AuroraMiddleware/skia2
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Xfermode: SSE2 implementation of colordodge&colorburn modes

Browse Source 
With SSE2 optimization, performance of the related benchmarks will improve
about 45% for Xfermode_ColorDodge and little for Xfermode_ColorBurn on
desktop i7-3770. The little performance improvement for
Xfermode_ColorBurn is due to the portable version may mostly go the fast
if branch while the SSE2 version do the calculation for all the three
if-else branches. Here are the data:
before:
Xfermode_ColorDodge   8888:  cmsecs =  73.71   565:  cmsecs =  82.88
 Xfermode_ColorBurn   8888:  cmsecs =  46.46   565:  cmsecs =  52.23
after:
Xfermode_ColorDodge   8888:  cmsecs =  39.70   565:  cmsecs =  47.45
 Xfermode_ColorBurn   8888:  cmsecs =  45.02   565:  cmsecs =  51.15

BUG=skia:
R=mtklein@google.com

Author: qiankun.miao@intel.com

Review URL: https://codereview.chromium.org/224823004

git-svn-id: http://skia.googlecode.com/svn/trunk@14377 2bbb7eff-a529-9590-31e7-b0007b416f81
This commit is contained in:
commit-bot@chromium.org 2014-04-25 09:44:53 +00:00
parent 6419a5edb8
commit e1ba93ee01
2 changed files with 118 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
src/opts/SkMath_opts_SSE2.h
View File

@ -10,6 +10,14 @@
#include <emmintrin.h>
// Because no _mm_div_epi32() in SSE2, we use float division to emulate.
// When using this function, make sure a and b don't exceed float's precision.
static inline __m128i shim_mm_div_epi32(const __m128i& a, const __m128i& b) {
__m128 x = _mm_cvtepi32_ps(a);
__m128 y = _mm_cvtepi32_ps(b);
return _mm_cvttps_epi32(_mm_div_ps(x, y));
}
// Portable version of SkSqrtBits is in SkMath.cpp.
static inline __m128i SkSqrtBits_SSE2(const __m128i& x, int count) {
__m128i root = _mm_setzero_si128();

112
src/opts/SkXfermode_opts_SSE2.cpp
View File

@ -283,6 +283,114 @@ static __m128i overlay_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
return SkPackARGB32_SSE2(a, r, g, b);
}
static inline __m128i colordodge_byte_SSE2(const __m128i& sc, const __m128i& dc,
const __m128i& sa, const __m128i& da) {
__m128i diff = _mm_sub_epi32(sa, sc);
__m128i ida = _mm_sub_epi32(_mm_set1_epi32(255), da);
__m128i isa = _mm_sub_epi32(_mm_set1_epi32(255), sa);
// if (0 == dc)
__m128i cmp1 = _mm_cmpeq_epi32(dc, _mm_setzero_si128());
__m128i rc1 = _mm_and_si128(cmp1, SkAlphaMulAlpha_SSE2(sc, ida));
// else if (0 == diff)
__m128i cmp2 = _mm_cmpeq_epi32(diff, _mm_setzero_si128());
__m128i cmp = _mm_andnot_si128(cmp1, cmp2);
__m128i tmp1 = _mm_mullo_epi16(sa, da);
__m128i tmp2 = _mm_mullo_epi16(sc, ida);
__m128i tmp3 = _mm_mullo_epi16(dc, isa);
__m128i rc2 = _mm_add_epi32(tmp1, tmp2);
rc2 = _mm_add_epi32(rc2, tmp3);
rc2 = clamp_div255round_SSE2(rc2);
rc2 = _mm_and_si128(cmp, rc2);
// else
__m128i cmp3 = _mm_or_si128(cmp1, cmp2);
__m128i value = _mm_mullo_epi16(dc, sa);
diff = shim_mm_div_epi32(value, diff);
__m128i tmp4 = SkMin32_SSE2(da, diff);
tmp4 = Multiply32_SSE2(sa, tmp4);
__m128i rc3 = _mm_add_epi32(tmp4, tmp2);
rc3 = _mm_add_epi32(rc3, tmp3);
rc3 = clamp_div255round_SSE2(rc3);
rc3 = _mm_andnot_si128(cmp3, rc3);
__m128i rc = _mm_or_si128(rc1, rc2);
rc = _mm_or_si128(rc, rc3);
return rc;
}
static __m128i colordodge_modeproc_SSE2(const __m128i& src,
const __m128i& dst) {
__m128i sa = SkGetPackedA32_SSE2(src);
__m128i da = SkGetPackedA32_SSE2(dst);
__m128i a = srcover_byte_SSE2(sa, da);
__m128i r = colordodge_byte_SSE2(SkGetPackedR32_SSE2(src),
SkGetPackedR32_SSE2(dst), sa, da);
__m128i g = colordodge_byte_SSE2(SkGetPackedG32_SSE2(src),
SkGetPackedG32_SSE2(dst), sa, da);
__m128i b = colordodge_byte_SSE2(SkGetPackedB32_SSE2(src),
SkGetPackedB32_SSE2(dst), sa, da);
return SkPackARGB32_SSE2(a, r, g, b);
}
static inline __m128i colorburn_byte_SSE2(const __m128i& sc, const __m128i& dc,
const __m128i& sa, const __m128i& da) {
__m128i ida = _mm_sub_epi32(_mm_set1_epi32(255), da);
__m128i isa = _mm_sub_epi32(_mm_set1_epi32(255), sa);
// if (dc == da)
__m128i cmp1 = _mm_cmpeq_epi32(dc, da);
__m128i tmp1 = _mm_mullo_epi16(sa, da);
__m128i tmp2 = _mm_mullo_epi16(sc, ida);
__m128i tmp3 = _mm_mullo_epi16(dc, isa);
__m128i rc1 = _mm_add_epi32(tmp1, tmp2);
rc1 = _mm_add_epi32(rc1, tmp3);
rc1 = clamp_div255round_SSE2(rc1);
rc1 = _mm_and_si128(cmp1, rc1);
// else if (0 == sc)
__m128i cmp2 = _mm_cmpeq_epi32(sc, _mm_setzero_si128());
__m128i rc2 = SkAlphaMulAlpha_SSE2(dc, isa);
__m128i cmp = _mm_andnot_si128(cmp1, cmp2);
rc2 = _mm_and_si128(cmp, rc2);
// else
__m128i cmp3 = _mm_or_si128(cmp1, cmp2);
__m128i tmp4 = _mm_sub_epi32(da, dc);
tmp4 = Multiply32_SSE2(tmp4, sa);
tmp4 = shim_mm_div_epi32(tmp4, sc);
__m128i tmp5 = _mm_sub_epi32(da, SkMin32_SSE2(da, tmp4));
tmp5 = Multiply32_SSE2(sa, tmp5);
__m128i rc3 = _mm_add_epi32(tmp5, tmp2);
rc3 = _mm_add_epi32(rc3, tmp3);
rc3 = clamp_div255round_SSE2(rc3);
rc3 = _mm_andnot_si128(cmp3, rc3);
__m128i rc = _mm_or_si128(rc1, rc2);
rc = _mm_or_si128(rc, rc3);
return rc;
}
static __m128i colorburn_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
__m128i sa = SkGetPackedA32_SSE2(src);
__m128i da = SkGetPackedA32_SSE2(dst);
__m128i a = srcover_byte_SSE2(sa, da);
__m128i r = colorburn_byte_SSE2(SkGetPackedR32_SSE2(src),
SkGetPackedR32_SSE2(dst), sa, da);
__m128i g = colorburn_byte_SSE2(SkGetPackedG32_SSE2(src),
SkGetPackedG32_SSE2(dst), sa, da);
__m128i b = colorburn_byte_SSE2(SkGetPackedB32_SSE2(src),
SkGetPackedB32_SSE2(dst), sa, da);
return SkPackARGB32_SSE2(a, r, g, b);
}
static inline __m128i hardlight_byte_SSE2(const __m128i& sc, const __m128i& dc,
const __m128i& sa, const __m128i& da) {
// if (2 * sc <= sa)
@ -620,8 +728,8 @@ SkXfermodeProcSIMD gSSE2XfermodeProcs[] = {
overlay_modeproc_SSE2,
NULL, // kDarken_Mode
NULL, // kLighten_Mode
NULL, // kColorDodge_Mode
NULL, // kColorBurn_Mode
colordodge_modeproc_SSE2,
colorburn_modeproc_SSE2,
hardlight_modeproc_SSE2,
softlight_modeproc_SSE2,
difference_modeproc_SSE2,