Clean up SkXfermode_opts.h
It seems that MSVC + __vectorcall don't play well together,
so back ourselves out into a situation where we don't need it.
- Inline transfermode functions. This removes the need for SK_VECTORCALL.
- Remove 565 destination specializations.
Blending into 565 is not speed-critical enough to merit the code bloat.
- Removing 565 specializations means a bunch of Sk4px code is now dead.
8888 xfermodes generally speed up a bit from inlining, smoothly ranging from no change down to 0.65x for the fastest functions like Plus or Modulate.
565 xfermodes generally slow down because we're doing 565 -> 8888 and 8888->565 conversion serially[1] and using the stack, smoothly ranging from no change up to 2x slower for the fastest functions like Plus and Modulate.
[1] the 565->8888 conversion is actually being autovectorized
BUG=skia:4765,skia:4776
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1565223002
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
No public API changes.
TBR=reed@google.com
Review URL: https://codereview.chromium.org/1565223002
This commit is contained in:
mtklein
Commit bot
parent
1a1efeacf7
commit
defa0daa6a
@ -300,14 +300,6 @@
|
||||
# endif
|
||||
#endif
|
||||
|
||||
#if defined(SK_BUILD_FOR_WIN) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
|
||||
#define SK_VECTORCALL __vectorcall
|
||||
#elif defined(SK_CPU_ARM32)
|
||||
#define SK_VECTORCALL __attribute__((pcs("aapcs-vfp")))
|
||||
#else
|
||||
#define SK_VECTORCALL
|
||||
#endif
|
||||
|
||||
//////////////////////////////////////////////////////////////////////
|
||||
|
||||
#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE1
|
||||
|
||||
@ -48,14 +48,6 @@ public:
|
||||
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 {
|
||||
@ -105,8 +97,8 @@ public:
|
||||
|
||||
// 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, typename Dst>
|
||||
static void MapSrc(int n, Dst* dst, const SkPMColor* src, const Fn& fn) {
|
||||
template <typename Fn>
|
||||
static void MapSrc(int n, SkPMColor* dst, const SkPMColor* src, const Fn& fn) {
|
||||
SkASSERT(dst);
|
||||
SkASSERT(src);
|
||||
// This looks a bit odd, but it helps loop-invariant hoisting across different calls to fn.
|
||||
@ -137,8 +129,8 @@ public:
|
||||
}
|
||||
|
||||
// As above, but with dst4' = fn(dst4, src4).
|
||||
template <typename Fn, typename Dst>
|
||||
static void MapDstSrc(int n, Dst* dst, const SkPMColor* src, const Fn& fn) {
|
||||
template <typename Fn>
|
||||
static void MapDstSrc(int n, SkPMColor* dst, const SkPMColor* src, const Fn& fn) {
|
||||
SkASSERT(dst);
|
||||
SkASSERT(src);
|
||||
while (n > 0) {
|
||||
@ -167,8 +159,8 @@ public:
|
||||
}
|
||||
|
||||
// As above, but with dst4' = fn(dst4, alpha4).
|
||||
template <typename Fn, typename Dst>
|
||||
static void MapDstAlpha(int n, Dst* dst, const SkAlpha* a, const Fn& fn) {
|
||||
template <typename Fn>
|
||||
static void MapDstAlpha(int n, SkPMColor* dst, const SkAlpha* a, const Fn& fn) {
|
||||
SkASSERT(dst);
|
||||
SkASSERT(a);
|
||||
while (n > 0) {
|
||||
@ -197,8 +189,8 @@ public:
|
||||
}
|
||||
|
||||
// As above, but with dst4' = fn(dst4, src4, alpha4).
|
||||
template <typename Fn, typename Dst>
|
||||
static void MapDstSrcAlpha(int n, Dst* dst, const SkPMColor* src, const SkAlpha* a,
|
||||
template <typename Fn>
|
||||
static void MapDstSrcAlpha(int n, SkPMColor* dst, const SkPMColor* src, const SkAlpha* a,
|
||||
const Fn& fn) {
|
||||
SkASSERT(dst);
|
||||
SkASSERT(src);
|
||||
|
||||
@ -95,76 +95,5 @@ inline Sk4px Sk4px::zeroAlphas() const {
|
||||
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
|
||||
|
||||
|
||||
@ -101,79 +101,4 @@ inline Sk4px Sk4px::zeroAlphas() const {
|
||||
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
|
||||
|
||||
@ -106,35 +106,4 @@ inline Sk4px Sk4px::zeroColors() const {
|
||||
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
|
||||
|
||||
@ -15,7 +15,9 @@
|
||||
namespace {
|
||||
|
||||
// Most xfermodes can be done most efficiently 4 pixels at a time in 8 or 16-bit fixed point.
|
||||
#define XFERMODE(Name) static Sk4px SK_VECTORCALL Name(Sk4px s, Sk4px d)
|
||||
#define XFERMODE(Xfermode) \
|
||||
struct Xfermode { Sk4px operator()(const Sk4px&, const Sk4px&) const; }; \
|
||||
inline Sk4px Xfermode::operator()(const Sk4px& s, const Sk4px& d) const
|
||||
|
||||
XFERMODE(Clear) { return Sk4px::DupPMColor(0); }
|
||||
XFERMODE(Src) { return s; }
|
||||
@ -23,13 +25,13 @@ XFERMODE(Dst) { return d; }
|
||||
XFERMODE(SrcIn) { return s.approxMulDiv255(d.alphas() ); }
|
||||
XFERMODE(SrcOut) { return s.approxMulDiv255(d.alphas().inv()); }
|
||||
XFERMODE(SrcOver) { return s + d.approxMulDiv255(s.alphas().inv()); }
|
||||
XFERMODE(DstIn) { return SrcIn (d,s); }
|
||||
XFERMODE(DstOut) { return SrcOut (d,s); }
|
||||
XFERMODE(DstOver) { return SrcOver(d,s); }
|
||||
XFERMODE(DstIn) { return SrcIn ()(d,s); }
|
||||
XFERMODE(DstOut) { return SrcOut ()(d,s); }
|
||||
XFERMODE(DstOver) { return SrcOver()(d,s); }
|
||||
|
||||
// [ S * Da + (1 - Sa) * D]
|
||||
XFERMODE(SrcATop) { return (s * d.alphas() + d * s.alphas().inv()).div255(); }
|
||||
XFERMODE(DstATop) { return SrcATop(d,s); }
|
||||
XFERMODE(DstATop) { return SrcATop()(d,s); }
|
||||
//[ S * (1 - Da) + (1 - Sa) * D ]
|
||||
XFERMODE(Xor) { return (s * d.alphas().inv() + d * s.alphas().inv()).div255(); }
|
||||
// [S + D ]
|
||||
@ -79,7 +81,7 @@ XFERMODE(HardLight) {
|
||||
auto colors = (both + isLite.thenElse(lite, dark)).div255();
|
||||
return alphas.zeroColors() + colors.zeroAlphas();
|
||||
}
|
||||
XFERMODE(Overlay) { return HardLight(d,s); }
|
||||
XFERMODE(Overlay) { return HardLight()(d,s); }
|
||||
|
||||
XFERMODE(Darken) {
|
||||
auto sa = s.alphas(),
|
||||
@ -110,7 +112,9 @@ XFERMODE(Lighten) {
|
||||
#undef XFERMODE
|
||||
|
||||
// Some xfermodes use math like divide or sqrt that's best done in floats 1 pixel at a time.
|
||||
#define XFERMODE(Name) static Sk4f SK_VECTORCALL Name(Sk4f d, Sk4f s)
|
||||
#define XFERMODE(Xfermode) \
|
||||
struct Xfermode { Sk4f operator()(const Sk4f&, const Sk4f&) const; }; \
|
||||
inline Sk4f Xfermode::operator()(const Sk4f& d, const Sk4f& s) const
|
||||
|
||||
static inline Sk4f a_rgb(const Sk4f& a, const Sk4f& rgb) {
|
||||
static_assert(SK_A32_SHIFT == 24, "");
|
||||
@ -181,15 +185,15 @@ XFERMODE(SoftLight) {
|
||||
|
||||
// A reasonable fallback mode for doing AA is to simply apply the transfermode first,
|
||||
// then linearly interpolate the AA.
|
||||
template <Sk4px (SK_VECTORCALL *Mode)(Sk4px, Sk4px)>
|
||||
static Sk4px SK_VECTORCALL xfer_aa(Sk4px s, Sk4px d, Sk4px aa) {
|
||||
Sk4px bw = Mode(s, d);
|
||||
template <typename Xfermode>
|
||||
static Sk4px xfer_aa(const Sk4px& s, const Sk4px& d, const Sk4px& aa) {
|
||||
Sk4px bw = Xfermode()(s, d);
|
||||
return (bw * aa + d * aa.inv()).div255();
|
||||
}
|
||||
|
||||
// For some transfermodes we specialize AA, either for correctness or performance.
|
||||
#define XFERMODE_AA(Name) \
|
||||
template <> Sk4px SK_VECTORCALL xfer_aa<Name>(Sk4px s, Sk4px d, Sk4px aa)
|
||||
#define XFERMODE_AA(Xfermode) \
|
||||
template <> Sk4px xfer_aa<Xfermode>(const Sk4px& s, const Sk4px& d, const Sk4px& aa)
|
||||
|
||||
// Plus' clamp needs to happen after AA. skia:3852
|
||||
XFERMODE_AA(Plus) { // [ clamp( (1-AA)D + (AA)(S+D) ) == clamp(D + AA*S) ]
|
||||
@ -198,95 +202,103 @@ XFERMODE_AA(Plus) { // [ clamp( (1-AA)D + (AA)(S+D) ) == clamp(D + AA*S) ]
|
||||
|
||||
#undef XFERMODE_AA
|
||||
|
||||
template <typename Xfermode>
|
||||
class Sk4pxXfermode : public SkProcCoeffXfermode {
|
||||
public:
|
||||
typedef Sk4px (SK_VECTORCALL *Proc4)(Sk4px, Sk4px);
|
||||
typedef Sk4px (SK_VECTORCALL *AAProc4)(Sk4px, Sk4px, Sk4px);
|
||||
|
||||
Sk4pxXfermode(const ProcCoeff& rec, SkXfermode::Mode mode, Proc4 proc4, AAProc4 aaproc4)
|
||||
: INHERITED(rec, mode)
|
||||
, fProc4(proc4)
|
||||
, fAAProc4(aaproc4) {}
|
||||
Sk4pxXfermode(const ProcCoeff& rec, SkXfermode::Mode mode)
|
||||
: INHERITED(rec, mode) {}
|
||||
|
||||
void xfer32(SkPMColor dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
|
||||
if (nullptr == aa) {
|
||||
Sk4px::MapDstSrc(n, dst, src, [&](const Sk4px& dst4, const Sk4px& src4) {
|
||||
return fProc4(src4, dst4);
|
||||
return Xfermode()(src4, dst4);
|
||||
});
|
||||
} else {
|
||||
Sk4px::MapDstSrcAlpha(n, dst, src, aa,
|
||||
[&](const Sk4px& dst4, const Sk4px& src4, const Sk4px& alpha) {
|
||||
return fAAProc4(src4, dst4, alpha);
|
||||
});
|
||||
return xfer_aa<Xfermode>(src4, dst4, alpha);
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
void xfer16(uint16_t dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
|
||||
if (nullptr == 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);
|
||||
});
|
||||
SkPMColor dst32[4];
|
||||
while (n >= 4) {
|
||||
dst32[0] = SkPixel16ToPixel32(dst[0]);
|
||||
dst32[1] = SkPixel16ToPixel32(dst[1]);
|
||||
dst32[2] = SkPixel16ToPixel32(dst[2]);
|
||||
dst32[3] = SkPixel16ToPixel32(dst[3]);
|
||||
|
||||
this->xfer32(dst32, src, 4, aa);
|
||||
|
||||
dst[0] = SkPixel32ToPixel16(dst32[0]);
|
||||
dst[1] = SkPixel32ToPixel16(dst32[1]);
|
||||
dst[2] = SkPixel32ToPixel16(dst32[2]);
|
||||
dst[3] = SkPixel32ToPixel16(dst32[3]);
|
||||
|
||||
dst += 4;
|
||||
src += 4;
|
||||
aa += aa ? 4 : 0;
|
||||
n -= 4;
|
||||
}
|
||||
while (n) {
|
||||
SkPMColor dst32 = SkPixel16ToPixel32(*dst);
|
||||
this->xfer32(&dst32, src, 1, aa);
|
||||
*dst = SkPixel32ToPixel16(dst32);
|
||||
|
||||
dst += 1;
|
||||
src += 1;
|
||||
aa += aa ? 1 : 0;
|
||||
n -= 1;
|
||||
}
|
||||
}
|
||||
|
||||
private:
|
||||
Proc4 fProc4;
|
||||
AAProc4 fAAProc4;
|
||||
typedef SkProcCoeffXfermode INHERITED;
|
||||
};
|
||||
|
||||
template <typename Xfermode>
|
||||
class Sk4fXfermode : public SkProcCoeffXfermode {
|
||||
public:
|
||||
typedef Sk4f (SK_VECTORCALL *ProcF)(Sk4f, Sk4f);
|
||||
Sk4fXfermode(const ProcCoeff& rec, SkXfermode::Mode mode, ProcF procf)
|
||||
: INHERITED(rec, mode)
|
||||
, fProcF(procf) {}
|
||||
Sk4fXfermode(const ProcCoeff& rec, SkXfermode::Mode mode)
|
||||
: INHERITED(rec, mode) {}
|
||||
|
||||
void xfer32(SkPMColor dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
|
||||
for (int i = 0; i < n; i++) {
|
||||
dst[i] = aa ? this->xfer32(dst[i], src[i], aa[i])
|
||||
: this->xfer32(dst[i], src[i]);
|
||||
dst[i] = Xfer32_1(dst[i], src[i], aa ? aa+i : nullptr);
|
||||
}
|
||||
}
|
||||
|
||||
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]);
|
||||
dst32 = Xfer32_1(dst32, src[i], aa ? aa+i : nullptr);
|
||||
dst[i] = SkPixel32ToPixel16(dst32);
|
||||
}
|
||||
}
|
||||
|
||||
private:
|
||||
static SkPMColor Xfer32_1(SkPMColor dst, const SkPMColor src, const SkAlpha* aa) {
|
||||
Sk4f d = Load(dst),
|
||||
s = Load(src),
|
||||
b = Xfermode()(d, s);
|
||||
if (aa) {
|
||||
Sk4f a = Sk4f(*aa) * Sk4f(1.0f/255);
|
||||
b = b*a + d*(Sk4f(1)-a);
|
||||
}
|
||||
return Round(b);
|
||||
}
|
||||
|
||||
static Sk4f Load(SkPMColor c) {
|
||||
return SkNx_cast<float>(Sk4b::Load((uint8_t*)&c)) * Sk4f(1.0f/255);
|
||||
}
|
||||
|
||||
static SkPMColor Round(const Sk4f& f) {
|
||||
SkPMColor c;
|
||||
SkNx_cast<uint8_t>(f * Sk4f(255) + Sk4f(0.5f)).store((uint8_t*)&c);
|
||||
return c;
|
||||
}
|
||||
inline SkPMColor xfer32(SkPMColor dst, SkPMColor src) const {
|
||||
return Round(fProcF(Load(dst), Load(src)));
|
||||
}
|
||||
|
||||
inline SkPMColor xfer32(SkPMColor dst, SkPMColor src, SkAlpha aa) const {
|
||||
Sk4f s(Load(src)),
|
||||
d(Load(dst)),
|
||||
b(fProcF(d,s));
|
||||
// We do aa in full float precision before going back down to bytes, because we can!
|
||||
Sk4f a = Sk4f(aa) * Sk4f(1.0f/255);
|
||||
b = b*a + d*(Sk4f(1)-a);
|
||||
return Round(b);
|
||||
}
|
||||
|
||||
ProcF fProcF;
|
||||
typedef SkProcCoeffXfermode INHERITED;
|
||||
};
|
||||
|
||||
@ -296,8 +308,8 @@ namespace SK_OPTS_NS {
|
||||
|
||||
static SkXfermode* create_xfermode(const ProcCoeff& rec, SkXfermode::Mode mode) {
|
||||
switch (mode) {
|
||||
#define CASE(Mode) \
|
||||
case SkXfermode::k##Mode##_Mode: return new Sk4pxXfermode(rec, mode, &Mode, &xfer_aa<Mode>)
|
||||
#define CASE(Xfermode) \
|
||||
case SkXfermode::k##Xfermode##_Mode: return new Sk4pxXfermode<Xfermode>(rec, mode)
|
||||
CASE(Clear);
|
||||
CASE(Src);
|
||||
CASE(Dst);
|
||||
@ -322,8 +334,8 @@ static SkXfermode* create_xfermode(const ProcCoeff& rec, SkXfermode::Mode mode)
|
||||
CASE(Lighten);
|
||||
#undef CASE
|
||||
|
||||
#define CASE(Mode) \
|
||||
case SkXfermode::k##Mode##_Mode: return new Sk4fXfermode(rec, mode, &Mode)
|
||||
#define CASE(Xfermode) \
|
||||
case SkXfermode::k##Xfermode##_Mode: return new Sk4fXfermode<Xfermode>(rec, mode)
|
||||
CASE(ColorDodge);
|
||||
CASE(ColorBurn);
|
||||
CASE(SoftLight);
|
||||
|
||||
Loading…
Reference in New Issue
Block a user