gfx/skia/trunk/src/effects/SkBlurMask.cpp@6474c204b198
gfx/skia/trunk/src/effects/SkBlurMask.cpp
Wed, 31 Dec 2014 06:09:35 +0100
- author
- Michael Schloh von Bennewitz <michael@schloh.com>
- date
- Wed, 31 Dec 2014 06:09:35 +0100
- changeset 0
- 6474c204b198
- permissions
- -rw-r--r--
Cloned upstream origin tor-browser at tor-browser-31.3.0esr-4.5-1-build1
revision ID fc1c9ff7c1b2defdbc039f12214767608f46423f for hacking purpose.
2 /*
3 * Copyright 2006 The Android Open Source Project
4 *
5 * Use of this source code is governed by a BSD-style license that can be
6 * found in the LICENSE file.
7 */
10 #include "SkBlurMask.h"
11 #include "SkMath.h"
12 #include "SkTemplates.h"
13 #include "SkEndian.h"
16 SkScalar SkBlurMask::ConvertRadiusToSigma(SkScalar radius) {
17 // This constant approximates the scaling done in the software path's
18 // "high quality" mode, in SkBlurMask::Blur() (1 / sqrt(3)).
19 // IMHO, it actually should be 1: we blur "less" than we should do
20 // according to the CSS and canvas specs, simply because Safari does the same.
21 // Firefox used to do the same too, until 4.0 where they fixed it. So at some
22 // point we should probably get rid of these scaling constants and rebaseline
23 // all the blur tests.
24 static const SkScalar kBLUR_SIGMA_SCALE = 0.57735f;
26 return radius ? kBLUR_SIGMA_SCALE * radius + 0.5f : 0.0f;
27 }
29 #define UNROLL_SEPARABLE_LOOPS
31 /**
32 * This function performs a box blur in X, of the given radius. If the
33 * "transpose" parameter is true, it will transpose the pixels on write,
34 * such that X and Y are swapped. Reads are always performed from contiguous
35 * memory in X, for speed. The destination buffer (dst) must be at least
36 * (width + leftRadius + rightRadius) * height bytes in size.
37 *
38 * This is what the inner loop looks like before unrolling, and with the two
39 * cases broken out separately (width < diameter, width >= diameter):
40 *
41 * if (width < diameter) {
42 * for (int x = 0; x < width; ++x) {
43 * sum += *right++;
44 * *dptr = (sum * scale + half) >> 24;
45 * dptr += dst_x_stride;
46 * }
47 * for (int x = width; x < diameter; ++x) {
48 * *dptr = (sum * scale + half) >> 24;
49 * dptr += dst_x_stride;
50 * }
51 * for (int x = 0; x < width; ++x) {
52 * *dptr = (sum * scale + half) >> 24;
53 * sum -= *left++;
54 * dptr += dst_x_stride;
55 * }
56 * } else {
57 * for (int x = 0; x < diameter; ++x) {
58 * sum += *right++;
59 * *dptr = (sum * scale + half) >> 24;
60 * dptr += dst_x_stride;
61 * }
62 * for (int x = diameter; x < width; ++x) {
63 * sum += *right++;
64 * *dptr = (sum * scale + half) >> 24;
65 * sum -= *left++;
66 * dptr += dst_x_stride;
67 * }
68 * for (int x = 0; x < diameter; ++x) {
69 * *dptr = (sum * scale + half) >> 24;
70 * sum -= *left++;
71 * dptr += dst_x_stride;
72 * }
73 * }
74 */
75 static int boxBlur(const uint8_t* src, int src_y_stride, uint8_t* dst,
76 int leftRadius, int rightRadius, int width, int height,
77 bool transpose)
78 {
79 int diameter = leftRadius + rightRadius;
80 int kernelSize = diameter + 1;
81 int border = SkMin32(width, diameter);
82 uint32_t scale = (1 << 24) / kernelSize;
83 int new_width = width + SkMax32(leftRadius, rightRadius) * 2;
84 int dst_x_stride = transpose ? height : 1;
85 int dst_y_stride = transpose ? 1 : new_width;
86 uint32_t half = 1 << 23;
87 for (int y = 0; y < height; ++y) {
88 uint32_t sum = 0;
89 uint8_t* dptr = dst + y * dst_y_stride;
90 const uint8_t* right = src + y * src_y_stride;
91 const uint8_t* left = right;
92 for (int x = 0; x < rightRadius - leftRadius; x++) {
93 *dptr = 0;
94 dptr += dst_x_stride;
95 }
96 #define LEFT_BORDER_ITER \
97 sum += *right++; \
98 *dptr = (sum * scale + half) >> 24; \
99 dptr += dst_x_stride;
101 int x = 0;
102 #ifdef UNROLL_SEPARABLE_LOOPS
103 for (; x < border - 16; x += 16) {
104 LEFT_BORDER_ITER
105 LEFT_BORDER_ITER
106 LEFT_BORDER_ITER
107 LEFT_BORDER_ITER
108 LEFT_BORDER_ITER
109 LEFT_BORDER_ITER
110 LEFT_BORDER_ITER
111 LEFT_BORDER_ITER
112 LEFT_BORDER_ITER
113 LEFT_BORDER_ITER
114 LEFT_BORDER_ITER
115 LEFT_BORDER_ITER
116 LEFT_BORDER_ITER
117 LEFT_BORDER_ITER
118 LEFT_BORDER_ITER
119 LEFT_BORDER_ITER
120 }
121 #endif
122 for (; x < border; ++x) {
123 LEFT_BORDER_ITER
124 }
125 #undef LEFT_BORDER_ITER
126 #define TRIVIAL_ITER \
127 *dptr = (sum * scale + half) >> 24; \
128 dptr += dst_x_stride;
129 x = width;
130 #ifdef UNROLL_SEPARABLE_LOOPS
131 for (; x < diameter - 16; x += 16) {
132 TRIVIAL_ITER
133 TRIVIAL_ITER
134 TRIVIAL_ITER
135 TRIVIAL_ITER
136 TRIVIAL_ITER
137 TRIVIAL_ITER
138 TRIVIAL_ITER
139 TRIVIAL_ITER
140 TRIVIAL_ITER
141 TRIVIAL_ITER
142 TRIVIAL_ITER
143 TRIVIAL_ITER
144 TRIVIAL_ITER
145 TRIVIAL_ITER
146 TRIVIAL_ITER
147 TRIVIAL_ITER
148 }
149 #endif
150 for (; x < diameter; ++x) {
151 TRIVIAL_ITER
152 }
153 #undef TRIVIAL_ITER
154 #define CENTER_ITER \
155 sum += *right++; \
156 *dptr = (sum * scale + half) >> 24; \
157 sum -= *left++; \
158 dptr += dst_x_stride;
160 x = diameter;
161 #ifdef UNROLL_SEPARABLE_LOOPS
162 for (; x < width - 16; x += 16) {
163 CENTER_ITER
164 CENTER_ITER
165 CENTER_ITER
166 CENTER_ITER
167 CENTER_ITER
168 CENTER_ITER
169 CENTER_ITER
170 CENTER_ITER
171 CENTER_ITER
172 CENTER_ITER
173 CENTER_ITER
174 CENTER_ITER
175 CENTER_ITER
176 CENTER_ITER
177 CENTER_ITER
178 CENTER_ITER
179 }
180 #endif
181 for (; x < width; ++x) {
182 CENTER_ITER
183 }
184 #undef CENTER_ITER
185 #define RIGHT_BORDER_ITER \
186 *dptr = (sum * scale + half) >> 24; \
187 sum -= *left++; \
188 dptr += dst_x_stride;
190 x = 0;
191 #ifdef UNROLL_SEPARABLE_LOOPS
192 for (; x < border - 16; x += 16) {
193 RIGHT_BORDER_ITER
194 RIGHT_BORDER_ITER
195 RIGHT_BORDER_ITER
196 RIGHT_BORDER_ITER
197 RIGHT_BORDER_ITER
198 RIGHT_BORDER_ITER
199 RIGHT_BORDER_ITER
200 RIGHT_BORDER_ITER
201 RIGHT_BORDER_ITER
202 RIGHT_BORDER_ITER
203 RIGHT_BORDER_ITER
204 RIGHT_BORDER_ITER
205 RIGHT_BORDER_ITER
206 RIGHT_BORDER_ITER
207 RIGHT_BORDER_ITER
208 RIGHT_BORDER_ITER
209 }
210 #endif
211 for (; x < border; ++x) {
212 RIGHT_BORDER_ITER
213 }
214 #undef RIGHT_BORDER_ITER
215 for (int x = 0; x < leftRadius - rightRadius; ++x) {
216 *dptr = 0;
217 dptr += dst_x_stride;
218 }
219 SkASSERT(sum == 0);
220 }
221 return new_width;
222 }
224 /**
225 * This variant of the box blur handles blurring of non-integer radii. It
226 * keeps two running sums: an outer sum for the rounded-up kernel radius, and
227 * an inner sum for the rounded-down kernel radius. For each pixel, it linearly
228 * interpolates between them. In float this would be:
229 * outer_weight * outer_sum / kernelSize +
230 * (1.0 - outer_weight) * innerSum / (kernelSize - 2)
231 *
232 * This is what the inner loop looks like before unrolling, and with the two
233 * cases broken out separately (width < diameter, width >= diameter):
234 *
235 * if (width < diameter) {
236 * for (int x = 0; x < width; x++) {
237 * inner_sum = outer_sum;
238 * outer_sum += *right++;
239 * *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
240 * dptr += dst_x_stride;
241 * }
242 * for (int x = width; x < diameter; ++x) {
243 * *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
244 * dptr += dst_x_stride;
245 * }
246 * for (int x = 0; x < width; x++) {
247 * inner_sum = outer_sum - *left++;
248 * *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
249 * dptr += dst_x_stride;
250 * outer_sum = inner_sum;
251 * }
252 * } else {
253 * for (int x = 0; x < diameter; x++) {
254 * inner_sum = outer_sum;
255 * outer_sum += *right++;
256 * *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
257 * dptr += dst_x_stride;
258 * }
259 * for (int x = diameter; x < width; ++x) {
260 * inner_sum = outer_sum - *left;
261 * outer_sum += *right++;
262 * *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
263 * dptr += dst_x_stride;
264 * outer_sum -= *left++;
265 * }
266 * for (int x = 0; x < diameter; x++) {
267 * inner_sum = outer_sum - *left++;
268 * *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
269 * dptr += dst_x_stride;
270 * outer_sum = inner_sum;
271 * }
272 * }
273 * }
274 * return new_width;
275 */
277 static int boxBlurInterp(const uint8_t* src, int src_y_stride, uint8_t* dst,
278 int radius, int width, int height,
279 bool transpose, uint8_t outer_weight)
280 {
281 int diameter = radius * 2;
282 int kernelSize = diameter + 1;
283 int border = SkMin32(width, diameter);
284 int inner_weight = 255 - outer_weight;
285 outer_weight += outer_weight >> 7;
286 inner_weight += inner_weight >> 7;
287 uint32_t outer_scale = (outer_weight << 16) / kernelSize;
288 uint32_t inner_scale = (inner_weight << 16) / (kernelSize - 2);
289 uint32_t half = 1 << 23;
290 int new_width = width + diameter;
291 int dst_x_stride = transpose ? height : 1;
292 int dst_y_stride = transpose ? 1 : new_width;
293 for (int y = 0; y < height; ++y) {
294 uint32_t outer_sum = 0, inner_sum = 0;
295 uint8_t* dptr = dst + y * dst_y_stride;
296 const uint8_t* right = src + y * src_y_stride;
297 const uint8_t* left = right;
298 int x = 0;
300 #define LEFT_BORDER_ITER \
301 inner_sum = outer_sum; \
302 outer_sum += *right++; \
303 *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24; \
304 dptr += dst_x_stride;
306 #ifdef UNROLL_SEPARABLE_LOOPS
307 for (;x < border - 16; x += 16) {
308 LEFT_BORDER_ITER
309 LEFT_BORDER_ITER
310 LEFT_BORDER_ITER
311 LEFT_BORDER_ITER
312 LEFT_BORDER_ITER
313 LEFT_BORDER_ITER
314 LEFT_BORDER_ITER
315 LEFT_BORDER_ITER
316 LEFT_BORDER_ITER
317 LEFT_BORDER_ITER
318 LEFT_BORDER_ITER
319 LEFT_BORDER_ITER
320 LEFT_BORDER_ITER
321 LEFT_BORDER_ITER
322 LEFT_BORDER_ITER
323 LEFT_BORDER_ITER
324 }
325 #endif
327 for (;x < border; ++x) {
328 LEFT_BORDER_ITER
329 }
330 #undef LEFT_BORDER_ITER
331 for (int x = width; x < diameter; ++x) {
332 *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
333 dptr += dst_x_stride;
334 }
335 x = diameter;
337 #define CENTER_ITER \
338 inner_sum = outer_sum - *left; \
339 outer_sum += *right++; \
340 *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24; \
341 dptr += dst_x_stride; \
342 outer_sum -= *left++;
344 #ifdef UNROLL_SEPARABLE_LOOPS
345 for (; x < width - 16; x += 16) {
346 CENTER_ITER
347 CENTER_ITER
348 CENTER_ITER
349 CENTER_ITER
350 CENTER_ITER
351 CENTER_ITER
352 CENTER_ITER
353 CENTER_ITER
354 CENTER_ITER
355 CENTER_ITER
356 CENTER_ITER
357 CENTER_ITER
358 CENTER_ITER
359 CENTER_ITER
360 CENTER_ITER
361 CENTER_ITER
362 }
363 #endif
364 for (; x < width; ++x) {
365 CENTER_ITER
366 }
367 #undef CENTER_ITER
369 #define RIGHT_BORDER_ITER \
370 inner_sum = outer_sum - *left++; \
371 *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24; \
372 dptr += dst_x_stride; \
373 outer_sum = inner_sum;
375 x = 0;
376 #ifdef UNROLL_SEPARABLE_LOOPS
377 for (; x < border - 16; x += 16) {
378 RIGHT_BORDER_ITER
379 RIGHT_BORDER_ITER
380 RIGHT_BORDER_ITER
381 RIGHT_BORDER_ITER
382 RIGHT_BORDER_ITER
383 RIGHT_BORDER_ITER
384 RIGHT_BORDER_ITER
385 RIGHT_BORDER_ITER
386 RIGHT_BORDER_ITER
387 RIGHT_BORDER_ITER
388 RIGHT_BORDER_ITER
389 RIGHT_BORDER_ITER
390 RIGHT_BORDER_ITER
391 RIGHT_BORDER_ITER
392 RIGHT_BORDER_ITER
393 RIGHT_BORDER_ITER
394 }
395 #endif
396 for (; x < border; ++x) {
397 RIGHT_BORDER_ITER
398 }
399 #undef RIGHT_BORDER_ITER
400 SkASSERT(outer_sum == 0 && inner_sum == 0);
401 }
402 return new_width;
403 }
405 static void get_adjusted_radii(SkScalar passRadius, int *loRadius, int *hiRadius)
406 {
407 *loRadius = *hiRadius = SkScalarCeilToInt(passRadius);
408 if (SkIntToScalar(*hiRadius) - passRadius > 0.5f) {
409 *loRadius = *hiRadius - 1;
410 }
411 }
413 #include "SkColorPriv.h"
415 static void merge_src_with_blur(uint8_t dst[], int dstRB,
416 const uint8_t src[], int srcRB,
417 const uint8_t blur[], int blurRB,
418 int sw, int sh) {
419 dstRB -= sw;
420 srcRB -= sw;
421 blurRB -= sw;
422 while (--sh >= 0) {
423 for (int x = sw - 1; x >= 0; --x) {
424 *dst = SkToU8(SkAlphaMul(*blur, SkAlpha255To256(*src)));
425 dst += 1;
426 src += 1;
427 blur += 1;
428 }
429 dst += dstRB;
430 src += srcRB;
431 blur += blurRB;
432 }
433 }
435 static void clamp_with_orig(uint8_t dst[], int dstRowBytes,
436 const uint8_t src[], int srcRowBytes,
437 int sw, int sh,
438 SkBlurMask::Style style) {
439 int x;
440 while (--sh >= 0) {
441 switch (style) {
442 case SkBlurMask::kSolid_Style:
443 for (x = sw - 1; x >= 0; --x) {
444 int s = *src;
445 int d = *dst;
446 *dst = SkToU8(s + d - SkMulDiv255Round(s, d));
447 dst += 1;
448 src += 1;
449 }
450 break;
451 case SkBlurMask::kOuter_Style:
452 for (x = sw - 1; x >= 0; --x) {
453 if (*src) {
454 *dst = SkToU8(SkAlphaMul(*dst, SkAlpha255To256(255 - *src)));
455 }
456 dst += 1;
457 src += 1;
458 }
459 break;
460 default:
461 SkDEBUGFAIL("Unexpected blur style here");
462 break;
463 }
464 dst += dstRowBytes - sw;
465 src += srcRowBytes - sw;
466 }
467 }
469 ///////////////////////////////////////////////////////////////////////////////
471 // we use a local function to wrap the class static method to work around
472 // a bug in gcc98
473 void SkMask_FreeImage(uint8_t* image);
474 void SkMask_FreeImage(uint8_t* image) {
475 SkMask::FreeImage(image);
476 }
478 bool SkBlurMask::BoxBlur(SkMask* dst, const SkMask& src,
479 SkScalar sigma, Style style, Quality quality,
480 SkIPoint* margin) {
482 if (src.fFormat != SkMask::kA8_Format) {
483 return false;
484 }
486 // Force high quality off for small radii (performance)
487 if (sigma <= SkIntToScalar(2)) {
488 quality = kLow_Quality;
489 }
491 SkScalar passRadius;
492 if (kHigh_Quality == quality) {
493 // For the high quality path the 3 pass box blur kernel width is
494 // 6*rad+1 while the full Gaussian width is 6*sigma.
495 passRadius = sigma - (1/6.0f);
496 } else {
497 // For the low quality path we only attempt to cover 3*sigma of the
498 // Gaussian blur area (1.5*sigma on each side). The single pass box
499 // blur's kernel size is 2*rad+1.
500 passRadius = 1.5f*sigma - 0.5f;
501 }
503 // highQuality: use three box blur passes as a cheap way
504 // to approximate a Gaussian blur
505 int passCount = (kHigh_Quality == quality) ? 3 : 1;
507 int rx = SkScalarCeilToInt(passRadius);
508 int outerWeight = 255 - SkScalarRoundToInt((SkIntToScalar(rx) - passRadius) * 255);
510 SkASSERT(rx >= 0);
511 SkASSERT((unsigned)outerWeight <= 255);
512 if (rx <= 0) {
513 return false;
514 }
516 int ry = rx; // only do square blur for now
518 int padx = passCount * rx;
519 int pady = passCount * ry;
521 if (margin) {
522 margin->set(padx, pady);
523 }
524 dst->fBounds.set(src.fBounds.fLeft - padx, src.fBounds.fTop - pady,
525 src.fBounds.fRight + padx, src.fBounds.fBottom + pady);
527 dst->fRowBytes = dst->fBounds.width();
528 dst->fFormat = SkMask::kA8_Format;
529 dst->fImage = NULL;
531 if (src.fImage) {
532 size_t dstSize = dst->computeImageSize();
533 if (0 == dstSize) {
534 return false; // too big to allocate, abort
535 }
537 int sw = src.fBounds.width();
538 int sh = src.fBounds.height();
539 const uint8_t* sp = src.fImage;
540 uint8_t* dp = SkMask::AllocImage(dstSize);
541 SkAutoTCallVProc<uint8_t, SkMask_FreeImage> autoCall(dp);
543 // build the blurry destination
544 SkAutoTMalloc<uint8_t> tmpBuffer(dstSize);
545 uint8_t* tp = tmpBuffer.get();
546 int w = sw, h = sh;
548 if (outerWeight == 255) {
549 int loRadius, hiRadius;
550 get_adjusted_radii(passRadius, &loRadius, &hiRadius);
551 if (kHigh_Quality == quality) {
552 // Do three X blurs, with a transpose on the final one.
553 w = boxBlur(sp, src.fRowBytes, tp, loRadius, hiRadius, w, h, false);
554 w = boxBlur(tp, w, dp, hiRadius, loRadius, w, h, false);
555 w = boxBlur(dp, w, tp, hiRadius, hiRadius, w, h, true);
556 // Do three Y blurs, with a transpose on the final one.
557 h = boxBlur(tp, h, dp, loRadius, hiRadius, h, w, false);
558 h = boxBlur(dp, h, tp, hiRadius, loRadius, h, w, false);
559 h = boxBlur(tp, h, dp, hiRadius, hiRadius, h, w, true);
560 } else {
561 w = boxBlur(sp, src.fRowBytes, tp, rx, rx, w, h, true);
562 h = boxBlur(tp, h, dp, ry, ry, h, w, true);
563 }
564 } else {
565 if (kHigh_Quality == quality) {
566 // Do three X blurs, with a transpose on the final one.
567 w = boxBlurInterp(sp, src.fRowBytes, tp, rx, w, h, false, outerWeight);
568 w = boxBlurInterp(tp, w, dp, rx, w, h, false, outerWeight);
569 w = boxBlurInterp(dp, w, tp, rx, w, h, true, outerWeight);
570 // Do three Y blurs, with a transpose on the final one.
571 h = boxBlurInterp(tp, h, dp, ry, h, w, false, outerWeight);
572 h = boxBlurInterp(dp, h, tp, ry, h, w, false, outerWeight);
573 h = boxBlurInterp(tp, h, dp, ry, h, w, true, outerWeight);
574 } else {
575 w = boxBlurInterp(sp, src.fRowBytes, tp, rx, w, h, true, outerWeight);
576 h = boxBlurInterp(tp, h, dp, ry, h, w, true, outerWeight);
577 }
578 }
580 dst->fImage = dp;
581 // if need be, alloc the "real" dst (same size as src) and copy/merge
582 // the blur into it (applying the src)
583 if (style == kInner_Style) {
584 // now we allocate the "real" dst, mirror the size of src
585 size_t srcSize = src.computeImageSize();
586 if (0 == srcSize) {
587 return false; // too big to allocate, abort
588 }
589 dst->fImage = SkMask::AllocImage(srcSize);
590 merge_src_with_blur(dst->fImage, src.fRowBytes,
591 sp, src.fRowBytes,
592 dp + passCount * (rx + ry * dst->fRowBytes),
593 dst->fRowBytes, sw, sh);
594 SkMask::FreeImage(dp);
595 } else if (style != kNormal_Style) {
596 clamp_with_orig(dp + passCount * (rx + ry * dst->fRowBytes),
597 dst->fRowBytes, sp, src.fRowBytes, sw, sh, style);
598 }
599 (void)autoCall.detach();
600 }
602 if (style == kInner_Style) {
603 dst->fBounds = src.fBounds; // restore trimmed bounds
604 dst->fRowBytes = src.fRowBytes;
605 }
607 return true;
608 }
610 /* Convolving a box with itself three times results in a piecewise
611 quadratic function:
613 0 x <= -1.5
614 9/8 + 3/2 x + 1/2 x^2 -1.5 < x <= -.5
615 3/4 - x^2 -.5 < x <= .5
616 9/8 - 3/2 x + 1/2 x^2 0.5 < x <= 1.5
617 0 1.5 < x
619 Mathematica:
621 g[x_] := Piecewise [ {
622 {9/8 + 3/2 x + 1/2 x^2 , -1.5 < x <= -.5},
623 {3/4 - x^2 , -.5 < x <= .5},
624 {9/8 - 3/2 x + 1/2 x^2 , 0.5 < x <= 1.5}
625 }, 0]
627 To get the profile curve of the blurred step function at the rectangle
628 edge, we evaluate the indefinite integral, which is piecewise cubic:
630 0 x <= -1.5
631 9/16 + 9/8 x + 3/4 x^2 + 1/6 x^3 -1.5 < x <= -0.5
632 1/2 + 3/4 x - 1/3 x^3 -.5 < x <= .5
633 7/16 + 9/8 x - 3/4 x^2 + 1/6 x^3 .5 < x <= 1.5
634 1 1.5 < x
636 in Mathematica code:
638 gi[x_] := Piecewise[ {
639 { 0 , x <= -1.5 },
640 { 9/16 + 9/8 x + 3/4 x^2 + 1/6 x^3, -1.5 < x <= -0.5 },
641 { 1/2 + 3/4 x - 1/3 x^3 , -.5 < x <= .5},
642 { 7/16 + 9/8 x - 3/4 x^2 + 1/6 x^3, .5 < x <= 1.5}
643 },1]
644 */
646 static float gaussianIntegral(float x) {
647 if (x > 1.5f) {
648 return 0.0f;
649 }
650 if (x < -1.5f) {
651 return 1.0f;
652 }
654 float x2 = x*x;
655 float x3 = x2*x;
657 if ( x > 0.5f ) {
658 return 0.5625f - (x3 / 6.0f - 3.0f * x2 * 0.25f + 1.125f * x);
659 }
660 if ( x > -0.5f ) {
661 return 0.5f - (0.75f * x - x3 / 3.0f);
662 }
663 return 0.4375f + (-x3 / 6.0f - 3.0f * x2 * 0.25f - 1.125f * x);
664 }
666 /* ComputeBlurProfile allocates and fills in an array of floating
667 point values between 0 and 255 for the profile signature of
668 a blurred half-plane with the given blur radius. Since we're
669 going to be doing screened multiplications (i.e., 1 - (1-x)(1-y))
670 all the time, we actually fill in the profile pre-inverted
671 (already done 255-x).
673 It's the responsibility of the caller to delete the
674 memory returned in profile_out.
675 */
677 void SkBlurMask::ComputeBlurProfile(SkScalar sigma, uint8_t **profile_out) {
678 int size = SkScalarCeilToInt(6*sigma);
680 int center = size >> 1;
681 uint8_t *profile = SkNEW_ARRAY(uint8_t, size);
683 float invr = 1.f/(2*sigma);
685 profile[0] = 255;
686 for (int x = 1 ; x < size ; ++x) {
687 float scaled_x = (center - x - .5f) * invr;
688 float gi = gaussianIntegral(scaled_x);
689 profile[x] = 255 - (uint8_t) (255.f * gi);
690 }
692 *profile_out = profile;
693 }
695 // TODO MAYBE: Maintain a profile cache to avoid recomputing this for
696 // commonly used radii. Consider baking some of the most common blur radii
697 // directly in as static data?
699 // Implementation adapted from Michael Herf's approach:
700 // http://stereopsis.com/shadowrect/
702 uint8_t SkBlurMask::ProfileLookup(const uint8_t *profile, int loc, int blurred_width, int sharp_width) {
703 int dx = SkAbs32(((loc << 1) + 1) - blurred_width) - sharp_width; // how far are we from the original edge?
704 int ox = dx >> 1;
705 if (ox < 0) {
706 ox = 0;
707 }
709 return profile[ox];
710 }
712 void SkBlurMask::ComputeBlurredScanline(uint8_t *pixels, const uint8_t *profile,
713 unsigned int width, SkScalar sigma) {
715 unsigned int profile_size = SkScalarCeilToInt(6*sigma);
716 SkAutoTMalloc<uint8_t> horizontalScanline(width);
718 unsigned int sw = width - profile_size;
719 // nearest odd number less than the profile size represents the center
720 // of the (2x scaled) profile
721 int center = ( profile_size & ~1 ) - 1;
723 int w = sw - center;
725 for (unsigned int x = 0 ; x < width ; ++x) {
726 if (profile_size <= sw) {
727 pixels[x] = ProfileLookup(profile, x, width, w);
728 } else {
729 float span = float(sw)/(2*sigma);
730 float giX = 1.5f - (x+.5f)/(2*sigma);
731 pixels[x] = (uint8_t) (255 * (gaussianIntegral(giX) - gaussianIntegral(giX + span)));
732 }
733 }
734 }
736 bool SkBlurMask::BlurRect(SkScalar sigma, SkMask *dst,
737 const SkRect &src, Style style,
738 SkIPoint *margin, SkMask::CreateMode createMode) {
739 int profile_size = SkScalarCeilToInt(6*sigma);
741 int pad = profile_size/2;
742 if (margin) {
743 margin->set( pad, pad );
744 }
746 dst->fBounds.set(SkScalarRoundToInt(src.fLeft - pad),
747 SkScalarRoundToInt(src.fTop - pad),
748 SkScalarRoundToInt(src.fRight + pad),
749 SkScalarRoundToInt(src.fBottom + pad));
751 dst->fRowBytes = dst->fBounds.width();
752 dst->fFormat = SkMask::kA8_Format;
753 dst->fImage = NULL;
755 int sw = SkScalarFloorToInt(src.width());
756 int sh = SkScalarFloorToInt(src.height());
758 if (createMode == SkMask::kJustComputeBounds_CreateMode) {
759 if (style == kInner_Style) {
760 dst->fBounds.set(SkScalarRoundToInt(src.fLeft),
761 SkScalarRoundToInt(src.fTop),
762 SkScalarRoundToInt(src.fRight),
763 SkScalarRoundToInt(src.fBottom)); // restore trimmed bounds
764 dst->fRowBytes = sw;
765 }
766 return true;
767 }
768 uint8_t *profile = NULL;
770 ComputeBlurProfile(sigma, &profile);
771 SkAutoTDeleteArray<uint8_t> ada(profile);
773 size_t dstSize = dst->computeImageSize();
774 if (0 == dstSize) {
775 return false; // too big to allocate, abort
776 }
778 uint8_t* dp = SkMask::AllocImage(dstSize);
780 dst->fImage = dp;
782 int dstHeight = dst->fBounds.height();
783 int dstWidth = dst->fBounds.width();
785 uint8_t *outptr = dp;
787 SkAutoTMalloc<uint8_t> horizontalScanline(dstWidth);
788 SkAutoTMalloc<uint8_t> verticalScanline(dstHeight);
790 ComputeBlurredScanline(horizontalScanline, profile, dstWidth, sigma);
791 ComputeBlurredScanline(verticalScanline, profile, dstHeight, sigma);
793 for (int y = 0 ; y < dstHeight ; ++y) {
794 for (int x = 0 ; x < dstWidth ; x++) {
795 unsigned int maskval = SkMulDiv255Round(horizontalScanline[x], verticalScanline[y]);
796 *(outptr++) = maskval;
797 }
798 }
800 if (style == kInner_Style) {
801 // now we allocate the "real" dst, mirror the size of src
802 size_t srcSize = (size_t)(src.width() * src.height());
803 if (0 == srcSize) {
804 return false; // too big to allocate, abort
805 }
806 dst->fImage = SkMask::AllocImage(srcSize);
807 for (int y = 0 ; y < sh ; y++) {
808 uint8_t *blur_scanline = dp + (y+pad)*dstWidth + pad;
809 uint8_t *inner_scanline = dst->fImage + y*sw;
810 memcpy(inner_scanline, blur_scanline, sw);
811 }
812 SkMask::FreeImage(dp);
814 dst->fBounds.set(SkScalarRoundToInt(src.fLeft),
815 SkScalarRoundToInt(src.fTop),
816 SkScalarRoundToInt(src.fRight),
817 SkScalarRoundToInt(src.fBottom)); // restore trimmed bounds
818 dst->fRowBytes = sw;
820 } else if (style == kOuter_Style) {
821 for (int y = pad ; y < dstHeight-pad ; y++) {
822 uint8_t *dst_scanline = dp + y*dstWidth + pad;
823 memset(dst_scanline, 0, sw);
824 }
825 } else if (style == kSolid_Style) {
826 for (int y = pad ; y < dstHeight-pad ; y++) {
827 uint8_t *dst_scanline = dp + y*dstWidth + pad;
828 memset(dst_scanline, 0xff, sw);
829 }
830 }
831 // normal and solid styles are the same for analytic rect blurs, so don't
832 // need to handle solid specially.
834 return true;
835 }
837 bool SkBlurMask::BlurRRect(SkScalar sigma, SkMask *dst,
838 const SkRRect &src, Style style,
839 SkIPoint *margin, SkMask::CreateMode createMode) {
840 // Temporary for now -- always fail, should cause caller to fall back
841 // to old path. Plumbing just to land API and parallelize effort.
843 return false;
844 }
846 // The "simple" blur is a direct implementation of separable convolution with a discrete
847 // gaussian kernel. It's "ground truth" in a sense; too slow to be used, but very
848 // useful for correctness comparisons.
850 bool SkBlurMask::BlurGroundTruth(SkScalar sigma, SkMask* dst, const SkMask& src,
851 Style style, SkIPoint* margin) {
853 if (src.fFormat != SkMask::kA8_Format) {
854 return false;
855 }
857 float variance = sigma * sigma;
859 int windowSize = SkScalarCeilToInt(sigma*6);
860 // round window size up to nearest odd number
861 windowSize |= 1;
863 SkAutoTMalloc<float> gaussWindow(windowSize);
865 int halfWindow = windowSize >> 1;
867 gaussWindow[halfWindow] = 1;
869 float windowSum = 1;
870 for (int x = 1 ; x <= halfWindow ; ++x) {
871 float gaussian = expf(-x*x / (2*variance));
872 gaussWindow[halfWindow + x] = gaussWindow[halfWindow-x] = gaussian;
873 windowSum += 2*gaussian;
874 }
876 // leave the filter un-normalized for now; we will divide by the normalization
877 // sum later;
879 int pad = halfWindow;
880 if (margin) {
881 margin->set( pad, pad );
882 }
884 dst->fBounds = src.fBounds;
885 dst->fBounds.outset(pad, pad);
887 dst->fRowBytes = dst->fBounds.width();
888 dst->fFormat = SkMask::kA8_Format;
889 dst->fImage = NULL;
891 if (src.fImage) {
893 size_t dstSize = dst->computeImageSize();
894 if (0 == dstSize) {
895 return false; // too big to allocate, abort
896 }
898 int srcWidth = src.fBounds.width();
899 int srcHeight = src.fBounds.height();
900 int dstWidth = dst->fBounds.width();
902 const uint8_t* srcPixels = src.fImage;
903 uint8_t* dstPixels = SkMask::AllocImage(dstSize);
904 SkAutoTCallVProc<uint8_t, SkMask_FreeImage> autoCall(dstPixels);
906 // do the actual blur. First, make a padded copy of the source.
907 // use double pad so we never have to check if we're outside anything
909 int padWidth = srcWidth + 4*pad;
910 int padHeight = srcHeight;
911 int padSize = padWidth * padHeight;
913 SkAutoTMalloc<uint8_t> padPixels(padSize);
914 memset(padPixels, 0, padSize);
916 for (int y = 0 ; y < srcHeight; ++y) {
917 uint8_t* padptr = padPixels + y * padWidth + 2*pad;
918 const uint8_t* srcptr = srcPixels + y * srcWidth;
919 memcpy(padptr, srcptr, srcWidth);
920 }
922 // blur in X, transposing the result into a temporary floating point buffer.
923 // also double-pad the intermediate result so that the second blur doesn't
924 // have to do extra conditionals.
926 int tmpWidth = padHeight + 4*pad;
927 int tmpHeight = padWidth - 2*pad;
928 int tmpSize = tmpWidth * tmpHeight;
930 SkAutoTMalloc<float> tmpImage(tmpSize);
931 memset(tmpImage, 0, tmpSize*sizeof(tmpImage[0]));
933 for (int y = 0 ; y < padHeight ; ++y) {
934 uint8_t *srcScanline = padPixels + y*padWidth;
935 for (int x = pad ; x < padWidth - pad ; ++x) {
936 float *outPixel = tmpImage + (x-pad)*tmpWidth + y + 2*pad; // transposed output
937 uint8_t *windowCenter = srcScanline + x;
938 for (int i = -pad ; i <= pad ; ++i) {
939 *outPixel += gaussWindow[pad+i]*windowCenter[i];
940 }
941 *outPixel /= windowSum;
942 }
943 }
945 // blur in Y; now filling in the actual desired destination. We have to do
946 // the transpose again; these transposes guarantee that we read memory in
947 // linear order.
949 for (int y = 0 ; y < tmpHeight ; ++y) {
950 float *srcScanline = tmpImage + y*tmpWidth;
951 for (int x = pad ; x < tmpWidth - pad ; ++x) {
952 float *windowCenter = srcScanline + x;
953 float finalValue = 0;
954 for (int i = -pad ; i <= pad ; ++i) {
955 finalValue += gaussWindow[pad+i]*windowCenter[i];
956 }
957 finalValue /= windowSum;
958 uint8_t *outPixel = dstPixels + (x-pad)*dstWidth + y; // transposed output
959 int integerPixel = int(finalValue + 0.5f);
960 *outPixel = SkClampMax( SkClampPos(integerPixel), 255 );
961 }
962 }
964 dst->fImage = dstPixels;
965 // if need be, alloc the "real" dst (same size as src) and copy/merge
966 // the blur into it (applying the src)
967 if (style == kInner_Style) {
968 // now we allocate the "real" dst, mirror the size of src
969 size_t srcSize = src.computeImageSize();
970 if (0 == srcSize) {
971 return false; // too big to allocate, abort
972 }
973 dst->fImage = SkMask::AllocImage(srcSize);
974 merge_src_with_blur(dst->fImage, src.fRowBytes,
975 srcPixels, src.fRowBytes,
976 dstPixels + pad*dst->fRowBytes + pad,
977 dst->fRowBytes, srcWidth, srcHeight);
978 SkMask::FreeImage(dstPixels);
979 } else if (style != kNormal_Style) {
980 clamp_with_orig(dstPixels + pad*dst->fRowBytes + pad,
981 dst->fRowBytes, srcPixels, src.fRowBytes, srcWidth, srcHeight, style);
982 }
983 (void)autoCall.detach();
984 }
986 if (style == kInner_Style) {
987 dst->fBounds = src.fBounds; // restore trimmed bounds
988 dst->fRowBytes = src.fRowBytes;
989 }
991 return true;
992 }
