Move SkColorMatrixFilter implementation to core.

GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1648933002

Review URL: https://codereview.chromium.org/1648933002

9 files changed, 479 insertions, 393 deletions

diff --git a/gyp/core.gypi b/gyp/core.gypi
index b80f43a72f..80f0eac956 100644
--- a/gyp/core.gypi
+++ b/gyp/core.gypi
@@ -71,6 +71,8 @@
         '<(skia_src_path)/core/SkColor.cpp',
         '<(skia_src_path)/core/SkColorFilter.cpp',
         '<(skia_src_path)/core/SkColorFilterShader.cpp',
+        '<(skia_src_path)/core/SkColorMatrixFilterRowMajor255.cpp',
+        '<(skia_src_path)/core/SkColorMatrixFilterRowMajor255.h',
         '<(skia_src_path)/core/SkColorShader.h',
         '<(skia_src_path)/core/SkColorTable.cpp',
         '<(skia_src_path)/core/SkComposeShader.cpp',
diff --git a/gyp/effects.gypi b/gyp/effects.gypi
index ea750b9bca..bbda38ae89 100644
--- a/gyp/effects.gypi
+++ b/gyp/effects.gypi
@@ -89,9 +89,9 @@
     '<(skia_include_path)/effects/SkBlurImageFilter.h',
     '<(skia_include_path)/effects/SkBlurMaskFilter.h',
     '<(skia_include_path)/effects/SkColorCubeFilter.h',
+    '<(skia_include_path)/effects/SkColorFilterImageFilter.h',
     '<(skia_include_path)/effects/SkColorMatrix.h',
     '<(skia_include_path)/effects/SkColorMatrixFilter.h',
-    '<(skia_include_path)/effects/SkColorFilterImageFilter.h',
     '<(skia_include_path)/effects/SkCornerPathEffect.h',
     '<(skia_include_path)/effects/SkDashPathEffect.h',
     '<(skia_include_path)/effects/SkDiscretePathEffect.h',
diff --git a/include/core/SkColorFilter.h b/include/core/SkColorFilter.h
index aece22e05a..14192c55ea 100644
--- a/include/core/SkColorFilter.h
+++ b/include/core/SkColorFilter.h
@@ -122,6 +122,11 @@ public:
      */
     static SkColorFilter* CreateComposeFilter(SkColorFilter* outer, SkColorFilter* inner);
 
+    /** Construct a color filter that transforms a color by a 4x5 matrix. The matrix is in row-
+     *  major order and the translation column is specified in unnormalized, 0...255, space.
+     */
+    static SkColorFilter* CreateMatrixFilterRowMajor255(const SkScalar array[20]);
+
     /**
      *  A subclass may implement this factory function to work with the GPU backend. It returns
      *  a GrFragmentProcessor that implemets the color filter in GPU shader code.
diff --git a/include/effects/SkColorMatrixFilter.h b/include/effects/SkColorMatrixFilter.h
index 3edf791f06..d71d2ed533 100644
--- a/include/effects/SkColorMatrixFilter.h
+++ b/include/effects/SkColorMatrixFilter.h
@@ -14,10 +14,10 @@
 class SK_API SkColorMatrixFilter : public SkColorFilter {
 public:
     static SkColorFilter* Create(const SkColorMatrix& cm) {
-        return new SkColorMatrixFilter(cm);
+        return SkColorFilter::CreateMatrixFilterRowMajor255(cm.fMat);
     }
     static SkColorFilter* Create(const SkScalar array[20]) {
-        return new SkColorMatrixFilter(array);
+        return SkColorFilter::CreateMatrixFilterRowMajor255(array);
     }
 
     /**
@@ -27,34 +27,6 @@ public:
      *  are ignored.
      */
     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;
-
-#if SK_SUPPORT_GPU
-    const GrFragmentProcessor* asFragmentProcessor(GrContext*) const override;
-#endif
-
-    SK_TO_STRING_OVERRIDE()
-
-    SK_DECLARE_PUBLIC_FLATTENABLE_DESERIALIZATION_PROCS(SkColorMatrixFilter)
-
-protected:
-    explicit SkColorMatrixFilter(const SkColorMatrix&);
-    explicit SkColorMatrixFilter(const SkScalar array[20]);
-    void flatten(SkWriteBuffer&) const override;
-
-private:
-    SkColorMatrix   fMatrix;
-    float           fTranspose[SkColorMatrix::kCount]; // for Sk4s
-    uint32_t        fFlags;
-
-    void initState(const SkScalar array[20]);
-
-    typedef SkColorFilter INHERITED;
 };
 
 #endif
diff --git a/src/core/SkColorMatrixFilterRowMajor255.cpp b/src/core/SkColorMatrixFilterRowMajor255.cpp
new file mode 100644
index 0000000000..6fdfa0b446
--- /dev/null
+++ b/src/core/SkColorMatrixFilterRowMajor255.cpp
@@ -0,0 +1,422 @@
+/*
+ * Copyright 2011 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "SkColorMatrixFilterRowMajor255.h"
+#include "SkColorPriv.h"
+#include "SkNx.h"
+#include "SkReadBuffer.h"
+#include "SkWriteBuffer.h"
+#include "SkUnPreMultiply.h"
+#include "SkString.h"
+
+#define SK_PMORDER_INDEX_A  (SK_A32_SHIFT / 8)
+#define SK_PMORDER_INDEX_R  (SK_R32_SHIFT / 8)
+#define SK_PMORDER_INDEX_G  (SK_G32_SHIFT / 8)
+#define SK_PMORDER_INDEX_B  (SK_B32_SHIFT / 8)
+
+static void transpose_to_pmorder(float dst[20], const float src[20]) {
+    const float* srcR = src + 0;
+    const float* srcG = src + 5;
+    const float* srcB = src + 10;
+    const float* srcA = src + 15;
+
+    for (int i = 0; i < 20; i += 4) {
+        dst[i + SK_PMORDER_INDEX_A] = *srcA++;
+        dst[i + SK_PMORDER_INDEX_R] = *srcR++;
+        dst[i + SK_PMORDER_INDEX_G] = *srcG++;
+        dst[i + SK_PMORDER_INDEX_B] = *srcB++;
+    }
+}
+
+// src is [20] but some compilers won't accept __restrict__ on anything
+// but an raw pointer or reference
+void SkColorMatrixFilterRowMajor255::initState(const SkScalar* SK_RESTRICT src) {
+    transpose_to_pmorder(fTranspose, src);
+
+    const float* array = fMatrix;
+
+    // check if we have to munge Alpha
+    bool changesAlpha = (array[15] || array[16] || array[17] || (array[18] - 1) || array[19]);
+    bool usesAlpha = (array[3] || array[8] || array[13]);
+
+    if (changesAlpha || usesAlpha) {
+        fFlags = changesAlpha ? 0 : kAlphaUnchanged_Flag;
+    } else {
+        fFlags = kAlphaUnchanged_Flag;
+    }
+    fFlags |= kSupports4f_Flag;
+}
+
+///////////////////////////////////////////////////////////////////////////////
+
+SkColorMatrixFilterRowMajor255::SkColorMatrixFilterRowMajor255(const SkScalar array[20]) {
+    memcpy(fMatrix, array, 20 * sizeof(SkScalar));
+    this->initState(array);
+}
+
+uint32_t SkColorMatrixFilterRowMajor255::getFlags() const {
+    return this->INHERITED::getFlags() | fFlags;
+}
+
+static Sk4f scale_rgb(float scale) {
+    static_assert(SkPM4f::A == 3, "Alpha is lane 3");
+    return Sk4f(scale, scale, scale, 1);
+}
+
+static Sk4f premul(const Sk4f& x) {
+    return x * scale_rgb(x.kth<SkPM4f::A>());
+}
+
+static Sk4f unpremul(const Sk4f& x) {
+    return x * scale_rgb(1 / x.kth<SkPM4f::A>());  // TODO: fast/approx invert?
+}
+
+static Sk4f clamp_0_1(const Sk4f& x) {
+    return Sk4f::Max(Sk4f::Min(x, Sk4f(1)), Sk4f(0));
+}
+
+static SkPMColor round(const Sk4f& x) {
+    SkPMColor c;
+    SkNx_cast<uint8_t>(x * Sk4f(255) + Sk4f(0.5f)).store(&c);
+    return c;
+}
+
+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(array + 0);
+    const Sk4f c1 = Sk4f::Load(array + 4);
+    const Sk4f c2 = Sk4f::Load(array + 8);
+    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(array + 16)*Sk4f(1.0f/255);
+
+    // todo: we could cache this in the constructor...
+    T matrix_translate_pmcolor = Adaptor::From4f(premul(clamp_0_1(c4)));
+
+    for (int i = 0; i < count; i++) {
+        Sk4f srcf = Adaptor::To4f(src[i]);
+        float srcA = srcf.kth<SkPM4f::A>();
+
+        if (0 == srcA) {
+            dst[i] = matrix_translate_pmcolor;
+            continue;
+        }
+        if (1 != srcA) {
+            srcf = unpremul(srcf);
+        }
+
+        Sk4f r4 = SkNx_dup<SK_R32_SHIFT/8>(srcf);
+        Sk4f g4 = SkNx_dup<SK_G32_SHIFT/8>(srcf);
+        Sk4f b4 = SkNx_dup<SK_B32_SHIFT/8>(srcf);
+        Sk4f a4 = SkNx_dup<SK_A32_SHIFT/8>(srcf);
+
+        // apply matrix
+        Sk4f dst4 = c0 * r4 + c1 * g4 + c2 * b4 + c3 * a4 + c4;
+
+        dst[i] = Adaptor::From4f(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(&c)) * Sk4f(1.0f/255);
+    }
+};
+void SkColorMatrixFilterRowMajor255::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);
+        return c;
+    }
+    static Sk4f To4f(const SkPM4f& c) {
+        return Sk4f::Load(&c);
+    }
+};
+void SkColorMatrixFilterRowMajor255::filterSpan4f(const SkPM4f src[], int count, SkPM4f dst[]) const {
+    filter_span<SkPM4fAdaptor>(fTranspose, src, count, dst);
+}
+
+///////////////////////////////////////////////////////////////////////////////
+
+void SkColorMatrixFilterRowMajor255::flatten(SkWriteBuffer& buffer) const {
+    SkASSERT(sizeof(fMatrix)/sizeof(SkScalar) == 20);
+    buffer.writeScalarArray(fMatrix, 20);
+}
+
+SkFlattenable* SkColorMatrixFilterRowMajor255::CreateProc(SkReadBuffer& buffer) {
+    SkScalar matrix[20];
+    if (buffer.readScalarArray(matrix, 20)) {
+        return new SkColorMatrixFilterRowMajor255(matrix);
+    }
+    return nullptr;
+}
+
+bool SkColorMatrixFilterRowMajor255::asColorMatrix(SkScalar matrix[20]) const {
+    if (matrix) {
+        memcpy(matrix, fMatrix, 20 * sizeof(SkScalar));
+    }
+    return true;
+}
+
+///////////////////////////////////////////////////////////////////////////////
+//  This code was duplicated from src/effects/SkColorMatrixc.cpp in order to be used in core.
+//////
+
+// To detect if we need to apply clamping after applying a matrix, we check if
+// any output component might go outside of [0, 255] for any combination of
+// input components in [0..255].
+// Each output component is an affine transformation of the input component, so
+// the minimum and maximum values are for any combination of minimum or maximum
+// values of input components (i.e. 0 or 255).
+// E.g. if R' = x*R + y*G + z*B + w*A + t
+// Then the maximum value will be for R=255 if x>0 or R=0 if x<0, and the
+// minimum value will be for R=0 if x>0 or R=255 if x<0.
+// Same goes for all components.
+static bool component_needs_clamping(const SkScalar row[5]) {
+    SkScalar maxValue = row[4] / 255;
+    SkScalar minValue = row[4] / 255;
+    for (int i = 0; i < 4; ++i) {
+        if (row[i] > 0)
+            maxValue += row[i];
+        else
+            minValue += row[i];
+    }
+    return (maxValue > 1) || (minValue < 0);
+}
+
+static bool needs_clamping(const SkScalar matrix[20]) {
+    return component_needs_clamping(matrix)
+        || component_needs_clamping(matrix+5)
+        || component_needs_clamping(matrix+10)
+        || component_needs_clamping(matrix+15);
+}
+
+static void set_concat(SkScalar result[20], const SkScalar outer[20], const SkScalar inner[20]) {
+    int index = 0;
+    for (int j = 0; j < 20; j += 5) {
+        for (int i = 0; i < 4; i++) {
+            result[index++] =   outer[j + 0] * inner[i + 0] +
+                                outer[j + 1] * inner[i + 5] +
+                                outer[j + 2] * inner[i + 10] +
+                                outer[j + 3] * inner[i + 15];
+        }
+        result[index++] =   outer[j + 0] * inner[4] +
+                            outer[j + 1] * inner[9] +
+                            outer[j + 2] * inner[14] +
+                            outer[j + 3] * inner[19] +
+                            outer[j + 4];
+    }
+}
+
+///////////////////////////////////////////////////////////////////////////////
+//  End duplication
+//////
+
+SkColorFilter* SkColorMatrixFilterRowMajor255::newComposed(const SkColorFilter* innerFilter) const {
+    SkScalar innerMatrix[20];
+    if (innerFilter->asColorMatrix(innerMatrix) && !needs_clamping(innerMatrix)) {
+        SkScalar concat[20];
+        set_concat(concat, fMatrix, innerMatrix);
+        return new SkColorMatrixFilterRowMajor255(concat);
+    }
+    return nullptr;
+}
+
+#if SK_SUPPORT_GPU
+#include "GrFragmentProcessor.h"
+#include "GrInvariantOutput.h"
+#include "glsl/GrGLSLFragmentProcessor.h"
+#include "glsl/GrGLSLFragmentShaderBuilder.h"
+#include "glsl/GrGLSLProgramDataManager.h"
+#include "glsl/GrGLSLUniformHandler.h"
+
+class ColorMatrixEffect : public GrFragmentProcessor {
+public:
+    static const GrFragmentProcessor* Create(const SkScalar matrix[20]) {
+        return new ColorMatrixEffect(matrix);
+    }
+
+    const char* name() const override { return "Color Matrix"; }
+
+    GR_DECLARE_FRAGMENT_PROCESSOR_TEST;
+
+    class GLSLProcessor : public GrGLSLFragmentProcessor {
+    public:
+        // this class always generates the same code.
+        static void GenKey(const GrProcessor&, const GrGLSLCaps&, GrProcessorKeyBuilder* b) {}
+
+        GLSLProcessor(const GrProcessor&) {}
+
+        virtual void emitCode(EmitArgs& args) override {
+            GrGLSLUniformHandler* uniformHandler = args.fUniformHandler;
+            fMatrixHandle = uniformHandler->addUniform(GrGLSLUniformHandler::kFragment_Visibility,
+                                                       kMat44f_GrSLType, kDefault_GrSLPrecision,
+                                                       "ColorMatrix");
+            fVectorHandle = uniformHandler->addUniform(GrGLSLUniformHandler::kFragment_Visibility,
+                                                       kVec4f_GrSLType, kDefault_GrSLPrecision,
+                                                       "ColorMatrixVector");
+
+            if (nullptr == args.fInputColor) {
+                // could optimize this case, but we aren't for now.
+                args.fInputColor = "vec4(1)";
+            }
+            GrGLSLFragmentBuilder* fragBuilder = args.fFragBuilder;
+            // The max() is to guard against 0 / 0 during unpremul when the incoming color is
+            // transparent black.
+            fragBuilder->codeAppendf("\tfloat nonZeroAlpha = max(%s.a, 0.00001);\n",
+                                     args.fInputColor);
+            fragBuilder->codeAppendf("\t%s = %s * vec4(%s.rgb / nonZeroAlpha, nonZeroAlpha) + %s;\n",
+                                     args.fOutputColor,
+                                     uniformHandler->getUniformCStr(fMatrixHandle),
+                                     args.fInputColor,
+                                     uniformHandler->getUniformCStr(fVectorHandle));
+            fragBuilder->codeAppendf("\t%s = clamp(%s, 0.0, 1.0);\n",
+                                     args.fOutputColor, args.fOutputColor);
+            fragBuilder->codeAppendf("\t%s.rgb *= %s.a;\n", args.fOutputColor, args.fOutputColor);
+        }
+
+    protected:
+        virtual void onSetData(const GrGLSLProgramDataManager& uniManager,
+                               const GrProcessor& proc) override {
+            const ColorMatrixEffect& cme = proc.cast<ColorMatrixEffect>();
+            const float* m = cme.fMatrix;
+            // The GL matrix is transposed from SkColorMatrix.
+            float mt[]  = {
+                m[0], m[5], m[10], m[15],
+                m[1], m[6], m[11], m[16],
+                m[2], m[7], m[12], m[17],
+                m[3], m[8], m[13], m[18],
+            };
+            static const float kScale = 1.0f / 255.0f;
+            float vec[] = {
+                m[4] * kScale, m[9] * kScale, m[14] * kScale, m[19] * kScale,
+            };
+            uniManager.setMatrix4fv(fMatrixHandle, 1, mt);
+            uniManager.set4fv(fVectorHandle, 1, vec);
+        }
+
+    private:
+        GrGLSLProgramDataManager::UniformHandle fMatrixHandle;
+        GrGLSLProgramDataManager::UniformHandle fVectorHandle;
+
+        typedef GrGLSLFragmentProcessor INHERITED;
+    };
+
+private:
+    ColorMatrixEffect(const SkScalar matrix[20]) {
+        memcpy(fMatrix, matrix, sizeof(SkScalar) * 20);
+        this->initClassID<ColorMatrixEffect>();
+    }
+
+    GrGLSLFragmentProcessor* onCreateGLSLInstance() const override {
+        return new GLSLProcessor(*this);
+    }
+
+    virtual void onGetGLSLProcessorKey(const GrGLSLCaps& caps,
+                                       GrProcessorKeyBuilder* b) const override {
+        GLSLProcessor::GenKey(*this, caps, b);
+    }
+
+    bool onIsEqual(const GrFragmentProcessor& s) const override {
+        const ColorMatrixEffect& cme = s.cast<ColorMatrixEffect>();
+        return 0 == memcmp(fMatrix, cme.fMatrix, sizeof(fMatrix));
+    }
+
+    void onComputeInvariantOutput(GrInvariantOutput* inout) const override {
+        // We only bother to check whether the alpha channel will be constant. If SkColorMatrix had
+        // type flags it might be worth checking the other components.
+
+        // The matrix is defined such the 4th row determines the output alpha. The first four
+        // columns of that row multiply the input r, g, b, and a, respectively, and the last column
+        // is the "translation".
+        static const uint32_t kRGBAFlags[] = {
+            kR_GrColorComponentFlag,
+            kG_GrColorComponentFlag,
+            kB_GrColorComponentFlag,
+            kA_GrColorComponentFlag
+        };
+        static const int kShifts[] = {
+            GrColor_SHIFT_R, GrColor_SHIFT_G, GrColor_SHIFT_B, GrColor_SHIFT_A,
+        };
+        enum {
+            kAlphaRowStartIdx = 15,
+            kAlphaRowTranslateIdx = 19,
+        };
+
+        SkScalar outputA = 0;
+        for (int i = 0; i < 4; ++i) {
+            // If any relevant component of the color to be passed through the matrix is non-const
+            // then we can't know the final result.
+            if (0 != fMatrix[kAlphaRowStartIdx + i]) {
+                if (!(inout->validFlags() & kRGBAFlags[i])) {
+                    inout->setToUnknown(GrInvariantOutput::kWill_ReadInput);
+                    return;
+                } else {
+                    uint32_t component = (inout->color() >> kShifts[i]) & 0xFF;
+                    outputA += fMatrix[kAlphaRowStartIdx + i] * component;
+                }
+            }
+        }
+        outputA += fMatrix[kAlphaRowTranslateIdx];
+        // We pin the color to [0,1]. This would happen to the *final* color output from the frag
+        // shader but currently the effect does not pin its own output. So in the case of over/
+        // underflow this may deviate from the actual result. Maybe the effect should pin its
+        // result if the matrix could over/underflow for any component?
+        inout->setToOther(kA_GrColorComponentFlag,
+                          static_cast<uint8_t>(SkScalarPin(outputA, 0, 255)) << GrColor_SHIFT_A,
+                          GrInvariantOutput::kWill_ReadInput);
+    }
+
+    SkScalar fMatrix[20];
+
+    typedef GrFragmentProcessor INHERITED;
+};
+
+GR_DEFINE_FRAGMENT_PROCESSOR_TEST(ColorMatrixEffect);
+
+const GrFragmentProcessor* ColorMatrixEffect::TestCreate(GrProcessorTestData* d) {
+    SkScalar colorMatrix[20];
+    for (size_t i = 0; i < SK_ARRAY_COUNT(colorMatrix); ++i) {
+        colorMatrix[i] = d->fRandom->nextSScalar1();
+    }
+    return ColorMatrixEffect::Create(colorMatrix);
+}
+
+const GrFragmentProcessor* SkColorMatrixFilterRowMajor255::asFragmentProcessor(GrContext*) const {
+    return ColorMatrixEffect::Create(fMatrix);
+}
+
+#endif
+
+#ifndef SK_IGNORE_TO_STRING
+void SkColorMatrixFilterRowMajor255::toString(SkString* str) const {
+    str->append("SkColorMatrixFilterRowMajor255: ");
+
+    str->append("matrix: (");
+    for (int i = 0; i < 20; ++i) {
+        str->appendScalar(fMatrix[i]);
+        if (i < 19) {
+            str->append(", ");
+        }
+    }
+    str->append(")");
+}
+#endif
+
+///////////////////////////////////////////////////////////////////////////////
+
+SkColorFilter* SkColorFilter::CreateMatrixFilterRowMajor255(const SkScalar array[20]) {
+    return new SkColorMatrixFilterRowMajor255(array);
+}
diff --git a/src/core/SkColorMatrixFilterRowMajor255.h b/src/core/SkColorMatrixFilterRowMajor255.h
new file mode 100644
index 0000000000..453dd41bff
--- /dev/null
+++ b/src/core/SkColorMatrixFilterRowMajor255.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef SkColorMatrixFilter_DEFINED
+#define SkColorMatrixFilter_DEFINED
+
+#include "SkColorFilter.h"
+
+class SK_API SkColorMatrixFilterRowMajor255 : public SkColorFilter {
+public:
+    explicit SkColorMatrixFilterRowMajor255(const SkScalar array[20]);
+
+    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;
+
+#if SK_SUPPORT_GPU
+    const GrFragmentProcessor* asFragmentProcessor(GrContext*) const override;
+#endif
+
+    SK_TO_STRING_OVERRIDE()
+
+    SK_DECLARE_PUBLIC_FLATTENABLE_DESERIALIZATION_PROCS(SkColorMatrixFilter)
+
+protected:
+    void flatten(SkWriteBuffer&) const override;
+
+private:
+    SkScalar        fMatrix[20];
+    float           fTranspose[20]; // for Sk4s
+    uint32_t        fFlags;
+
+    void initState(const SkScalar array[20]);
+
+    typedef SkColorFilter INHERITED;
+};
+
+#endif
diff --git a/src/effects/SkColorMatrixFilter.cpp b/src/effects/SkColorMatrixFilter.cpp
index a0878a5044..e122b6ac8e 100644
--- a/src/effects/SkColorMatrixFilter.cpp
+++ b/src/effects/SkColorMatrixFilter.cpp
@@ -6,366 +6,6 @@
  */
 
 #include "SkColorMatrixFilter.h"
-#include "SkColorMatrix.h"
-#include "SkColorPriv.h"
-#include "SkNx.h"
-#include "SkReadBuffer.h"
-#include "SkWriteBuffer.h"
-#include "SkUnPreMultiply.h"
-#include "SkString.h"
-
-#define SK_PMORDER_INDEX_A  (SK_A32_SHIFT / 8)
-#define SK_PMORDER_INDEX_R  (SK_R32_SHIFT / 8)
-#define SK_PMORDER_INDEX_G  (SK_G32_SHIFT / 8)
-#define SK_PMORDER_INDEX_B  (SK_B32_SHIFT / 8)
-
-static void transpose_to_pmorder(float dst[20], const float src[20]) {
-    const float* srcR = src + 0;
-    const float* srcG = src + 5;
-    const float* srcB = src + 10;
-    const float* srcA = src + 15;
-
-    for (int i = 0; i < 20; i += 4) {
-        dst[i + SK_PMORDER_INDEX_A] = *srcA++;
-        dst[i + SK_PMORDER_INDEX_R] = *srcR++;
-        dst[i + SK_PMORDER_INDEX_G] = *srcG++;
-        dst[i + SK_PMORDER_INDEX_B] = *srcB++;
-    }
-}
-
-// src is [20] but some compilers won't accept __restrict__ on anything
-// but an raw pointer or reference
-void SkColorMatrixFilter::initState(const SkScalar* SK_RESTRICT src) {
-    transpose_to_pmorder(fTranspose, src);
-
-    const float* array = fMatrix.fMat;
-
-    // check if we have to munge Alpha
-    bool changesAlpha = (array[15] || array[16] || array[17] || (array[18] - 1) || array[19]);
-    bool usesAlpha = (array[3] || array[8] || array[13]);
-
-    if (changesAlpha || usesAlpha) {
-        fFlags = changesAlpha ? 0 : kAlphaUnchanged_Flag;
-    } else {
-        fFlags = kAlphaUnchanged_Flag;
-    }
-    fFlags |= kSupports4f_Flag;
-}
-
-///////////////////////////////////////////////////////////////////////////////
-
-SkColorMatrixFilter::SkColorMatrixFilter(const SkColorMatrix& cm) : fMatrix(cm) {
-    this->initState(cm.fMat);
-}
-
-SkColorMatrixFilter::SkColorMatrixFilter(const SkScalar array[20]) {
-    memcpy(fMatrix.fMat, array, 20 * sizeof(SkScalar));
-    this->initState(array);
-}
-
-uint32_t SkColorMatrixFilter::getFlags() const {
-    return this->INHERITED::getFlags() | fFlags;
-}
-
-static Sk4f scale_rgb(float scale) {
-    static_assert(SkPM4f::A == 3, "Alpha is lane 3");
-    return Sk4f(scale, scale, scale, 1);
-}
-
-static Sk4f premul(const Sk4f& x) {
-    return x * scale_rgb(x.kth<SkPM4f::A>());
-}
-
-static Sk4f unpremul(const Sk4f& x) {
-    return x * scale_rgb(1 / x.kth<SkPM4f::A>());  // TODO: fast/approx invert?
-}
-
-static Sk4f clamp_0_1(const Sk4f& x) {
-    return Sk4f::Max(Sk4f::Min(x, Sk4f(1)), Sk4f(0));
-}
-
-static SkPMColor round(const Sk4f& x) {
-    SkPMColor c;
-    SkNx_cast<uint8_t>(x * Sk4f(255) + Sk4f(0.5f)).store(&c);
-    return c;
-}
-
-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(array + 0);
-    const Sk4f c1 = Sk4f::Load(array + 4);
-    const Sk4f c2 = Sk4f::Load(array + 8);
-    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(array + 16)*Sk4f(1.0f/255);
-
-    // todo: we could cache this in the constructor...
-    T matrix_translate_pmcolor = Adaptor::From4f(premul(clamp_0_1(c4)));
-
-    for (int i = 0; i < count; i++) {
-        Sk4f srcf = Adaptor::To4f(src[i]);
-        float srcA = srcf.kth<SkPM4f::A>();
-
-        if (0 == srcA) {
-            dst[i] = matrix_translate_pmcolor;
-            continue;
-        }
-        if (1 != srcA) {
-            srcf = unpremul(srcf);
-        }
-
-        Sk4f r4 = SkNx_dup<SK_R32_SHIFT/8>(srcf);
-        Sk4f g4 = SkNx_dup<SK_G32_SHIFT/8>(srcf);
-        Sk4f b4 = SkNx_dup<SK_B32_SHIFT/8>(srcf);
-        Sk4f a4 = SkNx_dup<SK_A32_SHIFT/8>(srcf);
-
-        // apply matrix
-        Sk4f dst4 = c0 * r4 + c1 * g4 + c2 * b4 + c3 * a4 + c4;
-
-        dst[i] = Adaptor::From4f(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(&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);
-        return c;
-    }
-    static Sk4f To4f(const SkPM4f& c) {
-        return Sk4f::Load(&c);
-    }
-};
-void SkColorMatrixFilter::filterSpan4f(const SkPM4f src[], int count, SkPM4f dst[]) const {
-    filter_span<SkPM4fAdaptor>(fTranspose, src, count, dst);
-}
-
-///////////////////////////////////////////////////////////////////////////////
-
-void SkColorMatrixFilter::flatten(SkWriteBuffer& buffer) const {
-    SkASSERT(sizeof(fMatrix.fMat)/sizeof(SkScalar) == 20);
-    buffer.writeScalarArray(fMatrix.fMat, 20);
-}
-
-SkFlattenable* SkColorMatrixFilter::CreateProc(SkReadBuffer& buffer) {
-    SkColorMatrix matrix;
-    if (buffer.readScalarArray(matrix.fMat, 20)) {
-        return Create(matrix);
-    }
-    return nullptr;
-}
-
-bool SkColorMatrixFilter::asColorMatrix(SkScalar matrix[20]) const {
-    if (matrix) {
-        memcpy(matrix, fMatrix.fMat, 20 * sizeof(SkScalar));
-    }
-    return true;
-}
-
-SkColorFilter* SkColorMatrixFilter::newComposed(const SkColorFilter* innerFilter) const {
-    SkScalar innerMatrix[20];
-    if (innerFilter->asColorMatrix(innerMatrix) && !SkColorMatrix::NeedsClamping(innerMatrix)) {
-        SkScalar concat[20];
-        SkColorMatrix::SetConcat(concat, fMatrix.fMat, innerMatrix);
-        return SkColorMatrixFilter::Create(concat);
-    }
-    return nullptr;
-}
-
-#if SK_SUPPORT_GPU
-#include "GrFragmentProcessor.h"
-#include "GrInvariantOutput.h"
-#include "glsl/GrGLSLFragmentProcessor.h"
-#include "glsl/GrGLSLFragmentShaderBuilder.h"
-#include "glsl/GrGLSLProgramDataManager.h"
-#include "glsl/GrGLSLUniformHandler.h"
-
-class ColorMatrixEffect : public GrFragmentProcessor {
-public:
-    static const GrFragmentProcessor* Create(const SkColorMatrix& matrix) {
-        return new ColorMatrixEffect(matrix);
-    }
-
-    const char* name() const override { return "Color Matrix"; }
-
-    GR_DECLARE_FRAGMENT_PROCESSOR_TEST;
-
-    class GLSLProcessor : public GrGLSLFragmentProcessor {
-    public:
-        // this class always generates the same code.
-        static void GenKey(const GrProcessor&, const GrGLSLCaps&, GrProcessorKeyBuilder* b) {}
-
-        GLSLProcessor(const GrProcessor&) {}
-
-        virtual void emitCode(EmitArgs& args) override {
-            GrGLSLUniformHandler* uniformHandler = args.fUniformHandler;
-            fMatrixHandle = uniformHandler->addUniform(GrGLSLUniformHandler::kFragment_Visibility,
-                                                       kMat44f_GrSLType, kDefault_GrSLPrecision,
-                                                       "ColorMatrix");
-            fVectorHandle = uniformHandler->addUniform(GrGLSLUniformHandler::kFragment_Visibility,
-                                                       kVec4f_GrSLType, kDefault_GrSLPrecision,
-                                                       "ColorMatrixVector");
-
-            if (nullptr == args.fInputColor) {
-                // could optimize this case, but we aren't for now.
-                args.fInputColor = "vec4(1)";
-            }
-            GrGLSLFragmentBuilder* fragBuilder = args.fFragBuilder;
-            // The max() is to guard against 0 / 0 during unpremul when the incoming color is
-            // transparent black.
-            fragBuilder->codeAppendf("\tfloat nonZeroAlpha = max(%s.a, 0.00001);\n",
-                                     args.fInputColor);
-            fragBuilder->codeAppendf("\t%s = %s * vec4(%s.rgb / nonZeroAlpha, nonZeroAlpha) + %s;\n",
-                                     args.fOutputColor,
-                                     uniformHandler->getUniformCStr(fMatrixHandle),
-                                     args.fInputColor,
-                                     uniformHandler->getUniformCStr(fVectorHandle));
-            fragBuilder->codeAppendf("\t%s = clamp(%s, 0.0, 1.0);\n",
-                                     args.fOutputColor, args.fOutputColor);
-            fragBuilder->codeAppendf("\t%s.rgb *= %s.a;\n", args.fOutputColor, args.fOutputColor);
-        }
-
-    protected:
-        virtual void onSetData(const GrGLSLProgramDataManager& uniManager,
-                               const GrProcessor& proc) override {
-            const ColorMatrixEffect& cme = proc.cast<ColorMatrixEffect>();
-            const float* m = cme.fMatrix.fMat;
-            // The GL matrix is transposed from SkColorMatrix.
-            float mt[]  = {
-                m[0], m[5], m[10], m[15],
-                m[1], m[6], m[11], m[16],
-                m[2], m[7], m[12], m[17],
-                m[3], m[8], m[13], m[18],
-            };
-            static const float kScale = 1.0f / 255.0f;
-            float vec[] = {
-                m[4] * kScale, m[9] * kScale, m[14] * kScale, m[19] * kScale,
-            };
-            uniManager.setMatrix4fv(fMatrixHandle, 1, mt);
-            uniManager.set4fv(fVectorHandle, 1, vec);
-        }
-
-    private:
-        GrGLSLProgramDataManager::UniformHandle fMatrixHandle;
-        GrGLSLProgramDataManager::UniformHandle fVectorHandle;
-
-        typedef GrGLSLFragmentProcessor INHERITED;
-    };
-
-private:
-    ColorMatrixEffect(const SkColorMatrix& matrix) : fMatrix(matrix) {
-        this->initClassID<ColorMatrixEffect>();
-    }
-
-    GrGLSLFragmentProcessor* onCreateGLSLInstance() const override {
-        return new GLSLProcessor(*this);
-    }
-
-    virtual void onGetGLSLProcessorKey(const GrGLSLCaps& caps,
-                                       GrProcessorKeyBuilder* b) const override {
-        GLSLProcessor::GenKey(*this, caps, b);
-    }
-
-    bool onIsEqual(const GrFragmentProcessor& s) const override {
-        const ColorMatrixEffect& cme = s.cast<ColorMatrixEffect>();
-        return cme.fMatrix == fMatrix;
-    }
-
-    void onComputeInvariantOutput(GrInvariantOutput* inout) const override {
-        // We only bother to check whether the alpha channel will be constant. If SkColorMatrix had
-        // type flags it might be worth checking the other components.
-
-        // The matrix is defined such the 4th row determines the output alpha. The first four
-        // columns of that row multiply the input r, g, b, and a, respectively, and the last column
-        // is the "translation".
-        static const uint32_t kRGBAFlags[] = {
-            kR_GrColorComponentFlag,
-            kG_GrColorComponentFlag,
-            kB_GrColorComponentFlag,
-            kA_GrColorComponentFlag
-        };
-        static const int kShifts[] = {
-            GrColor_SHIFT_R, GrColor_SHIFT_G, GrColor_SHIFT_B, GrColor_SHIFT_A,
-        };
-        enum {
-            kAlphaRowStartIdx = 15,
-            kAlphaRowTranslateIdx = 19,
-        };
-
-        SkScalar outputA = 0;
-        for (int i = 0; i < 4; ++i) {
-            // If any relevant component of the color to be passed through the matrix is non-const
-            // then we can't know the final result.
-            if (0 != fMatrix.fMat[kAlphaRowStartIdx + i]) {
-                if (!(inout->validFlags() & kRGBAFlags[i])) {
-                    inout->setToUnknown(GrInvariantOutput::kWill_ReadInput);
-                    return;
-                } else {
-                    uint32_t component = (inout->color() >> kShifts[i]) & 0xFF;
-                    outputA += fMatrix.fMat[kAlphaRowStartIdx + i] * component;
-                }
-            }
-        }
-        outputA += fMatrix.fMat[kAlphaRowTranslateIdx];
-        // We pin the color to [0,1]. This would happen to the *final* color output from the frag
-        // shader but currently the effect does not pin its own output. So in the case of over/
-        // underflow this may deviate from the actual result. Maybe the effect should pin its
-        // result if the matrix could over/underflow for any component?
-        inout->setToOther(kA_GrColorComponentFlag,
-                          static_cast<uint8_t>(SkScalarPin(outputA, 0, 255)) << GrColor_SHIFT_A,
-                          GrInvariantOutput::kWill_ReadInput);
-    }
-
-    SkColorMatrix fMatrix;
-
-    typedef GrFragmentProcessor INHERITED;
-};
-
-GR_DEFINE_FRAGMENT_PROCESSOR_TEST(ColorMatrixEffect);
-
-const GrFragmentProcessor* ColorMatrixEffect::TestCreate(GrProcessorTestData* d) {
-    SkColorMatrix colorMatrix;
-    for (size_t i = 0; i < SK_ARRAY_COUNT(colorMatrix.fMat); ++i) {
-        colorMatrix.fMat[i] = d->fRandom->nextSScalar1();
-    }
-    return ColorMatrixEffect::Create(colorMatrix);
-}
-
-const GrFragmentProcessor* SkColorMatrixFilter::asFragmentProcessor(GrContext*) const {
-    return ColorMatrixEffect::Create(fMatrix);
-}
-
-#endif
-
-#ifndef SK_IGNORE_TO_STRING
-void SkColorMatrixFilter::toString(SkString* str) const {
-    str->append("SkColorMatrixFilter: ");
-
-    str->append("matrix: (");
-    for (int i = 0; i < 20; ++i) {
-        str->appendScalar(fMatrix.fMat[i]);
-        if (i < 19) {
-            str->append(", ");
-        }
-    }
-    str->append(")");
-}
-#endif
-
-///////////////////////////////////////////////////////////////////////////////
 
 static SkScalar byte_to_scale(U8CPU byte) {
     if (0xFF == byte) {
diff --git a/src/ports/SkGlobalInitialization_default.cpp b/src/ports/SkGlobalInitialization_default.cpp
index 4c06817d4a..a2b91b4a93 100644
--- a/src/ports/SkGlobalInitialization_default.cpp
+++ b/src/ports/SkGlobalInitialization_default.cpp
@@ -17,7 +17,7 @@
 #include "SkBlurMaskFilter.h"
 #include "SkColorCubeFilter.h"
 #include "SkColorFilterImageFilter.h"
-#include "SkColorMatrixFilter.h"
+#include "SkColorMatrixFilterRowMajor255.h"
 #include "SkComposeImageFilter.h"
 #include "SkCornerPathEffect.h"
 #include "SkDashPathEffect.h"
@@ -70,7 +70,7 @@ void SkFlattenable::PrivateInitializer::InitEffects() {
 
     // ColorFilter
     SK_DEFINE_FLATTENABLE_REGISTRAR_ENTRY(SkColorCubeFilter)
-    SK_DEFINE_FLATTENABLE_REGISTRAR_ENTRY(SkColorMatrixFilter)
+    SK_DEFINE_FLATTENABLE_REGISTRAR_ENTRY(SkColorMatrixFilterRowMajor255)
     SK_DEFINE_FLATTENABLE_REGISTRAR_ENTRY(SkLumaColorFilter)
     SkAlphaThresholdFilter::InitializeFlattenables();
     SkArithmeticMode::InitializeFlattenables();
diff --git a/tests/ImageFilterTest.cpp b/tests/ImageFilterTest.cpp
index c9ef55944d..68a95fcb17 100644
--- a/tests/ImageFilterTest.cpp
+++ b/tests/ImageFilterTest.cpp
@@ -1,3 +1,4 @@
+
 /*
  * Copyright 2013 Google Inc.
  *