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

Add Sk4f::ToBytes(uint8_t[16], Sk4f, Sk4f, Sk4f, Sk4f)

Browse Source 
This is a big speedup for float -> byte.  E.g. gradient_linear_clamp_3color:
 x86-64 147µs -> 103µs  (Broadwell MBP)
 arm64 2.03ms -> 648µs  (Galaxy S6)
 armv7 1.12ms -> 489µs  (Galaxy S6, same device!)

BUG=skia:
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot;client.skia.android:Test-Android-GCC-Nexus9-CPU-Denver-Arm64-Debug-Trybot

Review URL: https://codereview.chromium.org/1483953002
This commit is contained in:
mtklein 2015-12-01 07:10:21 -08:00 committed by Commit bot
parent eeebdb538d
commit 9db43ac4ee
5 changed files with 37 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
bench/Sk4fBench.cpp
View File

@ -62,10 +62,7 @@ struct Sk4fGradientBench : public Benchmark {
c = b + dcdx,
d = c + dcdx;
for (size_t i = 0; i < SK_ARRAY_COUNT(fDevice); i += 4) {
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));
Sk4f::ToBytes((uint8_t*)(fDevice+i), a, b, c, d);
a = a + dcdx4;
b = b + dcdx4;
c = c + dcdx4;

9
src/core/SkNx.h
View File

@ -113,6 +113,15 @@ public:
fHi.toBytes(bytes+N/2);
}
// Some implementations can do this faster.
static void ToBytes(uint8_t bytes[4*N],
const SkNx& a, const SkNx& b, const SkNx& c, const SkNx& d) {
a.toBytes(bytes+0*N);
b.toBytes(bytes+1*N);
c.toBytes(bytes+2*N);
d.toBytes(bytes+3*N);
}
SkNx operator + (const SkNx& o) const { return SkNx(fLo + o.fLo, fHi + o.fHi); }
SkNx operator - (const SkNx& o) const { return SkNx(fLo - o.fLo, fHi - o.fHi); }
SkNx operator * (const SkNx& o) const { return SkNx(fLo * o.fLo, fHi * o.fHi); }

5
src/effects/gradients/SkLinearGradient.cpp
View File

@ -748,10 +748,15 @@ template <bool apply_alpha> void ramp(SkPMColor dstC[], int n, const Sk4f& c, co
Sk4f cd2 = cd0 + dc2;
Sk4f cd3 = cd1 + dc2;
while (n >= 4) {
if (!apply_alpha) {
Sk4f::ToBytes((uint8_t*)dstC, cd0, cd1, cd2, cd3);
dstC += 4;
} else {
*dstC++ = trunc_from_255<apply_alpha>(cd0);
*dstC++ = trunc_from_255<apply_alpha>(cd1);
*dstC++ = trunc_from_255<apply_alpha>(cd2);
*dstC++ = trunc_from_255<apply_alpha>(cd3);
}
cd0 = cd0 + dc4;
cd1 = cd1 + dc4;
cd2 = cd2 + dc4;

8
src/opts/SkNx_neon.h
View File

@ -165,6 +165,14 @@ public:
vst1_lane_u32((uint32_t*)bytes, (uint32x2_t)fix8, 0);
}
static void ToBytes(uint8_t bytes[16],
const SkNx& a, const SkNx& b, const SkNx& c, const SkNx& d) {
vst1q_u8(bytes, vuzpq_u8(vuzpq_u8((uint8x16_t)vcvtq_u32_f32(a.fVec),
(uint8x16_t)vcvtq_u32_f32(b.fVec)).val[0],
vuzpq_u8((uint8x16_t)vcvtq_u32_f32(c.fVec),
(uint8x16_t)vcvtq_u32_f32(d.fVec)).val[0]).val[0]);
}
SkNx approxInvert() const {
float32x4_t est0 = vrecpeq_f32(fVec),
est1 = vmulq_f32(vrecpsq_f32(est0, fVec), est0);

9
src/opts/SkNx_sse.h
View File

@ -132,6 +132,15 @@ public:
*(int*)bytes = _mm_cvtsi128_si32(fix8);
}
static void ToBytes(uint8_t bytes[16],
const SkNx& a, const SkNx& b, const SkNx& c, const SkNx& d) {
_mm_storeu_si128((__m128i*)bytes,
_mm_packus_epi16(_mm_packus_epi16(_mm_cvttps_epi32(a.fVec),
_mm_cvttps_epi32(b.fVec)),
_mm_packus_epi16(_mm_cvttps_epi32(c.fVec),
_mm_cvttps_epi32(d.fVec))));
}
SkNx operator + (const SkNx& o) const { return _mm_add_ps(fVec, o.fVec); }
SkNx operator - (const SkNx& o) const { return _mm_sub_ps(fVec, o.fVec); }
SkNx operator * (const SkNx& o) const { return _mm_mul_ps(fVec, o.fVec); }