From c70c7c02ff76ff7180263ba1d2a0e90a869f37ee Mon Sep 17 00:00:00 2001 From: "fbarchard@google.com" Date: Tue, 31 Mar 2015 23:59:27 +0000 Subject: [PATCH] scale to half size optimization for avx2 - use pmaddubsw instruction to horizontally add bytes, then pavgw to round and divide by 2. BUG=314 TESTED=libyuvTest.ScaleDownBy2* R=tpsiaki@google.com Review URL: https://webrtc-codereview.appspot.com/45909004 git-svn-id: http://libyuv.googlecode.com/svn/trunk@1352 16f28f9a-4ce2-e073-06de-1de4eb20be90 --- README.chromium | 2 +- include/libyuv/version.h | 2 +- source/scale_win.cc | 42 +++++++++++++++++++++------------------- 3 files changed, 24 insertions(+), 22 deletions(-) diff --git a/README.chromium b/README.chromium index facea334a..da3180f96 100644 --- a/README.chromium +++ b/README.chromium @@ -1,6 +1,6 @@ Name: libyuv URL: http://code.google.com/p/libyuv/ -Version: 1349 +Version: 1352 License: BSD License File: LICENSE diff --git a/include/libyuv/version.h b/include/libyuv/version.h index 39d66a2e9..a39afd306 100644 --- a/include/libyuv/version.h +++ b/include/libyuv/version.h @@ -11,6 +11,6 @@ #ifndef INCLUDE_LIBYUV_VERSION_H_ // NOLINT #define INCLUDE_LIBYUV_VERSION_H_ -#define LIBYUV_VERSION 1349 +#define LIBYUV_VERSION 1352 #endif // INCLUDE_LIBYUV_VERSION_H_ NOLINT diff --git a/source/scale_win.cc b/source/scale_win.cc index c8a474bcb..a9a68b668 100644 --- a/source/scale_win.cc +++ b/source/scale_win.cc @@ -236,22 +236,23 @@ void ScaleRowDown2Linear_AVX2(const uint8* src_ptr, ptrdiff_t src_stride, // src_stride mov edx, [esp + 12] // dst_ptr mov ecx, [esp + 16] // dst_width - vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0x00ff00ff - vpsrlw ymm5, ymm5, 8 + + vpcmpeqb ymm4, ymm4, ymm4 // '1' constant, 8b + vpsrlw ymm4, ymm4, 15 + vpackuswb ymm4, ymm4, ymm4 + vpxor ymm5, ymm5, ymm5 // constant 0 wloop: vmovdqu ymm0, [eax] vmovdqu ymm1, [eax + 32] lea eax, [eax + 64] - vpsrlw ymm2, ymm0, 8 // average columns (32 to 16 pixels) - vpsrlw ymm3, ymm1, 8 - vpand ymm0, ymm0, ymm5 - vpand ymm1, ymm1, ymm5 - vpavgw ymm0, ymm0, ymm2 - vpavgw ymm1, ymm1, ymm3 + vpmaddubsw ymm0, ymm0, ymm4 // average horizontally + vpmaddubsw ymm1, ymm1, ymm4 + vpavgw ymm0, ymm0, ymm5 // (x + 1) / 2 + vpavgw ymm1, ymm1, ymm5 vpackuswb ymm0, ymm0, ymm1 - vpermq ymm0, ymm0, 0xd8 // unmutate + vpermq ymm0, ymm0, 0xd8 // unmutate vpackuswb vmovdqu [edx], ymm0 lea edx, [edx + 32] @@ -273,24 +274,25 @@ void ScaleRowDown2Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride, mov esi, [esp + 4 + 8] // src_stride mov edx, [esp + 4 + 12] // dst_ptr mov ecx, [esp + 4 + 16] // dst_width - vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0x00ff00ff - vpsrlw ymm5, ymm5, 8 + + vpcmpeqb ymm4, ymm4, ymm4 // '1' constant, 8b + vpsrlw ymm4, ymm4, 15 + vpackuswb ymm4, ymm4, ymm4 + vpxor ymm5, ymm5, ymm5 // constant 0 wloop: - vmovdqu ymm0, [eax] + vmovdqu ymm0, [eax] // average rows vmovdqu ymm1, [eax + 32] - vpavgb ymm0, ymm0, [eax + esi] // average rows + vpavgb ymm0, ymm0, [eax + esi] vpavgb ymm1, ymm1, [eax + esi + 32] lea eax, [eax + 64] - vpsrlw ymm2, ymm0, 8 // average columns (32 to 16 pixels) - vpsrlw ymm3, ymm1, 8 - vpand ymm0, ymm0, ymm5 - vpand ymm1, ymm1, ymm5 - vpavgw ymm0, ymm0, ymm2 - vpavgw ymm1, ymm1, ymm3 + vpmaddubsw ymm0, ymm0, ymm4 // average horizontally + vpmaddubsw ymm1, ymm1, ymm4 + vpavgw ymm0, ymm0, ymm5 // (x + 1) / 2 + vpavgw ymm1, ymm1, ymm5 vpackuswb ymm0, ymm0, ymm1 - vpermq ymm0, ymm0, 0xd8 // unmutate + vpermq ymm0, ymm0, 0xd8 // unmutate vpackuswb vmovdqu [edx], ymm0 lea edx, [edx + 32]