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.
5 #include <cuda_runtime.h>
10 #include "vector_traits.hpp"
11 #include "grid_stride_range.hpp"
12 #include "execution.hpp"
13 #include "kernel_dispatcher.hpp"
15 #include "../cuda4dnn/csl/stream.hpp"
16 #include "../cuda4dnn/csl/tensor.hpp"
17 #include "../cuda4dnn/csl/span.hpp"
22 using namespace cv::dnn::cuda4dnn::csl;
23 using namespace cv::dnn::cuda4dnn::csl::device;
25 namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
28 template <class T, std::size_t N>
29 __global__ void concat_vec(
30 Span<T> output, size_type output_axis_size, index_type output_axis_offset,
31 View<T> input, size_type input_axis_size, size_type concat_size)
33 using vector_type = get_vector_type_t<T, N>;
35 auto output_vPtr = vector_type::get_pointer(output.data());
36 auto input_vPtr = vector_type::get_pointer(input.data());
38 /* we need to copy all the elements of input to some location in the output
39 * we copy blocks of size `total_concat_size` to some location in the output
41 const auto total_concat_size = concat_size * input_axis_size;
43 for (auto in_idx : grid_stride_range(input.size() / vector_type::size())) {
44 const index_type idx = in_idx * vector_type::size();
45 const index_type concat_num = idx / total_concat_size;
46 const index_type concat_index = idx % total_concat_size;
47 const index_type top_index = concat_index +
48 (concat_num * output_axis_size + output_axis_offset) * concat_size;
50 const auto out_idx = top_index / vector_type::size();
53 v_load(vec, input_vPtr[in_idx]);
54 v_store(output_vPtr[out_idx], vec);
58 template <class T, std::size_t Rank>
59 __global__ void concat_with_offsets(
60 Span<T> output, array<size_type, Rank> out_strides, array<index_type, Rank> out_offset,
61 View<T> input, array<size_type, Rank> in_strides)
63 for (auto i : grid_stride_range(input.size())) {
64 index_type in_index = i / in_strides[0];
65 index_type out_index = out_offset[0] + in_index;
66 index_type oidx = out_index * out_strides[0];
67 for (int j = 1; j < Rank; j++) {
68 in_index = (i % in_strides[j - 1]) / in_strides[j];
69 out_index = out_offset[j] + in_index;
70 oidx += out_index * out_strides[j];
73 output[oidx] = input[i];
78 template <class T, std::size_t N> static
79 void launch_vectorized_concat(const Stream& stream,
80 Span<T> output, size_type output_axis_size, index_type output_axis_offset,
81 View<T> input, size_type input_axis_size, size_type concat_size)
83 CV_Assert(is_fully_aligned<T>(output, N));
84 CV_Assert(is_fully_aligned<T>(input, N));
85 /* more assertions are required to fully check for vectorization possiblity; check concat() */
87 auto kernel = raw::concat_vec<T, N>;
88 auto policy = make_policy(kernel, input.size() / N, 0, stream);
89 launch_kernel(kernel, policy, output, output_axis_size, output_axis_offset, input, input_axis_size, concat_size);
95 TensorSpan<T> output, std::size_t output_axis_offset,
96 TensorView<T> input, std::size_t axis)
98 /* let's call the axis of interest as the channel axis for the purpose of the following discussion
99 * even though it can be any axis
101 * for each batch item:
102 * we move all the channels from the input (which together, for a single batch item, is contiguous)
103 * of a batch item to its corresponding contiguous place in the output
105 * for a valid vector operation:
106 * - the size of each copy block must be aligned
107 * - input must be aligned
108 * - all the destination locations in the output must be aligned
110 std::size_t concat_size = output.size_range(axis + 1, output.rank());
112 std::size_t input_axis_size = input.get_axis_size(axis);
113 std::size_t output_axis_size = output.get_axis_size(axis);
115 std::size_t copy_block_size = concat_size * input_axis_size;
116 std::size_t copy_block_stride = concat_size * output_axis_size;
117 std::size_t starting_offset = output_axis_offset * concat_size;
119 /* in a nutshell, all this concat operation does is copy several blocks of size `copy_block_size`
120 * to the output starting from `starting_offset` with blocks in the output strided by `copy_block_stride`
123 bool is_aligned_4 = copy_block_size % 4 == 0 && copy_block_stride % 4 == 0 && starting_offset % 4 == 0;
124 bool is_aligned_2 = copy_block_size % 2 == 0 && copy_block_stride % 2 == 0 && starting_offset % 2 == 0;
126 if (is_fully_aligned<T>(output, 4) && is_fully_aligned<T>(input, 4) && is_aligned_4) {
127 launch_vectorized_concat<T, 4>(stream, output, output_axis_size, output_axis_offset, input, input_axis_size, concat_size);
128 } else if (is_fully_aligned<T>(output, 2) && is_fully_aligned<T>(input, 2) && is_aligned_2) {
129 launch_vectorized_concat<T, 2>(stream, output, output_axis_size, output_axis_offset, input, input_axis_size, concat_size);
131 launch_vectorized_concat<T, 1>(stream, output, output_axis_size, output_axis_offset, input, input_axis_size, concat_size);
135 template void concat<__half>(const Stream&, TensorSpan<__half>, std::size_t, TensorView<__half>, std::size_t);
136 template void concat<float>(const Stream&, TensorSpan<float>, std::size_t, TensorView<float>, std::size_t);
138 template <class T, std::size_t Rank> static
139 void launch_concat_with_offsets(
140 const Stream& stream,
141 Span<T> output, const std::vector<std::size_t>& outStride, const std::vector<std::size_t>& outOffset,
142 View<T> input, const std::vector<std::size_t>& inStride)
144 CV_Assert(outStride.size() == Rank);
145 CV_Assert(outOffset.size() == Rank);
146 CV_Assert(inStride.size() == Rank);
148 array<size_type, Rank> outStride_k, inStride_k;
149 outStride_k.assign(std::begin(outStride), std::end(outStride));
150 inStride_k.assign(std::begin(inStride), std::end(inStride));
152 array<index_type, Rank> outOffset_k;
153 outOffset_k.assign(std::begin(outOffset), std::end(outOffset));
155 auto kernel = raw::concat_with_offsets<T, Rank>;
156 auto policy = make_policy(kernel, input.size(), 0, stream);
157 launch_kernel(kernel, policy, output, outStride_k, outOffset_k, input, inStride_k);
160 GENERATE_KERNEL_DISPATCHER(concat_with_offsets_dispatcher, launch_concat_with_offsets);
163 void concat_with_offsets(
164 const Stream& stream,
165 TensorSpan<T> output, TensorView<T> input,
166 std::vector<std::size_t> offsets)
168 CV_Assert(output.rank() == input.rank());
169 CV_Assert(output.rank() == offsets.size());
171 /* squeezable axes at the begining of both tensors can be eliminated
175 * Suppose an item's indices in the input tensor is [i1, i2, ...]. The indices in the output
176 * tensor will be [i1 + off1, i2 + off2, ...]. The concat operation essentially copies items
177 * from the input tensor to new locations in the output tensor.
179 * If the size of the first axis of the input and output tensor is unity, the input and output
180 * indices for all the elements will be of the form be [0, i2, ...] and [0, i2 + off2, ...]
181 * respectively. The first index does not contribute to the element's address calculation and
182 * hence does nothing apart from eating up few cycles.
184 while (input.get_axis_size(0) == 1 && output.get_axis_size(0) == 1) {
185 CV_Assert(offsets[0] == 0);
189 offsets.erase(std::begin(offsets));
191 CV_Assert(output.rank() == input.rank());
192 CV_Assert(output.rank() == offsets.size());
195 auto inShape = input.shape_as_vector();
196 auto outShape = output.shape_as_vector();
198 /* contiguous axes that undergo full copy can be combined into one axis
202 * Suppose an item's indices in the input tensor is [i1, i2, i3, ...]. Let the first two axes not undergo any
203 * concatenation. The indices in the output tensor will be [i1, i2, i3 + off3, ...].
205 * Each axis in the contiguous axes sequence will add an offset of iN * strideN. In the above example,
206 * the two axes add a total offset of `i1 * stride1 + i2 * stride2`. We can merge the two axes into one axis with
207 * a size of `size1 * size2`. The new offset added will be i12 * stride2` as the kernel iterates through `i12`.
208 * Note that `i12` is actually `(i1 * size2 + i2)` in the original tensor.
210 for (int i = 0; i < inShape.size(); i++) {
211 /* check if axis `i` requires any slicing */
212 if (offsets[i] == 0 && inShape[i] == outShape[i]) {
213 /* loop invariant: `i` is the first axis in the contiguous unsliced axis sequence */
215 int j = i + 1; /* `j` is the axis which we will attempt to merge */
216 while (j < inShape.size() && offsets[j] == 0 && inShape[j] == outShape[j]) {
217 /* `j` axis is also copied fully; merge `i` and `j` */
218 auto new_size = inShape[i] * inShape[j];
219 inShape[i] = new_size;
220 outShape[i] = new_size;
221 offsets[i] = 0; /* redundant */
223 /* delete axis `j` */
224 inShape.erase(std::begin(inShape) + j);
225 outShape.erase(std::begin(outShape) + j);
226 offsets.erase(std::begin(offsets) + j);
228 /* optimizations should not break the invariants */
229 CV_Assert(inShape.size() == outShape.size());
230 CV_Assert(inShape.size() == offsets.size());
231 CV_Assert(inShape[i] == outShape[i]);
232 CV_Assert(offsets[i] == 0);
237 auto rank = inShape.size();
239 std::vector<std::size_t> inStride(rank), outStride(rank);
241 outStride.back() = 1;
242 /* garbage, ..., garbage, 1 */
244 std::copy(std::begin(inShape) + 1, std::end(inShape), std::begin(inStride));
245 std::copy(std::begin(outShape) + 1, std::end(outShape), std::begin(outStride));
246 /* dim[0], dim[1], ..., dim[-1], 1 */
248 std::partial_sum(inStride.rbegin(), inStride.rend(), inStride.rbegin(), std::multiplies<int>());
249 std::partial_sum(outStride.rbegin(), outStride.rend(), outStride.rbegin(), std::multiplies<int>());
250 /* stride[0], stride[1], ..., stride[-2], 1 */
252 CV_Assert(1 <= rank && rank <= CSL_MAX_TENSOR_RANK);
253 concat_with_offsets_dispatcher<T, 1, CSL_MAX_TENSOR_RANK>(rank, stream, output, outStride, offsets, input, inStride);
256 template void concat_with_offsets(const Stream&, TensorSpan<__half>, TensorView<__half>, std::vector<std::size_t>);
257 template void concat_with_offsets(const Stream&, TensorSpan<float>, TensorView<float>, std::vector<std::size_t>);
259 }}}} /* namespace cv::dnn::cuda4dnn::kernels */