OSSForks/opencv
1
0
Fork 0
mirror of https://github.com/zebrajr/opencv.git synced 2025-12-06 12:19:50 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

Fix gaussianBlur5x5 performance regression

Browse Source
This commit is contained in:
Xue Zhang 2025-03-10 15:40:48 +05:30
parent 830d031213
commit accebdecf7
1 changed files with 25 additions and 66 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

91
3rdparty/fastcv/src/fastcv_hal_imgproc.cpp vendored
View File

@ -467,47 +467,30 @@ class FcvGaussianBlurLoop_Invoker : public cv::ParallelLoopBody
{ {
public: public:
FcvGaussianBlurLoop_Invoker(const cv::Mat& _src, cv::Mat& _dst, int _ksize, fcvBorderType _fcvBorder, int _fcvBorderValue) : FcvGaussianBlurLoop_Invoker(const cv::Mat& _src, cv::Mat& _dst, int _ksize, int _borderType, int _fcvBorderValue) :
cv::ParallelLoopBody(), src(_src),dst(_dst), ksize(_ksize), fcvBorder(_fcvBorder), fcvBorderValue(_fcvBorderValue) cv::ParallelLoopBody(), src(_src),dst(_dst), ksize(_ksize), borderType(_borderType), fcvBorderValue(_fcvBorderValue)
{ {
width = src.cols; width = src.cols;
height = src.rows; height = src.rows;
halfKernelSize = ksize / 2; halfKSize = ksize / 2;
fcvFuncType = FCV_MAKETYPE(ksize, src.depth()); fcvFuncType = FCV_MAKETYPE(ksize, src.depth());
} }
virtual void operator()(const cv::Range& range) const CV_OVERRIDE virtual void operator()(const cv::Range& range) const CV_OVERRIDE
{ {
int topLines = 0; int rangeHeight = range.end - range.start;
int rangeHeight = range.end-range.start; int paddedHeight = rangeHeight + halfKSize * 2;
int paddedHeight = rangeHeight; int paddedWidth = width;
// Need additional lines to be border. cv::Mat srcPadded = src(cv::Rect(0, range.start, paddedWidth, paddedHeight));
if(range.start != 0) cv::Mat dstPadded = dst(cv::Rect(0, range.start, paddedWidth, paddedHeight));
{
topLines += halfKernelSize;
paddedHeight += halfKernelSize;
}
if(range.end != height)
{
paddedHeight += halfKernelSize;
}
const cv::Mat srcPadded = src(cv::Rect(0, range.start - topLines, width, paddedHeight));
cv::Mat dstPadded = cv::Mat(paddedHeight, width, CV_8U);
if (fcvFuncType == FCV_MAKETYPE(3,CV_8U)) if (fcvFuncType == FCV_MAKETYPE(3,CV_8U))
fcvFilterGaussian3x3u8_v4(srcPadded.data, width, paddedHeight, srcPadded.step, dstPadded.data, dstPadded.step, fcvFilterGaussian3x3u8_v4(srcPadded.data, paddedWidth, paddedHeight, srcPadded.step, dstPadded.data, dstPadded.step,
fcvBorder, 0); fcvBorderType::FASTCV_BORDER_UNDEFINED, fcvBorderValue);
else if (fcvFuncType == FCV_MAKETYPE(5,CV_8U)) else if (fcvFuncType == FCV_MAKETYPE(5,CV_8U))
fcvFilterGaussian5x5u8_v3(srcPadded.data, width, paddedHeight, srcPadded.step, dstPadded.data, dstPadded.step, fcvFilterGaussian5x5u8_v3(srcPadded.data, paddedWidth, paddedHeight, srcPadded.step, dstPadded.data, dstPadded.step,
fcvBorder, 0); fcvBorderType::FASTCV_BORDER_UNDEFINED, fcvBorderValue);
// Only copy center part back to output image and ignore the padded lines
cv::Mat temp1 = dstPadded(cv::Rect(0, topLines, width, rangeHeight));
cv::Mat temp2 = dst(cv::Rect(0, range.start, width, rangeHeight));
temp1.copyTo(temp2);
} }
private: private:
@ -516,9 +499,9 @@ class FcvGaussianBlurLoop_Invoker : public cv::ParallelLoopBody
int width; int width;
int height; int height;
const int ksize; const int ksize;
int halfKernelSize; int halfKSize;
int fcvFuncType; int fcvFuncType;
fcvBorderType fcvBorder; int borderType;
int fcvBorderValue; int fcvBorderValue;
FcvGaussianBlurLoop_Invoker(const FcvGaussianBlurLoop_Invoker &); // = delete; FcvGaussianBlurLoop_Invoker(const FcvGaussianBlurLoop_Invoker &); // = delete;
@ -545,9 +528,9 @@ int fastcv_hal_gaussianBlurBinomial(
if (src_data == dst_data) if (src_data == dst_data)
CV_HAL_RETURN_NOT_IMPLEMENTED("Inplace is not supported"); CV_HAL_RETURN_NOT_IMPLEMENTED("Inplace is not supported");
// The input image width and height should greater than kernel size // The pixels of input image should larger than 320*240
if (((size_t)height <= ksize) || ((size_t)width <= ksize)) if((width*height) < (320*240))
CV_HAL_RETURN_NOT_IMPLEMENTED("Input image size should be larger than kernel size"); CV_HAL_RETURN_NOT_IMPLEMENTED("Input image size should be larger than 320*240");
// The input channel should be 1 // The input channel should be 1
if (cn != 1) if (cn != 1)
@ -560,35 +543,7 @@ int fastcv_hal_gaussianBlurBinomial(
INITIALIZATION_CHECK; INITIALIZATION_CHECK;
fcvStatus status = FASTCV_SUCCESS; fcvStatus status = FASTCV_SUCCESS;
fcvBorderType fcvBorder = fcvBorderType::FASTCV_BORDER_UNDEFINED; int fcvFuncType = FCV_MAKETYPE(ksize, depth);
int fcvFuncType = FCV_MAKETYPE(ksize,depth);
switch (border_type)
{
case cv::BorderTypes::BORDER_REPLICATE:
{
fcvBorder = fcvBorderType::FASTCV_BORDER_REPLICATE;
break;
}
// For constant border, there are no border value, OpenCV default value is 0
case cv::BorderTypes::BORDER_CONSTANT:
{
fcvBorder = fcvBorderType::FASTCV_BORDER_CONSTANT;
break;
}
case cv::BorderTypes::BORDER_REFLECT:
{
fcvBorder = fcvBorderType::FASTCV_BORDER_REFLECT;
break;
}
case cv::BorderTypes::BORDER_REFLECT_101:
{
fcvBorder = fcvBorderType::FASTCV_BORDER_REFLECT_V2;
break;
}
default:
CV_HAL_RETURN_NOT_IMPLEMENTED(cv::format("Border type:%s is not supported", borderToString(border_type)));
}
int nThreads = cv::getNumThreads(); int nThreads = cv::getNumThreads();
int nStripes = (nThreads > 1) ? ((height > 60) ? 3 * nThreads : 1) : 1; int nStripes = (nThreads > 1) ? ((height > 60) ? 3 * nThreads : 1) : 1;
@ -598,9 +553,13 @@ int fastcv_hal_gaussianBlurBinomial(
case FCV_MAKETYPE(3,CV_8U): case FCV_MAKETYPE(3,CV_8U):
case FCV_MAKETYPE(5,CV_8U): case FCV_MAKETYPE(5,CV_8U):
{ {
cv::Mat src = cv::Mat(height, width, CV_8UC1, (void*)src_data, src_step); cv::Mat src = cv::Mat(height, width, CV_8UC1, (void *)src_data, src_step);
cv::Mat dst = cv::Mat(height, width, CV_8UC1, (void*)dst_data, dst_step); cv::Mat dst = cv::Mat(height, width, CV_8UC1, (void *)dst_data, dst_step);
cv::parallel_for_(cv::Range(0, height), FcvGaussianBlurLoop_Invoker(src, dst, ksize, fcvBorder, 0), nStripes); cv::Mat src_tmp = cv::Mat(height + ksize - 1, width + ksize - 1, CV_8UC1);
cv::Mat dst_tmp = cv::Mat(height + ksize - 1, width + ksize - 1, CV_8UC1);
cv::copyMakeBorder(src, src_tmp, ksize / 2, ksize / 2, ksize / 2, ksize / 2, border_type);
cv::parallel_for_(cv::Range(0, height), FcvGaussianBlurLoop_Invoker(src_tmp, dst_tmp, ksize, border_type, 0), nStripes);
dst_tmp(cv::Rect(ksize / 2, ksize / 2, width, height)).copyTo(dst);
break; break;
} }
default: default: