[x265] [PATCH] asm: 10bpp blockcopy_ps integration for Luma and chroma partitions

murugan at multicorewareinc.com murugan at multicorewareinc.com
Tue Dec 10 11:15:37 CET 2013


# HG changeset patch
# User Murugan Vairavel <murugan at multicorewareinc.com>
# Date 1386670520 -19800
#      Tue Dec 10 15:45:20 2013 +0530
# Node ID a17508209a7041fef42fa8d8413af73084e6c452
# Parent  349641856c90412e395b8d6c607235ae8587a4dc
asm: 10bpp blockcopy_ps integration for Luma and chroma partitions

diff -r 349641856c90 -r a17508209a70 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue Dec 10 15:39:10 2013 +0530
+++ b/source/common/x86/asm-primitives.cpp	Tue Dec 10 15:45:20 2013 +0530
@@ -664,37 +664,8 @@
         CHROMA_PIXELSUB_PS(_sse2);
         LUMA_PIXELSUB(_sse2);
 
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_4x8] = x265_blockcopy_pp_2x4_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_4x16] = x265_blockcopy_pp_2x8_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_8x4] = x265_blockcopy_pp_4x2_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_8x8] = x265_blockcopy_pp_4x4_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_8x16] = x265_blockcopy_pp_4x8_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_8x32] = x265_blockcopy_pp_4x16_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_12x16] = x265_blockcopy_pp_6x8_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_16x4] = x265_blockcopy_pp_8x2_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_16x8] = x265_blockcopy_pp_8x4_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_16x12] = x265_blockcopy_pp_8x6_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_16x16] = x265_blockcopy_pp_8x8_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_16x32] = x265_blockcopy_pp_8x16_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_16x64] = x265_blockcopy_pp_8x32_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_24x32] = x265_blockcopy_pp_12x16_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_32x8] = x265_blockcopy_pp_16x4_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_32x16] = x265_blockcopy_pp_16x8_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_32x24] = x265_blockcopy_pp_16x12_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_32x32] = x265_blockcopy_pp_16x16_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_32x64] = x265_blockcopy_pp_16x32_sse2;
-        p.luma_copy_pp[LUMA_16x64] = x265_blockcopy_pp_16x64_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_48x64] = x265_blockcopy_pp_24x32_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_64x16] = x265_blockcopy_pp_32x8_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_64x32] = x265_blockcopy_pp_32x16_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_64x48] = x265_blockcopy_pp_32x24_sse2;
-        p.chroma[X265_CSP_I420].copy_pp[LUMA_64x64] = x265_blockcopy_pp_32x32_sse2;
-        p.luma_copy_pp[LUMA_32x64] = x265_blockcopy_pp_32x64_sse2;
-        p.luma_copy_pp[LUMA_48x64] = x265_blockcopy_pp_48x64_sse2;
-        p.luma_copy_pp[LUMA_64x16] = x265_blockcopy_pp_64x16_sse2;
-        p.luma_copy_pp[LUMA_64x32] = x265_blockcopy_pp_64x32_sse2;
-        p.luma_copy_pp[LUMA_64x48] = x265_blockcopy_pp_64x48_sse2;
-        p.luma_copy_pp[LUMA_64x64] = x265_blockcopy_pp_64x64_sse2;
+        CHROMA_BLOCKCOPY(_sse2);
+        LUMA_BLOCKCOPY(_sse2);
     }
     if (cpuMask & X265_CPU_SSSE3)
     {


More information about the x265-devel mailing list