565 support for SIMD xfermodes
This uses the most basic approach possible:
- to load an Sk4px from 565, convert to SkPMColors on the stack serially then load those SkPMColors.
- to store an Sk4px to 565, store to SkPMColors on the stack then convert to 565 serially.
Clearly, we can optimize these loads and stores. That's a TODO.
The code using SkPMFloat is the same idea but a little more long-term viable, as we're only operating on one pixel at a time anyway. We could probably write 565 <-> SkPMFloat methods, but I'd rather not until it's really compelling.
The speedups are varied but similar across SSE and NEON: a few uninteresting, many 50% faster, some 2x faster, and SoftLight ~4x faster.
This will cause minor GM diffs, but I don't think any layout test changes.
BUG=skia:
Committed: https://skia.googlesource.com/skia/+/942930dcaa51f66d82cdaf46ae62efebd16c8cd0
Committed: https://skia.googlesource.com/skia/+/860dcaa2ddfdadc050af4f943a84a9d499315066
Review URL: https://codereview.chromium.org/1245673002
diff --git a/src/core/Sk4px.h b/src/core/Sk4px.h
index e1d4dc1..24a21c66 100644
--- a/src/core/Sk4px.h
+++ b/src/core/Sk4px.h
@@ -10,6 +10,7 @@
#include "SkNx.h"
#include "SkColor.h"
+#include "SkColorPriv.h"
// This file may be included multiple times by .cpp files with different flags, leading
// to different definitions. Usually that doesn't matter because it's all inlined, but
@@ -47,6 +48,14 @@
void store2(SkPMColor[2]) const;
void store1(SkPMColor[1]) const;
+ // Same as above for 565.
+ static Sk4px Load4(const SkPMColor16 src[4]);
+ static Sk4px Load2(const SkPMColor16 src[2]);
+ static Sk4px Load1(const SkPMColor16 src[1]);
+ void store4(SkPMColor16 dst[4]) const;
+ void store2(SkPMColor16 dst[2]) const;
+ void store1(SkPMColor16 dst[1]) const;
+
// 1, 2, or 4 SkPMColors with 16-bit components.
// This is most useful as the result of a multiply, e.g. from mulWiden().
class Wide : public Sk16h {
@@ -99,8 +108,8 @@
// A generic driver that maps fn over a src array into a dst array.
// fn should take an Sk4px (4 src pixels) and return an Sk4px (4 dst pixels).
- template <typename Fn>
- static void MapSrc(int n, SkPMColor* dst, const SkPMColor* src, const Fn& fn) {
+ template <typename Fn, typename Dst>
+ static void MapSrc(int n, Dst* dst, const SkPMColor* src, const Fn& fn) {
// This looks a bit odd, but it helps loop-invariant hoisting across different calls to fn.
// Basically, we need to make sure we keep things inside a single loop.
while (n > 0) {
@@ -129,8 +138,8 @@
}
// As above, but with dst4' = fn(dst4, src4).
- template <typename Fn>
- static void MapDstSrc(int n, SkPMColor* dst, const SkPMColor* src, const Fn& fn) {
+ template <typename Fn, typename Dst>
+ static void MapDstSrc(int n, Dst* dst, const SkPMColor* src, const Fn& fn) {
while (n > 0) {
if (n >= 8) {
Sk4px dst0 = fn(Load4(dst+0), Load4(src+0)),
@@ -157,8 +166,8 @@
}
// As above, but with dst4' = fn(dst4, src4, alpha4).
- template <typename Fn>
- static void MapDstSrcAlpha(int n, SkPMColor* dst, const SkPMColor* src, const SkAlpha* a,
+ template <typename Fn, typename Dst>
+ static void MapDstSrcAlpha(int n, Dst* dst, const SkPMColor* src, const SkAlpha* a,
const Fn& fn) {
while (n > 0) {
if (n >= 8) {
diff --git a/src/core/Sk4pxXfermode.h b/src/core/Sk4pxXfermode.h
index fc0b643..ad822ed 100644
--- a/src/core/Sk4pxXfermode.h
+++ b/src/core/Sk4pxXfermode.h
@@ -222,6 +222,19 @@
}
}
+ void xfer16(uint16_t dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
+ if (NULL == aa) {
+ Sk4px::MapDstSrc(n, dst, src, [&](const Sk4px& dst4, const Sk4px& src4) {
+ return fProc4(src4, dst4);
+ });
+ } else {
+ Sk4px::MapDstSrcAlpha(n, dst, src, aa,
+ [&](const Sk4px& dst4, const Sk4px& src4, const Sk4px& alpha) {
+ return fAAProc4(src4, dst4, alpha);
+ });
+ }
+ }
+
private:
Proc4 fProc4;
AAProc4 fAAProc4;
@@ -237,19 +250,35 @@
void xfer32(SkPMColor dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
for (int i = 0; i < n; i++) {
- SkPMFloat s(src[i]),
- d(dst[i]),
- b(fProcF(s,d));
- if (aa) {
- // We do aa in full float precision before going back down to bytes, because we can!
- SkPMFloat a = Sk4f(aa[i]) * Sk4f(1.0f/255);
- b = b*a + d*(Sk4f(1)-a);
- }
- dst[i] = b.round();
+ dst[i] = aa ? this->xfer32(dst[i], src[i], aa[i])
+ : this->xfer32(dst[i], src[i]);
+ }
+ }
+
+ void xfer16(uint16_t dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
+ for (int i = 0; i < n; i++) {
+ SkPMColor dst32 = SkPixel16ToPixel32(dst[i]);
+ dst32 = aa ? this->xfer32(dst32, src[i], aa[i])
+ : this->xfer32(dst32, src[i]);
+ dst[i] = SkPixel32ToPixel16(dst32);
}
}
private:
+ inline SkPMColor xfer32(SkPMColor dst, SkPMColor src) const {
+ return fProcF(SkPMFloat(src), SkPMFloat(dst)).round();
+ }
+
+ inline SkPMColor xfer32(SkPMColor dst, SkPMColor src, SkAlpha aa) const {
+ SkPMFloat s(src),
+ d(dst),
+ b(fProcF(s,d));
+ // We do aa in full float precision before going back down to bytes, because we can!
+ SkPMFloat a = Sk4f(aa) * Sk4f(1.0f/255);
+ b = b*a + d*(Sk4f(1)-a);
+ return b.round();
+ }
+
ProcF fProcF;
typedef SkProcCoeffXfermode INHERITED;
};
diff --git a/src/opts/Sk4px_NEON.h b/src/opts/Sk4px_NEON.h
index cd6dea9..89841d9 100644
--- a/src/opts/Sk4px_NEON.h
+++ b/src/opts/Sk4px_NEON.h
@@ -89,5 +89,76 @@
return Sk16b(vbicq_u8(this->fVec, (uint8x16_t)vdupq_n_u32(0xFF << SK_A32_SHIFT)));
}
+static inline uint8x16_t widen_to_8888(uint16x4_t v) {
+ // RGB565 format: |R....|G.....|B....|
+ // Bit: 16 11 5 0
+
+ // First get each pixel into its own 32-bit lane.
+ // v == rgb3 rgb2 rgb1 rgb0
+ // spread == 0000 rgb3 0000 rgb2 0000 rgb1 0000 rgb0
+ uint32x4_t spread = vmovl_u16(v);
+
+ // Get each color independently, still in 565 precison but down at bit 0.
+ auto r5 = vshrq_n_u32(spread, 11),
+ g6 = vandq_u32(vdupq_n_u32(63), vshrq_n_u32(spread, 5)),
+ b5 = vandq_u32(vdupq_n_u32(31), spread);
+
+ // Scale 565 precision up to 8-bit each, filling low 323 bits with high bits of each component.
+ auto r8 = vorrq_u32(vshlq_n_u32(r5, 3), vshrq_n_u32(r5, 2)),
+ g8 = vorrq_u32(vshlq_n_u32(g6, 2), vshrq_n_u32(g6, 4)),
+ b8 = vorrq_u32(vshlq_n_u32(b5, 3), vshrq_n_u32(b5, 2));
+
+ // Now put all the 8-bit components into SkPMColor order.
+ return (uint8x16_t)vorrq_u32(vshlq_n_u32(r8, SK_R32_SHIFT), // TODO: one shift is zero...
+ vorrq_u32(vshlq_n_u32(g8, SK_G32_SHIFT),
+ vorrq_u32(vshlq_n_u32(b8, SK_B32_SHIFT),
+ vdupq_n_u32(0xFF << SK_A32_SHIFT))));
+}
+
+static inline uint16x4_t narrow_to_565(uint8x16_t w8x16) {
+ uint32x4_t w = (uint32x4_t)w8x16;
+
+ // Extract out top RGB 565 bits of each pixel, with no rounding.
+ auto r5 = vandq_u32(vdupq_n_u32(31), vshrq_n_u32(w, SK_R32_SHIFT + 3)),
+ g6 = vandq_u32(vdupq_n_u32(63), vshrq_n_u32(w, SK_G32_SHIFT + 2)),
+ b5 = vandq_u32(vdupq_n_u32(31), vshrq_n_u32(w, SK_B32_SHIFT + 3));
+
+ // Now put the bits in place in the low 16-bits of each 32-bit lane.
+ auto spread = vorrq_u32(vshlq_n_u32(r5, 11),
+ vorrq_u32(vshlq_n_u32(g6, 5),
+ b5));
+
+ // Pack the low 16-bits of our 128-bit register down into a 64-bit register.
+ // spread == 0000 rgb3 0000 rgb2 0000 rgb1 0000 rgb0
+ // v == rgb3 rgb2 rgb1 rgb0
+ auto v = vmovn_u32(spread);
+ return v;
+}
+
+
+inline Sk4px Sk4px::Load4(const SkPMColor16 src[4]) {
+ return Sk16b(widen_to_8888(vld1_u16(src)));
+}
+inline Sk4px Sk4px::Load2(const SkPMColor16 src[2]) {
+ auto src2 = ((uint32_t)src[0] )
+ | ((uint32_t)src[1] << 16);
+ return Sk16b(widen_to_8888(vcreate_u16(src2)));
+}
+inline Sk4px Sk4px::Load1(const SkPMColor16 src[1]) {
+ return Sk16b(widen_to_8888(vcreate_u16(src[0])));
+}
+
+inline void Sk4px::store4(SkPMColor16 dst[4]) const {
+ vst1_u16(dst, narrow_to_565(this->fVec));
+}
+inline void Sk4px::store2(SkPMColor16 dst[2]) const {
+ auto v = narrow_to_565(this->fVec);
+ dst[0] = vget_lane_u16(v, 0);
+ dst[1] = vget_lane_u16(v, 1);
+}
+inline void Sk4px::store1(SkPMColor16 dst[1]) const {
+ dst[0] = vget_lane_u16(narrow_to_565(this->fVec), 0);
+}
+
} // namespace
diff --git a/src/opts/Sk4px_SSE2.h b/src/opts/Sk4px_SSE2.h
index 3809c5e..5e97abf 100644
--- a/src/opts/Sk4px_SSE2.h
+++ b/src/opts/Sk4px_SSE2.h
@@ -93,4 +93,79 @@
return Sk16b(_mm_andnot_si128(_mm_set1_epi32(0xFF << SK_A32_SHIFT), this->fVec));
}
+static inline __m128i widen_low_half_to_8888(__m128i v) {
+ // RGB565 format: |R....|G.....|B....|
+ // Bit: 16 11 5 0
+
+ // First get each pixel into its own 32-bit lane.
+ // v == ____ ____ ____ ____ rgb3 rgb2 rgb1 rgb0
+ // spread == 0000 rgb3 0000 rgb2 0000 rgb1 0000 rgb0
+ auto spread = _mm_unpacklo_epi16(v, _mm_setzero_si128());
+
+ // Get each color independently, still in 565 precison but down at bit 0.
+ auto r5 = _mm_srli_epi32(spread, 11),
+ g6 = _mm_and_si128(_mm_set1_epi32(63), _mm_srli_epi32(spread, 5)),
+ b5 = _mm_and_si128(_mm_set1_epi32(31), spread);
+
+ // Scale 565 precision up to 8-bit each, filling low 323 bits with high bits of each component.
+ auto r8 = _mm_or_si128(_mm_slli_epi32(r5, 3), _mm_srli_epi32(r5, 2)),
+ g8 = _mm_or_si128(_mm_slli_epi32(g6, 2), _mm_srli_epi32(g6, 4)),
+ b8 = _mm_or_si128(_mm_slli_epi32(b5, 3), _mm_srli_epi32(b5, 2));
+
+ // Now put all the 8-bit components into SkPMColor order.
+ return _mm_or_si128(_mm_slli_epi32(r8, SK_R32_SHIFT), // TODO: one of these shifts is zero...
+ _mm_or_si128(_mm_slli_epi32(g8, SK_G32_SHIFT),
+ _mm_or_si128(_mm_slli_epi32(b8, SK_B32_SHIFT),
+ _mm_set1_epi32(0xFF << SK_A32_SHIFT))));
+}
+
+static inline __m128i narrow_to_565(__m128i w) {
+ // Extract out top RGB 565 bits of each pixel, with no rounding.
+ auto r5 = _mm_and_si128(_mm_set1_epi32(31), _mm_srli_epi32(w, SK_R32_SHIFT + 3)),
+ g6 = _mm_and_si128(_mm_set1_epi32(63), _mm_srli_epi32(w, SK_G32_SHIFT + 2)),
+ b5 = _mm_and_si128(_mm_set1_epi32(31), _mm_srli_epi32(w, SK_B32_SHIFT + 3));
+
+ // Now put the bits in place in the low 16-bits of each 32-bit lane.
+ auto spread = _mm_or_si128(_mm_slli_epi32(r5, 11),
+ _mm_or_si128(_mm_slli_epi32(g6, 5),
+ b5));
+
+ // We want to pack the bottom 16-bits of spread down into the low half of the register, v.
+ // spread == 0000 rgb3 0000 rgb2 0000 rgb1 0000 rgb0
+ // v == ____ ____ ____ ____ rgb3 rgb2 rgb1 rgb0
+
+ // Ideally now we'd use _mm_packus_epi32(spread, <anything>) to pack v. But that's from SSE4.
+ // With only SSE2, we need to use _mm_packs_epi32. That does signed saturation, and
+ // we need to preserve all 16 bits. So we pretend our data is signed by sign-extending first.
+ // TODO: is it faster to just _mm_shuffle_epi8 this when we have SSSE3?
+ auto signExtended = _mm_srai_epi32(_mm_slli_epi32(spread, 16), 16);
+ auto v = _mm_packs_epi32(signExtended, signExtended);
+ return v;
+}
+
+inline Sk4px Sk4px::Load4(const SkPMColor16 src[4]) {
+ return Sk16b(widen_low_half_to_8888(_mm_loadl_epi64((const __m128i*)src)));
+}
+inline Sk4px Sk4px::Load2(const SkPMColor16 src[2]) {
+ auto src2 = ((uint32_t)src[0] )
+ | ((uint32_t)src[1] << 16);
+ return Sk16b(widen_low_half_to_8888(_mm_cvtsi32_si128(src2)));
+}
+inline Sk4px Sk4px::Load1(const SkPMColor16 src[1]) {
+ return Sk16b(widen_low_half_to_8888(_mm_insert_epi16(_mm_setzero_si128(), src[0], 0)));
+}
+
+inline void Sk4px::store4(SkPMColor16 dst[4]) const {
+ _mm_storel_epi64((__m128i*)dst, narrow_to_565(this->fVec));
+}
+inline void Sk4px::store2(SkPMColor16 dst[2]) const {
+ uint32_t dst2 = _mm_cvtsi128_si32(narrow_to_565(this->fVec));
+ dst[0] = dst2;
+ dst[1] = dst2 >> 16;
+}
+inline void Sk4px::store1(SkPMColor16 dst[1]) const {
+ uint32_t dst2 = _mm_cvtsi128_si32(narrow_to_565(this->fVec));
+ dst[0] = dst2;
+}
+
} // namespace
diff --git a/src/opts/Sk4px_none.h b/src/opts/Sk4px_none.h
index ba13e58..540edb8 100644
--- a/src/opts/Sk4px_none.h
+++ b/src/opts/Sk4px_none.h
@@ -100,4 +100,35 @@
0,0,0, this->kth<15>());
}
+inline Sk4px Sk4px::Load4(const SkPMColor16 src[4]) {
+ SkPMColor src32[4];
+ for (int i = 0; i < 4; i++) { src32[i] = SkPixel16ToPixel32(src[i]); }
+ return Load4(src32);
+}
+inline Sk4px Sk4px::Load2(const SkPMColor16 src[2]) {
+ SkPMColor src32[2];
+ for (int i = 0; i < 2; i++) { src32[i] = SkPixel16ToPixel32(src[i]); }
+ return Load2(src32);
+}
+inline Sk4px Sk4px::Load1(const SkPMColor16 src[1]) {
+ SkPMColor src32 = SkPixel16ToPixel32(src[0]);
+ return Load1(&src32);
+}
+
+inline void Sk4px::store4(SkPMColor16 dst[4]) const {
+ SkPMColor dst32[4];
+ this->store4(dst32);
+ for (int i = 0; i < 4; i++) { dst[i] = SkPixel32ToPixel16(dst32[i]); }
+}
+inline void Sk4px::store2(SkPMColor16 dst[2]) const {
+ SkPMColor dst32[2];
+ this->store2(dst32);
+ for (int i = 0; i < 2; i++) { dst[i] = SkPixel32ToPixel16(dst32[i]); }
+}
+inline void Sk4px::store1(SkPMColor16 dst[1]) const {
+ SkPMColor dst32;
+ this->store1(&dst32);
+ dst[0] = SkPixel32ToPixel16(dst32);
+}
+
} // namespace
