Merge pull request #27299 from amd:fast_medianblur_simd

madanm3 · web-flow · commit 84ea77a4be59 · 2025-05-19T08:56:57.000+03:00
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
diff --git 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
@@ -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;
