1 /*******************************************************************************
2 * Copyright 2020 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 sum.cpp
18 /// > Annotated version: @ref sum_example_cpp
19 ///
20 /// @page sum_example_cpp_short
21 ///
22 /// This C++ API example demonstrates how to create and execute a
23 /// [Sum](@ref dev_guide_sum) primitive.
24 ///
25 /// Key optimizations included in this example:
26 /// - Identical memory formats for source (src) and destination (dst) tensors.
27 ///
28 /// @page sum_example_cpp Sum Primitive Example
29 /// @copydetails sum_example_cpp_short
30 ///
31 /// @include sum.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 
sum_example(dnnl::engine::kind engine_kind)47 void sum_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 = 3, // 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     // Number of src tensors.
75     const int num_src = 10;
76 
77     // Scaling factors.
78     std::vector<float> scales(num_src);
79     std::generate(
80             scales.begin(), scales.end(), [](int n = 0) { return sin(n); });
81 
82     // Create an array of memory descriptors and memory objects for src tensors.
83     std::vector<memory::desc> src_md;
84     std::vector<memory> src_mem;
85 
86     for (int n = 0; n < num_src; ++n) {
87         auto md = memory::desc(src_dims, dt::f32, tag::nchw);
88         auto mem = memory(md, engine);
89 
90         // Write data to memory object's handle.
91         write_to_dnnl_memory(src_data.data(), mem);
92 
93         src_md.push_back(md);
94         src_mem.push_back(mem);
95     }
96 
97     // Create primitive descriptor.
98     auto sum_pd = sum::primitive_desc(scales, src_md, engine);
99 
100     // Create the primitive.
101     auto sum_prim = sum(sum_pd);
102 
103     // Create memory object for dst.
104     auto dst_mem = memory(sum_pd.dst_desc(), engine);
105 
106     // Primitive arguments.
107     std::unordered_map<int, memory> sum_args;
108     sum_args.insert({DNNL_ARG_DST, dst_mem});
109     for (int n = 0; n < num_src; ++n) {
110         sum_args.insert({DNNL_ARG_MULTIPLE_SRC + n, src_mem[n]});
111     }
112 
113     // Primitive execution: sum.
114     sum_prim.execute(engine_stream, sum_args);
115 
116     // Wait for the computation to finalize.
117     engine_stream.wait();
118 
119     // Read data from memory object's handle.
120     read_from_dnnl_memory(dst_data.data(), dst_mem);
121 }
122 
main(int argc,char ** argv)123 int main(int argc, char **argv) {
124     return handle_example_errors(sum_example, parse_engine_kind(argc, argv));
125 }
126