From 5f66ebca9c41a17385341da4b553a8eb5f07edee Mon Sep 17 00:00:00 2001 From: Guillaume Wenzek Date: Fri, 29 Dec 2023 18:07:03 +0100 Subject: [PATCH] ggml : extend ggml_get_rows, ggml_repeat, ggml_concat (ggml/639) * add more int ops * ggml_compute_forward_dup_bytes * add tests * PR comments * tests : minor indentations --------- Co-authored-by: Georgi Gerganov --- ggml.c | 166 ++++++++++++++++++++++++++++++++++++- tests/test-backend-ops.cpp | 42 ++++++++-- 2 files changed, 198 insertions(+), 10 deletions(-) diff --git a/ggml.c b/ggml.c index bcec200f6..b124f14cc 100644 --- a/ggml.c +++ b/ggml.c @@ -4766,8 +4766,11 @@ struct ggml_tensor * ggml_get_rows( } // 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_4d(ctx, GGML_TYPE_F32, a->ne[0], b->ne[0], b->ne[1], b->ne[2]); + enum ggml_type type = GGML_TYPE_F32; + if (a->type == GGML_TYPE_I32) { + type = a->type; + } + struct ggml_tensor * result = ggml_new_tensor_4d(ctx, type, a->ne[0], b->ne[0], b->ne[1], b->ne[2]); result->op = GGML_OP_GET_ROWS; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; @@ -6938,14 +6941,165 @@ static void ggml_compute_forward_dup_f32( } } +// A simplified version of ggml_compute_forward_dup that doesn't do float upcasting, and just plain old memcpy. +static void ggml_compute_forward_dup_bytes( + 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(src0->type == dst->type); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) { + ggml_compute_forward_dup_same_cont(params, src0, dst); + return; + } + + GGML_TENSOR_UNARY_OP_LOCALS; + + const size_t type_size = ggml_type_size(src0->type); + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + + // parallelize by rows + const int nr = ne01; + // number of 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 (src0->type == dst->type && + ne00 == ne0 && + nb00 == type_size && nb0 == type_size) { + // copy by rows + const size_t rs = ne00 * type_size; + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ir0; i01 < ir1; i01++) { + memcpy( + ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), + rs); + } + } + } + return; + } + + if (ggml_is_contiguous(dst)) { + size_t id = 0; + char * dst_ptr = (char *) dst->data; + const size_t rs = ne00 * type_size; + + if (nb00 == type_size) { + // src0 is contigous on first dimension, copy by rows + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int64_t i01 = ir0; i01 < ir1; i01++) { + const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, rs); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else { + //printf("%s: this is not optimal - fix me\n", __func__); + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = (char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, type_size); + + id += type_size; + } + } + id += rs * (ne01 - ir1); + } + } + } + + return; + } + + // dst counters + + int64_t i10 = 0; + int64_t i11 = 0; + int64_t i12 = 0; + int64_t i13 = 0; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + memcpy(dst_ptr, src0_ptr, type_size); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } +} + 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); + if (src0->type == dst->type) { + ggml_compute_forward_dup_bytes(params, src0, dst); return; } + switch (src0->type) { case GGML_TYPE_F16: { @@ -8404,10 +8558,12 @@ static void ggml_compute_forward_repeat( struct ggml_tensor * dst) { switch (src0->type) { case GGML_TYPE_F16: + case GGML_TYPE_I16: { ggml_compute_forward_repeat_f16(params, src0, dst); } break; case GGML_TYPE_F32: + case GGML_TYPE_I32: { ggml_compute_forward_repeat_f32(params, src0, dst); } break; @@ -8550,6 +8706,7 @@ static void ggml_compute_forward_concat( struct ggml_tensor* dst) { switch (src0->type) { case GGML_TYPE_F32: + case GGML_TYPE_I32: { ggml_compute_forward_concat_f32(params, src0, src1, dst); } break; @@ -10674,6 +10831,7 @@ static void ggml_compute_forward_get_rows( ggml_compute_forward_get_rows_f16(params, src0, src1, dst); } break; case GGML_TYPE_F32: + case GGML_TYPE_I32: { ggml_compute_forward_get_rows_f32(params, src0, src1, dst); } break; diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index eff063b2d..44412cb94 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -58,6 +58,9 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m int64_t hist[16]; ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size, hist); ggml_backend_tensor_set(tensor, dataq.data(), 0, dataq.size()); + } else if (tensor->type == GGML_TYPE_I8 || tensor->type == GGML_TYPE_I16 || tensor->type == GGML_TYPE_I32) { + // This is going to create some weird integers though. + ggml_backend_tensor_set(tensor, data.data(), 0, ggml_nbytes(tensor)); } else { GGML_ASSERT(false); } @@ -87,8 +90,13 @@ static std::vector tensor_to_float(const ggml_tensor * t) { tv.push_back(*(float *) &buf[i]); } else if (t->type == GGML_TYPE_I32) { tv.push_back((float)*(int32_t *) &buf[i]); + } else if (t->type == GGML_TYPE_I16) { + tv.push_back((float)*(int16_t *) &buf[i]); + } else if (t->type == GGML_TYPE_I8) { + tv.push_back((float)*(int8_t *) &buf[i]); } else if (quantized) { - tt.to_float(&buf[i], vq.data(), bs); + std::vector vq(ggml_blck_size(t->type)); + tt.to_float(&buf[i], vq.data(), ggml_blck_size(t->type)); tv.insert(tv.end(), vq.begin(), vq.end()); } else { GGML_ASSERT(false); @@ -661,17 +669,26 @@ struct test_repeat : public test_case { struct test_dup : public test_case { const ggml_type type; const std::array ne; + const std::array permute; + bool _use_permute; std::string vars() override { - return VARS_TO_STR2(type, ne); + std::string v = VARS_TO_STR2(type, ne); + if (_use_permute) v += "," + VAR_TO_STR(permute); + return v; } test_dup(ggml_type type = GGML_TYPE_F32, - std::array ne = {10, 10, 10, 1}) - : type(type), ne(ne) {} + std::array ne = {10, 10, 10, 1}, + std::array permute = {0, 0, 0, 0}) + : type(type), ne(ne), permute(permute), + _use_permute(permute[0] + permute[1] + permute[2] + permute[3] > 0) {} ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data()); + if (_use_permute) { + src = ggml_permute(ctx, src, permute[0], permute[1], permute[2], permute[3]); + } ggml_tensor * out = ggml_dup(ctx, src); return out; } @@ -1450,14 +1467,26 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op } } } + for (int b : {1, 7}) { + for (bool v : {false, true}) { + test_cases.emplace_back(new test_get_rows(GGML_TYPE_I32, 256, 5, 4, b, v)); + } + } test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 1, 1})); test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {2, 1, 1, 1})); test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 2, 1, 1})); test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 2, 1})); test_cases.emplace_back(new test_repeat(GGML_TYPE_F32, {10, 10, 10, 10}, {1, 1, 1, 2})); + test_cases.emplace_back(new test_repeat(GGML_TYPE_I32, {10, 10, 10, 10}, {2, 1, 1, 1})); + test_cases.emplace_back(new test_repeat(GGML_TYPE_I16, {10, 10, 10, 10}, {1, 1, 1, 2})); - test_cases.emplace_back(new test_dup()); + test_cases.emplace_back(new test_dup(GGML_TYPE_F32)); + test_cases.emplace_back(new test_dup(GGML_TYPE_F16)); + test_cases.emplace_back(new test_dup(GGML_TYPE_I32)); + test_cases.emplace_back(new test_dup(GGML_TYPE_I16)); + test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {0, 2, 1, 3})); + test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {1, 2, 0, 3})); for (ggml_type type : all_types) { test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, type, {256, 10, 10, 1})); @@ -1565,7 +1594,8 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op test_cases.emplace_back(new test_alibi()); test_cases.emplace_back(new test_im2col()); - test_cases.emplace_back(new test_concat()); + test_cases.emplace_back(new test_concat(GGML_TYPE_F32)); + test_cases.emplace_back(new test_concat(GGML_TYPE_I32)); for (ggml_sort_order order : {GGML_SORT_ASC, GGML_SORT_DESC}) { test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {8, 1, 1, 1}, order));