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 "grid_stride_range.hpp"
11 #include "execution.hpp"
12 #include "kernel_dispatcher.hpp"
14 #include "../cuda4dnn/csl/stream.hpp"
15 #include "../cuda4dnn/csl/tensor.hpp"
16 #include "../cuda4dnn/csl/span.hpp"
18 #include <opencv2/core.hpp>
24 using namespace cv::dnn::cuda4dnn::csl;
25 using namespace cv::dnn::cuda4dnn::csl::device;
27 namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
30 template <class T, std::size_t Rank>
31 __global__ void slice(
32 Span<T> output, array<size_type, Rank> out_strides,
33 View<T> input, array<size_type, Rank> in_strides, array<index_type, Rank> in_offset)
35 for (auto i : grid_stride_range(output.size())) {
36 index_type out_index = i / out_strides[0];
37 index_type in_index = in_offset[0] + out_index;
38 index_type iidx = in_index * in_strides[0];
39 for (int j = 1; j < Rank; j++) {
40 out_index = (i % out_strides[j - 1]) / out_strides[j];
41 in_index = in_offset[j] + out_index;
42 iidx += in_index * in_strides[j];
45 output[i] = input[iidx];
50 template <class T, std::size_t Rank> static
53 Span<T> output, const std::vector<std::size_t>& outStride,
54 View<T> input, const std::vector<std::size_t>& inStride, const std::vector<std::size_t>& inOffset)
56 CV_Assert(outStride.size() == Rank);
57 CV_Assert(inStride.size() == Rank);
58 CV_Assert(inOffset.size() == Rank);
60 array<size_type, Rank> outStride_k, inStride_k;
61 outStride_k.assign(std::begin(outStride), std::end(outStride));
62 inStride_k.assign(std::begin(inStride), std::end(inStride));
64 array<index_type, Rank> inOffset_k;
65 inOffset_k.assign(std::begin(inOffset), std::end(inOffset));
67 auto kernel = raw::slice<T, Rank>;
68 auto policy = make_policy(kernel, output.size(), 0, stream);
69 launch_kernel(kernel, policy, output, outStride_k, input, inStride_k, inOffset_k);
72 GENERATE_KERNEL_DISPATCHER(slice_dispatcher, launch_slice);
75 void slice(const Stream& stream,
76 TensorSpan<T> output, TensorView<T> input,
77 std::vector<std::size_t> offsets)
79 CV_Assert(output.rank() == input.rank());
80 CV_Assert(output.rank() == offsets.size());
82 /* squeezable axes at the begining of both tensors can be eliminated
86 * Suppose an item's indices in the output tensor is [o1, o2, ...]. The indices in the input
87 * tensor will be [o1 + off1, o2 + off2, ...]. The rest of the elements in the input are igored.
89 * If the size of the first axis of the input and output tensor is unity, the input and output indices
90 * for all the elements will be of the form be [0, o2 + off2, ...] and [0, o2, ...] respectively. Note that
91 * there cannot be any ignored items since the axes have unit size. The first index does not contribute to the
92 * element's address calculation and hence does nothing apart from eating up few cycles.
94 while (input.get_axis_size(0) == 1 && output.get_axis_size(0) == 1) {
95 CV_Assert(offsets[0] == 0);
99 offsets.erase(std::begin(offsets));
101 CV_Assert(output.rank() == input.rank());
102 CV_Assert(output.rank() == offsets.size());
105 auto inShape = input.shape_as_vector();
106 auto outShape = output.shape_as_vector();
108 /* contiguous axes which do not undergo slicing can be combined into one axis
112 * Suppose an item's indices in the output tensor is [o1, o2, o3, ...]. Let the first two axes not undergo any
113 * slicing. The indices in the input tensor will be [o1, o2, o3 + off3, ...].
115 * Each axis in the contiguous unsliced axes sequence will add an offset of iN * strideN. In the above example,
116 * the two axes add a total offset of `o1 * stride1 + o2 * stride2`. We can merge the two axes into one axis with
117 * a size of `size1 * size2`. The new offset added will be o12 * stride2` as the kernel iterates through `o12`.
118 * Note that `o12` is actually `(o1 * size2 + o2)` in the original tensor.
120 for (int i = 0; i < inShape.size(); i++) {
121 /* check if axis `i` requires any slicing */
122 if (offsets[i] == 0 && inShape[i] == outShape[i]) {
123 /* loop invariant: `i` is the first axis in the contiguous unsliced axis sequence */
125 int j = i + 1; /* `j` is the axis which we will attempt to merge */
126 while (j < inShape.size() && offsets[j] == 0 && inShape[j] == outShape[j]) {
127 /* `j` axis is also unsliced; merge `i` and `j` */
128 auto new_size = inShape[i] * inShape[j];
129 inShape[i] = new_size;
130 outShape[i] = new_size;
131 offsets[i] = 0; /* redundant */
133 /* delete axis `j` */
134 inShape.erase(std::begin(inShape) + j);
135 outShape.erase(std::begin(outShape) + j);
136 offsets.erase(std::begin(offsets) + j);
138 /* optimizations should not break the invariants */
139 CV_Assert(inShape.size() == outShape.size());
140 CV_Assert(inShape.size() == offsets.size());
141 CV_Assert(inShape[i] == outShape[i]);
142 CV_Assert(offsets[i] == 0);
147 auto rank = inShape.size();
149 std::vector<std::size_t> inStride(rank), outStride(rank);
151 outStride.back() = 1;
152 /* garbage, ..., garbage, 1 */
154 std::copy(std::begin(inShape) + 1, std::end(inShape), std::begin(inStride));
155 std::copy(std::begin(outShape) + 1, std::end(outShape), std::begin(outStride));
156 /* dim[0], dim[1], ..., dim[-1], 1 */
158 std::partial_sum(inStride.rbegin(), inStride.rend(), inStride.rbegin(), std::multiplies<std::size_t>());
159 std::partial_sum(outStride.rbegin(), outStride.rend(), outStride.rbegin(), std::multiplies<std::size_t>());
160 /* stride[0], stride[1], ..., stride[-2], 1 */
162 CV_Assert(1 <= rank && rank <= CSL_MAX_TENSOR_RANK);
163 slice_dispatcher<T, 1, CSL_MAX_TENSOR_RANK>(rank, stream, output, outStride, input, inStride, offsets);
166 template void slice(const Stream&, TensorSpan<__half>, TensorView<__half>, std::vector<std::size_t>);
167 template void slice(const Stream&, TensorSpan<float>, TensorView<float>, std::vector<std::size_t>);
169 }}}} /* namespace cv::dnn::cuda4dnn::kernels */