diff options
author | mtklein <mtklein@chromium.org> | 2015-08-31 15:26:08 -0700 |
---|---|---|
committer | Commit bot <commit-bot@chromium.org> | 2015-08-31 15:26:08 -0700 |
commit | dde03ff89f58d7df9b9e37ca06c2ce8ea67ec7a2 (patch) | |
tree | c2fec1c1cf09d7c15454c566f2a368a259a2994d /bench | |
parent | 99138876a699a41637fe8c46ccdb0292dcabd7ce (diff) |
Clean up remaining users of SkPMFloat
This switches over SkXfermodes_opts.h and SkColorMatrixFilter to use Sk4f,
and converts the SkPMFloat benches to Sk4f benches.
No pixels should change here, and no code beyond the Sk4f_ benches should change speed.
The benches are faster than the old versions.
BUG=skia:4117
Review URL: https://codereview.chromium.org/1324743002
Diffstat (limited to 'bench')
-rw-r--r-- | bench/Sk4fBench.cpp (renamed from bench/PMFloatBench.cpp) | 48 |
1 files changed, 20 insertions, 28 deletions
diff --git a/bench/PMFloatBench.cpp b/bench/Sk4fBench.cpp index 540fdb7815..53978637f1 100644 --- a/bench/PMFloatBench.cpp +++ b/bench/Sk4fBench.cpp @@ -6,7 +6,8 @@ */ #include "Benchmark.h" -#include "SkPMFloat.h" +#include "SkColor.h" +#include "SkNx.h" // Used to prevent the compiler from optimizing away the whole loop. volatile uint32_t blackhole = 0; @@ -19,11 +20,10 @@ static uint32_t lcg_rand(uint32_t* seed) { return *seed; } -// I'm having better luck getting these to constant-propagate away as template parameters. -struct PMFloatRoundtripBench : public Benchmark { - PMFloatRoundtripBench() {} +struct Sk4fBytesRoundtripBench : public Benchmark { + Sk4fBytesRoundtripBench() {} - const char* onGetName() override { return "SkPMFloat_roundtrip"; } + const char* onGetName() override { return "Sk4f_roundtrip"; } bool isSuitableFor(Backend backend) override { return backend == kNonRendering_Backend; } void onDraw(const int loops, SkCanvas* canvas) override { @@ -31,32 +31,25 @@ struct PMFloatRoundtripBench : public Benchmark { uint32_t junk = 0; uint32_t seed = 0; for (int i = 0; i < loops; i++) { - SkPMColor color; - #ifdef SK_DEBUG - // Our SkASSERTs will remind us that it's technically required that we premultiply. - color = SkPreMultiplyColor(lcg_rand(&seed)); - #else - // But it's a lot faster not to, and this code won't really mind the non-PM colors. - color = lcg_rand(&seed); - #endif - - auto f = SkPMFloat::FromPMColor(color); - SkPMColor back = f.round(); + uint32_t color = lcg_rand(&seed), + back; + auto f = Sk4f::FromBytes((const uint8_t*)&color); + f.toBytes((uint8_t*)&back); junk ^= back; } blackhole ^= junk; } }; -DEF_BENCH(return new PMFloatRoundtripBench;) +DEF_BENCH(return new Sk4fBytesRoundtripBench;) -struct PMFloatGradientBench : public Benchmark { - const char* onGetName() override { return "PMFloat_gradient"; } +struct Sk4fGradientBench : public Benchmark { + const char* onGetName() override { return "Sk4f_gradient"; } bool isSuitableFor(Backend backend) override { return backend == kNonRendering_Backend; } SkPMColor fDevice[100]; void onDraw(const int loops, SkCanvas*) override { - Sk4f c0 = SkPMFloat::FromARGB(1, 1, 0, 0), - c1 = SkPMFloat::FromARGB(1, 0, 0, 1), + Sk4f c0(0,0,255,255), + c1(255,0,0,255), dc = c1 - c0, fx(0.1f), dx(0.002f), @@ -64,15 +57,15 @@ struct PMFloatGradientBench : public Benchmark { dcdx4(dcdx+dcdx+dcdx+dcdx); for (int n = 0; n < loops; n++) { - Sk4f a = c0 + dc*fx, + Sk4f a = c0 + dc*fx + Sk4f(0.5f), // add an extra 0.5f to get rounding for free. b = a + dcdx, c = b + dcdx, d = c + dcdx; for (size_t i = 0; i < SK_ARRAY_COUNT(fDevice); i += 4) { - fDevice[i+0] = SkPMFloat(a).round(); - fDevice[i+1] = SkPMFloat(b).round(); - fDevice[i+2] = SkPMFloat(c).round(); - fDevice[i+3] = SkPMFloat(d).round(); + a.toBytes((uint8_t*)(fDevice+i+0)); + b.toBytes((uint8_t*)(fDevice+i+1)); + c.toBytes((uint8_t*)(fDevice+i+2)); + d.toBytes((uint8_t*)(fDevice+i+3)); a = a + dcdx4; b = b + dcdx4; c = c + dcdx4; @@ -81,5 +74,4 @@ struct PMFloatGradientBench : public Benchmark { } } }; - -DEF_BENCH(return new PMFloatGradientBench;) +DEF_BENCH(return new Sk4fGradientBench;) |