mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-06-07 17:44:04 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge pull request #20933 from shubham-shahh:master

Browse Source 
Improved overall readability of the code

* grid_nms.cu: minor fix-ups

* Update grid_stride_range.hpp

* Update tf_importer.cpp
This commit is contained in:
Supernovae 2021-11-28 18:24:29 +05:30 committed by GitHub
parent 15073d63d9
commit b594ed99b8
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
modules/dnn/src/cuda/grid_nms.cu
View File

@ -68,7 +68,7 @@ namespace raw {
* to compute IOU(GROUP_B, GROUP_A). We still have to compute IOU(GROUP_A, GROUP_A) though since * to compute IOU(GROUP_B, GROUP_A). We still have to compute IOU(GROUP_A, GROUP_A) though since
* each group has many boxes and we need IOUs amongst boxes within a group. * each group has many boxes and we need IOUs amongst boxes within a group.
* *
* We arbitarily choose a scheme to exit : exit if group_i is greater than group_j. This way we only * We arbitrarily choose a scheme to exit : exit if group_i is greater than group_j. This way we only
* compute IOUs between groups once. While nearly half the blocks are wasted, it's ok since they exit * compute IOUs between groups once. While nearly half the blocks are wasted, it's ok since they exit
* early on and the working blocks are compute heavy. * early on and the working blocks are compute heavy.
*/ */
@ -92,7 +92,7 @@ namespace raw {
*/ */
/* The `j` box is fixed for each thread. All `i` boxes will be required for every thread. /* The `j` box is fixed for each thread. All `i` boxes will be required for every thread.
* We store the `i` boxes in shared memory to allow global memory coalesing. * We store the `i` boxes in shared memory to allow global memory coalescing.
*/ */
using vector_type = get_vector_type_t<T, 4>; using vector_type = get_vector_type_t<T, 4>;
__shared__ vector_type group_i_boxes[BLOCK_SIZE]; __shared__ vector_type group_i_boxes[BLOCK_SIZE];
@ -162,7 +162,7 @@ namespace raw {
* this loop has been highly tuned. Please profile and verify carefully before making changes. * this loop has been highly tuned. Please profile and verify carefully before making changes.
*/ */
/* UNROLL_SIZE is the number of boxes that must be processed per iteration. We manually unroll /* UNROLL_SIZE is the number of boxes that must be processed per iteration. We manually unroll
* the loop since the compiler cannot effectively unroll on its own preassumably due to presence * the loop since the compiler cannot effectively unroll on its own presumably due to presence
* of instructions forcing warp synchronization. * of instructions forcing warp synchronization.
*/ */
constexpr int UNROLL_SIZE = 4; constexpr int UNROLL_SIZE = 4;
@ -290,7 +290,7 @@ namespace raw {
if (boxes == 0) if (boxes == 0)
return; return;
/* We have a fixed number of threads and an arbitary number of boxes. We use an array of /* We have a fixed number of threads and an arbitrary number of boxes. We use an array of
* bits to store which boxes haven't been eliminated and which are still active. We organize * bits to store which boxes haven't been eliminated and which are still active. We organize
* the array of bits into a matrix of bits of the shape (num_rows, BLOCK_SIZE, 32) which * the array of bits into a matrix of bits of the shape (num_rows, BLOCK_SIZE, 32) which
* is equivalent to (num_rows, BLOCK_SIZE) where the type is a 32-bit unsigned integer. * is equivalent to (num_rows, BLOCK_SIZE) where the type is a 32-bit unsigned integer.
@ -464,4 +464,4 @@ std::size_t getGridNMSWorkspaceSizePerBatchItem(std::size_t num_classes, std::si
template void grid_nms(const Stream& stream, Span<unsigned int> workspace, TensorSpan<int> indices, TensorSpan<int> count, TensorView<__half> bboxes, int, bool normalized_bbox, float nms_threshold); template void grid_nms(const Stream& stream, Span<unsigned int> workspace, TensorSpan<int> indices, TensorSpan<int> count, TensorView<__half> bboxes, int, bool normalized_bbox, float nms_threshold);
template void grid_nms(const Stream& stream, Span<unsigned int> workspace, TensorSpan<int> indices, TensorSpan<int> count, TensorView<float> bboxes, int, bool normalized_bbox, float nms_threshold); template void grid_nms(const Stream& stream, Span<unsigned int> workspace, TensorSpan<int> indices, TensorSpan<int> count, TensorView<float> bboxes, int, bool normalized_bbox, float nms_threshold);
}}}} /* namespace cv::dnn::cuda4dnn::kernels */ }}}} /* namespace cv::dnn::cuda4dnn::kernels */

2
modules/dnn/src/cuda/grid_stride_range.hpp
View File

@ -36,7 +36,7 @@ public:
__device__ bool operator!=(const iterator& other) const { __device__ bool operator!=(const iterator& other) const {
/* NOTE HACK /* NOTE HACK
* 'pos' can move in large steps (see operator++) * 'pos' can move in large steps (see operator++)
* expansion of range for loop uses != as the loop conditioion * expansion of range for loop uses != as the loop condition
* => operator!= must return false if 'pos' crosses the end * => operator!= must return false if 'pos' crosses the end
*/ */
return pos < other.pos; return pos < other.pos;

2
modules/dnn/src/tensorflow/tf_importer.cpp
View File

@ -1154,7 +1154,7 @@ void TFImporter::parseExpandDims(tensorflow::GraphDef& net, const tensorflow::No
// Convert OpenCV's NHC to NCH first. // Convert OpenCV's NHC to NCH first.
if(outShapeSize == 3) if(outShapeSize == 3)
{ {
// If axis equal to outShapeSize, that mean we expand in Channel dimmension, and do not add permuteLayer. // If axis equal to outShapeSize, that mean we expand in Channel dimension, and do not add permuteLayer.
if(axis != outShapeSize) if(axis != outShapeSize)
{ {
int order[] = {0, 2, 1}; // From OpenCV's NHC to NCH. int order[] = {0, 2, 1}; // From OpenCV's NHC to NCH.