[x265] [PATCH] blockcopy_ss: 64x16, 64x32, 64x48, 64x64 AVX version of asm code, approx double speedup comapre to SSE

[sagar at multicorewareinc.com]

# HG changeset patch
# User Sagar Kotecha
# Date 1411477373 -19800
#      Tue Sep 23 18:32:53 2014 +0530
# Node ID 063c3005a181b22fb2b947ab3779479a6d2fae14
# Parent  166ed7f88b9205eb585f88c7ab43ab02b29ace69
blockcopy_ss: 64x16, 64x32, 64x48, 64x64 AVX version of asm code, approx double speedup comapre to SSE

diff -r 166ed7f88b92 -r 063c3005a181 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Tue Sep 23 12:47:02 2014 +0530
+++ b/source/common/x86/asm-primitives.cpp	Tue Sep 23 18:32:53 2014 +0530
@@ -1709,6 +1709,10 @@
 
         p.ssim_4x4x2_core = x265_pixel_ssim_4x4x2_core_avx;
         p.ssim_end_4 = x265_pixel_ssim_end4_avx;
+        p.luma_copy_ss[LUMA_64x16] = x265_blockcopy_ss_64x16_avx;
+        p.luma_copy_ss[LUMA_64x32] = x265_blockcopy_ss_64x32_avx;
+        p.luma_copy_ss[LUMA_64x48] = x265_blockcopy_ss_64x48_avx;
+        p.luma_copy_ss[LUMA_64x64] = x265_blockcopy_ss_64x64_avx;
     }
     if (cpuMask & X265_CPU_XOP)
     {
diff -r 166ed7f88b92 -r 063c3005a181 source/common/x86/blockcopy8.asm
--- a/source/common/x86/blockcopy8.asm	Tue Sep 23 12:47:02 2014 +0530
+++ b/source/common/x86/blockcopy8.asm	Tue Sep 23 18:32:53 2014 +0530
@@ -3317,6 +3317,69 @@
 BLOCKCOPY_SS_W64_H4 64, 48
 BLOCKCOPY_SS_W64_H4 64, 64
 
+;-----------------------------------------------------------------------------
+; void blockcopy_ss_%1x%2(int16_t *dest, intptr_t deststride, int16_t *src, intptr_t srcstride)
+;-----------------------------------------------------------------------------
+%macro BLOCKCOPY_SS_W64_H4_avx 2
+INIT_YMM avx
+cglobal blockcopy_ss_%1x%2, 4, 7, 4, dest, deststride, src, srcstride
+    mov     r4d, %2/4
+    add     r1, r1
+    add     r3, r3
+    lea     r5, [3 * r1]
+    lea     r6, [3 * r3]
+.loop:
+    movu    m0, [r2]
+    movu    m1, [r2 + 32]
+    movu    m2, [r2 + 64]
+    movu    m3, [r2 + 96]
+
+    movu    [r0], m0
+    movu    [r0 + 32], m1
+    movu    [r0 + 64], m2
+    movu    [r0 + 96], m3
+
+    movu    m0, [r2 + r3]
+    movu    m1, [r2 + r3 + 32]
+    movu    m2, [r2 + r3 + 64]
+    movu    m3, [r2 + r3 + 96]
+
+    movu    [r0 + r1], m0
+    movu    [r0 + r1 + 32], m1
+    movu    [r0 + r1 + 64], m2
+    movu    [r0 + r1 + 96], m3
+
+    movu    m0, [r2 + 2 * r3]
+    movu    m1, [r2 + 2 * r3 + 32]
+    movu    m2, [r2 + 2 * r3 + 64]
+    movu    m3, [r2 + 2 * r3 + 96]
+
+    movu    [r0 + 2 * r1], m0
+    movu    [r0 + 2 * r1 + 32], m1
+    movu    [r0 + 2 * r1 + 64], m2
+    movu    [r0 + 2 * r1 + 96], m3
+
+    movu    m0, [r2 + r6]
+    movu    m1, [r2 + r6 + 32]
+    movu    m2, [r2 + r6 + 64]
+    movu    m3, [r2 + r6 + 96]
+    lea     r2, [r2 + 4 * r3]
+
+    movu    [r0 + r5], m0
+    movu    [r0 + r5 + 32], m1
+    movu    [r0 + r5 + 64], m2
+    movu    [r0 + r5 + 96], m3
+    lea     r0, [r0 + 4 * r1]
+
+    dec     r4d
+    jnz     .loop
+    RET
+%endmacro
+
+BLOCKCOPY_SS_W64_H4_avx 64, 16
+BLOCKCOPY_SS_W64_H4_avx 64, 32
+BLOCKCOPY_SS_W64_H4_avx 64, 48
+BLOCKCOPY_SS_W64_H4_avx 64, 64
 
 ;-----------------------------------------------------------------------------
 ; void cvt32to16_shr(short *dst, int *src, intptr_t stride, int shift, int size)
diff -r 166ed7f88b92 -r 063c3005a181 source/common/x86/blockcopy8.h
--- a/source/common/x86/blockcopy8.h	Tue Sep 23 12:47:02 2014 +0530
+++ b/source/common/x86/blockcopy8.h	Tue Sep 23 18:32:53 2014 +0530
@@ -189,6 +189,10 @@
 void x265_blockcopy_ss_16x24_avx(int16_t *dest, intptr_t deststride, int16_t *src, intptr_t srcstride);
 void x265_blockcopy_ss_16x32_avx(int16_t *dest, intptr_t deststride, int16_t *src, intptr_t srcstride);
 void x265_blockcopy_ss_16x64_avx(int16_t *dest, intptr_t deststride, int16_t *src, intptr_t srcstride);
+void x265_blockcopy_ss_64x16_avx(int16_t *dest, intptr_t deststride, int16_t *src, intptr_t srcstride);
+void x265_blockcopy_ss_64x32_avx(int16_t *dest, intptr_t deststride, int16_t *src, intptr_t srcstride);
+void x265_blockcopy_ss_64x48_avx(int16_t *dest, intptr_t deststride, int16_t *src, intptr_t srcstride);
+void x265_blockcopy_ss_64x64_avx(int16_t *dest, intptr_t deststride, int16_t *src, intptr_t srcstride);
 
 #undef BLOCKCOPY_COMMON
 #undef BLOCKCOPY_SS_PP