1 // This file is part of OpenCV project.
2 // It is subject to the license terms in the LICENSE file found in the top-level directory
3 // of this distribution and at http://opencv.org/license.html.
4
5 #include <cuda_runtime.h>
6 #include <cuda_fp16.h>
7
8 #include "grid_stride_range.hpp"
9 #include "execution.hpp"
10 #include "vector_traits.hpp"
11
12 #include "../cuda4dnn/csl/stream.hpp"
13 #include "../cuda4dnn/csl/span.hpp"
14 #include "../cuda4dnn/csl/tensor.hpp"
15
16 #include <opencv2/core.hpp>
17
18 using namespace cv::dnn::cuda4dnn::csl;
19 using namespace cv::dnn::cuda4dnn::csl::device;
20
21 namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
22
23 namespace raw {
24 template <class T, std::size_t N>
input_shortcut_vec(Span<T> output,View<T> input,index_type c_input,View<T> from,index_type c_from,size_type channel_stride)25 __global__ void input_shortcut_vec(
26 Span<T> output,
27 View<T> input, index_type c_input, /* `c_input` = number of channels in `input` */
28 View<T> from, index_type c_from, /* `c_from` = number of channels in `from` */
29 size_type channel_stride /* common for both `input` and `from` */)
30 {
31 using vector_type = get_vector_type_t<T, N>;
32
33 auto output_vPtr = vector_type::get_pointer(output.data());
34 auto input_vPtr = vector_type::get_pointer(input.data());
35 auto from_vPtr = vector_type::get_pointer(from.data());
36
37 auto batch_stride_input = c_input * channel_stride;
38 auto batch_stride_from = c_from * channel_stride;
39
40 for (auto i : grid_stride_range(output.size() / vector_type::size())) {
41 const auto actual_idx = i * vector_type::size();
42 const auto b = actual_idx / batch_stride_input; /* `input` and `output` have the same shape */
43 const auto c = (actual_idx % batch_stride_input) / channel_stride;
44 const auto c_offset = actual_idx % channel_stride;
45
46 vector_type vec_input;
47 v_load(vec_input, input_vPtr[i]);
48
49 /* We can break down the shortcut operation into two steps:
50 * - copy `input` to `output`
51 * - add `from` to corresponding channels in `output`
52 *
53 * In this scheme, only some channels in the `output` differ from `input`. They differ in the channels
54 * which have a corresponding channel in `from`.
55 */
56 if (c < c_from) {
57 const auto from_actual_idx = b * batch_stride_from + c * channel_stride + c_offset;
58 const auto from_vec_idx = from_actual_idx / vector_type::size();
59
60 vector_type vec_from;
61 v_load(vec_from, from_vPtr[from_vec_idx]);
62 for (int j = 0; j < vector_type::size(); j++)
63 vec_input.data[j] += vec_from.data[j];
64 }
65
66 v_store(output_vPtr[i], vec_input);
67 }
68 }
69 }
70
71 template <class T, std::size_t N>
launch_vectorized_input_shortcut(const Stream & stream,Span<T> output,View<T> input,std::size_t c_input,View<T> from,std::size_t c_from,std::size_t channel_stride)72 void launch_vectorized_input_shortcut(const Stream& stream, Span<T> output, View<T> input, std::size_t c_input, View<T> from, std::size_t c_from, std::size_t channel_stride) {
73 CV_Assert(is_fully_aligned<T>(output, N));
74 CV_Assert(is_fully_aligned<T>(input, N));
75 CV_Assert(is_fully_aligned<T>(from, N));
76 CV_Assert(channel_stride % N == 0);
77
78 auto kernel = raw::input_shortcut_vec<T, N>;
79 auto policy = make_policy(kernel, output.size() / N, 0, stream);
80 launch_kernel(kernel, policy, output, input, c_input, from, c_from, channel_stride);
81 }
82
83 template <class T>
input_shortcut(const csl::Stream & stream,csl::TensorSpan<T> output,csl::TensorView<T> input,csl::TensorView<T> from)84 void input_shortcut(const csl::Stream& stream, csl::TensorSpan<T> output, csl::TensorView<T> input, csl::TensorView<T> from) {
85 CV_Assert(is_shape_same(output, input));
86 CV_Assert(output.rank() == from.rank());
87 for (int i = 0; i < output.rank(); i++) {
88 if (i != 1) {
89 CV_Assert(from.get_axis_size(i) == output.get_axis_size(i));
90 }
91 }
92
93 auto channel_stride = output.size_range(2, output.rank()); /* same for `output`, `input` and `from` */
94 auto c_input = input.get_axis_size(1);
95 auto c_from = from.get_axis_size(1);
96
97 if (is_fully_aligned<T>(output, 4) && is_fully_aligned<T>(input, 4) && is_fully_aligned<T>(from, 4) && channel_stride % 4 == 0) {
98 launch_vectorized_input_shortcut<T, 4>(stream, output, input, c_input, from, c_from, channel_stride);
99 } else if (is_fully_aligned<T>(output, 2) && is_fully_aligned<T>(input, 2) && is_fully_aligned<T>(from, 2) && channel_stride % 2 == 0) {
100 launch_vectorized_input_shortcut<T, 2>(stream, output, input, c_input, from, c_from, channel_stride);
101 } else {
102 launch_vectorized_input_shortcut<T, 1>(stream, output, input, c_input, from, c_from, channel_stride);
103 }
104 }
105
106 #if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
107 template void input_shortcut(const Stream&, TensorSpan<__half>, TensorView<__half>, TensorView<__half>);
108 #endif
109 template void input_shortcut(const Stream&, TensorSpan<float>, TensorView<float>, TensorView<float>);
110
111 }}}} /* namespace cv::dnn::cuda4dnn::kernels */
112