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 "src/gpu/GrTexture.h" |
10 | #include "src/gpu/GrTextureProxy.h" |
11 | #include "src/gpu/glsl/GrGLSLFragmentProcessor.h" |
12 | #include "src/gpu/glsl/GrGLSLFragmentShaderBuilder.h" |
13 | #include "src/gpu/glsl/GrGLSLProgramDataManager.h" |
14 | #include "src/gpu/glsl/GrGLSLUniformHandler.h" |
15 | |
16 | class GrGLMatrixConvolutionEffect : public GrGLSLFragmentProcessor { |
17 | public: |
18 | void emitCode(EmitArgs&) override; |
19 | |
20 | static inline void GenKey(const GrProcessor&, const GrShaderCaps&, GrProcessorKeyBuilder*); |
21 | |
22 | protected: |
23 | void onSetData(const GrGLSLProgramDataManager&, const GrFragmentProcessor&) override; |
24 | |
25 | private: |
26 | typedef GrGLSLProgramDataManager::UniformHandle UniformHandle; |
27 | |
28 | UniformHandle fKernelUni; |
29 | UniformHandle fImageIncrementUni; |
30 | UniformHandle fKernelOffsetUni; |
31 | UniformHandle fGainUni; |
32 | UniformHandle fBiasUni; |
33 | GrTextureDomain::GLDomain fDomain; |
34 | |
35 | typedef GrGLSLFragmentProcessor INHERITED; |
36 | }; |
37 | |
38 | void GrGLMatrixConvolutionEffect::emitCode(EmitArgs& args) { |
39 | const GrMatrixConvolutionEffect& mce = args.fFp.cast<GrMatrixConvolutionEffect>(); |
40 | const GrTextureDomain& domain = mce.domain(); |
41 | |
42 | int kWidth = mce.kernelSize().width(); |
43 | int kHeight = mce.kernelSize().height(); |
44 | |
45 | int arrayCount = (kWidth * kHeight + 3) / 4; |
46 | SkASSERT(4 * arrayCount >= kWidth * kHeight); |
47 | |
48 | GrGLSLUniformHandler* uniformHandler = args.fUniformHandler; |
49 | fImageIncrementUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf2_GrSLType, |
50 | "ImageIncrement" ); |
51 | fKernelUni = uniformHandler->addUniformArray(&mce, kFragment_GrShaderFlag, kHalf4_GrSLType, |
52 | "Kernel" , |
53 | arrayCount); |
54 | fKernelOffsetUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf2_GrSLType, |
55 | "KernelOffset" ); |
56 | fGainUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf_GrSLType, "Gain" ); |
57 | fBiasUni = uniformHandler->addUniform(&mce, kFragment_GrShaderFlag, kHalf_GrSLType, "Bias" ); |
58 | |
59 | const char* kernelOffset = uniformHandler->getUniformCStr(fKernelOffsetUni); |
60 | const char* imgInc = uniformHandler->getUniformCStr(fImageIncrementUni); |
61 | const char* kernel = uniformHandler->getUniformCStr(fKernelUni); |
62 | const char* gain = uniformHandler->getUniformCStr(fGainUni); |
63 | const char* bias = uniformHandler->getUniformCStr(fBiasUni); |
64 | |
65 | GrGLSLFPFragmentBuilder* fragBuilder = args.fFragBuilder; |
66 | SkString coords2D = fragBuilder->ensureCoords2D(args.fTransformedCoords[0].fVaryingPoint); |
67 | fragBuilder->codeAppend("half4 sum = half4(0, 0, 0, 0);" ); |
68 | fragBuilder->codeAppendf("float2 coord = %s - %s * %s;" , coords2D.c_str(), kernelOffset, imgInc); |
69 | fragBuilder->codeAppend("half4 c;" ); |
70 | |
71 | const char* kVecSuffix[4] = { ".x" , ".y" , ".z" , ".w" }; |
72 | for (int y = 0; y < kHeight; y++) { |
73 | for (int x = 0; x < kWidth; x++) { |
74 | GrGLSLShaderBuilder::ShaderBlock block(fragBuilder); |
75 | int offset = y*kWidth + x; |
76 | |
77 | fragBuilder->codeAppendf("half k = %s[%d]%s;" , kernel, offset / 4, |
78 | kVecSuffix[offset & 0x3]); |
79 | SkString coord; |
80 | coord.printf("coord + half2(%d, %d) * %s" , x, y, imgInc); |
81 | fDomain.sampleTexture(&mce, |
82 | fragBuilder, |
83 | uniformHandler, |
84 | args.fShaderCaps, |
85 | domain, |
86 | "c" , |
87 | coord, |
88 | args.fTexSamplers[0]); |
89 | if (!mce.convolveAlpha()) { |
90 | fragBuilder->codeAppend("c.rgb /= c.a;" ); |
91 | fragBuilder->codeAppend("c.rgb = saturate(c.rgb);" ); |
92 | } |
93 | fragBuilder->codeAppend("sum += c * k;" ); |
94 | } |
95 | } |
96 | if (mce.convolveAlpha()) { |
97 | fragBuilder->codeAppendf("%s = sum * %s + %s;" , args.fOutputColor, gain, bias); |
98 | fragBuilder->codeAppendf("%s.a = saturate(%s.a);" , args.fOutputColor, args.fOutputColor); |
99 | fragBuilder->codeAppendf("%s.rgb = clamp(%s.rgb, 0.0, %s.a);" , |
100 | args.fOutputColor, args.fOutputColor, args.fOutputColor); |
101 | } else { |
102 | fDomain.sampleTexture(&mce, |
103 | fragBuilder, |
104 | uniformHandler, |
105 | args.fShaderCaps, |
106 | domain, |
107 | "c" , |
108 | coords2D, |
109 | args.fTexSamplers[0]); |
110 | fragBuilder->codeAppendf("%s.a = c.a;" , args.fOutputColor); |
111 | fragBuilder->codeAppendf("%s.rgb = saturate(sum.rgb * %s + %s);" , args.fOutputColor, gain, bias); |
112 | fragBuilder->codeAppendf("%s.rgb *= %s.a;" , args.fOutputColor, args.fOutputColor); |
113 | } |
114 | fragBuilder->codeAppendf("%s *= %s;\n" , args.fOutputColor, args.fInputColor); |
115 | } |
116 | |
117 | void GrGLMatrixConvolutionEffect::GenKey(const GrProcessor& processor, |
118 | const GrShaderCaps&, GrProcessorKeyBuilder* b) { |
119 | const GrMatrixConvolutionEffect& m = processor.cast<GrMatrixConvolutionEffect>(); |
120 | SkASSERT(m.kernelSize().width() <= 0x7FFF && m.kernelSize().height() <= 0xFFFF); |
121 | uint32_t key = m.kernelSize().width() << 16 | m.kernelSize().height(); |
122 | key |= m.convolveAlpha() ? 1U << 31 : 0; |
123 | b->add32(key); |
124 | b->add32(GrTextureDomain::GLDomain::DomainKey(m.domain())); |
125 | } |
126 | |
127 | void GrGLMatrixConvolutionEffect::onSetData(const GrGLSLProgramDataManager& pdman, |
128 | const GrFragmentProcessor& processor) { |
129 | const GrMatrixConvolutionEffect& conv = processor.cast<GrMatrixConvolutionEffect>(); |
130 | const auto& view = conv.textureSampler(0).view(); |
131 | SkISize textureDims = view.proxy()->backingStoreDimensions(); |
132 | |
133 | float imageIncrement[2]; |
134 | float ySign = view.origin() == kTopLeft_GrSurfaceOrigin ? 1.0f : -1.0f; |
135 | imageIncrement[0] = 1.0f / textureDims.width(); |
136 | imageIncrement[1] = ySign / textureDims.height(); |
137 | pdman.set2fv(fImageIncrementUni, 1, imageIncrement); |
138 | pdman.set2fv(fKernelOffsetUni, 1, conv.kernelOffset()); |
139 | int kernelCount = conv.kernelSize().width() * conv.kernelSize().height(); |
140 | int arrayCount = (kernelCount + 3) / 4; |
141 | SkASSERT(4 * arrayCount >= kernelCount); |
142 | pdman.set4fv(fKernelUni, arrayCount, conv.kernel()); |
143 | pdman.set1f(fGainUni, conv.gain()); |
144 | pdman.set1f(fBiasUni, conv.bias()); |
145 | fDomain.setData(pdman, conv.domain(), view, conv.textureSampler(0).samplerState()); |
146 | } |
147 | |
148 | GrMatrixConvolutionEffect::GrMatrixConvolutionEffect(GrSurfaceProxyView srcView, |
149 | const SkIRect& srcBounds, |
150 | const SkISize& kernelSize, |
151 | const SkScalar* kernel, |
152 | SkScalar gain, |
153 | SkScalar bias, |
154 | const SkIPoint& kernelOffset, |
155 | GrTextureDomain::Mode tileMode, |
156 | bool convolveAlpha) |
157 | // To advertise either the modulation or opaqueness optimizations we'd have to examine the |
158 | // parameters. |
159 | : INHERITED(kGrMatrixConvolutionEffect_ClassID, kNone_OptimizationFlags) |
160 | , fCoordTransform(srcView.proxy(), srcView.origin()) |
161 | , fDomain(srcView.proxy(), GrTextureDomain::MakeTexelDomain(srcBounds, tileMode), |
162 | tileMode, tileMode) |
163 | , fTextureSampler(std::move(srcView)) |
164 | , fKernelSize(kernelSize) |
165 | , fGain(SkScalarToFloat(gain)) |
166 | , fBias(SkScalarToFloat(bias) / 255.0f) |
167 | , fConvolveAlpha(convolveAlpha) { |
168 | this->addCoordTransform(&fCoordTransform); |
169 | this->setTextureSamplerCnt(1); |
170 | for (int i = 0; i < kernelSize.width() * kernelSize.height(); i++) { |
171 | fKernel[i] = SkScalarToFloat(kernel[i]); |
172 | } |
173 | fKernelOffset[0] = static_cast<float>(kernelOffset.x()); |
174 | fKernelOffset[1] = static_cast<float>(kernelOffset.y()); |
175 | } |
176 | |
177 | GrMatrixConvolutionEffect::GrMatrixConvolutionEffect(const GrMatrixConvolutionEffect& that) |
178 | : INHERITED(kGrMatrixConvolutionEffect_ClassID, kNone_OptimizationFlags) |
179 | , fCoordTransform(that.fCoordTransform) |
180 | , fDomain(that.fDomain) |
181 | , fTextureSampler(that.fTextureSampler) |
182 | , fKernelSize(that.fKernelSize) |
183 | , fGain(that.fGain) |
184 | , fBias(that.fBias) |
185 | , fConvolveAlpha(that.fConvolveAlpha) { |
186 | this->addCoordTransform(&fCoordTransform); |
187 | this->setTextureSamplerCnt(1); |
188 | memcpy(fKernel, that.fKernel, sizeof(float) * fKernelSize.width() * fKernelSize.height()); |
189 | memcpy(fKernelOffset, that.fKernelOffset, sizeof(fKernelOffset)); |
190 | } |
191 | |
192 | std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::clone() const { |
193 | return std::unique_ptr<GrFragmentProcessor>(new GrMatrixConvolutionEffect(*this)); |
194 | } |
195 | |
196 | void GrMatrixConvolutionEffect::onGetGLSLProcessorKey(const GrShaderCaps& caps, |
197 | GrProcessorKeyBuilder* b) const { |
198 | GrGLMatrixConvolutionEffect::GenKey(*this, caps, b); |
199 | } |
200 | |
201 | GrGLSLFragmentProcessor* GrMatrixConvolutionEffect::onCreateGLSLInstance() const { |
202 | return new GrGLMatrixConvolutionEffect; |
203 | } |
204 | |
205 | bool GrMatrixConvolutionEffect::onIsEqual(const GrFragmentProcessor& sBase) const { |
206 | const GrMatrixConvolutionEffect& s = sBase.cast<GrMatrixConvolutionEffect>(); |
207 | return fKernelSize == s.kernelSize() && |
208 | !memcmp(fKernel, s.kernel(), |
209 | fKernelSize.width() * fKernelSize.height() * sizeof(float)) && |
210 | fGain == s.gain() && |
211 | fBias == s.bias() && |
212 | !memcmp(fKernelOffset, s.kernelOffset(), sizeof(fKernelOffset)) && |
213 | fConvolveAlpha == s.convolveAlpha() && |
214 | fDomain == s.domain(); |
215 | } |
216 | |
217 | static void fill_in_1D_gaussian_kernel_with_stride(float* kernel, int size, int stride, |
218 | float twoSigmaSqrd) { |
219 | SkASSERT(!SkScalarNearlyZero(twoSigmaSqrd, SK_ScalarNearlyZero)); |
220 | |
221 | const float sigmaDenom = 1.0f / twoSigmaSqrd; |
222 | const int radius = size / 2; |
223 | |
224 | float sum = 0.0f; |
225 | for (int i = 0; i < size; ++i) { |
226 | float term = static_cast<float>(i - radius); |
227 | // Note that the constant term (1/(sqrt(2*pi*sigma^2)) of the Gaussian |
228 | // is dropped here, since we renormalize the kernel below. |
229 | kernel[i * stride] = sk_float_exp(-term * term * sigmaDenom); |
230 | sum += kernel[i * stride]; |
231 | } |
232 | // Normalize the kernel |
233 | float scale = 1.0f / sum; |
234 | for (int i = 0; i < size; ++i) { |
235 | kernel[i * stride] *= scale; |
236 | } |
237 | } |
238 | |
239 | static void fill_in_2D_gaussian_kernel(float* kernel, int width, int height, |
240 | SkScalar sigmaX, SkScalar sigmaY) { |
241 | SkASSERT(width * height <= MAX_KERNEL_SIZE); |
242 | const float twoSigmaSqrdX = 2.0f * SkScalarToFloat(SkScalarSquare(sigmaX)); |
243 | const float twoSigmaSqrdY = 2.0f * SkScalarToFloat(SkScalarSquare(sigmaY)); |
244 | |
245 | // TODO: in all of these degenerate cases we're uploading (and using) a whole lot of zeros. |
246 | if (SkScalarNearlyZero(twoSigmaSqrdX, SK_ScalarNearlyZero) || |
247 | SkScalarNearlyZero(twoSigmaSqrdY, SK_ScalarNearlyZero)) { |
248 | // In this case the 2D Gaussian degenerates to a 1D Gaussian (in X or Y) or a point |
249 | SkASSERT(3 == width || 3 == height); |
250 | memset(kernel, 0, width*height*sizeof(float)); |
251 | |
252 | if (SkScalarNearlyZero(twoSigmaSqrdX, SK_ScalarNearlyZero) && |
253 | SkScalarNearlyZero(twoSigmaSqrdY, SK_ScalarNearlyZero)) { |
254 | // A point |
255 | SkASSERT(3 == width && 3 == height); |
256 | kernel[4] = 1.0f; |
257 | } else if (SkScalarNearlyZero(twoSigmaSqrdX, SK_ScalarNearlyZero)) { |
258 | // A 1D Gaussian in Y |
259 | SkASSERT(3 == width); |
260 | // Down the middle column of the kernel with a stride of width |
261 | fill_in_1D_gaussian_kernel_with_stride(&kernel[1], height, width, twoSigmaSqrdY); |
262 | } else { |
263 | // A 1D Gaussian in X |
264 | SkASSERT(SkScalarNearlyZero(twoSigmaSqrdY, SK_ScalarNearlyZero)); |
265 | SkASSERT(3 == height); |
266 | // Down the middle row of the kernel with a stride of 1 |
267 | fill_in_1D_gaussian_kernel_with_stride(&kernel[width], width, 1, twoSigmaSqrdX); |
268 | } |
269 | return; |
270 | } |
271 | |
272 | const float sigmaXDenom = 1.0f / twoSigmaSqrdX; |
273 | const float sigmaYDenom = 1.0f / twoSigmaSqrdY; |
274 | const int xRadius = width / 2; |
275 | const int yRadius = height / 2; |
276 | |
277 | float sum = 0.0f; |
278 | for (int x = 0; x < width; x++) { |
279 | float xTerm = static_cast<float>(x - xRadius); |
280 | xTerm = xTerm * xTerm * sigmaXDenom; |
281 | for (int y = 0; y < height; y++) { |
282 | float yTerm = static_cast<float>(y - yRadius); |
283 | float xyTerm = sk_float_exp(-(xTerm + yTerm * yTerm * sigmaYDenom)); |
284 | // Note that the constant term (1/(sqrt(2*pi*sigma^2)) of the Gaussian |
285 | // is dropped here, since we renormalize the kernel below. |
286 | kernel[y * width + x] = xyTerm; |
287 | sum += xyTerm; |
288 | } |
289 | } |
290 | // Normalize the kernel |
291 | float scale = 1.0f / sum; |
292 | for (int i = 0; i < width * height; ++i) { |
293 | kernel[i] *= scale; |
294 | } |
295 | } |
296 | |
297 | // Static function to create a 2D convolution |
298 | std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::MakeGaussian( |
299 | GrSurfaceProxyView srcView, |
300 | const SkIRect& srcBounds, |
301 | const SkISize& kernelSize, |
302 | SkScalar gain, |
303 | SkScalar bias, |
304 | const SkIPoint& kernelOffset, |
305 | GrTextureDomain::Mode tileMode, |
306 | bool convolveAlpha, |
307 | SkScalar sigmaX, |
308 | SkScalar sigmaY) { |
309 | // SkGpuBlurUtils is not as aggressive as it once was about avoiding texture domains. |
310 | // Check for a trivial case here where the domain can be avoided. TODO: Use GrTextureEffect |
311 | // here which includes this and more. |
312 | if (tileMode == GrTextureDomain::kClamp_Mode && !srcView.proxy()->isFullyLazy() && |
313 | srcBounds.contains(SkIRect::MakeSize(srcView.proxy()->backingStoreDimensions()))) { |
314 | tileMode = GrTextureDomain::kIgnore_Mode; |
315 | } |
316 | float kernel[MAX_KERNEL_SIZE]; |
317 | |
318 | fill_in_2D_gaussian_kernel(kernel, kernelSize.width(), kernelSize.height(), sigmaX, sigmaY); |
319 | |
320 | return std::unique_ptr<GrFragmentProcessor>( |
321 | new GrMatrixConvolutionEffect(std::move(srcView), srcBounds, kernelSize, kernel, |
322 | gain, bias, kernelOffset, tileMode, convolveAlpha)); |
323 | } |
324 | |
325 | GR_DEFINE_FRAGMENT_PROCESSOR_TEST(GrMatrixConvolutionEffect); |
326 | |
327 | #if GR_TEST_UTILS |
328 | std::unique_ptr<GrFragmentProcessor> GrMatrixConvolutionEffect::TestCreate(GrProcessorTestData* d) { |
329 | auto [view, ct, at] = d->randomView(); |
330 | |
331 | int width = d->fRandom->nextRangeU(1, MAX_KERNEL_SIZE); |
332 | int height = d->fRandom->nextRangeU(1, MAX_KERNEL_SIZE / width); |
333 | SkISize kernelSize = SkISize::Make(width, height); |
334 | std::unique_ptr<SkScalar[]> kernel(new SkScalar[width * height]); |
335 | for (int i = 0; i < width * height; i++) { |
336 | kernel.get()[i] = d->fRandom->nextSScalar1(); |
337 | } |
338 | SkScalar gain = d->fRandom->nextSScalar1(); |
339 | SkScalar bias = d->fRandom->nextSScalar1(); |
340 | |
341 | uint32_t kernalOffsetX = d->fRandom->nextRangeU(0, kernelSize.width()); |
342 | uint32_t kernalOffsetY = d->fRandom->nextRangeU(0, kernelSize.height()); |
343 | SkIPoint kernelOffset = SkIPoint::Make(kernalOffsetX, kernalOffsetY); |
344 | |
345 | uint32_t boundsX = d->fRandom->nextRangeU(0, view.width()); |
346 | uint32_t boundsY = d->fRandom->nextRangeU(0, view.height()); |
347 | uint32_t boundsW = d->fRandom->nextRangeU(0, view.width()); |
348 | uint32_t boundsH = d->fRandom->nextRangeU(0, view.height()); |
349 | SkIRect bounds = SkIRect::MakeXYWH(boundsX, boundsY, boundsW, boundsH); |
350 | |
351 | GrTextureDomain::Mode tileMode = |
352 | static_cast<GrTextureDomain::Mode>(d->fRandom->nextRangeU(0, 2)); |
353 | bool convolveAlpha = d->fRandom->nextBool(); |
354 | |
355 | return GrMatrixConvolutionEffect::Make(std::move(view), |
356 | bounds, |
357 | kernelSize, |
358 | kernel.get(), |
359 | gain, |
360 | bias, |
361 | kernelOffset, |
362 | tileMode, |
363 | convolveAlpha); |
364 | } |
365 | #endif |
366 | |