Administrator/opencv-MIRROR
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-09-11 04:43:22 -05:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #29801 from amd:fast_sobel2d_opt

Browse Source 
imgproc: fused spatialGradient + dispatched SIMD Canny
This commit is contained in:
Alexander Smorkalov
2026-08-31 12:50:54 +03:00
committed by GitHub
parent 1e82d103e5 b0cfda3187
commit 6e4c669be8
12 changed files with 1065 additions and 587 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
modules/imgproc/CMakeLists.txt
View File

@@ -2,11 +2,13 @@ set(the_description "Image Processing")
ocv_add_dispatched_file(accum SSE4_1 AVX AVX2 AVX512_SKX AVX512_ICL)
ocv_add_dispatched_file(bilateral_filter SSE2 AVX2 AVX512_SKX AVX512_ICL)
ocv_add_dispatched_file(box_filter SSE2 SSE4_1 AVX2 AVX512_SKX)
ocv_add_dispatched_file(canny SSE4_1 AVX2 AVX512_SKX AVX512_ICL)
ocv_add_dispatched_file(filter SSE2 SSE4_1 AVX2)
ocv_add_dispatched_file(color_hsv SSE2 SSE4_1 AVX2 AVX512_SKX AVX512_ICL)
ocv_add_dispatched_file(color_rgb SSE2 SSE4_1 AVX2 AVX512_SKX AVX512_ICL)
ocv_add_dispatched_file(color_yuv SSE2 SSE4_1 AVX2 AVX512_SKX AVX512_ICL)
ocv_add_dispatched_file(median_blur SSE2 SSE4_1 AVX2 AVX512_SKX AVX512_ICL)
ocv_add_dispatched_file(spatialgradient SSE4_1 AVX2 AVX512_SKX AVX512_ICL)
ocv_add_dispatched_file(morph SSE2 SSE4_1 AVX2)
ocv_add_dispatched_file(smooth SSE2 SSE4_1 AVX2 AVX512_ICL)
ocv_add_dispatched_file(sumpixels SSE2 AVX2 AVX512_SKX)

21
modules/imgproc/include/opencv2/imgproc.hpp
View File

@@ -1818,21 +1818,22 @@ This is a fused variant of #Sobel: instead of two separate calls
Sobel( src, dx, ddepth, 1, 0, ksize, scale );
Sobel( src, dy, ddepth, 0, 1, ksize, scale );
@endcode
it produces both first-order derivatives in one traversal of the source. For an 8-bit single-channel
whole-image (non-ROI) source with @p ksize = 3, @p ddepth = CV_16S, @p scale = 1, and
#BORDER_DEFAULT (#BORDER_REFLECT_101) or #BORDER_REPLICATE, a dedicated fused 3x3 stencil kernel is
used (including a HAL fast path) that reads each source sample once and shares it between the dx and
dy computations; the result is bit-identical to the two #Sobel calls above. All other cases
(@p ddepth = CV_32F, @p ksize = 5, scaled int16 output, floating-point source, other border types,
or ROI/sub-matrix input) currently fall back to the two equivalent #Sobel passes. @note Fused fast
paths for CV_32F output and @p ksize = 5 are added in a follow-up change.
it produces both first-order derivatives in one traversal of the source. The fused single-pass
kernels apply to an 8-bit single-channel (CV_8UC1) source with @p ksize = 3 or 5 and a
reflect/replicate border (#BORDER_DEFAULT / #BORDER_REFLECT_101, #BORDER_REFLECT, #BORDER_REPLICATE),
for both @p ddepth = CV_16S (unit @p scale) and @p ddepth = CV_32F (any @p scale); full-width
row-range ROIs are supported as well. In these cases each source sample is read once and shared
between the dx and dy computations, and the result is bit-identical to the two #Sobel calls above.
The remaining cases (scaled int16 output, floating-point source, #BORDER_CONSTANT/#BORDER_WRAP, or a
column-offset/partial-width ROI) fall back to the two equivalent #Sobel passes.
@param src input image; single-channel, 8-bit (CV_8UC1) for the fused fast paths (CV_32FC1 is
accepted via the fallback).
@param dx output image with the first-order derivative in x (depth @p ddepth, same size as src).
@param dy output image with the first-order derivative in y (depth @p ddepth, same size as src).
@param ksize size of the Sobel kernel; fused fast paths require 3 (5 uses the Sobel fallback).
Also accepts -1 (Scharr) and 7 for Sobel-compatible callers such as #HoughCircles.
@param ksize size of the Sobel kernel; fused fast paths require 3 or 5 (other sizes use the Sobel
fallback). Also accepts 1, -1 (Scharr), and 7 for Sobel-compatible callers such as
#HoughCircles and #cv::segmentation::IntelligentScissorsMB.
@param borderType pixel extrapolation method, see #BorderTypes. #BORDER_WRAP is not supported.
@param ddepth output image depth; CV_16S or CV_32F.
@param scale optional scale factor applied to the computed derivatives.

256
modules/imgproc/src/canny.cpp
View File

@@ -12,6 +12,7 @@
//
// Copyright (C) 2000, Intel Corporation, all rights reserved.
// Copyright (C) 2014, Itseez Inc., all rights reserved.
// Copyright (C) 2026, Advanced Micro Devices, Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
@@ -42,7 +43,7 @@
#include "precomp.hpp"
#include "opencl_kernels_imgproc.hpp"
#include "opencv2/core/hal/intrin.hpp"
#include "canny.hpp"
#include <deque>
namespace cv
@@ -223,19 +224,10 @@ public:
parallelCanny(const Mat &_src, Mat &_map, std::deque<uchar*> &borderPeaksParallel,
int _low, int _high, int _aperture_size, bool _L2gradient) :
src(_src), src2(_src), map(_map), _borderPeaksParallel(borderPeaksParallel),
low(_low), high(_high), aperture_size(_aperture_size), L2gradient(_L2gradient)
low(_low), high(_high), aperture_size(_aperture_size), L2gradient(_L2gradient),
simdWidth(canny_simd_width())
{
#if (CV_SIMD || CV_SIMD_SCALABLE)
for(int i = 0; i < VTraits<v_int8>::vlanes(); ++i)
{
smask[i] = 0;
smask[i + VTraits<v_int8>::vlanes()] = (schar)-1;
}
if (true)
_map.create(src.rows + 2, (int)alignSize((size_t)(src.cols + CV_SIMD_WIDTH + 1), CV_SIMD_WIDTH), CV_8UC1);
else
#endif
_map.create(src.rows + 2, src.cols + 2, CV_8UC1);
_map.create(src.rows + 2, (int)alignSize((size_t)(src.cols + simdWidth + 1), simdWidth), CV_8UC1);
map = _map;
map.row(0).setTo(1);
map.row(src.rows + 1).setTo(1);
@@ -247,19 +239,10 @@ public:
parallelCanny(const Mat &_dx, const Mat &_dy, Mat &_map, std::deque<uchar*> &borderPeaksParallel,
int _low, int _high, bool _L2gradient) :
src(_dx), src2(_dy), map(_map), _borderPeaksParallel(borderPeaksParallel),
low(_low), high(_high), aperture_size(0), L2gradient(_L2gradient)
low(_low), high(_high), aperture_size(0), L2gradient(_L2gradient),
simdWidth(canny_simd_width())
{
#if (CV_SIMD || CV_SIMD_SCALABLE)
for(int i = 0; i < VTraits<v_int8>::vlanes(); ++i)
{
smask[i] = 0;
smask[i + VTraits<v_int8>::vlanes()] = (schar)-1;
}
if (true)
_map.create(src.rows + 2, (int)alignSize((size_t)(src.cols + CV_SIMD_WIDTH + 1), CV_SIMD_WIDTH), CV_8UC1);
else
#endif
_map.create(src.rows + 2, src.cols + 2, CV_8UC1);
_map.create(src.rows + 2, (int)alignSize((size_t)(src.cols + simdWidth + 1), simdWidth), CV_8UC1);
map = _map;
map.row(0).setTo(1);
map.row(src.rows + 1).setTo(1);
@@ -290,12 +273,8 @@ public:
CV_TRACE_REGION("gradient")
if(needGradient)
{
if (aperture_size == 7)
{
scale = 1 / 16.0;
}
Sobel(src.rowRange(rowStart, rowEnd), dx, CV_16S, 1, 0, aperture_size, scale, 0, BORDER_REPLICATE);
Sobel(src.rowRange(rowStart, rowEnd), dy, CV_16S, 0, 1, aperture_size, scale, 0, BORDER_REPLICATE);
scale = (aperture_size == 7) ? (1 / 16.0) : 1.0;
spatialGradient(src.rowRange(rowStart, rowEnd), dx, dy, aperture_size, BORDER_REPLICATE, CV_16S, scale);
}
else
{
@@ -315,17 +294,10 @@ public:
}
// _mag_p: previous row, _mag_a: actual row, _mag_n: next row
#if (CV_SIMD || CV_SIMD_SCALABLE)
AutoBuffer<int> buffer(3 * (mapstep * cn + CV_SIMD_WIDTH));
_mag_p = alignPtr(buffer.data() + 1, CV_SIMD_WIDTH);
_mag_a = alignPtr(_mag_p + mapstep * cn, CV_SIMD_WIDTH);
_mag_n = alignPtr(_mag_a + mapstep * cn, CV_SIMD_WIDTH);
#else
AutoBuffer<int> buffer(3 * (mapstep * cn));
_mag_p = buffer.data() + 1;
_mag_a = _mag_p + mapstep * cn;
_mag_n = _mag_a + mapstep * cn;
#endif
AutoBuffer<int> buffer(3 * (mapstep * cn + simdWidth));
_mag_p = alignPtr(buffer.data() + 1, simdWidth);
_mag_a = alignPtr(_mag_p + mapstep * cn, simdWidth);
_mag_n = alignPtr(_mag_a + mapstep * cn, simdWidth);
// For the first time when just 2 rows are filled and for left and right borders
if(rowStart == boundaries.start)
@@ -352,50 +324,7 @@ public:
_dx = dx.ptr<short>(i - rowStart);
_dy = dy.ptr<short>(i - rowStart);
if (L2gradient)
{
int j = 0, width = src.cols * cn;
#if (CV_SIMD || CV_SIMD_SCALABLE)
for ( ; j <= width - VTraits<v_int16>::vlanes(); j += VTraits<v_int16>::vlanes())
{
v_int16 v_dx = vx_load((const short*)(_dx + j));
v_int16 v_dy = vx_load((const short*)(_dy + j));
v_int32 v_dxp_low, v_dxp_high;
v_int32 v_dyp_low, v_dyp_high;
v_expand(v_dx, v_dxp_low, v_dxp_high);
v_expand(v_dy, v_dyp_low, v_dyp_high);
v_store_aligned((int *)(_mag_n + j), v_add(v_mul(v_dxp_low, v_dxp_low), v_mul(v_dyp_low, v_dyp_low)));
v_store_aligned((int *)(_mag_n + j + VTraits<v_int32>::vlanes()), v_add(v_mul(v_dxp_high, v_dxp_high), v_mul(v_dyp_high, v_dyp_high)));
}
#endif
for ( ; j < width; ++j)
_mag_n[j] = int(_dx[j])*_dx[j] + int(_dy[j])*_dy[j];
}
else
{
int j = 0, width = src.cols * cn;
#if (CV_SIMD || CV_SIMD_SCALABLE)
for(; j <= width - VTraits<v_int16>::vlanes(); j += VTraits<v_int16>::vlanes())
{
v_int16 v_dx = vx_load((const short *)(_dx + j));
v_int16 v_dy = vx_load((const short *)(_dy + j));
v_dx = v_reinterpret_as_s16(v_abs(v_dx));
v_dy = v_reinterpret_as_s16(v_abs(v_dy));
v_int32 v_dx_ml, v_dy_ml, v_dx_mh, v_dy_mh;
v_expand(v_dx, v_dx_ml, v_dx_mh);
v_expand(v_dy, v_dy_ml, v_dy_mh);
v_store_aligned((int *)(_mag_n + j), v_add(v_dx_ml, v_dy_ml));
v_store_aligned((int *)(_mag_n + j + VTraits<v_int32>::vlanes()), v_add(v_dx_mh, v_dy_mh));
}
#endif
for ( ; j < width; ++j)
_mag_n[j] = std::abs(int(_dx[j])) + std::abs(int(_dy[j]));
}
canny_calc_magnitude(_dx, _dy, _mag_n, src.cols * cn, L2gradient);
if(cn > 1)
{
@@ -434,12 +363,7 @@ public:
// From here actual src row is (i - 1)
// Set left and right border to 1
#if (CV_SIMD || CV_SIMD_SCALABLE)
if (true)
_pmap = map.ptr<uchar>(i) + CV_SIMD_WIDTH;
else
#endif
_pmap = map.ptr<uchar>(i) + 1;
_pmap = map.ptr<uchar>(i) + simdWidth;
_pmap[src.cols] =_pmap[-1] = 1;
@@ -454,109 +378,7 @@ public:
_dy = _dy_a;
}
const int TG22 = 13573;
int j = 0;
#if (CV_SIMD || CV_SIMD_SCALABLE)
{
const v_int32 v_low = vx_setall_s32(low);
const v_int8 v_one = vx_setall_s8(1);
for (; j <= src.cols - VTraits<v_int8>::vlanes(); j += VTraits<v_int8>::vlanes())
{
v_store_aligned((signed char*)(_pmap + j), v_one);
v_int8 v_cmp = v_pack(v_pack(v_gt(vx_load_aligned((const int *)(_mag_a + j)), v_low),
v_gt(vx_load_aligned((const int *)(_mag_a + j + VTraits<v_int32>::vlanes())), v_low)),
v_pack(v_gt(vx_load_aligned((const int *)(_mag_a + j + 2 * VTraits<v_int32>::vlanes())), v_low),
v_gt(vx_load_aligned((const int *)(_mag_a + j + 3 * VTraits<v_int32>::vlanes())), v_low)));
while (v_check_any(v_cmp))
{
int l = v_scan_forward(v_cmp);
v_cmp = v_and(v_cmp, vx_load(smask + VTraits<v_int8>::vlanes() - 1 - l));
int k = j + l;
int m = _mag_a[k];
short xs = _dx[k];
short ys = _dy[k];
int x = (int)std::abs(xs);
int y = (int)std::abs(ys) << 15;
int tg22x = x * TG22;
if (y < tg22x)
{
if (m > _mag_a[k - 1] && m >= _mag_a[k + 1])
{
CANNY_CHECK(m, high, (_pmap+k), stack);
}
}
else
{
int tg67x = tg22x + (x << 16);
if (y > tg67x)
{
if (m > _mag_p[k] && m >= _mag_n[k])
{
CANNY_CHECK(m, high, (_pmap+k), stack);
}
}
else
{
int s = (xs ^ ys) < 0 ? -1 : 1;
if(m > _mag_p[k - s] && m > _mag_n[k + s])
{
CANNY_CHECK(m, high, (_pmap+k), stack);
}
}
}
}
}
}
#endif
for (; j < src.cols; j++)
{
int m = _mag_a[j];
if (m > low)
{
short xs = _dx[j];
short ys = _dy[j];
int x = (int)std::abs(xs);
int y = (int)std::abs(ys) << 15;
int tg22x = x * TG22;
if (y < tg22x)
{
if (m > _mag_a[j - 1] && m >= _mag_a[j + 1])
{
CANNY_CHECK(m, high, (_pmap+j), stack);
continue;
}
}
else
{
int tg67x = tg22x + (x << 16);
if (y > tg67x)
{
if (m > _mag_p[j] && m >= _mag_n[j])
{
CANNY_CHECK(m, high, (_pmap+j), stack);
continue;
}
}
else
{
int s = (xs ^ ys) < 0 ? -1 : 1;
if(m > _mag_p[j - s] && m > _mag_n[j + s])
{
CANNY_CHECK(m, high, (_pmap+j), stack);
continue;
}
}
}
}
_pmap[j] = 1;
}
canny_nms_row(_mag_a, _mag_p, _mag_n, _dx, _dy, _pmap, src.cols, low, high, stack);
}
// Not for first row of first slice or last row of last slice
@@ -611,10 +433,8 @@ private:
bool L2gradient, needGradient;
ptrdiff_t mapstep;
int cn;
int simdWidth;
mutable Mutex mutex;
#if (CV_SIMD || CV_SIMD_SCALABLE)
schar smask[2*VTraits<v_int8>::max_nlanes];
#endif
};
class finalPass : public ParallelLoopBody
@@ -631,44 +451,14 @@ public:
void operator()(const Range &boundaries) const CV_OVERRIDE
{
// the final pass, form the final image
// the final pass, form the final image. The map's left padding matches the
// width used by parallelCanny when it was allocated (same dispatched value).
const int simdWidth = canny_simd_width();
for (int i = boundaries.start; i < boundaries.end; i++)
{
int j = 0;
uchar *pdst = dst.ptr<uchar>(i);
const uchar *pmap = map.ptr<uchar>(i + 1);
#if (CV_SIMD || CV_SIMD_SCALABLE)
if (true)
pmap += CV_SIMD_WIDTH;
else
#endif
pmap += 1;
#if (CV_SIMD || CV_SIMD_SCALABLE)
{
const v_uint8 v_zero = vx_setzero_u8();
const v_uint8 v_ff = v_not(v_zero);
const v_uint8 v_two = vx_setall_u8(2);
for (; j <= dst.cols - VTraits<v_uint8>::vlanes(); j += VTraits<v_uint8>::vlanes())
{
v_uint8 v_pmap = vx_load_aligned((const unsigned char*)(pmap + j));
v_pmap = v_select(v_eq(v_pmap, v_two), v_ff, v_zero);
v_store((pdst + j), v_pmap);
}
if (j <= dst.cols - VTraits<v_uint8>::vlanes()/2)
{
v_uint8 v_pmap = vx_load_low((const unsigned char*)(pmap + j));
v_pmap = v_select(v_eq(v_pmap, v_two), v_ff, v_zero);
v_store_low((pdst + j), v_pmap);
j += VTraits<v_uint8>::vlanes()/2;
}
}
#endif
for (; j < dst.cols; j++)
{
pdst[j] = (uchar)-(pmap[j] >> 1);
}
const uchar *pmap = map.ptr<uchar>(i + 1) + simdWidth;
canny_finalize_row(pmap, pdst, dst.cols);
}
}

85
modules/imgproc/src/canny.dispatch.cpp Normal file
View File

@@ -0,0 +1,85 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000, Intel Corporation, all rights reserved.
// Copyright (C) 2014, Itseez Inc., all rights reserved.
// Copyright (C) 2026, Advanced Micro Devices, Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
#include <deque>
#include "opencv2/core/hal/intrin.hpp"
#include "canny.hpp"
#include "canny.simd.hpp"
#include "canny.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX512_SKX,...,BASELINE based on CMakeLists.txt content
namespace cv {
int canny_simd_width()
{
CV_CPU_DISPATCH(canny_simd_width, (), CV_CPU_DISPATCH_MODES_ALL);
}
void canny_calc_magnitude(const short* _dx, const short* _dy, int* _mag_n,
int width, bool L2gradient)
{
CV_INSTRUMENT_REGION();
CV_CPU_DISPATCH(canny_calc_magnitude, (_dx, _dy, _mag_n, width, L2gradient),
CV_CPU_DISPATCH_MODES_ALL);
}
void canny_nms_row(const int* _mag_a, const int* _mag_p, const int* _mag_n,
const short* _dx, const short* _dy, uchar* _pmap,
int width, int low, int high, std::deque<uchar*>& stack)
{
CV_INSTRUMENT_REGION();
CV_CPU_DISPATCH(canny_nms_row, (_mag_a, _mag_p, _mag_n, _dx, _dy, _pmap, width, low, high, stack),
CV_CPU_DISPATCH_MODES_ALL);
}
void canny_finalize_row(const uchar* pmap, uchar* pdst, int width)
{
CV_INSTRUMENT_REGION();
CV_CPU_DISPATCH(canny_finalize_row, (pmap, pdst, width),
CV_CPU_DISPATCH_MODES_ALL);
}
} // namespace cv

24
modules/imgproc/src/canny.hpp Normal file
View File

@@ -0,0 +1,24 @@
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html
// Copyright (C) 2026, Advanced Micro Devices, Inc., all rights reserved.
#ifndef OPENCV_IMGPROC_CANNY_HPP
#define OPENCV_IMGPROC_CANNY_HPP
#include <deque>
namespace cv {
int canny_simd_width();
void canny_calc_magnitude(const short* _dx, const short* _dy, int* _mag_n,
int width, bool L2gradient);
void canny_nms_row(const int* _mag_a, const int* _mag_p, const int* _mag_n,
const short* _dx, const short* _dy, uchar* _pmap,
int width, int low, int high, std::deque<uchar*>& stack);
void canny_finalize_row(const uchar* pmap, uchar* pdst, int width);
} // namespace cv
#endif // OPENCV_IMGPROC_CANNY_HPP

285
modules/imgproc/src/canny.simd.hpp Normal file
View File

@@ -0,0 +1,285 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000, Intel Corporation, all rights reserved.
// Copyright (C) 2014, Itseez Inc., all rights reserved.
// Copyright (C) 2026, Advanced Micro Devices, Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "precomp.hpp"
#include "opencv2/core/hal/intrin.hpp"
#include <deque>
namespace cv {
CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN
int canny_simd_width();
void canny_calc_magnitude(const short* _dx, const short* _dy, int* _mag_n,
int width, bool L2gradient);
void canny_nms_row(const int* _mag_a, const int* _mag_p, const int* _mag_n,
const short* _dx, const short* _dy, uchar* _pmap,
int width, int low, int high, std::deque<uchar*>& stack);
void canny_finalize_row(const uchar* pmap, uchar* pdst, int width);
#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
#define CANNY_PUSH(map, stack) *map = 2, stack.push_back(map)
#define CANNY_CHECK(m, high, map, stack) \
if (m > high) \
CANNY_PUSH(map, stack); \
else \
*map = 0
int canny_simd_width()
{
#if (CV_SIMD || CV_SIMD_SCALABLE)
return VTraits<v_uint8>::vlanes();
#else
return 1;
#endif
}
// stage 1: magnitude of the gradient for one row (L1 = |dx|+|dy|, L2 = dx*dx+dy*dy).
// width == src.cols * cn.
void canny_calc_magnitude(const short* _dx, const short* _dy, int* _mag_n,
int width, bool L2gradient)
{
int j = 0;
if (L2gradient)
{
#if (CV_SIMD || CV_SIMD_SCALABLE)
for ( ; j <= width - VTraits<v_int16>::vlanes(); j += VTraits<v_int16>::vlanes())
{
v_int16 v_dx = vx_load((const short*)(_dx + j));
v_int16 v_dy = vx_load((const short*)(_dy + j));
v_int32 v_dxp_low, v_dxp_high;
v_int32 v_dyp_low, v_dyp_high;
v_expand(v_dx, v_dxp_low, v_dxp_high);
v_expand(v_dy, v_dyp_low, v_dyp_high);
v_store_aligned((int *)(_mag_n + j), v_add(v_mul(v_dxp_low, v_dxp_low), v_mul(v_dyp_low, v_dyp_low)));
v_store_aligned((int *)(_mag_n + j + VTraits<v_int32>::vlanes()), v_add(v_mul(v_dxp_high, v_dxp_high), v_mul(v_dyp_high, v_dyp_high)));
}
#endif
for ( ; j < width; ++j)
_mag_n[j] = int(_dx[j])*_dx[j] + int(_dy[j])*_dy[j];
}
else
{
#if (CV_SIMD || CV_SIMD_SCALABLE)
for(; j <= width - VTraits<v_int16>::vlanes(); j += VTraits<v_int16>::vlanes())
{
v_int16 v_dx = vx_load((const short *)(_dx + j));
v_int16 v_dy = vx_load((const short *)(_dy + j));
v_dx = v_reinterpret_as_s16(v_abs(v_dx));
v_dy = v_reinterpret_as_s16(v_abs(v_dy));
v_int32 v_dx_ml, v_dy_ml, v_dx_mh, v_dy_mh;
v_expand(v_dx, v_dx_ml, v_dx_mh);
v_expand(v_dy, v_dy_ml, v_dy_mh);
v_store_aligned((int *)(_mag_n + j), v_add(v_dx_ml, v_dy_ml));
v_store_aligned((int *)(_mag_n + j + VTraits<v_int32>::vlanes()), v_add(v_dx_mh, v_dy_mh));
}
#endif
for ( ; j < width; ++j)
_mag_n[j] = std::abs(int(_dx[j])) + std::abs(int(_dy[j]));
}
}
// stage 2: non-maxima suppression + double thresholding for one row.
// Fills _pmap with: 0 (maybe edge), 1 (not edge), 2 (strong edge). Strong edges
// are pushed onto `stack` for the subsequent hysteresis pass.
void canny_nms_row(const int* _mag_a, const int* _mag_p, const int* _mag_n,
const short* _dx, const short* _dy, uchar* _pmap,
int width, int low, int high, std::deque<uchar*>& stack)
{
const int TG22 = 13573;
int j = 0;
#if (CV_SIMD || CV_SIMD_SCALABLE)
// smask[i] selects, after a v_scan_forward hit at lane l, all lanes >= l so
// that the already-handled lanes are cleared from the candidate mask.
schar smask[2 * VTraits<v_int8>::max_nlanes];
for (int i = 0; i < VTraits<v_int8>::vlanes(); ++i)
{
smask[i] = 0;
smask[i + VTraits<v_int8>::vlanes()] = (schar)-1;
}
const v_int32 v_low = vx_setall_s32(low);
const v_int8 v_one = vx_setall_s8(1);
for (; j <= width - VTraits<v_int8>::vlanes(); j += VTraits<v_int8>::vlanes())
{
v_store_aligned((signed char*)(_pmap + j), v_one);
v_int8 v_cmp = v_pack(v_pack(v_gt(vx_load_aligned((const int *)(_mag_a + j)), v_low),
v_gt(vx_load_aligned((const int *)(_mag_a + j + VTraits<v_int32>::vlanes())), v_low)),
v_pack(v_gt(vx_load_aligned((const int *)(_mag_a + j + 2 * VTraits<v_int32>::vlanes())), v_low),
v_gt(vx_load_aligned((const int *)(_mag_a + j + 3 * VTraits<v_int32>::vlanes())), v_low)));
while (v_check_any(v_cmp))
{
int l = v_scan_forward(v_cmp);
v_cmp = v_and(v_cmp, vx_load(smask + VTraits<v_int8>::vlanes() - 1 - l));
int k = j + l;
int m = _mag_a[k];
short xs = _dx[k];
short ys = _dy[k];
int x = (int)std::abs(xs);
int y = (int)std::abs(ys) << 15;
int tg22x = x * TG22;
if (y < tg22x)
{
if (m > _mag_a[k - 1] && m >= _mag_a[k + 1])
{
CANNY_CHECK(m, high, (_pmap+k), stack);
}
}
else
{
int tg67x = tg22x + (x << 16);
if (y > tg67x)
{
if (m > _mag_p[k] && m >= _mag_n[k])
{
CANNY_CHECK(m, high, (_pmap+k), stack);
}
}
else
{
int s = (xs ^ ys) < 0 ? -1 : 1;
if(m > _mag_p[k - s] && m > _mag_n[k + s])
{
CANNY_CHECK(m, high, (_pmap+k), stack);
}
}
}
}
}
#endif
for (; j < width; j++)
{
int m = _mag_a[j];
if (m > low)
{
short xs = _dx[j];
short ys = _dy[j];
int x = (int)std::abs(xs);
int y = (int)std::abs(ys) << 15;
int tg22x = x * TG22;
if (y < tg22x)
{
if (m > _mag_a[j - 1] && m >= _mag_a[j + 1])
{
CANNY_CHECK(m, high, (_pmap+j), stack);
continue;
}
}
else
{
int tg67x = tg22x + (x << 16);
if (y > tg67x)
{
if (m > _mag_p[j] && m >= _mag_n[j])
{
CANNY_CHECK(m, high, (_pmap+j), stack);
continue;
}
}
else
{
int s = (xs ^ ys) < 0 ? -1 : 1;
if(m > _mag_p[j - s] && m > _mag_n[j + s])
{
CANNY_CHECK(m, high, (_pmap+j), stack);
continue;
}
}
}
}
_pmap[j] = 1;
}
}
// final pass: turn the edge map (value 2 == edge) into a 0/255 output row.
void canny_finalize_row(const uchar* pmap, uchar* pdst, int width)
{
int j = 0;
#if (CV_SIMD || CV_SIMD_SCALABLE)
{
const v_uint8 v_zero = vx_setzero_u8();
const v_uint8 v_ff = v_not(v_zero);
const v_uint8 v_two = vx_setall_u8(2);
for (; j <= width - VTraits<v_uint8>::vlanes(); j += VTraits<v_uint8>::vlanes())
{
v_uint8 v_pmap = vx_load_aligned((const unsigned char*)(pmap + j));
v_pmap = v_select(v_eq(v_pmap, v_two), v_ff, v_zero);
v_store((pdst + j), v_pmap);
}
if (j <= width - VTraits<v_uint8>::vlanes()/2)
{
v_uint8 v_pmap = vx_load_low((const unsigned char*)(pmap + j));
v_pmap = v_select(v_eq(v_pmap, v_two), v_ff, v_zero);
v_store_low((pdst + j), v_pmap);
j += VTraits<v_uint8>::vlanes()/2;
}
}
#endif
for (; j < width; j++)
{
pdst[j] = (uchar)-(pmap[j] >> 1);
}
}
#endif // CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
CV_CPU_OPTIMIZATION_NAMESPACE_END
} // namespace cv

33
modules/imgproc/src/corner.cpp
View File

@@ -255,25 +255,7 @@ cornerEigenValsVecs( const Mat& src, Mat& eigenv, int block_size,
CV_Assert( src.type() == CV_8UC1 || src.type() == CV_32FC1 );
Mat Dx, Dy;
// TODO(follow-up PR): route CV_32F gradients through the fused spatialGradient once its
// CV_32F/ksize=5 fast paths land. Disabled for now so this stays on the tuned Sobel path.
#if 0
if( aperture_size == 3 || aperture_size == 5 )
{
spatialGradient( src, Dx, Dy, aperture_size, borderType, CV_32F, scale );
}
else
#endif
if( aperture_size > 0 )
{
Sobel( src, Dx, CV_32F, 1, 0, aperture_size, scale, 0, borderType );
Sobel( src, Dy, CV_32F, 0, 1, aperture_size, scale, 0, borderType );
}
else
{
Scharr( src, Dx, CV_32F, 1, 0, scale, 0, borderType );
Scharr( src, Dy, CV_32F, 0, 1, scale, 0, borderType );
}
spatialGradient( src, Dx, Dy, aperture_size, borderType, CV_32F, scale );
Size size = src.size();
Mat cov( size, CV_32FC3 );
@@ -376,18 +358,7 @@ static bool extractCovData(InputArray _src, UMat & Dx, UMat & Dy, int depth,
return k.run(2, globalsize, localsize, false);
}
else
{
if (aperture_size > 0)
{
Sobel(_src, Dx, CV_32F, 1, 0, aperture_size, scale, 0, borderType);
Sobel(_src, Dy, CV_32F, 0, 1, aperture_size, scale, 0, borderType);
}
else
{
Scharr(_src, Dx, CV_32F, 1, 0, scale, 0, borderType);
Scharr(_src, Dy, CV_32F, 0, 1, scale, 0, borderType);
}
}
spatialGradient(_src, Dx, Dy, aperture_size, borderType, CV_32F, scale);
return true;
}

6
modules/imgproc/src/generalized_hough.cpp
View File

@@ -116,13 +116,7 @@ namespace
CV_Assert( cannyLowThresh_ > 0 && cannyLowThresh_ < cannyHighThresh_ );
Canny(src, edges, cannyLowThresh_, cannyHighThresh_);
// TODO(follow-up PR): use the fused spatialGradient once its CV_32F fast path lands.
#if 0
spatialGradient(src, dx, dy, 3, BORDER_DEFAULT, CV_32F);
#else
Sobel(src, dx, CV_32F, 1, 0);
Sobel(src, dy, CV_32F, 0, 1);
#endif
}
void GeneralizedHoughBase::setTemplateImpl(InputArray templ, Point templCenter)

13
modules/imgproc/src/intelligent_scissors.cpp
View File

@@ -145,18 +145,7 @@ struct IntelligentScissorsMB::Impl
if (!Ix_.empty())
return;
initGrayscale_(image);
// TODO(follow-up PR): use the fused spatialGradient once its CV_32F/ksize=5 fast paths land.
#if 0
if (sobelKernelSize == 3 || sobelKernelSize == 5)
{
spatialGradient(grayscale_, Ix_, Iy_, sobelKernelSize, BORDER_DEFAULT, CV_32F);
}
else
#endif
{
Sobel(grayscale_, Ix_, CV_32FC1, 1, 0, sobelKernelSize);
Sobel(grayscale_, Iy_, CV_32FC1, 0, 1, sobelKernelSize);
}
spatialGradient(grayscale_, Ix_, Iy_, sobelKernelSize, BORDER_DEFAULT, CV_32F);
}
Mat image_magnitude_;
void initImageMagnitude_(InputArray image)

399
modules/imgproc/src/spatialgradient.cpp
View File

@@ -12,6 +12,7 @@
//
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
// Copyright (C) 2026, Advanced Micro Devices, Inc., all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
@@ -43,329 +44,141 @@
#include "precomp.hpp"
#include "opencv2/core/hal/intrin.hpp"
#include "spatialgradient.simd.hpp"
#include "spatialgradient.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL based on CMakeLists.txt content
namespace cv
{
/* NOTE:
*
* Sobel-x: -1 0 1
* -2 0 2
* -1 0 1
*
* Sobel-y: -1 -2 -1
* 0 0 0
* 1 2 1
*/
#if (CV_SIMD || CV_SIMD_SCALABLE)
template <typename T>
static inline void spatialGradientKernel_vec( T& vx, T& vy,
const T& v00, const T& v01, const T& v02,
const T& v10, const T& v12,
const T& v20, const T& v21, const T& v22 )
// Baseline dispatchers: pick the best CPU variant of each fused kernel at runtime.
static void dispatchSep3x3_16s(const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
short* dx, short* dy, size_t dst_step, int borderType)
{
// vx = (v22 - v00) + (v02 - v20) + 2 * (v12 - v10)
// vy = (v22 - v00) + (v20 - v02) + 2 * (v21 - v01)
T tmp_add = v_sub(v22, v00),
tmp_sub = v_sub(v02, v20),
tmp_x = v_sub(v12, v10),
tmp_y = v_sub(v21, v01);
vx = v_add(v_add(v_add(tmp_add, tmp_sub), tmp_x), tmp_x);
vy = v_add(v_add(v_sub(tmp_add, tmp_sub), tmp_y), tmp_y);
}
#endif
template <typename T>
static inline void spatialGradientKernel( T& vx, T& vy,
const T& v00, const T& v01, const T& v02,
const T& v10, const T& v12,
const T& v20, const T& v21, const T& v22 )
{
// vx = (v22 - v00) + (v02 - v20) + 2 * (v12 - v10)
// vy = (v22 - v00) + (v20 - v02) + 2 * (v21 - v01)
T tmp_add = v22 - v00,
tmp_sub = v02 - v20,
tmp_x = v12 - v10,
tmp_y = v21 - v01;
vx = tmp_add + tmp_sub + tmp_x + tmp_x;
vy = tmp_add - tmp_sub + tmp_y + tmp_y;
CV_CPU_DISPATCH(spatialGradientSep3x3_16s,
(src, src_step, srcRows, srcCols, rowStart, rowEnd, dx, dy, dst_step, borderType),
CV_CPU_DISPATCH_MODES_ALL);
}
#if defined(__GNUC__) || defined(__clang__)
# define CV_SPATIALGRAD_NOINLINE __attribute__((noinline))
#elif defined(_MSC_VER)
# define CV_SPATIALGRAD_NOINLINE __declspec(noinline)
#else
# define CV_SPATIALGRAD_NOINLINE
#endif
static void dispatchSep5x5_16s(const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
short* dx, short* dy, size_t dst_step, int borderType)
{
CV_CPU_DISPATCH(spatialGradientSep5x5_16s,
(src, src_step, srcRows, srcCols, rowStart, rowEnd, dx, dy, dst_step, borderType),
CV_CPU_DISPATCH_MODES_ALL);
}
// Fused single-pass 3x3 Sobel over the whole image: 8-bit input, CV_16S output,
// unit scale, BORDER_REFLECT_101 / BORDER_REPLICATE. Kept in its own (non-inlined)
// function so the hot vectorized loop is codegen'd in isolation from the public
// entry point's dispatch/fallback logic. Folding both into one frame perturbs loop
// alignment and costs ~10% on some microarchitectures (observed on AMD Zen5/Turin).
static CV_SPATIALGRAD_NOINLINE
void spatialGradientFused3x3_8u16s( Mat src, OutputArray _dx, OutputArray _dy, int bt );
static void dispatchSep3x3_32f(const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
float* dx, float* dy, size_t dst_step, float scale, int borderType)
{
CV_CPU_DISPATCH(spatialGradientSep3x3_32f,
(src, src_step, srcRows, srcCols, rowStart, rowEnd, dx, dy, dst_step, scale, borderType),
CV_CPU_DISPATCH_MODES_ALL);
}
static void dispatchSep5x5_32f(const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
float* dx, float* dy, size_t dst_step, float scale, int borderType)
{
CV_CPU_DISPATCH(spatialGradientSep5x5_32f,
(src, src_step, srcRows, srcCols, rowStart, rowEnd, dx, dy, dst_step, scale, borderType),
CV_CPU_DISPATCH_MODES_ALL);
}
// 3x3 CV_8U -> CV_16S via cv_hal_spatialGradient when a platform HAL provides it
// (e.g. RISC-V RVV). Returns false if the HAL is not implemented so the caller
// falls back to the fused separable kernel.
static bool spatialGradient3x3_tryHal( const Mat& src, OutputArray _dx, OutputArray _dy, int bt )
{
_dx.create( src.size(), CV_16SC1 );
_dy.create( src.size(), CV_16SC1 );
Mat dx = _dx.getMat(), dy = _dy.getMat();
int res = cv_hal_spatialGradient( src.data, src.step,
dx.ptr<short>(), dx.step,
dy.ptr<short>(), dy.step,
src.cols, src.rows, 3, bt );
if ( res == CV_HAL_ERROR_OK )
return true;
if ( res != CV_HAL_ERROR_NOT_IMPLEMENTED )
CV_Error_( cv::Error::StsInternal,
("HAL implementation spatialGradient ==> cv_hal_spatialGradient returned %d (0x%08x)", res, res) );
return false;
}
void spatialGradient( InputArray _src, OutputArray _dx, OutputArray _dy,
int ksize, int borderType, int ddepth, double scale )
{
CV_INSTRUMENT_REGION();
// Prepare InputArray src
Mat src = _src.getMat();
CV_Assert( !src.empty() );
CV_Assert( ksize == -1 || ksize == 3 || ksize == 5 || ksize == 7 );
CV_Assert( ksize == -1 || ksize == 1 || ksize == 3 || ksize == 5 || ksize == 7 );
CV_Assert( ddepth == CV_16S || ddepth == CV_32F );
// Fused single-pass 3x3 Sobel fast path (existing vectorized kernel): 8-bit source,
// CV_16S output, unit scale, whole-image (non-ROI), BORDER_REFLECT_101/REPLICATE.
// NOTE: fused fast paths for CV_32F output and ksize == 5 (plus ROI-aware slicing and
// other border types) are added in a follow-up PR; until then those cases are computed
// with the two equivalent cv::Sobel() passes below.
Size wholeSize;
Point ofs;
src.locateROI( wholeSize, ofs );
const bool entireParent = ( ofs.x == 0 && ofs.y == 0 &&
src.cols == wholeSize.width && src.rows == wholeSize.height );
const bool isolated = ( borderType & BORDER_ISOLATED ) != 0;
const int bt = borderType & ~BORDER_ISOLATED;
const int bt = borderType & ~BORDER_ISOLATED;
const bool fastPath = ( ksize == 3 && ddepth == CV_16S && scale == 1.0
// (1) Optional platform HAL for whole-image 3x3 CV_8U->CV_16S at unit scale.
const bool halEligible = ( ksize == 3 && ddepth == CV_16S && scale == 1.0
&& src.type() == CV_8UC1 && entireParent && !isolated
&& ( bt == BORDER_DEFAULT || bt == BORDER_REPLICATE ) );
&& ( bt == BORDER_REFLECT_101 || bt == BORDER_REPLICATE ) );
if ( halEligible && spatialGradient3x3_tryHal( src, _dx, _dy, bt ) )
return;
if ( !fastPath )
// (2) Fused separable single-pass path (CV_8U source): 3x3/5x5, CV_16S (unit scale) and
// CV_32F (any scale), reflect/replicate borders and full-width row-range ROIs. Each
// source sample is read once and shared between dx and dy; output is bit-identical to
// two cv::Sobel() passes. CV_16S fuses only at unit scale (the kernel emits the
// unscaled int16 result); a scaled int16 request falls back to keep cv::Sobel rounding.
const bool fusableBorder = ( bt == BORDER_REPLICATE || bt == BORDER_REFLECT
|| bt == BORDER_REFLECT_101 );
const bool fullWidthRoi = ( ofs.x == 0 && src.cols == wholeSize.width );
const bool scaleOk = ( ddepth == CV_32F ) || ( scale == 1.0 );
if ( (ksize == 3 || ksize == 5) && src.type() == CV_8UC1 && fusableBorder && fullWidthRoi && !isolated && scaleOk )
{
Sobel( _src, _dx, ddepth, 1, 0, ksize, scale, 0, borderType );
Sobel( _src, _dy, ddepth, 0, 1, ksize, scale, 0, borderType );
_dx.create( src.size(), CV_MAKETYPE(ddepth, 1) );
_dy.create( src.size(), CV_MAKETYPE(ddepth, 1) );
Mat dx = _dx.getMat(), dy = _dy.getMat();
const uchar* base = src.ptr<uchar>(0) - (size_t)ofs.y * src.step;
const int parentRows = wholeSize.height;
const int rowStart = ofs.y;
const int rowEnd = ofs.y + src.rows;
if ( ddepth == CV_32F )
{
const float fscale = (float)scale;
if ( ksize == 3 )
dispatchSep3x3_32f( base, src.step, parentRows, src.cols, rowStart, rowEnd,
dx.ptr<float>(), dy.ptr<float>(), dx.step1(), fscale, bt );
else
dispatchSep5x5_32f( base, src.step, parentRows, src.cols, rowStart, rowEnd,
dx.ptr<float>(), dy.ptr<float>(), dx.step1(), fscale, bt );
}
else
{
if ( ksize == 3 )
dispatchSep3x3_16s( base, src.step, parentRows, src.cols, rowStart, rowEnd,
dx.ptr<short>(), dy.ptr<short>(), dx.step1(), bt );
else
dispatchSep5x5_16s( base, src.step, parentRows, src.cols, rowStart, rowEnd,
dx.ptr<short>(), dy.ptr<short>(), dx.step1(), bt );
}
return;
}
spatialGradientFused3x3_8u16s( src, _dx, _dy, bt );
}
static CV_SPATIALGRAD_NOINLINE
void spatialGradientFused3x3_8u16s( Mat src, OutputArray _dx, OutputArray _dy, int bt )
{
// Prepare OutputArrays dx, dy
_dx.create( src.size(), CV_16SC1 );
_dy.create( src.size(), CV_16SC1 );
Mat dx = _dx.getMat(),
dy = _dy.getMat();
CALL_HAL(spatialGradient, cv_hal_spatialGradient,
src.data, src.step,
dx.ptr<short>(), dx.step,
dy.ptr<short>(), dy.step,
src.cols, src.rows,
3, bt);
// Get dimensions
const int H = src.rows,
W = src.cols;
// Row, column indices
int i = 0,
j = 0;
// Handle border types
int i_top = 0, // Case for H == 1 && W == 1 && BORDER_REPLICATE
i_bottom = H - 1,
j_offl = 0, // j offset from 0th pixel to reach -1st pixel
j_offr = 0; // j offset from W-1th pixel to reach Wth pixel
if ( bt == BORDER_DEFAULT ) // Equiv. to BORDER_REFLECT_101
{
if ( H > 1 )
{
i_top = 1;
i_bottom = H - 2;
}
if ( W > 1 )
{
j_offl = 1;
j_offr = -1;
}
}
int i_start = 0;
int j_start = 0;
#if (CV_SIMD || CV_SIMD_SCALABLE)
// Characters in variable names have the following meanings:
// u: unsigned char
// s: signed int
//
// [row][column]
// m: offset -1
// n: offset 0
// p: offset 1
// Example: umn is offset -1 in row and offset 0 in column
for ( i = 0; i < H - 1; i += 2 )
{
uchar *p_src = src.ptr<uchar>(i == 0 ? i_top : i - 1);
uchar *c_src = src.ptr<uchar>(i);
uchar *n_src = src.ptr<uchar>(i+1);
uchar *m_src = src.ptr<uchar>(i == H - 2 ? i_bottom : i + 2);
short *c_dx = dx.ptr<short>(i);
short *c_dy = dy.ptr<short>(i);
short *n_dx = dx.ptr<short>(i+1);
short *n_dy = dy.ptr<short>(i+1);
// Process rest of columns 16-column chunks at a time
for ( j = 1; j < W - VTraits<v_uint8>::vlanes(); j += VTraits<v_uint8>::vlanes())
{
// Load top row for 3x3 Sobel filter
v_uint8 v_um = vx_load(&p_src[j-1]);
v_uint8 v_un = vx_load(&p_src[j]);
v_uint8 v_up = vx_load(&p_src[j+1]);
v_uint16 v_um1, v_um2, v_un1, v_un2, v_up1, v_up2;
v_expand(v_um, v_um1, v_um2);
v_expand(v_un, v_un1, v_un2);
v_expand(v_up, v_up1, v_up2);
v_int16 v_s1m1 = v_reinterpret_as_s16(v_um1);
v_int16 v_s1m2 = v_reinterpret_as_s16(v_um2);
v_int16 v_s1n1 = v_reinterpret_as_s16(v_un1);
v_int16 v_s1n2 = v_reinterpret_as_s16(v_un2);
v_int16 v_s1p1 = v_reinterpret_as_s16(v_up1);
v_int16 v_s1p2 = v_reinterpret_as_s16(v_up2);
// Load second row for 3x3 Sobel filter
v_um = vx_load(&c_src[j-1]);
v_un = vx_load(&c_src[j]);
v_up = vx_load(&c_src[j+1]);
v_expand(v_um, v_um1, v_um2);
v_expand(v_un, v_un1, v_un2);
v_expand(v_up, v_up1, v_up2);
v_int16 v_s2m1 = v_reinterpret_as_s16(v_um1);
v_int16 v_s2m2 = v_reinterpret_as_s16(v_um2);
v_int16 v_s2n1 = v_reinterpret_as_s16(v_un1);
v_int16 v_s2n2 = v_reinterpret_as_s16(v_un2);
v_int16 v_s2p1 = v_reinterpret_as_s16(v_up1);
v_int16 v_s2p2 = v_reinterpret_as_s16(v_up2);
// Load third row for 3x3 Sobel filter
v_um = vx_load(&n_src[j-1]);
v_un = vx_load(&n_src[j]);
v_up = vx_load(&n_src[j+1]);
v_expand(v_um, v_um1, v_um2);
v_expand(v_un, v_un1, v_un2);
v_expand(v_up, v_up1, v_up2);
v_int16 v_s3m1 = v_reinterpret_as_s16(v_um1);
v_int16 v_s3m2 = v_reinterpret_as_s16(v_um2);
v_int16 v_s3n1 = v_reinterpret_as_s16(v_un1);
v_int16 v_s3n2 = v_reinterpret_as_s16(v_un2);
v_int16 v_s3p1 = v_reinterpret_as_s16(v_up1);
v_int16 v_s3p2 = v_reinterpret_as_s16(v_up2);
// dx & dy for rows 1, 2, 3
v_int16 v_sdx1, v_sdy1;
spatialGradientKernel_vec<v_int16>( v_sdx1, v_sdy1,
v_s1m1, v_s1n1, v_s1p1,
v_s2m1, v_s2p1,
v_s3m1, v_s3n1, v_s3p1 );
v_int16 v_sdx2, v_sdy2;
spatialGradientKernel_vec<v_int16>( v_sdx2, v_sdy2,
v_s1m2, v_s1n2, v_s1p2,
v_s2m2, v_s2p2,
v_s3m2, v_s3n2, v_s3p2 );
// Store
v_store(&c_dx[j], v_sdx1);
v_store(&c_dx[j+VTraits<v_int16>::vlanes()], v_sdx2);
v_store(&c_dy[j], v_sdy1);
v_store(&c_dy[j+VTraits<v_int16>::vlanes()], v_sdy2);
// Load fourth row for 3x3 Sobel filter
v_um = vx_load(&m_src[j-1]);
v_un = vx_load(&m_src[j]);
v_up = vx_load(&m_src[j+1]);
v_expand(v_um, v_um1, v_um2);
v_expand(v_un, v_un1, v_un2);
v_expand(v_up, v_up1, v_up2);
v_int16 v_s4m1 = v_reinterpret_as_s16(v_um1);
v_int16 v_s4m2 = v_reinterpret_as_s16(v_um2);
v_int16 v_s4n1 = v_reinterpret_as_s16(v_un1);
v_int16 v_s4n2 = v_reinterpret_as_s16(v_un2);
v_int16 v_s4p1 = v_reinterpret_as_s16(v_up1);
v_int16 v_s4p2 = v_reinterpret_as_s16(v_up2);
// dx & dy for rows 2, 3, 4
spatialGradientKernel_vec<v_int16>( v_sdx1, v_sdy1,
v_s2m1, v_s2n1, v_s2p1,
v_s3m1, v_s3p1,
v_s4m1, v_s4n1, v_s4p1 );
spatialGradientKernel_vec<v_int16>( v_sdx2, v_sdy2,
v_s2m2, v_s2n2, v_s2p2,
v_s3m2, v_s3p2,
v_s4m2, v_s4n2, v_s4p2 );
// Store
v_store(&n_dx[j], v_sdx1);
v_store(&n_dx[j+VTraits<v_int16>::vlanes()], v_sdx2);
v_store(&n_dy[j], v_sdy1);
v_store(&n_dy[j+VTraits<v_int16>::vlanes()], v_sdy2);
}
}
i_start = i;
j_start = j;
#endif
int j_p, j_n;
uchar v00, v01, v02, v10, v11, v12, v20, v21, v22;
for ( i = 0; i < H; i++ )
{
uchar *p_src = src.ptr<uchar>(i == 0 ? i_top : i - 1);
uchar *c_src = src.ptr<uchar>(i);
uchar *n_src = src.ptr<uchar>(i == H - 1 ? i_bottom : i + 1);
short *c_dx = dx.ptr<short>(i);
short *c_dy = dy.ptr<short>(i);
// Process left-most column
j = 0;
j_p = j + j_offl;
j_n = 1;
if ( j_n >= W ) j_n = j + j_offr;
v00 = p_src[j_p]; v01 = p_src[j]; v02 = p_src[j_n];
v10 = c_src[j_p]; v11 = c_src[j]; v12 = c_src[j_n];
v20 = n_src[j_p]; v21 = n_src[j]; v22 = n_src[j_n];
spatialGradientKernel<short>( c_dx[0], c_dy[0], v00, v01, v02, v10,
v12, v20, v21, v22 );
v00 = v01; v10 = v11; v20 = v21;
v01 = v02; v11 = v12; v21 = v22;
// Process middle columns
j = i >= i_start ? 1 : j_start;
j_p = j - 1;
v00 = p_src[j_p]; v01 = p_src[j];
v10 = c_src[j_p]; v11 = c_src[j];
v20 = n_src[j_p]; v21 = n_src[j];
for ( ; j < W - 1; j++ )
{
// Get values for next column
j_n = j + 1; v02 = p_src[j_n]; v12 = c_src[j_n]; v22 = n_src[j_n];
spatialGradientKernel<short>( c_dx[j], c_dy[j], v00, v01, v02, v10,
v12, v20, v21, v22 );
// Move values back one column for next iteration
v00 = v01; v10 = v11; v20 = v21;
v01 = v02; v11 = v12; v21 = v22;
}
// Process right-most column
if ( j < W )
{
j_n = j + j_offr; v02 = p_src[j_n]; v12 = c_src[j_n]; v22 = n_src[j_n];
spatialGradientKernel<short>( c_dx[j], c_dy[j], v00, v01, v02, v10,
v12, v20, v21, v22 );
}
}
// (3) General fallback: two separable cv::Sobel passes.
Sobel( _src, _dx, ddepth, 1, 0, ksize, scale, 0, borderType );
Sobel( _src, _dy, ddepth, 0, 1, ksize, scale, 0, borderType );
}
}

524
modules/imgproc/src/spatialgradient.simd.hpp Normal file
View File

@@ -0,0 +1,524 @@
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html
// Copyright (C) 2026, Advanced Micro Devices, Inc., all rights reserved.
#include "precomp.hpp"
#include "opencv2/core/hal/intrin.hpp"
namespace cv {
CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN
// Fused single-pass Sobel gradient entry points (per-CPU-variant dispatched).
// Row bounds [rowStart, rowEnd) are absolute in the parent buffer so a full-width
// ROI reads the real neighbouring rows at its top/bottom edges (matches cv::Sobel).
// dst_step is in elements. Output is bit-identical to two cv::Sobel() passes.
void spatialGradientSep3x3_16s(const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
short* dx, short* dy, size_t dst_step, int borderType);
void spatialGradientSep5x5_16s(const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
short* dx, short* dy, size_t dst_step, int borderType);
void spatialGradientSep3x3_32f(const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
float* dx, float* dy, size_t dst_step, float scale, int borderType);
void spatialGradientSep5x5_32f(const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
float* dx, float* dy, size_t dst_step, float scale, int borderType);
#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
// Output store for the fused kernels.
// - short: store the int16 result (scale is always 1 here).
// - float: widen to int32, convert to float, multiply by scale, store.
template<typename DstT> struct SobelSepStore;
template<> struct SobelSepStore<short>
{
explicit SobelSepStore(double) {}
#if (CV_SIMD || CV_SIMD_SCALABLE)
inline void store(short* p, int j, const v_int16& v) const { v_store(p + j, v); }
#endif
inline void scalar(short* p, int j, int v) const { p[j] = (short)v; }
};
template<> struct SobelSepStore<float>
{
float s;
explicit SobelSepStore(double scale) : s((float)scale) {}
#if (CV_SIMD || CV_SIMD_SCALABLE)
inline void store(float* p, int j, const v_int16& v) const
{
v_int32 lo, hi;
v_expand(v, lo, hi);
v_float32 vs = vx_setall_f32(s);
v_store(p + j, v_mul(v_cvt_f32(lo), vs));
v_store(p + j + VTraits<v_float32>::vlanes(), v_mul(v_cvt_f32(hi), vs));
}
#endif
inline void scalar(float* p, int j, int v) const { p[j] = s * (float)v; }
};
// Fused 3x3 Sobel:
// hd = src[x+1] - src[x-1]; hs = src[x-1] + 2*src[x] + src[x+1]
// dx = hd[y-1] + 2*hd[y] + hd[y+1]; dy = hs[y+1] - hs[y-1]
template<typename DstT>
static void sobelSep3x3_( const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
DstT* dx, DstT* dy, size_t dst_step, double scale, int borderType )
{
const int width = srcCols;
const int outRows = rowEnd - rowStart;
if (width <= 0 || outRows <= 0)
return;
const SobelSepStore<DstT> t_store(scale);
cv::AutoBuffer<short> _buf((size_t)width * 6);
short* hd[3];
short* hs[3];
for (int k = 0; k < 3; ++k)
{
hd[k] = _buf.data() + (size_t)k * 2 * width;
hs[k] = hd[k] + width;
}
auto cidx = [&](int x) { return borderInterpolate(x, width, borderType); };
auto horiz = [&](const uchar* s, short* hdo, short* hso)
{
{
int l = s[cidx(-1)], c = s[0], r = s[cidx(1)];
hdo[0] = (short)(r - l);
hso[0] = (short)(l + 2 * c + r);
}
int j = 1;
#if (CV_SIMD || CV_SIMD_SCALABLE)
for (; j + VTraits<v_int16>::vlanes() <= width - 1; j += VTraits<v_int16>::vlanes())
{
v_int16 c = v_reinterpret_as_s16(vx_load_expand(s + j));
v_int16 l = v_reinterpret_as_s16(vx_load_expand(s + j - 1));
v_int16 r = v_reinterpret_as_s16(vx_load_expand(s + j + 1));
v_store(hdo + j, v_sub(r, l));
v_store(hso + j, v_add(v_add(l, r), v_add(c, c)));
}
#endif
for (; j < width - 1; ++j)
{
int l = s[j - 1], c = s[j], r = s[j + 1];
hdo[j] = (short)(r - l);
hso[j] = (short)(l + 2 * c + r);
}
if (width > 1)
{
int x = width - 1;
int l = s[x - 1], c = s[x], r = s[cidx(x + 1)];
hdo[x] = (short)(r - l);
hso[x] = (short)(l + 2 * c + r);
}
};
auto combine = [&](const short* hdm, const short* hd0, const short* hdp,
const short* hsm, const short* hsp,
DstT* odx, DstT* ody)
{
int j = 0;
#if (CV_SIMD || CV_SIMD_SCALABLE)
for (; j <= width - VTraits<v_int16>::vlanes(); j += VTraits<v_int16>::vlanes())
{
v_int16 a = vx_load(hdm + j);
v_int16 b = vx_load(hd0 + j);
v_int16 c = vx_load(hdp + j);
t_store.store(odx, j, v_add(v_add(a, c), v_add(b, b)));
v_int16 p = vx_load(hsp + j);
v_int16 m = vx_load(hsm + j);
t_store.store(ody, j, v_sub(p, m));
}
#endif
for (; j < width; ++j)
{
t_store.scalar(odx, j, hdm[j] + 2 * hd0[j] + hdp[j]);
t_store.scalar(ody, j, hsp[j] - hsm[j]);
}
};
for (int r = rowStart - 1; r <= rowEnd; ++r)
{
int rc = borderInterpolate(r, srcRows, borderType);
int slot = ((r % 3) + 3) % 3;
horiz(src + (size_t)rc * src_step, hd[slot], hs[slot]);
int a = r - 1;
if (a >= rowStart && a < rowEnd)
{
int sm = (((a - 1) % 3) + 3) % 3;
int s0 = ((a % 3) + 3) % 3;
int sp = (((a + 1) % 3) + 3) % 3;
combine(hd[sm], hd[s0], hd[sp], hs[sm], hs[sp],
dx + (size_t)(a - rowStart) * dst_step,
dy + (size_t)(a - rowStart) * dst_step);
}
}
}
// Fused 5x5 Sobel (getSobelKernels ksize=5): smoothing [1,4,6,4,1], derivative [-1,-2,0,2,1].
// hd = s[x+2] + 2*s[x+1] - 2*s[x-1] - s[x-2]
// hs = s[x-2] + 4*s[x-1] + 6*s[x] + 4*s[x+1] + s[x+2]
// dx = hd[y-2] + 4*hd[y-1] + 6*hd[y] + 4*hd[y+1] + hd[y+2]
// dy = -hs[y-2] - 2*hs[y-1] + 2*hs[y+1] + hs[y+2]
template<typename DstT>
static void sobelSep5x5_( const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
DstT* dx, DstT* dy, size_t dst_step, double scale, int borderType )
{
const int width = srcCols;
const int outRows = rowEnd - rowStart;
if (width <= 0 || outRows <= 0)
return;
const SobelSepStore<DstT> t_store(scale);
cv::AutoBuffer<short> _buf((size_t)width * 10);
short* hd[5];
short* hs[5];
for (int k = 0; k < 5; ++k)
{
hd[k] = _buf.data() + (size_t)k * 2 * width;
hs[k] = hd[k] + width;
}
auto cidx = [&](int x) { return borderInterpolate(x, width, borderType); };
auto horiz = [&](const uchar* s, short* hdo, short* hso)
{
auto scalarCol = [&](int x)
{
int m2 = s[cidx(x - 2)], m1 = s[cidx(x - 1)], c = s[x];
int p1 = s[cidx(x + 1)], p2 = s[cidx(x + 2)];
hdo[x] = (short)(p2 + 2 * p1 - 2 * m1 - m2);
hso[x] = (short)(m2 + 4 * m1 + 6 * c + 4 * p1 + p2);
};
int j = 0;
for (; j < 2 && j < width; ++j)
scalarCol(j);
#if (CV_SIMD || CV_SIMD_SCALABLE)
for (; j + VTraits<v_int16>::vlanes() <= width - 2; j += VTraits<v_int16>::vlanes())
{
v_int16 m2 = v_reinterpret_as_s16(vx_load_expand(s + j - 2));
v_int16 m1 = v_reinterpret_as_s16(vx_load_expand(s + j - 1));
v_int16 c = v_reinterpret_as_s16(vx_load_expand(s + j));
v_int16 p1 = v_reinterpret_as_s16(vx_load_expand(s + j + 1));
v_int16 p2 = v_reinterpret_as_s16(vx_load_expand(s + j + 2));
v_store(hdo + j, v_sub(v_sub(v_add(p2, v_add(p1, p1)), v_add(m1, m1)), m2));
v_int16 c6 = v_add(v_shl<2>(c), v_add(c, c));
v_store(hso + j, v_add(v_add(v_add(m2, p2), v_shl<2>(v_add(m1, p1))), c6));
}
#endif
for (; j < width; ++j)
scalarCol(j);
};
auto combine = [&](const short* hdm2, const short* hdm1, const short* hd0,
const short* hdp1, const short* hdp2,
const short* hsm2, const short* hsm1,
const short* hsp1, const short* hsp2,
DstT* odx, DstT* ody)
{
int j = 0;
#if (CV_SIMD || CV_SIMD_SCALABLE)
for (; j <= width - VTraits<v_int16>::vlanes(); j += VTraits<v_int16>::vlanes())
{
v_int16 a = vx_load(hdm2 + j), b = vx_load(hdm1 + j), c = vx_load(hd0 + j);
v_int16 d = vx_load(hdp1 + j), e = vx_load(hdp2 + j);
v_int16 c6 = v_add(v_shl<2>(c), v_add(c, c));
t_store.store(odx, j, v_add(v_add(v_add(a, e), v_shl<2>(v_add(b, d))), c6));
v_int16 g = vx_load(hsm2 + j), h = vx_load(hsm1 + j);
v_int16 p = vx_load(hsp1 + j), q = vx_load(hsp2 + j);
t_store.store(ody, j, v_add(v_sub(q, g), v_shl<1>(v_sub(p, h))));
}
#endif
for (; j < width; ++j)
{
t_store.scalar(odx, j, hdm2[j] + 4 * hdm1[j] + 6 * hd0[j] + 4 * hdp1[j] + hdp2[j]);
t_store.scalar(ody, j, -hsm2[j] - 2 * hsm1[j] + 2 * hsp1[j] + hsp2[j]);
}
};
for (int r = rowStart - 2; r <= rowEnd + 1; ++r)
{
int rc = borderInterpolate(r, srcRows, borderType);
int slot = ((r % 5) + 5) % 5;
horiz(src + (size_t)rc * src_step, hd[slot], hs[slot]);
int a = r - 2;
if (a >= rowStart && a < rowEnd)
{
int s_m2 = (((a - 2) % 5) + 5) % 5;
int s_m1 = (((a - 1) % 5) + 5) % 5;
int s_0 = ((a % 5) + 5) % 5;
int s_p1 = (((a + 1) % 5) + 5) % 5;
int s_p2 = (((a + 2) % 5) + 5) % 5;
combine(hd[s_m2], hd[s_m1], hd[s_0], hd[s_p1], hd[s_p2],
hs[s_m2], hs[s_m1], hs[s_p1], hs[s_p2],
dx + (size_t)(a - rowStart) * dst_step,
dy + (size_t)(a - rowStart) * dst_step);
}
}
}
// CV_32F output must match cv::Sobel(..., CV_32F, ..., scale, ...), which folds scale
// into the smoothing kernel before sepFilter2D (not as a final post-multiply). The float
// horizontal pass carries scale so the result is bit-exact with cv::Sobel.
static void sobelSep3x3_f32( const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
float* dx, float* dy, size_t dst_step, float scale, int borderType )
{
const int width = srcCols;
if (width <= 0 || rowEnd <= rowStart)
return;
const float sk0 = 2.f * scale;
const float sk1 = scale;
const float dk0 = sk0;
const float dk1 = sk1;
cv::AutoBuffer<float> _buf((size_t)width * 6);
float* hd[3];
float* hs[3];
for (int k = 0; k < 3; ++k)
{
hd[k] = _buf.data() + (size_t)k * 2 * width;
hs[k] = hd[k] + width;
}
auto cidx = [&](int x) { return borderInterpolate(x, width, borderType); };
auto horiz = [&](const uchar* s, float* hdo, float* hso)
{
{
float l = (float)s[cidx(-1)], c = (float)s[0], r = (float)s[cidx(1)];
hdo[0] = r - l;
hso[0] = sk0 * c + sk1 * (l + r);
}
int j = 1;
#if (CV_SIMD || CV_SIMD_SCALABLE)
v_float32 v_sk0 = vx_setall_f32(sk0);
v_float32 v_sk1 = vx_setall_f32(sk1);
for (; j + VTraits<v_float32>::vlanes() <= width - 1; j += VTraits<v_float32>::vlanes())
{
v_float32 l = v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(s + j - 1)));
v_float32 c = v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(s + j)));
v_float32 r = v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(s + j + 1)));
v_store(hdo + j, v_sub(r, l));
v_store(hso + j, v_muladd(v_add(l, r), v_sk1, v_mul(c, v_sk0)));
}
#endif
for (; j < width - 1; ++j)
{
float l = (float)s[j - 1], c = (float)s[j], r = (float)s[j + 1];
hdo[j] = r - l;
hso[j] = sk0 * c + sk1 * (l + r);
}
if (width > 1)
{
int x = width - 1;
float l = (float)s[x - 1], c = (float)s[x], r = (float)s[cidx(x + 1)];
hdo[x] = r - l;
hso[x] = sk0 * c + sk1 * (l + r);
}
};
auto combine = [&](const float* hdm, const float* hd0, const float* hdp,
const float* hsm, const float* hsp,
float* odx, float* ody)
{
int j = 0;
#if (CV_SIMD || CV_SIMD_SCALABLE)
v_float32 v_dk0 = vx_setall_f32(dk0);
v_float32 v_dk1 = vx_setall_f32(dk1);
for (; j <= width - VTraits<v_float32>::vlanes(); j += VTraits<v_float32>::vlanes())
{
v_float32 hd0v = vx_load(hd0 + j);
v_store(odx + j, v_muladd(v_add(vx_load(hdm + j), vx_load(hdp + j)), v_dk1, v_mul(hd0v, v_dk0)));
v_store(ody + j, v_sub(vx_load(hsp + j), vx_load(hsm + j)));
}
#endif
for (; j < width; ++j)
{
odx[j] = dk0 * hd0[j] + dk1 * (hdm[j] + hdp[j]);
ody[j] = hsp[j] - hsm[j];
}
};
for (int r = rowStart - 1; r <= rowEnd; ++r)
{
int rc = borderInterpolate(r, srcRows, borderType);
int slot = ((r % 3) + 3) % 3;
horiz(src + (size_t)rc * src_step, hd[slot], hs[slot]);
int a = r - 1;
if (a >= rowStart && a < rowEnd)
{
int sm = (((a - 1) % 3) + 3) % 3;
int s0 = ((a % 3) + 3) % 3;
int sp = (((a + 1) % 3) + 3) % 3;
combine(hd[sm], hd[s0], hd[sp], hs[sm], hs[sp],
dx + (size_t)(a - rowStart) * dst_step,
dy + (size_t)(a - rowStart) * dst_step);
}
}
}
static void sobelSep5x5_f32( const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
float* dx, float* dy, size_t dst_step, float scale, int borderType )
{
const int width = srcCols;
if (width <= 0 || rowEnd <= rowStart)
return;
const float sk0 = 6.f * scale;
const float sk1 = 4.f * scale;
const float sk2 = scale;
const float dk0 = sk0;
const float dk1 = sk1;
const float dk2 = sk2;
cv::AutoBuffer<float> _buf((size_t)width * 10);
float* hd[5];
float* hs[5];
for (int k = 0; k < 5; ++k)
{
hd[k] = _buf.data() + (size_t)k * 2 * width;
hs[k] = hd[k] + width;
}
auto cidx = [&](int x) { return borderInterpolate(x, width, borderType); };
auto horiz = [&](const uchar* s, float* hdo, float* hso)
{
auto scalarCol = [&](int x)
{
float m2 = (float)s[cidx(x - 2)], m1 = (float)s[cidx(x - 1)], c = (float)s[x];
float p1 = (float)s[cidx(x + 1)], p2 = (float)s[cidx(x + 2)];
hdo[x] = p2 + 2.f * p1 - 2.f * m1 - m2;
hso[x] = sk2 * (m2 + p2) + sk1 * (m1 + p1) + sk0 * c;
};
int j = 0;
for (; j < 2 && j < width; ++j)
scalarCol(j);
#if (CV_SIMD || CV_SIMD_SCALABLE)
v_float32 v_sk0 = vx_setall_f32(sk0);
v_float32 v_sk1 = vx_setall_f32(sk1);
v_float32 v_sk2 = vx_setall_f32(sk2);
for (; j + VTraits<v_float32>::vlanes() <= width - 2; j += VTraits<v_float32>::vlanes())
{
v_float32 m2 = v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(s + j - 2)));
v_float32 m1 = v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(s + j - 1)));
v_float32 c = v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(s + j)));
v_float32 p1 = v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(s + j + 1)));
v_float32 p2 = v_cvt_f32(v_reinterpret_as_s32(vx_load_expand_q(s + j + 2)));
v_store(hdo + j, v_sub(v_sub(v_add(p2, v_add(p1, p1)), v_add(m1, m1)), m2));
v_store(hso + j, v_muladd(v_add(m2, p2), v_sk2,
v_muladd(v_add(m1, p1), v_sk1, v_mul(c, v_sk0))));
}
#endif
for (; j < width; ++j)
scalarCol(j);
};
auto combine = [&](const float* hdm2, const float* hdm1, const float* hd0,
const float* hdp1, const float* hdp2,
const float* hsm2, const float* hsm1,
const float* hsp1, const float* hsp2,
float* odx, float* ody)
{
int j = 0;
#if (CV_SIMD || CV_SIMD_SCALABLE)
v_float32 v_dk0 = vx_setall_f32(dk0);
v_float32 v_dk1 = vx_setall_f32(dk1);
v_float32 v_dk2 = vx_setall_f32(dk2);
v_float32 v2 = vx_setall_f32(2.f);
for (; j <= width - VTraits<v_float32>::vlanes(); j += VTraits<v_float32>::vlanes())
{
v_float32 hd0v = vx_load(hd0 + j);
v_store(odx + j, v_muladd(v_add(vx_load(hdm2 + j), vx_load(hdp2 + j)), v_dk2,
v_muladd(v_add(vx_load(hdm1 + j), vx_load(hdp1 + j)), v_dk1,
v_mul(hd0v, v_dk0))));
v_store(ody + j, v_add(v_sub(vx_load(hsp2 + j), vx_load(hsm2 + j)),
v_mul(v2, v_sub(vx_load(hsp1 + j), vx_load(hsm1 + j)))));
}
#endif
for (; j < width; ++j)
{
odx[j] = dk0 * hd0[j] + dk1 * (hdm1[j] + hdp1[j]) + dk2 * (hdm2[j] + hdp2[j]);
ody[j] = (hsp2[j] - hsm2[j]) + 2.f * (hsp1[j] - hsm1[j]);
}
};
for (int r = rowStart - 2; r <= rowEnd + 1; ++r)
{
int rc = borderInterpolate(r, srcRows, borderType);
int slot = ((r % 5) + 5) % 5;
horiz(src + (size_t)rc * src_step, hd[slot], hs[slot]);
int a = r - 2;
if (a >= rowStart && a < rowEnd)
{
int s_m2 = (((a - 2) % 5) + 5) % 5;
int s_m1 = (((a - 1) % 5) + 5) % 5;
int s_0 = ((a % 5) + 5) % 5;
int s_p1 = (((a + 1) % 5) + 5) % 5;
int s_p2 = (((a + 2) % 5) + 5) % 5;
combine(hd[s_m2], hd[s_m1], hd[s_0], hd[s_p1], hd[s_p2],
hs[s_m2], hs[s_m1], hs[s_p1], hs[s_p2],
dx + (size_t)(a - rowStart) * dst_step,
dy + (size_t)(a - rowStart) * dst_step);
}
}
}
void spatialGradientSep3x3_16s(const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
short* dx, short* dy, size_t dst_step, int borderType)
{
sobelSep3x3_<short>(src, src_step, srcRows, srcCols, rowStart, rowEnd,
dx, dy, dst_step, 1.0, borderType);
}
void spatialGradientSep5x5_16s(const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
short* dx, short* dy, size_t dst_step, int borderType)
{
sobelSep5x5_<short>(src, src_step, srcRows, srcCols, rowStart, rowEnd,
dx, dy, dst_step, 1.0, borderType);
}
void spatialGradientSep3x3_32f(const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
float* dx, float* dy, size_t dst_step, float scale, int borderType)
{
sobelSep3x3_f32(src, src_step, srcRows, srcCols, rowStart, rowEnd,
dx, dy, dst_step, scale, borderType);
}
void spatialGradientSep5x5_32f(const uchar* src, size_t src_step, int srcRows, int srcCols,
int rowStart, int rowEnd,
float* dx, float* dy, size_t dst_step, float scale, int borderType)
{
sobelSep5x5_f32(src, src_step, srcRows, srcCols, rowStart, rowEnd,
dx, dy, dst_step, scale, borderType);
}
#endif // CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
CV_CPU_OPTIMIZATION_NAMESPACE_END
} // namespace cv

4
modules/imgproc/test/test_filter.cpp
View File

@@ -1941,7 +1941,7 @@ TEST(Imgproc_SpatialGradient, fused_accuracy)
{
Mat parent(200, 160, CV_8UC1);
rng.fill(parent, RNG::UNIFORM, 0, 256);
for (int ks : {3, 5})
for (int ks : {1, 3, 5})
for (int b : {BORDER_DEFAULT, BORDER_REPLICATE, BORDER_REFLECT, BORDER_CONSTANT})
for (int ddepth : {CV_16S, CV_32F})
{
@@ -1957,7 +1957,7 @@ TEST(Imgproc_SpatialGradient, fused_accuracy)
// invalid aperture sizes must be rejected
Mat src(16, 16, CV_8UC1), dx, dy;
EXPECT_ANY_THROW(spatialGradient(src, dx, dy, 1));
EXPECT_ANY_THROW(spatialGradient(src, dx, dy, 2));
}
// Reproduces parallelCanny's per-slice row splitting and checks each slice's