Use sk_srgb_to_linear_trunc in SkColorXform_opts

This gives us a little more control over instruction order, allowing
us to pipeline the muls and get better performance.  Technically,
clang should be able to do this for us anyway...

Performance on HP z620 (201295.jpg):
toSRGB: 371us -> 356us

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2175413002
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Review-Url: https://codereview.chromium.org/2175413002
This commit is contained in:
msarett 2016-07-25 18:58:11 -07:00 committed by Commit bot
parent 3418c0e797
commit a18c3b8306

View File

@ -16,14 +16,19 @@
namespace SK_OPTS_NS { namespace SK_OPTS_NS {
static Sk4i linear_to_2dot2(const Sk4f& x) { // Strange that we need a wrapper on SkNx_cast to use as a function ptr.
static Sk4i Sk4f_trunc(const Sk4f& x) {
return SkNx_cast<int>(x);
}
static Sk4f linear_to_2dot2(const Sk4f& x) {
// x^(29/64) is a very good approximation of the true value, x^(1/2.2). // x^(29/64) is a very good approximation of the true value, x^(1/2.2).
auto x2 = x.rsqrt(), // x^(-1/2) auto x2 = x.rsqrt(), // x^(-1/2)
x32 = x2.rsqrt().rsqrt().rsqrt().rsqrt(), // x^(-1/32) x32 = x2.rsqrt().rsqrt().rsqrt().rsqrt(), // x^(-1/32)
x64 = x32.rsqrt(); // x^(+1/64) x64 = x32.rsqrt(); // x^(+1/64)
// 29 = 32 - 2 - 1 // 29 = 32 - 2 - 1
return Sk4f_round(sk_clamp_0_255(255.0f * x2.invert() * x32 * x64.invert())); return 255.0f * x2.invert() * x32 * x64.invert();
} }
enum DstGamma { enum DstGamma {
@ -74,16 +79,22 @@ static void color_xform_RGB1(void* dst, const uint32_t* src, int len,
auto store_4 = [&dstReds, &dstGreens, &dstBlues, &dst, &dstTables] { auto store_4 = [&dstReds, &dstGreens, &dstBlues, &dst, &dstTables] {
if (kSRGB_DstGamma == kDstGamma || k2Dot2_DstGamma == kDstGamma) { if (kSRGB_DstGamma == kDstGamma || k2Dot2_DstGamma == kDstGamma) {
Sk4i (*linear_to_curve)(const Sk4f&) = Sk4f (*linear_to_curve)(const Sk4f&) = (kSRGB_DstGamma == kDstGamma) ?
(kSRGB_DstGamma == kDstGamma) ? sk_linear_to_srgb : linear_to_2dot2; sk_linear_to_srgb_needs_trunc : linear_to_2dot2;
Sk4i (*float_to_int)(const Sk4f&) = (kSRGB_DstGamma == kDstGamma) ?
Sk4f_trunc : Sk4f_round;
auto reds = linear_to_curve(dstReds); dstReds = linear_to_curve(dstReds);
auto greens = linear_to_curve(dstGreens); dstGreens = linear_to_curve(dstGreens);
auto blues = linear_to_curve(dstBlues); dstBlues = linear_to_curve(dstBlues);
auto rgba = (reds << SK_R32_SHIFT) dstReds = sk_clamp_0_255(dstReds);
| (greens << SK_G32_SHIFT) dstGreens = sk_clamp_0_255(dstGreens);
| (blues << SK_B32_SHIFT) dstBlues = sk_clamp_0_255(dstBlues);
auto rgba = (float_to_int(dstReds) << SK_R32_SHIFT)
| (float_to_int(dstGreens) << SK_G32_SHIFT)
| (float_to_int(dstBlues) << SK_B32_SHIFT)
| (Sk4i{0xFF} << SK_A32_SHIFT); | (Sk4i{0xFF} << SK_A32_SHIFT);
rgba.store((uint32_t*) dst); rgba.store((uint32_t*) dst);
@ -146,13 +157,15 @@ static void color_xform_RGB1(void* dst, const uint32_t* src, int len,
auto dstPixel = rXgXbX*r + rYgYbY*g + rZgZbZ*b + rTgTbT; auto dstPixel = rXgXbX*r + rYgYbY*g + rZgZbZ*b + rTgTbT;
if (kSRGB_DstGamma == kDstGamma || k2Dot2_DstGamma == kDstGamma) { if (kSRGB_DstGamma == kDstGamma || k2Dot2_DstGamma == kDstGamma) {
Sk4i (*linear_to_curve)(const Sk4f&) = Sk4f (*linear_to_curve)(const Sk4f&) = (kSRGB_DstGamma == kDstGamma) ?
(kSRGB_DstGamma == kDstGamma) ? sk_linear_to_srgb : linear_to_2dot2; sk_linear_to_srgb_needs_trunc : linear_to_2dot2;
Sk4i (*float_to_int)(const Sk4f&) = (kSRGB_DstGamma == kDstGamma) ?
Sk4f_trunc : Sk4f_round;
auto pixel = linear_to_curve(dstPixel); dstPixel = sk_clamp_0_255(linear_to_curve(dstPixel));
uint32_t rgba; uint32_t rgba;
SkNx_cast<uint8_t>(pixel).store(&rgba); SkNx_cast<uint8_t>(float_to_int(dstPixel)).store(&rgba);
rgba |= 0xFF000000; rgba |= 0xFF000000;
*((uint32_t*) dst) = SkSwizzle_RGBA_to_PMColor(rgba); *((uint32_t*) dst) = SkSwizzle_RGBA_to_PMColor(rgba);
dst = SkTAddOffset<void>(dst, sizeof(uint32_t)); dst = SkTAddOffset<void>(dst, sizeof(uint32_t));