1 /* 2 * Copyright 2011 The Android Open Source Project 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 "SkBitmap.h" 9 #include "SkBlurImageFilter.h" 10 #include "SkColorPriv.h" 11 #include "SkReadBuffer.h" 12 #include "SkWriteBuffer.h" 13 #include "SkGpuBlurUtils.h" 14 #include "SkBlurImage_opts.h" 15 #if SK_SUPPORT_GPU 16 #include "GrContext.h" 17 #endif 18 19 // This rather arbitrary-looking value results in a maximum box blur kernel size 20 // of 1000 pixels on the raster path, which matches the WebKit and Firefox 21 // implementations. Since the GPU path does not compute a box blur, putting 22 // the limit on sigma ensures consistent behaviour between the GPU and 23 // raster paths. 24 #define MAX_SIGMA SkIntToScalar(532) 25 26 SkBlurImageFilter::SkBlurImageFilter(SkReadBuffer& buffer) 27 : INHERITED(1, buffer) { 28 fSigma.fWidth = buffer.readScalar(); 29 fSigma.fHeight = buffer.readScalar(); 30 buffer.validate(SkScalarIsFinite(fSigma.fWidth) && 31 SkScalarIsFinite(fSigma.fHeight) && 32 (fSigma.fWidth >= 0) && 33 (fSigma.fHeight >= 0)); 34 } 35 36 SkBlurImageFilter::SkBlurImageFilter(SkScalar sigmaX, 37 SkScalar sigmaY, 38 SkImageFilter* input, 39 const CropRect* cropRect) 40 : INHERITED(input, cropRect), fSigma(SkSize::Make(sigmaX, sigmaY)) { 41 SkASSERT(sigmaX >= 0 && sigmaY >= 0); 42 } 43 44 void SkBlurImageFilter::flatten(SkWriteBuffer& buffer) const { 45 this->INHERITED::flatten(buffer); 46 buffer.writeScalar(fSigma.fWidth); 47 buffer.writeScalar(fSigma.fHeight); 48 } 49 50 enum BlurDirection { 51 kX, kY 52 }; 53 54 /** 55 * 56 * In order to make memory accesses cache-friendly, we reorder the passes to 57 * use contiguous memory reads wherever possible. 58 * 59 * For example, the 6 passes of the X-and-Y blur case are rewritten as 60 * follows. Instead of 3 passes in X and 3 passes in Y, we perform 61 * 2 passes in X, 1 pass in X transposed to Y on write, 2 passes in X, 62 * then 1 pass in X transposed to Y on write. 63 * 64 * +----+ +----+ +----+ +---+ +---+ +---+ +----+ 65 * + AB + ----> | AB | ----> | AB | -----> | A | ----> | A | ----> | A | -----> | AB | 66 * +----+ blurX +----+ blurX +----+ blurXY | B | blurX | B | blurX | B | blurXY +----+ 67 * +---+ +---+ +---+ 68 * 69 * In this way, two of the y-blurs become x-blurs applied to transposed 70 * images, and all memory reads are contiguous. 71 */ 72 73 template<BlurDirection srcDirection, BlurDirection dstDirection> 74 static void boxBlur(const SkPMColor* src, int srcStride, SkPMColor* dst, int kernelSize, 75 int leftOffset, int rightOffset, int width, int height) 76 { 77 int rightBorder = SkMin32(rightOffset + 1, width); 78 int srcStrideX = srcDirection == kX ? 1 : srcStride; 79 int dstStrideX = dstDirection == kX ? 1 : height; 80 int srcStrideY = srcDirection == kX ? srcStride : 1; 81 int dstStrideY = dstDirection == kX ? width : 1; 82 uint32_t scale = (1 << 24) / kernelSize; 83 uint32_t half = 1 << 23; 84 for (int y = 0; y < height; ++y) { 85 int sumA = 0, sumR = 0, sumG = 0, sumB = 0; 86 const SkPMColor* p = src; 87 for (int i = 0; i < rightBorder; ++i) { 88 sumA += SkGetPackedA32(*p); 89 sumR += SkGetPackedR32(*p); 90 sumG += SkGetPackedG32(*p); 91 sumB += SkGetPackedB32(*p); 92 p += srcStrideX; 93 } 94 95 const SkPMColor* sptr = src; 96 SkColor* dptr = dst; 97 for (int x = 0; x < width; ++x) { 98 *dptr = SkPackARGB32((sumA * scale + half) >> 24, 99 (sumR * scale + half) >> 24, 100 (sumG * scale + half) >> 24, 101 (sumB * scale + half) >> 24); 102 if (x >= leftOffset) { 103 SkColor l = *(sptr - leftOffset * srcStrideX); 104 sumA -= SkGetPackedA32(l); 105 sumR -= SkGetPackedR32(l); 106 sumG -= SkGetPackedG32(l); 107 sumB -= SkGetPackedB32(l); 108 } 109 if (x + rightOffset + 1 < width) { 110 SkColor r = *(sptr + (rightOffset + 1) * srcStrideX); 111 sumA += SkGetPackedA32(r); 112 sumR += SkGetPackedR32(r); 113 sumG += SkGetPackedG32(r); 114 sumB += SkGetPackedB32(r); 115 } 116 sptr += srcStrideX; 117 if (srcDirection == kY) { 118 SK_PREFETCH(sptr + (rightOffset + 1) * srcStrideX); 119 } 120 dptr += dstStrideX; 121 } 122 src += srcStrideY; 123 dst += dstStrideY; 124 } 125 } 126 127 static void getBox3Params(SkScalar s, int *kernelSize, int* kernelSize3, int *lowOffset, 128 int *highOffset) 129 { 130 float pi = SkScalarToFloat(SK_ScalarPI); 131 int d = static_cast<int>(floorf(SkScalarToFloat(s) * 3.0f * sqrtf(2.0f * pi) / 4.0f + 0.5f)); 132 *kernelSize = d; 133 if (d % 2 == 1) { 134 *lowOffset = *highOffset = (d - 1) / 2; 135 *kernelSize3 = d; 136 } else { 137 *highOffset = d / 2; 138 *lowOffset = *highOffset - 1; 139 *kernelSize3 = d + 1; 140 } 141 } 142 143 bool SkBlurImageFilter::onFilterImage(Proxy* proxy, 144 const SkBitmap& source, const Context& ctx, 145 SkBitmap* dst, SkIPoint* offset) const { 146 SkBitmap src = source; 147 SkIPoint srcOffset = SkIPoint::Make(0, 0); 148 if (getInput(0) && !getInput(0)->filterImage(proxy, source, ctx, &src, &srcOffset)) { 149 return false; 150 } 151 152 if (src.colorType() != kN32_SkColorType) { 153 return false; 154 } 155 156 SkIRect srcBounds, dstBounds; 157 if (!this->applyCropRect(ctx, proxy, src, &srcOffset, &srcBounds, &src)) { 158 return false; 159 } 160 161 SkAutoLockPixels alp(src); 162 if (!src.getPixels()) { 163 return false; 164 } 165 166 if (!dst->allocPixels(src.info().makeWH(srcBounds.width(), srcBounds.height()))) { 167 return false; 168 } 169 dst->getBounds(&dstBounds); 170 171 SkVector sigma = SkVector::Make(fSigma.width(), fSigma.height()); 172 ctx.ctm().mapVectors(&sigma, 1); 173 sigma.fX = SkMinScalar(sigma.fX, MAX_SIGMA); 174 sigma.fY = SkMinScalar(sigma.fY, MAX_SIGMA); 175 176 int kernelSizeX, kernelSizeX3, lowOffsetX, highOffsetX; 177 int kernelSizeY, kernelSizeY3, lowOffsetY, highOffsetY; 178 getBox3Params(sigma.x(), &kernelSizeX, &kernelSizeX3, &lowOffsetX, &highOffsetX); 179 getBox3Params(sigma.y(), &kernelSizeY, &kernelSizeY3, &lowOffsetY, &highOffsetY); 180 181 if (kernelSizeX < 0 || kernelSizeY < 0) { 182 return false; 183 } 184 185 if (kernelSizeX == 0 && kernelSizeY == 0) { 186 src.copyTo(dst, dst->colorType()); 187 offset->fX = srcBounds.fLeft; 188 offset->fY = srcBounds.fTop; 189 return true; 190 } 191 192 SkBitmap temp; 193 if (!temp.allocPixels(dst->info())) { 194 return false; 195 } 196 197 offset->fX = srcBounds.fLeft; 198 offset->fY = srcBounds.fTop; 199 srcBounds.offset(-srcOffset); 200 const SkPMColor* s = src.getAddr32(srcBounds.left(), srcBounds.top()); 201 SkPMColor* t = temp.getAddr32(0, 0); 202 SkPMColor* d = dst->getAddr32(0, 0); 203 int w = dstBounds.width(), h = dstBounds.height(); 204 int sw = src.rowBytesAsPixels(); 205 SkBoxBlurProc boxBlurX, boxBlurY, boxBlurXY, boxBlurYX; 206 if (!SkBoxBlurGetPlatformProcs(&boxBlurX, &boxBlurY, &boxBlurXY, &boxBlurYX)) { 207 boxBlurX = boxBlur<kX, kX>; 208 boxBlurY = boxBlur<kY, kY>; 209 boxBlurXY = boxBlur<kX, kY>; 210 boxBlurYX = boxBlur<kY, kX>; 211 } 212 213 if (kernelSizeX > 0 && kernelSizeY > 0) { 214 boxBlurX(s, sw, t, kernelSizeX, lowOffsetX, highOffsetX, w, h); 215 boxBlurX(t, w, d, kernelSizeX, highOffsetX, lowOffsetX, w, h); 216 boxBlurXY(d, w, t, kernelSizeX3, highOffsetX, highOffsetX, w, h); 217 boxBlurX(t, h, d, kernelSizeY, lowOffsetY, highOffsetY, h, w); 218 boxBlurX(d, h, t, kernelSizeY, highOffsetY, lowOffsetY, h, w); 219 boxBlurXY(t, h, d, kernelSizeY3, highOffsetY, highOffsetY, h, w); 220 } else if (kernelSizeX > 0) { 221 boxBlurX(s, sw, d, kernelSizeX, lowOffsetX, highOffsetX, w, h); 222 boxBlurX(d, w, t, kernelSizeX, highOffsetX, lowOffsetX, w, h); 223 boxBlurX(t, w, d, kernelSizeX3, highOffsetX, highOffsetX, w, h); 224 } else if (kernelSizeY > 0) { 225 boxBlurYX(s, sw, d, kernelSizeY, lowOffsetY, highOffsetY, h, w); 226 boxBlurX(d, h, t, kernelSizeY, highOffsetY, lowOffsetY, h, w); 227 boxBlurXY(t, h, d, kernelSizeY3, highOffsetY, highOffsetY, h, w); 228 } 229 return true; 230 } 231 232 233 void SkBlurImageFilter::computeFastBounds(const SkRect& src, SkRect* dst) const { 234 if (getInput(0)) { 235 getInput(0)->computeFastBounds(src, dst); 236 } else { 237 *dst = src; 238 } 239 240 dst->outset(SkScalarMul(fSigma.width(), SkIntToScalar(3)), 241 SkScalarMul(fSigma.height(), SkIntToScalar(3))); 242 } 243 244 bool SkBlurImageFilter::onFilterBounds(const SkIRect& src, const SkMatrix& ctm, 245 SkIRect* dst) const { 246 SkIRect bounds = src; 247 if (getInput(0) && !getInput(0)->filterBounds(src, ctm, &bounds)) { 248 return false; 249 } 250 SkVector sigma = SkVector::Make(fSigma.width(), fSigma.height()); 251 ctm.mapVectors(&sigma, 1); 252 bounds.outset(SkScalarCeilToInt(SkScalarMul(sigma.x(), SkIntToScalar(3))), 253 SkScalarCeilToInt(SkScalarMul(sigma.y(), SkIntToScalar(3)))); 254 *dst = bounds; 255 return true; 256 } 257 258 bool SkBlurImageFilter::filterImageGPU(Proxy* proxy, const SkBitmap& src, const Context& ctx, 259 SkBitmap* result, SkIPoint* offset) const { 260 #if SK_SUPPORT_GPU 261 SkBitmap input = src; 262 SkIPoint srcOffset = SkIPoint::Make(0, 0); 263 if (getInput(0) && !getInput(0)->getInputResultGPU(proxy, src, ctx, &input, &srcOffset)) { 264 return false; 265 } 266 SkIRect rect; 267 if (!this->applyCropRect(ctx, proxy, input, &srcOffset, &rect, &input)) { 268 return false; 269 } 270 GrTexture* source = input.getTexture(); 271 SkVector sigma = SkVector::Make(fSigma.width(), fSigma.height()); 272 ctx.ctm().mapVectors(&sigma, 1); 273 sigma.fX = SkMinScalar(sigma.fX, MAX_SIGMA); 274 sigma.fY = SkMinScalar(sigma.fY, MAX_SIGMA); 275 offset->fX = rect.fLeft; 276 offset->fY = rect.fTop; 277 rect.offset(-srcOffset); 278 SkAutoTUnref<GrTexture> tex(SkGpuBlurUtils::GaussianBlur(source->getContext(), 279 source, 280 false, 281 SkRect::Make(rect), 282 true, 283 sigma.x(), 284 sigma.y())); 285 WrapTexture(tex, rect.width(), rect.height(), result); 286 return true; 287 #else 288 SkDEBUGFAIL("Should not call in GPU-less build"); 289 return false; 290 #endif 291 } 292
