coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2026-02-11 04:39:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

compare SumSquareError_SSE2 ported to gcc

Browse Source 
BUG=none
TEST=media_unittest
Review URL: http://webrtc-codereview.appspot.com/279005

git-svn-id: http://libyuv.googlecode.com/svn/trunk@79 16f28f9a-4ce2-e073-06de-1de4eb20be90
This commit is contained in:
fbarchard@google.com 2011-11-15 18:44:31 +00:00
parent 2430e04e0a
commit 4cf70bd6db
2 changed files with 55 additions and 31 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 Name: libyuv
URL: http://code.google.com/p/libyuv/ URL: http://code.google.com/p/libyuv/
Version: 78 Version: 79
License: BSD License: BSD
License File: LICENSE License File: LICENSE

84
source/compare.cc
View File

@ -18,7 +18,8 @@
namespace libyuv { namespace libyuv {
#if defined(__ARM_NEON__) && !defined(COVERAGE_ENABLED) #if defined(__ARM_NEON__) && \
!defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR)
#define HAS_SUMSQUAREERROR_NEON #define HAS_SUMSQUAREERROR_NEON
static uint32 SumSquareError_NEON(const uint8* src_a, static uint32 SumSquareError_NEON(const uint8* src_a,
@ -58,10 +59,8 @@ static uint32 SumSquareError_NEON(const uint8* src_a,
return sse; return sse;
} }
#elif (defined(WIN32) || defined(__x86_64__) || defined(__i386__)) \ #elif defined(WIN32) && \
&& !defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR) !defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR)
#if defined(WIN32) && !defined(COVERAGE_ENABLED)
#define HAS_SUMSQUAREERROR_SSE2 #define HAS_SUMSQUAREERROR_SSE2
__declspec(naked) __declspec(naked)
static uint32 SumSquareError_SSE2(const uint8* src_a, static uint32 SumSquareError_SSE2(const uint8* src_a,
@ -103,41 +102,63 @@ static uint32 SumSquareError_SSE2(const uint8* src_a,
#elif (defined(__x86_64__) || defined(__i386__)) && \ #elif (defined(__x86_64__) || defined(__i386__)) && \
!defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR) !defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR)
// DISABLE #define HAS_SUMSQUAREERROR_SSE2
//#define HAS_SUMSQUAREERROR_SSE2
// DISABLE
#if HAS_SUMSQUAREERROR_SSE2
static uint32 SumSquareError_SSE2(const uint8* src_a, static uint32 SumSquareError_SSE2(const uint8* src_a,
const uint8* src_b, int count) { const uint8* src_b, int count) {
volatile uint32 sse; uint32 sse;
asm volatile ( asm volatile (
" \n" "pxor %%xmm0,%%xmm0 \n"
"pxor %%xmm5,%%xmm5 \n"
"sub %0,%1 \n"
"1: \n"
"movdqa (%0),%%xmm1 \n"
"movdqa (%0,%1,1),%%xmm2 \n"
"lea 0x10(%0),%0 \n"
"movdqa %%xmm1,%%xmm3 \n"
"psubusb %%xmm2,%%xmm1 \n"
"psubusb %%xmm3,%%xmm2 \n"
"por %%xmm2,%%xmm1 \n"
"movdqa %%xmm1,%%xmm2 \n"
"punpcklbw %%xmm5,%%xmm1 \n"
"punpckhbw %%xmm5,%%xmm2 \n"
"pmaddwd %%xmm1,%%xmm1 \n"
"pmaddwd %%xmm2,%%xmm2 \n"
"paddd %%xmm1,%%xmm0 \n"
"paddd %%xmm2,%%xmm0 \n"
"sub $0x10,%2 \n"
"ja 1b \n"
"pshufd $0xee,%%xmm0,%%xmm1 \n"
"paddd %%xmm1,%%xmm0 \n"
"pshufd $0x1,%%xmm0,%%xmm1 \n"
"paddd %%xmm1,%%xmm0 \n"
"movd %%xmm0,%3 \n"
: "+r"(src_a), // %0 : "+r"(src_a), // %0
"+r"(src_b), // %1 "+r"(src_b), // %1
"+r"(count), // %2 "+r"(count), // %2
"=r"(sse) // %3 "=g"(sse) // %3
: :
: "memory", "cc" : "memory", "cc"
#if defined(__SSE2__) #if defined(__SSE2__)
, "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" , "xmm0", "xmm1", "xmm2", "xmm5"
#endif #endif
); );
return sse; return sse;
} }
#endif #endif
#endif
#endif
static uint32 SumSquareError_C(const uint8* src_a, static uint32 SumSquareError_C(const uint8* src_a,
const uint8* src_b, int count) { const uint8* src_b, int count) {
uint32 udiff = 0u; uint32 sse = 0u;
for (int x = 0; x < count; ++x) { for (int x = 0; x < count; ++x) {
int diff = src_a[0] - src_b[0]; int diff = src_a[0] - src_b[0];
udiff += static_cast<uint32>(diff * diff); sse += static_cast<uint32>(diff * diff);
src_a += 1; src_a += 1;
src_b += 1; src_b += 1;
} }
return udiff; return sse;
} }
uint64 ComputeSumSquareError(const uint8* src_a, uint64 ComputeSumSquareError(const uint8* src_a,
@ -157,22 +178,25 @@ uint64 ComputeSumSquareError(const uint8* src_a,
{ {
SumSquareError = SumSquareError_C; SumSquareError = SumSquareError_C;
} }
const int kBlockSize = 4096; const int kBlockSize = 32768;
uint64 diff = 0; uint64 sse = 0;
while (count >= kBlockSize) { while (count >= kBlockSize) {
diff += SumSquareError(src_a, src_b, kBlockSize); sse += SumSquareError(src_a, src_b, kBlockSize);
src_a += kBlockSize; src_a += kBlockSize;
src_b += kBlockSize; src_b += kBlockSize;
count -= kBlockSize; count -= kBlockSize;
} }
if (count > 0) { int remainder = count & ~15;
if (count % 16 == 0) { if (remainder) {
diff += static_cast<uint64>(SumSquareError(src_a, src_b, count)); sse += SumSquareError(src_a, src_b, remainder);
} else { src_a += remainder;
diff += static_cast<uint64>(SumSquareError_C(src_a, src_b, count)); src_b += remainder;
} count -= remainder;
} }
return diff; if (count) {
sse += SumSquareError_C(src_a, src_b, count);
}
return sse;
} }
uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a, uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
@ -192,7 +216,7 @@ uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
uint64 sse = 0; uint64 sse = 0;
for (int h = 0; h < height; ++h) { for (int h = 0; h < height; ++h) {
sse += static_cast<uint64>(SumSquareError(src_a, src_b, width)); sse += SumSquareError(src_a, src_b, width);
src_a += stride_a; src_a += stride_a;
src_b += stride_b; src_b += stride_b;
} }