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 | /// @example matmul.cpp |
18 | /// > Annotated version: @ref matmul_example_cpp |
19 | /// |
20 | /// @page matmul_example_cpp_short |
21 | /// |
22 | /// This C++ API example demonstrates how to create and execute a |
23 | /// [MatMul](@ref dev_guide_matmul) primitive. |
24 | /// |
25 | /// Key optimizations included in this example: |
26 | /// - Primitive attributes with fused post-ops. |
27 | /// |
28 | /// @page matmul_example_cpp Matmul Primitive Example |
29 | /// @copydetails matmul_example_cpp_short |
30 | /// |
31 | /// @include matmul.cpp |
32 | |
33 | #include <algorithm> |
34 | #include <cmath> |
35 | #include <iostream> |
36 | #include <string> |
37 | #include <vector> |
38 | |
39 | #include "example_utils.hpp" |
40 | #include "oneapi/dnnl/dnnl.hpp" |
41 | |
42 | using namespace dnnl; |
43 | |
44 | using tag = memory::format_tag; |
45 | using dt = memory::data_type; |
46 | |
47 | void matmul_example(dnnl::engine::kind engine_kind) { |
48 | |
49 | // Create execution dnnl::engine. |
50 | dnnl::engine engine(engine_kind, 0); |
51 | |
52 | // Create dnnl::stream. |
53 | dnnl::stream engine_stream(engine); |
54 | |
55 | // Tensor dimensions. |
56 | const memory::dim MB = 3, // batch size |
57 | M = 128, K = 256, N = 512; |
58 | |
59 | // Source (src), weights, bias, and destination (dst) tensors dimensions. |
60 | memory::dims src_dims = {MB, M, K}; |
61 | memory::dims weights_dims = {MB, K, N}; |
62 | memory::dims bias_dims = {1, 1, N}; |
63 | memory::dims dst_dims = {MB, M, N}; |
64 | |
65 | // Allocate buffers. |
66 | std::vector<float> src_data(product(src_dims)); |
67 | std::vector<float> weights_data(product(weights_dims)); |
68 | std::vector<float> bias_data(product(bias_dims)); |
69 | std::vector<float> dst_data(product(dst_dims)); |
70 | |
71 | // Initialize src, weights, bias. |
72 | std::generate(src_data.begin(), src_data.end(), []() { |
73 | static int i = 0; |
74 | return std::cos(i++ / 10.f); |
75 | }); |
76 | std::generate(weights_data.begin(), weights_data.end(), []() { |
77 | static int i = 0; |
78 | return std::sin(i++ * 2.f); |
79 | }); |
80 | std::generate(bias_data.begin(), bias_data.end(), []() { |
81 | static int i = 0; |
82 | return std::tanh(float(i++)); |
83 | }); |
84 | |
85 | // Create memory descriptors and memory objects for src, weights, bias, and |
86 | // dst. |
87 | auto src_md = memory::desc(src_dims, dt::f32, tag::abc); |
88 | auto weights_md = memory::desc(weights_dims, dt::f32, tag::abc); |
89 | auto bias_md = memory::desc(bias_dims, dt::f32, tag::abc); |
90 | auto dst_md = memory::desc(dst_dims, dt::f32, tag::abc); |
91 | |
92 | auto src_mem = memory(src_md, engine); |
93 | auto weights_mem = memory(weights_md, engine); |
94 | auto bias_mem = memory(bias_md, engine); |
95 | auto dst_mem = memory(dst_md, engine); |
96 | |
97 | // Write data to memory object's handles. |
98 | write_to_dnnl_memory(src_data.data(), src_mem); |
99 | write_to_dnnl_memory(weights_data.data(), weights_mem); |
100 | write_to_dnnl_memory(bias_data.data(), bias_mem); |
101 | |
102 | // Create primitive post-ops (ReLU). |
103 | const float alpha = 0.f; |
104 | const float beta = 0.f; |
105 | post_ops matmul_ops; |
106 | matmul_ops.append_eltwise(algorithm::eltwise_relu, alpha, beta); |
107 | primitive_attr matmul_attr; |
108 | matmul_attr.set_post_ops(matmul_ops); |
109 | |
110 | // Create primitive descriptor. |
111 | auto matmul_pd = matmul::primitive_desc( |
112 | engine, src_md, weights_md, bias_md, dst_md, matmul_attr); |
113 | |
114 | // Create the primitive. |
115 | auto matmul_prim = matmul(matmul_pd); |
116 | |
117 | // Primitive arguments. |
118 | std::unordered_map<int, memory> matmul_args; |
119 | matmul_args.insert({DNNL_ARG_SRC, src_mem}); |
120 | matmul_args.insert({DNNL_ARG_WEIGHTS, weights_mem}); |
121 | matmul_args.insert({DNNL_ARG_BIAS, bias_mem}); |
122 | matmul_args.insert({DNNL_ARG_DST, dst_mem}); |
123 | |
124 | // Primitive execution: matrix multiplication with ReLU. |
125 | matmul_prim.execute(engine_stream, matmul_args); |
126 | |
127 | // Wait for the computation to finalize. |
128 | engine_stream.wait(); |
129 | |
130 | // Read data from memory object's handle. |
131 | read_from_dnnl_memory(dst_data.data(), dst_mem); |
132 | } |
133 | |
134 | int main(int argc, char **argv) { |
135 | return handle_example_errors(matmul_example, parse_engine_kind(argc, argv)); |
136 | } |
137 | |