[x265] [PATCH] blockfill_s_32x32 avx2 asm code: performance improved from 1354.05 cycles to 705.81 cycles, over sse version of asm code

[praveen at multicorewareinc.com]

# HG changeset patch
# User Praveen Tiwari
# Date 1412055614 -19800
# Node ID fb46cfd8ee611339da6af16f3c0de426eca3628f
# Parent  975078c41433328fa35913ad46a95cd9c78a8bb2
blockfill_s_32x32 avx2 asm code: performance improved from 1354.05 cycles to 705.81 cycles, over sse version of asm code

diff -r 975078c41433 -r fb46cfd8ee61 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue Sep 30 10:57:12 2014 +0530
+++ b/source/common/x86/asm-primitives.cpp	Tue Sep 30 11:10:14 2014 +0530
@@ -1761,6 +1761,7 @@
         p.copy_cnt[BLOCK_32x32] = x265_copy_cnt_32_avx2;
 
         p.blockfill_s[BLOCK_16x16] = x265_blockfill_s_16x16_avx2;
+        p.blockfill_s[BLOCK_32x32] = x265_blockfill_s_32x32_avx2;
 
         p.cvt32to16_shl[BLOCK_4x4] = x265_cvt32to16_shl_4_avx2;
         p.cvt32to16_shl[BLOCK_8x8] = x265_cvt32to16_shl_8_avx2;
diff -r 975078c41433 -r fb46cfd8ee61 source/common/x86/blockcopy8.asm
--- a/source/common/x86/blockcopy8.asm	Tue Sep 30 10:57:12 2014 +0530
+++ b/source/common/x86/blockcopy8.asm	Tue Sep 30 11:10:14 2014 +0530
@@ -1902,7 +1902,85 @@
 
 BLOCKFILL_S_W32_H4 32, 32
 
-
+INIT_YMM avx2
+cglobal blockfill_s_32x32, 3, 4, 1
+add          r1, r1
+lea          r3, [3 * r1]
+movd         xm0, r2d
+vpbroadcastw m0, xm0
+
+movu       [r0], m0
+movu       [r0 + 32], m0
+movu       [r0 + r1], m0
+movu       [r0 + r1 + 32], m0
+movu       [r0 + 2 * r1], m0
+movu       [r0 + 2 * r1 + 32], m0
+movu       [r0 + r3], m0
+movu       [r0 + r3 + 32], m0
+lea        r0, [r0 + 4 * r1]
+movu       [r0], m0
+movu       [r0 + 32], m0
+movu       [r0 + r1], m0
+movu       [r0 + r1 + 32], m0
+movu       [r0 + 2 * r1], m0
+movu       [r0 + 2 * r1 + 32], m0
+movu       [r0 + r3], m0
+movu       [r0 + r3 + 32], m0
+lea        r0, [r0 + 4 * r1]
+movu       [r0], m0
+movu       [r0 + 32], m0
+movu       [r0 + r1], m0
+movu       [r0 + r1 + 32], m0
+movu       [r0 + 2 * r1], m0
+movu       [r0 + 2 * r1 + 32], m0
+movu       [r0 + r3], m0
+movu       [r0 + r3 + 32], m0
+lea        r0, [r0 + 4 * r1]
+movu       [r0], m0
+movu       [r0 + 32], m0
+movu       [r0 + r1], m0
+movu       [r0 + r1 + 32], m0
+movu       [r0 + 2 * r1], m0
+movu       [r0 + 2 * r1 + 32], m0
+movu       [r0 + r3], m0
+movu       [r0 + r3 + 32], m0
+lea        r0, [r0 + 4 * r1]
+movu       [r0], m0
+movu       [r0 + 32], m0
+movu       [r0 + r1], m0
+movu       [r0 + r1 + 32], m0
+movu       [r0 + 2 * r1], m0
+movu       [r0 + 2 * r1 + 32], m0
+movu       [r0 + r3], m0
+movu       [r0 + r3 + 32], m0
+lea        r0, [r0 + 4 * r1]
+movu       [r0], m0
+movu       [r0 + 32], m0
+movu       [r0 + r1], m0
+movu       [r0 + r1 + 32], m0
+movu       [r0 + 2 * r1], m0
+movu       [r0 + 2 * r1 + 32], m0
+movu       [r0 + r3], m0
+movu       [r0 + r3 + 32], m0
+lea        r0, [r0 + 4 * r1]
+movu       [r0], m0
+movu       [r0 + 32], m0
+movu       [r0 + r1], m0
+movu       [r0 + r1 + 32], m0
+movu       [r0 + 2 * r1], m0
+movu       [r0 + 2 * r1 + 32], m0
+movu       [r0 + r3], m0
+movu       [r0 + r3 + 32], m0
+lea        r0, [r0 + 4 * r1]
+movu       [r0], m0
+movu       [r0 + 32], m0
+movu       [r0 + r1], m0
+movu       [r0 + r1 + 32], m0
+movu       [r0 + 2 * r1], m0
+movu       [r0 + 2 * r1 + 32], m0
+movu       [r0 + r3], m0
+movu       [r0 + r3 + 32], m0
+RET
 
 ;-----------------------------------------------------------------------------
 ; void blockcopy_ps_2x4(int16_t *dest, intptr_t destStride, pixel *src, intptr_t srcStride);
diff -r 975078c41433 -r fb46cfd8ee61 source/common/x86/blockcopy8.h
--- a/source/common/x86/blockcopy8.h	Tue Sep 30 10:57:12 2014 +0530
+++ b/source/common/x86/blockcopy8.h	Tue Sep 30 11:10:14 2014 +0530
@@ -202,6 +202,7 @@
 void x265_blockcopy_pp_32x64_avx(pixel * a, intptr_t stridea, pixel * b, intptr_t strideb);
 
 void x265_blockfill_s_16x16_avx2(int16_t *dst, intptr_t dstride, int16_t val);
+void x265_blockfill_s_32x32_avx2(int16_t *dst, intptr_t dstride, int16_t val);
 
 #undef BLOCKCOPY_COMMON
 #undef BLOCKCOPY_SS_PP