| 1 | /* |
| 2 | * Copyright 2018 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 | |
| 8 | /************************************************************************************************** |
| 9 | *** This file was autogenerated from GrCircleBlurFragmentProcessor.fp; do not modify. |
| 10 | **************************************************************************************************/ |
| 11 | #include "GrCircleBlurFragmentProcessor.h" |
| 12 | |
| 13 | #include "include/gpu/GrRecordingContext.h" |
| 14 | #include "src/gpu/GrBitmapTextureMaker.h" |
| 15 | #include "src/gpu/GrProxyProvider.h" |
| 16 | #include "src/gpu/GrRecordingContextPriv.h" |
| 17 | |
| 18 | // Computes an unnormalized half kernel (right side). Returns the summation of all the half |
| 19 | // kernel values. |
| 20 | static float make_unnormalized_half_kernel(float* halfKernel, int halfKernelSize, float sigma) { |
| 21 | const float invSigma = 1.f / sigma; |
| 22 | const float b = -0.5f * invSigma * invSigma; |
| 23 | float tot = 0.0f; |
| 24 | // Compute half kernel values at half pixel steps out from the center. |
| 25 | float t = 0.5f; |
| 26 | for (int i = 0; i < halfKernelSize; ++i) { |
| 27 | float value = expf(t * t * b); |
| 28 | tot += value; |
| 29 | halfKernel[i] = value; |
| 30 | t += 1.f; |
| 31 | } |
| 32 | return tot; |
| 33 | } |
| 34 | |
| 35 | // Create a Gaussian half-kernel (right side) and a summed area table given a sigma and number |
| 36 | // of discrete steps. The half kernel is normalized to sum to 0.5. |
| 37 | static void make_half_kernel_and_summed_table(float* halfKernel, |
| 38 | float* summedHalfKernel, |
| 39 | int halfKernelSize, |
| 40 | float sigma) { |
| 41 | // The half kernel should sum to 0.5 not 1.0. |
| 42 | const float tot = 2.f * make_unnormalized_half_kernel(halfKernel, halfKernelSize, sigma); |
| 43 | float sum = 0.f; |
| 44 | for (int i = 0; i < halfKernelSize; ++i) { |
| 45 | halfKernel[i] /= tot; |
| 46 | sum += halfKernel[i]; |
| 47 | summedHalfKernel[i] = sum; |
| 48 | } |
| 49 | } |
| 50 | |
| 51 | // Applies the 1D half kernel vertically at points along the x axis to a circle centered at the |
| 52 | // origin with radius circleR. |
| 53 | void apply_kernel_in_y(float* results, |
| 54 | int numSteps, |
| 55 | float firstX, |
| 56 | float circleR, |
| 57 | int halfKernelSize, |
| 58 | const float* summedHalfKernelTable) { |
| 59 | float x = firstX; |
| 60 | for (int i = 0; i < numSteps; ++i, x += 1.f) { |
| 61 | if (x < -circleR || x > circleR) { |
| 62 | results[i] = 0; |
| 63 | continue; |
| 64 | } |
| 65 | float y = sqrtf(circleR * circleR - x * x); |
| 66 | // In the column at x we exit the circle at +y and -y |
| 67 | // The summed table entry j is actually reflects an offset of j + 0.5. |
| 68 | y -= 0.5f; |
| 69 | int yInt = SkScalarFloorToInt(y); |
| 70 | SkASSERT(yInt >= -1); |
| 71 | if (y < 0) { |
| 72 | results[i] = (y + 0.5f) * summedHalfKernelTable[0]; |
| 73 | } else if (yInt >= halfKernelSize - 1) { |
| 74 | results[i] = 0.5f; |
| 75 | } else { |
| 76 | float yFrac = y - yInt; |
| 77 | results[i] = (1.f - yFrac) * summedHalfKernelTable[yInt] + |
| 78 | yFrac * summedHalfKernelTable[yInt + 1]; |
| 79 | } |
| 80 | } |
| 81 | } |
| 82 | |
| 83 | // Apply a Gaussian at point (evalX, 0) to a circle centered at the origin with radius circleR. |
| 84 | // This relies on having a half kernel computed for the Gaussian and a table of applications of |
| 85 | // the half kernel in y to columns at (evalX - halfKernel, evalX - halfKernel + 1, ..., evalX + |
| 86 | // halfKernel) passed in as yKernelEvaluations. |
| 87 | static uint8_t eval_at(float evalX, |
| 88 | float circleR, |
| 89 | const float* halfKernel, |
| 90 | int halfKernelSize, |
| 91 | const float* yKernelEvaluations) { |
| 92 | float acc = 0; |
| 93 | |
| 94 | float x = evalX - halfKernelSize; |
| 95 | for (int i = 0; i < halfKernelSize; ++i, x += 1.f) { |
| 96 | if (x < -circleR || x > circleR) { |
| 97 | continue; |
| 98 | } |
| 99 | float verticalEval = yKernelEvaluations[i]; |
| 100 | acc += verticalEval * halfKernel[halfKernelSize - i - 1]; |
| 101 | } |
| 102 | for (int i = 0; i < halfKernelSize; ++i, x += 1.f) { |
| 103 | if (x < -circleR || x > circleR) { |
| 104 | continue; |
| 105 | } |
| 106 | float verticalEval = yKernelEvaluations[i + halfKernelSize]; |
| 107 | acc += verticalEval * halfKernel[i]; |
| 108 | } |
| 109 | // Since we applied a half kernel in y we multiply acc by 2 (the circle is symmetric about |
| 110 | // the x axis). |
| 111 | return SkUnitScalarClampToByte(2.f * acc); |
| 112 | } |
| 113 | |
| 114 | // This function creates a profile of a blurred circle. It does this by computing a kernel for |
| 115 | // half the Gaussian and a matching summed area table. The summed area table is used to compute |
| 116 | // an array of vertical applications of the half kernel to the circle along the x axis. The |
| 117 | // table of y evaluations has 2 * k + n entries where k is the size of the half kernel and n is |
| 118 | // the size of the profile being computed. Then for each of the n profile entries we walk out k |
| 119 | // steps in each horizontal direction multiplying the corresponding y evaluation by the half |
| 120 | // kernel entry and sum these values to compute the profile entry. |
| 121 | static void create_circle_profile(uint8_t* weights, |
| 122 | float sigma, |
| 123 | float circleR, |
| 124 | int profileTextureWidth) { |
| 125 | const int numSteps = profileTextureWidth; |
| 126 | |
| 127 | // The full kernel is 6 sigmas wide. |
| 128 | int halfKernelSize = SkScalarCeilToInt(6.0f * sigma); |
| 129 | // round up to next multiple of 2 and then divide by 2 |
| 130 | halfKernelSize = ((halfKernelSize + 1) & ~1) >> 1; |
| 131 | |
| 132 | // Number of x steps at which to apply kernel in y to cover all the profile samples in x. |
| 133 | int numYSteps = numSteps + 2 * halfKernelSize; |
| 134 | |
| 135 | SkAutoTArray<float> bulkAlloc(halfKernelSize + halfKernelSize + numYSteps); |
| 136 | float* halfKernel = bulkAlloc.get(); |
| 137 | float* summedKernel = bulkAlloc.get() + halfKernelSize; |
| 138 | float* yEvals = bulkAlloc.get() + 2 * halfKernelSize; |
| 139 | make_half_kernel_and_summed_table(halfKernel, summedKernel, halfKernelSize, sigma); |
| 140 | |
| 141 | float firstX = -halfKernelSize + 0.5f; |
| 142 | apply_kernel_in_y(yEvals, numYSteps, firstX, circleR, halfKernelSize, summedKernel); |
| 143 | |
| 144 | for (int i = 0; i < numSteps - 1; ++i) { |
| 145 | float evalX = i + 0.5f; |
| 146 | weights[i] = eval_at(evalX, circleR, halfKernel, halfKernelSize, yEvals + i); |
| 147 | } |
| 148 | // Ensure the tail of the Gaussian goes to zero. |
| 149 | weights[numSteps - 1] = 0; |
| 150 | } |
| 151 | |
| 152 | static void create_half_plane_profile(uint8_t* profile, int profileWidth) { |
| 153 | SkASSERT(!(profileWidth & 0x1)); |
| 154 | // The full kernel is 6 sigmas wide. |
| 155 | float sigma = profileWidth / 6.f; |
| 156 | int halfKernelSize = profileWidth / 2; |
| 157 | |
| 158 | SkAutoTArray<float> halfKernel(halfKernelSize); |
| 159 | |
| 160 | // The half kernel should sum to 0.5. |
| 161 | const float tot = 2.f * make_unnormalized_half_kernel(halfKernel.get(), halfKernelSize, sigma); |
| 162 | float sum = 0.f; |
| 163 | // Populate the profile from the right edge to the middle. |
| 164 | for (int i = 0; i < halfKernelSize; ++i) { |
| 165 | halfKernel[halfKernelSize - i - 1] /= tot; |
| 166 | sum += halfKernel[halfKernelSize - i - 1]; |
| 167 | profile[profileWidth - i - 1] = SkUnitScalarClampToByte(sum); |
| 168 | } |
| 169 | // Populate the profile from the middle to the left edge (by flipping the half kernel and |
| 170 | // continuing the summation). |
| 171 | for (int i = 0; i < halfKernelSize; ++i) { |
| 172 | sum += halfKernel[i]; |
| 173 | profile[halfKernelSize - i - 1] = SkUnitScalarClampToByte(sum); |
| 174 | } |
| 175 | // Ensure tail goes to 0. |
| 176 | profile[profileWidth - 1] = 0; |
| 177 | } |
| 178 | |
| 179 | static std::unique_ptr<GrFragmentProcessor> create_profile_effect(GrRecordingContext* context, |
| 180 | const SkRect& circle, |
| 181 | float sigma, |
| 182 | float* solidRadius, |
| 183 | float* textureRadius) { |
| 184 | float circleR = circle.width() / 2.0f; |
| 185 | if (circleR < SK_ScalarNearlyZero) { |
| 186 | return nullptr; |
| 187 | } |
| 188 | // Profile textures are cached by the ratio of sigma to circle radius and by the size of the |
| 189 | // profile texture (binned by powers of 2). |
| 190 | SkScalar sigmaToCircleRRatio = sigma / circleR; |
| 191 | // When sigma is really small this becomes a equivalent to convolving a Gaussian with a |
| 192 | // half-plane. Similarly, in the extreme high ratio cases circle becomes a point WRT to the |
| 193 | // Guassian and the profile texture is a just a Gaussian evaluation. However, we haven't yet |
| 194 | // implemented this latter optimization. |
| 195 | sigmaToCircleRRatio = std::min(sigmaToCircleRRatio, 8.f); |
| 196 | SkFixed sigmaToCircleRRatioFixed; |
| 197 | static const SkScalar kHalfPlaneThreshold = 0.1f; |
| 198 | bool useHalfPlaneApprox = false; |
| 199 | if (sigmaToCircleRRatio <= kHalfPlaneThreshold) { |
| 200 | useHalfPlaneApprox = true; |
| 201 | sigmaToCircleRRatioFixed = 0; |
| 202 | *solidRadius = circleR - 3 * sigma; |
| 203 | *textureRadius = 6 * sigma; |
| 204 | } else { |
| 205 | // Convert to fixed point for the key. |
| 206 | sigmaToCircleRRatioFixed = SkScalarToFixed(sigmaToCircleRRatio); |
| 207 | // We shave off some bits to reduce the number of unique entries. We could probably |
| 208 | // shave off more than we do. |
| 209 | sigmaToCircleRRatioFixed &= ~0xff; |
| 210 | sigmaToCircleRRatio = SkFixedToScalar(sigmaToCircleRRatioFixed); |
| 211 | sigma = circleR * sigmaToCircleRRatio; |
| 212 | *solidRadius = 0; |
| 213 | *textureRadius = circleR + 3 * sigma; |
| 214 | } |
| 215 | |
| 216 | static constexpr int kProfileTextureWidth = 512; |
| 217 | // This would be kProfileTextureWidth/textureRadius if it weren't for the fact that we do |
| 218 | // the calculation of the profile coord in a coord space that has already been scaled by |
| 219 | // 1 / textureRadius. This is done to avoid overflow in length(). |
| 220 | SkMatrix texM = SkMatrix::Scale(kProfileTextureWidth, 1.f); |
| 221 | |
| 222 | static const GrUniqueKey::Domain kDomain = GrUniqueKey::GenerateDomain(); |
| 223 | GrUniqueKey key; |
| 224 | GrUniqueKey::Builder builder(&key, kDomain, 1, "1-D Circular Blur" ); |
| 225 | builder[0] = sigmaToCircleRRatioFixed; |
| 226 | builder.finish(); |
| 227 | |
| 228 | GrProxyProvider* proxyProvider = context->priv().proxyProvider(); |
| 229 | if (sk_sp<GrTextureProxy> blurProfile = proxyProvider->findOrCreateProxyByUniqueKey(key)) { |
| 230 | GrSwizzle swizzle = context->priv().caps()->getReadSwizzle(blurProfile->backendFormat(), |
| 231 | GrColorType::kAlpha_8); |
| 232 | GrSurfaceProxyView profileView{std::move(blurProfile), kTopLeft_GrSurfaceOrigin, swizzle}; |
| 233 | return GrTextureEffect::Make(std::move(profileView), kPremul_SkAlphaType, texM); |
| 234 | } |
| 235 | |
| 236 | SkBitmap bm; |
| 237 | if (!bm.tryAllocPixels(SkImageInfo::MakeA8(kProfileTextureWidth, 1))) { |
| 238 | return nullptr; |
| 239 | } |
| 240 | |
| 241 | if (useHalfPlaneApprox) { |
| 242 | create_half_plane_profile(bm.getAddr8(0, 0), kProfileTextureWidth); |
| 243 | } else { |
| 244 | // Rescale params to the size of the texture we're creating. |
| 245 | SkScalar scale = kProfileTextureWidth / *textureRadius; |
| 246 | create_circle_profile(bm.getAddr8(0, 0), sigma * scale, circleR * scale, |
| 247 | kProfileTextureWidth); |
| 248 | } |
| 249 | |
| 250 | bm.setImmutable(); |
| 251 | |
| 252 | GrBitmapTextureMaker maker(context, bm, GrImageTexGenPolicy::kNew_Uncached_Budgeted); |
| 253 | auto profileView = maker.view(GrMipmapped::kNo); |
| 254 | if (!profileView) { |
| 255 | return nullptr; |
| 256 | } |
| 257 | proxyProvider->assignUniqueKeyToProxy(key, profileView.asTextureProxy()); |
| 258 | return GrTextureEffect::Make(std::move(profileView), kPremul_SkAlphaType, texM); |
| 259 | } |
| 260 | |
| 261 | std::unique_ptr<GrFragmentProcessor> GrCircleBlurFragmentProcessor::Make( |
| 262 | std::unique_ptr<GrFragmentProcessor> inputFP, |
| 263 | GrRecordingContext* context, |
| 264 | const SkRect& circle, |
| 265 | float sigma) { |
| 266 | float solidRadius; |
| 267 | float textureRadius; |
| 268 | std::unique_ptr<GrFragmentProcessor> profile = |
| 269 | create_profile_effect(context, circle, sigma, &solidRadius, &textureRadius); |
| 270 | if (!profile) { |
| 271 | return nullptr; |
| 272 | } |
| 273 | return std::unique_ptr<GrFragmentProcessor>(new GrCircleBlurFragmentProcessor( |
| 274 | std::move(inputFP), circle, solidRadius, textureRadius, std::move(profile))); |
| 275 | } |
| 276 | #include "src/core/SkUtils.h" |
| 277 | #include "src/gpu/GrTexture.h" |
| 278 | #include "src/gpu/glsl/GrGLSLFragmentProcessor.h" |
| 279 | #include "src/gpu/glsl/GrGLSLFragmentShaderBuilder.h" |
| 280 | #include "src/gpu/glsl/GrGLSLProgramBuilder.h" |
| 281 | #include "src/sksl/SkSLCPP.h" |
| 282 | #include "src/sksl/SkSLUtil.h" |
| 283 | class GrGLSLCircleBlurFragmentProcessor : public GrGLSLFragmentProcessor { |
| 284 | public: |
| 285 | GrGLSLCircleBlurFragmentProcessor() {} |
| 286 | void emitCode(EmitArgs& args) override { |
| 287 | GrGLSLFPFragmentBuilder* fragBuilder = args.fFragBuilder; |
| 288 | const GrCircleBlurFragmentProcessor& _outer = |
| 289 | args.fFp.cast<GrCircleBlurFragmentProcessor>(); |
| 290 | (void)_outer; |
| 291 | auto circleRect = _outer.circleRect; |
| 292 | (void)circleRect; |
| 293 | auto solidRadius = _outer.solidRadius; |
| 294 | (void)solidRadius; |
| 295 | auto textureRadius = _outer.textureRadius; |
| 296 | (void)textureRadius; |
| 297 | circleDataVar = args.fUniformHandler->addUniform(&_outer, kFragment_GrShaderFlag, |
| 298 | kHalf4_GrSLType, "circleData" ); |
| 299 | fragBuilder->codeAppendf( |
| 300 | R"SkSL(; |
| 301 | half2 vec = half2((sk_FragCoord.xy - float2(%s.xy)) * float(%s.w)); |
| 302 | half dist = length(vec) + (0.5 - %s.z) * %s.w;)SkSL" , |
| 303 | args.fUniformHandler->getUniformCStr(circleDataVar), |
| 304 | args.fUniformHandler->getUniformCStr(circleDataVar), |
| 305 | args.fUniformHandler->getUniformCStr(circleDataVar), |
| 306 | args.fUniformHandler->getUniformCStr(circleDataVar)); |
| 307 | SkString _sample13902 = this->invokeChild(0, args); |
| 308 | fragBuilder->codeAppendf( |
| 309 | R"SkSL( |
| 310 | half4 inputColor = %s;)SkSL" , |
| 311 | _sample13902.c_str()); |
| 312 | SkString _coords13950("float2(half2(dist, 0.5))" ); |
| 313 | SkString _sample13950 = this->invokeChild(1, args, _coords13950.c_str()); |
| 314 | fragBuilder->codeAppendf( |
| 315 | R"SkSL( |
| 316 | %s = inputColor * %s.w; |
| 317 | )SkSL" , |
| 318 | args.fOutputColor, _sample13950.c_str()); |
| 319 | } |
| 320 | |
| 321 | private: |
| 322 | void onSetData(const GrGLSLProgramDataManager& data, |
| 323 | const GrFragmentProcessor& _proc) override { |
| 324 | const GrCircleBlurFragmentProcessor& _outer = _proc.cast<GrCircleBlurFragmentProcessor>(); |
| 325 | auto circleRect = _outer.circleRect; |
| 326 | (void)circleRect; |
| 327 | auto solidRadius = _outer.solidRadius; |
| 328 | (void)solidRadius; |
| 329 | auto textureRadius = _outer.textureRadius; |
| 330 | (void)textureRadius; |
| 331 | UniformHandle& circleData = circleDataVar; |
| 332 | (void)circleData; |
| 333 | |
| 334 | data.set4f(circleData, circleRect.centerX(), circleRect.centerY(), solidRadius, |
| 335 | 1.f / textureRadius); |
| 336 | } |
| 337 | UniformHandle circleDataVar; |
| 338 | }; |
| 339 | GrGLSLFragmentProcessor* GrCircleBlurFragmentProcessor::onCreateGLSLInstance() const { |
| 340 | return new GrGLSLCircleBlurFragmentProcessor(); |
| 341 | } |
| 342 | void GrCircleBlurFragmentProcessor::onGetGLSLProcessorKey(const GrShaderCaps& caps, |
| 343 | GrProcessorKeyBuilder* b) const {} |
| 344 | bool GrCircleBlurFragmentProcessor::onIsEqual(const GrFragmentProcessor& other) const { |
| 345 | const GrCircleBlurFragmentProcessor& that = other.cast<GrCircleBlurFragmentProcessor>(); |
| 346 | (void)that; |
| 347 | if (circleRect != that.circleRect) return false; |
| 348 | if (solidRadius != that.solidRadius) return false; |
| 349 | if (textureRadius != that.textureRadius) return false; |
| 350 | return true; |
| 351 | } |
| 352 | GrCircleBlurFragmentProcessor::GrCircleBlurFragmentProcessor( |
| 353 | const GrCircleBlurFragmentProcessor& src) |
| 354 | : INHERITED(kGrCircleBlurFragmentProcessor_ClassID, src.optimizationFlags()) |
| 355 | , circleRect(src.circleRect) |
| 356 | , solidRadius(src.solidRadius) |
| 357 | , textureRadius(src.textureRadius) { |
| 358 | this->cloneAndRegisterAllChildProcessors(src); |
| 359 | } |
| 360 | std::unique_ptr<GrFragmentProcessor> GrCircleBlurFragmentProcessor::clone() const { |
| 361 | return std::make_unique<GrCircleBlurFragmentProcessor>(*this); |
| 362 | } |
| 363 | #if GR_TEST_UTILS |
| 364 | SkString GrCircleBlurFragmentProcessor::onDumpInfo() const { |
| 365 | return SkStringPrintf("(circleRect=half4(%f, %f, %f, %f), solidRadius=%f, textureRadius=%f)" , |
| 366 | circleRect.left(), circleRect.top(), circleRect.right(), |
| 367 | circleRect.bottom(), solidRadius, textureRadius); |
| 368 | } |
| 369 | #endif |
| 370 | GR_DEFINE_FRAGMENT_PROCESSOR_TEST(GrCircleBlurFragmentProcessor); |
| 371 | #if GR_TEST_UTILS |
| 372 | std::unique_ptr<GrFragmentProcessor> GrCircleBlurFragmentProcessor::TestCreate( |
| 373 | GrProcessorTestData* testData) { |
| 374 | SkScalar wh = testData->fRandom->nextRangeScalar(100.f, 1000.f); |
| 375 | SkScalar sigma = testData->fRandom->nextRangeF(1.f, 10.f); |
| 376 | SkRect circle = SkRect::MakeWH(wh, wh); |
| 377 | return GrCircleBlurFragmentProcessor::Make(testData->inputFP(), testData->context(), circle, |
| 378 | sigma); |
| 379 | } |
| 380 | #endif |
| 381 | |