coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2025-12-07 17:26:49 +08:00
Code Issues Packages Projects Releases Wiki Activity

Interpolate Row ported to AVX2 GCC/NaCL.

Browse Source 
BUG=269
TESTED=nacl build
R=brucedawson@google.com, tpsiaki@google.com

Review URL: https://webrtc-codereview.appspot.com/25329004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1193 16f28f9a-4ce2-e073-06de-1de4eb20be90
This commit is contained in:
fbarchard@google.com 2014-12-09 22:21:53 +00:00
parent d5b5594f8a
commit e0bb4c26e2
4 changed files with 109 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
README.chromium
View File

@ -1,6 +1,6 @@
Name: libyuv
URL: http://code.google.com/p/libyuv/
Version: 1191
Version: 1193
License: BSD
License File: LICENSE

2
include/libyuv/row.h
View File

@ -196,6 +196,7 @@ extern "C" {
#define HAS_I422TOARGBROW_AVX2
#define HAS_I422TOBGRAROW_AVX2
#define HAS_I422TORGBAROW_AVX2
#define HAS_INTERPOLATEROW_AVX2
#define HAS_MERGEUVROW_AVX2
#define HAS_MIRRORROW_AVX2
#define HAS_SPLITUVROW_AVX2
@ -218,7 +219,6 @@ extern "C" {
// TODO(fbarchard): Port to gcc.
#if !defined(LIBYUV_DISABLE_X86) && defined(VISUALC_HAS_AVX2)
#define HAS_ARGBTOUVROW_AVX2
#define HAS_INTERPOLATEROW_AVX2
#endif // defined(VISUALC_HAS_AVX2)
// The following are Yasm x86 only:

2
include/libyuv/version.h
View File

@ -11,6 +11,6 @@
#ifndef INCLUDE_LIBYUV_VERSION_H_ // NOLINT
#define INCLUDE_LIBYUV_VERSION_H_
#define LIBYUV_VERSION 1191
#define LIBYUV_VERSION 1193
#endif // INCLUDE_LIBYUV_VERSION_H_ NOLINT

106
source/row_posix.cc
View File

@ -5135,6 +5135,112 @@ void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
}
#endif // HAS_INTERPOLATEROW_SSSE3
#ifdef HAS_INTERPOLATEROW_AVX2
// Bilinear filter 32x2 -> 32x1
void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
ptrdiff_t src_stride, int dst_width,
int source_y_fraction) {
asm volatile (
"shr %3 \n"
"cmp $0x0,%3 \n"
"je 100f \n"
"sub %1,%0 \n"
"cmp $0x20,%3 \n"
"je 75f \n"
"cmp $0x40,%3 \n"
"je 50f \n"
"cmp $0x60,%3 \n"
"je 25f \n"
"vmovd %3,%%xmm0 \n"
"neg %3 \n"
"add $0x80,%3 \n"
"vmovd %3,%%xmm5 \n"
"vpunpcklbw %%xmm0,%%xmm5,%%xmm5 \n"
"vpunpcklwd %%xmm5,%%xmm5,%%xmm5 \n"
"vpxor %%ymm0,%%ymm0,%%ymm0 \n"
"vpermd %%ymm5,%%ymm0,%%ymm5 \n"
// General purpose row blend.
LABELALIGN
"1: \n"
"vmovdqu " MEMACCESS(1) ",%%ymm0 \n"
MEMOPREG(vmovdqu,0x00,1,4,1,ymm2)
"vpunpckhbw %%ymm2,%%ymm0,%%ymm1 \n"
"vpunpcklbw %%ymm2,%%ymm0,%%ymm0 \n"
"vpmaddubsw %%ymm5,%%ymm0,%%ymm0 \n"
"vpmaddubsw %%ymm5,%%ymm1,%%ymm1 \n"
"vpsrlw $0x7,%%ymm0,%%ymm0 \n"
"vpsrlw $0x7,%%ymm1,%%ymm1 \n"
"vpackuswb %%ymm1,%%ymm0,%%ymm0 \n"
MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
"lea " MEMLEA(0x20,1) ",%1 \n"
"sub $0x20,%2 \n"
"jg 1b \n"
"jmp 99f \n"
// Blend 25 / 75.
LABELALIGN
"25: \n"
"vmovdqu " MEMACCESS(1) ",%%ymm0 \n"
MEMOPREG(vmovdqu,0x00,1,4,1,ymm1)
"vpavgb %%ymm1,%%ymm0,%%ymm0 \n"
"vpavgb %%ymm1,%%ymm0,%%ymm0 \n"
MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
"lea " MEMLEA(0x20,1) ",%1 \n"
"sub $0x20,%2 \n"
"jg 25b \n"
"jmp 99f \n"
// Blend 50 / 50.
LABELALIGN
"50: \n"
"vmovdqu " MEMACCESS(1) ",%%ymm0 \n"
VMEMOPREG(vpavgb,0x00,1,4,1,ymm0,ymm0) // vpavgb (%1,%4,1),%%ymm0,%%ymm0
MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
"lea " MEMLEA(0x20,1) ",%1 \n"
"sub $0x20,%2 \n"
"jg 50b \n"
"jmp 99f \n"
// Blend 75 / 25.
LABELALIGN
"75: \n"
"vmovdqu " MEMACCESS(1) ",%%ymm1 \n"
MEMOPREG(vmovdqu,0x00,1,4,1,ymm0)
"vpavgb %%ymm1,%%ymm0,%%ymm0 \n"
"vpavgb %%ymm1,%%ymm0,%%ymm0 \n"
MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
"lea " MEMLEA(0x20,1) ",%1 \n"
"sub $0x20,%2 \n"
"jg 75b \n"
"jmp 99f \n"
// Blend 100 / 0 - Copy row unchanged.
LABELALIGN
"100: \n"
"rep movsb " MEMMOVESTRING(1,0) " \n"
"jmp 999f \n"
"99: \n"
"vzeroupper \n"
"999: \n"
: "+D"(dst_ptr), // %0
"+S"(src_ptr), // %1
"+c"(dst_width), // %2
"+r"(source_y_fraction) // %3
: "r"((intptr_t)(src_stride)) // %4
: "memory", "cc"
#if defined(__native_client__) && defined(__x86_64__)
, "r14"
#endif
#if defined(__SSE2__)
, "xmm0", "xmm1", "xmm2", "xmm5"
#endif
);
}
#endif // HAS_INTERPOLATEROW_AVX2
#ifdef HAS_INTERPOLATEROW_SSE2
// Bilinear filter 16x2 -> 16x1
void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,