones, dot); - } + bx = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4)); + by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16)); + bx = _mm_sub_epi8(bx, off); + const __m128i i32_1 = mul_sum_i8_pairs(bx, by); // Convert int32_t to float - __m256 p = _mm256_cvtepi32_ps( _mm256_set_m128i( i32[0], i32[1] )); + __m256 p = _mm256_cvtepi32_ps(_mm256_set_m128i(i32_0, i32_1)); + // Apply the scale, and accumulate acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc); } *s = hsum_float_8(acc); +#elif defined(__SSSE3__) + // set constants + const __m128i lowMask = _mm_set1_epi8(0xF); + const __m128i off = _mm_set1_epi8(8); + + // Initialize accumulator with zeros + __m128 acc_0 = _mm_setzero_ps(); + __m128 acc_1 = _mm_setzero_ps(); + __m128 acc_2 = _mm_setzero_ps(); + __m128 acc_3 = _mm_setzero_ps(); + + // First round without accumulation + { + _mm_prefetch(&x[0] + sizeof(block_q4_0), _MM_HINT_T0); + _mm_prefetch(&y[0] + sizeof(block_q8_0), _MM_HINT_T0); + + // Compute combined scale for the block 0 and 1 + const __m128 d_0_1 = _mm_mul_ps( _mm_set1_ps( x[0].d ), _mm_set1_ps( y[0].d ) ); + + const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[0].qs); + + __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1); + __m128i by_0 = _mm_loadu_si128((const __m128i *)y[0].qs); + bx_0 = _mm_sub_epi8(bx_0, off); + const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0); + + __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4)); + __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[0].qs + 16)); + bx_1 = _mm_sub_epi8(bx_1, off); + const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1); + + _mm_prefetch(&x[1] + sizeof(block_q4_0), _MM_HINT_T0); + _mm_prefetch(&y[1] + sizeof(block_q8_0), _MM_HINT_T0); + + // Compute combined scale for the block 2 and 3 + const __m128 d_2_3 = _mm_mul_ps( _mm_set1_ps( x[1].d ), _mm_set1_ps( y[1].d ) ); + + const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[1].qs); + + __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3); + __m128i by_2 = _mm_loadu_si128((const __m128i *)y[1].qs); + bx_2 = _mm_sub_epi8(bx_2, off); + const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2); + + __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4)); + __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[1].qs + 16)); + bx_3 = _mm_sub_epi8(bx_3, off); + const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3); + + // Convert int32_t to float + __m128 p0 = _mm_cvtepi32_ps(i32_0); + __m128 p1 = _mm_cvtepi32_ps(i32_1); + __m128 p2 = _mm_cvtepi32_ps(i32_2); + __m128 p3 = _mm_cvtepi32_ps(i32_3); + + // Apply the scale + acc_0 = _mm_mul_ps( d_0_1, p0 ); + acc_1 = _mm_mul_ps( d_0_1, p1 ); + acc_2 = _mm_mul_ps( d_2_3, p2 ); + acc_3 = _mm_mul_ps( d_2_3, p3 ); + } + + // Main loop + for (int i = 2; i < nb; i+=2) { + _mm_prefetch(&x[i] + sizeof(block_q4_0), _MM_HINT_T0); + _mm_prefetch(&y[i] + sizeof(block_q8_0), _MM_HINT_T0); + + // Compute combined scale for the block 0 and 1 + const __m128 d_0_1 = _mm_mul_ps( _mm_set1_ps( x[i].d ), _mm_set1_ps( y[i].d ) ); + + const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[i].qs); + + __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1); + __m128i by_0 = _mm_loadu_si128((const __m128i *)y[i].qs); + bx_0 = _mm_sub_epi8(bx_0, off); + const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0); + + __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4)); + __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[i].qs + 16)); + bx_1 = _mm_sub_epi8(bx_1, off); + const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1); + + _mm_prefetch(&x[i] + 2 * sizeof(block_q4_0), _MM_HINT_T0); + _mm_prefetch(&y[i] + 2 * sizeof(block_q8_0), _MM_HINT_T0); + + // Compute combined scale for the block 2 and 3 + const __m128 d_2_3 = _mm_mul_ps( _mm_set1_ps( x[i + 1].d ), _mm_set1_ps( y[i + 1].d ) ); + + const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[i + 1].qs); + + __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3); + __m128i by_2 = _mm_loadu_si128((const __m128i *)y[i + 1].qs); + bx_2 = _mm_sub_epi8(bx_2, off); + const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2); + + __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4)); + __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[i + 1].qs + 16)); + bx_3 = _mm_sub_epi8(bx_3, off); + const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3); + + // Convert int32_t to float + __m128 p0 = _mm_cvtepi32_ps(i32_0); + __m128 p1 = _mm_cvtepi32_ps(i32_1); + __m128 p2 = _mm_cvtepi32_ps(i32_2); + __m128 p3 = _mm_cvtepi32_ps(i32_3); + + // Apply the scale + __m128 p0_d = _mm_mul_ps( d_0_1, p0 ); + __m128 p1_d = _mm_mul_ps( d_0_1, p1 ); + __m128 p2_d = _mm_mul_ps( d_2_3, p2 ); + __m128 p3_d = _mm_mul_ps( d_2_3, p3 ); + + // Acummulate + acc_0 = _mm_add_ps(p0_d, acc_0); + acc_1 = _mm_add_ps(p1_d, acc_1); + acc_2 = _mm_add_ps(p2_d, acc_2); + acc_3 = _mm_add_ps(p3_d, acc_3); + } + + *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3); #else // scalar float sumf = 0.0; + for (int i = 0; i < nb; i++) { - const float d0 = x[i].d; - const float d1 = y[i].d; - - const uint8_t * restrict p0 = x[i].qs; - const int8_t * restrict p1 = y[i].qs; - int sumi = 0; - for (int j = 0; j < QK8_0/2; j++) { - const uint8_t v0 = p0[j]; - const int i0 = (int8_t) (v0 & 0x0F) - 8; - const int i1 = (int8_t) (v0 >> 4) - 8; + for (int j = 0; j < qk/2; ++j) { + const int v0 = (x[i].qs[j] & 0x0F) - 8; + const int v1 = (x[i].qs[j] >> 4) - 8; - const int i2 = p1[2*j + 0]; - const int i3 = p1[2*j + 1]; - - sumi += i0*i2 + i1*i3; + sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); } - sumf += d0*d1*sumi; + + sumf += (x[i].d*y[i].d)*sumi; } + *s = sumf; #endif } static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_1; + const int qk = QK8_1; + const int nb = n / qk; - assert(n % QK8_1 == 0); + assert(n % qk == 0); assert(nb % 2 == 0); const block_q4_1 * restrict x = vx; const block_q8_1 * restrict y = vy; - // TODO: add AVX / WASM SIMD / etc + // TODO: add WASM SIMD #if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); @@ -2914,7 +2307,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * const block_q8_1 * restrict y0 = &y[i + 0]; const block_q8_1 * restrict y1 = &y[i + 1]; - summs += x0->m * (y0->s0 + y0->s1) + x1->m * (y1->s0 + y1->s1); + summs += x0->m * y0->s + x1->m * y1->s; const uint8x16_t m4b = vdupq_n_u8(0x0F); @@ -2927,12 +2320,6 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - // interleave - const int8x16_t v0_0lz = vzip1q_s8(v0_0l, v0_0h); - const int8x16_t v0_0hz = vzip2q_s8(v0_0l, v0_0h); - const int8x16_t v0_1lz = vzip1q_s8(v0_1l, v0_1h); - const int8x16_t v0_1hz = vzip2q_s8(v0_1l, v0_1h); - // load y const int8x16_t v1_0l = vld1q_s8(y0->qs); const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); @@ -2941,21 +2328,21 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * #if defined(__ARM_FEATURE_DOTPROD) // dot product into int32x4_t - const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l), v0_0hz, v1_0h); - const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l), v0_1hz, v1_1h); + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h); sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); #else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0h)); - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1h)); const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); @@ -2979,7 +2366,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * const float * d0 = &x[i].d; const float * d1 = &y[i].d; - summs += x[i].m * (y[i].s0 + y[i].s1); + summs += x[i].m * y[i].s; const __m256 d0v = _mm256_broadcast_ss( d0 ); const __m256 d1v = _mm256_broadcast_ss( d1 ); @@ -3001,77 +2388,86 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * #else // scalar float sumf = 0.0; + for (int i = 0; i < nb; i++) { - const float d0 = x[i].d; - const float m0 = x[i].m; - const float d1 = y[i].d; + int sumi = 0; - const uint8_t * restrict p0 = x[i].qs; - const int8_t * restrict p1 = y[i].qs; + for (int j = 0; j < qk/2; ++j) { + const int v0 = (x[i].qs[j] & 0x0F); + const int v1 = (x[i].qs[j] >> 4); - // TODO: this is very slow .. - for (int j = 0; j < QK8_1/2; j++) { - const uint8_t v0 = p0[j]; - - const float f0 = d0*(v0 & 0x0F) + m0; - const float f1 = d0*(v0 >> 4) + m0; - - const float f2 = d1*p1[2*j + 0]; - const float f3 = d1*p1[2*j + 1]; - - sumf += f0*f2 + f1*f3; + sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); } + + sumf += (x[i].d*y[i].d)*sumi + x[i].m*y[i].s; } + *s = sumf; #endif } -static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_0; +static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int qk = QK8_0; + const int nb = n / qk; - assert(n % QK8_0 == 0); + assert(n % qk == 0); assert(nb % 2 == 0); - assert(QK8_0 == 2*QK4_2); + assert(qk == QK5_0); - const block_q4_2 * restrict x = vx; + const block_q5_0 * restrict x = vx; const block_q8_0 * restrict y = vy; #if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); - for (int i = 0; i < nb; i += 2) { - const block_q4_2 * restrict x0_0 = &x[2*(i + 0) + 0]; - const block_q4_2 * restrict x0_1 = &x[2*(i + 0) + 1]; - const block_q4_2 * restrict x1_0 = &x[2*(i + 1) + 0]; - const block_q4_2 * restrict x1_1 = &x[2*(i + 1) + 1]; + uint32_t qh0; + uint32_t qh1; - const block_q8_0 * restrict y0 = &y[i + 0]; + uint64_t tmp0[4]; + uint64_t tmp1[4]; + + for (int i = 0; i < nb; i += 2) { + const block_q5_0 * restrict x0 = &x[i]; + const block_q5_0 * restrict x1 = &x[i + 1]; + const block_q8_0 * restrict y0 = &y[i]; const block_q8_0 * restrict y1 = &y[i + 1]; - const uint8x16_t m4b = vdupq_n_u8(0x0F); - const int8x16_t s8b = vdupq_n_s8(0x8); + const uint8x16_t m4b = vdupq_n_u8(0x0F); - const uint8x16_t v0_0 = vcombine_u8(vld1_u8(x0_0->qs), vld1_u8(x0_1->qs)); - const uint8x16_t v0_1 = vcombine_u8(vld1_u8(x1_0->qs), vld1_u8(x1_1->qs)); + // extract the 5th bit via lookup table ((!b) << 4) + memcpy(&qh0, x0->qh, sizeof(qh0)); + memcpy(&qh1, x1->qh, sizeof(qh1)); + + tmp0[0] = table_b2b_1[(qh0 >> 0) & 0xFF]; + tmp0[1] = table_b2b_1[(qh0 >> 8) & 0xFF]; + tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF]; + tmp0[3] = table_b2b_1[(qh0 >> 24) ]; + + tmp1[0] = table_b2b_1[(qh1 >> 0) & 0xFF]; + tmp1[1] = table_b2b_1[(qh1 >> 8) & 0xFF]; + tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF]; + tmp1[3] = table_b2b_1[(qh1 >> 24) ]; + + const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0)); + const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2)); + const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0)); + const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2)); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); // 4-bit -> 8-bit - const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); - const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); - const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); - const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - // sub 8 - const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b); - const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b); - const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b); - const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b); - - // interleave - const int8x16_t v0_0lz = vzip1q_s8(v0_0ls, v0_0hs); - const int8x16_t v0_0hz = vzip2q_s8(v0_0ls, v0_0hs); - const int8x16_t v0_1lz = vzip1q_s8(v0_1ls, v0_1hs); - const int8x16_t v0_1hz = vzip2q_s8(v0_1ls, v0_1hs); + // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero) + const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0); + const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0); + const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1); + const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1); // load y const int8x16_t v1_0l = vld1q_s8(y0->qs); @@ -3079,187 +2475,45 @@ static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * const int8x16_t v1_1l = vld1q_s8(y1->qs); const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + const float x0d = GGML_FP16_TO_FP32(x0->d); + const float x1d = GGML_FP16_TO_FP32(x1->d); + #if defined(__ARM_FEATURE_DOTPROD) - sumv0 = vmlaq_n_f32(sumv0, vaddq_f32( - vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l)), GGML_FP16_TO_FP32(x0_0->d)), - vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0hz, v1_0h)), GGML_FP16_TO_FP32(x0_1->d))), y0->d); - - sumv1 = vmlaq_n_f32(sumv1, vaddq_f32( - vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l)), GGML_FP16_TO_FP32(x1_0->d)), - vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1hz, v1_1h)), GGML_FP16_TO_FP32(x1_1->d))), y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), + vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), x0d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), + vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), x1d*y1->d); #else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h)); - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h)); const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - sumv0 = vmlaq_n_f32(sumv0, vaddq_f32( - vmulq_n_f32(vcvtq_f32_s32(pl0), GGML_FP16_TO_FP32(x0_0->d)), - vmulq_n_f32(vcvtq_f32_s32(ph0), GGML_FP16_TO_FP32(x0_1->d))), y0->d); - - sumv1 = vmlaq_n_f32(sumv1, vaddq_f32( - vmulq_n_f32(vcvtq_f32_s32(pl1), GGML_FP16_TO_FP32(x1_0->d)), - vmulq_n_f32(vcvtq_f32_s32(ph1), GGML_FP16_TO_FP32(x1_1->d))), y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1d*y1->d); #endif } *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); -#elif defined(__AVX2__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - - // Main loop - for (int i = 0; i < nb; i++) { - /* Compute combined scale for the block */ - const __m128 d0 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 0].d)); - const __m128 d1 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 1].d)); - const __m256 d = _mm256_mul_ps(_mm256_set_m128(d1, d0), _mm256_broadcast_ss(&y[i].d)); - - __m128i bx0 = bytes_from_nibbles_16(x[2*i + 0].qs); - __m128i bx1 = bytes_from_nibbles_16(x[2*i + 1].qs); - __m256i bx = _mm256_set_m128i(bx1, bx0); - - // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. - const __m256i off = _mm256_set1_epi8(8); - bx = _mm256_sub_epi8(bx, off); - - __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); - - const __m256 q = mul_sum_i8_pairs_float(bx, by); - - /* Multiply q with scale and accumulate */ - acc = _mm256_fmadd_ps(d, q, acc); - } - - *s = hsum_float_8(acc); -#else - // scalar - float sumf = 0.0; - for (int i = 0; i < nb; i++) { - const uint8_t * restrict x0 = x[2*i + 0].qs; - const uint8_t * restrict x1 = x[2*i + 1].qs; - const int8_t * restrict y0 = y[i].qs; - - const float d0 = GGML_FP16_TO_FP32(x[2*i + 0].d); - const float d1 = GGML_FP16_TO_FP32(x[2*i + 1].d); - - int sumi_0 = 0; - int sumi_1 = 0; - - for (int j = 0; j < QK8_0/4; j++) { - const uint8_t v0 = x0[j]; - const uint8_t v1 = x1[j]; - - const int i0_0 = (int8_t) (v0 & 0x0F) - 8; - const int i1_0 = (int8_t) (v0 >> 4) - 8; - - const int i0_1 = (int8_t) (v1 & 0x0F) - 8; - const int i1_1 = (int8_t) (v1 >> 4) - 8; - - const int i2_0 = y0[2*j + 0]; - const int i3_0 = y0[2*j + 1]; - - const int i2_1 = y0[2*(j + QK8_0/4) + 0]; - const int i3_1 = y0[2*(j + QK8_0/4) + 1]; - - sumi_0 += i0_0*i2_0 + i1_0*i3_0; - sumi_1 += i0_1*i2_1 + i1_1*i3_1; - } - - sumf += (d0 * y[i].d) * sumi_0; - sumf += (d1 * y[i].d) * sumi_1; - } - *s = sumf; -#endif -} - -static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_0; - - assert(n % QK8_0 == 0); - assert(nb % 2 == 0); - assert(QK8_0 == QK5_0); - - const block_q5_0 * restrict x = vx; - const block_q8_0 * restrict y = vy; - -#if defined(__ARM_NEON) - float32x4_t sumv = vdupq_n_f32(0.0f); - - uint64_t tmp[4]; - - for (int i = 0; i < nb; ++i) { - const block_q5_0 * restrict x0 = &x[i]; - const block_q8_0 * restrict y0 = &y[i]; - - const uint8x16_t m4b = vdupq_n_u8(0x0F); - const int8x16_t s16b = vdupq_n_s8(0x10); - - // extract the 5th bit - uint32_t qh; - memcpy(&qh, x0->qh, sizeof(qh)); - - tmp[0] = table_b2b_u[(qh >> 0) & 0xFF]; - tmp[1] = table_b2b_u[(qh >> 8) & 0xFF]; - tmp[2] = table_b2b_u[(qh >> 16) & 0xFF]; - tmp[3] = table_b2b_u[(qh >> 24) ]; - - const int8x16_t qhl = vld1q_s8((const int8_t *)(tmp + 0)); - const int8x16_t qhh = vld1q_s8((const int8_t *)(tmp + 2)); - - const uint8x16_t v0 = vld1q_u8(x0->qs); - - // 4-bit -> 8-bit - const int8x16_t v0l = vreinterpretq_s8_u8(vandq_u8 (v0, m4b)); - const int8x16_t v0h = vreinterpretq_s8_u8(vshrq_n_u8(v0, 4)); - - // interleave - const int8x16_t v0lz = vzip1q_s8(v0l, v0h); - const int8x16_t v0hz = vzip2q_s8(v0l, v0h); - - // add high bit and sub 16 - const int8x16_t v0lf = vsubq_s8(vorrq_s8(v0lz, qhl), s16b); - const int8x16_t v0hf = vsubq_s8(vorrq_s8(v0hz, qhh), s16b); - - // load y - const int8x16_t v1l = vld1q_s8(y0->qs); - const int8x16_t v1h = vld1q_s8(y0->qs + 16); - - const float x0d = GGML_FP16_TO_FP32(x0->d); - -#if defined(__ARM_FEATURE_DOTPROD) - sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), v0lf, v1l), - vdotq_s32(vdupq_n_s32(0), v0hf, v1h))), x0d*y0->d); -#else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0lf), vget_low_s8 (v1l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0lf), vget_high_s8(v1l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0hf), vget_low_s8 (v1h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0hf), vget_high_s8(v1h)); - - const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); - const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); - - sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); -#endif - } - - *s = vaddvq_f32(sumv); #elif defined(__wasm_simd128__) v128_t sumv = wasm_f32x4_splat(0.0f); + uint32_t qh; uint64_t tmp[4]; + // TODO: check if unrolling this is better for (int i = 0; i < nb; ++i) { const block_q5_0 * restrict x0 = &x[i]; const block_q8_0 * restrict y0 = &y[i]; @@ -3268,13 +2522,12 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * const v128_t s16b = wasm_i8x16_splat(0x10); // extract the 5th bit - uint32_t qh; memcpy(&qh, x0->qh, sizeof(qh)); - tmp[0] = table_b2b_u[(qh >> 0) & 0xFF]; - tmp[1] = table_b2b_u[(qh >> 8) & 0xFF]; - tmp[2] = table_b2b_u[(qh >> 16) & 0xFF]; - tmp[3] = table_b2b_u[(qh >> 24) ]; + tmp[0] = table_b2b_1[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_1[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_1[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_1[(qh >> 24) ]; const v128_t qhl = wasm_v128_load(tmp + 0); const v128_t qhh = wasm_v128_load(tmp + 2); @@ -3285,13 +2538,9 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * const v128_t v0l = wasm_v128_and (v0, m4b); const v128_t v0h = wasm_u8x16_shr(v0, 4); - // interleave - const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23); - const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31); - - // add high bit and sub 16 - const v128_t v0lf = wasm_i8x16_sub(wasm_v128_or(v0lz, qhl), s16b); - const v128_t v0hf = wasm_i8x16_sub(wasm_v128_or(v0hz, qhh), s16b); + // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero) + const v128_t v0lf = wasm_i8x16_sub(v0l, qhl); + const v128_t v0hf = wasm_i8x16_sub(v0h, qhh); // load y const v128_t v1l = wasm_v128_load(y0->qs); @@ -3347,134 +2596,161 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * #else // scalar float sumf = 0.0; - for (int i = 0; i < nb; i++) { - const uint8_t * restrict x0 = x[i].qs; - const int8_t * restrict y0 = y[i].qs; + for (int i = 0; i < nb; i++) { uint32_t qh; memcpy(&qh, x[i].qh, sizeof(qh)); - const float d = GGML_FP16_TO_FP32(x[i].d); + int sumi = 0; - int sxy = 0; + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; + const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12)); - for (int j = 0; j < QK8_0/2; j++) { - const uint8_t v0 = x0[j]; + const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16; + const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; - const int x0_0h = ((qh & (1u << (2*j + 0))) >> (2*j + 0)) << 4; - const int x1_0h = ((qh & (1u << (2*j + 1))) >> (2*j + 1)) << 4; - - const int x0_0 = ((v0 & 0x0F) | x0_0h) - 16; - const int x1_0 = ((v0 >> 4) | x1_0h) - 16; - - const int y0_0 = y0[2*j + 0]; - const int y1_0 = y0[2*j + 1]; - - sxy += x0_0*y0_0 + x1_0*y1_0; + sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]); } - sumf += (d*sxy)*y[i].d; + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi; } + *s = sumf; #endif } static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_1; + const int qk = QK8_1; + const int nb = n / qk; - assert(n % QK8_1 == 0); + assert(n % qk == 0); assert(nb % 2 == 0); - assert(QK8_1 == QK5_1); + assert(qk == QK5_1); const block_q5_1 * restrict x = vx; const block_q8_1 * restrict y = vy; #if defined(__ARM_NEON) - float32x4_t sumv = vdupq_n_f32(0.0f); + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); - float summs = 0.0f; + float summs0 = 0.0f; + float summs1 = 0.0f; - uint64_t tmp[4]; + uint32_t qh0; + uint32_t qh1; - for (int i = 0; i < nb; ++i) { + uint64_t tmp0[4]; + uint64_t tmp1[4]; + + for (int i = 0; i < nb; i += 2) { const block_q5_1 * restrict x0 = &x[i]; + const block_q5_1 * restrict x1 = &x[i + 1]; const block_q8_1 * restrict y0 = &y[i]; + const block_q8_1 * restrict y1 = &y[i + 1]; - summs += GGML_FP16_TO_FP32(x0->m) * (y0->s0 + y0->s1); + const uint8x16_t m4b = vdupq_n_u8(0x0F); - // extract the 5th bit - uint32_t qh; - memcpy(&qh, x0->qh, sizeof(qh)); + summs0 += GGML_FP16_TO_FP32(x0->m) * y0->s; + summs1 += GGML_FP16_TO_FP32(x1->m) * y1->s; - tmp[0] = table_b2b_u[(qh >> 0) & 0xFF]; - tmp[1] = table_b2b_u[(qh >> 8) & 0xFF]; - tmp[2] = table_b2b_u[(qh >> 16) & 0xFF]; - tmp[3] = table_b2b_u[(qh >> 24) ]; + // extract the 5th bit via lookup table ((b) << 4) + memcpy(&qh0, x0->qh, sizeof(qh0)); + memcpy(&qh1, x1->qh, sizeof(qh1)); - const int8x16_t qhl = vld1q_s8((const int8_t *)(tmp + 0)); - const int8x16_t qhh = vld1q_s8((const int8_t *)(tmp + 2)); + tmp0[0] = table_b2b_0[(qh0 >> 0) & 0xFF]; + tmp0[1] = table_b2b_0[(qh0 >> 8) & 0xFF]; + tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF]; + tmp0[3] = table_b2b_0[(qh0 >> 24) ]; - const uint8x16_t v0 = vld1q_u8(x0->qs); + tmp1[0] = table_b2b_0[(qh1 >> 0) & 0xFF]; + tmp1[1] = table_b2b_0[(qh1 >> 8) & 0xFF]; + tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF]; + tmp1[3] = table_b2b_0[(qh1 >> 24) ]; + + const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0)); + const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2)); + const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0)); + const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2)); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); // 4-bit -> 8-bit - const int8x16_t v0l = vreinterpretq_s8_u8(vandq_u8 (v0, vdupq_n_u8(0x0F))); - const int8x16_t v0h = vreinterpretq_s8_u8(vshrq_n_u8(v0, 4)); + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - // interleave - const int8x16_t v0lz = vzip1q_s8(v0l, v0h); - const int8x16_t v0hz = vzip2q_s8(v0l, v0h); - - // add - const int8x16_t v0lf = vorrq_s8(v0lz, qhl); - const int8x16_t v0hf = vorrq_s8(v0hz, qhh); + // add high bit + const int8x16_t v0_0lf = vorrq_s8(v0_0l, qhl0); + const int8x16_t v0_0hf = vorrq_s8(v0_0h, qhh0); + const int8x16_t v0_1lf = vorrq_s8(v0_1l, qhl1); + const int8x16_t v0_1hf = vorrq_s8(v0_1h, qhh1); // load y - const int8x16_t v1l = vld1q_s8(y0->qs); - const int8x16_t v1h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); const float x0d = GGML_FP16_TO_FP32(x0->d); + const float x1d = GGML_FP16_TO_FP32(x1->d); #if defined(__ARM_FEATURE_DOTPROD) - sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), v0lf, v1l), - vdotq_s32(vdupq_n_s32(0), v0hf, v1h))), x0d*y0->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), + vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), x0d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), + vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), x1d*y1->d); #else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0lf), vget_low_s8 (v1l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0lf), vget_high_s8(v1l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0hf), vget_low_s8 (v1h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0hf), vget_high_s8(v1h)); + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h)); const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1d*y1->d); #endif } - *s = vaddvq_f32(sumv) + summs; + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1; #elif defined(__wasm_simd128__) v128_t sumv = wasm_f32x4_splat(0.0f); float summs = 0.0f; + uint32_t qh; uint64_t tmp[4]; + // TODO: check if unrolling this is better for (int i = 0; i < nb; ++i) { const block_q5_1 * restrict x0 = &x[i]; const block_q8_1 * restrict y0 = &y[i]; - summs += GGML_FP16_TO_FP32(x0->m) * (y0->s0 + y0->s1); + summs += GGML_FP16_TO_FP32(x0->m) * y0->s; const v128_t m4b = wasm_i8x16_splat(0x0F); // extract the 5th bit - uint32_t qh; memcpy(&qh, x0->qh, sizeof(qh)); - tmp[0] = table_b2b_u[(qh >> 0) & 0xFF]; - tmp[1] = table_b2b_u[(qh >> 8) & 0xFF]; - tmp[2] = table_b2b_u[(qh >> 16) & 0xFF]; - tmp[3] = table_b2b_u[(qh >> 24) ]; + tmp[0] = table_b2b_0[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_0[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_0[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_0[(qh >> 24) ]; const v128_t qhl = wasm_v128_load(tmp + 0); const v128_t qhh = wasm_v128_load(tmp + 2); @@ -3487,13 +2763,9 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * static bool x = true; - // interleave - const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23); - const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31); - // add high bit - const v128_t v0lf = wasm_v128_or(v0lz, qhl); - const v128_t v0hf = wasm_v128_or(v0hz, qhh); + const v128_t v0lf = wasm_v128_or(v0l, qhl); + const v128_t v0hf = wasm_v128_or(v0h, qhh); // load y const v128_t v1l = wasm_v128_load(y0->qs); @@ -3526,13 +2798,14 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * #elif defined(__AVX2__) // Initialize accumulator with zeros __m256 acc = _mm256_setzero_ps(); + float summs = 0.0f; // Main loop for (int i = 0; i < nb; i++) { const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)); - summs += GGML_FP16_TO_FP32(x[i].m) * (y[i].s0 + y[i].s1); + summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; __m256i bx = bytes_from_nibbles_32(x[i].qs); __m256i bxhi = bytes_from_bits_32(x[i].qh); @@ -3549,36 +2822,26 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * *s = hsum_float_8(acc) + summs; #else + // scalar float sumf = 0.0; for (int i = 0; i < nb; i++) { - const uint8_t * restrict x0 = x[i].qs; - const int8_t * restrict y0 = y[i].qs; - uint32_t qh; memcpy(&qh, x[i].qh, sizeof(qh)); - const float d = GGML_FP16_TO_FP32(x[i].d); - const float m = GGML_FP16_TO_FP32(x[i].m); + int sumi = 0; - int sxy = 0; + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - for (int j = 0; j < QK8_1/2; j++) { - const uint8_t v0 = x0[j]; + const int32_t x0 = (x[i].qs[j] & 0xF) | xh_0; + const int32_t x1 = (x[i].qs[j] >> 4) | xh_1; - const int x0_0h = ((qh & (1u << (2*j + 0))) >> (2*j + 0)) << 4; - const int x1_0h = ((qh & (1u << (2*j + 1))) >> (2*j + 1)) << 4; - - const int x0_0 = (v0 & 0x0F) | x0_0h; - const int x1_0 = (v0 >> 4) | x1_0h; - - const int y0_0 = y0[2*j + 0]; - const int y1_0 = y0[2*j + 1]; - - sxy += x0_0*y0_0 + x1_0*y1_0; + sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]); } - sumf += (d*sxy)*y[i].d + m*(y[i].s0 + y[i].s1); + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; } *s = sumf; @@ -3586,11 +2849,11 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * } static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_0; + const int qk = QK8_0; + const int nb = n / qk; - assert(n % QK8_0 == 0); + assert(n % qk == 0); assert(nb % 2 == 0); - assert(QK8_0 == QK8_0); const block_q8_0 * restrict x = vx; const block_q8_0 * restrict y = vy; @@ -3670,16 +2933,10 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * float sumf = 0.0; for (int i = 0; i < nb; i++) { - const int8_t * restrict x0 = x[i].qs; - const int8_t * restrict y0 = y[i].qs; - int sumi = 0; - for (int j = 0; j < QK8_0; j++) { - const int v0 = x0[j]; - const int v1 = y0[j]; - - sumi += v0*v1; + for (int j = 0; j < qk; j++) { + sumi += x[i].qs[j]*y[i].qs[j]; } sumf += (x[i].d*y[i].d)*sumi; @@ -3808,6 +3065,7 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float v) { inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, x, x); *s = sqrtf(*s); } inline static void ggml_vec_sqr_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i]; } inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); } +inline static void ggml_vec_log_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]); } inline static void ggml_vec_abs_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); } inline static void ggml_vec_sgn_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); } inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; } @@ -3849,12 +3107,12 @@ inline static float ggml_silu_f32(float x) { return x/(1.0f + expf(-x)); } -inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { - const uint16_t * i16 = (const uint16_t *) x; - for (int i = 0; i < n; ++i) { - y[i] = table_silu_f16[i16[i]]; - } -} +//inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { +// const uint16_t * i16 = (const uint16_t *) x; +// for (int i = 0; i < n; ++i) { +// y[i] = table_silu_f16[i16[i]]; +// } +//} #ifdef GGML_SILU_FP16 inline static void ggml_vec_silu_f32(const int n, float * y, const float * x) { @@ -3873,6 +3131,29 @@ inline static void ggml_vec_silu_f32(const int n, float * y, const float * x) { } #endif +inline static float ggml_silu_backward_f32(float x, float dy) { + const float s = 1.0f/(1.0f + expf(-x)); + return dy*s*(1.0f + x*(1.0f - s)); +} + +#ifdef GGML_SILU_FP16 +inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) { + for (int i = 0; i < n; ++i) { + // we did not use x[i] to compute forward silu but its f16 equivalent + // take derivative at f16 of x[i]: + ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); + float usedx = GGML_FP16_TO_FP32(fp16); + dx[i] = ggml_silu_backward_f32(usedx, dy[i]); + } +} +#else +inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) { + for (int i = 0; i < n; ++i) { + dx[i] = ggml_silu_backward_f32(x[i], dy[i]); + } +} +#endif + inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) { #ifndef GGML_USE_ACCELERATE ggml_float sum = 0.0; @@ -3943,7 +3224,6 @@ static const int GGML_BLCK_SIZE[GGML_TYPE_COUNT] = { [GGML_TYPE_F16] = 1, [GGML_TYPE_Q4_0] = QK4_0, [GGML_TYPE_Q4_1] = QK4_1, - [GGML_TYPE_Q4_2] = QK4_2, [GGML_TYPE_Q5_0] = QK5_0, [GGML_TYPE_Q5_1] = QK5_1, [GGML_TYPE_Q8_0] = QK8_0, @@ -3959,7 +3239,6 @@ static const size_t GGML_TYPE_SIZE[GGML_TYPE_COUNT] = { [GGML_TYPE_F16] = sizeof(ggml_fp16_t), [GGML_TYPE_Q4_0] = sizeof(block_q4_0), [GGML_TYPE_Q4_1] = sizeof(block_q4_1), - [GGML_TYPE_Q4_2] = sizeof(block_q4_2), [GGML_TYPE_Q5_0] = sizeof(block_q5_0), [GGML_TYPE_Q5_1] = sizeof(block_q5_1), [GGML_TYPE_Q8_0] = sizeof(block_q8_0), @@ -3976,7 +3255,6 @@ static const char * GGML_TYPE_NAME[GGML_TYPE_COUNT] = { [GGML_TYPE_F16] = "f16", [GGML_TYPE_Q4_0] = "q4_0", [GGML_TYPE_Q4_1] = "q4_1", - [GGML_TYPE_Q4_2] = "q4_2", [GGML_TYPE_Q5_0] = "q5_0", [GGML_TYPE_Q5_1] = "q5_1", [GGML_TYPE_Q8_0] = "q8_0", @@ -3992,7 +3270,6 @@ static bool GGML_IS_QUANTIZED[GGML_TYPE_COUNT] = { [GGML_TYPE_F16] = false, [GGML_TYPE_Q4_0] = true, [GGML_TYPE_Q4_1] = true, - [GGML_TYPE_Q4_2] = true, [GGML_TYPE_Q5_0] = true, [GGML_TYPE_Q5_1] = true, [GGML_TYPE_Q8_0] = true, @@ -4008,12 +3285,16 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { "DUP", "ADD", + "ADD1", + "ACC", "SUB", "MUL", "DIV", "SQR", "SQRT", + "LOG", "SUM", + "SUM_ROWS", "MEAN", "REPEAT", "ABS", @@ -4023,12 +3304,15 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { "RELU", "GELU", "SILU", + "SILU_BACK", "NORM", "RMS_NORM", + "RMS_NORM_BACK", "MUL_MAT", "SCALE", + "SET", "CPY", "CONT", "RESHAPE", @@ -4036,9 +3320,13 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { "PERMUTE", "TRANSPOSE", "GET_ROWS", + "GET_ROWS_BACK", + "DIAG", "DIAG_MASK_INF", + "DIAG_MASK_ZERO", "SOFT_MAX", "ROPE", + "ROPE_BACK", "ALIBI", "CONV_1D_1S", "CONV_1D_2S", @@ -4050,19 +3338,23 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { "MAP_BINARY", }; -static_assert(GGML_OP_COUNT == 39, "GGML_OP_COUNT != 39"); +static_assert(GGML_OP_COUNT == 50, "GGML_OP_COUNT != 50"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", "x", "x+y", + "x+y", + "view(x,nb,offset)+=y->x", "x-y", "x*y", "x/y", "x^2", "√x", + "log(x)", "Σx", + "Σx_k", "Σx/n", "repeat(x)", "abs(x)", @@ -4072,12 +3364,15 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "relu(x)", "gelu(x)", "silu(x)", + "silu_back(x)", "norm(x)", "rms_norm(x)", + "rms_norm_back(x)", "X*Y", "x*v", + "y-\\>view(x)", "x-\\>y", "cont(x)", "reshape(x)", @@ -4085,9 +3380,13 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "permute(x)", "transpose(x)", "get_rows(x)", + "get_rows_back(x)", + "diag(x)", "diag_mask_inf(x)", + "diag_mask_zero(x)", "soft_max(x)", "rope(x)", + "rope_back(x)", "alibi(x)", "conv_1d_1s(x)", "conv_1d_2s(x)", @@ -4099,7 +3398,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "f(x,y)", }; -static_assert(GGML_OP_COUNT == 39, "GGML_OP_COUNT != 39"); +static_assert(GGML_OP_COUNT == 50, "GGML_OP_COUNT != 50"); static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN"); static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN"); @@ -4277,7 +3576,6 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_F16: wtype = GGML_TYPE_F16; break; case GGML_FTYPE_MOSTLY_Q4_0: wtype = GGML_TYPE_Q4_0; break; case GGML_FTYPE_MOSTLY_Q4_1: wtype = GGML_TYPE_Q4_1; break; - case GGML_FTYPE_MOSTLY_Q4_2: wtype = GGML_TYPE_Q4_2; break; case GGML_FTYPE_MOSTLY_Q5_0: wtype = GGML_TYPE_Q5_0; break; case GGML_FTYPE_MOSTLY_Q5_1: wtype = GGML_TYPE_Q5_1; break; case GGML_FTYPE_MOSTLY_Q8_0: wtype = GGML_TYPE_Q8_0; break; @@ -4338,9 +3636,9 @@ static inline int ggml_up32(int n) { return (n + 31) & ~31; } -static inline int ggml_up64(int n) { - return (n + 63) & ~63; -} +//static inline int ggml_up64(int n) { +// return (n + 63) & ~63; +//} static inline int ggml_up(int n, int m) { // assert m is a power of 2 @@ -4496,6 +3794,20 @@ size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch) return result; } +// IMPORTANT: +// when creating "opt" tensors, always save and load the scratch buffer +// this is an error prone process, but it is necessary to support inplace +// operators when using scratch buffers +// TODO: implement a better way +void ggml_scratch_save(struct ggml_context * ctx) { + ctx->scratch_save = ctx->scratch; + ctx->scratch.data = NULL; +} + +void ggml_scratch_load(struct ggml_context * ctx) { + ctx->scratch = ctx->scratch_save; +} + //////////////////////////////////////////////////////////////////////////////// struct ggml_tensor * ggml_new_tensor_impl( @@ -4584,6 +3896,7 @@ struct ggml_tensor * ggml_new_tensor_impl( *result = (struct ggml_tensor) { /*.type =*/ type, + /*.backend =*/ GGML_BACKEND_CPU, /*.n_dims =*/ n_dims, /*.ne =*/ { 1, 1, 1, 1 }, /*.nb =*/ { 0, 0, 0, 0 }, @@ -4666,12 +3979,11 @@ struct ggml_tensor * ggml_new_tensor_4d( } struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) { - ctx->scratch_save = ctx->scratch; - ctx->scratch.data = NULL; + ggml_scratch_save(ctx); struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 1); - ctx->scratch = ctx->scratch_save; + ggml_scratch_load(ctx); ggml_set_i32(result, value); @@ -4679,12 +3991,11 @@ struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) { } struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value) { - ctx->scratch_save = ctx->scratch; - ctx->scratch.data = NULL; + ggml_scratch_save(ctx); struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1); - ctx->scratch = ctx->scratch_save; + ggml_scratch_load(ctx); ggml_set_f32(result, value); @@ -5050,6 +4361,113 @@ struct ggml_tensor * ggml_add_inplace( return ggml_add_impl(ctx, a, b, true); } +// ggml_add1 + +struct ggml_tensor * ggml_add1_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + bool inplace) { + GGML_ASSERT(ggml_is_scalar(b)); + GGML_ASSERT(ggml_is_padded_1d(a)); + + bool is_node = false; + + if (!inplace && (a->grad || b->grad)) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_ADD1; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + + return result; +} + +struct ggml_tensor * ggml_add1( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b) { + return ggml_add1_impl(ctx, a, b, false); +} + +struct ggml_tensor * ggml_add1_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b) { + return ggml_add1_impl(ctx, a, b, true); +} + +// ggml_acc + +struct ggml_tensor * ggml_acc_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset, + bool inplace) { + GGML_ASSERT(ggml_nelements(b) <= ggml_nelements(a)); + GGML_ASSERT(ggml_is_contiguous(a)); + GGML_ASSERT(a->type == GGML_TYPE_F32); + GGML_ASSERT(b->type == GGML_TYPE_F32); + + bool is_node = false; + + if (!inplace && (a->grad || b->grad)) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + ggml_scratch_save(ctx); + + struct ggml_tensor * c = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 5); + + ((int32_t *) c->data)[0] = nb1; + ((int32_t *) c->data)[1] = nb2; + ((int32_t *) c->data)[2] = nb3; + ((int32_t *) c->data)[3] = offset; + ((int32_t *) c->data)[4] = inplace ? 1 : 0; + + ggml_scratch_load(ctx); + + result->op = GGML_OP_ACC; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + result->opt[0] = c; + + return result; +} + +struct ggml_tensor * ggml_acc( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset) { + return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, false); +} + +struct ggml_tensor * ggml_acc_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset) { + return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, true); +} + // ggml_sub struct ggml_tensor * ggml_sub_impl( @@ -5243,6 +4661,41 @@ struct ggml_tensor * ggml_sqrt_inplace( return ggml_sqrt_impl(ctx, a, true); } + +// ggml_log + +struct ggml_tensor * ggml_log_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + bool inplace) { + bool is_node = false; + + if (!inplace && (a->grad)) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_LOG; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + return result; +} + +struct ggml_tensor * ggml_log( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_log_impl(ctx, a, false); +} + +struct ggml_tensor * ggml_log_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_log_impl(ctx, a, true); +} + // ggml_sum struct ggml_tensor * ggml_sum( @@ -5264,6 +4717,33 @@ struct ggml_tensor * ggml_sum( return result; } + +// ggml_sum_rows + +struct ggml_tensor * ggml_sum_rows( + struct ggml_context * ctx, + struct ggml_tensor * a) { + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + int64_t ne[4] = {1,1,1,1}; + for (int i=1; in_dims; ++i) { + ne[i] = a->ne[i]; + } + + struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, a->n_dims, ne); + + result->op = GGML_OP_SUM_ROWS; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + return result; +} + // ggml_mean struct ggml_tensor * ggml_mean( @@ -5554,6 +5034,29 @@ struct ggml_tensor * ggml_silu_inplace( return ggml_silu_impl(ctx, a, true); } +// ggml_silu_back + +struct ggml_tensor * ggml_silu_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b) { + bool is_node = false; + + if (a->grad || b->grad) { + // TODO: implement backward + is_node = true; + } + + struct ggml_tensor * result = ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_SILU_BACK; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + + return result; +} + // ggml_norm struct ggml_tensor * ggml_norm_impl( @@ -5596,7 +5099,6 @@ struct ggml_tensor * ggml_rms_norm_impl( bool is_node = false; if (!inplace && (a->grad)) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5622,6 +5124,28 @@ struct ggml_tensor * ggml_rms_norm_inplace( return ggml_rms_norm_impl(ctx, a, true); } +struct ggml_tensor * ggml_rms_norm_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b) { + bool is_node = false; + + if (a->grad) { + // TODO: implement backward + is_node = true; + } + + struct ggml_tensor * result = ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_RMS_NORM_BACK; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + + return result; +} + + // ggml_mul_mat struct ggml_tensor * ggml_mul_mat( @@ -5661,13 +5185,10 @@ struct ggml_tensor * ggml_scale_impl( bool is_node = false; if (!inplace && (a->grad || b->grad)) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } - // TODO: when implement backward, fix this: - //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - struct ggml_tensor * result = ggml_view_tensor(ctx, a); + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); result->op = GGML_OP_SCALE; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; @@ -5691,6 +5212,106 @@ struct ggml_tensor * ggml_scale_inplace( return ggml_scale_impl(ctx, a, b, true); } +// ggml_set + +struct ggml_tensor * ggml_set_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset, + bool inplace) { + GGML_ASSERT(ggml_nelements(a) >= ggml_nelements(b)); + + bool is_node = false; + + if (!inplace && (a->grad || b->grad)) { + is_node = true; + } + + // make a view of the destination + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + ggml_scratch_save(ctx); + + struct ggml_tensor * c = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 5); + + (( int32_t * ) c->data)[0] = nb1; + (( int32_t * ) c->data)[1] = nb2; + (( int32_t * ) c->data)[2] = nb3; + (( int32_t * ) c->data)[3] = offset; + (( int32_t * ) c->data)[4] = inplace ? 1 : 0; + + ggml_scratch_load(ctx); + + result->op = GGML_OP_SET; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + result->opt[0] = c; + + return result; +} + +struct ggml_tensor * ggml_set( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset) { + return ggml_set_impl(ctx, a, b, nb1, nb2, nb3, offset, false); +} + +struct ggml_tensor * ggml_set_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset) { + return ggml_set_impl(ctx, a, b, nb1, nb2, nb3, offset, true); +} + +struct ggml_tensor * ggml_set_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t offset) { + return ggml_set_impl(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], offset, false); +} + +struct ggml_tensor * ggml_set_1d_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t offset) { + return ggml_set_impl(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], offset, true); +} + +struct ggml_tensor * ggml_set_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t offset) { + return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, false); +} + +struct ggml_tensor * ggml_set_2d_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t offset) { + return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, false); +} + + // ggml_cpy struct ggml_tensor * ggml_cpy_impl( @@ -5703,7 +5324,6 @@ struct ggml_tensor * ggml_cpy_impl( bool is_node = false; if (!inplace && (a->grad || b->grad)) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5741,7 +5361,6 @@ struct ggml_tensor * ggml_cont_impl( bool is_node = false; if (!inplace && a->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5779,11 +5398,15 @@ struct ggml_tensor * ggml_reshape( bool is_node = false; - if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward + if (a->grad) { is_node = true; } + if (b->grad) { + // gradient propagation is not supported + //GGML_ASSERT(false); + } + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a->data); result->op = GGML_OP_RESHAPE; @@ -5794,6 +5417,30 @@ struct ggml_tensor * ggml_reshape( return result; } +struct ggml_tensor * ggml_reshape_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0) { + GGML_ASSERT(ggml_is_contiguous(a)); + GGML_ASSERT(ggml_nelements(a) == ne0); + + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + const int64_t ne[1] = { ne0 }; + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a->data); + + result->op = GGML_OP_RESHAPE; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + return result; +} + struct ggml_tensor * ggml_reshape_2d( struct ggml_context * ctx, struct ggml_tensor * a, @@ -5805,7 +5452,6 @@ struct ggml_tensor * ggml_reshape_2d( bool is_node = false; if (a->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5832,7 +5478,6 @@ struct ggml_tensor * ggml_reshape_3d( bool is_node = false; if (a->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5847,6 +5492,34 @@ struct ggml_tensor * ggml_reshape_3d( return result; } + +struct ggml_tensor * ggml_reshape_4d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3) { + GGML_ASSERT(ggml_is_contiguous(a)); + GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2*ne3); + + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + const int64_t ne[4] = { ne0, ne1, ne2, ne3 }; + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a->data); + + result->op = GGML_OP_RESHAPE; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + return result; +} + // ggml_view_1d struct ggml_tensor * ggml_view_1d( @@ -5854,16 +5527,23 @@ struct ggml_tensor * ggml_view_1d( struct ggml_tensor * a, int64_t ne0, size_t offset) { + + bool is_node = false; + if (a->grad) { - GGML_ASSERT(false); // gradient propagation is not supported + is_node = true; } struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, &ne0, (char *) a->data + offset); result->op = GGML_OP_VIEW; - result->grad = NULL; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src0 = a; - result->src1 = NULL; // TODO: maybe store the offset here? + result->src1 = NULL; + + if (is_node) { + memcpy(result->padding, &offset, sizeof(offset)); + } return result; } @@ -5877,8 +5557,11 @@ struct ggml_tensor * ggml_view_2d( int64_t ne1, size_t nb1, size_t offset) { + + bool is_node = false; + if (a->grad) { - GGML_ASSERT(false); // gradient propagation is not supported + is_node = true; } const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, 1, 1 }; @@ -5890,9 +5573,13 @@ struct ggml_tensor * ggml_view_2d( result->nb[3] = result->nb[2]; result->op = GGML_OP_VIEW; - result->grad = NULL; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src0 = a; - result->src1 = NULL; // TODO: maybe store the offset here? + result->src1 = NULL; + + if (is_node) { + memcpy(result->padding, &offset, sizeof(offset)); + } return result; } @@ -5908,8 +5595,11 @@ struct ggml_tensor * ggml_view_3d( size_t nb1, size_t nb2, size_t offset) { + + bool is_node = false; + if (a->grad) { - GGML_ASSERT(false); // gradient propagation is not supported + is_node = true; } const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, 1 }; @@ -5921,9 +5611,53 @@ struct ggml_tensor * ggml_view_3d( result->nb[3] = result->nb[2]*ne2; result->op = GGML_OP_VIEW; - result->grad = NULL; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src0 = a; - result->src1 = NULL; // TODO: maybe store the offset here? + result->src1 = NULL; + + if (is_node) { + memcpy(result->padding, &offset, sizeof(offset)); + } + + return result; +} + +// ggml_view_4d + +struct ggml_tensor * ggml_view_4d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset) { + + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, ne3 }; + + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, (char *) a->data + offset); + + result->nb[1] = nb1; + result->nb[2] = nb2; + result->nb[3] = nb3; + + result->op = GGML_OP_VIEW; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + if (is_node) { + memcpy(result->padding, &offset, sizeof(offset)); + } return result; } @@ -5952,7 +5686,6 @@ struct ggml_tensor * ggml_permute( bool is_node = false; if (a->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -5984,7 +5717,14 @@ struct ggml_tensor * ggml_permute( result->op = GGML_OP_PERMUTE; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src0 = a; - result->src1 = NULL; // TODO: maybe store the permutation here? + result->src1 = NULL; + + if (is_node) { + result->padding[0] = axis0; + result->padding[1] = axis1; + result->padding[2] = axis2; + result->padding[3] = axis3; + } return result; } @@ -5997,7 +5737,6 @@ struct ggml_tensor * ggml_transpose( bool is_node = false; if (a->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -6028,7 +5767,6 @@ struct ggml_tensor * ggml_get_rows( bool is_node = false; if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } @@ -6044,24 +5782,82 @@ struct ggml_tensor * ggml_get_rows( return result; } -// ggml_diag_mask_inf +// ggml_get_rows_back -struct ggml_tensor * ggml_diag_mask_inf( +struct ggml_tensor * ggml_get_rows_back( struct ggml_context * ctx, struct ggml_tensor * a, - int n_past) { + struct ggml_tensor * b, + struct ggml_tensor * c) { + GGML_ASSERT(ggml_is_matrix(a) && ggml_is_vector(b) && b->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_is_matrix(c) && (a->ne[0] == c->ne[0])); + bool is_node = false; - if (a->grad) { - GGML_ASSERT(false); // TODO: implement backward + if (a->grad || b->grad) { is_node = true; } - // TODO: when implement backward, fix this: - //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - struct ggml_tensor * result = ggml_view_tensor(ctx, a); - struct ggml_tensor * b = ggml_new_i32(ctx, n_past); - ggml_set_name(b, "n_past"); + // TODO: implement non F32 return + //struct ggml_tensor * result = ggml_new_tensor_2d(ctx, a->type, a->ne[0], b->ne[0]); + struct ggml_tensor * result = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, c->ne[0], c->ne[1]); + + result->op = GGML_OP_GET_ROWS_BACK; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + result->opt[0] = c; + + return result; +} + +// ggml_diag + +struct ggml_tensor * ggml_diag( + struct ggml_context * ctx, + struct ggml_tensor * a) { + GGML_ASSERT(a->ne[1] == 1); + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + const int64_t ne[4] = { a->ne[0], a->ne[0], a->ne[2], a->ne[3] }; + struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, MAX(a->n_dims, 2), ne); + + result->op = GGML_OP_DIAG; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + return result; +} + + +// ggml_diag_mask_inf + +struct ggml_tensor * ggml_diag_mask_inf_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + bool inplace) { + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + ggml_scratch_save(ctx); + + struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2); + + ((int32_t *) b->data)[0] = n_past; + ((int32_t *) b->data)[1] = inplace ? 1 : 0; + + ggml_scratch_load(ctx); result->op = GGML_OP_DIAG_MASK_INF; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; @@ -6071,21 +5867,81 @@ struct ggml_tensor * ggml_diag_mask_inf( return result; } -// ggml_soft_max - -struct ggml_tensor * ggml_soft_max( +struct ggml_tensor * ggml_diag_mask_inf( struct ggml_context * ctx, - struct ggml_tensor * a) { + struct ggml_tensor * a, + int n_past) { + return ggml_diag_mask_inf_impl(ctx, a, n_past, false); +} + + +struct ggml_tensor * ggml_diag_mask_inf_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past) { + return ggml_diag_mask_inf_impl(ctx, a, n_past, true); +} + +// ggml_diag_mask_zero + +struct ggml_tensor * ggml_diag_mask_zero_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + bool inplace) { bool is_node = false; if (a->grad) { - GGML_ASSERT(false); // TODO: implement backward is_node = true; } - // TODO: when implement backward, fix this: - //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - struct ggml_tensor * result = ggml_view_tensor(ctx, a); + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + ggml_scratch_save(ctx); + + struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2); + ggml_set_name(b, "n_past, inplace"); + + ((int32_t *) b->data)[0] = n_past; + ((int32_t *) b->data)[1] = inplace ? 1 : 0; + + ggml_scratch_load(ctx); + + result->op = GGML_OP_DIAG_MASK_ZERO; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + + return result; +} + +struct ggml_tensor * ggml_diag_mask_zero( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past) { + return ggml_diag_mask_zero_impl(ctx, a, n_past, false); +} + +struct ggml_tensor * ggml_diag_mask_zero_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past) { + return ggml_diag_mask_zero_impl(ctx, a, n_past, true); +} + +// ggml_soft_max + +struct ggml_tensor * ggml_soft_max_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + bool inplace) { + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); result->op = GGML_OP_SOFT_MAX; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; @@ -6095,14 +5951,80 @@ struct ggml_tensor * ggml_soft_max( return result; } +struct ggml_tensor * ggml_soft_max( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_soft_max_impl(ctx, a, false); +} + +struct ggml_tensor * ggml_soft_max_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_soft_max_impl(ctx, a, true); +} + // ggml_rope +struct ggml_tensor * ggml_rope_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode, + bool inplace) { + GGML_ASSERT(n_past >= 0); + bool is_node = false; + + if (!inplace && a->grad) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + ggml_scratch_save(ctx); + + struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 3); + + ((int32_t *) b->data)[0] = n_past; + ((int32_t *) b->data)[1] = n_dims; + ((int32_t *) b->data)[2] = mode; + + ggml_scratch_load(ctx); + + result->op = GGML_OP_ROPE; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + + return result; +} + struct ggml_tensor * ggml_rope( struct ggml_context * ctx, struct ggml_tensor * a, int n_past, int n_dims, int mode) { + return ggml_rope_impl(ctx, a, n_past, n_dims, mode, false); +} + +struct ggml_tensor * ggml_rope_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode) { + return ggml_rope_impl(ctx, a, n_past, n_dims, mode, true); +} + +// ggml_rope_back + +struct ggml_tensor * ggml_rope_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode) { GGML_ASSERT(n_past >= 0); bool is_node = false; @@ -6111,17 +6033,20 @@ struct ggml_tensor * ggml_rope( is_node = true; } - // TODO: when implement backward, fix this: - //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - struct ggml_tensor * result = ggml_view_tensor(ctx, a); + struct ggml_tensor * result = ggml_dup_tensor(ctx, a); + + ggml_scratch_save(ctx); struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 3); + ggml_set_name(b, "n_past, n_dims, mode"); + ((int32_t *) b->data)[0] = n_past; ((int32_t *) b->data)[1] = n_dims; ((int32_t *) b->data)[2] = mode; - ggml_set_name(b, "n_past, n_dims, mode"); - result->op = GGML_OP_ROPE; + ggml_scratch_load(ctx); + + result->op = GGML_OP_ROPE_BACK; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src0 = a; result->src1 = b; @@ -6148,10 +6073,15 @@ struct ggml_tensor * ggml_alibi( //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); struct ggml_tensor * result = ggml_view_tensor(ctx, a); + ggml_scratch_save(ctx); + struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2); + ((int32_t *) b->data)[0] = n_past; ((int32_t *) b->data)[1] = n_head; + ggml_scratch_load(ctx); + result->op = GGML_OP_ALIBI; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src0 = a; @@ -6375,6 +6305,38 @@ void ggml_set_param( // ggml_compute_forward_dup +static void ggml_compute_forward_dup_same_cont( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + GGML_ASSERT(src0->type == dst->type); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const size_t nb00 = src0->nb[0]; + const size_t nb0 = dst->nb[0]; + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + // parallelize by elements + const int ne = ggml_nelements(dst); + const int dr = (ne + nth - 1) / nth; + const int ie0 = dr * ith; + const int ie1 = MIN(ie0 + dr, ne); + + if (ie0 < ie1) { + memcpy( + ((char *) dst->data + ie0*nb0), + ((char *) src0->data + ie0*nb00), + (ie1 - ie0) * GGML_TYPE_SIZE[src0->type]); + } + +} static void ggml_compute_forward_dup_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, @@ -6409,17 +6371,7 @@ static void ggml_compute_forward_dup_f16( const int nth = params->nth; // number of threads if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) { - // parallelize by elements - const int ne = ggml_nelements(dst); - const int dr = (ne + nth - 1) / nth; - const int ie0 = dr * ith; - const int ie1 = MIN(ie0 + dr, ne); - - memcpy( - ((char *) dst->data + ie0*nb0), - ((char *) src0->data + ie0*nb00), - (ie1 - ie0) * GGML_TYPE_SIZE[src0->type]); - + ggml_compute_forward_dup_same_cont(params, src0, dst); return; } @@ -6708,17 +6660,7 @@ static void ggml_compute_forward_dup_f32( const int nth = params->nth; // number of threads if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) { - // parallelize by elements - const int ne = ggml_nelements(dst); - const int dr = (ne + nth - 1) / nth; - const int ie0 = dr * ith; - const int ie1 = MIN(ie0 + dr, ne); - - memcpy( - ((char *) dst->data + ie0*nb0), - ((char *) src0->data + ie0*nb00), - (ie1 - ie0) * GGML_TYPE_SIZE[src0->type]); - + ggml_compute_forward_dup_same_cont(params, src0, dst); return; } @@ -6973,6 +6915,10 @@ static void ggml_compute_forward_dup( const struct ggml_compute_params * params, const struct ggml_tensor * src0, struct ggml_tensor * dst) { + if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) { + ggml_compute_forward_dup_same_cont(params, src0, dst); + return; + } switch (src0->type) { case GGML_TYPE_F16: { @@ -7005,44 +6951,73 @@ static void ggml_compute_forward_add_f32( const int ith = params->ith; const int nth = params->nth; - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; const size_t nb00 = src0->nb[0]; const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; const size_t nb10 = src1->nb[0]; const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; const size_t nb0 = dst->nb[0]; const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; GGML_ASSERT( nb0 == sizeof(float)); GGML_ASSERT(nb00 == sizeof(float)); + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + if (nb10 == sizeof(float)) { - for (int j = ith; j < n; j += nth) { + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + #ifdef GGML_USE_ACCELERATE vDSP_vadd( - (float *) ((char *) src0->data + j*nb01), 1, - (float *) ((char *) src1->data + j*nb11), 1, - (float *) ((char *) dst->data + j*nb1), 1, nc); + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, + ne0); #else - ggml_vec_add_f32(nc, - (float *) ((char *) dst->data + j*nb1), - (float *) ((char *) src0->data + j*nb01), - (float *) ((char *) src1->data + j*nb11)); + ggml_vec_add_f32(ne0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); #endif + // } + // } } } else { // src1 is not contiguous - for (int j = ith; j < n; j += nth) { - float * dst_ptr = (float *) ((char *) dst->data + j*nb1); - float * src0_ptr = (float *) ((char *) src0->data + j*nb01); - for (int i = 0; i < nc; i++) { - float * src1_ptr = (float *) ((char *) src1->data + j*nb11 + i*nb10); + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - dst_ptr[i] = src0_ptr[i] + *src1_ptr; + float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i0 = 0; i0 < ne0; i0++) { + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10); + + dst_ptr[i0] = src0_ptr[i0] + *src1_ptr; } } } @@ -7062,17 +7037,25 @@ static void ggml_compute_forward_add_f16_f32( const int ith = params->ith; const int nth = params->nth; - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; const size_t nb00 = src0->nb[0]; const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; const size_t nb10 = src1->nb[0]; const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; const size_t nb0 = dst->nb[0]; const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F32); @@ -7081,13 +7064,26 @@ static void ggml_compute_forward_add_f16_f32( GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + if (nb10 == sizeof(float)) { - for (int j = ith; j < n; j += nth) { - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + j*nb1); - ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01); - for (int i = 0; i < nc; i++) { - float * src1_ptr = (float *) ((char *) src1->data + j*nb11 + i*nb10); - dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + *src1_ptr); + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]); } } } @@ -7111,32 +7107,53 @@ static void ggml_compute_forward_add_f16_f16( const int ith = params->ith; const int nth = params->nth; - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; const size_t nb00 = src0->nb[0]; const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; const size_t nb10 = src1->nb[0]; const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; const size_t nb0 = dst->nb[0]; const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F16); - GGML_ASSERT(dst->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F16); GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + if (nb10 == sizeof(ggml_fp16_t)) { - for (int j = ith; j < n; j += nth) { - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + j*nb1); - ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01); - for (int i = 0; i < nc; i++) { - ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + j*nb11 + i*nb10); - dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + GGML_FP16_TO_FP32(*src1_ptr)); + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + GGML_FP16_TO_FP32(src1_ptr[i])); } } } @@ -7157,50 +7174,36 @@ static void ggml_compute_forward_add_q_f32( return; } + const int nr = ggml_nrows(src0); const int64_t ne00 = src0->ne[0]; const int64_t ne01 = src0->ne[1]; const int64_t ne02 = src0->ne[2]; - const int64_t ne03 = src0->ne[3]; + //const int64_t ne03 = src0->ne[3]; - //const int64_t ne10 = src1->ne[0]; - //const int64_t ne11 = src1->ne[1]; - const int64_t ne12 = src1->ne[2]; - const int64_t ne13 = src1->ne[3]; + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; - //const int64_t ne0 = dst->ne[0]; - //const int64_t ne1 = dst->ne[1]; - const int64_t ne2 = dst->ne[2]; - const int64_t ne3 = dst->ne[3]; + const size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; - const int nb00 = src0->nb[0]; - const int nb01 = src0->nb[1]; - const int nb02 = src0->nb[2]; - const int nb03 = src0->nb[3]; - - const int nb10 = src1->nb[0]; - const int nb11 = src1->nb[1]; - const int nb12 = src1->nb[2]; - const int nb13 = src1->nb[3]; - - const int nb0 = dst->nb[0]; - const int nb1 = dst->nb[1]; - const int nb2 = dst->nb[2]; - const int nb3 = dst->nb[3]; + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; const int ith = params->ith; const int nth = params->nth; - GGML_ASSERT(ne02 == ne12); - GGML_ASSERT(ne03 == ne13); - GGML_ASSERT(ne2 == ne12); - GGML_ASSERT(ne3 == ne13); - const enum ggml_type type = src0->type; dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q; quantize_row_q_t const quantize_row_q = quantize_fns[type].quantize_row_q; // we don't support permuted src0 or src1 - GGML_ASSERT(nb00 == (int) GGML_TYPE_SIZE[type]); + GGML_ASSERT(nb00 == GGML_TYPE_SIZE[type]); GGML_ASSERT(nb10 == sizeof(float)); // dst cannot be transposed or permuted @@ -7212,9 +7215,6 @@ static void ggml_compute_forward_add_q_f32( GGML_ASSERT(dst->type == src0->type); GGML_ASSERT(src1->type == GGML_TYPE_F32); - // total rows in src0 - const int nr = ne01*ne02*ne03; - // rows per thread const int dr = (nr + nth - 1)/nth; @@ -7278,7 +7278,6 @@ static void ggml_compute_forward_add( } break; case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: - case GGML_TYPE_Q4_2: case GGML_TYPE_Q5_0: case GGML_TYPE_Q5_1: case GGML_TYPE_Q8_0: @@ -7292,6 +7291,428 @@ static void ggml_compute_forward_add( } } +// ggml_compute_forward_add1 + +static void ggml_compute_forward_add1_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + +#ifdef GGML_USE_ACCELERATE + UNUSED(ggml_vec_add1_f32); + + vDSP_vadd( + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, + (float *) ((char *) src1->data), 0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, + ne0); +#else + ggml_vec_add1_f32(ne0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), + *(float *) src1->data); +#endif + } +} + +static void ggml_compute_forward_add1_f16_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + // scalar to add + const float v = *(float *) src1->data; + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F16); + + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); + } + } +} + +static void ggml_compute_forward_add1_f16_f16( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + // scalar to add + const float v = GGML_FP16_TO_FP32(*(ggml_fp16_t *) src1->data); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F16); + + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); + } + } +} + +static void ggml_compute_forward_add1_q_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + // scalar to add + const float v = *(float *) src1->data; + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + const enum ggml_type type = src0->type; + dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q; + quantize_row_q_t const quantize_row_q = quantize_fns[type].quantize_row_q; + + // we don't support permuted src0 + GGML_ASSERT(nb00 == GGML_TYPE_SIZE[type]); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + GGML_ASSERT(ggml_is_quantized(src0->type)); + GGML_ASSERT(dst->type == src0->type); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float * wdata = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32) * ith; + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + void * src0_row = (void *) ((char *) src0->data + (i1*nb01 + i2*nb02 + i3*nb03)); + void * dst_row = (void *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb0 )); + + assert(ne0 % 32 == 0); + + // unquantize row from src0 to temp buffer + dequantize_row_q(src0_row, wdata, ne0); + // add src1 + ggml_vec_acc1_f32(ne0, wdata, v); + // quantize row to dst + quantize_row_q(wdata, dst_row, ne0); + } +} + +static void ggml_compute_forward_add1( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_add1_f32(params, src0, src1, dst); + } break; + case GGML_TYPE_F16: + { + if (src1->type == GGML_TYPE_F16) { + ggml_compute_forward_add1_f16_f16(params, src0, src1, dst); + } + else if (src1->type == GGML_TYPE_F32) { + ggml_compute_forward_add1_f16_f32(params, src0, src1, dst); + } + else { + GGML_ASSERT(false); + } + } break; + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + { + ggml_compute_forward_add1_q_f32(params, src0, src1, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + + +// ggml_compute_forward_acc + +static void ggml_compute_forward_acc_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + + GGML_ASSERT(opt0->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_nelements(opt0) == 5); + + // view src0 and dst with these strides and data offset inbytes during acc + // nb0 is implicitely element_size because src0 and dst are contiguous + size_t nb1 = ((int32_t *) opt0->data)[0]; + size_t nb2 = ((int32_t *) opt0->data)[1]; + size_t nb3 = ((int32_t *) opt0->data)[2]; + size_t offset = ((int32_t *) opt0->data)[3]; + bool inplace = (bool) ((int32_t *) opt0->data)[4]; + + if (!inplace && (params->type == GGML_TASK_INIT)) { + // memcpy needs to be synchronized across threads to avoid race conditions. + // => do it in INIT phase + memcpy( + ((char *) dst->data), + ((char *) src0->data), + ggml_nbytes(dst)); + } + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src1); + const int nc = src1->ne[0]; + + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + const int64_t ne13 = src1->ne[3]; + + const size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; + + // src0 and dst as viewed during acc + const size_t nb0 = ggml_element_size(src0); + + const size_t nb00 = nb0; + const size_t nb01 = nb1; + const size_t nb02 = nb2; + const size_t nb03 = nb3; + + GGML_ASSERT(offset + (ne10 == 0 ? 0 : ne10-1)*nb0 + (ne11 == 0 ? 0 : ne11-1)*nb1 + (ne12 == 0 ? 0 : ne12-1)*nb2 + (ne13 == 0 ? 0 : ne13-1)*nb3 < ggml_nbytes(dst)); + GGML_ASSERT(offset + (ne10 == 0 ? 0 : ne10-1)*nb00 + (ne11 == 0 ? 0 : ne11-1)*nb01 + (ne12 == 0 ? 0 : ne12-1)*nb02 + (ne13 == 0 ? 0 : ne13-1)*nb03 < ggml_nbytes(src0)); + + GGML_ASSERT(nb10 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are viewed with shape of src1 and offset + // => same indices + const int i3 = ir/(ne12*ne11); + const int i2 = (ir - i3*ne12*ne11)/ne11; + const int i1 = (ir - i3*ne12*ne11 - i2*ne11); + +#ifdef GGML_USE_ACCELERATE + vDSP_vadd( + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + offset), 1, + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), 1, nc); +#else + ggml_vec_add_f32(nc, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + offset), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); +#endif + } +} + +static void ggml_compute_forward_acc( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_acc_f32(params, src0, src1, opt0, dst); + } break; + case GGML_TYPE_F16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_sub static void ggml_compute_forward_sub_f32( @@ -7306,18 +7727,68 @@ static void ggml_compute_forward_sub_f32( return; } - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; - assert( dst->nb[0] == sizeof(float)); - assert(src0->nb[0] == sizeof(float)); - assert(src1->nb[0] == sizeof(float)); + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; - for (int i = 0; i < n; i++) { - ggml_vec_sub_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1])), - (float *) ((char *) src1->data + i*(src1->nb[1]))); + const size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + if (nb10 == sizeof(float)) { + for (int ir = 0; ir < nr; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + +#ifdef GGML_USE_ACCELERATE + vDSP_vsub( + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, + ne0); +#else + ggml_vec_sub_f32(ne0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); +#endif + // } + // } + } + } else { + // src1 is not contiguous + for (int ir = 0; ir < nr; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i0 = 0; i0 < ne0; i0++) { + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10); + + dst_ptr[i0] = src0_ptr[i0] - *src1_ptr; + } + } } } @@ -7345,25 +7816,78 @@ static void ggml_compute_forward_mul_f32( const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { - assert(params->ith == 0); assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } + const int ith = params->ith; + const int nth = params->nth; - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; - assert( dst->nb[0] == sizeof(float)); - assert(src0->nb[0] == sizeof(float)); - assert(src1->nb[0] == sizeof(float)); + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; - for (int i = 0; i < n; i++) { - ggml_vec_mul_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1])), - (float *) ((char *) src1->data + i*(src1->nb[1]))); + const size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + if (nb10 == sizeof(float)) { + for (int ir = ith; ir < nr; ir += nth) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + +#ifdef GGML_USE_ACCELERATE + UNUSED(ggml_vec_mul_f32); + + vDSP_vmul( + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, + ne0); +#else + ggml_vec_mul_f32(ne0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); +#endif + // } + // } + } + } else { + // src1 is not contiguous + for (int ir = ith; ir < nr; ir += nth) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i0 = 0; i0 < ne0; i0++) { + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10); + + dst_ptr[i0] = src0_ptr[i0] * (*src1_ptr); + } + } } } @@ -7398,18 +7922,68 @@ static void ggml_compute_forward_div_f32( return; } - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + const int64_t ne0 = src0->ne[0]; + const int64_t ne1 = src0->ne[1]; + const int64_t ne2 = src0->ne[2]; - assert( dst->nb[0] == sizeof(float)); - assert(src0->nb[0] == sizeof(float)); - assert(src1->nb[0] == sizeof(float)); + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; - for (int i = 0; i < n; i++) { - ggml_vec_div_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1])), - (float *) ((char *) src1->data + i*(src1->nb[1]))); + const size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + if (nb10 == sizeof(float)) { + for (int ir = 0; ir < nr; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + +#ifdef GGML_USE_ACCELERATE + vDSP_vdiv( + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, + ne0); +#else + ggml_vec_div_f32(ne0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); +#endif + // } + // } + } + } else { + // src1 is not contiguous + for (int ir = 0; ir < nr; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i0 = 0; i0 < ne0; i0++) { + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10); + + dst_ptr[i0] = src0_ptr[i0] / (*src1_ptr); + } + } } } @@ -7514,6 +8088,49 @@ static void ggml_compute_forward_sqrt( } } + +// ggml_compute_forward_log + +static void ggml_compute_forward_log_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + GGML_ASSERT(params->ith == 0); + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + GGML_ASSERT( dst->nb[0] == sizeof(float)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_log_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_log( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_log_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_sum static void ggml_compute_forward_sum_f32( @@ -7571,6 +8188,73 @@ static void ggml_compute_forward_sum( } } +// ggml_compute_forward_sum_rows + +static void ggml_compute_forward_sum_rows_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + GGML_ASSERT(params->ith == 0); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(dst->nb[0] == sizeof(float)); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + const int64_t ne03 = src0->ne[3]; + + const int64_t ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t ne3 = dst->ne[3]; + + GGML_ASSERT(ne0 == 1); + GGML_ASSERT(ne1 == ne01); + GGML_ASSERT(ne2 == ne02); + GGML_ASSERT(ne3 == ne03); + + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + for (int64_t i3 = 0; i3 < ne03; i3++) { + for (int64_t i2 = 0; i2 < ne02; i2++) { + for (int64_t i1 = 0; i1 < ne01; i1++) { + float* src_row = (float *) ((char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03); + float* dst_row = (float *) ((char *) dst->data + i1*nb1 + i2*nb2 + i3*nb3); + float row_sum = 0; + ggml_vec_sum_f32(ne00, &row_sum, src_row); + dst_row[0] = row_sum; + } + } + } +} + +static void ggml_compute_forward_sum_rows( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sum_rows_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_mean static void ggml_compute_forward_mean_f32( @@ -7648,37 +8332,58 @@ static void ggml_compute_forward_repeat_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, struct ggml_tensor * dst) { - assert(params->ith == 0); - assert(ggml_can_repeat(src0, dst)); + GGML_ASSERT(params->ith == 0); + GGML_ASSERT(ggml_can_repeat(src0, dst)); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - // TODO: implement support for rank > 2 tensors - assert(src0->ne[2] == 1); - assert(src0->ne[3] == 1); - assert( dst->ne[2] == 1); - assert( dst->ne[3] == 1); + const int64_t ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t ne3 = dst->ne[3]; - const int nc = dst->ne[0]; - const int nr = dst->ne[1]; - const int nc0 = src0->ne[0]; - const int nr0 = src0->ne[1]; - const int ncr = nc/nc0; // guaranteed to be an integer due to the check in ggml_can_repeat - const int nrr = nr/nr0; // guaranteed to be an integer due to the check in ggml_can_repeat + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + const int64_t ne03 = src0->ne[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + // guaranteed to be an integer due to the check in ggml_can_repeat + const int nr0 = (int)(ne0/ne00); + const int nr1 = (int)(ne1/ne01); + const int nr2 = (int)(ne2/ne02); + const int nr3 = (int)(ne3/ne03); // TODO: support for transposed / permuted tensors - assert( dst->nb[0] == sizeof(float)); - assert(src0->nb[0] == sizeof(float)); + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); // TODO: maybe this is not optimal? - for (int i = 0; i < nrr; i++) { - for (int j = 0; j < ncr; j++) { - for (int k = 0; k < nr0; k++) { - ggml_vec_cpy_f32(nc0, - (float *) ((char *) dst->data + (i*nr0 + k)*( dst->nb[1]) + j*nc0*( dst->nb[0])), - (float *) ((char *) src0->data + ( k)*(src0->nb[1]))); + for (int i3 = 0; i3 < nr3; i3++) { + for (int k3 = 0; k3 < ne03; k3++) { + for (int i2 = 0; i2 < nr2; i2++) { + for (int k2 = 0; k2 < ne02; k2++) { + for (int i1 = 0; i1 < nr1; i1++) { + for (int k1 = 0; k1 < ne01; k1++) { + for (int i0 = 0; i0 < nr0; i0++) { + ggml_vec_cpy_f32(ne00, + (float *) ((char *) dst->data + (i3*ne03 + k3)*nb3 + (i2*ne02 + k2)*nb2 + (i1*ne01 + k1)*nb1 + (i0*ne00)*nb0), + (float *) ((char *) src0->data + ( k3)*nb03 + ( k2)*nb02 + ( k1)*nb01)); + } + } + } + } } } } @@ -8031,6 +8736,70 @@ static void ggml_compute_forward_silu( } +// ggml_compute_forward_silu_back + +static void ggml_compute_forward_silu_back_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * grad, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_is_contiguous(grad)); + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_are_same_shape(src0, grad)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_silu_backward_f32(nc, + (float *) ((char *) dst->data + i1*( dst->nb[1])), + (float *) ((char *) src0->data + i1*(src0->nb[1])), + (float *) ((char *) grad->data + i1*(grad->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k]; + UNUSED(x); + assert(!isnan(x)); + assert(!isinf(x)); + } +#endif + } +} + +static void ggml_compute_forward_silu_back( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * grad, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_silu_back_f32(params, src0, grad, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_norm static void ggml_compute_forward_norm_f32( @@ -8185,6 +8954,195 @@ static void ggml_compute_forward_rms_norm( } +static void ggml_compute_forward_rms_norm_back_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst) && ggml_are_same_shape(src0, src1)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + const int64_t ne03 = src0->ne[3]; + + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; + + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + const float eps = 1e-6f; // TODO: make this a parameter + + // TODO: optimize + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ith; i01 < ne01; i01 += nth) { + // src1 is same shape as src0 => same indices + const int64_t i11 = i01; + const int64_t i12 = i02; + const int64_t i13 = i03; + + const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + const float * dz = (float *) ((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13); + + ggml_float sum_xx = 0.0; + ggml_float sum_xdz = 0.0; + + for (int64_t i00 = 0; i00 < ne00; i00++) { + sum_xx += (ggml_float)(x[i00] * x[i00]); + sum_xdz += (ggml_float)(x[i00] * dz[i00]); + } + + //const float mean = (float)(sum_xx)/ne00; + const float mean_eps = (float)(sum_xx)/ne00 + eps; + const float sum_eps = (float)(sum_xx) + eps*ne00; + //const float mean_xdz = (float)(sum_xdz)/ne00; + // we could cache rms from forward pass to improve performance. + // to do this implement ggml_rms and compose ggml_rms_norm using ggml_rms. + //const float rms = sqrtf(mean_eps); + const float rrms = 1.0f / sqrtf(mean_eps); + //const float scale = -rrms/(ne00 * mean_eps); // -1/(n*rms**3) + + { + // z = rms_norm(x) + // + // rms_norm(src0) = + // scale( + // src0, + // div( + // 1, + // sqrt( + // add( + // scale( + // sum( + // sqr( + // src0)), + // (1.0/N)), + // eps)))); + + // postorder: + // ## op args grad + // 00 param src0 grad[#00] + // 01 const 1 + // 02 sqr (#00) grad[#02] + // 03 sum (#02) grad[#03] + // 04 const 1/N + // 05 scale (#03, #04) grad[#05] + // 06 const eps + // 07 add (#05, #06) grad[#07] + // 08 sqrt (#07) grad[#08] + // 09 div (#01,#08) grad[#09] + // 10 scale (#00,#09) grad[#10] + // + // backward pass, given grad[#10] + // #10: scale + // grad[#00] += scale(grad[#10],#09) + // grad[#09] += sum(mul(grad[#10],#00)) + // #09: div + // grad[#08] += neg(mul(grad[#09], div(#09,#08))) + // #08: sqrt + // grad[#07] += mul(grad[#08], div(0.5, #08)) + // #07: add + // grad[#05] += grad[#07] + // #05: scale + // grad[#03] += scale(grad[#05],#04) + // #03: sum + // grad[#02] += repeat(grad[#03], #02) + // #02: + // grad[#00] += scale(mul(#00, grad[#02]), 2.0) + // + // substitute and simplify: + // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, grad[#02]), 2.0) + // grad[#02] = repeat(grad[#03], #02) + // grad[#02] = repeat(scale(grad[#05],#04), #02) + // grad[#02] = repeat(scale(grad[#07],#04), #02) + // grad[#02] = repeat(scale(mul(grad[#08], div(0.5, #08)),#04), #02) + // grad[#02] = repeat(scale(mul(neg(mul(grad[#09], div(#09,#08))), div(0.5, #08)),#04), #02) + // grad[#02] = repeat(scale(mul(neg(mul(sum(mul(grad[#10],#00)), div(#09,#08))), div(0.5, #08)),#04), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(#09,#08) * div(0.5, #08) * (1/N)), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(div(#01,#08),#08) * div(0.5, #08) * (1/N)), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(1,#08*#08) * div(0.5, #08) * (1/N)), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N)), #02) + // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, grad[#02]), 2.0) + // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, repeat(-(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N)), #02)), 2.0) + // grad[#00] = scale(grad(#10), #09) + scale(scale(#00, -(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N))), 2.0) + // grad[#00] = scale(grad(#10), #09) + scale(#00, -(sum(mul(grad[#10],#00)) * div(1,#07) * div(1,#08) * (1/N))) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,#07*#08) * (-1/N)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,#07*#08) * (-1/N)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,mean_eps*rms) * (-1/N)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*mean_eps)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*(sum_xx/N+eps))) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*sum_xx+rms*N*eps)) + // grad[#00] = scale(dz, rrms) + scale(x, sum(mul(dz,x)) * div(-1,rms*N*mean_eps)) + // grad[#00] = scale(dz, rrms) + scale(x, sum_xdz * div(-1,rms*N*mean_eps)) + // a = b*c + d*e + // a = b*c*f/f + d*e*f/f + // a = (b*c*f + d*e*f)*(1/f) + // a = (b*c*(1/c) + d*e*(1/c))*(1/(1/c)) + // a = (b + d*e/c)*c + // b = dz, c = rrms, d = x, e = sum_xdz * div(-1,rms*N*mean_eps) + // a = (dz + x*sum_xdz * div(-1,rms*N*mean_eps)/rrms)*rrms + // a = (dz + x*sum_xdz * div(-1,rms*N*mean_eps)*rms)*rrms + // a = (dz + x*sum_xdz * div(-rms,rms*N*mean_eps))*rrms + // a = (dz + x*sum_xdz * div(-1,N*mean_eps))*rrms + // a = (dz + x*div(-sum_xdz,N*mean_eps))*rrms + // a = (dz + x*div(-mean_xdz,mean_eps))*rrms + // grad[#00] = scale(dz + scale(x, div(-mean_xdz,mean_eps)),rrms) + // grad[#00] = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) + // dx = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) + } + // dx = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) + // post-order: + // dx := x + // dx := scale(dx,-mean_xdz/mean_eps) + // dx := add(dx, dz) + // dx := scale(dx, rrms) + float * dx = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3); + + ggml_vec_cpy_f32 (ne00, dx, x); + // ggml_vec_scale_f32(ne00, dx, -mean_xdz/mean_eps); + ggml_vec_scale_f32(ne00, dx, (float)(-sum_xdz)/sum_eps); + ggml_vec_acc_f32 (ne00, dx, dz); + ggml_vec_scale_f32(ne00, dx, rrms); + } + } + } +} + +static void ggml_compute_forward_rms_norm_back( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_rms_norm_back_f32(params, src0, src1, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + + // ggml_compute_forward_mul_mat #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) @@ -8833,7 +9791,6 @@ static void ggml_compute_forward_mul_mat( switch (src0->type) { case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: - case GGML_TYPE_Q4_2: case GGML_TYPE_Q5_0: case GGML_TYPE_Q5_1: case GGML_TYPE_Q8_0: @@ -8888,8 +9845,17 @@ static void ggml_compute_forward_scale_f32( const int ir0 = dr*ith; const int ir1 = MIN(ir0 + dr, nr); + const size_t nb01 = src0->nb[1]; + + const size_t nb1 = dst->nb[1]; + + for (int i1 = ir0; i1 < ir1; i1++) { - ggml_vec_scale_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), v); + if (dst->data != src0->data) { + // src0 is same shape as dst => same indices + memcpy((char *)dst->data + i1*nb1, (char *)src0->data + i1*nb01, nc * sizeof(float)); + } + ggml_vec_scale_f32(nc, (float *) ((char *) dst->data + i1*nb1), v); } } @@ -8910,6 +9876,115 @@ static void ggml_compute_forward_scale( } } +// ggml_compute_forward_set + +static void ggml_compute_forward_set_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + + GGML_ASSERT(opt0->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_nelements(opt0) == 5); + + // view src0 and dst with these strides and data offset inbytes during set + // nb0 is implicitely element_size because src0 and dst are contiguous + size_t nb1 = ((int32_t *) opt0->data)[0]; + size_t nb2 = ((int32_t *) opt0->data)[1]; + size_t nb3 = ((int32_t *) opt0->data)[2]; + size_t offset = ((int32_t *) opt0->data)[3]; + bool inplace = (bool) ((int32_t *) opt0->data)[4]; + + if (!inplace && (params->type == GGML_TASK_INIT)) { + // memcpy needs to be synchronized across threads to avoid race conditions. + // => do it in INIT phase + memcpy( + ((char *) dst->data), + ((char *) src0->data), + ggml_nbytes(dst)); + } + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src1); + const int nc = src1->ne[0]; + + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + const int64_t ne13 = src1->ne[3]; + + const size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + const size_t nb12 = src1->nb[2]; + const size_t nb13 = src1->nb[3]; + + // src0 and dst as viewed during set + const size_t nb0 = ggml_element_size(src0); + + const int im0 = (ne10 == 0 ? 0 : ne10-1); + const int im1 = (ne11 == 0 ? 0 : ne11-1); + const int im2 = (ne12 == 0 ? 0 : ne12-1); + const int im3 = (ne13 == 0 ? 0 : ne13-1); + + GGML_ASSERT(offset + im0*nb0 + im1*nb1 + im2*nb2 + im3*nb3 < ggml_nbytes(dst)); + + GGML_ASSERT(nb10 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are viewed with shape of src1 and offset + // => same indices + const int i3 = ir/(ne12*ne11); + const int i2 = (ir - i3*ne12*ne11)/ne11; + const int i1 = (ir - i3*ne12*ne11 - i2*ne11); + + ggml_vec_cpy_f32(nc, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); + } +} + +static void ggml_compute_forward_set( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_set_f32(params, src0, src1, opt0, dst); + } break; + case GGML_TYPE_F16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_cpy static void ggml_compute_forward_cpy( @@ -9064,7 +10139,6 @@ static void ggml_compute_forward_get_rows( switch (src0->type) { case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: - case GGML_TYPE_Q4_2: case GGML_TYPE_Q5_0: case GGML_TYPE_Q5_1: case GGML_TYPE_Q8_0: @@ -9105,22 +10179,226 @@ static void ggml_compute_forward_get_rows( //} } -// ggml_compute_forward_diag_mask_inf +// ggml_compute_forward_get_rows_back -static void ggml_compute_forward_diag_mask_inf_f32( +static void ggml_compute_forward_get_rows_back_f32_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - assert(params->ith == 0); - assert(src1->type == GGML_TYPE_I32); - assert(ggml_nelements(src1) == 1); + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + GGML_ASSERT(params->ith == 0); + GGML_ASSERT(ggml_are_same_shape(opt0, dst)); + GGML_ASSERT(ggml_is_contiguous(opt0)); + GGML_ASSERT(ggml_is_contiguous(dst)); + + ggml_compute_forward_dup_same_cont(params, opt0, dst); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - const int n_past = ((int32_t *) src1->data)[0]; + const int nc = src0->ne[0]; + const int nr = ggml_nelements(src1); + + GGML_ASSERT( dst->ne[0] == nc); + GGML_ASSERT(src0->nb[0] == sizeof(ggml_fp16_t)); + + for (int i = 0; i < nr; ++i) { + const int r = ((int32_t *) src1->data)[i]; + + for (int j = 0; j < nc; ++j) { + ggml_fp16_t v = ((ggml_fp16_t *) ((char *) src0->data + i*src0->nb[1]))[j]; + ((float *) ((char *) dst->data + r*dst->nb[1]))[j] += GGML_FP16_TO_FP32(v); + } + } +} + +static void ggml_compute_forward_get_rows_back_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + GGML_ASSERT(params->ith == 0); + GGML_ASSERT(ggml_are_same_shape(opt0, dst)); + GGML_ASSERT(ggml_is_contiguous(opt0)); + GGML_ASSERT(ggml_is_contiguous(dst)); + + ggml_compute_forward_dup_same_cont(params, opt0, dst); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int nc = src0->ne[0]; + const int nr = ggml_nelements(src1); + + GGML_ASSERT( dst->ne[0] == nc); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < nr; ++i) { + const int r = ((int32_t *) src1->data)[i]; + + ggml_vec_add_f32(nc, + (float *) ((char *) dst->data + r*dst->nb[1]), + (float *) ((char *) dst->data + r*dst->nb[1]), + (float *) ((char *) src0->data + i*src0->nb[1])); + } +} + + +static void ggml_compute_forward_get_rows_back( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * opt0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_get_rows_back_f32_f16(params, src0, src1, opt0, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_get_rows_back_f32(params, src0, src1, opt0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } + + //static bool first = true; + //printf("ne0 = %d, ne1 = %d, ne2 = %d\n", dst->ne[0], dst->ne[1], dst->ne[2]); + //if (first) { + // first = false; + //} else { + // for (int k = 0; k < dst->ne[1]; ++k) { + // for (int j = 0; j < dst->ne[0]/16; ++j) { + // for (int i = 0; i < 16; ++i) { + // printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]); + // } + // printf("\n"); + // } + // printf("\n"); + // } + // printf("\n"); + // exit(0); + //} +} + +// ggml_compute_forward_diag + +static void ggml_compute_forward_diag_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + GGML_ASSERT(params->ith == 0); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + // TODO: handle transposed/permuted matrices + + const int ne00 = src0->ne[0]; + const int ne01 = src0->ne[1]; + const int ne02 = src0->ne[2]; + const int ne03 = src0->ne[3]; + const int ne0 = dst->ne[0]; + const int ne1 = dst->ne[1]; + const int ne2 = dst->ne[2]; + const int ne3 = dst->ne[3]; + GGML_ASSERT(ne00 == ne0); + GGML_ASSERT(ne00 == ne1); + GGML_ASSERT(ne01 == 1); + GGML_ASSERT(ne02 == ne2); + GGML_ASSERT(ne03 == ne3); + + const int nb00 = src0->nb[0]; + //const int nb01 = src0->nb[1]; + const int nb02 = src0->nb[2]; + const int nb03 = src0->nb[3]; + const int nb0 = dst->nb[0]; + const int nb1 = dst->nb[1]; + const int nb2 = dst->nb[2]; + const int nb3 = dst->nb[3]; + + GGML_ASSERT(nb00 == sizeof(float)); + GGML_ASSERT(nb0 == sizeof(float)); + + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = 0; i2 < ne2; i2++) { + for (int i1 = 0; i1 < ne1; i1++) { + float * d = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + float * s = (float *)((char *) src0->data + i3*nb03 + i2*nb02); + for (int i0 = 0; i0 < i1; i0++) { + d[i0] = 0; + } + d[i1] = s[i1]; + for (int i0 = i1+1; i0 < ne0; i0++) { + d[i0] = 0; + } + } + } + } +} + +static void ggml_compute_forward_diag( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_diag_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + +// ggml_compute_forward_diag_mask_inf + +static void ggml_compute_forward_diag_mask_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst, + const float value) { + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 2); + + const int n_past = ((int32_t *) src1->data)[0]; + const bool inplace = (bool)((int32_t *) src1->data)[1]; + + if (params->type == GGML_TASK_INIT) { + // TODO: this hack is not good, need a better way to handle this + if (!inplace) { + // use the init task to copy src -> dst + struct ggml_compute_params params_cpy = *params; + + params_cpy.ith = 0; + params_cpy.nth = 1; + params_cpy.type = GGML_TASK_COMPUTE; + + ggml_compute_forward_dup_same_cont(¶ms_cpy, src0, dst); + } + + return; + } + + if (params->type == GGML_TASK_FINALIZE) { + return; + } + + const int ith = params->ith; + const int nth = params->nth; + + assert(n_past >= 0); // TODO: handle transposed/permuted matrices @@ -9133,10 +10411,10 @@ static void ggml_compute_forward_diag_mask_inf_f32( assert(src0->nb[0] == sizeof(float)); for (int k = 0; k < nz; k++) { - for (int j = 0; j < nr; j++) { + for (int j = ith; j < nr; j += nth) { for (int i = n_past; i < nc; i++) { if (i > n_past + j) { - *(float *)((char *) dst->data + k*dst->nb[2] + j*dst->nb[1] + i*dst->nb[0]) = -INFINITY; + *(float *)((char *) dst->data + k*dst->nb[2] + j*dst->nb[1] + i*dst->nb[0]) = value; } } } @@ -9151,7 +10429,24 @@ static void ggml_compute_forward_diag_mask_inf( switch (src0->type) { case GGML_TYPE_F32: { - ggml_compute_forward_diag_mask_inf_f32(params, src0, src1, dst); + ggml_compute_forward_diag_mask_f32(params, src0, src1, dst, -INFINITY); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + +static void ggml_compute_forward_diag_mask_zero( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_diag_mask_f32(params, src0, src1, dst, 0); } break; default: { @@ -9190,44 +10485,44 @@ static void ggml_compute_forward_soft_max_f32( const int ir1 = MIN(ir0 + dr, nr); for (int i1 = ir0; i1 < ir1; i1++) { - float *p = (float *)((char *) dst->data + i1*dst->nb[1]); + float *sp = (float *)((char *) src0->data + i1*src0->nb[1]); + float *dp = (float *)((char *) dst->data + i1*dst->nb[1]); #ifndef NDEBUG for (int i = 0; i < nc; ++i) { //printf("p[%d] = %f\n", i, p[i]); - assert(!isnan(p[i])); + assert(!isnan(sp[i])); } #endif float max = -INFINITY; - ggml_vec_max_f32(nc, &max, p); + ggml_vec_max_f32(nc, &max, sp); ggml_float sum = 0.0; uint16_t scvt; for (int i = 0; i < nc; i++) { - //printf("p[%3d] = %8.4f\n", i, p[i]); - if (p[i] == -INFINITY) { - p[i] = 0.0f; + if (sp[i] == -INFINITY) { + dp[i] = 0.0f; } else { - //const float val = (p[i] == -INFINITY) ? 0.0 : exp(p[i] - max); - ggml_fp16_t s = GGML_FP32_TO_FP16(p[i] - max); + // const float val = (sp[i] == -INFINITY) ? 0.0 : exp(sp[i] - max); + ggml_fp16_t s = GGML_FP32_TO_FP16(sp[i] - max); memcpy(&scvt, &s, sizeof(scvt)); const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]); sum += (ggml_float)val; - p[i] = val; + dp[i] = val; } } assert(sum > 0.0); sum = 1.0/sum; - ggml_vec_scale_f32(nc, p, sum); + ggml_vec_scale_f32(nc, dp, sum); #ifndef NDEBUG for (int i = 0; i < nc; ++i) { - assert(!isnan(p[i])); - assert(!isinf(p[i])); + assert(!isnan(dp[i])); + assert(!isinf(dp[i])); } #endif } @@ -9267,6 +10562,8 @@ static void ggml_compute_forward_alibi_f32( const int n_past = ((int32_t *) src1->data)[0]; const int n_head = ((int32_t *) src1->data)[1]; + assert(n_past >= 0); + const int ne0 = src0->ne[0]; // all_seq_len = n_past + ne1 const int ne1 = src0->ne[1]; // seq_len_without_past //const int ne2 = src0->ne[2]; // n_head -> this is k @@ -9305,7 +10602,7 @@ static void ggml_compute_forward_alibi_f32( m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1); } - pdst[0] = (j+1) * m_k + src[0]; + pdst[0] = i * m_k + src[0]; } } } @@ -9328,6 +10625,8 @@ static void ggml_compute_forward_alibi_f16( const int n_past = ((int32_t *) src1->data)[0]; const int n_head = ((int32_t *) src1->data)[1]; + assert(n_past >= 0); + const int ne0 = src0->ne[0]; // all_seq_len = n_past + ne1 const int ne1 = src0->ne[1]; // seq_len_without_past //const int ne2 = src0->ne[2]; // n_head -> this is k @@ -9367,7 +10666,7 @@ static void ggml_compute_forward_alibi_f16( } // we return F32 - pdst[0] = (j+1) * m_k + GGML_FP16_TO_FP32(src[0]); + pdst[0] = i * m_k + GGML_FP16_TO_FP32(src[0]); } } } @@ -9389,7 +10688,6 @@ static void ggml_compute_forward_alibi( } break; case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: - case GGML_TYPE_Q4_2: case GGML_TYPE_Q5_0: case GGML_TYPE_Q5_1: case GGML_TYPE_Q8_0: @@ -9411,8 +10709,8 @@ static void ggml_compute_forward_rope_f32( const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { - assert(src1->type == GGML_TYPE_I32); - assert(ggml_nelements(src1) == 3); + GGML_ASSERT(src1->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_nelements(src1) == 3); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; @@ -9422,25 +10720,35 @@ static void ggml_compute_forward_rope_f32( const int n_dims = ((int32_t *) src1->data)[1]; const int mode = ((int32_t *) src1->data)[2]; - //const int64_t ne0 = src0->ne[0]; - const int64_t ne1 = src0->ne[1]; - const int64_t ne2 = src0->ne[2]; - const int64_t ne3 = src0->ne[3]; + assert(n_past >= 0); - const int nb0 = src0->nb[0]; - const int nb1 = src0->nb[1]; - const int nb2 = src0->nb[2]; - const int nb3 = src0->nb[3]; + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const int64_t ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t ne3 = dst->ne[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); //printf("n_past = %d, ne2 = %d\n", n_past, ne2); - assert(nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); const int ith = params->ith; const int nth = params->nth; - const int nr = ggml_nrows(src0); + const int nr = ggml_nrows(dst); + + GGML_ASSERT(n_dims <= ne0); + GGML_ASSERT(n_dims % 2 == 0); // rows per thread const int dr = (nr + nth - 1)/nth; @@ -9458,37 +10766,50 @@ static void ggml_compute_forward_rope_f32( for (int64_t i3 = 0; i3 < ne3; i3++) { for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) { - const int p = ((mode & 1) == 0 ? n_past + i2 : i2); + const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2); for (int64_t i1 = 0; i1 < ne1; i1++) { if (ir++ < ir0) continue; if (ir > ir1) break; float theta = (float)p; - for (int i0 = 0; i0 < n_dims; i0 += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + if (!is_neox) { + for (int64_t i0 = 0; i0 < ne0; i0 += 2) { + const float cos_theta = cosf(theta); + const float sin_theta = sinf(theta); - theta *= theta_scale; + theta *= theta_scale; - if (!is_neox) { - const float * const src = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); const float x0 = src[0]; const float x1 = src[1]; dst_data[0] = x0*cos_theta - x1*sin_theta; dst_data[1] = x0*sin_theta + x1*cos_theta; - } else { - const float * const src = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); - float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + } + } else { + // TODO: this is probably wrong, but I can't figure it out .. + // ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28 + for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { + for (int64_t ic = 0; ic < n_dims; ic += 2) { + const float cos_theta = cosf(theta); + const float sin_theta = sinf(theta); - const float x0 = src[0]; - const float x1 = src[n_dims/2]; + theta *= theta_scale; - dst_data[0] = x0*cos_theta - x1*sin_theta; - dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta; + const int64_t i0 = ib*n_dims + ic/2; + + const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims/2]; + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta; + } } } } @@ -9501,8 +10822,8 @@ static void ggml_compute_forward_rope_f16( const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { - assert(src1->type == GGML_TYPE_I32); - assert(ggml_nelements(src1) == 3); + GGML_ASSERT(src1->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_nelements(src1) == 3); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; @@ -9512,25 +10833,35 @@ static void ggml_compute_forward_rope_f16( const int n_dims = ((int32_t *) src1->data)[1]; const int mode = ((int32_t *) src1->data)[2]; - //const int64_t ne0 = src0->ne[0]; - const int64_t ne1 = src0->ne[1]; - const int64_t ne2 = src0->ne[2]; - const int64_t ne3 = src0->ne[3]; + assert(n_past >= 0); - const int nb0 = src0->nb[0]; - const int nb1 = src0->nb[1]; - const int nb2 = src0->nb[2]; - const int nb3 = src0->nb[3]; + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const int64_t ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t ne3 = dst->ne[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); //printf("n_past = %d, ne2 = %d\n", n_past, ne2); - assert(nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb0 == sizeof(ggml_fp16_t)); const int ith = params->ith; const int nth = params->nth; - const int nr = ggml_nrows(src0); + const int nr = ggml_nrows(dst); + + GGML_ASSERT(n_dims <= ne0); + GGML_ASSERT(n_dims % 2 == 0); // rows per thread const int dr = (nr + nth - 1)/nth; @@ -9548,37 +10879,50 @@ static void ggml_compute_forward_rope_f16( for (int64_t i3 = 0; i3 < ne3; i3++) { for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) { - const int p = ((mode & 1) == 0 ? n_past + i2 : i2); + const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2); for (int64_t i1 = 0; i1 < ne1; i1++) { if (ir++ < ir0) continue; if (ir > ir1) break; float theta = (float)p; - for (int i0 = 0; i0 < n_dims; i0 += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + if (!is_neox) { + for (int64_t i0 = 0; i0 < ne0; i0 += 2) { + const float cos_theta = cosf(theta); + const float sin_theta = sinf(theta); - theta *= theta_scale; + theta *= theta_scale; - if (!is_neox) { - const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); const float x0 = GGML_FP16_TO_FP32(src[0]); const float x1 = GGML_FP16_TO_FP32(src[1]); dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); - } else { - const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); - ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + } + } else { + // TODO: this is probably wrong, but I can't figure it out .. + // ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28 + for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { + for (int64_t ic = 0; ic < n_dims; ic += 2) { + const float cos_theta = cosf(theta); + const float sin_theta = sinf(theta); - const float x0 = GGML_FP16_TO_FP32(src[0]); - const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]); + theta *= theta_scale; - dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); - dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + const int64_t i0 = ib*n_dims + ic/2; + + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float x0 = GGML_FP16_TO_FP32(src[0]); + const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]); + + dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + } } } } @@ -9607,6 +10951,255 @@ static void ggml_compute_forward_rope( } } +// ggml_compute_forward_rope_back + +static void ggml_compute_forward_rope_back_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 3); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + // y = rope(x, src1) + // dx = rope_back(dy, src1) + // src0 is dy, src1 contains options + + const int n_past = ((int32_t *) src1->data)[0]; + const int n_dims = ((int32_t *) src1->data)[1]; + const int mode = ((int32_t *) src1->data)[2]; + + assert(n_past >= 0); + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const int64_t ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t ne3 = dst->ne[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + + //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); + //printf("n_past = %d, ne2 = %d\n", n_past, ne2); + + assert(nb0 == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(dst); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + // row index used to determine which thread to use + int ir = 0; + + const float theta_scale = powf(10000.0, -2.0f/n_dims); + + const bool is_neox = mode & 2; + + for (int64_t i3 = 0; i3 < ne3; i3++) { + for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) { + const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2); + for (int64_t i1 = 0; i1 < ne1; i1++) { + if (ir++ < ir0) continue; + if (ir > ir1) break; + + float theta = (float)p; + + if (!is_neox) { + for (int64_t i0 = 0; i0 < ne0; i0 += 2) { + const float cos_theta = cosf(theta); + const float sin_theta = sinf(theta); + + theta *= theta_scale; + + const float * const dy = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + float * dx = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float dy0 = dy[0]; + const float dy1 = dy[1]; + + dx[0] = dy0*cos_theta + dy1*sin_theta; + dx[1] = - dy0*sin_theta + dy1*cos_theta; + } + } else { + for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { + for (int64_t ic = 0; ic < n_dims; ic += 2) { + const float cos_theta = cosf(theta); + const float sin_theta = sinf(theta); + + theta *= theta_scale; + + const int64_t i0 = ib*n_dims + ic/2; + + const float * const dy = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + float * dx = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float dy0 = dy[0]; + const float dy1 = dy[n_dims/2]; + + dx[0] = dy0*cos_theta + dy1*sin_theta; + dx[n_dims/2] = - dy0*sin_theta + dy1*cos_theta; + } + } + } + } + } + } +} + +static void ggml_compute_forward_rope_back_f16( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 3); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + // y = rope(x, src1) + // dx = rope_back(dy, src1) + // src0 is dy, src1 contains options + + const int n_past = ((int32_t *) src1->data)[0]; + const int n_dims = ((int32_t *) src1->data)[1]; + const int mode = ((int32_t *) src1->data)[2]; + + assert(n_past >= 0); + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + const size_t nb02 = src0->nb[2]; + const size_t nb03 = src0->nb[3]; + + const int64_t ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t ne3 = dst->ne[3]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + + //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); + //printf("n_past = %d, ne2 = %d\n", n_past, ne2); + + assert(nb0 == sizeof(ggml_fp16_t)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(dst); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + // row index used to determine which thread to use + int ir = 0; + + const float theta_scale = powf(10000.0, -2.0f/n_dims); + + const bool is_neox = mode & 2; + + for (int64_t i3 = 0; i3 < ne3; i3++) { + for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) { + const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2); + for (int64_t i1 = 0; i1 < ne1; i1++) { + if (ir++ < ir0) continue; + if (ir > ir1) break; + + float theta = (float)p; + + if (!is_neox) { + for (int64_t i0 = 0; i0 < ne0; i0 += 2) { + const float cos_theta = cosf(theta); + const float sin_theta = sinf(theta); + + theta *= theta_scale; + + const ggml_fp16_t * const dy = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + ggml_fp16_t * dx = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float dy0 = GGML_FP16_TO_FP32(dy[0]); + const float dy1 = GGML_FP16_TO_FP32(dy[1]); + + dx[0] = GGML_FP32_TO_FP16( dy0*cos_theta + dy1*sin_theta); + dx[1] = GGML_FP32_TO_FP16(-dy0*sin_theta + dy1*cos_theta); + } + } else { + for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { + for (int64_t ic = 0; ic < n_dims; ic += 2) { + const float cos_theta = cosf(theta); + const float sin_theta = sinf(theta); + + theta *= theta_scale; + + const int64_t i0 = ib*n_dims + ic/2; + + const ggml_fp16_t * const dy = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + ggml_fp16_t * dx = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float dy0 = GGML_FP16_TO_FP32(dy[0]); + const float dy1 = GGML_FP16_TO_FP32(dy[n_dims/2]); + + dx[0] = GGML_FP32_TO_FP16( dy0*cos_theta + dy1*sin_theta); + dx[n_dims/2] = GGML_FP32_TO_FP16(-dy0*sin_theta + dy1*cos_theta); + } + } + } + } + } + } +} + +static void ggml_compute_forward_rope_back( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_rope_back_f16(params, src0, src1, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_rope_back_f32(params, src0, src1, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_conv_1d_1s static void ggml_compute_forward_conv_1d_1s_f16_f32( @@ -10926,6 +12519,14 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_add(params, tensor->src0, tensor->src1, tensor); } break; + case GGML_OP_ADD1: + { + ggml_compute_forward_add1(params, tensor->src0, tensor->src1, tensor); + } break; + case GGML_OP_ACC: + { + ggml_compute_forward_acc(params, tensor->src0, tensor->src1, tensor->opt[0], tensor); + } break; case GGML_OP_SUB: { ggml_compute_forward_sub(params, tensor->src0, tensor->src1, tensor); @@ -10946,10 +12547,18 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_sqrt(params, tensor->src0, tensor); } break; + case GGML_OP_LOG: + { + ggml_compute_forward_log(params, tensor->src0, tensor); + } break; case GGML_OP_SUM: { ggml_compute_forward_sum(params, tensor->src0, tensor); } break; + case GGML_OP_SUM_ROWS: + { + ggml_compute_forward_sum_rows(params, tensor->src0, tensor); + } break; case GGML_OP_MEAN: { ggml_compute_forward_mean(params, tensor->src0, tensor); @@ -10986,6 +12595,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_silu(params, tensor->src0, tensor); } break; + case GGML_OP_SILU_BACK: + { + ggml_compute_forward_silu_back(params, tensor->src0, tensor->src1, tensor); + } break; case GGML_OP_NORM: { ggml_compute_forward_norm(params, tensor->src0, tensor); @@ -10994,6 +12607,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_rms_norm(params, tensor->src0, tensor); } break; + case GGML_OP_RMS_NORM_BACK: + { + ggml_compute_forward_rms_norm_back(params, tensor->src0, tensor->src1, tensor); + } break; case GGML_OP_MUL_MAT: { ggml_compute_forward_mul_mat(params, tensor->src0, tensor->src1, tensor); @@ -11002,6 +12619,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_scale(params, tensor->src0, tensor->src1, tensor); } break; + case GGML_OP_SET: + { + ggml_compute_forward_set(params, tensor->src0, tensor->src1, tensor->opt[0], tensor); + } break; case GGML_OP_CPY: { ggml_compute_forward_cpy(params, tensor->src0, tensor); @@ -11030,10 +12651,22 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_get_rows(params, tensor->src0, tensor->src1, tensor); } break; + case GGML_OP_GET_ROWS_BACK: + { + ggml_compute_forward_get_rows_back(params, tensor->src0, tensor->src1, tensor->opt[0], tensor); + } break; + case GGML_OP_DIAG: + { + ggml_compute_forward_diag(params, tensor->src0, tensor); + } break; case GGML_OP_DIAG_MASK_INF: { ggml_compute_forward_diag_mask_inf(params, tensor->src0, tensor->src1, tensor); } break; + case GGML_OP_DIAG_MASK_ZERO: + { + ggml_compute_forward_diag_mask_zero(params, tensor->src0, tensor->src1, tensor); + } break; case GGML_OP_SOFT_MAX: { ggml_compute_forward_soft_max(params, tensor->src0, tensor); @@ -11042,6 +12675,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_rope(params, tensor->src0, tensor->src1, tensor); } break; + case GGML_OP_ROPE_BACK: + { + ggml_compute_forward_rope_back(params, tensor->src0, tensor->src1, tensor); + } break; case GGML_OP_ALIBI: { ggml_compute_forward_alibi(params, tensor->src0, tensor->src1, tensor); @@ -11110,6 +12747,48 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor src1->grad = ggml_add_impl(ctx, src1->grad, tensor->grad, inplace); } } break; + case GGML_OP_ADD1: + { + if (src0->grad) { + src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace); + } + if (src1->grad) { + src1->grad = ggml_add_impl(ctx, + src1->grad, + ggml_mean(ctx, tensor->grad), // TODO: should probably be sum instead of mean + inplace); + } + } break; + case GGML_OP_ACC: + { + if (src0->grad) { + src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace); + } + if (src1->grad) { + GGML_ASSERT(ggml_nelements(tensor->opt[0]) == 5); + GGML_ASSERT(tensor->opt[0]->type == GGML_TYPE_I32); + const size_t nb1 = (( int32_t * ) tensor->opt[0]->data)[0]; + const size_t nb2 = (( int32_t * ) tensor->opt[0]->data)[1]; + const size_t nb3 = (( int32_t * ) tensor->opt[0]->data)[2]; + const size_t offset = (( int32_t * ) tensor->opt[0]->data)[3]; + + struct ggml_tensor * tensor_grad_view = ggml_view_4d(ctx, + tensor->grad, + src1->grad->ne[0], + src1->grad->ne[1], + src1->grad->ne[2], + src1->grad->ne[3], + nb1, nb2, nb3, offset); + + src1->grad = + ggml_add_impl(ctx, + src1->grad, + ggml_reshape(ctx, + ggml_cont(ctx, tensor_grad_view), + src1->grad), + inplace); + } + } break; case GGML_OP_SUB: { if (src0->grad) { @@ -11161,31 +12840,57 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor src0->grad = ggml_add_impl(ctx, src0->grad, - ggml_mul(ctx, + ggml_scale(ctx, ggml_mul(ctx, src0, tensor->grad), - ggml_repeat(ctx, ggml_new_f32(ctx, 2.0f), src0)), + ggml_new_f32(ctx, 2.0f)), inplace); } } break; case GGML_OP_SQRT: + { + if (src0->grad) { + src0->grad = + ggml_add_impl(ctx, + src0->grad, + ggml_mul(ctx, + tensor->grad, // this was not catched by test_grad because in test_grad tensor->grad is 1 + ggml_div(ctx, + ggml_repeat(ctx, ggml_new_f32(ctx, 0.5f), tensor), + tensor)), + inplace); + } + } break; + case GGML_OP_LOG: { if (src0->grad) { src0->grad = ggml_add_impl(ctx, src0->grad, ggml_div(ctx, - ggml_repeat(ctx, ggml_new_f32(ctx, 0.5f), tensor), - tensor), + tensor->grad, + src0), inplace); } } break; case GGML_OP_SUM: + { + if (src0->grad) { + src0->grad = + ggml_add1_impl(ctx, + src0->grad, + tensor->grad, + inplace); + } + } break; + case GGML_OP_SUM_ROWS: { if (src0->grad) { src0->grad = ggml_add_impl(ctx, src0->grad, - ggml_repeat(ctx, tensor->grad, src0->grad), + ggml_repeat(ctx, + tensor->grad, + src0->grad), inplace); } } break; @@ -11195,11 +12900,44 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor } break; case GGML_OP_REPEAT: { + // necessary for llama if (src0->grad) { + GGML_ASSERT(src0->n_dims == 1 || src0->n_dims == 2); + const int nc = tensor->ne[0]; + const int nr = tensor->ne[1]; + const int nc0 = src0->ne[0]; + const int nr0 = src0->ne[1]; + const int ncr = nc/nc0; // guaranteed to be an integer due to the check in ggml_can_repeat + const int nrr = nr/nr0; // guaranteed to be an integer due to the check in ggml_can_repeat + // tensor->grad [nc,nr,1,1] + // reshape [nc0,nc/nc0,nr0,nr/nr0] + // permute [nc0,nr0,nc/nc0,nr/nr0] + // substitute [nc0,nr0,ncr,nrr] + // reshape [nc0*nr0,ncr*nrr,1,1] + // transpose [ncr*nrr,nc0*nr0,1,1] + // sum rows [1,nc0*nr0,1,1] + // transpose [nc0*nr0,1,1] + // reshape [nc0,nr0,1,1] reshape_1d or reshape_2d + // add to src0->grad + + int64_t ne[4] = {nc0,ncr,nr0,nrr}; + + struct ggml_tensor* F00 = tensor->grad; + struct ggml_tensor* F01 = ggml_reshape (ctx, F00, ggml_new_tensor(ctx,tensor->grad->type,4,ne)); + struct ggml_tensor* F02 = ggml_permute (ctx, F01, 0,2,1,3); + struct ggml_tensor* F03 = ggml_cont (ctx, F02); + struct ggml_tensor* F04 = ggml_reshape_2d(ctx, F03, nc0*nr0, ncr*nrr); + struct ggml_tensor* F05 = ggml_transpose (ctx, F04); + struct ggml_tensor* F06 = ggml_cont (ctx, F05); + struct ggml_tensor* F07 = ggml_sum_rows (ctx, F06); + struct ggml_tensor* F08 = ggml_transpose (ctx, F07); + struct ggml_tensor* F09 = ggml_cont (ctx, F08); + struct ggml_tensor* F10 = ggml_reshape (ctx, F09, src0->grad); + src0->grad = ggml_add_impl(ctx, src0->grad, - ggml_sum(ctx, tensor->grad), + F10, inplace); } } break; @@ -11253,6 +12991,16 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor GGML_ASSERT(false); // TODO: not implemented } break; case GGML_OP_SILU: + { + // necessary for llama + if (src0->grad) { + src0->grad = ggml_add_impl(ctx, + src0->grad, + ggml_silu_back(ctx, src0, tensor->grad), + inplace); + } + } break; + case GGML_OP_SILU_BACK: { GGML_ASSERT(false); // TODO: not implemented } break; @@ -11261,68 +13009,372 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor GGML_ASSERT(false); // TODO: not implemented } break; case GGML_OP_RMS_NORM: + { + // necessary for llama + if (src0->grad) { + src0->grad = ggml_add_impl(ctx, + src0->grad, + ggml_rms_norm_back(ctx, src0, tensor->grad), + inplace); + } + } break; + case GGML_OP_RMS_NORM_BACK: { GGML_ASSERT(false); // TODO: not implemented } break; case GGML_OP_MUL_MAT: { + // https://cs231n.github.io/optimization-2/#staged + // # forward pass + // s0 = np.random.randn(5, 10) + // s1 = np.random.randn(10, 3) + // t = s0.dot(s1) + + // # now suppose we had the gradient on t from above in the circuit + // dt = np.random.randn(*t.shape) # same shape as t + // ds0 = dt.dot(s1.T) #.T gives the transpose of the matrix + // ds1 = t.T.dot(dt) + + // tensor.shape [m,p] + // src0.shape [n,m] + // src1.shape [n,p] + + // necessary for llama if (src0->grad) { // TODO: this requires outer product - ggml_out_prod(ctx, src1, tensor->grad); - GGML_ASSERT(false); + src0->grad = + ggml_add_impl(ctx, + src0->grad, + // ds0 = dt.dot(s1.T) + // ggml_out_prod(ctx, // [n,m] + // src1, // [n,p] + // tensor->grad), // [m,p] + // for now just using A*B==(B.T*A.T).T + ggml_cont(ctx, // [n,m] + ggml_transpose(ctx, // [n,m] + ggml_mul_mat(ctx, // [m,n] + ggml_cont(ctx, // [p,m] + ggml_transpose(ctx, // [p,m] + tensor->grad)), // [m,p] + ggml_cont(ctx, // [p,n] + ggml_transpose(ctx, // [p,n] + src1))))), // [n,p] + inplace); } if (src1->grad) { src1->grad = ggml_add_impl(ctx, src1->grad, - ggml_mul_mat(ctx, - ggml_cont(ctx, ggml_transpose(ctx, src0)), - tensor->grad), + // ds1 = s0.T.dot(dt): + ggml_mul_mat(ctx, // [n,p] + ggml_cont(ctx, // [m,n] + ggml_transpose(ctx, src0)), // [m,n] + tensor->grad), // [m,p] inplace); } } break; case GGML_OP_SCALE: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + src0->grad = + ggml_add_impl(ctx, + src0->grad, + ggml_scale_impl(ctx, tensor->grad, src1, false), + inplace); + } + if (src1->grad) { + src1->grad = + ggml_add_impl(ctx, + src1->grad, + ggml_sum(ctx, ggml_mul_impl(ctx, tensor->grad, src0, false)), + inplace); + } + } break; + case GGML_OP_SET: + { + GGML_ASSERT(ggml_nelements(tensor->opt[0]) == 5); + GGML_ASSERT(tensor->opt[0]->type == GGML_TYPE_I32); + const size_t nb1 = (( int32_t * ) tensor->opt[0]->data)[0]; + const size_t nb2 = (( int32_t * ) tensor->opt[0]->data)[1]; + const size_t nb3 = (( int32_t * ) tensor->opt[0]->data)[2]; + const size_t offset = (( int32_t * ) tensor->opt[0]->data)[3]; + + struct ggml_tensor * tensor_grad_view = NULL; + + if (src0->grad || src1->grad) { + GGML_ASSERT(src0->type == tensor->type); + GGML_ASSERT(tensor->grad->type == tensor->type); + GGML_ASSERT(tensor->grad->type == src1->grad->type); + + tensor_grad_view = ggml_view_4d(ctx, + tensor->grad, + src1->grad->ne[0], + src1->grad->ne[1], + src1->grad->ne[2], + src1->grad->ne[3], + nb1, nb2, nb3, offset); + } + + if (src0->grad) { + src0->grad = ggml_add_impl(ctx, + src0->grad, + ggml_acc_impl(ctx, + tensor->grad, + ggml_neg(ctx, tensor_grad_view), + nb1, nb2, nb3, offset, false), + inplace); + } + + if (src1->grad) { + src1->grad = + ggml_add_impl(ctx, + src1->grad, + ggml_reshape(ctx, + ggml_cont(ctx, tensor_grad_view), + src1->grad), + inplace); + } } break; case GGML_OP_CPY: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + // cpy overwrites value of src1 by src0 and returns view(src1) + // the overwriting is mathematically equivalent to: + // tensor = src0 * 1 + src1 * 0 + if (src0->grad) { + // dsrc0 = dtensor * 1 + src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace); + } + if (src1->grad) { + // dsrc1 = dtensor * 0 -> noop + } } break; case GGML_OP_CONT: { - GGML_ASSERT(false); // TODO: not implemented + // same as cpy + if (src0->grad) { + GGML_ASSERT(ggml_is_contiguous(src0->grad)); + GGML_ASSERT(ggml_is_contiguous(tensor->grad)); + src0->grad = ggml_add_impl(ctx, src0->grad, tensor->grad, inplace); + } } break; case GGML_OP_RESHAPE: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + src0->grad = + ggml_add_impl(ctx, src0->grad, + ggml_reshape(ctx, tensor->grad, src0->grad), + inplace); + } } break; case GGML_OP_VIEW: { - GGML_ASSERT(false); // not supported + // necessary for llama + if (src0->grad) { + size_t offset; + memcpy(&offset, tensor->padding, sizeof(offset)); + + size_t nb1 = tensor->nb[1]; + size_t nb2 = tensor->nb[2]; + size_t nb3 = tensor->nb[3]; + + if (src0->type != src0->grad->type) { + // gradient is typically F32, but src0 could be other type + size_t ng = ggml_element_size(src0->grad); + size_t n0 = ggml_element_size(src0); + GGML_ASSERT(offset % n0 == 0); + GGML_ASSERT(nb1 % n0 == 0); + GGML_ASSERT(nb2 % n0 == 0); + GGML_ASSERT(nb3 % n0 == 0); + offset = (offset / n0) * ng; + nb1 = (nb1 / n0) * ng; + nb2 = (nb2 / n0) * ng; + nb3 = (nb3 / n0) * ng; + } + + src0->grad = ggml_acc_impl(ctx, src0->grad, tensor->grad, nb1, nb2, nb3, offset, inplace); + } } break; case GGML_OP_PERMUTE: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + int axis0 = tensor->padding[0] & 0x3; + int axis1 = tensor->padding[1] & 0x3; + int axis2 = tensor->padding[2] & 0x3; + int axis3 = tensor->padding[3] & 0x3; + int axes_backward[4] = {0,0,0,0}; + axes_backward[axis0] = 0; + axes_backward[axis1] = 1; + axes_backward[axis2] = 2; + axes_backward[axis3] = 3; + src0->grad = + ggml_add_impl(ctx, src0->grad, + ggml_permute(ctx, + tensor->grad, + axes_backward[0], + axes_backward[1], + axes_backward[2], + axes_backward[3]), + inplace); + } } break; case GGML_OP_TRANSPOSE: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + src0->grad = + ggml_add_impl(ctx, src0->grad, + ggml_transpose(ctx, tensor->grad), + inplace); + } } break; case GGML_OP_GET_ROWS: + { + // necessary for llama (only for tokenizer) + if (src0->grad) { + src0->grad = + ggml_add_impl(ctx, src0->grad, + ggml_get_rows_back(ctx, tensor->grad, src1, src0->grad), + inplace); + } + if (src1->grad) { + // noop + } + } break; + case GGML_OP_GET_ROWS_BACK: + { + GGML_ASSERT(false); // TODO: not implemented + } break; + case GGML_OP_DIAG: { GGML_ASSERT(false); // TODO: not implemented } break; case GGML_OP_DIAG_MASK_INF: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 2); + const int n_past = ((int32_t *) src1->data)[0]; + src0->grad = + ggml_add_impl(ctx, src0->grad, + ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false), + inplace); + } + if (src1->grad) { + // noop + } + } break; + case GGML_OP_DIAG_MASK_ZERO: + { + // necessary for llama + if (src0->grad) { + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 2); + const int n_past = ((int32_t *) src1->data)[0]; + src0->grad = + ggml_add_impl(ctx, src0->grad, + ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false), + inplace); + } + if (src1->grad) { + // noop + } } break; case GGML_OP_SOFT_MAX: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + // y = softmax(x) + // + // Jii = yi - yi*yi + // Jij = -yi*yj + // J = diag(y)-y.*y + // dx = J * dy + // dxk = sum(Jkj * dyk) + + int64_t ne2[4] = { + tensor->ne[0], + 1, + tensor->ne[1]*tensor->ne[2], + tensor->ne[3] + }; + struct ggml_tensor * tensor2 = ggml_cont(ctx, + ggml_reshape_4d(ctx, + ggml_cont(ctx, tensor), + ne2[0], ne2[1], ne2[2], ne2[3])); + + struct ggml_tensor * grad2 = ggml_cont(ctx, + ggml_reshape_4d(ctx, + ggml_cont(ctx, tensor->grad), + ne2[0], ne2[1], ne2[2], ne2[3])); + + struct ggml_tensor * tensor2_t = ggml_cont(ctx, // [1,ne0,ne1*ne2,ne3] + ggml_permute(ctx, // [1,ne0,ne1*ne2,ne3] + tensor2, // [ne0,1,ne1*ne2,ne3] + 1, 0, 2, 3)); + + src0->grad = + ggml_add_impl(ctx, + src0->grad, // [ne0,ne1,ne2,ne3] + ggml_reshape(ctx, // [ne0,ne1,ne2,ne3] + ggml_mul_mat(ctx, // [ne0,1,ne1*ne2,ne3] + ggml_sub(ctx, // [ne0,ne0,ne1*ne2,ne3] + ggml_diag(ctx, // [ne0,ne0,ne1*ne2,ne3] + tensor2), // [ne0,1,ne1*ne2,ne3] + ggml_mul_mat(ctx, // [ne0,ne0,ne1*ne2,ne3] + tensor2_t, // [1,ne0,ne1*ne2,ne3] + tensor2_t)), // [1,ne0,ne1*ne2,ne3] + grad2), // [ne0,1,ne1*ne2,ne3] + src0->grad), + inplace); + } } break; case GGML_OP_ROPE: { - GGML_ASSERT(false); // TODO: not implemented + // necessary for llama + if (src0->grad) { + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 3); + const int n_past = ((int32_t *) src1->data)[0]; + const int n_dims = ((int32_t *) src1->data)[1]; + const int mode = ((int32_t *) src1->data)[2]; + src0->grad = ggml_add_impl(ctx, + src0->grad, + ggml_rope_back(ctx, + tensor->grad, + n_past, + n_dims, + mode), + inplace); + } + if (src1->grad) { + // noop + } + } break; + case GGML_OP_ROPE_BACK: + { + if (src0->grad) { + assert(src1->type == GGML_TYPE_I32); + assert(ggml_nelements(src1) == 3); + const int n_past = ((int32_t *) src1->data)[0]; + const int n_dims = ((int32_t *) src1->data)[1]; + const int mode = ((int32_t *) src1->data)[2]; + src0->grad = ggml_add_impl(ctx, + src0->grad, + ggml_rope(ctx, + tensor->grad, + n_past, + n_dims, + mode), + inplace); + } + if (src1->grad) { + // noop + } } break; case GGML_OP_CONV_1D_1S: { @@ -11536,7 +13588,11 @@ typedef int ggml_lock_t; #define ggml_lock_init(x) UNUSED(x) #define ggml_lock_destroy(x) UNUSED(x) +#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64)) +#define ggml_lock_lock(x) _mm_pause() +#else #define ggml_lock_lock(x) UNUSED(x) +#endif #define ggml_lock_unlock(x) UNUSED(x) #define GGML_LOCK_INITIALIZER 0 @@ -11676,6 +13732,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) work_size = MAX(work_size, cur); } break; case GGML_OP_ADD: + case GGML_OP_ADD1: { node->n_tasks = n_threads; @@ -11685,14 +13742,27 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) cur = GGML_TYPE_SIZE[GGML_TYPE_F32] * node->src0->ne[0] * n_threads; } + work_size = MAX(work_size, cur); + } break; + case GGML_OP_ACC: + { + node->n_tasks = n_threads; + + size_t cur = 0; + + if (ggml_is_quantized(node->src0->type)) { + cur = GGML_TYPE_SIZE[GGML_TYPE_F32] * node->src1->ne[0] * n_threads; + } + work_size = MAX(work_size, cur); } break; case GGML_OP_SUB: - case GGML_OP_MUL: case GGML_OP_DIV: case GGML_OP_SQR: case GGML_OP_SQRT: + case GGML_OP_LOG: case GGML_OP_SUM: + case GGML_OP_SUM_ROWS: case GGML_OP_MEAN: case GGML_OP_REPEAT: case GGML_OP_ABS: @@ -11703,16 +13773,13 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) { node->n_tasks = 1; } break; + case GGML_OP_MUL: case GGML_OP_GELU: - { - node->n_tasks = n_threads; - } break; case GGML_OP_SILU: - { - node->n_tasks = n_threads; - } break; + case GGML_OP_SILU_BACK: case GGML_OP_NORM: case GGML_OP_RMS_NORM: + case GGML_OP_RMS_NORM_BACK: { node->n_tasks = n_threads; } break; @@ -11778,21 +13845,23 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) { node->n_tasks = n_threads; } break; + case GGML_OP_SET: case GGML_OP_CONT: case GGML_OP_RESHAPE: case GGML_OP_VIEW: case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: case GGML_OP_GET_ROWS: - case GGML_OP_DIAG_MASK_INF: + case GGML_OP_GET_ROWS_BACK: + case GGML_OP_DIAG: + case GGML_OP_DIAG_MASK_ZERO: { node->n_tasks = 1; } break; + case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: - { - node->n_tasks = n_threads; - } break; case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: { node->n_tasks = n_threads; } break; @@ -12929,7 +14998,7 @@ enum ggml_opt_result ggml_opt( // build forward + backward compute graphs struct ggml_cgraph gf = ggml_build_forward (f); - struct ggml_cgraph gb = ggml_build_backward(ctx, &gf, false); + struct ggml_cgraph gb = ggml_build_backward(ctx, &gf, true); switch (params.type) { case GGML_OPT_ADAM: @@ -12965,15 +15034,15 @@ size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * assert(k % QK4_0 == 0); const int nb = k / QK4_0; - for (int j = 0; j < n; j += k) { - block_q4_0 * restrict y = (block_q4_0 *)dst + j/QK4_0; + for (int b = 0; b < n; b += k) { + block_q4_0 * restrict y = (block_q4_0 *) dst + b/QK4_0; - quantize_row_q4_0_reference(src + j, y, k); + quantize_row_q4_0_reference(src + b, y, k); for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK4_0; l += 2) { - const uint8_t vi0 = y[i].qs[l/2] & 0x0F; - const uint8_t vi1 = y[i].qs[l/2] >> 4; + for (int j = 0; j < QK4_0; j += 2) { + const uint8_t vi0 = y[i].qs[j/2] & 0x0F; + const uint8_t vi1 = y[i].qs[j/2] >> 4; hist[vi0]++; hist[vi1]++; @@ -12988,15 +15057,15 @@ size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * assert(k % QK4_1 == 0); const int nb = k / QK4_1; - for (int j = 0; j < n; j += k) { - block_q4_1 * restrict y = (block_q4_1 *)dst + j/QK4_1; + for (int b = 0; b < n; b += k) { + block_q4_1 * restrict y = (block_q4_1 *) dst + b/QK4_1; - quantize_row_q4_1_reference(src + j, y, k); + quantize_row_q4_1_reference(src + b, y, k); for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK4_1; l += 2) { - const uint8_t vi0 = y[i].qs[l/2] & 0x0F; - const uint8_t vi1 = y[i].qs[l/2] >> 4; + for (int j = 0; j < QK4_1; j += 2) { + const uint8_t vi0 = y[i].qs[j/2] & 0x0F; + const uint8_t vi1 = y[i].qs[j/2] >> 4; hist[vi0]++; hist[vi1]++; @@ -13007,49 +15076,26 @@ size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * return (n/QK4_1*sizeof(block_q4_1)); } -size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t * hist) { - assert(k % QK4_2 == 0); - const int nb = k / QK4_2; - - for (int j = 0; j < n; j += k) { - block_q4_2 * restrict y = (block_q4_2 *)dst + j/QK4_2; - - quantize_row_q4_2_reference(src + j, y, k); - - for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK4_2; l += 2) { - const uint8_t vi0 = y[i].qs[l/2] & 0x0F; - const uint8_t vi1 = y[i].qs[l/2] >> 4; - - hist[vi0]++; - hist[vi1]++; - } - } - } - - return (n/QK4_2*sizeof(block_q4_2)); -} - size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist) { assert(k % QK5_0 == 0); const int nb = k / QK5_0; - for (int j = 0; j < n; j += k) { - block_q5_0 * restrict y = (block_q5_0 *)dst + j/QK5_0; + for (int b = 0; b < n; b += k) { + block_q5_0 * restrict y = (block_q5_0 *)dst + b/QK5_0; - quantize_row_q5_0_reference(src + j, y, k); + quantize_row_q5_0_reference(src + b, y, k); for (int i = 0; i < nb; i++) { uint32_t qh; memcpy(&qh, &y[i].qh, sizeof(qh)); - for (int l = 0; l < QK5_0; l += 2) { - const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4; - const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4; + for (int j = 0; j < QK5_0; j += 2) { + const uint8_t vh0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; + const uint8_t vh1 = ((qh & (1u << (j + 16))) >> (j + 12)); // cast to 16 bins - const uint8_t vi0 = ((y[i].qs[l/2] & 0x0F) | vh0) / 2; - const uint8_t vi1 = ((y[i].qs[l/2] >> 4) | vh1) / 2; + const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2; + const uint8_t vi1 = ((y[i].qs[j/2] >> 4) | vh1) / 2; hist[vi0]++; hist[vi1]++; @@ -13064,22 +15110,22 @@ size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * assert(k % QK5_1 == 0); const int nb = k / QK5_1; - for (int j = 0; j < n; j += k) { - block_q5_1 * restrict y = (block_q5_1 *)dst + j/QK5_1; + for (int b = 0; b < n; b += k) { + block_q5_1 * restrict y = (block_q5_1 *)dst + b/QK5_1; - quantize_row_q5_1_reference(src + j, y, k); + quantize_row_q5_1_reference(src + b, y, k); for (int i = 0; i < nb; i++) { uint32_t qh; memcpy(&qh, &y[i].qh, sizeof(qh)); - for (int l = 0; l < QK5_1; l += 2) { - const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4; - const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4; + for (int j = 0; j < QK5_1; j += 2) { + const uint8_t vh0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; + const uint8_t vh1 = ((qh & (1u << (j + 16))) >> (j + 12)); // cast to 16 bins - const uint8_t vi0 = ((y[i].qs[l/2] & 0x0F) | vh0) / 2; - const uint8_t vi1 = ((y[i].qs[l/2] >> 4) | vh1) / 2; + const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2; + const uint8_t vi1 = ((y[i].qs[j/2] >> 4) | vh1) / 2; hist[vi0]++; hist[vi1]++; @@ -13094,14 +15140,14 @@ size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * assert(k % QK8_0 == 0); const int nb = k / QK8_0; - for (int j = 0; j < n; j += k) { - block_q8_0 * restrict y = (block_q8_0 *)dst + j/QK8_0; + for (int b = 0; b < n; b += k) { + block_q8_0 * restrict y = (block_q8_0 *)dst + b/QK8_0; - quantize_row_q8_0_reference(src + j, y, k); + quantize_row_q8_0_reference(src + b, y, k); for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK8_0; ++l) { - const int8_t vi = y[i].qs[l]; + for (int j = 0; j < QK8_0; ++j) { + const int8_t vi = y[i].qs[j]; hist[vi/16 + 8]++; } @@ -13126,12 +15172,6 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i block_q4_1 * block = (block_q4_1*)dst + start / QK4_1; result = ggml_quantize_q4_1(src + start, block, n, n, hist); } break; - case GGML_TYPE_Q4_2: - { - GGML_ASSERT(start % QK4_2 == 0); - block_q4_2 * block = (block_q4_2*)dst + start / QK4_2; - result = ggml_quantize_q4_2(src + start, block, n, n, hist); - } break; case GGML_TYPE_Q5_0: { GGML_ASSERT(start % QK5_0 == 0); diff --git a/ggml.h b/ggml.h index 508dd69..255541d 100644 --- a/ggml.h +++ b/ggml.h @@ -190,9 +190,12 @@ #define GGML_FILE_MAGIC 0x67676d6c // "ggml" #define GGML_FILE_VERSION 1 +#define GGML_QNT_VERSION 1 // bump this on quantization format changes +#define GGML_QNT_VERSION_FACTOR 1000 // do not change this + #define GGML_MAX_DIMS 4 #define GGML_MAX_NODES 4096 -#define GGML_MAX_PARAMS 16 +#define GGML_MAX_PARAMS 256 #define GGML_MAX_CONTEXTS 64 #define GGML_MAX_OPT 4 #define GGML_DEFAULT_N_THREADS 4 @@ -231,7 +234,7 @@ extern "C" { GGML_TYPE_F16 = 1, GGML_TYPE_Q4_0 = 2, GGML_TYPE_Q4_1 = 3, - GGML_TYPE_Q4_2 = 4, + // GGML_TYPE_Q4_2 = 4, support has been removed // GGML_TYPE_Q4_3 (5) support has been removed GGML_TYPE_Q5_0 = 6, GGML_TYPE_Q5_1 = 7, @@ -243,6 +246,11 @@ extern "C" { GGML_TYPE_COUNT, }; + enum ggml_backend { + GGML_BACKEND_CPU = 0, + GGML_BACKEND_CUDA = 1, + }; + // model file types enum ggml_ftype { GGML_FTYPE_UNKNOWN = -1, @@ -251,7 +259,6 @@ extern "C" { GGML_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors GGML_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 - GGML_FTYPE_MOSTLY_Q4_2 = 5, // except 1d tensors GGML_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors GGML_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors GGML_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors @@ -263,12 +270,16 @@ extern "C" { GGML_OP_DUP, GGML_OP_ADD, + GGML_OP_ADD1, + GGML_OP_ACC, GGML_OP_SUB, GGML_OP_MUL, GGML_OP_DIV, GGML_OP_SQR, GGML_OP_SQRT, + GGML_OP_LOG, GGML_OP_SUM, + GGML_OP_SUM_ROWS, GGML_OP_MEAN, GGML_OP_REPEAT, GGML_OP_ABS, @@ -278,12 +289,15 @@ extern "C" { GGML_OP_RELU, GGML_OP_GELU, GGML_OP_SILU, + GGML_OP_SILU_BACK, GGML_OP_NORM, // normalize GGML_OP_RMS_NORM, + GGML_OP_RMS_NORM_BACK, GGML_OP_MUL_MAT, GGML_OP_SCALE, + GGML_OP_SET, GGML_OP_CPY, GGML_OP_CONT, GGML_OP_RESHAPE, @@ -291,9 +305,13 @@ extern "C" { GGML_OP_PERMUTE, GGML_OP_TRANSPOSE, GGML_OP_GET_ROWS, + GGML_OP_GET_ROWS_BACK, + GGML_OP_DIAG, GGML_OP_DIAG_MASK_INF, + GGML_OP_DIAG_MASK_ZERO, GGML_OP_SOFT_MAX, GGML_OP_ROPE, + GGML_OP_ROPE_BACK, GGML_OP_ALIBI, GGML_OP_CONV_1D_1S, GGML_OP_CONV_1D_2S, @@ -322,7 +340,8 @@ extern "C" { // n-dimensional tensor struct ggml_tensor { - enum ggml_type type; + enum ggml_type type; + enum ggml_backend backend; int n_dims; int64_t ne[GGML_MAX_DIMS]; // number of elements @@ -353,7 +372,7 @@ extern "C" { char name[32]; - char padding[8]; // TODO: remove and add padding to name? + char padding[16]; }; // computation graph @@ -497,6 +516,29 @@ extern "C" { struct ggml_tensor * a, struct ggml_tensor * b); + GGML_API struct ggml_tensor * ggml_add1( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + GGML_API struct ggml_tensor * ggml_acc( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset); + + GGML_API struct ggml_tensor * ggml_acc_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset); + GGML_API struct ggml_tensor * ggml_sub( struct ggml_context * ctx, struct ggml_tensor * a, @@ -520,12 +562,24 @@ extern "C" { struct ggml_context * ctx, struct ggml_tensor * a); + GGML_API struct ggml_tensor * ggml_log( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_log_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + // return scalar - // TODO: compute sum along rows GGML_API struct ggml_tensor * ggml_sum( struct ggml_context * ctx, struct ggml_tensor * a); + // sums along rows, with input shape [a,b,c,d] return shape [1,b,c,d] + GGML_API struct ggml_tensor * ggml_sum_rows( + struct ggml_context * ctx, + struct ggml_tensor * a); + // mean along rows GGML_API struct ggml_tensor * ggml_mean( struct ggml_context * ctx, @@ -567,6 +621,13 @@ extern "C" { struct ggml_context * ctx, struct ggml_tensor * a); + // a - x + // b - dy + GGML_API struct ggml_tensor * ggml_silu_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + // normalize along rows // TODO: eps is hardcoded to 1e-5 for now GGML_API struct ggml_tensor * ggml_norm( @@ -577,6 +638,13 @@ extern "C" { struct ggml_context * ctx, struct ggml_tensor * a); + // a - x + // b - dy + GGML_API struct ggml_tensor * ggml_rms_norm_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + // A: m rows, n columns // B: p rows, n columns (i.e. we transpose it internally) // result is m columns, p rows @@ -589,12 +657,66 @@ extern "C" { // operations on tensors without backpropagation // - // in-place, returns view(a) GGML_API struct ggml_tensor * ggml_scale( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b); + // in-place, returns view(a) + GGML_API struct ggml_tensor * ggml_scale_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + + // b -> view(a,offset,nb1,nb2,3), return modified a + GGML_API struct ggml_tensor * ggml_set( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset); + + // b -> view(a,offset,nb1,nb2,3), return view(a) + GGML_API struct ggml_tensor * ggml_set_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t nb2, + size_t nb3, + size_t offset); + + GGML_API struct ggml_tensor * ggml_set_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t offset); + + GGML_API struct ggml_tensor * ggml_set_1d_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t offset); + + // b -> view(a,offset,nb1,nb2,3), return modified a + GGML_API struct ggml_tensor * ggml_set_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t offset); + + // b -> view(a,offset,nb1,nb2,3), return view(a) + GGML_API struct ggml_tensor * ggml_set_2d_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + size_t nb1, + size_t offset); + + // a -> b, return view(b) GGML_API struct ggml_tensor * ggml_cpy( struct ggml_context * ctx, @@ -615,6 +737,11 @@ extern "C" { // return view(a) // TODO: when we start computing gradient, make a copy instead of view + GGML_API struct ggml_tensor * ggml_reshape_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0); + GGML_API struct ggml_tensor * ggml_reshape_2d( struct ggml_context * ctx, struct ggml_tensor * a, @@ -630,6 +757,14 @@ extern "C" { int64_t ne1, int64_t ne2); + GGML_API struct ggml_tensor * ggml_reshape_4d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3); + // offset in bytes GGML_API struct ggml_tensor * ggml_view_1d( struct ggml_context * ctx, @@ -655,6 +790,18 @@ extern "C" { size_t nb2, // slice stride in bytes size_t offset); + GGML_API struct ggml_tensor * ggml_view_4d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3, + size_t nb1, // row stride in bytes + size_t nb2, // slice stride in bytes + size_t nb3, + size_t offset); + GGML_API struct ggml_tensor * ggml_permute( struct ggml_context * ctx, struct ggml_tensor * a, @@ -673,20 +820,50 @@ extern "C" { struct ggml_tensor * a, struct ggml_tensor * b); + GGML_API struct ggml_tensor * ggml_get_rows_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c); + + GGML_API struct ggml_tensor * ggml_diag( + struct ggml_context * ctx, + struct ggml_tensor * a); + // set elements above the diagonal to -INF - // in-place, returns view(a) GGML_API struct ggml_tensor * ggml_diag_mask_inf( struct ggml_context * ctx, struct ggml_tensor * a, int n_past); // in-place, returns view(a) + GGML_API struct ggml_tensor * ggml_diag_mask_inf_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past); + + // set elements above the diagonal to 0 + GGML_API struct ggml_tensor * ggml_diag_mask_zero( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past); + + // in-place, returns view(a) + GGML_API struct ggml_tensor * gml_diag_mask_zero_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past); + GGML_API struct ggml_tensor * ggml_soft_max( struct ggml_context * ctx, struct ggml_tensor * a); - // rotary position embedding // in-place, returns view(a) + GGML_API struct ggml_tensor * ggml_soft_max_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // rotary position embedding // if mode & 1 == 1, skip n_past elements // if mode & 2 == 1, GPT-NeoX style // TODO: avoid creating a new tensor every time @@ -697,6 +874,23 @@ extern "C" { int n_dims, int mode); + // in-place, returns view(a) + GGML_API struct ggml_tensor * ggml_rope_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode); + + // rotary position embedding backward, i.e compute dx from dy + // a - dy + GGML_API struct ggml_tensor * ggml_rope_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + int n_past, + int n_dims, + int mode); + // alibi position embedding // in-place, returns view(a) struct ggml_tensor * ggml_alibi( @@ -741,13 +935,13 @@ extern "C" { GGML_API struct ggml_tensor * ggml_map_unary_f32( struct ggml_context * ctx, struct ggml_tensor * a, - const ggml_unary_op_f32_t fun); + ggml_unary_op_f32_t fun); GGML_API struct ggml_tensor * ggml_map_binary_f32( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, - const ggml_binary_op_f32_t fun); + ggml_binary_op_f32_t fun); // // automatic differentiation @@ -876,7 +1070,6 @@ extern "C" { GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist); diff --git a/whisper.cpp b/whisper.cpp index 158aa0b..1cdd350 100644 --- a/whisper.cpp +++ b/whisper.cpp @@ -291,15 +291,6 @@ static const std::map> MEM_REQ_MODEL = { { MODEL_LARGE, 1124ull*MB }, }, }, - { GGML_TYPE_Q4_2, - { - { MODEL_TINY, 26ull*MB }, - { MODEL_BASE, 50ull*MB }, - { MODEL_SMALL, 154ull*MB }, - { MODEL_MEDIUM, 470ull*MB }, - { MODEL_LARGE, 940ull*MB }, - }, - }, { GGML_TYPE_Q5_0, { { MODEL_TINY, 30ull*MB }, @@ -861,6 +852,10 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con model.type = e_model::MODEL_LARGE; } + const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR; + + hparams.ftype %= GGML_QNT_VERSION_FACTOR; + // for the big tensors, we have the option to store the data in 16-bit floats or quantized // in order to save memory and also to speed up the computation wctx.wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype)); @@ -882,6 +877,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con fprintf(stderr, "%s: n_text_layer = %d\n", __func__, hparams.n_text_layer); fprintf(stderr, "%s: n_mels = %d\n", __func__, hparams.n_mels); fprintf(stderr, "%s: ftype = %d\n", __func__, model.hparams.ftype); + fprintf(stderr, "%s: qntvr = %d\n", __func__, qntvr); fprintf(stderr, "%s: type = %d\n", __func__, model.type); // print memory requirements @@ -1106,7 +1102,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con ctx_size += n_text_layer*( n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // cross_attn_ln_1_b } - ctx_size += (15 + 15*n_audio_layer + 24*n_text_layer)*256; // object overhead + ctx_size += (15 + 15*n_audio_layer + 24*n_text_layer)*512; // object overhead fprintf(stderr, "%s: model ctx = %7.2f MB\n", __func__, ctx_size/(1024.0*1024.0)); } @@ -1554,14 +1550,14 @@ static bool whisper_encode_internal( Qcur), Qcur); - //Qcur = ggml_scale(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25))); + //Qcur = ggml_scale_inplace(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25))); // note: no bias for Key struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, layer.attn_k_w, cur); - //Kcur = ggml_scale(ctx0, Kcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25))); + //Kcur = ggml_scale_inplace(ctx0, Kcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25))); struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, layer.attn_v_w, @@ -1621,12 +1617,12 @@ static bool whisper_encode_internal( struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); struct ggml_tensor * KQ_scaled = - ggml_scale(ctx0, + ggml_scale_inplace(ctx0, KQ, ggml_new_f32(ctx0, 1.0f/sqrt(float(n_state)/n_head)) ); - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_scaled); + struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_scaled); struct ggml_tensor * V = ggml_cpy(ctx0, @@ -1809,7 +1805,7 @@ static bool whisper_encode_internal( layer.cross_attn_k_w, cur); - Kcross = ggml_scale(ctx0, Kcross, ggml_new_f32(ctx0, pow(float(n_state) / n_head, -0.25))); + Kcross = ggml_scale_inplace(ctx0, Kcross, ggml_new_f32(ctx0, pow(float(n_state) / n_head, -0.25))); wstate.use_buf(ctx0, 1); @@ -1956,14 +1952,14 @@ static bool whisper_decode_internal( Qcur), Qcur); - Qcur = ggml_scale(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25))); + Qcur = ggml_scale_inplace(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25))); // note: no bias for Key struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, layer.attn_k_w, cur); - Kcur = ggml_scale(ctx0, Kcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25))); + Kcur = ggml_scale_inplace(ctx0, Kcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25))); // store key and value to memory { @@ -2012,14 +2008,14 @@ static bool whisper_decode_internal( struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); //struct ggml_tensor * KQ_scaled = - // ggml_scale(ctx0, + // ggml_scale_inplace(ctx0, // KQ, // ggml_new_f32(ctx0, 1.0f/sqrt(float(n_state)/n_head)) // ); - struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ, n_past); + struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ, n_past); - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); + struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); struct ggml_tensor * V = ggml_view_3d(ctx0, kv_self.v, @@ -2083,7 +2079,7 @@ static bool whisper_decode_internal( Qcur), Qcur); - Qcur = ggml_scale(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25))); + Qcur = ggml_scale_inplace(ctx0, Qcur, ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25))); // Kcross is already scaled struct ggml_tensor * Kcross = @@ -2123,15 +2119,15 @@ static bool whisper_decode_internal( struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); //struct ggml_tensor * KQ_scaled = - // ggml_scale(ctx0, + // ggml_scale_inplace(ctx0, // KQ, // ggml_new_f32(ctx0, 1.0f/sqrt(float(n_state)/n_head)) // ); // no masking for cross-attention - //struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past); + //struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past); - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ); + struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ); struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); @@ -4903,7 +4899,7 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) { // b: N*N*sizeof(float) // c: N*N*sizeof(float) // when F16 is used, there is an extra work buffer of size N*N*sizeof(float) - std::vector buf(4llu*N_max*N_max*sizeof(float) + 4*256); + std::vector buf(4llu*N_max*N_max*sizeof(float) + 4*512); // put a bunch of random data in the buffer for (size_t i = 0; i < buf.size(); i++) buf[i] = i; @@ -4911,7 +4907,6 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) { for (int j = 0; j < (int) sizes.size(); j++) { int n_q4_0 = 0; int n_q4_1 = 0; - int n_q4_2 = 0; int n_q5_0 = 0; int n_q5_1 = 0; int n_q8_0 = 0; @@ -4921,7 +4916,6 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) { // GFLOPS/s double s_q4_0 = 0.0; double s_q4_1 = 0.0; - double s_q4_2 = 0.0; double s_q5_0 = 0.0; double s_q5_1 = 0.0; double s_q8_0 = 0.0; @@ -4930,18 +4924,17 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) { const size_t N = sizes[j]; - for (int k = 0; k < 8; ++k) { + for (int k = 0; k < 7; ++k) { const ggml_type wtype = k == 0 ? GGML_TYPE_Q4_0 : k == 1 ? GGML_TYPE_Q4_1 : - k == 2 ? GGML_TYPE_Q4_2 : - k == 3 ? GGML_TYPE_Q5_0 : - k == 4 ? GGML_TYPE_Q5_1 : - k == 5 ? GGML_TYPE_Q8_0 : - k == 6 ? GGML_TYPE_F16 : GGML_TYPE_F32; + k == 2 ? GGML_TYPE_Q5_0 : + k == 3 ? GGML_TYPE_Q5_1 : + k == 4 ? GGML_TYPE_Q8_0 : + k == 5 ? GGML_TYPE_F16 : GGML_TYPE_F32; - double & s = k == 0 ? s_q4_0 : k == 1 ? s_q4_1 : k == 2 ? s_q4_2 : k == 3 ? s_q5_0 : k == 4 ? s_q5_1 : k == 5 ? s_q8_0 : k == 6 ? s_fp16 : /*k == 7*/ s_fp32; - int & n = k == 0 ? n_q4_0 : k == 1 ? n_q4_1 : k == 2 ? n_q4_2 : k == 3 ? n_q5_0 : k == 4 ? n_q5_1 : k == 5 ? n_q8_0 : k == 6 ? n_fp16 : /*k == 7*/ n_fp32; + double & s = k == 0 ? s_q4_0 : k == 1 ? s_q4_1 : k == 2 ? s_q5_0 : k == 3 ? s_q5_1 : k == 4 ? s_q8_0 : k == 5 ? s_fp16 : /*k == 6*/ s_fp32; + int & n = k == 0 ? n_q4_0 : k == 1 ? n_q4_1 : k == 2 ? n_q5_0 : k == 3 ? n_q5_1 : k == 4 ? n_q8_0 : k == 5 ? n_fp16 : /*k == 6*/ n_fp32; struct ggml_init_params gparams = { /*.mem_size =*/ buf.size(), @@ -4985,9 +4978,9 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) { s = ((2.0*N*N*N*n)/tsum)*1e-9; } - // Q4_0 | Q4_1 | Q4_2 - snprintf(strbuf, sizeof(strbuf), "%4zu x %4zu: Q4_0 %7.1f GFLOPS (%3d runs) | Q4_1 %7.1f GFLOPS (%3d runs) | Q4_2 %7.1f GFLOPS (%3d runs)\n", - N, N, s_q4_0, n_q4_0, s_q4_1, n_q4_1, s_q4_2, n_q4_2); + // Q4_0 | Q4_1 + snprintf(strbuf, sizeof(strbuf), "%4zu x %4zu: Q4_0 %7.1f GFLOPS (%3d runs) | Q4_1 %7.1f GFLOPS (%3d runs)\n", + N, N, s_q4_0, n_q4_0, s_q4_1, n_q4_1); s += strbuf; // Q5_0 | Q5_1 | Q8_0