Added SkComposeShader GPU implementation

src/core/SkComposeShader.cpp

Index: src/core/SkComposeShader.cpp
diff --git a/src/core/SkComposeShader.cpp b/src/core/SkComposeShader.cpp
index c16d7ffc578f352f948d432934c68c41851c6623..a188e7908c549d5af1c7faa54030187bf52d10d2 100644
--- a/src/core/SkComposeShader.cpp
+++ b/src/core/SkComposeShader.cpp
@@ -194,6 +194,171 @@ void SkComposeShader::ComposeShaderContext::shadeSpan(int x, int y, SkPMColor re
     }
 }
 
+#if SK_SUPPORT_GPU
+
+#include "SkGr.h"
+#include "GrProcessor.h"
+#include "gl/GrGLBlend.h"
+#include "gl/builders/GrGLProgramBuilder.h"
+
+/////////////////////////////////////////////////////////////////////
+
+class GrComposeEffect : public GrFragmentProcessor {
+public:
+
+    static GrFragmentProcessor* Create(GrFragmentProcessor* fpA, GrFragmentProcessor* fpB,
+                                       SkXfermode::Mode mode) {
+        return SkNEW_ARGS(GrComposeEffect, (fpA, fpB, mode));
+    }
+    const char* name() const override {return fName.c_str(); }
+    void onGetGLProcessorKey(const GrGLSLCaps& caps, GrProcessorKeyBuilder* b) const override;
+
+    SkXfermode::Mode getMode() const { return fMode; }
+    int getShaderAChildIndex() const { return fShaderAChildIndex; }
+    int getShaderBChildIndex() const { return fShaderBChildIndex; }
+
+protected:
+    bool onIsEqual(const GrFragmentProcessor&) const override;
+    void onComputeInvariantOutput(GrInvariantOutput* inout) const override;
+
+private:
+    GrComposeEffect(GrFragmentProcessor* fpA, GrFragmentProcessor* fpB, SkXfermode::Mode mode)
+        : fMode(mode) {
+        this->initClassID<GrComposeEffect>();
+        fShaderAChildIndex = this->registerChildProcessor(fpA);
+        fShaderBChildIndex = this->registerChildProcessor(fpB);
+        fName.printf("ComposeShader (Mode: %s)", SkXfermode::ModeName(fMode));

[bsalomon, 2015/08/27 20:12:24]

Maybe we should init this in name()? I believe it is only called when building
the shader.

[wangyix, 2015/08/31 14:10:05]

I'm not sure what you mean by this.  On another note, it seems all other
shaders' name() function just returns the name of the shader and nothing more,
so maybe I could change name() to just 'return "ComposeShader";' for
consistency?

[bsalomon, 2015/08/31 14:48:35]

Let's go with that. I just don't want the constructor to pay for printf() when
the name string is only really used for shader debugging.

[wangyix, 2015/08/31 18:22:09]

Done.

+    }
+
+    GrGLFragmentProcessor* onCreateGLInstance() const override;
+
+    SkString fName;
+    int fShaderAChildIndex;
+    int fShaderBChildIndex;
+    SkXfermode::Mode fMode;
+
+    typedef GrFragmentProcessor INHERITED;
+};
+
+/////////////////////////////////////////////////////////////////////
+
+class GrGLComposeEffect : public GrGLFragmentProcessor {
+public:
+    GrGLComposeEffect(const GrProcessor& processor) {
+        const GrComposeEffect& cs = processor.cast<GrComposeEffect>();
+        fShaderAChildIndex = cs.getShaderAChildIndex();
+        fShaderBChildIndex = cs.getShaderBChildIndex();
+        fMode = cs.getMode();
+    }
+
+    void emitCode(EmitArgs&) override;
+
+private:
+    int fShaderAChildIndex;
+    int fShaderBChildIndex;
+    SkXfermode::Mode fMode;
+
+    typedef GrGLFragmentProcessor INHERITED;
+};
+
+bool GrComposeEffect::onIsEqual(const GrFragmentProcessor& other) const {
+    const GrComposeEffect& cs = other.cast<GrComposeEffect>();
+    return fMode == cs.fMode;
+}
+
+void GrComposeEffect::onComputeInvariantOutput(GrInvariantOutput* inout) const {
+    inout->setToUnknown(GrInvariantOutput::kWill_ReadInput);
+}
+
+void GrComposeEffect::onGetGLProcessorKey(const GrGLSLCaps& caps, GrProcessorKeyBuilder* b) const {
+    b->add32(fMode);
+}
+
+GrGLFragmentProcessor* GrComposeEffect::onCreateGLInstance() const{
+    return SkNEW_ARGS(GrGLComposeEffect, (*this));
+}
+
+/////////////////////////////////////////////////////////////////////
+
+void GrGLComposeEffect::emitCode(EmitArgs& args) {
+
+    GrGLFragmentBuilder* fsBuilder = args.fBuilder->getFragmentShaderBuilder();
+
+    // Store alpha of input color and un-premultiply the input color by its alpha. We will
+    // re-multiply by this alpha after blending the output colors of the two child procs.
+    // This is because we don't want the paint's alpha to affect either child proc's output
+    // before the blend; we want to apply the paint's alpha AFTER the blend. This mirrors the
+    // software implementation of SkComposeShader.
+    SkString inputAlpha("inputAlpha");
+    fsBuilder->codeAppendf("float %s = %s.a;\n", inputAlpha.c_str(), args.fInputColor);
+    fsBuilder->codeAppendf("%s /= %s.a;\n", args.fInputColor, args.fInputColor);
+
+    // emit the code of the two child shaders
+    SkString mangledOutputColorA;
+    this->emitChild(fShaderAChildIndex, args.fInputColor, &mangledOutputColorA, args);
+    SkString mangledOutputColorB;
+    this->emitChild(fShaderBChildIndex, args.fInputColor, &mangledOutputColorB, args);
+
+    // emit blend code
+    fsBuilder->codeAppend("{\n");
+    fsBuilder->codeAppendf("// Compose Xfer Mode: %s\n", SkXfermode::ModeName(fMode));
+    GrGLBlend::AppendPorterDuffBlend(fsBuilder, mangledOutputColorB.c_str(),
+                                     mangledOutputColorA.c_str(), args.fOutputColor, fMode);
+    fsBuilder->codeAppend("}\n");
+
+    // re-multiply the output color by the input color's alpha
+    fsBuilder->codeAppendf("%s *= %s;\n", args.fOutputColor, inputAlpha.c_str());
+}
+
+/////////////////////////////////////////////////////////////////////
+
+bool SkComposeShader::asFragmentProcessor(GrContext* context, const SkPaint& paint,
+                                          const SkMatrix& viewM, const SkMatrix* localMatrix,
+                                          GrColor* paintColor,
+                                          GrProcessorDataManager* procDataManager,
+                                          GrFragmentProcessor** fp) const {
+    // NULL fMode implies src-over
+    SkXfermode::Mode mode = SkXfermode::Mode::kSrcOver_Mode;
+    if (fMode) {
+        // Fragment processor will only support coefficient modes that aren't kClear_Mode.
+        // This is because GrGLBlend::AppendPorterDuffBlend(), which emits the blend code in the
+        // shader, only supports those xfer modes.
+        SkXfermode::Coeff srcCoeff, dstCoeff;
+        if (!fMode->asMode(&mode) || !SkXfermode::ModeAsCoeff(mode, &srcCoeff, &dstCoeff)
+            || mode == SkXfermode::Mode::kClear_Mode) {

[wangyix, 2015/08/31 14:10:05]

What should the behavior be for kClear_Mode under the new asFragmentProcessor()
signature?  Should NULL be returned for asFragmentProcessor()?

Also, if the mode is kSrc_Mode or kDst_Mode, should I just return the src or dst
child's fragment processor directly instead of a GrComposeEffect?

[egdaniel, 2015/08/31 14:43:05]

Did AppendPorterDuffBlend() ever get fixed to support clear? That definitely
should be done first.

As for optimizations that can be done here, I don't think I would return NULL
for the FP here on clear. If you have a clear Xfermode then what you will end up
with is all trans black for each pixel. I don't think returning NULL would give
you the same output? Maybe a const color processor that does trans black?

The src and dst could also be optimized as you said. However personally I say
lets make sure it works without any of these specific mode optimizations first
then add those in. 

[bsalomon, 2015/08/31 14:48:35]

Hm... I think making kSrc, kDst, kClear FPs is maybe a bad idea that should
always fail. If we're doing that we can probably use a simpler shader that will
result in less duplication. (e.g. use a const color fp for clear). It might be
better to make it assert in the append so that the author sees the issue and
writes the correct optimization.

The clear optimization should probably just happen in the ComposeEffect's
factory function, since every caller (when this is moved out of this cpp file)
can benefit from it. The kSrc and kDst optimization probably should happen in
this asFP function since we can skip making either the dst or src FP entirely.

[wangyix, 2015/08/31 15:59:24]

If the clear optimization happens in GrComposeEffect::Create(), then we can't
skip making the src and dst FPs.

I also feel that SkComposeShader shouldn't support kClear, kSrc, kDst in the
first place.  Maybe add warnings or asserts for those 3 cases, but still have
code that optimizes for them in SkComposeShader::asFragmentProcessor() so those
modes still work in release mode?

+            return false;
+        }
+    }
+    GrFragmentProcessor* fpA = NULL;
+    if (!fShaderA->asFragmentProcessor(context, paint, viewM, localMatrix, paintColor,
+        procDataManager, &fpA) || !fpA) {
+        return false;
+    }
+    GrFragmentProcessor* fpB = NULL;
+    if (!fShaderB->asFragmentProcessor(context, paint, viewM, localMatrix, paintColor,
+        procDataManager, &fpB) || !fpB) {
+        fpA->unref();
+        return false;
+    }
+
+    *fp = GrComposeEffect::Create(fpA, fpB, mode);
+    fpA->unref();
+    fpB->unref();
+
+    return true;
+}
+
+
+
+#else

[bsalomon, 2015/08/28 16:15:30]

Let's just put the function decl in the header behind a #if SK_SUPPORT_GPU. (I
got rid of all these bogus impls in the the refactor of asFragmentProcessor)

[wangyix, 2015/08/31 18:22:09]

Done.

+bool SkComposeShader::asFragmentProcessor(GrContext*, const SkPaint&, const SkMatrix&,
+                                          const SkMatrix*, GrColor*, GrProcessorDataManager*,
+                                          GrFragmentProcessor**) const {
+    SkDEBUGFAIL("Should not call in GPU-less build");
+    return false;
+}
+#endif
+
 #ifndef SK_IGNORE_TO_STRING
 void SkComposeShader::toString(SkString* str) const {
     str->append("SkComposeShader: (");