experiment: float color components

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1617813002

Review URL: https://codereview.chromium.org/1617813002

include/core/SkColor.h

@@ -160,4 +160,44 @@ SK_API SkPMColor SkPreMultiplyColor(SkColor c);
 */
 typedef SkPMColor (*SkXfermodeProc)(SkPMColor src, SkPMColor dst);
 
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+/*
+ *  The float values are 0...1 premultiplied
+ */
+struct SK_ATTRIBUTE(aligned(16)) SkPM4f {
+    float fVec[4];
+
+    float a() const { return fVec[SK_A32_SHIFT/8]; }
+};
+
+/*
+ *  The float values are 0...1 unpremultiplied
+ */
+struct SkColor4f {
+    float fA;
+    float fR;
+    float fG;
+    float fB;
+
+    bool operator==(const SkColor4f& other) const {
+        return fA == other.fA && fR == other.fR && fG == other.fG && fB == other.fB;
+    }
+    bool operator!=(const SkColor4f& other) const {
+        return !(*this == other);
+    }
+
+    const float* vec() const { return &fA; }
+    float* vec() { return &fA; }
+
+    static SkColor4f Pin(float a, float r, float g, float b);
+    static SkColor4f FromColor(SkColor);
+
+    SkColor4f pin() const {
+        return Pin(fA, fR, fG, fB);
+    }
+
+    SkPM4f premul() const;
+};
+
 #endif

include/core/SkColorFilter.h

@@ -68,10 +68,13 @@ public:
     */
     virtual void filterSpan(const SkPMColor src[], int count, SkPMColor result[]) const = 0;
 
+    virtual void filterSpan4f(const SkPM4f src[], int count, SkPM4f result[]) const;
+
     enum Flags {
         /** If set the filter methods will not change the alpha channel of the colors.
         */
-        kAlphaUnchanged_Flag = 0x01,
+        kAlphaUnchanged_Flag = 1 << 0,
+        kSupports4f_Flag     = 1 << 1,
     };
 
     /** Returns the flags for this filter. Override in subclasses to return custom flags.

include/core/SkColorPriv.h

@@ -71,6 +71,11 @@
      SK_B32_SHIFT == SK_BGRA_B32_SHIFT)
 
 
+#define SK_A_INDEX  (SK_A32_SHIFT/8)
+#define SK_R_INDEX  (SK_R32_SHIFT/8)
+#define SK_G_INDEX  (SK_G32_SHIFT/8)
+#define SK_B_INDEX  (SK_B32_SHIFT/8)
+
 #if defined(SK_PMCOLOR_IS_RGBA) && !LOCAL_PMCOLOR_SHIFTS_EQUIVALENT_TO_RGBA
     #error "SK_PMCOLOR_IS_RGBA does not match SK_*32_SHIFT values"
 #endif

include/core/SkShader.h

@@ -81,6 +81,7 @@ public:
             shadeSpan().
          */
         kConstInY32_Flag = 1 << 1,
+        kSupports4f_Flag = 1 << 2,
     };
 
     /**
@@ -127,6 +128,8 @@ public:
          */
         virtual void shadeSpan(int x, int y, SkPMColor[], int count) = 0;
 
+        virtual void shadeSpan4f(int x, int y, SkPM4f[], int count);
+
         /**
          * The const void* ctx is only const because all the implementations are const.
          * This can be changed to non-const if a new shade proc needs to change the ctx.

include/effects/SkColorMatrixFilter.h

@@ -29,6 +29,7 @@ public:
     static SkColorFilter* CreateLightingFilter(SkColor mul, SkColor add);
 
     void filterSpan(const SkPMColor src[], int count, SkPMColor[]) const override;
+    void filterSpan4f(const SkPM4f src[], int count, SkPM4f[]) const override;
     uint32_t getFlags() const override;
     bool asColorMatrix(SkScalar matrix[20]) const override;
     SkColorFilter* newComposed(const SkColorFilter*) const override;

src/core/SkColor.cpp

@@ -100,3 +100,37 @@ SkColor SkHSVToColor(U8CPU a, const SkScalar hsv[3]) {
     }
     return SkColorSetARGB(a, r, g, b);
 }
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+#include "SkNx.h"
+
+SkColor4f SkColor4f::Pin(float a, float r, float g, float b) {
+    SkColor4f c4;
+    Sk4f::Min(Sk4f::Max(Sk4f(a, r, g, b), Sk4f(0)), Sk4f(1)).store(c4.vec());
+    return c4;
+}
+
+SkColor4f SkColor4f::FromColor(SkColor c) {
+    Sk4f value = SkNx_shuffle<3,2,1,0>(SkNx_cast<float>(Sk4b::Load((const uint8_t*)&c)));
+    SkColor4f c4;
+    (value * Sk4f(1.0f / 255)).store(c4.vec());
+    return c4;
+}
+
+SkPM4f SkColor4f::premul() const {
+    auto src = Sk4f::Load(this->pin().vec());
+    float srcAlpha = src.kth<0>();  // need the pinned version of our alpha
+    src = src * Sk4f(1, srcAlpha, srcAlpha, srcAlpha);
+
+#ifdef SK_PMCOLOR_IS_BGRA
+    // ARGB -> BGRA
+    Sk4f dst = SkNx_shuffle<3,2,1,0>(src);
+#else
+    // ARGB -> RGBA
+    Sk4f dst = SkNx_shuffle<1,2,3,0>(src);
+#endif
+
+    SkPM4f pm4;
+    dst.store(pm4.fVec);
+    return pm4;
+}

src/core/SkColorFilter.cpp

@@ -35,6 +35,10 @@ SkColor SkColorFilter::filterColor(SkColor c) const {
     return SkUnPreMultiply::PMColorToColor(dst);
 }
 
+void SkColorFilter::filterSpan4f(const SkPM4f[], int count, SkPM4f[]) const {
+    SkASSERT(false && "filterSpan4f called but not implemented");
+}
+
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 
 /*

src/core/SkColorShader.h

@@ -36,8 +36,10 @@ public:
         uint32_t getFlags() const override;
         void shadeSpan(int x, int y, SkPMColor span[], int count) override;
         void shadeSpanAlpha(int x, int y, uint8_t alpha[], int count) override;
+        void shadeSpan4f(int x, int y, SkPM4f[], int count) override;
 
     private:
+        SkPM4f      fPM4f;
         SkPMColor   fPMColor;
         uint32_t    fFlags;
 

src/core/SkMathPriv.h

@@ -78,4 +78,8 @@ static inline unsigned SkDiv255Round(unsigned prod) {
     return (prod + (prod >> 8)) >> 8;
 }
 
+static inline float SkPinToUnitFloat(float x) {
+    return SkTMin(SkTMax(x, 0.0f), 1.0f);
+}
+
 #endif

src/core/SkShader.cpp

@@ -117,6 +117,10 @@ SkShader::Context::ShadeProc SkShader::Context::asAShadeProc(void** ctx) {
     return nullptr;
 }
 
+void SkShader::Context::shadeSpan4f(int x, int y, SkPM4f[], int count) {
+    SkASSERT(false && "shadeSpan4f called but not implemented");
+}
+
 #include "SkColorPriv.h"
 
 #define kTempColorQuadCount 6   // balance between speed (larger) and saving stack-space
@@ -279,7 +283,11 @@ SkColorShader::ColorShaderContext::ColorShaderContext(const SkColorShader& shade
     }
     fPMColor = SkPackARGB32(a, r, g, b);
 
-    fFlags = kConstInY32_Flag;
+    SkColor4f c4 = SkColor4f::FromColor(shader.fColor);
+    c4.fA *= rec.fPaint->getAlpha() / 255.0f;
+    fPM4f = c4.premul();
+
+    fFlags = kConstInY32_Flag | kSupports4f_Flag;
     if (255 == a) {
         fFlags |= kOpaqueAlpha_Flag;
     }
@@ -293,6 +301,12 @@ void SkColorShader::ColorShaderContext::shadeSpanAlpha(int x, int y, uint8_t alp
     memset(alpha, SkGetPackedA32(fPMColor), count);
 }
 
+void SkColorShader::ColorShaderContext::shadeSpan4f(int x, int y, SkPM4f span[], int count) {
+    for (int i = 0; i < count; ++i) {
+        span[i] = fPM4f;
+    }
+}
+
 SkShader::GradientType SkColorShader::asAGradient(GradientInfo* info) const {
     if (info) {
         if (info->fColors && info->fColorCount >= 1) {

src/effects/SkColorMatrixFilter.cpp

@@ -45,10 +45,11 @@ void SkColorMatrixFilter::initState(const SkScalar* SK_RESTRICT src) {
     bool usesAlpha = (array[3] || array[8] || array[13]);
 
     if (changesAlpha || usesAlpha) {
-        fFlags = changesAlpha ? 0 : SkColorFilter::kAlphaUnchanged_Flag;
+        fFlags = changesAlpha ? 0 : kAlphaUnchanged_Flag;
     } else {
-        fFlags = SkColorFilter::kAlphaUnchanged_Flag;
+        fFlags = kAlphaUnchanged_Flag;
     }
+    fFlags |= kSupports4f_Flag;
 }
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -89,28 +90,28 @@ static SkPMColor round(const Sk4f& x) {
     return c;
 }
 
-void SkColorMatrixFilter::filterSpan(const SkPMColor src[], int count, SkPMColor dst[]) const {
+template <typename Adaptor, typename T>
+void filter_span(const float array[], const T src[], int count, T dst[]) {
     // c0-c3 are already in [0,1].
-    const Sk4f c0 = Sk4f::Load(fTranspose + 0);
+    const Sk4f c0 = Sk4f::Load(array + 0);
-    const Sk4f c1 = Sk4f::Load(fTranspose + 4);
+    const Sk4f c1 = Sk4f::Load(array + 4);
-    const Sk4f c2 = Sk4f::Load(fTranspose + 8);
+    const Sk4f c2 = Sk4f::Load(array + 8);
-    const Sk4f c3 = Sk4f::Load(fTranspose + 12);
+    const Sk4f c3 = Sk4f::Load(array + 12);
     // c4 (the translate vector) is in [0, 255].  Bring it back to [0,1].
-    const Sk4f c4 = Sk4f::Load(fTranspose + 16)*Sk4f(1.0f/255);
+    const Sk4f c4 = Sk4f::Load(array + 16)*Sk4f(1.0f/255);
 
     // todo: we could cache this in the constructor...
-    SkPMColor matrix_translate_pmcolor = round(premul(clamp_0_1(c4)));
+    T matrix_translate_pmcolor = Adaptor::From4f(premul(clamp_0_1(c4)));
 
     for (int i = 0; i < count; i++) {
-        const SkPMColor src_c = src[i];
+        Sk4f srcf = Adaptor::To4f(src[i]);
-        if (0 == src_c) {
+        float srcA = srcf.kth<SK_A32_SHIFT/8>();
+
+        if (0 == srcA) {
             dst[i] = matrix_translate_pmcolor;
             continue;
         }
-
+        if (1 == srcA) {
-        Sk4f srcf = SkNx_cast<float>(Sk4b::Load((const uint8_t*)&src_c)) * Sk4f(1.0f/255);
-
-        if (0xFF != SkGetPackedA32(src_c)) {
             srcf = unpremul(srcf);
         }
 
@@ -122,11 +123,36 @@ void SkColorMatrixFilter::filterSpan(const SkPMColor src[], int count, SkPMColor
         // apply matrix
         Sk4f dst4 = c0 * r4 + c1 * g4 + c2 * b4 + c3 * a4 + c4;
 
-        // clamp, re-premul, and write
+        dst[i] = Adaptor::From4f(premul(clamp_0_1(dst4)));
-        dst[i] = round(premul(clamp_0_1(dst4)));
     }
 }
 
+struct SkPMColorAdaptor {
+    static SkPMColor From4f(const Sk4f& c4) {
+        return round(c4);
+    }
+    static Sk4f To4f(SkPMColor c) {
+        return SkNx_cast<float>(Sk4b::Load((const uint8_t*)&c)) * Sk4f(1.0f/255);
+    }
+};
+void SkColorMatrixFilter::filterSpan(const SkPMColor src[], int count, SkPMColor dst[]) const {
+    filter_span<SkPMColorAdaptor>(fTranspose, src, count, dst);
+}
+
+struct SkPM4fAdaptor {
+    static SkPM4f From4f(const Sk4f& c4) {
+        SkPM4f c;
+        c4.store(c.fVec);
+        return c;
+    }
+    static Sk4f To4f(const SkPM4f& c) {
+        return Sk4f::Load(c.fVec);
+    }
+};
+void SkColorMatrixFilter::filterSpan4f(const SkPM4f src[], int count, SkPM4f dst[]) const {
+    filter_span<SkPM4fAdaptor>(fTranspose, src, count, dst);
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 
 void SkColorMatrixFilter::flatten(SkWriteBuffer& buffer) const {

tests/SkColor4fTest.cpp

@@ -0,0 +1,62 @@
+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "SkColor.h"
+#include "SkShader.h"
+#include "SkColorMatrixFilter.h"
+#include "Test.h"
+#include "SkRandom.h"
+
+DEF_TEST(SkColor4f_FromColor, reporter) {
+    const struct {
+        SkColor     fC;
+        SkColor4f   fC4;
+    } recs[] = {
+        { SK_ColorBLACK, { 1, 0, 0, 0 } },
+        { SK_ColorWHITE, { 1, 1, 1, 1 } },
+        { SK_ColorRED,   { 1, 1, 0, 0 } },
+        { SK_ColorGREEN, { 1, 0, 1, 0 } },
+        { SK_ColorBLUE,  { 1, 0, 0, 1 } },
+        { 0,             { 0, 0, 0, 0 } },
+        { 0x55AAFF00,    { 1/3.0f, 2/3.0f, 1, 0 } },
+    };
+
+    for (const auto& r : recs) {
+        SkColor4f c4 = SkColor4f::FromColor(r.fC);
+        REPORTER_ASSERT(reporter, c4 == r.fC4);
+    }
+}
+
+static bool nearly_equal(float a, float b) {
+    const float kTolerance = 1.0f / (1 << 20);
+    return fabsf(a - b) < kTolerance;
+}
+
+DEF_TEST(SkColor4f_premul, reporter) {
+    SkRandom rand;
+
+    for (int i = 0; i < 1000000; ++i) {
+        // First just test opaque colors, so that the premul should be exact
+        SkColor4f c4 {
+            1, rand.nextUScalar1(), rand.nextUScalar1(), rand.nextUScalar1()
+        };
+        SkPM4f pm4 = c4.premul();
+        REPORTER_ASSERT(reporter, pm4.fVec[SK_A_INDEX] == c4.fA);
+        REPORTER_ASSERT(reporter, pm4.fVec[SK_R_INDEX] == c4.fA * c4.fR);
+        REPORTER_ASSERT(reporter, pm4.fVec[SK_G_INDEX] == c4.fA * c4.fG);
+        REPORTER_ASSERT(reporter, pm4.fVec[SK_B_INDEX] == c4.fA * c4.fB);
+
+        // We compare with a tolerance, in case our premul multiply is implemented at slightly
+        // different precision than the test code.
+        c4.fA = rand.nextUScalar1();
+        pm4 = c4.premul();
+        REPORTER_ASSERT(reporter, pm4.fVec[SK_A_INDEX] == c4.fA);
+        REPORTER_ASSERT(reporter, nearly_equal(pm4.fVec[SK_R_INDEX], c4.fA * c4.fR));
+        REPORTER_ASSERT(reporter, nearly_equal(pm4.fVec[SK_G_INDEX], c4.fA * c4.fG));
+        REPORTER_ASSERT(reporter, nearly_equal(pm4.fVec[SK_B_INDEX], c4.fA * c4.fB));
+    }
+}