#include "SkUtilsArm.h"
#include "SkWriteBuffer.h"
-// When implemented, the Sk4f and Sk4px xfermodes beat src/opts/SkXfermodes_opts_SSE2's.
-// When implemented, the Sk4px, but not Sk4f, xfermodes beat src/opts/SkXfermodes_arm_neon's.
-#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
- #define SK_4F_XFERMODES_ARE_FAST
- #define SK_4PX_XFERMODES_ARE_FAST
-#elif defined(SK_ARM_HAS_NEON)
- #define SK_4PX_XFERMODES_ARE_FAST
+#if SK_CPU_X86 && SK_CPU_SSE_LEVEL < SK_CPU_SSE_LEVEL_SSE2
+ #warning "SkXfermode will be much faster if you compile with support for SSE2."
+#endif
+
+#if SK_CPU_X86 || defined(SK_ARM_HAS_NEON)
+ #define SK_USE_4PX_XFERMODES
#endif
#if !SK_ARM_NEON_IS_NONE
///////////////////////////////////////////////////////////////////////////////
-/* These modes can merge coverage into src-alpha
- *
-{ dst_modeproc, SkXfermode::kZero_Coeff, SkXfermode::kOne_Coeff },
-{ srcover_modeproc, SkXfermode::kOne_Coeff, SkXfermode::kISA_Coeff },
-{ dstover_modeproc, SkXfermode::kIDA_Coeff, SkXfermode::kOne_Coeff },
-{ dstout_modeproc, SkXfermode::kZero_Coeff, SkXfermode::kISA_Coeff },
-{ srcatop_modeproc, SkXfermode::kDA_Coeff, SkXfermode::kISA_Coeff },
-{ xor_modeproc, SkXfermode::kIDA_Coeff, SkXfermode::kISA_Coeff },
-{ plus_modeproc, SkXfermode::kOne_Coeff, SkXfermode::kOne_Coeff },
-{ screen_modeproc, SkXfermode::kOne_Coeff, SkXfermode::kISC_Coeff },
-*/
-
-static const float gInv255 = 0.0039215683f; // (1.0f / 255) - ULP == SkBits2Float(0x3B808080)
-
-static Sk4f ramp(const Sk4f& v0, const Sk4f& v1, const Sk4f& t) {
- return v0 + (v1 - v0) * t;
-}
-
-static Sk4f clamp_255(const Sk4f& value) {
- return Sk4f::Min(Sk4f(255), value);
-}
-
-static Sk4f clamp_0_255(const Sk4f& value) {
- return Sk4f::Max(Sk4f(0), Sk4f::Min(Sk4f(255), value));
-}
-
-/**
- * Some modes can, due to very slight numerical error, generate "invalid" pmcolors...
- *
- * e.g.
- * alpha = 100.9999
- * red = 101
- *
- * or
- * alpha = 255.0001
- *
- * If we know we're going to write-out the values as bytes, we can relax these somewhat,
- * since we only really need to enforce that the bytes are valid premul...
- *
- * To that end, this method asserts that the resulting pmcolor will be valid, but does not call
- * SkPMFloat::isValid(), as that would fire sometimes, but not result in a bad pixel.
- */
-static inline SkPMFloat check_as_pmfloat(const Sk4f& value) {
- SkPMFloat pm = value;
-#ifdef SK_DEBUG
- (void)pm.round();
-#endif
- return pm;
-}
-
#define XFERMODE(Name) \
struct Name { \
static Sk4px Xfer(const Sk4px&, const Sk4px&); \
XFERMODE(DstOut) { return SrcOut ::Xfer(d,s); }
XFERMODE(DstOver) { return SrcOver::Xfer(d,s); }
-#undef XFERMODE
-
-// kSrcATop_Mode, //!< [Da, Sc * Da + (1 - Sa) * Dc]
-struct SrcATop4f {
- static SkPMFloat Xfer(const SkPMFloat& src, const SkPMFloat& dst) {
- const Sk4f inv255(gInv255);
- return check_as_pmfloat(dst + (src * Sk4f(dst.a()) - dst * Sk4f(src.a())) * inv255);
- }
- static Sk4px Xfer(const Sk4px& src, const Sk4px& dst) {
- return Sk4px::Wide(src.mulWiden(dst.alphas()) + dst.mulWiden(src.alphas().inv()))
- .div255RoundNarrow();
- }
- static const bool kFoldCoverageIntoSrcAlpha = true;
- static const SkXfermode::Mode kMode = SkXfermode::kSrcATop_Mode;
-};
-
-// kDstATop_Mode, //!< [Sa, Sa * Dc + Sc * (1 - Da)]
-struct DstATop4f {
- static SkPMFloat Xfer(const SkPMFloat& src, const SkPMFloat& dst) {
- return SrcATop4f::Xfer(dst, src);
- }
- static Sk4px Xfer(const Sk4px& src, const Sk4px& dst) {
- return SrcATop4f::Xfer(dst, src);
- }
- static const bool kFoldCoverageIntoSrcAlpha = false;
- static const SkXfermode::Mode kMode = SkXfermode::kDstATop_Mode;
-};
-
-// kXor_Mode [Sa + Da - 2 * Sa * Da, Sc * (1 - Da) + (1 - Sa) * Dc]
-struct Xor4f {
- static SkPMFloat Xfer(const SkPMFloat& src, const SkPMFloat& dst) {
- const Sk4f inv255(gInv255);
- return check_as_pmfloat(src + dst - (src * Sk4f(dst.a()) + dst * Sk4f(src.a())) * inv255);
- }
- static Sk4px Xfer(const Sk4px& src, const Sk4px& dst) {
- return Sk4px::Wide(src.mulWiden(dst.alphas().inv()) + dst.mulWiden(src.alphas().inv()))
- .div255RoundNarrow();
- }
- static const bool kFoldCoverageIntoSrcAlpha = true;
- static const SkXfermode::Mode kMode = SkXfermode::kXor_Mode;
-};
-
-// kPlus_Mode [Sa + Da, Sc + Dc]
-struct Plus4f {
- static SkPMFloat Xfer(const SkPMFloat& src, const SkPMFloat& dst) {
- return check_as_pmfloat(clamp_255(src + dst));
- }
- static Sk4px Xfer(const Sk4px& src, const Sk4px& dst) {
- return src.saturatedAdd(dst);
- }
- static const bool kFoldCoverageIntoSrcAlpha = false;
- static const SkXfermode::Mode kMode = SkXfermode::kPlus_Mode;
-};
-
-// kModulate_Mode [Sa * Da, Sc * Dc]
-struct Modulate4f {
- static SkPMFloat Xfer(const SkPMFloat& src, const SkPMFloat& dst) {
- const Sk4f inv255(gInv255);
- return check_as_pmfloat(src * dst * inv255);
- }
- static Sk4px Xfer(const Sk4px& src, const Sk4px& dst) {
- return src.fastMulDiv255Round(dst);
- }
- static const bool kFoldCoverageIntoSrcAlpha = false;
- static const SkXfermode::Mode kMode = SkXfermode::kModulate_Mode;
-};
-
-// kScreen_Mode [S + D - S * D]
-struct Screen4f {
- static SkPMFloat Xfer(const SkPMFloat& src, const SkPMFloat& dst) {
- const Sk4f inv255(gInv255);
- return check_as_pmfloat(src + dst - src * dst * inv255);
- }
- static Sk4px Xfer(const Sk4px& src, const Sk4px& dst) {
- // Doing the math as S + (1-S)*D or S + (D - S*D) means the add and subtract can be done
- // in 8-bit space without overflow. S + (1-S)*D is a touch faster because inv() is cheap.
- return src + dst.fastMulDiv255Round(src.inv());
- }
- static const bool kFoldCoverageIntoSrcAlpha = true;
- static const SkXfermode::Mode kMode = SkXfermode::kScreen_Mode;
-};
-
-struct Multiply4f {
- static SkPMFloat Xfer(const SkPMFloat& src, const SkPMFloat& dst) {
- const Sk4f inv255(gInv255);
- Sk4f sa = Sk4f(src.a());
- Sk4f da = Sk4f(dst.a());
- Sk4f sc = src;
- Sk4f dc = dst;
- Sk4f rc = sc + dc + (sc * (dc - da) - dc * sa) * inv255;
- // ra = srcover(sa, da), but the calc for rc happens to accomplish this for us
- return check_as_pmfloat(clamp_0_255(rc));
- }
- static Sk4px Xfer(const Sk4px& src, const Sk4px& dst) {
- return Sk4px::Wide(src.mulWiden(dst.alphas().inv()) +
- dst.mulWiden(src.alphas().inv()) +
- src.mulWiden(dst))
- .div255RoundNarrow();
- }
- static const bool kFoldCoverageIntoSrcAlpha = false;
- static const SkXfermode::Mode kMode = SkXfermode::kMultiply_Mode;
-};
-
-// [ sa + da - sa*da, sc + dc - 2*min(sc*da, dc*sa) ] (And notice sa*da == min(sa*da, da*sa).)
-struct Difference4f {
- static SkPMFloat Xfer(const SkPMFloat& src, const SkPMFloat& dst) {
- const Sk4f inv255(gInv255);
- Sk4f sa = Sk4f(src.a());
- Sk4f da = Sk4f(dst.a());
- Sk4f sc = src;
- Sk4f dc = dst;
- Sk4f min = Sk4f::Min(sc * da, dc * sa) * inv255;
- Sk4f ra = sc + dc - min;
- return check_as_pmfloat(ra - min * SkPMFloat(0, 1, 1, 1));
- }
- static Sk4px Xfer(const Sk4px& src, const Sk4px& dst) {
- auto m = Sk4px::Wide(Sk16h::Min(src.mulWiden(dst.alphas()), dst.mulWiden(src.alphas())))
- .div255RoundNarrow();
- // There's no chance of underflow, and if we subtract m before adding src+dst, no overflow.
- return (src - m) + (dst - m.zeroAlphas());
- }
- static const bool kFoldCoverageIntoSrcAlpha = false;
- static const SkXfermode::Mode kMode = SkXfermode::kDifference_Mode;
-};
-
-// [ sa + da - sa*da, sc + dc - 2*sc*dc ]
-struct Exclusion4f {
- static SkPMFloat Xfer(const SkPMFloat& src, const SkPMFloat& dst) {
- const Sk4f inv255(gInv255);
- Sk4f sc = src;
- Sk4f dc = dst;
- Sk4f prod = sc * dc * inv255;
- Sk4f ra = sc + dc - prod;
- return check_as_pmfloat(ra - prod * SkPMFloat(0, 1, 1, 1));
- }
- static Sk4px Xfer(const Sk4px& src, const Sk4px& dst) {
- auto p = src.fastMulDiv255Round(dst);
- // There's no chance of underflow, and if we subtract p before adding src+dst, no overflow.
- return (src - p) + (dst - p.zeroAlphas());
- }
- static const bool kFoldCoverageIntoSrcAlpha = false;
- static const SkXfermode::Mode kMode = SkXfermode::kExclusion_Mode;
-};
-
-template <typename ProcType>
-class SkT4fXfermode : public SkProcCoeffXfermode {
-public:
- static SkXfermode* Create(const ProcCoeff& rec) {
- return SkNEW_ARGS(SkT4fXfermode, (rec));
- }
-
- void xfer32(SkPMColor dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
- if (NULL == aa) {
- for (int i = 0; i < n; ++i) {
- dst[i] = ProcType::Xfer(SkPMFloat(src[i]), SkPMFloat(dst[i])).round();
- }
- } else {
- for (int i = 0; i < n; ++i) {
- const Sk4f aa4 = Sk4f(aa[i] * gInv255);
- SkPMFloat dstF(dst[i]);
- SkPMFloat srcF(src[i]);
- Sk4f res;
- if (ProcType::kFoldCoverageIntoSrcAlpha) {
- Sk4f src4 = srcF;
- res = ProcType::Xfer(src4 * aa4, dstF);
- } else {
- res = ramp(dstF, ProcType::Xfer(srcF, dstF), aa4);
- }
- dst[i] = SkPMFloat(res).round();
- }
- }
- }
-
-private:
- SkT4fXfermode(const ProcCoeff& rec) : SkProcCoeffXfermode(rec, ProcType::kMode) {}
+// [ S * Da + (1 - Sa) * D]
+XFERMODE(SrcATop) {
+ return Sk4px::Wide(s.mulWiden(d.alphas()) + d.mulWiden(s.alphas().inv()))
+ .div255RoundNarrow();
+}
+XFERMODE(DstATop) { return SrcATop::Xfer(d,s); }
+//[ S * (1 - Da) + (1 - Sa) * D ]
+XFERMODE(Xor) {
+ return Sk4px::Wide(s.mulWiden(d.alphas().inv()) + d.mulWiden(s.alphas().inv()))
+ .div255RoundNarrow();
+}
+// [S + D ]
+XFERMODE(Plus) { return s.saturatedAdd(d); }
+// [S * D ]
+XFERMODE(Modulate) { return s.fastMulDiv255Round(d); }
+// [S + D - S * D]
+XFERMODE(Screen) {
+ // Doing the math as S + (1-S)*D or S + (D - S*D) means the add and subtract can be done
+ // in 8-bit space without overflow. S + (1-S)*D is a touch faster because inv() is cheap.
+ return s + d.fastMulDiv255Round(s.inv());
+}
+XFERMODE(Multiply) {
+ return Sk4px::Wide(s.mulWiden(d.alphas().inv()) +
+ d.mulWiden(s.alphas().inv()) +
+ s.mulWiden(d))
+ .div255RoundNarrow();
+}
+// [ Sa + Da - Sa*Da, Sc + Dc - 2*min(Sc*Da, Dc*Sa) ] (And notice Sa*Da == min(Sa*Da, Da*Sa).)
+XFERMODE(Difference) {
+ auto m = Sk4px::Wide(Sk16h::Min(s.mulWiden(d.alphas()), d.mulWiden(s.alphas())))
+ .div255RoundNarrow();
+ // There's no chance of underflow, and if we subtract m before adding s+d, no overflow.
+ return (s - m) + (d - m.zeroAlphas());
+}
+// [ Sa + Da - Sa*Da, Sc + Dc - 2*Sc*Dc ]
+XFERMODE(Exclusion) {
+ auto p = s.fastMulDiv255Round(d);
+ // There's no chance of underflow, and if we subtract p before adding src+dst, no overflow.
+ return (s - p) + (d - p.zeroAlphas());
+}
- typedef SkProcCoeffXfermode INHERITED;
-};
+#undef XFERMODE
template <typename ProcType>
class SkT4pxXfermode : public SkProcCoeffXfermode {
} else {
Sk4px::MapDstSrcAlpha(n, dst, src, aa,
[&](const Sk4px& dst4, const Sk4px& src4, const Sk16b& alpha) {
- // We can't exploit kFoldCoverageIntoSrcAlpha. That requires >=24-bit intermediates.
Sk4px res4 = ProcType::Xfer(src4, dst4);
return Sk4px::Wide(res4.mulWiden(alpha) + dst4.mulWiden(Sk4px(alpha).inv()))
.div255RoundNarrow();
rec.fProc = pp;
}
-#if defined(SK_4PX_XFERMODES_ARE_FAST) && !defined(SK_PREFER_LEGACY_FLOAT_XFERMODES)
+#if defined(SK_USE_4PX_XFERMODES)
switch (mode) {
case SkXfermode::kClear_Mode: return SkT4pxXfermode<Clear>::Create(rec);
case SkXfermode::kSrc_Mode: return SkT4pxXfermode<Src>::Create(rec);
case SkXfermode::kDstIn_Mode: return SkT4pxXfermode<DstIn>::Create(rec);
case SkXfermode::kSrcOut_Mode: return SkT4pxXfermode<SrcOut>::Create(rec);
case SkXfermode::kDstOut_Mode: return SkT4pxXfermode<DstOut>::Create(rec);
-
- case SkXfermode::kSrcATop_Mode: return SkT4pxXfermode<SrcATop4f>::Create(rec);
- case SkXfermode::kDstATop_Mode: return SkT4pxXfermode<DstATop4f>::Create(rec);
- case SkXfermode::kXor_Mode: return SkT4pxXfermode<Xor4f>::Create(rec);
- case SkXfermode::kPlus_Mode: return SkT4pxXfermode<Plus4f>::Create(rec);
- case SkXfermode::kModulate_Mode: return SkT4pxXfermode<Modulate4f>::Create(rec);
- case SkXfermode::kScreen_Mode: return SkT4pxXfermode<Screen4f>::Create(rec);
- case SkXfermode::kMultiply_Mode: return SkT4pxXfermode<Multiply4f>::Create(rec);
- case SkXfermode::kDifference_Mode: return SkT4pxXfermode<Difference4f>::Create(rec);
- case SkXfermode::kExclusion_Mode: return SkT4pxXfermode<Exclusion4f>::Create(rec);
- default: break;
- }
-#endif
-
-#if defined(SK_4F_XFERMODES_ARE_FAST)
- switch (mode) {
- case SkXfermode::kSrcATop_Mode: return SkT4fXfermode<SrcATop4f>::Create(rec);
- case SkXfermode::kDstATop_Mode: return SkT4fXfermode<DstATop4f>::Create(rec);
- case SkXfermode::kXor_Mode: return SkT4fXfermode<Xor4f>::Create(rec);
- case SkXfermode::kPlus_Mode: return SkT4fXfermode<Plus4f>::Create(rec);
- case SkXfermode::kModulate_Mode: return SkT4fXfermode<Modulate4f>::Create(rec);
- case SkXfermode::kScreen_Mode: return SkT4fXfermode<Screen4f>::Create(rec);
- case SkXfermode::kMultiply_Mode: return SkT4fXfermode<Multiply4f>::Create(rec);
- case SkXfermode::kDifference_Mode: return SkT4fXfermode<Difference4f>::Create(rec);
- case SkXfermode::kExclusion_Mode: return SkT4fXfermode<Exclusion4f>::Create(rec);
+ case SkXfermode::kSrcATop_Mode: return SkT4pxXfermode<SrcATop>::Create(rec);
+ case SkXfermode::kDstATop_Mode: return SkT4pxXfermode<DstATop>::Create(rec);
+ case SkXfermode::kXor_Mode: return SkT4pxXfermode<Xor>::Create(rec);
+ case SkXfermode::kPlus_Mode: return SkT4pxXfermode<Plus>::Create(rec);
+ case SkXfermode::kModulate_Mode: return SkT4pxXfermode<Modulate>::Create(rec);
+ case SkXfermode::kScreen_Mode: return SkT4pxXfermode<Screen>::Create(rec);
+ case SkXfermode::kMultiply_Mode: return SkT4pxXfermode<Multiply>::Create(rec);
+ case SkXfermode::kDifference_Mode: return SkT4pxXfermode<Difference>::Create(rec);
+ case SkXfermode::kExclusion_Mode: return SkT4pxXfermode<Exclusion>::Create(rec);
default: break;
}
#endif
return prod;
}
-static inline __m128i saturated_add_SSE2(const __m128i& a, const __m128i& b) {
- __m128i sum = _mm_add_epi32(a, b);
- __m128i cmp = _mm_cmpgt_epi32(sum, _mm_set1_epi32(255));
-
- sum = _mm_or_si128(_mm_and_si128(cmp, _mm_set1_epi32(255)),
- _mm_andnot_si128(cmp, sum));
- return sum;
-}
-
-static inline __m128i clamp_signed_byte_SSE2(const __m128i& n) {
- __m128i cmp1 = _mm_cmplt_epi32(n, _mm_setzero_si128());
- __m128i cmp2 = _mm_cmpgt_epi32(n, _mm_set1_epi32(255));
- __m128i ret = _mm_and_si128(cmp2, _mm_set1_epi32(255));
-
- __m128i cmp = _mm_or_si128(cmp1, cmp2);
- ret = _mm_or_si128(_mm_and_si128(cmp, ret), _mm_andnot_si128(cmp, n));
-
- return ret;
-}
-
static inline __m128i clamp_div255round_SSE2(const __m128i& prod) {
// test if > 0
__m128i cmp1 = _mm_cmpgt_epi32(prod, _mm_setzero_si128());
return ret;
}
-
-static __m128i srcover_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i isa = _mm_sub_epi32(_mm_set1_epi32(256), SkGetPackedA32_SSE2(src));
- return _mm_add_epi32(src, SkAlphaMulQ_SSE2(dst, isa));
-}
-
-static __m128i dstover_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i ida = _mm_sub_epi32(_mm_set1_epi32(256), SkGetPackedA32_SSE2(dst));
- return _mm_add_epi32(dst, SkAlphaMulQ_SSE2(src, ida));
-}
-
-static __m128i srcin_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i da = SkGetPackedA32_SSE2(dst);
- return SkAlphaMulQ_SSE2(src, SkAlpha255To256_SSE2(da));
-}
-
-static __m128i dstin_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i sa = SkGetPackedA32_SSE2(src);
- return SkAlphaMulQ_SSE2(dst, SkAlpha255To256_SSE2(sa));
-}
-
-static __m128i srcout_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i ida = _mm_sub_epi32(_mm_set1_epi32(256), SkGetPackedA32_SSE2(dst));
- return SkAlphaMulQ_SSE2(src, ida);
-}
-
-static __m128i dstout_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i isa = _mm_sub_epi32(_mm_set1_epi32(256), SkGetPackedA32_SSE2(src));
- return SkAlphaMulQ_SSE2(dst, isa);
-}
-
-static __m128i srcatop_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i sa = SkGetPackedA32_SSE2(src);
- __m128i da = SkGetPackedA32_SSE2(dst);
- __m128i isa = _mm_sub_epi32(_mm_set1_epi32(255), sa);
-
- __m128i a = da;
-
- __m128i r1 = SkAlphaMulAlpha_SSE2(da, SkGetPackedR32_SSE2(src));
- __m128i r2 = SkAlphaMulAlpha_SSE2(isa, SkGetPackedR32_SSE2(dst));
- __m128i r = _mm_add_epi32(r1, r2);
-
- __m128i g1 = SkAlphaMulAlpha_SSE2(da, SkGetPackedG32_SSE2(src));
- __m128i g2 = SkAlphaMulAlpha_SSE2(isa, SkGetPackedG32_SSE2(dst));
- __m128i g = _mm_add_epi32(g1, g2);
-
- __m128i b1 = SkAlphaMulAlpha_SSE2(da, SkGetPackedB32_SSE2(src));
- __m128i b2 = SkAlphaMulAlpha_SSE2(isa, SkGetPackedB32_SSE2(dst));
- __m128i b = _mm_add_epi32(b1, b2);
-
- return SkPackARGB32_SSE2(a, r, g, b);
-}
-
-static __m128i dstatop_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i sa = SkGetPackedA32_SSE2(src);
- __m128i da = SkGetPackedA32_SSE2(dst);
- __m128i ida = _mm_sub_epi32(_mm_set1_epi32(255), da);
-
- __m128i a = sa;
-
- __m128i r1 = SkAlphaMulAlpha_SSE2(ida, SkGetPackedR32_SSE2(src));
- __m128i r2 = SkAlphaMulAlpha_SSE2(sa, SkGetPackedR32_SSE2(dst));
- __m128i r = _mm_add_epi32(r1, r2);
-
- __m128i g1 = SkAlphaMulAlpha_SSE2(ida, SkGetPackedG32_SSE2(src));
- __m128i g2 = SkAlphaMulAlpha_SSE2(sa, SkGetPackedG32_SSE2(dst));
- __m128i g = _mm_add_epi32(g1, g2);
-
- __m128i b1 = SkAlphaMulAlpha_SSE2(ida, SkGetPackedB32_SSE2(src));
- __m128i b2 = SkAlphaMulAlpha_SSE2(sa, SkGetPackedB32_SSE2(dst));
- __m128i b = _mm_add_epi32(b1, b2);
-
- return SkPackARGB32_SSE2(a, r, g, b);
-}
-
-static __m128i xor_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i sa = SkGetPackedA32_SSE2(src);
- __m128i da = SkGetPackedA32_SSE2(dst);
- __m128i isa = _mm_sub_epi32(_mm_set1_epi32(255), sa);
- __m128i ida = _mm_sub_epi32(_mm_set1_epi32(255), da);
-
- __m128i a1 = _mm_add_epi32(sa, da);
- __m128i a2 = SkAlphaMulAlpha_SSE2(sa, da);
- a2 = _mm_slli_epi32(a2, 1);
- __m128i a = _mm_sub_epi32(a1, a2);
-
- __m128i r1 = SkAlphaMulAlpha_SSE2(ida, SkGetPackedR32_SSE2(src));
- __m128i r2 = SkAlphaMulAlpha_SSE2(isa, SkGetPackedR32_SSE2(dst));
- __m128i r = _mm_add_epi32(r1, r2);
-
- __m128i g1 = SkAlphaMulAlpha_SSE2(ida, SkGetPackedG32_SSE2(src));
- __m128i g2 = SkAlphaMulAlpha_SSE2(isa, SkGetPackedG32_SSE2(dst));
- __m128i g = _mm_add_epi32(g1, g2);
-
- __m128i b1 = SkAlphaMulAlpha_SSE2(ida, SkGetPackedB32_SSE2(src));
- __m128i b2 = SkAlphaMulAlpha_SSE2(isa, SkGetPackedB32_SSE2(dst));
- __m128i b = _mm_add_epi32(b1, b2);
-
- return SkPackARGB32_SSE2(a, r, g, b);
-}
-
-static __m128i plus_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i b = saturated_add_SSE2(SkGetPackedB32_SSE2(src),
- SkGetPackedB32_SSE2(dst));
- __m128i g = saturated_add_SSE2(SkGetPackedG32_SSE2(src),
- SkGetPackedG32_SSE2(dst));
- __m128i r = saturated_add_SSE2(SkGetPackedR32_SSE2(src),
- SkGetPackedR32_SSE2(dst));
- __m128i a = saturated_add_SSE2(SkGetPackedA32_SSE2(src),
- SkGetPackedA32_SSE2(dst));
- return SkPackARGB32_SSE2(a, r, g, b);
-}
-
-static __m128i modulate_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i a = SkAlphaMulAlpha_SSE2(SkGetPackedA32_SSE2(src),
- SkGetPackedA32_SSE2(dst));
- __m128i r = SkAlphaMulAlpha_SSE2(SkGetPackedR32_SSE2(src),
- SkGetPackedR32_SSE2(dst));
- __m128i g = SkAlphaMulAlpha_SSE2(SkGetPackedG32_SSE2(src),
- SkGetPackedG32_SSE2(dst));
- __m128i b = SkAlphaMulAlpha_SSE2(SkGetPackedB32_SSE2(src),
- SkGetPackedB32_SSE2(dst));
- return SkPackARGB32_SSE2(a, r, g, b);
-}
-
static inline __m128i SkMin32_SSE2(const __m128i& a, const __m128i& b) {
__m128i cmp = _mm_cmplt_epi32(a, b);
return _mm_or_si128(_mm_and_si128(cmp, a), _mm_andnot_si128(cmp, b));
}
-static inline __m128i blendfunc_multiply_byte_SSE2(const __m128i& sc, const __m128i& dc,
- const __m128i& sa, const __m128i& da) {
- // sc * (255 - da)
- __m128i ret1 = _mm_sub_epi32(_mm_set1_epi32(255), da);
- ret1 = _mm_mullo_epi16(sc, ret1);
-
- // dc * (255 - sa)
- __m128i ret2 = _mm_sub_epi32(_mm_set1_epi32(255), sa);
- ret2 = _mm_mullo_epi16(dc, ret2);
-
- // sc * dc
- __m128i ret3 = _mm_mullo_epi16(sc, dc);
-
- __m128i ret = _mm_add_epi32(ret1, ret2);
- ret = _mm_add_epi32(ret, ret3);
-
- return clamp_div255round_SSE2(ret);
-}
-
-static __m128i multiply_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i sa = SkGetPackedA32_SSE2(src);
- __m128i da = SkGetPackedA32_SSE2(dst);
- __m128i a = srcover_byte_SSE2(sa, da);
-
- __m128i sr = SkGetPackedR32_SSE2(src);
- __m128i dr = SkGetPackedR32_SSE2(dst);
- __m128i r = blendfunc_multiply_byte_SSE2(sr, dr, sa, da);
-
- __m128i sg = SkGetPackedG32_SSE2(src);
- __m128i dg = SkGetPackedG32_SSE2(dst);
- __m128i g = blendfunc_multiply_byte_SSE2(sg, dg, sa, da);
-
-
- __m128i sb = SkGetPackedB32_SSE2(src);
- __m128i db = SkGetPackedB32_SSE2(dst);
- __m128i b = blendfunc_multiply_byte_SSE2(sb, db, sa, da);
-
- return SkPackARGB32_SSE2(a, r, g, b);
-}
-
-static __m128i screen_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i a = srcover_byte_SSE2(SkGetPackedA32_SSE2(src),
- SkGetPackedA32_SSE2(dst));
- __m128i r = srcover_byte_SSE2(SkGetPackedR32_SSE2(src),
- SkGetPackedR32_SSE2(dst));
- __m128i g = srcover_byte_SSE2(SkGetPackedG32_SSE2(src),
- SkGetPackedG32_SSE2(dst));
- __m128i b = srcover_byte_SSE2(SkGetPackedB32_SSE2(src),
- SkGetPackedB32_SSE2(dst));
- return SkPackARGB32_SSE2(a, r, g, b);
-}
-
// Portable version overlay_byte() is in SkXfermode.cpp.
static inline __m128i overlay_byte_SSE2(const __m128i& sc, const __m128i& dc,
const __m128i& sa, const __m128i& da) {
return SkPackARGB32_SSE2(a, r, g, b);
}
-static inline __m128i difference_byte_SSE2(const __m128i& sc, const __m128i& dc,
- const __m128i& sa, const __m128i& da) {
- __m128i tmp1 = _mm_mullo_epi16(sc, da);
- __m128i tmp2 = _mm_mullo_epi16(dc, sa);
- __m128i tmp = SkMin32_SSE2(tmp1, tmp2);
-
- __m128i ret1 = _mm_add_epi32(sc, dc);
- __m128i ret2 = _mm_slli_epi32(SkDiv255Round_SSE2(tmp), 1);
- __m128i ret = _mm_sub_epi32(ret1, ret2);
-
- ret = clamp_signed_byte_SSE2(ret);
- return ret;
-}
-
-static __m128i difference_modeproc_SSE2(const __m128i& src,
- const __m128i& dst) {
- __m128i sa = SkGetPackedA32_SSE2(src);
- __m128i da = SkGetPackedA32_SSE2(dst);
-
- __m128i a = srcover_byte_SSE2(sa, da);
- __m128i r = difference_byte_SSE2(SkGetPackedR32_SSE2(src),
- SkGetPackedR32_SSE2(dst), sa, da);
- __m128i g = difference_byte_SSE2(SkGetPackedG32_SSE2(src),
- SkGetPackedG32_SSE2(dst), sa, da);
- __m128i b = difference_byte_SSE2(SkGetPackedB32_SSE2(src),
- SkGetPackedB32_SSE2(dst), sa, da);
- return SkPackARGB32_SSE2(a, r, g, b);
-}
-
-static inline __m128i exclusion_byte_SSE2(const __m128i& sc, const __m128i& dc,
- const __m128i&, __m128i&) {
- __m128i tmp1 = _mm_mullo_epi16(_mm_set1_epi32(255), sc); // 255 * sc
- __m128i tmp2 = _mm_mullo_epi16(_mm_set1_epi32(255), dc); // 255 * dc
- tmp1 = _mm_add_epi32(tmp1, tmp2);
- tmp2 = _mm_mullo_epi16(sc, dc); // sc * dc
- tmp2 = _mm_slli_epi32(tmp2, 1); // 2 * sc * dc
-
- __m128i r = _mm_sub_epi32(tmp1, tmp2);
- return clamp_div255round_SSE2(r);
-}
-
-static __m128i exclusion_modeproc_SSE2(const __m128i& src, const __m128i& dst) {
- __m128i sa = SkGetPackedA32_SSE2(src);
- __m128i da = SkGetPackedA32_SSE2(dst);
-
- __m128i a = srcover_byte_SSE2(sa, da);
- __m128i r = exclusion_byte_SSE2(SkGetPackedR32_SSE2(src),
- SkGetPackedR32_SSE2(dst), sa, da);
- __m128i g = exclusion_byte_SSE2(SkGetPackedG32_SSE2(src),
- SkGetPackedG32_SSE2(dst), sa, da);
- __m128i b = exclusion_byte_SSE2(SkGetPackedB32_SSE2(src),
- SkGetPackedB32_SSE2(dst), sa, da);
- return SkPackARGB32_SSE2(a, r, g, b);
-}
////////////////////////////////////////////////////////////////////////////////
typedef __m128i (*SkXfermodeProcSIMD)(const __m128i& src, const __m128i& dst);
-extern SkXfermodeProcSIMD gSSE2XfermodeProcs[];
-
void SkSSE2ProcCoeffXfermode::xfer32(SkPMColor dst[], const SkPMColor src[],
int count, const SkAlpha aa[]) const {
SkASSERT(dst && src && count >= 0);
}
#endif
-////////////////////////////////////////////////////////////////////////////////
-
-// 4 pixels modeprocs with SSE2
-SkXfermodeProcSIMD gSSE2XfermodeProcs[] = {
- NULL, // kClear_Mode
- NULL, // kSrc_Mode
- NULL, // kDst_Mode
- srcover_modeproc_SSE2,
- dstover_modeproc_SSE2,
- srcin_modeproc_SSE2,
- dstin_modeproc_SSE2,
- srcout_modeproc_SSE2,
- dstout_modeproc_SSE2,
- srcatop_modeproc_SSE2,
- dstatop_modeproc_SSE2,
- xor_modeproc_SSE2,
- plus_modeproc_SSE2,
- modulate_modeproc_SSE2,
- screen_modeproc_SSE2,
-
- overlay_modeproc_SSE2,
- darken_modeproc_SSE2,
- lighten_modeproc_SSE2,
- colordodge_modeproc_SSE2,
- colorburn_modeproc_SSE2,
- hardlight_modeproc_SSE2,
- softlight_modeproc_SSE2,
- difference_modeproc_SSE2,
- exclusion_modeproc_SSE2,
- multiply_modeproc_SSE2,
-
- NULL, // kHue_Mode
- NULL, // kSaturation_Mode
- NULL, // kColor_Mode
- NULL, // kLuminosity_Mode
-};
-
SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl_SSE2(const ProcCoeff& rec,
SkXfermode::Mode mode) {
- void* procSIMD = reinterpret_cast<void*>(gSSE2XfermodeProcs[mode]);
-
- if (procSIMD != NULL) {
- return SkNEW_ARGS(SkSSE2ProcCoeffXfermode, (rec, mode, procSIMD));
+ SkXfermodeProcSIMD proc = nullptr;
+ // TODO(mtklein): implement these Sk4px.
+ switch (mode) {
+ case SkProcCoeffXfermode::kOverlay_Mode: proc = overlay_modeproc_SSE2; break;
+ case SkProcCoeffXfermode::kDarken_Mode: proc = darken_modeproc_SSE2; break;
+ case SkProcCoeffXfermode::kLighten_Mode: proc = lighten_modeproc_SSE2; break;
+ case SkProcCoeffXfermode::kColorDodge_Mode: proc = colordodge_modeproc_SSE2; break;
+ case SkProcCoeffXfermode::kColorBurn_Mode: proc = colorburn_modeproc_SSE2; break;
+ case SkProcCoeffXfermode::kHardLight_Mode: proc = hardlight_modeproc_SSE2; break;
+ case SkProcCoeffXfermode::kSoftLight_Mode: proc = softlight_modeproc_SSE2; break;
+ default: break;
}
- return NULL;
+ return proc ? SkNEW_ARGS(SkSSE2ProcCoeffXfermode, (rec, mode, (void*)proc)) : nullptr;
}
projects / platform / upstream / libSkiaSharp.git / commitdiff