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 batch_normalization.cpp
18/// > Annotated version: @ref batch_normalization_example_cpp
19///
20/// @page batch_normalization_example_cpp_short
21///
22/// This C++ API example demonstrates how to create and execute a
23/// [Batch Normalization](@ref dev_guide_batch_normalization) primitive in
24/// forward training propagation mode.
25///
26/// Key optimizations included in this example:
27/// - In-place primitive execution;
28/// - Source memory format for an optimized primitive implementation;
29/// - Fused post-ops via operation descriptor flags;
30///
31/// @page batch_normalization_example_cpp Batch Normalization Primitive Example
32/// @copydetails batch_normalization_example_cpp_short
33///
34/// @include batch_normalization.cpp
35
36#include <algorithm>
37#include <cmath>
38#include <iostream>
39#include <string>
40#include <vector>
41
42#include "example_utils.hpp"
43#include "oneapi/dnnl/dnnl.hpp"
44
45using namespace dnnl;
46
47using tag = memory::format_tag;
48using dt = memory::data_type;
49
50void batch_normalization_example(dnnl::engine::kind engine_kind) {
51
52 // Create execution dnnl::engine.
53 dnnl::engine engine(engine_kind, 0);
54
55 // Create dnnl::stream.
56 dnnl::stream engine_stream(engine);
57
58 // Tensor dimensions.
59 const memory::dim N = 3, // batch size
60 IC = 3, // channels
61 IH = 227, // tensor height
62 IW = 227; // tensor width
63
64 // Source (src) and destination (dst) tensors dimensions.
65 memory::dims src_dims = {N, IC, IH, IW};
66
67 // Scale/shift tensor dimensions.
68 memory::dims scaleshift_dims = {IC};
69
70 // Allocate buffers.
71 std::vector<float> src_data(product(src_dims));
72 std::vector<float> scale_data(product(scaleshift_dims));
73 std::vector<float> shift_data(product(scaleshift_dims));
74
75 // Initialize src.
76 std::generate(src_data.begin(), src_data.end(), []() {
77 static int i = 0;
78 return std::cos(i++ / 10.f);
79 });
80
81 // Initialize scale.
82 std::generate(scale_data.begin(), scale_data.end(), []() {
83 static int i = 0;
84 return std::sin(i++ * 2.f);
85 });
86
87 // Initialize shift.
88 std::generate(shift_data.begin(), shift_data.end(), []() {
89 static int i = 0;
90 return std::tan(float(i++));
91 });
92
93 // Create src and scale/shift memory descriptors and memory objects.
94 auto src_md = memory::desc(src_dims, dt::f32, tag::nchw);
95 auto dst_md = memory::desc(src_dims, dt::f32, tag::nchw);
96 auto scaleshift_md = memory::desc(scaleshift_dims, dt::f32, tag::x);
97
98 auto src_mem = memory(src_md, engine);
99 auto scale_mem = memory(scaleshift_md, engine);
100 auto shift_mem = memory(scaleshift_md, engine);
101
102 // Write data to memory object's handle.
103 write_to_dnnl_memory(src_data.data(), src_mem);
104 write_to_dnnl_memory(scale_data.data(), scale_mem);
105 write_to_dnnl_memory(shift_data.data(), shift_mem);
106
107 // Create primitive descriptor.
108 auto bnorm_pd = batch_normalization_forward::primitive_desc(engine,
109 prop_kind::forward_training, src_md, dst_md, 1.e-10f,
110 normalization_flags::use_scale | normalization_flags::use_shift
111 | normalization_flags::fuse_norm_relu);
112
113 // Create memory objects using memory descriptors created by the primitive
114 // descriptor: mean, variance, workspace.
115 // NOTE: Here, the ReLU post-ops require a workspace for later usage in
116 // backward propagation mode.
117 auto mean_mem = memory(bnorm_pd.mean_desc(), engine);
118 auto variance_mem = memory(bnorm_pd.variance_desc(), engine);
119 auto workspace_mem = memory(bnorm_pd.workspace_desc(), engine);
120
121 // Create the primitive.
122 auto bnorm_prim = batch_normalization_forward(bnorm_pd);
123
124 // Primitive arguments. Set up in-place execution by assigning src as DST.
125 std::unordered_map<int, memory> bnorm_args;
126 bnorm_args.insert({DNNL_ARG_SRC, src_mem});
127 bnorm_args.insert({DNNL_ARG_MEAN, mean_mem});
128 bnorm_args.insert({DNNL_ARG_VARIANCE, variance_mem});
129 bnorm_args.insert({DNNL_ARG_SCALE, scale_mem});
130 bnorm_args.insert({DNNL_ARG_SHIFT, shift_mem});
131 bnorm_args.insert({DNNL_ARG_WORKSPACE, workspace_mem});
132 bnorm_args.insert({DNNL_ARG_DST, src_mem});
133
134 // Primitive execution: batch normalization with ReLU.
135 bnorm_prim.execute(engine_stream, bnorm_args);
136
137 // Wait for the computation to finalize.
138 engine_stream.wait();
139
140 // Read data from memory object's handle.
141 read_from_dnnl_memory(src_data.data(), src_mem);
142}
143
144int main(int argc, char **argv) {
145 return handle_example_errors(
146 batch_normalization_example, parse_engine_kind(argc, argv));
147}
148