src/gpu/ganesh/effects/GrSkSLFP.cpp

/*
 * Copyright 2018 Google Inc.
 *
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE file.
 */

#include "src/gpu/ganesh/effects/GrSkSLFP.h"

#include "include/core/SkAlphaType.h"
#include "include/core/SkColor.h"
#include "include/core/SkData.h"
#include "include/core/SkString.h"
#include "include/core/SkSurfaceProps.h"
#include "include/effects/SkOverdrawColorFilter.h"
#include "include/effects/SkRuntimeEffect.h"
#include "include/private/SkSLSampleUsage.h"
#include "include/private/base/SkMalloc.h"
#include "include/private/base/SkTo.h"
#include "include/private/gpu/ganesh/GrTypesPriv.h"
#include "src/base/SkArenaAlloc.h"
#include "src/base/SkRandom.h"
#include "src/core/SkColorSpacePriv.h"
#include "src/core/SkRasterPipeline.h"
#include "src/core/SkRasterPipelineOpContexts.h"
#include "src/core/SkRasterPipelineOpList.h"
#include "src/core/SkRuntimeEffectPriv.h"
#include "src/core/SkSLTypeShared.h"
#include "src/gpu/KeyBuilder.h"
#include "src/gpu/ganesh/GrColorInfo.h"
#include "src/gpu/ganesh/GrColorSpaceXform.h"
#include "src/gpu/ganesh/GrFragmentProcessors.h"
#include "src/gpu/ganesh/GrShaderVar.h"
#include "src/gpu/ganesh/glsl/GrGLSLFragmentShaderBuilder.h"
#include "src/gpu/ganesh/glsl/GrGLSLUniformHandler.h"
#include "src/sksl/SkSLString.h"
#include "src/sksl/SkSLUtil.h"
#include "src/sksl/codegen/SkSLPipelineStageCodeGenerator.h"
#include "src/sksl/codegen/SkSLRasterPipelineBuilder.h"
#include "src/sksl/ir/SkSLProgram.h"
#include "src/sksl/ir/SkSLType.h"
#include "src/sksl/ir/SkSLVarDeclarations.h"
#include "src/sksl/ir/SkSLVariable.h"

#include <algorithm>

namespace SkSL { class Context; }
struct GrShaderCaps;

class GrSkSLFP::Impl : public ProgramImpl {
public:
    void emitCode(EmitArgs& args) override {
        const GrSkSLFP& fp            = args.fFp.cast<GrSkSLFP>();
        const SkSL::Program& program  = *fp.fEffect->fBaseProgram;

        class FPCallbacks : public SkSL::PipelineStage::Callbacks {
        public:
            FPCallbacks(Impl* self,
                        EmitArgs& args,
                        const char* inputColor,
                        const SkSL::Context& context,
                        const uint8_t* uniformData,
                        const Specialized* specialized)
                    : fSelf(self)
                    , fArgs(args)
                    , fInputColor(inputColor)
                    , fContext(context)
                    , fUniformData(uniformData)
                    , fSpecialized(specialized) {}

            std::string declareUniform(const SkSL::VarDeclaration* decl) override {
                const SkSL::Variable* var = decl->var();
                if (var->type().isOpaque()) {
                    // Nothing to do. The only opaque types we should see are children, and those
                    // are handled specially.
                    SkASSERT(var->type().isEffectChild());
                    return std::string(var->name());
                }

                const SkSL::Type* type = &var->type();
                size_t sizeInBytes = type->slotCount() * sizeof(float);
                const float* floatData = reinterpret_cast<const float*>(fUniformData);
                const int* intData = reinterpret_cast<const int*>(fUniformData);
                fUniformData += sizeInBytes;

                bool isArray = false;
                if (type->isArray()) {
                    type = &type->componentType();
                    isArray = true;
                }

                SkSLType gpuType;
                SkAssertResult(SkSL::type_to_sksltype(fContext, *type, &gpuType));

                if (*fSpecialized++ == Specialized::kYes) {
                    SkASSERTF(!isArray, "specializing array uniforms is not allowed");
                    std::string value = SkSLTypeString(gpuType);
                    value.append("(");

                    bool isFloat = SkSLTypeIsFloatType(gpuType);
                    size_t slots = type->slotCount();
                    for (size_t i = 0; i < slots; ++i) {
                        value.append(isFloat ? skstd::to_string(floatData[i])
                                             : std::to_string(intData[i]));
                        value.append(",");
                    }
                    value.back() = ')';
                    return value;
                }

                const char* uniformName = nullptr;
                auto handle =
                        fArgs.fUniformHandler->addUniformArray(&fArgs.fFp.cast<GrSkSLFP>(),
                                                               kFragment_GrShaderFlag,
                                                               gpuType,
                                                               SkString(var->name()).c_str(),
                                                               isArray ? var->type().columns() : 0,
                                                               &uniformName);
                fSelf->fUniformHandles.push_back(handle);
                return std::string(uniformName);
            }

            std::string getMangledName(const char* name) override {
                return std::string(fArgs.fFragBuilder->getMangledFunctionName(name).c_str());
            }

            void defineFunction(const char* decl, const char* body, bool isMain) override {
                if (isMain) {
                    fArgs.fFragBuilder->codeAppend(body);
                } else {
                    fArgs.fFragBuilder->emitFunction(decl, body);
                }
            }

            void declareFunction(const char* decl) override {
                fArgs.fFragBuilder->emitFunctionPrototype(decl);
            }

            void defineStruct(const char* definition) override {
                fArgs.fFragBuilder->definitionAppend(definition);
            }

            void declareGlobal(const char* declaration) override {
                fArgs.fFragBuilder->definitionAppend(declaration);
            }

            std::string sampleShader(int index, std::string coords) override {
                // If the child was sampled using the coords passed to main (and they are never
                // modified), then we will have marked the child as PassThrough. The code generator
                // doesn't know that, and still supplies coords. Inside invokeChild, we assert that
                // any coords passed for a PassThrough child match args.fSampleCoords exactly.
                //
                // Normally, this is valid. Here, we *copied* the sample coords to a local variable
                // (so that they're mutable in the runtime effect SkSL). Thus, the coords string we
                // get here is the name of the local copy, and fSampleCoords still points to the
                // unmodified original (which might be a varying, for example).
                // To prevent the assert, we pass the empty string in this case. Note that for
                // children sampled like this, invokeChild doesn't even use the coords parameter,
                // except for that assert.
                const GrFragmentProcessor* child = fArgs.fFp.childProcessor(index);
                if (child && child->sampleUsage().isPassThrough()) {
                    coords.clear();
                }
                return child ? std::string(fSelf->invokeChild(index, fInputColor, fArgs, coords)
                                                   .c_str())
                             : std::string("half4(0)");
            }

            std::string sampleColorFilter(int index, std::string color) override {
                return std::string(fSelf->invokeChild(index,
                                                      color.empty() ? fInputColor : color.c_str(),
                                                      fArgs)
                                           .c_str());
            }

            std::string sampleBlender(int index, std::string src, std::string dst) override {
                if (!fSelf->childProcessor(index)) {
                    return SkSL::String::printf("blend_src_over(%s, %s)", src.c_str(), dst.c_str());
                }
                return std::string(
                        fSelf->invokeChild(index, src.c_str(), dst.c_str(), fArgs).c_str());
            }

            // These intrinsics take and return 3-component vectors, but child FPs operate on
            // 4-component vectors. We use swizzles here to paper over the difference.
            std::string toLinearSrgb(std::string color) override {
                const GrSkSLFP& fp = fArgs.fFp.cast<GrSkSLFP>();
                if (fp.fToLinearSrgbChildIndex < 0) {
                    return color;
                }
                color = SkSL::String::printf("(%s).rgb1", color.c_str());
                SkString xformedColor = fSelf->invokeChild(
                        fp.fToLinearSrgbChildIndex, color.c_str(), fArgs);
                return SkSL::String::printf("(%s).rgb", xformedColor.c_str());
            }

            std::string fromLinearSrgb(std::string color) override {
                const GrSkSLFP& fp = fArgs.fFp.cast<GrSkSLFP>();
                if (fp.fFromLinearSrgbChildIndex < 0) {
                    return color;
                }
                color = SkSL::String::printf("(%s).rgb1", color.c_str());
                SkString xformedColor = fSelf->invokeChild(
                        fp.fFromLinearSrgbChildIndex, color.c_str(), fArgs);
                return SkSL::String::printf("(%s).rgb", xformedColor.c_str());
            }

            Impl*                         fSelf;
            EmitArgs&                     fArgs;
            const char*                   fInputColor;
            const SkSL::Context&          fContext;
            const uint8_t*                fUniformData;
            const Specialized*            fSpecialized;
            int                           fUniformIndex = 0;
        };

        // If we have an input child, we invoke it now, and make the result of that be the "input
        // color" for all other purposes later (eg, the default passed via sample calls, etc.)
        if (fp.fInputChildIndex >= 0) {
            args.fFragBuilder->codeAppendf("%s = %s;\n",
                                           args.fInputColor,
                                           this->invokeChild(fp.fInputChildIndex, args).c_str());
        }

        if (fp.fEffect->allowBlender()) {
            // If we have an dest-color child, we invoke it now, and make the result of that be the
            // "dest color" for all other purposes later.
            if (fp.fDestColorChildIndex >= 0) {
                args.fFragBuilder->codeAppendf(
                        "%s = %s;\n",
                        args.fDestColor,
                        this->invokeChild(fp.fDestColorChildIndex, args.fDestColor, args).c_str());
            }
        } else {
            // We're not making a blender, so we don't expect a dest-color child FP to exist.
            SkASSERT(fp.fDestColorChildIndex < 0);
        }

        // Snap off a global copy of the input color at the start of main. We need this when
        // we call child processors (particularly from helper functions, which can't "see" the
        // parameter to main). Even from within main, if the code mutates the parameter, calls to
        // sample should still be passing the original color (by default).
        SkString inputColorName;
        if (fp.fEffect->samplesOutsideMain()) {
            GrShaderVar inputColorCopy(args.fFragBuilder->getMangledFunctionName("inColor"),
                                       SkSLType::kHalf4);
            args.fFragBuilder->declareGlobal(inputColorCopy);
            inputColorName = inputColorCopy.getName();
            args.fFragBuilder->codeAppendf("%s = %s;\n", inputColorName.c_str(), args.fInputColor);
        } else {
            inputColorName = args.fFragBuilder->newTmpVarName("inColor");
            args.fFragBuilder->codeAppendf(
                    "half4 %s = %s;\n", inputColorName.c_str(), args.fInputColor);
        }

        // Copy the incoming coords to a local variable. Code in main might modify the coords
        // parameter. fSampleCoord could be a varying, so writes to it would be illegal.
        const char* coords = "float2(0)";
        SkString coordsVarName;
        if (fp.usesSampleCoordsDirectly()) {
            coordsVarName = args.fFragBuilder->newTmpVarName("coords");
            coords = coordsVarName.c_str();
            args.fFragBuilder->codeAppendf("float2 %s = %s;\n", coords, args.fSampleCoord);
        }

        FPCallbacks callbacks(this,
                              args,
                              inputColorName.c_str(),
                              *program.fContext,
                              fp.uniformData(),
                              fp.specialized());
        SkSL::PipelineStage::ConvertProgram(
                program, coords, args.fInputColor, args.fDestColor, &callbacks);
    }

private:
    void onSetData(const GrGLSLProgramDataManager& pdman,
                   const GrFragmentProcessor& _proc) override {
        const GrSkSLFP& outer = _proc.cast<GrSkSLFP>();
        pdman.setRuntimeEffectUniforms(outer.fEffect->uniforms(),
                                       SkSpan(fUniformHandles),
                                       SkSpan(outer.specialized(), outer.uniformCount()),
                                       outer.uniformData());
    }

    std::vector<UniformHandle> fUniformHandles;
};

std::unique_ptr<GrSkSLFP> GrSkSLFP::MakeWithData(
        sk_sp<SkRuntimeEffect> effect,
        const char* name,
        sk_sp<SkColorSpace> dstColorSpace,
        std::unique_ptr<GrFragmentProcessor> inputFP,
        std::unique_ptr<GrFragmentProcessor> destColorFP,
        const sk_sp<const SkData>& uniforms,
        SkSpan<std::unique_ptr<GrFragmentProcessor>> childFPs) {
    if (uniforms->size() != effect->uniformSize()) {
        return nullptr;
    }
    size_t uniformSize = uniforms->size();
    size_t specializedSize = effect->uniforms().size() * sizeof(Specialized);
    std::unique_ptr<GrSkSLFP> fp(new (uniformSize + specializedSize)
                                         GrSkSLFP(std::move(effect), name, OptFlags::kNone));
    sk_careful_memcpy(fp->uniformData(), uniforms->data(), uniformSize);
    for (auto& childFP : childFPs) {
        fp->addChild(std::move(childFP), /*mergeOptFlags=*/true);
    }
    if (inputFP) {
        fp->setInput(std::move(inputFP));
    }
    if (destColorFP) {
        fp->setDestColorFP(std::move(destColorFP));
    }
    if (fp->fEffect->usesColorTransform() && dstColorSpace) {
        fp->addColorTransformChildren(dstColorSpace.get());
    }
    return fp;
}

GrFragmentProcessor::OptimizationFlags GrSkSLFP::DetermineOptimizationFlags(
        OptFlags of, SkRuntimeEffect* effect) {
    OptimizationFlags optFlags = static_cast<OptimizationFlags>(of);
    if (SkRuntimeEffectPriv::SupportsConstantOutputForConstantInput(effect)) {
        optFlags |= kConstantOutputForConstantInput_OptimizationFlag;
    }
    return optFlags;
}

GrSkSLFP::GrSkSLFP(sk_sp<SkRuntimeEffect> effect, const char* name, OptFlags optFlags)
        : INHERITED(kGrSkSLFP_ClassID, DetermineOptimizationFlags(optFlags, effect.get()))
        , fEffect(std::move(effect))
        , fName(name)
        , fUniformSize(SkToU32(fEffect->uniformSize())) {
    std::fill_n(this->specialized(), this->uniformCount(), Specialized::kNo);
    if (fEffect->usesSampleCoords()) {
        this->setUsesSampleCoordsDirectly();
    }
    if (fEffect->allowBlender()) {
        this->setIsBlendFunction();
    }
}

GrSkSLFP::GrSkSLFP(const GrSkSLFP& other)
        : INHERITED(other)
        , fEffect(other.fEffect)
        , fName(other.fName)
        , fUniformSize(other.fUniformSize)
        , fInputChildIndex(other.fInputChildIndex)
        , fDestColorChildIndex(other.fDestColorChildIndex)
        , fToLinearSrgbChildIndex(other.fToLinearSrgbChildIndex)
        , fFromLinearSrgbChildIndex(other.fFromLinearSrgbChildIndex) {
    std::copy_n(other.specialized(), this->uniformCount(), this->specialized());
    sk_careful_memcpy(this->uniformData(), other.uniformData(), fUniformSize);
}

void GrSkSLFP::addChild(std::unique_ptr<GrFragmentProcessor> child, bool mergeOptFlags) {
    SkASSERTF(fInputChildIndex == -1, "all addChild calls must happen before setInput");
    SkASSERTF(fDestColorChildIndex == -1, "all addChild calls must happen before setDestColorFP");
    int childIndex = this->numChildProcessors();
    SkASSERT((size_t)childIndex < fEffect->fSampleUsages.size());
    if (mergeOptFlags) {
        this->mergeOptimizationFlags(ProcessorOptimizationFlags(child.get()));
    }
    this->clearConstantOutputForConstantInputFlag();
    this->registerChild(std::move(child), fEffect->fSampleUsages[childIndex]);
}

void GrSkSLFP::setInput(std::unique_ptr<GrFragmentProcessor> input) {
    SkASSERTF(fInputChildIndex == -1, "setInput should not be called more than once");
    fInputChildIndex = this->numChildProcessors();
    SkASSERT((size_t)fInputChildIndex >= fEffect->fSampleUsages.size());
    this->mergeOptimizationFlags(ProcessorOptimizationFlags(input.get()));
    this->registerChild(std::move(input), SkSL::SampleUsage::PassThrough());
}

void GrSkSLFP::setDestColorFP(std::unique_ptr<GrFragmentProcessor> destColorFP) {
    SkASSERTF(fEffect->allowBlender(), "dest colors are only used by blend effects");
    SkASSERTF(fDestColorChildIndex == -1, "setDestColorFP should not be called more than once");
    fDestColorChildIndex = this->numChildProcessors();
    SkASSERT((size_t)fDestColorChildIndex >= fEffect->fSampleUsages.size());
    this->mergeOptimizationFlags(ProcessorOptimizationFlags(destColorFP.get()));
    this->registerChild(std::move(destColorFP), SkSL::SampleUsage::PassThrough());
}

void GrSkSLFP::addColorTransformChildren(SkColorSpace* dstColorSpace) {
    SkASSERTF(fToLinearSrgbChildIndex == -1 && fFromLinearSrgbChildIndex == -1,
              "addColorTransformChildren should not be called more than once");

    // We use child FPs for the color transforms. They're really just code snippets that get
    // invoked, but each one injects a collection of uniforms and helper functions. Doing it
    // this way leverages per-FP name mangling to avoid conflicts.
    auto workingToLinear = GrColorSpaceXformEffect::Make(nullptr,
                                                         dstColorSpace,
                                                         kUnpremul_SkAlphaType,
                                                         sk_srgb_linear_singleton(),
                                                         kUnpremul_SkAlphaType);
    auto linearToWorking = GrColorSpaceXformEffect::Make(nullptr,
                                                         sk_srgb_linear_singleton(),
                                                         kUnpremul_SkAlphaType,
                                                         dstColorSpace,
                                                         kUnpremul_SkAlphaType);

    fToLinearSrgbChildIndex = this->numChildProcessors();
    SkASSERT((size_t)fToLinearSrgbChildIndex >= fEffect->fSampleUsages.size());
    this->registerChild(std::move(workingToLinear), SkSL::SampleUsage::PassThrough());

    fFromLinearSrgbChildIndex = this->numChildProcessors();
    SkASSERT((size_t)fFromLinearSrgbChildIndex >= fEffect->fSampleUsages.size());
    this->registerChild(std::move(linearToWorking), SkSL::SampleUsage::PassThrough());
}

std::unique_ptr<GrFragmentProcessor::ProgramImpl> GrSkSLFP::onMakeProgramImpl() const {
    return std::make_unique<Impl>();
}

void GrSkSLFP::onAddToKey(const GrShaderCaps& caps, skgpu::KeyBuilder* b) const {
    // In the unlikely event of a hash collision, we also include the uniform size in the key.
    // That ensures that we will (at worst) use the wrong program, but one that expects the same
    // amount of uniform data.
    b->add32(fEffect->hash());
    b->add32(fUniformSize);

    const Specialized* specialized = this->specialized();
    const uint8_t* uniformData = this->uniformData();
    size_t uniformCount = this->uniformCount();
    auto iter = fEffect->uniforms().begin();

    for (size_t i = 0; i < uniformCount; ++i, ++iter) {
        bool specialize = specialized[i] == Specialized::kYes;
        b->addBool(specialize, "specialize");
        if (specialize) {
            b->addBytes(iter->sizeInBytes(), uniformData + iter->offset, iter->name);
        }
    }
}

bool GrSkSLFP::onIsEqual(const GrFragmentProcessor& other) const {
    const GrSkSLFP& sk = other.cast<GrSkSLFP>();
    const size_t specializedSize = this->uniformCount() * sizeof(Specialized);
    return fEffect->hash() == sk.fEffect->hash() &&
           this->uniformCount() == sk.uniformCount() &&
           fUniformSize == sk.fUniformSize &&
           !sk_careful_memcmp(this->uniformData(),
                              sk.uniformData(),
                              fUniformSize + specializedSize);
}

std::unique_ptr<GrFragmentProcessor> GrSkSLFP::clone() const {
    return std::unique_ptr<GrFragmentProcessor>(new (UniformPayloadSize(fEffect.get()))
                                                        GrSkSLFP(*this));
}

SkPMColor4f GrSkSLFP::constantOutputForConstantInput(const SkPMColor4f& inputColor) const {
    SkPMColor4f color = (fInputChildIndex >= 0)
            ? ConstantOutputForConstantInput(this->childProcessor(fInputChildIndex), inputColor)
            : inputColor;

    class ConstantOutputForConstantInput_SkRPCallbacks : public SkSL::RP::Callbacks {
    public:
        bool appendShader(int index) override {
           SkDEBUGFAIL("constant-output-for-constant-input unsupported when child shaders present");
           return false;
        }
        bool appendColorFilter(int index) override {
           SkDEBUGFAIL("constant-output-for-constant-input unsupported when child shaders present");
           return false;
        }
        bool appendBlender(int index) override {
           SkDEBUGFAIL("constant-output-for-constant-input unsupported when child shaders present");
           return false;
        }
        void toLinearSrgb(const void* color) override { /* identity color conversion */ }
        void fromLinearSrgb(const void* color) override { /* identity color conversion */ }
    };

    if (const SkSL::RP::Program* program = fEffect->getRPProgram(/*debugTrace=*/nullptr)) {
        // No color conversion is happening here, so we can use untransformed uniforms.
        SkSpan<const float> uniforms{reinterpret_cast<const float*>(this->uniformData()),
                                     fUniformSize / sizeof(float)};
        SkSTArenaAlloc<2048> alloc;  // sufficient for a tiny SkSL program
        SkRasterPipeline pipeline(&alloc);
        pipeline.appendConstantColor(&alloc, color.vec());
        ConstantOutputForConstantInput_SkRPCallbacks callbacks;
        if (program->appendStages(&pipeline, &alloc, &callbacks, uniforms)) {
            SkPMColor4f outputColor;
            SkRasterPipelineContexts::MemoryCtx outputCtx = {&outputColor, 0};
            pipeline.append(SkRasterPipelineOp::store_f32, &outputCtx);
            pipeline.run(0, 0, 1, 1);
            return outputColor;
        }
    }

    // We weren't able to run the Raster Pipeline program.
    return color;
}

/**************************************************************************************************/

GR_DEFINE_FRAGMENT_PROCESSOR_TEST(GrSkSLFP)

#if defined(GPU_TEST_UTILS)

std::unique_ptr<GrFragmentProcessor> GrSkSLFP::TestCreate(GrProcessorTestData* d) {
    SkColor colors[SkOverdrawColorFilter::kNumColors];
    for (SkColor& c : colors) {
        c = d->fRandom->nextU();
    }
    auto filter = SkOverdrawColorFilter::MakeWithSkColors(colors);
    SkSurfaceProps props; // default props for testing
    auto [success, fp] = GrFragmentProcessors::Make(
            d->surfaceDrawContext(), filter.get(), /*inputFP=*/nullptr, GrColorInfo{}, props);
    SkASSERT(success);
    return std::move(fp);
}

#endif