Revert "use pipeline for non-opt xfermodes"

This reverts commit e93cf97175d348ce1400762bdd8d9acabdd29766.

Reason for revert: unblock Chrome roll

Original change's description:
> use pipeline for non-opt xfermodes
> 
> Produces slightly different results for Hue, Saturation, Color, Luminosity,
> seemingly around the aa edging.
> 
> Bug: skia:
> Change-Id: I6364818c9788863e5fad6d14cad4797d073dbea3
> Reviewed-on: https://skia-review.googlesource.com/19554
> Reviewed-by: Mike Klein <mtklein@google.com>
> Commit-Queue: Mike Reed <reed@google.com>

TBR=mtklein@google.com,fmalita@chromium.org,reed@google.com

Change-Id: Ib85a51753d21ce778fa5cfedd6b7d1b5b2b87096
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: skia:
Reviewed-on: https://skia-review.googlesource.com/19745
Reviewed-by: Mike Klein <mtklein@chromium.org>
Commit-Queue: Mike Klein <mtklein@chromium.org>

1 files changed, 662 insertions, 34 deletions

diff --git a/src/core/SkXfermode.cpp b/src/core/SkXfermode.cpp
index 54ac5891a3..0dd9e87cdf 100644
--- a/src/core/SkXfermode.cpp
+++ b/src/core/SkXfermode.cpp
@@ -6,16 +6,16 @@
  */
 
 #include "SkBlendModePriv.h"
+#include "SkXfermode_proccoeff.h"
 #include "SkColorPriv.h"
 #include "SkMathPriv.h"
 #include "SkOnce.h"
 #include "SkOpts.h"
-#include "SkPM4f.h"
 #include "SkRasterPipeline.h"
 #include "SkReadBuffer.h"
 #include "SkString.h"
 #include "SkWriteBuffer.h"
-#include "SkXfermodePriv.h"
+#include "SkPM4f.h"
 
 #if SK_SUPPORT_GPU
 #include "GrFragmentProcessor.h"
@@ -24,50 +24,668 @@
 #include "effects/GrXfermodeFragmentProcessor.h"
 #endif
 
-///////////////////////////////////////////////////////////////////////////////////////////////////
+#define SkAlphaMulAlpha(a, b)   SkMulDiv255Round(a, b)
 
-class SkProcCoeffXfermode : public SkXfermode {
-public:
-    SkProcCoeffXfermode(SkBlendMode mode) : fMode(mode) {}
+static inline unsigned saturated_add(unsigned a, unsigned b) {
+    SkASSERT(a <= 255);
+    SkASSERT(b <= 255);
+    unsigned sum = a + b;
+    if (sum > 255) {
+        sum = 255;
+    }
+    return sum;
+}
 
-    void xfer32(SkPMColor dst[], const SkPMColor src[], int count,
-                const SkAlpha aa[]) const override {
-        SkASSERT(dst && src && count >= 0);
+static inline int clamp_signed_byte(int n) {
+    if (n < 0) {
+        n = 0;
+    } else if (n > 255) {
+        n = 255;
+    }
+    return n;
+}
 
-        bool needs_swap = (kN32_SkColorType == kBGRA_8888_SkColorType) &&
-                          SkBlendMode_CaresAboutRBOrder(fMode);
+static inline int clamp_div255round(int prod) {
+    if (prod <= 0) {
+        return 0;
+    } else if (prod >= 255*255) {
+        return 255;
+    } else {
+        return SkDiv255Round(prod);
+    }
+}
 
-        SkRasterPipeline_<256> p;
+///////////////////////////////////////////////////////////////////////////////
 
-        p.append(SkRasterPipeline::load_8888, &dst);
-        if (needs_swap) {
-            p.append(SkRasterPipeline::swap_rb);
-        }
-        p.append(SkRasterPipeline::move_src_dst);
-        p.append(SkRasterPipeline::load_8888, &src);
-        if (needs_swap) {
-            p.append(SkRasterPipeline::swap_rb);
-        }
+//  kClear_Mode,    //!< [0, 0]
+static SkPMColor clear_modeproc(SkPMColor src, SkPMColor dst) {
+    return 0;
+}
 
-        SkBlendMode_AppendStagesNoClamp(fMode, &p);
-        if (aa) {
-            p.append(SkRasterPipeline::lerp_u8, &aa);
-        }
-        SkBlendMode_AppendClampIfNeeded(fMode, &p);
+//  kSrc_Mode,      //!< [Sa, Sc]
+static SkPMColor src_modeproc(SkPMColor src, SkPMColor dst) {
+    return src;
+}
+
+//  kDst_Mode,      //!< [Da, Dc]
+static SkPMColor dst_modeproc(SkPMColor src, SkPMColor dst) {
+    return dst;
+}
+
+//  kSrcOver_Mode,  //!< [Sa + Da - Sa*Da, Sc + (1 - Sa)*Dc]
+static SkPMColor srcover_modeproc(SkPMColor src, SkPMColor dst) {
+#if 0
+    // this is the old, more-correct way, but it doesn't guarantee that dst==255
+    // will always stay opaque
+    return src + SkAlphaMulQ(dst, SkAlpha255To256(255 - SkGetPackedA32(src)));
+#else
+    // this is slightly faster, but more importantly guarantees that dst==255
+    // will always stay opaque
+    return src + SkAlphaMulQ(dst, 256 - SkGetPackedA32(src));
+#endif
+}
 
-        if (needs_swap) {
-            p.append(SkRasterPipeline::swap_rb);
+//  kDstOver_Mode,  //!< [Sa + Da - Sa*Da, Dc + (1 - Da)*Sc]
+static SkPMColor dstover_modeproc(SkPMColor src, SkPMColor dst) {
+    // this is the reverse of srcover, just flipping src and dst
+    // see srcover's comment about the 256 for opaqueness guarantees
+    return dst + SkAlphaMulQ(src, 256 - SkGetPackedA32(dst));
+}
+
+//  kSrcIn_Mode,    //!< [Sa * Da, Sc * Da]
+static SkPMColor srcin_modeproc(SkPMColor src, SkPMColor dst) {
+    return SkAlphaMulQ(src, SkAlpha255To256(SkGetPackedA32(dst)));
+}
+
+//  kDstIn_Mode,    //!< [Sa * Da, Sa * Dc]
+static SkPMColor dstin_modeproc(SkPMColor src, SkPMColor dst) {
+    return SkAlphaMulQ(dst, SkAlpha255To256(SkGetPackedA32(src)));
+}
+
+//  kSrcOut_Mode,   //!< [Sa * (1 - Da), Sc * (1 - Da)]
+static SkPMColor srcout_modeproc(SkPMColor src, SkPMColor dst) {
+    return SkAlphaMulQ(src, SkAlpha255To256(255 - SkGetPackedA32(dst)));
+}
+
+//  kDstOut_Mode,   //!< [Da * (1 - Sa), Dc * (1 - Sa)]
+static SkPMColor dstout_modeproc(SkPMColor src, SkPMColor dst) {
+    return SkAlphaMulQ(dst, SkAlpha255To256(255 - SkGetPackedA32(src)));
+}
+
+//  kSrcATop_Mode,  //!< [Da, Sc * Da + (1 - Sa) * Dc]
+static SkPMColor srcatop_modeproc(SkPMColor src, SkPMColor dst) {
+    unsigned sa = SkGetPackedA32(src);
+    unsigned da = SkGetPackedA32(dst);
+    unsigned isa = 255 - sa;
+
+    return SkPackARGB32(da,
+                        SkAlphaMulAlpha(da, SkGetPackedR32(src)) +
+                            SkAlphaMulAlpha(isa, SkGetPackedR32(dst)),
+                        SkAlphaMulAlpha(da, SkGetPackedG32(src)) +
+                            SkAlphaMulAlpha(isa, SkGetPackedG32(dst)),
+                        SkAlphaMulAlpha(da, SkGetPackedB32(src)) +
+                            SkAlphaMulAlpha(isa, SkGetPackedB32(dst)));
+}
+
+//  kDstATop_Mode,  //!< [Sa, Sa * Dc + Sc * (1 - Da)]
+static SkPMColor dstatop_modeproc(SkPMColor src, SkPMColor dst) {
+    unsigned sa = SkGetPackedA32(src);
+    unsigned da = SkGetPackedA32(dst);
+    unsigned ida = 255 - da;
+
+    return SkPackARGB32(sa,
+                        SkAlphaMulAlpha(ida, SkGetPackedR32(src)) +
+                            SkAlphaMulAlpha(sa, SkGetPackedR32(dst)),
+                        SkAlphaMulAlpha(ida, SkGetPackedG32(src)) +
+                            SkAlphaMulAlpha(sa, SkGetPackedG32(dst)),
+                        SkAlphaMulAlpha(ida, SkGetPackedB32(src)) +
+                            SkAlphaMulAlpha(sa, SkGetPackedB32(dst)));
+}
+
+//  kXor_Mode   [Sa + Da - 2 * Sa * Da, Sc * (1 - Da) + (1 - Sa) * Dc]
+static SkPMColor xor_modeproc(SkPMColor src, SkPMColor dst) {
+    unsigned sa = SkGetPackedA32(src);
+    unsigned da = SkGetPackedA32(dst);
+    unsigned isa = 255 - sa;
+    unsigned ida = 255 - da;
+
+    return SkPackARGB32(sa + da - (SkAlphaMulAlpha(sa, da) << 1),
+                        SkAlphaMulAlpha(ida, SkGetPackedR32(src)) +
+                            SkAlphaMulAlpha(isa, SkGetPackedR32(dst)),
+                        SkAlphaMulAlpha(ida, SkGetPackedG32(src)) +
+                            SkAlphaMulAlpha(isa, SkGetPackedG32(dst)),
+                        SkAlphaMulAlpha(ida, SkGetPackedB32(src)) +
+                            SkAlphaMulAlpha(isa, SkGetPackedB32(dst)));
+}
+
+///////////////////////////////////////////////////////////////////////////////
+
+// kPlus_Mode
+static SkPMColor plus_modeproc(SkPMColor src, SkPMColor dst) {
+    unsigned b = saturated_add(SkGetPackedB32(src), SkGetPackedB32(dst));
+    unsigned g = saturated_add(SkGetPackedG32(src), SkGetPackedG32(dst));
+    unsigned r = saturated_add(SkGetPackedR32(src), SkGetPackedR32(dst));
+    unsigned a = saturated_add(SkGetPackedA32(src), SkGetPackedA32(dst));
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kModulate_Mode
+static SkPMColor modulate_modeproc(SkPMColor src, SkPMColor dst) {
+    int a = SkAlphaMulAlpha(SkGetPackedA32(src), SkGetPackedA32(dst));
+    int r = SkAlphaMulAlpha(SkGetPackedR32(src), SkGetPackedR32(dst));
+    int g = SkAlphaMulAlpha(SkGetPackedG32(src), SkGetPackedG32(dst));
+    int b = SkAlphaMulAlpha(SkGetPackedB32(src), SkGetPackedB32(dst));
+    return SkPackARGB32(a, r, g, b);
+}
+
+static inline int srcover_byte(int a, int b) {
+    return a + b - SkAlphaMulAlpha(a, b);
+}
+
+// kMultiply_Mode
+// B(Cb, Cs) = Cb x Cs
+// multiply uses its own version of blendfunc_byte because sa and da are not needed
+static int blendfunc_multiply_byte(int sc, int dc, int sa, int da) {
+    return clamp_div255round(sc * (255 - da)  + dc * (255 - sa)  + sc * dc);
+}
+
+static SkPMColor multiply_modeproc(SkPMColor src, SkPMColor dst) {
+    int sa = SkGetPackedA32(src);
+    int da = SkGetPackedA32(dst);
+    int a = srcover_byte(sa, da);
+    int r = blendfunc_multiply_byte(SkGetPackedR32(src), SkGetPackedR32(dst), sa, da);
+    int g = blendfunc_multiply_byte(SkGetPackedG32(src), SkGetPackedG32(dst), sa, da);
+    int b = blendfunc_multiply_byte(SkGetPackedB32(src), SkGetPackedB32(dst), sa, da);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kScreen_Mode
+static SkPMColor screen_modeproc(SkPMColor src, SkPMColor dst) {
+    int a = srcover_byte(SkGetPackedA32(src), SkGetPackedA32(dst));
+    int r = srcover_byte(SkGetPackedR32(src), SkGetPackedR32(dst));
+    int g = srcover_byte(SkGetPackedG32(src), SkGetPackedG32(dst));
+    int b = srcover_byte(SkGetPackedB32(src), SkGetPackedB32(dst));
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kOverlay_Mode
+static inline int overlay_byte(int sc, int dc, int sa, int da) {
+    int tmp = sc * (255 - da) + dc * (255 - sa);
+    int rc;
+    if (2 * dc <= da) {
+        rc = 2 * sc * dc;
+    } else {
+        rc = sa * da - 2 * (da - dc) * (sa - sc);
+    }
+    return clamp_div255round(rc + tmp);
+}
+static SkPMColor overlay_modeproc(SkPMColor src, SkPMColor dst) {
+    int sa = SkGetPackedA32(src);
+    int da = SkGetPackedA32(dst);
+    int a = srcover_byte(sa, da);
+    int r = overlay_byte(SkGetPackedR32(src), SkGetPackedR32(dst), sa, da);
+    int g = overlay_byte(SkGetPackedG32(src), SkGetPackedG32(dst), sa, da);
+    int b = overlay_byte(SkGetPackedB32(src), SkGetPackedB32(dst), sa, da);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kDarken_Mode
+static inline int darken_byte(int sc, int dc, int sa, int da) {
+    int sd = sc * da;
+    int ds = dc * sa;
+    if (sd < ds) {
+        // srcover
+        return sc + dc - SkDiv255Round(ds);
+    } else {
+        // dstover
+        return dc + sc - SkDiv255Round(sd);
+    }
+}
+static SkPMColor darken_modeproc(SkPMColor src, SkPMColor dst) {
+    int sa = SkGetPackedA32(src);
+    int da = SkGetPackedA32(dst);
+    int a = srcover_byte(sa, da);
+    int r = darken_byte(SkGetPackedR32(src), SkGetPackedR32(dst), sa, da);
+    int g = darken_byte(SkGetPackedG32(src), SkGetPackedG32(dst), sa, da);
+    int b = darken_byte(SkGetPackedB32(src), SkGetPackedB32(dst), sa, da);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kLighten_Mode
+static inline int lighten_byte(int sc, int dc, int sa, int da) {
+    int sd = sc * da;
+    int ds = dc * sa;
+    if (sd > ds) {
+        // srcover
+        return sc + dc - SkDiv255Round(ds);
+    } else {
+        // dstover
+        return dc + sc - SkDiv255Round(sd);
+    }
+}
+static SkPMColor lighten_modeproc(SkPMColor src, SkPMColor dst) {
+    int sa = SkGetPackedA32(src);
+    int da = SkGetPackedA32(dst);
+    int a = srcover_byte(sa, da);
+    int r = lighten_byte(SkGetPackedR32(src), SkGetPackedR32(dst), sa, da);
+    int g = lighten_byte(SkGetPackedG32(src), SkGetPackedG32(dst), sa, da);
+    int b = lighten_byte(SkGetPackedB32(src), SkGetPackedB32(dst), sa, da);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kColorDodge_Mode
+static inline int colordodge_byte(int sc, int dc, int sa, int da) {
+    int diff = sa - sc;
+    int rc;
+    if (0 == dc) {
+        return SkAlphaMulAlpha(sc, 255 - da);
+    } else if (0 == diff) {
+        rc = sa * da + sc * (255 - da) + dc * (255 - sa);
+    } else {
+        diff = dc * sa / diff;
+        rc = sa * ((da < diff) ? da : diff) + sc * (255 - da) + dc * (255 - sa);
+    }
+    return clamp_div255round(rc);
+}
+static SkPMColor colordodge_modeproc(SkPMColor src, SkPMColor dst) {
+    int sa = SkGetPackedA32(src);
+    int da = SkGetPackedA32(dst);
+    int a = srcover_byte(sa, da);
+    int r = colordodge_byte(SkGetPackedR32(src), SkGetPackedR32(dst), sa, da);
+    int g = colordodge_byte(SkGetPackedG32(src), SkGetPackedG32(dst), sa, da);
+    int b = colordodge_byte(SkGetPackedB32(src), SkGetPackedB32(dst), sa, da);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kColorBurn_Mode
+static inline int colorburn_byte(int sc, int dc, int sa, int da) {
+    int rc;
+    if (dc == da) {
+        rc = sa * da + sc * (255 - da) + dc * (255 - sa);
+    } else if (0 == sc) {
+        return SkAlphaMulAlpha(dc, 255 - sa);
+    } else {
+        int tmp = (da - dc) * sa / sc;
+        rc = sa * (da - ((da < tmp) ? da : tmp))
+            + sc * (255 - da) + dc * (255 - sa);
+    }
+    return clamp_div255round(rc);
+}
+static SkPMColor colorburn_modeproc(SkPMColor src, SkPMColor dst) {
+    int sa = SkGetPackedA32(src);
+    int da = SkGetPackedA32(dst);
+    int a = srcover_byte(sa, da);
+    int r = colorburn_byte(SkGetPackedR32(src), SkGetPackedR32(dst), sa, da);
+    int g = colorburn_byte(SkGetPackedG32(src), SkGetPackedG32(dst), sa, da);
+    int b = colorburn_byte(SkGetPackedB32(src), SkGetPackedB32(dst), sa, da);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kHardLight_Mode
+static inline int hardlight_byte(int sc, int dc, int sa, int da) {
+    int rc;
+    if (2 * sc <= sa) {
+        rc = 2 * sc * dc;
+    } else {
+        rc = sa * da - 2 * (da - dc) * (sa - sc);
+    }
+    return clamp_div255round(rc + sc * (255 - da) + dc * (255 - sa));
+}
+static SkPMColor hardlight_modeproc(SkPMColor src, SkPMColor dst) {
+    int sa = SkGetPackedA32(src);
+    int da = SkGetPackedA32(dst);
+    int a = srcover_byte(sa, da);
+    int r = hardlight_byte(SkGetPackedR32(src), SkGetPackedR32(dst), sa, da);
+    int g = hardlight_byte(SkGetPackedG32(src), SkGetPackedG32(dst), sa, da);
+    int b = hardlight_byte(SkGetPackedB32(src), SkGetPackedB32(dst), sa, da);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// returns 255 * sqrt(n/255)
+static U8CPU sqrt_unit_byte(U8CPU n) {
+    return SkSqrtBits(n, 15+4);
+}
+
+// kSoftLight_Mode
+static inline int softlight_byte(int sc, int dc, int sa, int da) {
+    int m = da ? dc * 256 / da : 0;
+    int rc;
+    if (2 * sc <= sa) {
+        rc = dc * (sa + ((2 * sc - sa) * (256 - m) >> 8));
+    } else if (4 * dc <= da) {
+        int tmp = (4 * m * (4 * m + 256) * (m - 256) >> 16) + 7 * m;
+        rc = dc * sa + (da * (2 * sc - sa) * tmp >> 8);
+    } else {
+        int tmp = sqrt_unit_byte(m) - m;
+        rc = dc * sa + (da * (2 * sc - sa) * tmp >> 8);
+    }
+    return clamp_div255round(rc + sc * (255 - da) + dc * (255 - sa));
+}
+static SkPMColor softlight_modeproc(SkPMColor src, SkPMColor dst) {
+    int sa = SkGetPackedA32(src);
+    int da = SkGetPackedA32(dst);
+    int a = srcover_byte(sa, da);
+    int r = softlight_byte(SkGetPackedR32(src), SkGetPackedR32(dst), sa, da);
+    int g = softlight_byte(SkGetPackedG32(src), SkGetPackedG32(dst), sa, da);
+    int b = softlight_byte(SkGetPackedB32(src), SkGetPackedB32(dst), sa, da);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kDifference_Mode
+static inline int difference_byte(int sc, int dc, int sa, int da) {
+    int tmp = SkMin32(sc * da, dc * sa);
+    return clamp_signed_byte(sc + dc - 2 * SkDiv255Round(tmp));
+}
+static SkPMColor difference_modeproc(SkPMColor src, SkPMColor dst) {
+    int sa = SkGetPackedA32(src);
+    int da = SkGetPackedA32(dst);
+    int a = srcover_byte(sa, da);
+    int r = difference_byte(SkGetPackedR32(src), SkGetPackedR32(dst), sa, da);
+    int g = difference_byte(SkGetPackedG32(src), SkGetPackedG32(dst), sa, da);
+    int b = difference_byte(SkGetPackedB32(src), SkGetPackedB32(dst), sa, da);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kExclusion_Mode
+static inline int exclusion_byte(int sc, int dc, int, int) {
+    // this equations is wacky, wait for SVG to confirm it
+    //int r = sc * da + dc * sa - 2 * sc * dc + sc * (255 - da) + dc * (255 - sa);
+
+    // The above equation can be simplified as follows
+    int r = 255*(sc + dc) - 2 * sc * dc;
+    return clamp_div255round(r);
+}
+static SkPMColor exclusion_modeproc(SkPMColor src, SkPMColor dst) {
+    int sa = SkGetPackedA32(src);
+    int da = SkGetPackedA32(dst);
+    int a = srcover_byte(sa, da);
+    int r = exclusion_byte(SkGetPackedR32(src), SkGetPackedR32(dst), sa, da);
+    int g = exclusion_byte(SkGetPackedG32(src), SkGetPackedG32(dst), sa, da);
+    int b = exclusion_byte(SkGetPackedB32(src), SkGetPackedB32(dst), sa, da);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// The CSS compositing spec introduces the following formulas:
+// (See https://dvcs.w3.org/hg/FXTF/rawfile/tip/compositing/index.html#blendingnonseparable)
+// SkComputeLuminance is similar to this formula but it uses the new definition from Rec. 709
+// while PDF and CG uses the one from Rec. Rec. 601
+// See http://www.glennchan.info/articles/technical/hd-versus-sd-color-space/hd-versus-sd-color-space.htm
+static inline int Lum(int r, int g, int b)
+{
+    return SkDiv255Round(r * 77 + g * 150 + b * 28);
+}
+
+static inline int min2(int a, int b) { return a < b ? a : b; }
+static inline int max2(int a, int b) { return a > b ? a : b; }
+#define minimum(a, b, c) min2(min2(a, b), c)
+#define maximum(a, b, c) max2(max2(a, b), c)
+
+static inline int Sat(int r, int g, int b) {
+    return maximum(r, g, b) - minimum(r, g, b);
+}
+
+static inline void setSaturationComponents(int* Cmin, int* Cmid, int* Cmax, int s) {
+    if(*Cmax > *Cmin) {
+        *Cmid =  SkMulDiv(*Cmid - *Cmin, s, *Cmax - *Cmin);
+        *Cmax = s;
+    } else {
+        *Cmax = 0;
+        *Cmid = 0;
+    }
+
+    *Cmin = 0;
+}
+
+static inline void SetSat(int* r, int* g, int* b, int s) {
+    if(*r <= *g) {
+        if(*g <= *b) {
+            setSaturationComponents(r, g, b, s);
+        } else if(*r <= *b) {
+            setSaturationComponents(r, b, g, s);
+        } else {
+            setSaturationComponents(b, r, g, s);
         }
-        p.append(SkRasterPipeline::store_8888, &dst);
-        p.run(0, 0, count);
+    } else if(*r <= *b) {
+        setSaturationComponents(g, r, b, s);
+    } else if(*g <= *b) {
+        setSaturationComponents(g, b, r, s);
+    } else {
+        setSaturationComponents(b, g, r, s);
+    }
+}
+
+static inline void clipColor(int* r, int* g, int* b, int a) {
+    int L = Lum(*r, *g, *b);
+    int n = minimum(*r, *g, *b);
+    int x = maximum(*r, *g, *b);
+    int denom;
+    if ((n < 0) && (denom = L - n)) { // Compute denom and make sure it's non zero
+       *r = L + SkMulDiv(*r - L, L, denom);
+       *g = L + SkMulDiv(*g - L, L, denom);
+       *b = L + SkMulDiv(*b - L, L, denom);
+    }
+
+    if ((x > a) && (denom = x - L)) { // Compute denom and make sure it's non zero
+       int numer = a - L;
+       *r = L + SkMulDiv(*r - L, numer, denom);
+       *g = L + SkMulDiv(*g - L, numer, denom);
+       *b = L + SkMulDiv(*b - L, numer, denom);
+    }
+}
+
+static inline void SetLum(int* r, int* g, int* b, int a, int l) {
+  int d = l - Lum(*r, *g, *b);
+  *r +=  d;
+  *g +=  d;
+  *b +=  d;
+
+  clipColor(r, g, b, a);
+}
+
+// non-separable blend modes are done in non-premultiplied alpha
+#define  blendfunc_nonsep_byte(sc, dc, sa, da, blendval) \
+  clamp_div255round(sc * (255 - da) +  dc * (255 - sa) + blendval)
+
+// kHue_Mode
+// B(Cb, Cs) = SetLum(SetSat(Cs, Sat(Cb)), Lum(Cb))
+// Create a color with the hue of the source color and the saturation and luminosity of the backdrop color.
+static SkPMColor hue_modeproc(SkPMColor src, SkPMColor dst) {
+    int sr = SkGetPackedR32(src);
+    int sg = SkGetPackedG32(src);
+    int sb = SkGetPackedB32(src);
+    int sa = SkGetPackedA32(src);
+
+    int dr = SkGetPackedR32(dst);
+    int dg = SkGetPackedG32(dst);
+    int db = SkGetPackedB32(dst);
+    int da = SkGetPackedA32(dst);
+    int Sr, Sg, Sb;
+
+    if(sa && da) {
+        Sr = sr * sa;
+        Sg = sg * sa;
+        Sb = sb * sa;
+        SetSat(&Sr, &Sg, &Sb, Sat(dr, dg, db) * sa);
+        SetLum(&Sr, &Sg, &Sb, sa * da, Lum(dr, dg, db) * sa);
+    } else {
+        Sr = 0;
+        Sg = 0;
+        Sb = 0;
+    }
+
+    int a = srcover_byte(sa, da);
+    int r = blendfunc_nonsep_byte(sr, dr, sa, da, Sr);
+    int g = blendfunc_nonsep_byte(sg, dg, sa, da, Sg);
+    int b = blendfunc_nonsep_byte(sb, db, sa, da, Sb);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kSaturation_Mode
+// B(Cb, Cs) = SetLum(SetSat(Cb, Sat(Cs)), Lum(Cb))
+// Create a color with the saturation of the source color and the hue and luminosity of the backdrop color.
+static SkPMColor saturation_modeproc(SkPMColor src, SkPMColor dst) {
+    int sr = SkGetPackedR32(src);
+    int sg = SkGetPackedG32(src);
+    int sb = SkGetPackedB32(src);
+    int sa = SkGetPackedA32(src);
+
+    int dr = SkGetPackedR32(dst);
+    int dg = SkGetPackedG32(dst);
+    int db = SkGetPackedB32(dst);
+    int da = SkGetPackedA32(dst);
+    int Dr, Dg, Db;
+
+    if(sa && da) {
+        Dr = dr * sa;
+        Dg = dg * sa;
+        Db = db * sa;
+        SetSat(&Dr, &Dg, &Db, Sat(sr, sg, sb) * da);
+        SetLum(&Dr, &Dg, &Db, sa * da, Lum(dr, dg, db) * sa);
+    } else {
+        Dr = 0;
+        Dg = 0;
+        Db = 0;
+    }
+
+    int a = srcover_byte(sa, da);
+    int r = blendfunc_nonsep_byte(sr, dr, sa, da, Dr);
+    int g = blendfunc_nonsep_byte(sg, dg, sa, da, Dg);
+    int b = blendfunc_nonsep_byte(sb, db, sa, da, Db);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kColor_Mode
+// B(Cb, Cs) = SetLum(Cs, Lum(Cb))
+// Create a color with the hue and saturation of the source color and the luminosity of the backdrop color.
+static SkPMColor color_modeproc(SkPMColor src, SkPMColor dst) {
+    int sr = SkGetPackedR32(src);
+    int sg = SkGetPackedG32(src);
+    int sb = SkGetPackedB32(src);
+    int sa = SkGetPackedA32(src);
+
+    int dr = SkGetPackedR32(dst);
+    int dg = SkGetPackedG32(dst);
+    int db = SkGetPackedB32(dst);
+    int da = SkGetPackedA32(dst);
+    int Sr, Sg, Sb;
+
+    if(sa && da) {
+        Sr = sr * da;
+        Sg = sg * da;
+        Sb = sb * da;
+        SetLum(&Sr, &Sg, &Sb, sa * da, Lum(dr, dg, db) * sa);
+    } else {
+        Sr = 0;
+        Sg = 0;
+        Sb = 0;
+    }
+
+    int a = srcover_byte(sa, da);
+    int r = blendfunc_nonsep_byte(sr, dr, sa, da, Sr);
+    int g = blendfunc_nonsep_byte(sg, dg, sa, da, Sg);
+    int b = blendfunc_nonsep_byte(sb, db, sa, da, Sb);
+    return SkPackARGB32(a, r, g, b);
+}
+
+// kLuminosity_Mode
+// B(Cb, Cs) = SetLum(Cb, Lum(Cs))
+// Create a color with the luminosity of the source color and the hue and saturation of the backdrop color.
+static SkPMColor luminosity_modeproc(SkPMColor src, SkPMColor dst) {
+    int sr = SkGetPackedR32(src);
+    int sg = SkGetPackedG32(src);
+    int sb = SkGetPackedB32(src);
+    int sa = SkGetPackedA32(src);
+
+    int dr = SkGetPackedR32(dst);
+    int dg = SkGetPackedG32(dst);
+    int db = SkGetPackedB32(dst);
+    int da = SkGetPackedA32(dst);
+    int Dr, Dg, Db;
+
+    if(sa && da) {
+        Dr = dr * sa;
+        Dg = dg * sa;
+        Db = db * sa;
+        SetLum(&Dr, &Dg, &Db, sa * da, Lum(sr, sg, sb) * da);
+    } else {
+        Dr = 0;
+        Dg = 0;
+        Db = 0;
     }
 
-private:
-    const SkBlendMode fMode;
+    int a = srcover_byte(sa, da);
+    int r = blendfunc_nonsep_byte(sr, dr, sa, da, Dr);
+    int g = blendfunc_nonsep_byte(sg, dg, sa, da, Dg);
+    int b = blendfunc_nonsep_byte(sb, db, sa, da, Db);
+    return SkPackARGB32(a, r, g, b);
+}
 
-    typedef SkXfermode INHERITED;
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+const SkXfermodeProc gProcs[] = {
+    clear_modeproc,
+    src_modeproc,
+    dst_modeproc,
+    srcover_modeproc,
+    dstover_modeproc,
+    srcin_modeproc,
+    dstin_modeproc,
+    srcout_modeproc,
+    dstout_modeproc,
+    srcatop_modeproc,
+    dstatop_modeproc,
+    xor_modeproc,
+
+    plus_modeproc,
+    modulate_modeproc,
+    screen_modeproc,
+    overlay_modeproc,
+    darken_modeproc,
+    lighten_modeproc,
+    colordodge_modeproc,
+    colorburn_modeproc,
+    hardlight_modeproc,
+    softlight_modeproc,
+    difference_modeproc,
+    exclusion_modeproc,
+    multiply_modeproc,
+    hue_modeproc,
+    saturation_modeproc,
+    color_modeproc,
+    luminosity_modeproc,
 };
 
+///////////////////////////////////////////////////////////////////////////////
+///////////////////////////////////////////////////////////////////////////////
+
+void SkProcCoeffXfermode::xfer32(SkPMColor* SK_RESTRICT dst,
+                                 const SkPMColor* SK_RESTRICT src, int count,
+                                 const SkAlpha* SK_RESTRICT aa) const {
+    SkASSERT(dst && src && count >= 0);
+
+    SkXfermodeProc proc = fProc;
+
+    if (proc) {
+        if (nullptr == aa) {
+            for (int i = count - 1; i >= 0; --i) {
+                dst[i] = proc(src[i], dst[i]);
+            }
+        } else {
+            for (int i = count - 1; i >= 0; --i) {
+                unsigned a = aa[i];
+                if (0 != a) {
+                    SkPMColor dstC = dst[i];
+                    SkPMColor C = proc(src[i], dstC);
+                    if (a != 0xFF) {
+                        C = SkFourByteInterp(C, dstC, a);
+                    }
+                    dst[i] = C;
+                }
+            }
+        }
+    }
+}
+
 const char* SkBlendMode_Name(SkBlendMode mode) {
     SkASSERT((unsigned) mode <= (unsigned)SkBlendMode::kLastMode);
     const char* gModeStrings[] = {
@@ -94,20 +712,30 @@ sk_sp<SkXfermode> SkXfermode::Make(SkBlendMode mode) {
     }
 
     const int COUNT_BLENDMODES = (int)SkBlendMode::kLastMode + 1;
+    SkASSERT(SK_ARRAY_COUNT(gProcs) == COUNT_BLENDMODES);
 
     static SkOnce        once[COUNT_BLENDMODES];
     static SkXfermode* cached[COUNT_BLENDMODES];
 
     once[(int)mode]([mode] {
+        SkXfermodeProc proc = gProcs[(int)mode];
         if (auto xfermode = SkOpts::create_xfermode(mode)) {
             cached[(int)mode] = xfermode;
         } else {
-            cached[(int)mode] = new SkProcCoeffXfermode(mode);
+            cached[(int)mode] = new SkProcCoeffXfermode(proc, mode);
         }
     });
     return sk_ref_sp(cached[(int)mode]);
 }
 
+SkXfermodeProc SkXfermode::GetProc(SkBlendMode mode) {
+    SkXfermodeProc  proc = nullptr;
+    if ((unsigned)mode <= (unsigned)SkBlendMode::kLastMode) {
+        proc = gProcs[(unsigned)mode];
+    }
+    return proc;
+}
+
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 
 bool SkXfermode::IsOpaque(SkBlendMode mode, SrcColorOpacity opacityType) {