1 | /* |
2 | * Copyright 2012 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 | #include "src/gpu/effects/GrGaussianConvolutionFragmentProcessor.h" |
9 | |
10 | #include "src/gpu/GrTexture.h" |
11 | #include "src/gpu/GrTextureProxy.h" |
12 | #include "src/gpu/glsl/GrGLSLFragmentProcessor.h" |
13 | #include "src/gpu/glsl/GrGLSLFragmentShaderBuilder.h" |
14 | #include "src/gpu/glsl/GrGLSLProgramDataManager.h" |
15 | #include "src/gpu/glsl/GrGLSLUniformHandler.h" |
16 | |
17 | // For brevity |
18 | using UniformHandle = GrGLSLProgramDataManager::UniformHandle; |
19 | using Direction = GrGaussianConvolutionFragmentProcessor::Direction; |
20 | |
21 | class GrGLConvolutionEffect : public GrGLSLFragmentProcessor { |
22 | public: |
23 | void emitCode(EmitArgs&) override; |
24 | |
25 | static inline void GenKey(const GrProcessor&, const GrShaderCaps&, GrProcessorKeyBuilder*); |
26 | |
27 | protected: |
28 | void onSetData(const GrGLSLProgramDataManager&, const GrFragmentProcessor&) override; |
29 | |
30 | private: |
31 | UniformHandle fKernelUni; |
32 | UniformHandle fImageIncrementUni; |
33 | UniformHandle fBoundsUni; |
34 | |
35 | typedef GrGLSLFragmentProcessor INHERITED; |
36 | }; |
37 | |
38 | void GrGLConvolutionEffect::emitCode(EmitArgs& args) { |
39 | const GrGaussianConvolutionFragmentProcessor& ce = |
40 | args.fFp.cast<GrGaussianConvolutionFragmentProcessor>(); |
41 | |
42 | GrGLSLUniformHandler* uniformHandler = args.fUniformHandler; |
43 | fImageIncrementUni = uniformHandler->addUniform(&ce, kFragment_GrShaderFlag, kHalf2_GrSLType, |
44 | "ImageIncrement" ); |
45 | if (ce.useBounds()) { |
46 | fBoundsUni = uniformHandler->addUniform(&ce, kFragment_GrShaderFlag, kHalf2_GrSLType, |
47 | "Bounds" ); |
48 | } |
49 | |
50 | int width = ce.width(); |
51 | |
52 | int arrayCount = (width + 3) / 4; |
53 | SkASSERT(4 * arrayCount >= width); |
54 | |
55 | fKernelUni = uniformHandler->addUniformArray(&ce, kFragment_GrShaderFlag, kHalf4_GrSLType, |
56 | "Kernel" , arrayCount); |
57 | |
58 | GrGLSLFPFragmentBuilder* fragBuilder = args.fFragBuilder; |
59 | SkString coords2D = fragBuilder->ensureCoords2D(args.fTransformedCoords[0].fVaryingPoint); |
60 | |
61 | fragBuilder->codeAppendf("%s = half4(0, 0, 0, 0);" , args.fOutputColor); |
62 | |
63 | const GrShaderVar& kernel = uniformHandler->getUniformVariable(fKernelUni); |
64 | const char* imgInc = uniformHandler->getUniformCStr(fImageIncrementUni); |
65 | |
66 | fragBuilder->codeAppendf("float2 coord = %s - %d.0 * %s;" , coords2D.c_str(), ce.radius(), imgInc); |
67 | fragBuilder->codeAppend("float2 coordSampled = half2(0, 0);" ); |
68 | |
69 | // Manually unroll loop because some drivers don't; yields 20-30% speedup. |
70 | const char* kVecSuffix[4] = {".x" , ".y" , ".z" , ".w" }; |
71 | for (int i = 0; i < width; i++) { |
72 | SkString index; |
73 | SkString kernelIndex; |
74 | index.appendS32(i / 4); |
75 | kernel.appendArrayAccess(index.c_str(), &kernelIndex); |
76 | kernelIndex.append(kVecSuffix[i & 0x3]); |
77 | |
78 | fragBuilder->codeAppend("coordSampled = coord;" ); |
79 | if (ce.useBounds()) { |
80 | // We used to compute a bool indicating whether we're in bounds or not, cast it to a |
81 | // float, and then mul weight*texture_sample by the float. However, the Adreno 430 seems |
82 | // to have a bug that caused corruption. |
83 | const char* bounds = uniformHandler->getUniformCStr(fBoundsUni); |
84 | const char* component = ce.direction() == Direction::kY ? "y" : "x" ; |
85 | |
86 | switch (ce.mode()) { |
87 | case GrTextureDomain::kClamp_Mode: { |
88 | fragBuilder->codeAppendf("coordSampled.%s = clamp(coord.%s, %s.x, %s.y);\n" , |
89 | component, component, bounds, bounds); |
90 | break; |
91 | } |
92 | // Deferring implementing kMirrorRepeat until we use DomainEffects as |
93 | // child processors. Fallback to Repeat. |
94 | case GrTextureDomain::kMirrorRepeat_Mode: |
95 | case GrTextureDomain::kRepeat_Mode: { |
96 | fragBuilder->codeAppendf("coordSampled.%s = " |
97 | "mod(coord.%s - %s.x, %s.y - %s.x) + %s.x;\n" , |
98 | component, component, bounds, bounds, bounds, bounds); |
99 | break; |
100 | } |
101 | case GrTextureDomain::kDecal_Mode: { |
102 | fragBuilder->codeAppendf("if (coord.%s >= %s.x && coord.%s <= %s.y) {" , |
103 | component, bounds, component, bounds); |
104 | break; |
105 | } |
106 | default: { |
107 | SK_ABORT("Unsupported operation." ); |
108 | } |
109 | } |
110 | } |
111 | fragBuilder->codeAppendf("%s += " , args.fOutputColor); |
112 | fragBuilder->appendTextureLookup(args.fTexSamplers[0], "coordSampled" ); |
113 | fragBuilder->codeAppendf(" * %s;\n" , kernelIndex.c_str()); |
114 | if (GrTextureDomain::kDecal_Mode == ce.mode()) { |
115 | fragBuilder->codeAppend("}" ); |
116 | } |
117 | fragBuilder->codeAppendf("coord += %s;\n" , imgInc); |
118 | } |
119 | fragBuilder->codeAppendf("%s *= %s;\n" , args.fOutputColor, args.fInputColor); |
120 | } |
121 | |
122 | void GrGLConvolutionEffect::onSetData(const GrGLSLProgramDataManager& pdman, |
123 | const GrFragmentProcessor& processor) { |
124 | const GrGaussianConvolutionFragmentProcessor& conv = |
125 | processor.cast<GrGaussianConvolutionFragmentProcessor>(); |
126 | const auto& view = conv.textureSampler(0).view(); |
127 | GrSurfaceProxy* proxy = view.proxy(); |
128 | GrTexture& texture = *proxy->peekTexture(); |
129 | |
130 | float imageIncrement[2] = {0}; |
131 | float ySign = view.origin() != kTopLeft_GrSurfaceOrigin ? 1.0f : -1.0f; |
132 | switch (conv.direction()) { |
133 | case Direction::kX: |
134 | imageIncrement[0] = 1.0f / texture.width(); |
135 | break; |
136 | case Direction::kY: |
137 | imageIncrement[1] = ySign / texture.height(); |
138 | break; |
139 | default: |
140 | SK_ABORT("Unknown filter direction." ); |
141 | } |
142 | pdman.set2fv(fImageIncrementUni, 1, imageIncrement); |
143 | if (conv.useBounds()) { |
144 | float bounds[2] = {0}; |
145 | bounds[0] = conv.bounds()[0]; |
146 | bounds[1] = conv.bounds()[1]; |
147 | if (GrTextureDomain::kClamp_Mode == conv.mode()) { |
148 | bounds[0] += SK_ScalarHalf; |
149 | bounds[1] -= SK_ScalarHalf; |
150 | } |
151 | if (Direction::kX == conv.direction()) { |
152 | SkScalar inv = SkScalarInvert(SkIntToScalar(texture.width())); |
153 | bounds[0] *= inv; |
154 | bounds[1] *= inv; |
155 | } else { |
156 | SkScalar inv = SkScalarInvert(SkIntToScalar(texture.height())); |
157 | if (view.origin() != kTopLeft_GrSurfaceOrigin) { |
158 | float tmp = bounds[0]; |
159 | bounds[0] = 1.0f - (inv * bounds[1]); |
160 | bounds[1] = 1.0f - (inv * tmp); |
161 | } else { |
162 | bounds[0] *= inv; |
163 | bounds[1] *= inv; |
164 | } |
165 | } |
166 | |
167 | SkASSERT(bounds[0] <= bounds[1]); |
168 | pdman.set2f(fBoundsUni, bounds[0], bounds[1]); |
169 | } |
170 | int width = conv.width(); |
171 | |
172 | int arrayCount = (width + 3) / 4; |
173 | SkASSERT(4 * arrayCount >= width); |
174 | pdman.set4fv(fKernelUni, arrayCount, conv.kernel()); |
175 | } |
176 | |
177 | void GrGLConvolutionEffect::GenKey(const GrProcessor& processor, const GrShaderCaps&, |
178 | GrProcessorKeyBuilder* b) { |
179 | const GrGaussianConvolutionFragmentProcessor& conv = |
180 | processor.cast<GrGaussianConvolutionFragmentProcessor>(); |
181 | uint32_t key = conv.radius(); |
182 | key <<= 3; |
183 | if (conv.useBounds()) { |
184 | key |= Direction::kY == conv.direction() ? 0x4 : 0x0; |
185 | } |
186 | key |= static_cast<uint32_t>(conv.mode()); |
187 | b->add32(key); |
188 | } |
189 | |
190 | /////////////////////////////////////////////////////////////////////////////// |
191 | static void fill_in_1D_gaussian_kernel(float* kernel, int width, float gaussianSigma, int radius) { |
192 | const float twoSigmaSqrd = 2.0f * gaussianSigma * gaussianSigma; |
193 | if (SkScalarNearlyZero(twoSigmaSqrd, SK_ScalarNearlyZero)) { |
194 | for (int i = 0; i < width; ++i) { |
195 | kernel[i] = 0.0f; |
196 | } |
197 | return; |
198 | } |
199 | |
200 | const float denom = 1.0f / twoSigmaSqrd; |
201 | |
202 | float sum = 0.0f; |
203 | for (int i = 0; i < width; ++i) { |
204 | float x = static_cast<float>(i - radius); |
205 | // Note that the constant term (1/(sqrt(2*pi*sigma^2)) of the Gaussian |
206 | // is dropped here, since we renormalize the kernel below. |
207 | kernel[i] = sk_float_exp(-x * x * denom); |
208 | sum += kernel[i]; |
209 | } |
210 | // Normalize the kernel |
211 | float scale = 1.0f / sum; |
212 | for (int i = 0; i < width; ++i) { |
213 | kernel[i] *= scale; |
214 | } |
215 | } |
216 | |
217 | GrGaussianConvolutionFragmentProcessor::GrGaussianConvolutionFragmentProcessor( |
218 | GrSurfaceProxyView view, |
219 | SkAlphaType alphaType, |
220 | Direction direction, |
221 | int radius, |
222 | float gaussianSigma, |
223 | GrTextureDomain::Mode mode, |
224 | int bounds[2]) |
225 | : INHERITED(kGrGaussianConvolutionFragmentProcessor_ClassID, |
226 | ModulateForSamplerOptFlags(alphaType, mode == GrTextureDomain::kDecal_Mode)) |
227 | , fCoordTransform(view.proxy(), view.origin()) |
228 | , fTextureSampler(std::move(view)) |
229 | , fRadius(radius) |
230 | , fDirection(direction) |
231 | , fMode(mode) { |
232 | this->addCoordTransform(&fCoordTransform); |
233 | this->setTextureSamplerCnt(1); |
234 | SkASSERT(radius <= kMaxKernelRadius); |
235 | |
236 | fill_in_1D_gaussian_kernel(fKernel, this->width(), gaussianSigma, this->radius()); |
237 | // SkGpuBlurUtils is not as aggressive as it once was about avoiding domains. So we check |
238 | // here if we can omit the domain. TODO: remove this when this effect uses a child to |
239 | // sample the texture. |
240 | auto samplerProxy = fTextureSampler.proxy(); |
241 | if (!samplerProxy->isFullyLazy()) { |
242 | int wh = (fDirection == Direction::kX) ? samplerProxy->backingStoreDimensions().width() |
243 | : samplerProxy->backingStoreDimensions().height(); |
244 | if (bounds[0] == 0 && bounds[1] == wh) { |
245 | bool useSampler = false; |
246 | GrSamplerState::WrapMode samplerMode = GrSamplerState::WrapMode::kClamp; |
247 | switch (fMode) { |
248 | case GrTextureDomain::kClamp_Mode: |
249 | case GrTextureDomain::kIgnore_Mode: |
250 | useSampler = true; |
251 | break; |
252 | case GrTextureDomain::kRepeat_Mode: |
253 | useSampler = true; |
254 | samplerMode = GrSamplerState::WrapMode::kRepeat; |
255 | break; |
256 | case GrTextureDomain::kMirrorRepeat_Mode: |
257 | useSampler = true; |
258 | samplerMode = GrSamplerState::WrapMode::kMirrorRepeat; |
259 | break; |
260 | case GrTextureDomain::kDecal_Mode: |
261 | // Not sure if we support this in HW without having GrCaps here. |
262 | // Just wait until we replace this with GrTextureEffect. |
263 | break; |
264 | } |
265 | if (useSampler) { |
266 | fMode = GrTextureDomain::kIgnore_Mode; |
267 | if (fDirection == Direction::kX) { |
268 | fTextureSampler.samplerState().setWrapModeX(samplerMode); |
269 | } else { |
270 | fTextureSampler.samplerState().setWrapModeY(samplerMode); |
271 | } |
272 | } |
273 | } |
274 | } |
275 | memcpy(fBounds, bounds, sizeof(fBounds)); |
276 | } |
277 | |
278 | GrGaussianConvolutionFragmentProcessor::GrGaussianConvolutionFragmentProcessor( |
279 | const GrGaussianConvolutionFragmentProcessor& that) |
280 | : INHERITED(kGrGaussianConvolutionFragmentProcessor_ClassID, that.optimizationFlags()) |
281 | , fCoordTransform(that.fCoordTransform) |
282 | , fTextureSampler(that.fTextureSampler) |
283 | , fRadius(that.fRadius) |
284 | , fDirection(that.fDirection) |
285 | , fMode(that.fMode) { |
286 | this->addCoordTransform(&fCoordTransform); |
287 | this->setTextureSamplerCnt(1); |
288 | memcpy(fKernel, that.fKernel, that.width() * sizeof(float)); |
289 | memcpy(fBounds, that.fBounds, sizeof(fBounds)); |
290 | } |
291 | |
292 | void GrGaussianConvolutionFragmentProcessor::onGetGLSLProcessorKey(const GrShaderCaps& caps, |
293 | GrProcessorKeyBuilder* b) const { |
294 | GrGLConvolutionEffect::GenKey(*this, caps, b); |
295 | } |
296 | |
297 | GrGLSLFragmentProcessor* GrGaussianConvolutionFragmentProcessor::onCreateGLSLInstance() const { |
298 | return new GrGLConvolutionEffect; |
299 | } |
300 | |
301 | bool GrGaussianConvolutionFragmentProcessor::onIsEqual(const GrFragmentProcessor& sBase) const { |
302 | const GrGaussianConvolutionFragmentProcessor& s = |
303 | sBase.cast<GrGaussianConvolutionFragmentProcessor>(); |
304 | return (this->radius() == s.radius() && this->direction() == s.direction() && |
305 | this->mode() == s.mode() && |
306 | 0 == memcmp(fBounds, s.fBounds, sizeof(fBounds)) && |
307 | 0 == memcmp(fKernel, s.fKernel, this->width() * sizeof(float))); |
308 | } |
309 | |
310 | /////////////////////////////////////////////////////////////////////////////// |
311 | |
312 | GR_DEFINE_FRAGMENT_PROCESSOR_TEST(GrGaussianConvolutionFragmentProcessor); |
313 | |
314 | #if GR_TEST_UTILS |
315 | std::unique_ptr<GrFragmentProcessor> GrGaussianConvolutionFragmentProcessor::TestCreate( |
316 | GrProcessorTestData* d) { |
317 | auto [view, ct, at] = d->randomView(); |
318 | |
319 | int bounds[2]; |
320 | int modeIdx = d->fRandom->nextRangeU(0, GrTextureDomain::kModeCount-1); |
321 | |
322 | Direction dir; |
323 | if (d->fRandom->nextBool()) { |
324 | dir = Direction::kX; |
325 | bounds[0] = d->fRandom->nextRangeU(0, view.width()-2); |
326 | bounds[1] = d->fRandom->nextRangeU(bounds[0]+1, view.width()-1); |
327 | } else { |
328 | dir = Direction::kY; |
329 | bounds[0] = d->fRandom->nextRangeU(0, view.height()-2); |
330 | bounds[1] = d->fRandom->nextRangeU(bounds[0]+1, view.height()-1); |
331 | } |
332 | |
333 | int radius = d->fRandom->nextRangeU(1, kMaxKernelRadius); |
334 | float sigma = radius / 3.f; |
335 | |
336 | return GrGaussianConvolutionFragmentProcessor::Make(std::move(view), at, dir, radius, sigma, |
337 | static_cast<GrTextureDomain::Mode>(modeIdx), |
338 | bounds); |
339 | } |
340 | #endif |
341 | |