| 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/GrBitmapTextureMaker.h" |
| 11 | #include "src/gpu/GrContextPriv.h" |
| 12 | #include "src/gpu/GrProxyProvider.h" |
| 13 | #include "src/gpu/GrRecordingContextPriv.h" |
| 14 | #include "src/gpu/GrTexture.h" |
| 15 | #include "src/gpu/GrTextureProxy.h" |
| 16 | #include "src/gpu/effects/GrTextureEffect.h" |
| 17 | #include "src/gpu/glsl/GrGLSLFragmentProcessor.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 GrGLMatrixConvolutionEffect : public GrGLSLFragmentProcessor { |
| 23 | public: |
| 24 | void emitCode(EmitArgs&) override; |
| 25 | |
| 26 | static inline void GenKey(const GrProcessor&, const GrShaderCaps&, GrProcessorKeyBuilder*); |
| 27 | |
| 28 | protected: |
| 29 | void onSetData(const GrGLSLProgramDataManager&, const GrFragmentProcessor&) override; |
| 30 | |
| 31 | private: |
| 32 | typedef GrGLSLProgramDataManager::UniformHandle UniformHandle; |
| 33 | |
| 34 | void emitKernelBlock(EmitArgs&, SkIPoint); |
| 35 | |
| 36 | UniformHandle fKernelUni; |
| 37 | UniformHandle fKernelOffsetUni; |
| 38 | UniformHandle fGainUni; |
| 39 | UniformHandle fBiasUni; |
| 40 | UniformHandle fKernelBiasUni; |
| 41 | |
| 42 | typedef GrGLSLFragmentProcessor INHERITED; |
| 43 | }; |
| 44 | |
| 45 | GrMatrixConvolutionEffect::KernelWrapper::MakeResult |
| 46 | GrMatrixConvolutionEffect::KernelWrapper::Make(GrRecordingContext* context, |
| 47 | SkISize size, |
| 48 | const GrCaps& caps, |
| 49 | const SkScalar* values) { |
| 50 | if (!context || !values || size.isEmpty()) { |
| 51 | return {}; |
| 52 | } |
| 53 | const int length = size.area(); |
| 54 | // Small kernel -> just fill the array. |
| 55 | KernelWrapper result(size); |
| 56 | if (length <= kMaxUniformSize) { |
| 57 | for (int i = 0; i < length; i++) { |
| 58 | result.fArray[i] = SkScalarToFloat(values[i]); |
| 59 | } |
| 60 | return {result, nullptr}; |
| 61 | } |
| 62 | |
| 63 | BiasAndGain& scalableSampler = result.fBiasAndGain; |
| 64 | bool useA16 = |
| 65 | context->defaultBackendFormat(kA16_float_SkColorType, GrRenderable::kNo).isValid(); |
| 66 | SkScalar min = values[0]; |
| 67 | if (!useA16) { |
| 68 | // Determine min and max values to figure out inner gain & bias. |
| 69 | SkScalar max = values[0]; |
| 70 | for (int i = 1; i < length; i++) { |
| 71 | if (values[i] < min) { |
| 72 | min = values[i]; |
| 73 | } |
| 74 | if (values[i] > max) { |
| 75 | max = values[i]; |
| 76 | } |
| 77 | } |
| 78 | // Treat near-0 gain (i.e. box blur) as 1, and let the kernelBias |
| 79 | // move everything up to the final value. |
| 80 | const SkScalar computedGain = max - min; |
| 81 | scalableSampler.fGain = |
| 82 | SkScalarNearlyZero(computedGain) ? 1.0f : SkScalarToFloat(computedGain); |
| 83 | // Inner bias is pre-inner-gain so we divide that out. |
| 84 | scalableSampler.fBias = SkScalarToFloat(min) / scalableSampler.fGain; |
| 85 | } |
| 86 | |
| 87 | // TODO: Pick cache or dont-cache based on observed perf. |
| 88 | static constexpr bool kCacheKernelTexture = true; |
| 89 | |
| 90 | GrUniqueKey key; |
| 91 | if (kCacheKernelTexture) { |
| 92 | static const GrUniqueKey::Domain kDomain = GrUniqueKey::GenerateDomain(); |
| 93 | GrUniqueKey::Builder builder(&key, kDomain, length, "Matrix Convolution Kernel" ); |
| 94 | // Texture cache key is the exact content of the kernel. |
| 95 | static_assert(sizeof(float) == 4); |
| 96 | for (int i = 0; i < length; i++) { |
| 97 | builder[i] = *(const uint32_t*)&values[i]; |
| 98 | } |
| 99 | builder.finish(); |
| 100 | } |
| 101 | |
| 102 | // Find or create a texture. |
| 103 | GrProxyProvider* proxyProvider = context->priv().proxyProvider(); |
| 104 | GrSurfaceProxyView view; |
| 105 | SkColorType colorType = useA16 ? kA16_float_SkColorType : kAlpha_8_SkColorType; |
| 106 | sk_sp<GrTextureProxy> cachedKernel; |
| 107 | if (kCacheKernelTexture && (cachedKernel = proxyProvider->findOrCreateProxyByUniqueKey(key))) { |
| 108 | GrSwizzle swizzle = |
| 109 | context->priv().caps()->getReadSwizzle(cachedKernel->backendFormat(), |
| 110 | SkColorTypeToGrColorType(colorType)); |
| 111 | view = {std::move(cachedKernel), kTopLeft_GrSurfaceOrigin, swizzle}; |
| 112 | } else { |
| 113 | SkBitmap bm; |
| 114 | auto info = SkImageInfo::Make({length, 1}, colorType, kPremul_SkAlphaType, nullptr); |
| 115 | if (!bm.tryAllocPixels(info)) { |
| 116 | return {}; |
| 117 | } |
| 118 | for (int i = 0; i < length; i++) { |
| 119 | if (useA16) { |
| 120 | *bm.getAddr16(i, 0) = SkFloatToHalf(values[i]); |
| 121 | } else { |
| 122 | *bm.getAddr8(i, 0) = |
| 123 | SkScalarRoundToInt((values[i] - min) / scalableSampler.fGain * 255); |
| 124 | } |
| 125 | } |
| 126 | bm.setImmutable(); |
| 127 | GrBitmapTextureMaker maker(context, bm, GrImageTexGenPolicy::kNew_Uncached_Budgeted); |
| 128 | view = maker.view(GrMipmapped::kNo); |
| 129 | if (!view) { |
| 130 | return {}; |
| 131 | } |
| 132 | if (kCacheKernelTexture) { |
| 133 | proxyProvider->assignUniqueKeyToProxy(key, view.asTextureProxy()); |
| 134 | } |
| 135 | } |
| 136 | auto kernelFP = GrTextureEffect::Make(std::move(view), kUnknown_SkAlphaType); |
| 137 | return {result, std::move(kernelFP)}; |
| 138 | } |
| 139 | |
| 140 | bool GrMatrixConvolutionEffect::KernelWrapper::operator==(const KernelWrapper& k) const { |
| 141 | if (fSize != k.fSize) { |
| 142 | return false; |
| 143 | } else if (this->isSampled()) { |
| 144 | return fBiasAndGain == k.fBiasAndGain; |
| 145 | } else { |
| 146 | return std::equal(fArray.begin(), fArray.begin() + fSize.area(), k.fArray.begin()); |
| 147 | } |
| 148 | } |
| 149 | |
| 150 | bool GrMatrixConvolutionEffect::KernelWrapper::BiasAndGain::operator==( |
| 151 | const BiasAndGain& k) const { |
| 152 | return fGain == k.fGain && fBias == k.fBias; |
| 153 | } |
| 154 | |
| 155 | // For sampled kernels, emit a for loop that does all the kernel accumulation. |
| 156 | // For uniform kernels, emit a single iteration. Function is called repeatedly in a for loop. |
| 157 | // loc is ignored for sampled kernels. |
| 158 | void GrGLMatrixConvolutionEffect::emitKernelBlock(EmitArgs& args, SkIPoint loc) { |
| 159 | const GrMatrixConvolutionEffect& mce = args.fFp.cast<GrMatrixConvolutionEffect>(); |
| 160 | GrGLSLFPFragmentBuilder* fragBuilder = args.fFragBuilder; |
| 161 | GrGLSLUniformHandler* uniformHandler = args.fUniformHandler; |
| 162 | int kernelWidth = mce.kernelSize().width(); |
| 163 | int kernelHeight = mce.kernelSize().height(); |
| 164 | int kernelArea = kernelWidth * kernelHeight; |
| 165 | |
| 166 | if (mce.kernelIsSampled()) { |
| 167 | fragBuilder->codeAppendf("for (int i = 0; i < %d; ++i)" , (int)kernelArea); |
| 168 | } |
| 169 | |
| 170 | GrGLSLShaderBuilder::ShaderBlock block(fragBuilder); |
| 171 | |
| 172 | fragBuilder->codeAppend("half k;" ); |
| 173 | fragBuilder->codeAppend("half2 sourceOffset;" ); |
| 174 | if (mce.kernelIsSampled()) { |
| 175 | const char* kernelBias = uniformHandler->getUniformCStr(fKernelBiasUni); |
| 176 | SkString kernelCoord = SkStringPrintf("float2(float(i) + 0.5, 0.5)" ); |
| 177 | SkString kernelSample = this->invokeChild(1, args, kernelCoord.c_str()); |
| 178 | fragBuilder->codeAppendf("k = %s.w + %s;" , kernelSample.c_str(), kernelBias); |
| 179 | fragBuilder->codeAppendf("sourceOffset.y = floor(i / %d);" , kernelWidth); |
| 180 | fragBuilder->codeAppendf("sourceOffset.x = i - sourceOffset.y * %d;" , kernelWidth); |
| 181 | } else { |
| 182 | fragBuilder->codeAppendf("sourceOffset = half2(%d, %d);" , loc.x(), loc.y()); |
| 183 | int offset = loc.y() * kernelWidth + loc.x(); |
| 184 | static constexpr const char kVecSuffix[][4] = { ".x" , ".y" , ".z" , ".w" }; |
| 185 | const char* kernel = uniformHandler->getUniformCStr(fKernelUni); |
| 186 | fragBuilder->codeAppendf("k = %s[%d]%s;" , kernel, offset / 4, |
| 187 | kVecSuffix[offset & 0x3]); |
| 188 | } |
| 189 | |
| 190 | auto sample = this->invokeChild(0, args, "coord + sourceOffset" ); |
| 191 | fragBuilder->codeAppendf("half4 c = %s;" , sample.c_str()); |
| 192 | if (!mce.convolveAlpha()) { |
| 193 | fragBuilder->codeAppend("c = unpremul(c);" ); |
| 194 | fragBuilder->codeAppend("c.rgb = saturate(c.rgb);" ); |
| 195 | } |
| 196 | fragBuilder->codeAppend("sum += c * k;" ); |
| 197 | } |
| 198 | |
| 199 | void GrGLMatrixConvolutionEffect::emitCode(EmitArgs& args) { |
| 200 | const GrMatrixConvolutionEffect& mce = args.fFp.cast<GrMatrixConvolutionEffect>(); |
| 201 | |
| 202 | int kernelWidth = mce.kernelSize().width(); |
| 203 | int kernelHeight = mce.kernelSize().height(); |
| 204 | |
| 205 | int arrayCount = (kernelWidth * kernelHeight + 3) / 4; |
| 206 | SkASSERT(4 * arrayCount >= kernelWidth * kernelHeight); |
| 207 | |
| 208 | GrGLSLUniformHandler* uniformHandler = args.fUniformHandler; |
| 209 | if (mce.kernelIsSampled()) { |
| 210 | fKernelBiasUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, |
| 211 | kHalf_GrSLType, "KernelBias" ); |
| 212 | } else { |
| 213 | fKernelUni = uniformHandler->addUniformArray(&mce, kFragment_GrShaderFlag, |
| 214 | kHalf4_GrSLType, "Kernel" , arrayCount); |
| 215 | } |
| 216 | fKernelOffsetUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf2_GrSLType, |
| 217 | "KernelOffset" ); |
| 218 | fGainUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf_GrSLType, "Gain" ); |
| 219 | fBiasUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf_GrSLType, "Bias" ); |
| 220 | |
| 221 | const char* kernelOffset = uniformHandler->getUniformCStr(fKernelOffsetUni); |
| 222 | const char* gain = uniformHandler->getUniformCStr(fGainUni); |
| 223 | const char* bias = uniformHandler->getUniformCStr(fBiasUni); |
| 224 | |
| 225 | GrGLSLFPFragmentBuilder* fragBuilder = args.fFragBuilder; |
| 226 | fragBuilder->codeAppend("half4 sum = half4(0, 0, 0, 0);" ); |
| 227 | fragBuilder->codeAppendf("float2 coord = %s - %s;" , args.fSampleCoord, kernelOffset); |
| 228 | |
| 229 | if (mce.kernelIsSampled()) { |
| 230 | this->emitKernelBlock(args, {}); |
| 231 | } else { |
| 232 | for (int x = 0; x < kernelWidth; ++x) { |
| 233 | for (int y = 0; y < kernelHeight; ++y) { |
| 234 | this->emitKernelBlock(args, SkIPoint::Make(x, y)); |
| 235 | } |
| 236 | } |
| 237 | } |
| 238 | |
| 239 | if (mce.convolveAlpha()) { |
| 240 | fragBuilder->codeAppendf("%s = sum * %s + %s;" , args.fOutputColor, gain, bias); |
| 241 | fragBuilder->codeAppendf("%s.a = saturate(%s.a);" , args.fOutputColor, args.fOutputColor); |
| 242 | fragBuilder->codeAppendf("%s.rgb = clamp(%s.rgb, 0.0, %s.a);" , |
| 243 | args.fOutputColor, args.fOutputColor, args.fOutputColor); |
| 244 | } else { |
| 245 | auto sample = this->invokeChild(0, args); |
| 246 | fragBuilder->codeAppendf("half4 c = %s;" , sample.c_str()); |
| 247 | fragBuilder->codeAppendf("%s.a = c.a;" , args.fOutputColor); |
| 248 | fragBuilder->codeAppendf("%s.rgb = saturate(sum.rgb * %s + %s);" , args.fOutputColor, gain, bias); |
| 249 | fragBuilder->codeAppendf("%s.rgb *= %s.a;" , args.fOutputColor, args.fOutputColor); |
| 250 | } |
| 251 | fragBuilder->codeAppendf("%s *= %s;\n" , args.fOutputColor, args.fInputColor); |
| 252 | } |
| 253 | |
| 254 | void GrGLMatrixConvolutionEffect::GenKey(const GrProcessor& processor, |
| 255 | const GrShaderCaps&, GrProcessorKeyBuilder* b) { |
| 256 | const GrMatrixConvolutionEffect& m = processor.cast<GrMatrixConvolutionEffect>(); |
| 257 | SkASSERT(m.kernelSize().width() <= 0x7FFF && m.kernelSize().height() <= 0xFFFF); |
| 258 | uint32_t key = m.kernelSize().width() << 16 | m.kernelSize().height(); |
| 259 | key |= m.convolveAlpha() ? 1U << 31 : 0; |
| 260 | b->add32(key); |
| 261 | } |
| 262 | |
| 263 | void GrGLMatrixConvolutionEffect::onSetData(const GrGLSLProgramDataManager& pdman, |
| 264 | const GrFragmentProcessor& processor) { |
| 265 | const GrMatrixConvolutionEffect& conv = processor.cast<GrMatrixConvolutionEffect>(); |
| 266 | pdman.set2f(fKernelOffsetUni, conv.kernelOffset().fX, conv.kernelOffset().fY); |
| 267 | float totalGain = conv.gain(); |
| 268 | if (conv.kernelIsSampled()) { |
| 269 | totalGain *= conv.kernelSampleGain(); |
| 270 | pdman.set1f(fKernelBiasUni, conv.kernelSampleBias()); |
| 271 | } else { |
| 272 | int kernelCount = conv.kernelSize().area(); |
| 273 | int arrayCount = (kernelCount + 3) / 4; |
| 274 | SkASSERT(4 * arrayCount >= kernelCount); |
| 275 | pdman.set4fv(fKernelUni, arrayCount, conv.kernel()); |
| 276 | } |
| 277 | pdman.set1f(fBiasUni, conv.bias()); |
| 278 | pdman.set1f(fGainUni, totalGain); |
| 279 | } |
| 280 | |
| 281 | GrMatrixConvolutionEffect::GrMatrixConvolutionEffect(std::unique_ptr<GrFragmentProcessor> child, |
| 282 | const KernelWrapper& kernel, |
| 283 | std::unique_ptr<GrFragmentProcessor> kernelFP, |
| 284 | SkScalar gain, |
| 285 | SkScalar bias, |
| 286 | const SkIPoint& kernelOffset, |
| 287 | bool convolveAlpha) |
| 288 | // To advertise either the modulation or opaqueness optimizations we'd have to examine the |
| 289 | // parameters. |
| 290 | : INHERITED(kGrMatrixConvolutionEffect_ClassID, kNone_OptimizationFlags) |
| 291 | , fKernel(kernel) |
| 292 | , fGain(SkScalarToFloat(gain)) |
| 293 | , fBias(SkScalarToFloat(bias) / 255.0f) |
| 294 | , fConvolveAlpha(convolveAlpha) { |
| 295 | this->registerChild(std::move(child), SkSL::SampleUsage::Explicit()); |
| 296 | this->registerChild(std::move(kernelFP), SkSL::SampleUsage::Explicit()); |
| 297 | fKernelOffset = {static_cast<float>(kernelOffset.x()), |
| 298 | static_cast<float>(kernelOffset.y())}; |
| 299 | this->setUsesSampleCoordsDirectly(); |
| 300 | } |
| 301 | |
| 302 | GrMatrixConvolutionEffect::GrMatrixConvolutionEffect(const GrMatrixConvolutionEffect& that) |
| 303 | : INHERITED(kGrMatrixConvolutionEffect_ClassID, kNone_OptimizationFlags) |
| 304 | , fKernel(that.fKernel) |
| 305 | , fGain(that.fGain) |
| 306 | , fBias(that.fBias) |
| 307 | , fKernelOffset(that.fKernelOffset) |
| 308 | , fConvolveAlpha(that.fConvolveAlpha) { |
| 309 | this->cloneAndRegisterAllChildProcessors(that); |
| 310 | this->setUsesSampleCoordsDirectly(); |
| 311 | } |
| 312 | |
| 313 | std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::clone() const { |
| 314 | return std::unique_ptr<GrFragmentProcessor>(new GrMatrixConvolutionEffect(*this)); |
| 315 | } |
| 316 | |
| 317 | void GrMatrixConvolutionEffect::onGetGLSLProcessorKey(const GrShaderCaps& caps, |
| 318 | GrProcessorKeyBuilder* b) const { |
| 319 | GrGLMatrixConvolutionEffect::GenKey(*this, caps, b); |
| 320 | } |
| 321 | |
| 322 | GrGLSLFragmentProcessor* GrMatrixConvolutionEffect::onCreateGLSLInstance() const { |
| 323 | return new GrGLMatrixConvolutionEffect; |
| 324 | } |
| 325 | |
| 326 | bool GrMatrixConvolutionEffect::onIsEqual(const GrFragmentProcessor& sBase) const { |
| 327 | const GrMatrixConvolutionEffect& s = sBase.cast<GrMatrixConvolutionEffect>(); |
| 328 | return fKernel == s.fKernel && |
| 329 | fGain == s.gain() && |
| 330 | fBias == s.bias() && |
| 331 | fKernelOffset == s.kernelOffset() && |
| 332 | fConvolveAlpha == s.convolveAlpha(); |
| 333 | } |
| 334 | |
| 335 | static void fill_in_1D_gaussian_kernel_with_stride(float* kernel, int size, int stride, |
| 336 | float twoSigmaSqrd) { |
| 337 | SkASSERT(!SkScalarNearlyZero(twoSigmaSqrd, SK_ScalarNearlyZero)); |
| 338 | |
| 339 | const float sigmaDenom = 1.0f / twoSigmaSqrd; |
| 340 | const int radius = size / 2; |
| 341 | |
| 342 | float sum = 0.0f; |
| 343 | for (int i = 0; i < size; ++i) { |
| 344 | float term = static_cast<float>(i - radius); |
| 345 | // Note that the constant term (1/(sqrt(2*pi*sigma^2)) of the Gaussian |
| 346 | // is dropped here, since we renormalize the kernel below. |
| 347 | kernel[i * stride] = sk_float_exp(-term * term * sigmaDenom); |
| 348 | sum += kernel[i * stride]; |
| 349 | } |
| 350 | // Normalize the kernel |
| 351 | float scale = 1.0f / sum; |
| 352 | for (int i = 0; i < size; ++i) { |
| 353 | kernel[i * stride] *= scale; |
| 354 | } |
| 355 | } |
| 356 | |
| 357 | static void fill_in_2D_gaussian_kernel(float* kernel, int width, int height, |
| 358 | SkScalar sigmaX, SkScalar sigmaY) { |
| 359 | const float twoSigmaSqrdX = 2.0f * SkScalarToFloat(SkScalarSquare(sigmaX)); |
| 360 | const float twoSigmaSqrdY = 2.0f * SkScalarToFloat(SkScalarSquare(sigmaY)); |
| 361 | |
| 362 | // TODO: in all of these degenerate cases we're uploading (and using) a whole lot of zeros. |
| 363 | if (SkScalarNearlyZero(twoSigmaSqrdX, SK_ScalarNearlyZero) || |
| 364 | SkScalarNearlyZero(twoSigmaSqrdY, SK_ScalarNearlyZero)) { |
| 365 | // In this case the 2D Gaussian degenerates to a 1D Gaussian (in X or Y) or a point |
| 366 | SkASSERT(3 == width || 3 == height); |
| 367 | std::fill_n(kernel, width*height, 0); |
| 368 | |
| 369 | if (SkScalarNearlyZero(twoSigmaSqrdX, SK_ScalarNearlyZero) && |
| 370 | SkScalarNearlyZero(twoSigmaSqrdY, SK_ScalarNearlyZero)) { |
| 371 | // A point |
| 372 | SkASSERT(3 == width && 3 == height); |
| 373 | kernel[4] = 1.0f; |
| 374 | } else if (SkScalarNearlyZero(twoSigmaSqrdX, SK_ScalarNearlyZero)) { |
| 375 | // A 1D Gaussian in Y |
| 376 | SkASSERT(3 == width); |
| 377 | // Down the middle column of the kernel with a stride of width |
| 378 | fill_in_1D_gaussian_kernel_with_stride(&kernel[1], height, width, twoSigmaSqrdY); |
| 379 | } else { |
| 380 | // A 1D Gaussian in X |
| 381 | SkASSERT(SkScalarNearlyZero(twoSigmaSqrdY, SK_ScalarNearlyZero)); |
| 382 | SkASSERT(3 == height); |
| 383 | // Down the middle row of the kernel with a stride of 1 |
| 384 | fill_in_1D_gaussian_kernel_with_stride(&kernel[width], width, 1, twoSigmaSqrdX); |
| 385 | } |
| 386 | return; |
| 387 | } |
| 388 | |
| 389 | const float sigmaXDenom = 1.0f / twoSigmaSqrdX; |
| 390 | const float sigmaYDenom = 1.0f / twoSigmaSqrdY; |
| 391 | const int xRadius = width / 2; |
| 392 | const int yRadius = height / 2; |
| 393 | |
| 394 | float sum = 0.0f; |
| 395 | for (int x = 0; x < width; x++) { |
| 396 | float xTerm = static_cast<float>(x - xRadius); |
| 397 | xTerm = xTerm * xTerm * sigmaXDenom; |
| 398 | for (int y = 0; y < height; y++) { |
| 399 | float yTerm = static_cast<float>(y - yRadius); |
| 400 | float xyTerm = sk_float_exp(-(xTerm + yTerm * yTerm * sigmaYDenom)); |
| 401 | // Note that the constant term (1/(sqrt(2*pi*sigma^2)) of the Gaussian |
| 402 | // is dropped here, since we renormalize the kernel below. |
| 403 | kernel[y * width + x] = xyTerm; |
| 404 | sum += xyTerm; |
| 405 | } |
| 406 | } |
| 407 | // Normalize the kernel |
| 408 | float scale = 1.0f / sum; |
| 409 | for (int i = 0; i < width * height; ++i) { |
| 410 | kernel[i] *= scale; |
| 411 | } |
| 412 | } |
| 413 | |
| 414 | std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::Make(GrRecordingContext* context, |
| 415 | GrSurfaceProxyView srcView, |
| 416 | const SkIRect& srcBounds, |
| 417 | const SkISize& kernelSize, |
| 418 | const SkScalar* kernel, |
| 419 | SkScalar gain, |
| 420 | SkScalar bias, |
| 421 | const SkIPoint& kernelOffset, |
| 422 | GrSamplerState::WrapMode wm, |
| 423 | bool convolveAlpha, |
| 424 | const GrCaps& caps) { |
| 425 | auto [kernelWrapper, kernelFP] = KernelWrapper::Make(context, kernelSize, caps, kernel); |
| 426 | if (!kernelWrapper.isValid()) { |
| 427 | return nullptr; |
| 428 | } |
| 429 | GrSamplerState sampler(wm, GrSamplerState::Filter::kNearest); |
| 430 | auto child = GrTextureEffect::MakeSubset(std::move(srcView), kPremul_SkAlphaType, SkMatrix::I(), |
| 431 | sampler, SkRect::Make(srcBounds), caps); |
| 432 | return std::unique_ptr<GrFragmentProcessor>( |
| 433 | new GrMatrixConvolutionEffect(std::move(child), kernelWrapper, std::move(kernelFP), |
| 434 | gain, bias, kernelOffset, convolveAlpha)); |
| 435 | } |
| 436 | |
| 437 | std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::MakeGaussian( |
| 438 | GrRecordingContext* context, |
| 439 | GrSurfaceProxyView srcView, |
| 440 | const SkIRect& srcBounds, |
| 441 | const SkISize& kernelSize, |
| 442 | SkScalar gain, |
| 443 | SkScalar bias, |
| 444 | const SkIPoint& kernelOffset, |
| 445 | GrSamplerState::WrapMode wm, |
| 446 | bool convolveAlpha, |
| 447 | SkScalar sigmaX, |
| 448 | SkScalar sigmaY, |
| 449 | const GrCaps& caps) { |
| 450 | SkAutoSTMalloc<32, float> kernel(kernelSize.area()); |
| 451 | fill_in_2D_gaussian_kernel(kernel.get(), kernelSize.width(), kernelSize.height(), |
| 452 | sigmaX, sigmaY); |
| 453 | return Make(context, std::move(srcView), srcBounds, kernelSize, kernel.get(), |
| 454 | gain, bias, kernelOffset, wm, convolveAlpha, caps); |
| 455 | } |
| 456 | |
| 457 | GR_DEFINE_FRAGMENT_PROCESSOR_TEST(GrMatrixConvolutionEffect); |
| 458 | |
| 459 | #if GR_TEST_UTILS |
| 460 | std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::TestCreate(GrProcessorTestData* d) { |
| 461 | auto [view, ct, at] = d->randomView(); |
| 462 | |
| 463 | static constexpr size_t kMaxTestKernelSize = 2 * kMaxUniformSize; |
| 464 | int width = d->fRandom->nextRangeU(1, kMaxTestKernelSize); |
| 465 | int height = d->fRandom->nextRangeU(1, kMaxTestKernelSize / width); |
| 466 | SkISize kernelSize = SkISize::Make(width, height); |
| 467 | std::unique_ptr<SkScalar[]> kernel(new SkScalar[width * height]); |
| 468 | for (int i = 0; i < width * height; i++) { |
| 469 | kernel.get()[i] = d->fRandom->nextSScalar1(); |
| 470 | } |
| 471 | SkScalar gain = d->fRandom->nextSScalar1(); |
| 472 | SkScalar bias = d->fRandom->nextSScalar1(); |
| 473 | |
| 474 | uint32_t kernalOffsetX = d->fRandom->nextRangeU(0, kernelSize.width()); |
| 475 | uint32_t kernalOffsetY = d->fRandom->nextRangeU(0, kernelSize.height()); |
| 476 | SkIPoint kernelOffset = SkIPoint::Make(kernalOffsetX, kernalOffsetY); |
| 477 | |
| 478 | uint32_t boundsX = d->fRandom->nextRangeU(0, view.width()); |
| 479 | uint32_t boundsY = d->fRandom->nextRangeU(0, view.height()); |
| 480 | uint32_t boundsW = d->fRandom->nextRangeU(0, view.width()); |
| 481 | uint32_t boundsH = d->fRandom->nextRangeU(0, view.height()); |
| 482 | SkIRect bounds = SkIRect::MakeXYWH(boundsX, boundsY, boundsW, boundsH); |
| 483 | |
| 484 | auto wm = static_cast<GrSamplerState::WrapMode>( |
| 485 | d->fRandom->nextULessThan(GrSamplerState::kWrapModeCount)); |
| 486 | bool convolveAlpha = d->fRandom->nextBool(); |
| 487 | return GrMatrixConvolutionEffect::Make(d->context(), |
| 488 | std::move(view), |
| 489 | bounds, |
| 490 | kernelSize, |
| 491 | kernel.get(), |
| 492 | gain, |
| 493 | bias, |
| 494 | kernelOffset, |
| 495 | wm, |
| 496 | convolveAlpha, |
| 497 | *d->caps()); |
| 498 | } |
| 499 | #endif |
| 500 | |