From 84ea77a4be59732e8d641ab24e7c46f8dc888404 Mon Sep 17 00:00:00 2001 From: Madan mohan Manokar Date: Mon, 19 May 2025 11:26:57 +0530 Subject: [PATCH] Merge pull request #27299 from amd:fast_medianblur_simd imgproc: medianblur: Performance improvement #27299 * Bottleneck in non-vectorized path reduced. * AVX512 dispatch added for medianblur. ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV - [x] The PR is proposed to the proper branch - [ ] There is a reference to the original bug report and related work - [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [ ] The feature is well documented and sample code can be built with the project CMake --- modules/imgproc/CMakeLists.txt | 2 +- modules/imgproc/src/median_blur.simd.hpp | 21 +++++++++++++++++++-- 2 files changed, 20 insertions(+), 3 deletions(-) diff --git a/modules/imgproc/CMakeLists.txt b/modules/imgproc/CMakeLists.txt index 7bf6087a9a..d56de63f80 100644 --- a/modules/imgproc/CMakeLists.txt +++ b/modules/imgproc/CMakeLists.txt @@ -6,7 +6,7 @@ ocv_add_dispatched_file(filter SSE2 SSE4_1 AVX2) ocv_add_dispatched_file(color_hsv SSE2 SSE4_1 AVX2) ocv_add_dispatched_file(color_rgb SSE2 SSE4_1 AVX2) ocv_add_dispatched_file(color_yuv SSE2 SSE4_1 AVX2) -ocv_add_dispatched_file(median_blur SSE2 SSE4_1 AVX2) +ocv_add_dispatched_file(median_blur SSE2 SSE4_1 AVX2 AVX512_SKX) ocv_add_dispatched_file(morph SSE2 SSE4_1 AVX2) ocv_add_dispatched_file(smooth SSE2 SSE4_1 AVX2) ocv_add_dispatched_file(sumpixels SSE2 AVX2 AVX512_SKX) diff --git a/modules/imgproc/src/median_blur.simd.hpp b/modules/imgproc/src/median_blur.simd.hpp index 7cc0aa693c..a3e7a101bb 100644 --- a/modules/imgproc/src/median_blur.simd.hpp +++ b/modules/imgproc/src/median_blur.simd.hpp @@ -13,6 +13,7 @@ // Copyright (C) 2000-2008, 2018, Intel Corporation, all rights reserved. // Copyright (C) 2009, Willow Garage Inc., all rights reserved. // Copyright (C) 2014-2015, Itseez Inc., all rights reserved. +// Copyright (C) 2025, Advanced Micro Devices, all rights reserved. // Third party copyrights are property of their respective owners. // // Redistribution and use in source and binary forms, with or without modification, @@ -693,8 +694,16 @@ medianBlur_SortNet( const Mat& _src, Mat& _dst, int m ) #else int nlanes = 1; #endif - for( ; j <= size.width - nlanes - cn; j += nlanes ) + for (; j < size.width - cn; j += nlanes) { + //handling tail in vectorized path itself + if ( j > size.width - cn - nlanes ) { + if (j == cn || src == dst) { + break; + } + j = size.width - cn - nlanes; + } + VT p0 = vop.load(row0+j-cn), p1 = vop.load(row0+j), p2 = vop.load(row0+j+cn); VT p3 = vop.load(row1+j-cn), p4 = vop.load(row1+j), p5 = vop.load(row1+j+cn); VT p6 = vop.load(row2+j-cn), p7 = vop.load(row2+j), p8 = vop.load(row2+j+cn); @@ -705,6 +714,7 @@ medianBlur_SortNet( const Mat& _src, Mat& _dst, int m ) vop(p3, p6); vop(p1, p4); vop(p2, p5); vop(p4, p7); vop(p4, p2); vop(p6, p4); vop(p4, p2); vop.store(dst+j, p4); + } limit = size.width; @@ -798,8 +808,14 @@ medianBlur_SortNet( const Mat& _src, Mat& _dst, int m ) #else int nlanes = 1; #endif - for( ; j <= size.width - nlanes - cn*2; j += nlanes ) + for( ; j < size.width - cn*2; j += nlanes) { + if ( j > size.width - cn*2 - nlanes ) { + if (j == cn*2 || src == dst) { + break; + } + j = size.width - cn*2 - nlanes; + } VT p0 = vop.load(row[0]+j-cn*2), p5 = vop.load(row[1]+j-cn*2), p10 = vop.load(row[2]+j-cn*2), p15 = vop.load(row[3]+j-cn*2), p20 = vop.load(row[4]+j-cn*2); VT p1 = vop.load(row[0]+j-cn*1), p6 = vop.load(row[1]+j-cn*1), p11 = vop.load(row[2]+j-cn*1), p16 = vop.load(row[3]+j-cn*1), p21 = vop.load(row[4]+j-cn*1); VT p2 = vop.load(row[0]+j-cn*0), p7 = vop.load(row[1]+j-cn*0), p12 = vop.load(row[2]+j-cn*0), p17 = vop.load(row[3]+j-cn*0), p22 = vop.load(row[4]+j-cn*0); @@ -830,6 +846,7 @@ medianBlur_SortNet( const Mat& _src, Mat& _dst, int m ) vop(p13, p17); vop(p3, p15); vop(p11, p23); vop(p11, p15); vop(p7, p19); vop(p7, p11); vop(p11, p13); vop(p11, p12); vop.store(dst+j, p12); + } limit = size.width;