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 shuffle.cpp
18/// > Annotated version: @ref shuffle_example_cpp
19///
20/// @page shuffle_example_cpp_short
21///
22/// This C++ API example demonstrates how to create and execute a
23/// [Shuffle](@ref dev_guide_shuffle) primitive.
24///
25/// Key optimizations included in this example:
26/// - Shuffle along axis 1 (channels).
27///
28/// @page shuffle_example_cpp Shuffle Primitive Example
29/// @copydetails shuffle_example_cpp_short
30///
31/// @include shuffle.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
42using namespace dnnl;
43
44using tag = memory::format_tag;
45using dt = memory::data_type;
46
47void shuffle_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 N = 3, // batch size
57 IC = 72, // channels
58 IH = 227, // tensor height
59 IW = 227; // tensor width
60
61 // Source (src) and destination (dst) tensors dimensions.
62 memory::dims src_dims = {N, IC, IH, IW};
63
64 // Allocate buffers.
65 std::vector<float> src_data(product(src_dims));
66 std::vector<float> dst_data(product(src_dims));
67
68 // Initialize src.
69 std::generate(src_data.begin(), src_data.end(), []() {
70 static int i = 0;
71 return std::cos(i++ / 10.f);
72 });
73
74 // Shuffle axis and group size.
75 const int shuffle_axis = 1;
76 const int group_size = 4;
77
78 // Create memory descriptor and memory objects for src and dst.
79 auto src_md = memory::desc(src_dims, dt::f32, tag::nchw);
80 auto dst_md = memory::desc(src_dims, dt::f32, tag::nchw);
81 auto src_mem = memory(src_md, engine);
82
83 auto dst_mem = memory({src_dims, dt::f32, tag::abcd}, engine);
84
85 // Write data to memory object's handle.
86 write_to_dnnl_memory(src_data.data(), src_mem);
87
88 // Create primitive descriptor.
89 auto shuffle_pd = shuffle_forward::primitive_desc(engine,
90 prop_kind::forward_training, src_md, dst_md, shuffle_axis,
91 group_size);
92
93 // Create the primitive.
94 auto shuffle_prim = shuffle_forward(shuffle_pd);
95
96 // Primitive arguments.
97 std::unordered_map<int, memory> shuffle_args;
98 shuffle_args.insert({DNNL_ARG_SRC, src_mem});
99 shuffle_args.insert({DNNL_ARG_DST, dst_mem});
100
101 // Primitive execution: shuffle.
102 shuffle_prim.execute(engine_stream, shuffle_args);
103
104 // Wait for the computation to finalize.
105 engine_stream.wait();
106
107 // Read data from memory object.
108 read_from_dnnl_memory(dst_data.data(), dst_mem);
109}
110
111int main(int argc, char **argv) {
112 return handle_example_errors(
113 shuffle_example, parse_engine_kind(argc, argv));
114}
115