diff options
author | Mike Klein <mtklein@chromium.org> | 2016-10-06 15:06:38 -0400 |
---|---|---|
committer | Skia Commit-Bot <skia-commit-bot@chromium.org> | 2016-10-07 12:52:29 +0000 |
commit | 1aebdaee0e2aa4324509fd3ad4c40c21703ae4a2 (patch) | |
tree | c5ffae6c59217f3d228891177e1d50d7f784801a /src/core/SkHalf.h | |
parent | 2766cc567d5c939730fadd2d865e4bdf05477263 (diff) |
SkRasterPipeline: 8x pipelines
Bench runtime changes:
sRGB: 7194 -> 3735 = 1.93x faster
F16: 6531 -> 2559 = 2.55x faster
Instead of building 4x and 1-3x pipelines and then maybe 8x and 1-7x, instead build either the short ones or the long ones, but not both. If we just take care to use a compatible run_pipeline(), there's some cross-module type disagreement but everything works out in the end.
Oddly, a few places that looked like they'd be faster using SkNx_fma() or Sk4f_round()/Sk8f_round() are actually faster the long way, e.g. multiply, add 0.5, truncate. Curious! In all the other places you see here that I've used SkNx_fma(), it's been a significant speedup.
This folds in a couple refactors and cleanups that I've been meaning to do. Hope you don't mind... if find the new code considerably easier to read than the old code.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2990
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
Change-Id: I1c82e5755d8e44cc0b9c6673d04b117f85d71a3a
Reviewed-on: https://skia-review.googlesource.com/2990
Reviewed-by: Matt Sarett <msarett@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
Diffstat (limited to 'src/core/SkHalf.h')
-rw-r--r-- | src/core/SkHalf.h | 29 |
1 files changed, 29 insertions, 0 deletions
diff --git a/src/core/SkHalf.h b/src/core/SkHalf.h index dd978a2347..e71cb8750a 100644 --- a/src/core/SkHalf.h +++ b/src/core/SkHalf.h @@ -11,6 +11,10 @@ #include "SkNx.h" #include "SkTypes.h" +#if !defined(_MSC_VER) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX2 + #include <x86intrin.h> +#endif + // 16-bit floating point value // format is 1 bit sign, 5 bits exponent, 10 bits mantissa // only used for storage @@ -85,4 +89,29 @@ static inline Sk4h SkFloatToHalf_finite_ftz(const Sk4f& fs) { #endif } +static inline Sk8f SkHalfToFloat_finite_ftz(const Sk8h& hs) { +#if !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX2 + return _mm256_cvtph_ps(hs.fVec); + +#else + uint64_t parts[2]; + hs.store(parts); + return SkNx_join(SkHalfToFloat_finite_ftz(parts[0]), + SkHalfToFloat_finite_ftz(parts[1])); + +#endif +} + +static inline Sk8h SkFloatToHalf_finite_ftz(const Sk8f& fs) { +#if !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_AVX2 + return _mm256_cvtps_ph(fs.fVec, _MM_FROUND_CUR_DIRECTION); + +#else + uint64_t parts[2]; + SkFloatToHalf_finite_ftz(fs.fLo).store(parts+0); + SkFloatToHalf_finite_ftz(fs.fHi).store(parts+1); + return Sk8h::Load(parts); +#endif +} + #endif |