diff options
| author |  mtklein <mtklein@chromium.org> | 2016-01-11 06:16:26 -0800 |
|---|---|---|
| committer |  Commit bot <commit-bot@chromium.org> | 2016-01-11 06:16:26 -0800 |
| commit | 3d00db39c6637607a2a1f184be2f31e4fe17d9ad (patch) | |
| tree | 5e74d6c753b64b31588c34119185a28f09849949 /src/codec/SkSwizzler.cpp | |
| parent | 4645e4687d39b87a430e925c2d209dca70250652 (diff) | |
Might as well inline these premultiplies.
We're paying quite a bit of function-call overhead per pixel.
On one test image we spend 3.5% of our total time in swizzle_rgba_to_n32_premul() and 8.8% of our total time in SkPreMultiplyARGB(). That turns into just 8.8% of our total time in swizzle_rgba_to_n32_premul() after inlining.
That's about a 30% speedup.
This will make SIMD procs look worse, so it's nice to land this first.
BUG=skia:4767
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1571923002
Review URL: https://codereview.chromium.org/1571923002
Diffstat (limited to 'src/codec/SkSwizzler.cpp')
| -rw-r--r-- | src/codec/SkSwizzler.cpp | 8 |
1 files changed, 4 insertions, 4 deletions
diff --git a/src/codec/SkSwizzler.cpp b/src/codec/SkSwizzler.cpp index d783380294..aa58120451 100644 --- a/src/codec/SkSwizzler.cpp +++ b/src/codec/SkSwizzler.cpp @@ -395,7 +395,7 @@ static SkSwizzler::ResultAlpha swizzle_bgra_to_n32_premul( for (int x = 0; x < dstWidth; x++) { uint8_t alpha = src[3]; UPDATE_RESULT_ALPHA(alpha); - dst[x] = SkPreMultiplyARGB(alpha, src[2], src[1], src[0]); + dst[x] = SkPremultiplyARGBInline(alpha, src[2], src[1], src[0]); src += deltaSrc; } return COMPUTE_RESULT_ALPHA; @@ -440,7 +440,7 @@ static SkSwizzler::ResultAlpha swizzle_rgba_to_n32_premul( for (int x = 0; x < dstWidth; x++) { unsigned alpha = src[3]; UPDATE_RESULT_ALPHA(alpha); - dst[x] = SkPreMultiplyARGB(alpha, src[0], src[1], src[2]); + dst[x] = SkPremultiplyARGBInline(alpha, src[0], src[1], src[2]); src += deltaSrc; } return COMPUTE_RESULT_ALPHA; @@ -473,7 +473,7 @@ static SkSwizzler::ResultAlpha swizzle_rgba_to_n32_premul_skipZ( unsigned alpha = src[3]; UPDATE_RESULT_ALPHA(alpha); if (0 != alpha) { - dst[x] = SkPreMultiplyARGB(alpha, src[0], src[1], src[2]); + dst[x] = SkPremultiplyARGBInline(alpha, src[0], src[1], src[2]); } src += deltaSrc; } @@ -779,7 +779,7 @@ SkSwizzler* SkSwizzler::CreateSwizzler(SkSwizzler::SrcConfig sc, // Store bpp in bytes if it is an even multiple, otherwise use bits int srcBPP = SkIsAlign8(BitsPerPixel(sc)) ? BytesPerPixel(sc) : BitsPerPixel(sc); int dstBPP = SkColorTypeBytesPerPixel(dstInfo.colorType()); - + int srcOffset = 0; int srcWidth = dstInfo.width(); int dstOffset = 0; |