/src/skia/src/gpu/effects/GrMatrixConvolutionEffect.cpp

Source (jump to first uncovered line)
/*
 * Copyright 2014 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/effects/GrMatrixConvolutionEffect.h"

#include "include/private/SkHalf.h"
#include "src/gpu/GrDirectContextPriv.h"
#include "src/gpu/GrProxyProvider.h"
#include "src/gpu/GrRecordingContextPriv.h"
#include "src/gpu/GrTexture.h"
#include "src/gpu/GrTextureProxy.h"
#include "src/gpu/GrThreadSafeCache.h"
#include "src/gpu/SkGr.h"
#include "src/gpu/effects/GrTextureEffect.h"
#include "src/gpu/glsl/GrGLSLFragmentShaderBuilder.h"
#include "src/gpu/glsl/GrGLSLProgramDataManager.h"
#include "src/gpu/glsl/GrGLSLUniformHandler.h"

class GrMatrixConvolutionEffect::Impl : public ProgramImpl {
public:
    void emitCode(EmitArgs&) override;

private:
    void onSetData(const GrGLSLProgramDataManager&, const GrFragmentProcessor&) override;

    typedef GrGLSLProgramDataManager::UniformHandle UniformHandle;

    void emitKernelBlock(EmitArgs&, SkIPoint);

    UniformHandle               fKernelUni;
    UniformHandle               fKernelOffsetUni;
    UniformHandle               fGainUni;
    UniformHandle               fBiasUni;
    UniformHandle               fKernelBiasUni;

    using INHERITED = ProgramImpl;
};

GrMatrixConvolutionEffect::KernelWrapper::MakeResult
GrMatrixConvolutionEffect::KernelWrapper::Make(GrRecordingContext* rContext,
                                               SkISize size,
                                               const GrCaps& caps,
                                               const SkScalar* values) {
    if (!rContext || !values || size.isEmpty()) {
        return {};
    }

    const int length = size.area();
    // Small kernel -> just fill the array.
    KernelWrapper result(size);
    if (length <= kMaxUniformSize) {
        for (int i = 0; i < length; i++) {
            result.fArray[i] = SkScalarToFloat(values[i]);
        }
        return {result, nullptr};
    }

    BiasAndGain& scalableSampler = result.fBiasAndGain;
    bool useA16 =
        rContext->defaultBackendFormat(kA16_float_SkColorType, GrRenderable::kNo).isValid();
    SkScalar min = values[0];
    if (!useA16) {
        // Determine min and max values to figure out inner gain & bias.
        SkScalar max = values[0];
        for (int i = 1; i < length; i++) {
            if (values[i] < min) {
                min = values[i];
            }
            if (values[i] > max) {
                max = values[i];
            }
        }
        // Treat near-0 gain (i.e. box blur) as 1, and let the kernelBias
        // move everything up to the final value.
        const SkScalar computedGain = max - min;
        scalableSampler.fGain =
            SkScalarNearlyZero(computedGain) ? 1.0f : SkScalarToFloat(computedGain);
        // Inner bias is pre-inner-gain so we divide that out.
        scalableSampler.fBias = SkScalarToFloat(min) / scalableSampler.fGain;
    }

    // TODO: Pick cache or dont-cache based on observed perf.
    static constexpr bool kCacheKernelTexture = true;

    GrUniqueKey key;
    if (kCacheKernelTexture) {
        static const GrUniqueKey::Domain kDomain = GrUniqueKey::GenerateDomain();
        GrUniqueKey::Builder builder(&key, kDomain, length, "Matrix Convolution Kernel");
        // Texture cache key is the exact content of the kernel.
        static_assert(sizeof(float) == 4);
        for (int i = 0; i < length; i++) {
            builder[i] = *(const uint32_t*)&values[i];
        }
        builder.finish();
    }

    // Find or create a texture.
    auto threadSafeCache = rContext->priv().threadSafeCache();

    SkColorType colorType = useA16 ? kA16_float_SkColorType : kAlpha_8_SkColorType;

    GrSurfaceProxyView view;
    if (kCacheKernelTexture && (view = threadSafeCache->find(key))) {
        SkASSERT(view.origin() == kTopLeft_GrSurfaceOrigin);
        auto kernelFP = GrTextureEffect::Make(std::move(view), kUnknown_SkAlphaType);
        return {result, std::move(kernelFP)};
    }

    SkBitmap bm;
    auto info = SkImageInfo::Make({length, 1}, colorType, kPremul_SkAlphaType, nullptr);
    if (!bm.tryAllocPixels(info)) {
        return {};
    }
    for (int i = 0; i < length; i++) {
        if (useA16) {
            *bm.getAddr16(i, 0) = SkFloatToHalf(values[i]);
        } else {
            *bm.getAddr8(i, 0) =
                SkScalarRoundToInt((values[i] - min) / scalableSampler.fGain * 255);
        }
    }
    bm.setImmutable();

    view = std::get<0>(GrMakeUncachedBitmapProxyView(rContext, bm));
    if (!view) {
        return {};
    }

    if (kCacheKernelTexture) {
        view = threadSafeCache->add(key, view);
    }

    SkASSERT(view.origin() == kTopLeft_GrSurfaceOrigin);
    auto kernelFP = GrTextureEffect::Make(std::move(view), kUnknown_SkAlphaType);
    return {result, std::move(kernelFP)};
}

bool GrMatrixConvolutionEffect::KernelWrapper::operator==(const KernelWrapper& k) const {
    if (fSize != k.fSize) {
        return false;
    } else if (this->isSampled()) {
        return fBiasAndGain == k.fBiasAndGain;
    } else {
        return std::equal(fArray.begin(), fArray.begin() + fSize.area(), k.fArray.begin());
    }
}

bool GrMatrixConvolutionEffect::KernelWrapper::BiasAndGain::operator==(
                                                                const BiasAndGain& k) const {
    return fGain == k.fGain && fBias == k.fBias;
}

// For sampled kernels, emit a for loop that does all the kernel accumulation.
// For uniform kernels, emit a single iteration. Function is called repeatedly in a for loop.
// loc is ignored for sampled kernels.
void GrMatrixConvolutionEffect::Impl::emitKernelBlock(EmitArgs& args, SkIPoint loc) {
    const GrMatrixConvolutionEffect& mce = args.fFp.cast<GrMatrixConvolutionEffect>();
    GrGLSLFPFragmentBuilder* fragBuilder = args.fFragBuilder;
    GrGLSLUniformHandler* uniformHandler = args.fUniformHandler;
    int kernelWidth = mce.fKernel.size().width();
    int kernelHeight = mce.fKernel.size().height();
    int kernelArea = kernelWidth * kernelHeight;

    if (mce.fKernel.isSampled()) {
        fragBuilder->codeAppendf("for (int i = 0; i < %d; ++i)", (int)kernelArea);
    }

    GrGLSLShaderBuilder::ShaderBlock block(fragBuilder);

    fragBuilder->codeAppend("half k;");
    fragBuilder->codeAppend("half2 sourceOffset;");
    if (mce.fKernel.isSampled()) {
        const char* kernelBias = uniformHandler->getUniformCStr(fKernelBiasUni);
        SkString kernelSample = this->invokeChild(1, args, "float2(float(i) + 0.5, 0.5)");
        fragBuilder->codeAppendf("k = %s.w + %s;", kernelSample.c_str(), kernelBias);
        fragBuilder->codeAppendf("sourceOffset.y = floor(half(i) / %d);", kernelWidth);
        fragBuilder->codeAppendf("sourceOffset.x = half(i) - sourceOffset.y * %d;", kernelWidth);
    } else {
        fragBuilder->codeAppendf("sourceOffset = half2(%d, %d);", loc.x(), loc.y());
        int offset = loc.y() * kernelWidth + loc.x();
        const char* kernel = uniformHandler->getUniformCStr(fKernelUni);
        fragBuilder->codeAppendf("k = %s[%d][%d];", kernel, offset / 4, offset & 0x3);
    }

    auto sample = this->invokeChild(0, args, "coord + sourceOffset");
    fragBuilder->codeAppendf("half4 c = %s;", sample.c_str());
    if (!mce.fConvolveAlpha) {
        fragBuilder->codeAppend("c = unpremul(c);");
        fragBuilder->codeAppend("c.rgb = saturate(c.rgb);");
    }
    fragBuilder->codeAppend("sum += c * k;");
}

void GrMatrixConvolutionEffect::Impl::emitCode(EmitArgs& args) {
    const GrMatrixConvolutionEffect& mce = args.fFp.cast<GrMatrixConvolutionEffect>();

    int kernelWidth = mce.fKernel.size().width();
    int kernelHeight = mce.fKernel.size().height();

    int arrayCount = (kernelWidth * kernelHeight + 3) / 4;
    SkASSERT(4 * arrayCount >= kernelWidth * kernelHeight);

    GrGLSLUniformHandler* uniformHandler = args.fUniformHandler;
    if (mce.fKernel.isSampled()) {
        fKernelBiasUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag,
                                                    kHalf_GrSLType, "KernelBias");
    } else {
        fKernelUni = uniformHandler->addUniformArray(&mce, kFragment_GrShaderFlag,
                                                     kHalf4_GrSLType, "Kernel", arrayCount);
    }
    fKernelOffsetUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf2_GrSLType,
                                                  "KernelOffset");
    fGainUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf_GrSLType, "Gain");
    fBiasUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf_GrSLType, "Bias");

    const char* kernelOffset = uniformHandler->getUniformCStr(fKernelOffsetUni);
    const char* gain = uniformHandler->getUniformCStr(fGainUni);
    const char* bias = uniformHandler->getUniformCStr(fBiasUni);

    GrGLSLFPFragmentBuilder* fragBuilder = args.fFragBuilder;
    fragBuilder->codeAppend("half4 sum = half4(0);");
    fragBuilder->codeAppendf("float2 coord = %s - %s;", args.fSampleCoord, kernelOffset);

    if (mce.fKernel.isSampled()) {
        this->emitKernelBlock(args, {});
    } else {
        for (int x = 0; x < kernelWidth; ++x) {
            for (int y = 0; y < kernelHeight; ++y) {
                this->emitKernelBlock(args, SkIPoint::Make(x, y));
            }
        }
    }

    fragBuilder->codeAppendf("half4 color;");
    if (mce.fConvolveAlpha) {
        fragBuilder->codeAppendf("color = sum * %s + %s;", gain, bias);
        fragBuilder->codeAppendf("color.a = saturate(color.a);");
        fragBuilder->codeAppendf("color.rgb = clamp(color.rgb, 0.0, color.a);");
    } else {
        auto sample = this->invokeChild(0, args);
        fragBuilder->codeAppendf("half4 c = %s;", sample.c_str());
        fragBuilder->codeAppendf("color.a = c.a;");
        fragBuilder->codeAppendf("color.rgb = saturate(sum.rgb * %s + %s);", gain, bias);
        fragBuilder->codeAppendf("color.rgb *= color.a;");
    }
    fragBuilder->codeAppendf("return color;");
}

void GrMatrixConvolutionEffect::Impl::onSetData(const GrGLSLProgramDataManager& pdman,
                                                const GrFragmentProcessor& processor) {
    const GrMatrixConvolutionEffect& conv = processor.cast<GrMatrixConvolutionEffect>();
    pdman.set2f(fKernelOffsetUni, conv.fKernelOffset.fX, conv.fKernelOffset.fY);
    float totalGain = conv.fGain;
    if (conv.fKernel.isSampled()) {
        totalGain *= conv.fKernel.biasAndGain().fGain;
        pdman.set1f(fKernelBiasUni, conv.fKernel.biasAndGain().fBias);
    } else {
        int kernelCount = conv.fKernel.size().area();
        int arrayCount = (kernelCount + 3) / 4;
        SkASSERT(4 * arrayCount >= kernelCount);
        pdman.set4fv(fKernelUni, arrayCount, conv.fKernel.array().data());
    }
    pdman.set1f(fBiasUni, conv.fBias);
    pdman.set1f(fGainUni, totalGain);
}

GrMatrixConvolutionEffect::GrMatrixConvolutionEffect(std::unique_ptr<GrFragmentProcessor> child,
                                                     const KernelWrapper& kernel,
                                                     std::unique_ptr<GrFragmentProcessor> kernelFP,
                                                     SkScalar gain,
                                                     SkScalar bias,
                                                     const SkIPoint& kernelOffset,
                                                     bool convolveAlpha)
        // To advertise either the modulation or opaqueness optimizations we'd have to examine the
        // parameters.
        : INHERITED(kGrMatrixConvolutionEffect_ClassID, kNone_OptimizationFlags)
        , fKernel(kernel)
        , fGain(SkScalarToFloat(gain))
        , fBias(SkScalarToFloat(bias) / 255.0f)
        , fConvolveAlpha(convolveAlpha) {
    this->registerChild(std::move(child), SkSL::SampleUsage::Explicit());
    this->registerChild(std::move(kernelFP), SkSL::SampleUsage::Explicit());
    fKernelOffset = {static_cast<float>(kernelOffset.x()),
                     static_cast<float>(kernelOffset.y())};
    this->setUsesSampleCoordsDirectly();
}

GrMatrixConvolutionEffect::GrMatrixConvolutionEffect(const GrMatrixConvolutionEffect& that)
        : INHERITED(that)
        , fKernel(that.fKernel)
        , fGain(that.fGain)
        , fBias(that.fBias)
        , fKernelOffset(that.fKernelOffset)
        , fConvolveAlpha(that.fConvolveAlpha) {}

std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::clone() const {
    return std::unique_ptr<GrFragmentProcessor>(new GrMatrixConvolutionEffect(*this));
}

void GrMatrixConvolutionEffect::onAddToKey(const GrShaderCaps& caps,
                                           GrProcessorKeyBuilder* b) const {
    SkASSERT(this->fKernel.size().width() <= 0x7FFF && this->fKernel.size().height() <= 0xFFFF);
    uint32_t key = this->fKernel.size().width() << 16 | this->fKernel.size().height();
    key |= fConvolveAlpha ? 1U << 31 : 0;
    b->add32(key);
}

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

bool GrMatrixConvolutionEffect::onIsEqual(const GrFragmentProcessor& sBase) const {
    const GrMatrixConvolutionEffect& s = sBase.cast<GrMatrixConvolutionEffect>();
    return fKernel == s.fKernel             &&
           fGain == s.fGain                 &&
           fBias == s.fBias                 &&
           fKernelOffset == s.fKernelOffset &&
           fConvolveAlpha == s.fConvolveAlpha;
}

std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::Make(GrRecordingContext* context,
                                                                     GrSurfaceProxyView srcView,
                                                                     const SkIRect& srcBounds,
                                                                     const SkISize& kernelSize,
                                                                     const SkScalar* kernel,
                                                                     SkScalar gain,
                                                                     SkScalar bias,
                                                                     const SkIPoint& kernelOffset,
                                                                     GrSamplerState::WrapMode wm,
                                                                     bool convolveAlpha,
                                                                     const GrCaps& caps) {
    auto [kernelWrapper, kernelFP] = KernelWrapper::Make(context, kernelSize, caps, kernel);
    if (!kernelWrapper.isValid()) {
        return nullptr;
    }
    GrSamplerState sampler(wm, GrSamplerState::Filter::kNearest);
    auto child = GrTextureEffect::MakeSubset(std::move(srcView), kPremul_SkAlphaType, SkMatrix::I(),
                                             sampler, SkRect::Make(srcBounds), caps);
    return std::unique_ptr<GrFragmentProcessor>(
            new GrMatrixConvolutionEffect(std::move(child), kernelWrapper, std::move(kernelFP),
                                          gain, bias, kernelOffset, convolveAlpha));
}

GR_DEFINE_FRAGMENT_PROCESSOR_TEST(GrMatrixConvolutionEffect);

#if GR_TEST_UTILS
std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::TestCreate(GrProcessorTestData* d) {
    auto [view, ct, at] = d->randomView();

    static constexpr size_t kMaxTestKernelSize = 2 * kMaxUniformSize;
    int width = d->fRandom->nextRangeU(1, kMaxTestKernelSize);
    int height = d->fRandom->nextRangeU(1, kMaxTestKernelSize / width);
    SkISize kernelSize = SkISize::Make(width, height);
    std::unique_ptr<SkScalar[]> kernel(new SkScalar[width * height]);
    for (int i = 0; i < width * height; i++) {
        kernel.get()[i] = d->fRandom->nextSScalar1();
    }
    SkScalar gain = d->fRandom->nextSScalar1();
    SkScalar bias = d->fRandom->nextSScalar1();

    uint32_t kernalOffsetX = d->fRandom->nextRangeU(0, kernelSize.width());
    uint32_t kernalOffsetY = d->fRandom->nextRangeU(0, kernelSize.height());
    SkIPoint kernelOffset = SkIPoint::Make(kernalOffsetX, kernalOffsetY);

    uint32_t boundsX = d->fRandom->nextRangeU(0, view.width());
    uint32_t boundsY = d->fRandom->nextRangeU(0, view.height());
    uint32_t boundsW = d->fRandom->nextRangeU(0, view.width());
    uint32_t boundsH = d->fRandom->nextRangeU(0, view.height());
    SkIRect bounds = SkIRect::MakeXYWH(boundsX, boundsY, boundsW, boundsH);

    auto wm = static_cast<GrSamplerState::WrapMode>(
            d->fRandom->nextULessThan(GrSamplerState::kWrapModeCount));
    bool convolveAlpha = d->fRandom->nextBool();
    return GrMatrixConvolutionEffect::Make(d->context(),
                                           std::move(view),
                                           bounds,
                                           kernelSize,
                                           kernel.get(),
                                           gain,
                                           bias,
                                           kernelOffset,
                                           wm,
                                           convolveAlpha,
                                           *d->caps());
}
#endif

Coverage Report

Created: 2021-08-22 09:07

Line	Count	Source (jump to first uncovered line)
1		/*
2		* Copyright 2014 Google Inc.
3		*
4		* Use of this source code is governed by a BSD-style license that can be
5		* found in the LICENSE file.
6		*/
7		#include "src/gpu/effects/GrMatrixConvolutionEffect.h"
8
9		#include "include/private/SkHalf.h"
10		#include "src/gpu/GrDirectContextPriv.h"
11		#include "src/gpu/GrProxyProvider.h"
12		#include "src/gpu/GrRecordingContextPriv.h"
13		#include "src/gpu/GrTexture.h"
14		#include "src/gpu/GrTextureProxy.h"
15		#include "src/gpu/GrThreadSafeCache.h"
16		#include "src/gpu/SkGr.h"
17		#include "src/gpu/effects/GrTextureEffect.h"
18		#include "src/gpu/glsl/GrGLSLFragmentShaderBuilder.h"
19		#include "src/gpu/glsl/GrGLSLProgramDataManager.h"
20		#include "src/gpu/glsl/GrGLSLUniformHandler.h"
21
22		class GrMatrixConvolutionEffect::Impl : public ProgramImpl {
23		public:
24		void emitCode(EmitArgs&) override;
25
26		private:
27		void onSetData(const GrGLSLProgramDataManager&, const GrFragmentProcessor&) override;
28
29		typedef GrGLSLProgramDataManager::UniformHandle UniformHandle;
30
31		void emitKernelBlock(EmitArgs&, SkIPoint);
32
33		UniformHandle fKernelUni;
34		UniformHandle fKernelOffsetUni;
35		UniformHandle fGainUni;
36		UniformHandle fBiasUni;
37		UniformHandle fKernelBiasUni;
38
39		using INHERITED = ProgramImpl;
40		};
41
42		GrMatrixConvolutionEffect::KernelWrapper::MakeResult
43		GrMatrixConvolutionEffect::KernelWrapper::Make(GrRecordingContext* rContext,
44		SkISize size,
45		const GrCaps& caps,
46	592	const SkScalar* values) {
47	592	if (!rContext \|\| !values \|\| size.isEmpty()) {
48	0	return {};
49	0	}
50
51	592	const int length = size.area();
52		// Small kernel -> just fill the array.
53	592	KernelWrapper result(size);
54	592	if (length <= kMaxUniformSize) {
55	5.78k	for (int i = 0; i < length; i++) {
56	5.18k	result.fArray[i] = SkScalarToFloat(values[i]);
57	5.18k	}
58	592	return {result, nullptr};
59	592	}
60
61	0	BiasAndGain& scalableSampler = result.fBiasAndGain;
62	0	bool useA16 =
63	0	rContext->defaultBackendFormat(kA16_float_SkColorType, GrRenderable::kNo).isValid();
64	0	SkScalar min = values[0];
65	0	if (!useA16) {
66		// Determine min and max values to figure out inner gain & bias.
67	0	SkScalar max = values[0];
68	0	for (int i = 1; i < length; i++) {
69	0	if (values[i] < min) {
70	0	min = values[i];
71	0	}
72	0	if (values[i] > max) {
73	0	max = values[i];
74	0	}
75	0	}
76		// Treat near-0 gain (i.e. box blur) as 1, and let the kernelBias
77		// move everything up to the final value.
78	0	const SkScalar computedGain = max - min;
79	0	scalableSampler.fGain =
80	0	SkScalarNearlyZero(computedGain) ? 1.0f : SkScalarToFloat(computedGain);
81		// Inner bias is pre-inner-gain so we divide that out.
82	0	scalableSampler.fBias = SkScalarToFloat(min) / scalableSampler.fGain;
83	0	}
84
85		// TODO: Pick cache or dont-cache based on observed perf.
86	0	static constexpr bool kCacheKernelTexture = true;
87
88	0	GrUniqueKey key;
89	0	if (kCacheKernelTexture) {
90	0	static const GrUniqueKey::Domain kDomain = GrUniqueKey::GenerateDomain();
91	0	GrUniqueKey::Builder builder(&key, kDomain, length, "Matrix Convolution Kernel");
92		// Texture cache key is the exact content of the kernel.
93	0	static_assert(sizeof(float) == 4);
94	0	for (int i = 0; i < length; i++) {
95	0	builder[i] = (const uint32_t)&values[i];
96	0	}
97	0	builder.finish();
98	0	}
99
100		// Find or create a texture.
101	0	auto threadSafeCache = rContext->priv().threadSafeCache();
102
103	0	SkColorType colorType = useA16 ? kA16_float_SkColorType : kAlpha_8_SkColorType;
104
105	0	GrSurfaceProxyView view;
106	0	if (kCacheKernelTexture && (view = threadSafeCache->find(key))) {
107	0	SkASSERT(view.origin() == kTopLeft_GrSurfaceOrigin);
108	0	auto kernelFP = GrTextureEffect::Make(std::move(view), kUnknown_SkAlphaType);
109	0	return {result, std::move(kernelFP)};
110	0	}
111
112	0	SkBitmap bm;
113	0	auto info = SkImageInfo::Make({length, 1}, colorType, kPremul_SkAlphaType, nullptr);
114	0	if (!bm.tryAllocPixels(info)) {
115	0	return {};
116	0	}
117	0	for (int i = 0; i < length; i++) {
118	0	if (useA16) {
119	0	*bm.getAddr16(i, 0) = SkFloatToHalf(values[i]);
120	0	} else {
121	0	*bm.getAddr8(i, 0) =
122	0	SkScalarRoundToInt((values[i] - min) / scalableSampler.fGain * 255);
123	0	}
124	0	}
125	0	bm.setImmutable();
126
127	0	view = std::get<0>(GrMakeUncachedBitmapProxyView(rContext, bm));
128	0	if (!view) {
129	0	return {};
130	0	}
131
132	0	if (kCacheKernelTexture) {
133	0	view = threadSafeCache->add(key, view);
134	0	}
135
136	0	SkASSERT(view.origin() == kTopLeft_GrSurfaceOrigin);
137	0	auto kernelFP = GrTextureEffect::Make(std::move(view), kUnknown_SkAlphaType);
138	0	return {result, std::move(kernelFP)};
139	0	}
140
141	0	bool GrMatrixConvolutionEffect::KernelWrapper::operator==(const KernelWrapper& k) const {
142	0	if (fSize != k.fSize) {
143	0	return false;
144	0	} else if (this->isSampled()) {
145	0	return fBiasAndGain == k.fBiasAndGain;
146	0	} else {
147	0	return std::equal(fArray.begin(), fArray.begin() + fSize.area(), k.fArray.begin());
148	0	}
149	0	}
150
151		bool GrMatrixConvolutionEffect::KernelWrapper::BiasAndGain::operator==(
152	0	const BiasAndGain& k) const {
153	0	return fGain == k.fGain && fBias == k.fBias;
154	0	}
155
156		// For sampled kernels, emit a for loop that does all the kernel accumulation.
157		// For uniform kernels, emit a single iteration. Function is called repeatedly in a for loop.
158		// loc is ignored for sampled kernels.
159	0	void GrMatrixConvolutionEffect::Impl::emitKernelBlock(EmitArgs& args, SkIPoint loc) {
160	0	const GrMatrixConvolutionEffect& mce = args.fFp.cast<GrMatrixConvolutionEffect>();
161	0	GrGLSLFPFragmentBuilder* fragBuilder = args.fFragBuilder;
162	0	GrGLSLUniformHandler* uniformHandler = args.fUniformHandler;
163	0	int kernelWidth = mce.fKernel.size().width();
164	0	int kernelHeight = mce.fKernel.size().height();
165	0	int kernelArea = kernelWidth * kernelHeight;
166
167	0	if (mce.fKernel.isSampled()) {
168	0	fragBuilder->codeAppendf("for (int i = 0; i < %d; ++i)", (int)kernelArea);
169	0	}
170
171	0	GrGLSLShaderBuilder::ShaderBlock block(fragBuilder);
172
173	0	fragBuilder->codeAppend("half k;");
174	0	fragBuilder->codeAppend("half2 sourceOffset;");
175	0	if (mce.fKernel.isSampled()) {
176	0	const char* kernelBias = uniformHandler->getUniformCStr(fKernelBiasUni);
177	0	SkString kernelSample = this->invokeChild(1, args, "float2(float(i) + 0.5, 0.5)");
178	0	fragBuilder->codeAppendf("k = %s.w + %s;", kernelSample.c_str(), kernelBias);
179	0	fragBuilder->codeAppendf("sourceOffset.y = floor(half(i) / %d);", kernelWidth);
180	0	fragBuilder->codeAppendf("sourceOffset.x = half(i) - sourceOffset.y * %d;", kernelWidth);
181	0	} else {
182	0	fragBuilder->codeAppendf("sourceOffset = half2(%d, %d);", loc.x(), loc.y());
183	0	int offset = loc.y() * kernelWidth + loc.x();
184	0	const char* kernel = uniformHandler->getUniformCStr(fKernelUni);
185	0	fragBuilder->codeAppendf("k = %s[%d][%d];", kernel, offset / 4, offset & 0x3);
186	0	}
187
188	0	auto sample = this->invokeChild(0, args, "coord + sourceOffset");
189	0	fragBuilder->codeAppendf("half4 c = %s;", sample.c_str());
190	0	if (!mce.fConvolveAlpha) {
191	0	fragBuilder->codeAppend("c = unpremul(c);");
192	0	fragBuilder->codeAppend("c.rgb = saturate(c.rgb);");
193	0	}
194	0	fragBuilder->codeAppend("sum += c * k;");
195	0	}
196
197	0	void GrMatrixConvolutionEffect::Impl::emitCode(EmitArgs& args) {
198	0	const GrMatrixConvolutionEffect& mce = args.fFp.cast<GrMatrixConvolutionEffect>();
199
200	0	int kernelWidth = mce.fKernel.size().width();
201	0	int kernelHeight = mce.fKernel.size().height();
202
203	0	int arrayCount = (kernelWidth * kernelHeight + 3) / 4;
204	0	SkASSERT(4 * arrayCount >= kernelWidth * kernelHeight);
205
206	0	GrGLSLUniformHandler* uniformHandler = args.fUniformHandler;
207	0	if (mce.fKernel.isSampled()) {
208	0	fKernelBiasUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag,
209	0	kHalf_GrSLType, "KernelBias");
210	0	} else {
211	0	fKernelUni = uniformHandler->addUniformArray(&mce, kFragment_GrShaderFlag,
212	0	kHalf4_GrSLType, "Kernel", arrayCount);
213	0	}
214	0	fKernelOffsetUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf2_GrSLType,
215	0	"KernelOffset");
216	0	fGainUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf_GrSLType, "Gain");
217	0	fBiasUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf_GrSLType, "Bias");
218
219	0	const char* kernelOffset = uniformHandler->getUniformCStr(fKernelOffsetUni);
220	0	const char* gain = uniformHandler->getUniformCStr(fGainUni);
221	0	const char* bias = uniformHandler->getUniformCStr(fBiasUni);
222
223	0	GrGLSLFPFragmentBuilder* fragBuilder = args.fFragBuilder;
224	0	fragBuilder->codeAppend("half4 sum = half4(0);");
225	0	fragBuilder->codeAppendf("float2 coord = %s - %s;", args.fSampleCoord, kernelOffset);
226
227	0	if (mce.fKernel.isSampled()) {
228	0	this->emitKernelBlock(args, {});
229	0	} else {
230	0	for (int x = 0; x < kernelWidth; ++x) {
231	0	for (int y = 0; y < kernelHeight; ++y) {
232	0	this->emitKernelBlock(args, SkIPoint::Make(x, y));
233	0	}
234	0	}
235	0	}
236
237	0	fragBuilder->codeAppendf("half4 color;");
238	0	if (mce.fConvolveAlpha) {
239	0	fragBuilder->codeAppendf("color = sum * %s + %s;", gain, bias);
240	0	fragBuilder->codeAppendf("color.a = saturate(color.a);");
241	0	fragBuilder->codeAppendf("color.rgb = clamp(color.rgb, 0.0, color.a);");
242	0	} else {
243	0	auto sample = this->invokeChild(0, args);
244	0	fragBuilder->codeAppendf("half4 c = %s;", sample.c_str());
245	0	fragBuilder->codeAppendf("color.a = c.a;");
246	0	fragBuilder->codeAppendf("color.rgb = saturate(sum.rgb * %s + %s);", gain, bias);
247	0	fragBuilder->codeAppendf("color.rgb *= color.a;");
248	0	}
249	0	fragBuilder->codeAppendf("return color;");
250	0	} Unexecuted instantiation: GrMatrixConvolutionEffect::Impl::emitCode(GrFragmentProcessor::ProgramImpl::EmitArgs&) Unexecuted instantiation: GrMatrixConvolutionEffect::Impl::emitCode(GrFragmentProcessor::ProgramImpl::EmitArgs&)
251
252		void GrMatrixConvolutionEffect::Impl::onSetData(const GrGLSLProgramDataManager& pdman,
253	0	const GrFragmentProcessor& processor) {
254	0	const GrMatrixConvolutionEffect& conv = processor.cast<GrMatrixConvolutionEffect>();
255	0	pdman.set2f(fKernelOffsetUni, conv.fKernelOffset.fX, conv.fKernelOffset.fY);
256	0	float totalGain = conv.fGain;
257	0	if (conv.fKernel.isSampled()) {
258	0	totalGain *= conv.fKernel.biasAndGain().fGain;
259	0	pdman.set1f(fKernelBiasUni, conv.fKernel.biasAndGain().fBias);
260	0	} else {
261	0	int kernelCount = conv.fKernel.size().area();
262	0	int arrayCount = (kernelCount + 3) / 4;
263	0	SkASSERT(4 * arrayCount >= kernelCount);
264	0	pdman.set4fv(fKernelUni, arrayCount, conv.fKernel.array().data());
265	0	}
266	0	pdman.set1f(fBiasUni, conv.fBias);
267	0	pdman.set1f(fGainUni, totalGain);
268	0	} Unexecuted instantiation: GrMatrixConvolutionEffect::Impl::onSetData(GrGLSLProgramDataManager const&, GrFragmentProcessor const&) Unexecuted instantiation: GrMatrixConvolutionEffect::Impl::onSetData(GrGLSLProgramDataManager const&, GrFragmentProcessor const&)
269
270		GrMatrixConvolutionEffect::GrMatrixConvolutionEffect(std::unique_ptr<GrFragmentProcessor> child,
271		const KernelWrapper& kernel,
272		std::unique_ptr<GrFragmentProcessor> kernelFP,
273		SkScalar gain,
274		SkScalar bias,
275		const SkIPoint& kernelOffset,
276		bool convolveAlpha)
277		// To advertise either the modulation or opaqueness optimizations we'd have to examine the
278		// parameters.
279		: INHERITED(kGrMatrixConvolutionEffect_ClassID, kNone_OptimizationFlags)
280		, fKernel(kernel)
281		, fGain(SkScalarToFloat(gain))
282		, fBias(SkScalarToFloat(bias) / 255.0f)
283	592	, fConvolveAlpha(convolveAlpha) {
284	592	this->registerChild(std::move(child), SkSL::SampleUsage::Explicit());
285	592	this->registerChild(std::move(kernelFP), SkSL::SampleUsage::Explicit());
286	592	fKernelOffset = {static_cast<float>(kernelOffset.x()),
287	592	static_cast<float>(kernelOffset.y())};
288	592	this->setUsesSampleCoordsDirectly();
289	592	}
290
291		GrMatrixConvolutionEffect::GrMatrixConvolutionEffect(const GrMatrixConvolutionEffect& that)
292		: INHERITED(that)
293		, fKernel(that.fKernel)
294		, fGain(that.fGain)
295		, fBias(that.fBias)
296		, fKernelOffset(that.fKernelOffset)
297	0	, fConvolveAlpha(that.fConvolveAlpha) {}
298
299	0	std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::clone() const {
300	0	return std::unique_ptr<GrFragmentProcessor>(new GrMatrixConvolutionEffect(*this));
301	0	}
302
303		void GrMatrixConvolutionEffect::onAddToKey(const GrShaderCaps& caps,
304	0	GrProcessorKeyBuilder* b) const {
305	0	SkASSERT(this->fKernel.size().width() <= 0x7FFF && this->fKernel.size().height() <= 0xFFFF);
306	0	uint32_t key = this->fKernel.size().width() << 16 \| this->fKernel.size().height();
307	0	key \|= fConvolveAlpha ? 1U << 31 : 0;
308	0	b->add32(key);
309	0	} Unexecuted instantiation: GrMatrixConvolutionEffect::onAddToKey(GrShaderCaps const&, GrProcessorKeyBuilder) const Unexecuted instantiation: GrMatrixConvolutionEffect::onAddToKey(GrShaderCaps const&, GrProcessorKeyBuilder) const
310
311		std::unique_ptr<GrFragmentProcessor::ProgramImpl>
312	0	GrMatrixConvolutionEffect::onMakeProgramImpl() const {
313	0	return std::make_unique<Impl>();
314	0	}
315
316	0	bool GrMatrixConvolutionEffect::onIsEqual(const GrFragmentProcessor& sBase) const {
317	0	const GrMatrixConvolutionEffect& s = sBase.cast<GrMatrixConvolutionEffect>();
318	0	return fKernel == s.fKernel &&
319	0	fGain == s.fGain &&
320	0	fBias == s.fBias &&
321	0	fKernelOffset == s.fKernelOffset &&
322	0	fConvolveAlpha == s.fConvolveAlpha;
323	0	}
324
325		std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::Make(GrRecordingContext* context,
326		GrSurfaceProxyView srcView,
327		const SkIRect& srcBounds,
328		const SkISize& kernelSize,
329		const SkScalar* kernel,
330		SkScalar gain,
331		SkScalar bias,
332		const SkIPoint& kernelOffset,
333		GrSamplerState::WrapMode wm,
334		bool convolveAlpha,
335	592	const GrCaps& caps) {
336	592	auto [kernelWrapper, kernelFP] = KernelWrapper::Make(context, kernelSize, caps, kernel);
337	592	if (!kernelWrapper.isValid()) {
338	0	return nullptr;
339	0	}
340	592	GrSamplerState sampler(wm, GrSamplerState::Filter::kNearest);
341	592	auto child = GrTextureEffect::MakeSubset(std::move(srcView), kPremul_SkAlphaType, SkMatrix::I(),
342	592	sampler, SkRect::Make(srcBounds), caps);
343	592	return std::unique_ptr<GrFragmentProcessor>(
344	592	new GrMatrixConvolutionEffect(std::move(child), kernelWrapper, std::move(kernelFP),
345	592	gain, bias, kernelOffset, convolveAlpha));
346	592	}
347
348		GR_DEFINE_FRAGMENT_PROCESSOR_TEST(GrMatrixConvolutionEffect);
349
350		#if GR_TEST_UTILS
351	0	std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::TestCreate(GrProcessorTestData* d) {
352	0	auto [view, ct, at] = d->randomView();
353
354	0	static constexpr size_t kMaxTestKernelSize = 2 * kMaxUniformSize;
355	0	int width = d->fRandom->nextRangeU(1, kMaxTestKernelSize);
356	0	int height = d->fRandom->nextRangeU(1, kMaxTestKernelSize / width);
357	0	SkISize kernelSize = SkISize::Make(width, height);
358	0	std::unique_ptr<SkScalar[]> kernel(new SkScalar[width * height]);
359	0	for (int i = 0; i < width * height; i++) {
360	0	kernel.get()[i] = d->fRandom->nextSScalar1();
361	0	}
362	0	SkScalar gain = d->fRandom->nextSScalar1();
363	0	SkScalar bias = d->fRandom->nextSScalar1();
364
365	0	uint32_t kernalOffsetX = d->fRandom->nextRangeU(0, kernelSize.width());
366	0	uint32_t kernalOffsetY = d->fRandom->nextRangeU(0, kernelSize.height());
367	0	SkIPoint kernelOffset = SkIPoint::Make(kernalOffsetX, kernalOffsetY);
368
369	0	uint32_t boundsX = d->fRandom->nextRangeU(0, view.width());
370	0	uint32_t boundsY = d->fRandom->nextRangeU(0, view.height());
371	0	uint32_t boundsW = d->fRandom->nextRangeU(0, view.width());
372	0	uint32_t boundsH = d->fRandom->nextRangeU(0, view.height());
373	0	SkIRect bounds = SkIRect::MakeXYWH(boundsX, boundsY, boundsW, boundsH);
374
375	0	auto wm = static_cast<GrSamplerState::WrapMode>(
376	0	d->fRandom->nextULessThan(GrSamplerState::kWrapModeCount));
377	0	bool convolveAlpha = d->fRandom->nextBool();
378	0	return GrMatrixConvolutionEffect::Make(d->context(),
379	0	std::move(view),
380	0	bounds,
381	0	kernelSize,
382	0	kernel.get(),
383	0	gain,
384	0	bias,
385	0	kernelOffset,
386	0	wm,
387	0	convolveAlpha,
388	0	*d->caps());
389	0	}
390		#endif