1 | /******************************************************************************* |
---|---|
2 | * Copyright 2020-2022 Intel Corporation |
3 | * |
4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
5 | * you may not use this file except in compliance with the License. |
6 | * You may obtain a copy of the License at |
7 | * |
8 | * http://www.apache.org/licenses/LICENSE-2.0 |
9 | * |
10 | * Unless required by applicable law or agreed to in writing, software |
11 | * distributed under the License is distributed on an "AS IS" BASIS, |
12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
13 | * See the License for the specific language governing permissions and |
14 | * limitations under the License. |
15 | *******************************************************************************/ |
16 | |
17 | #ifndef GPU_GEMM_GPU_GEMM_HPP |
18 | #define GPU_GEMM_GPU_GEMM_HPP |
19 | |
20 | #include "common/c_types_map.hpp" |
21 | #include "common/primitive.hpp" |
22 | #include "gpu/gemm/gpu_gemm_exec_types.hpp" |
23 | #include "gpu/gpu_primitive.hpp" |
24 | |
25 | namespace dnnl { |
26 | namespace impl { |
27 | namespace gpu { |
28 | |
29 | struct gpu_gemm_t : public gpu_primitive_t { |
30 | using gpu_primitive_t::gpu_primitive_t; |
31 | virtual status_t execute(const gemm_exec_ctx_t &ctx) const = 0; |
32 | status_t execute(const exec_ctx_t &ctx) const override { |
33 | gemm_exec_args_t gemm_args; |
34 | // TODO: we have to swap a and b because |
35 | // - gemm primitive is created with row major desc, |
36 | // - parameters to gemm are passed as row major |
37 | // - but gemm implementation assumes column major |
38 | gemm_args.a = &CTX_IN_STORAGE(DNNL_ARG_B); |
39 | gemm_args.b = &CTX_IN_STORAGE(DNNL_ARG_A); |
40 | gemm_args.c = &CTX_OUT_STORAGE(DNNL_ARG_C); |
41 | gemm_args.a_zero_point |
42 | = &CTX_IN_STORAGE(DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_B); |
43 | gemm_args.b_zero_point |
44 | = &CTX_IN_STORAGE(DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_A); |
45 | gemm_args.c_zero_point |
46 | = &CTX_IN_STORAGE(DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_C); |
47 | gemm_args.a_scales = &CTX_IN_STORAGE(DNNL_ARG_ATTR_SCALES | DNNL_ARG_B); |
48 | gemm_args.b_scales = &CTX_IN_STORAGE(DNNL_ARG_ATTR_SCALES | DNNL_ARG_A); |
49 | gemm_args.c_scales = &CTX_IN_STORAGE(DNNL_ARG_ATTR_SCALES | DNNL_ARG_C); |
50 | gemm_exec_ctx_t gemm_ctx(ctx, gemm_args); |
51 | return execute(gemm_ctx); |
52 | } |
53 | }; |
54 | |
55 | inline const gpu_gemm_t *gpu_gemm(const std::shared_ptr<primitive_t> &p) { |
56 | return utils::downcast<gpu_gemm_t *>(p.get()); |
57 | } |
58 | |
59 | } // namespace gpu |
60 | } // namespace impl |
61 | } // namespace dnnl |
62 | |
63 | #endif |
64 |