1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/skia/trunk/src/effects/SkBlurMask.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,992 @@
1.4 +
1.5 +/*
1.6 + * Copyright 2006 The Android Open Source Project
1.7 + *
1.8 + * Use of this source code is governed by a BSD-style license that can be
1.9 + * found in the LICENSE file.
1.10 + */
1.11 +
1.12 +
1.13 +#include "SkBlurMask.h"
1.14 +#include "SkMath.h"
1.15 +#include "SkTemplates.h"
1.16 +#include "SkEndian.h"
1.17 +
1.18 +
1.19 +SkScalar SkBlurMask::ConvertRadiusToSigma(SkScalar radius) {
1.20 + // This constant approximates the scaling done in the software path's
1.21 + // "high quality" mode, in SkBlurMask::Blur() (1 / sqrt(3)).
1.22 + // IMHO, it actually should be 1: we blur "less" than we should do
1.23 + // according to the CSS and canvas specs, simply because Safari does the same.
1.24 + // Firefox used to do the same too, until 4.0 where they fixed it. So at some
1.25 + // point we should probably get rid of these scaling constants and rebaseline
1.26 + // all the blur tests.
1.27 + static const SkScalar kBLUR_SIGMA_SCALE = 0.57735f;
1.28 +
1.29 + return radius ? kBLUR_SIGMA_SCALE * radius + 0.5f : 0.0f;
1.30 +}
1.31 +
1.32 +#define UNROLL_SEPARABLE_LOOPS
1.33 +
1.34 +/**
1.35 + * This function performs a box blur in X, of the given radius. If the
1.36 + * "transpose" parameter is true, it will transpose the pixels on write,
1.37 + * such that X and Y are swapped. Reads are always performed from contiguous
1.38 + * memory in X, for speed. The destination buffer (dst) must be at least
1.39 + * (width + leftRadius + rightRadius) * height bytes in size.
1.40 + *
1.41 + * This is what the inner loop looks like before unrolling, and with the two
1.42 + * cases broken out separately (width < diameter, width >= diameter):
1.43 + *
1.44 + * if (width < diameter) {
1.45 + * for (int x = 0; x < width; ++x) {
1.46 + * sum += *right++;
1.47 + * *dptr = (sum * scale + half) >> 24;
1.48 + * dptr += dst_x_stride;
1.49 + * }
1.50 + * for (int x = width; x < diameter; ++x) {
1.51 + * *dptr = (sum * scale + half) >> 24;
1.52 + * dptr += dst_x_stride;
1.53 + * }
1.54 + * for (int x = 0; x < width; ++x) {
1.55 + * *dptr = (sum * scale + half) >> 24;
1.56 + * sum -= *left++;
1.57 + * dptr += dst_x_stride;
1.58 + * }
1.59 + * } else {
1.60 + * for (int x = 0; x < diameter; ++x) {
1.61 + * sum += *right++;
1.62 + * *dptr = (sum * scale + half) >> 24;
1.63 + * dptr += dst_x_stride;
1.64 + * }
1.65 + * for (int x = diameter; x < width; ++x) {
1.66 + * sum += *right++;
1.67 + * *dptr = (sum * scale + half) >> 24;
1.68 + * sum -= *left++;
1.69 + * dptr += dst_x_stride;
1.70 + * }
1.71 + * for (int x = 0; x < diameter; ++x) {
1.72 + * *dptr = (sum * scale + half) >> 24;
1.73 + * sum -= *left++;
1.74 + * dptr += dst_x_stride;
1.75 + * }
1.76 + * }
1.77 + */
1.78 +static int boxBlur(const uint8_t* src, int src_y_stride, uint8_t* dst,
1.79 + int leftRadius, int rightRadius, int width, int height,
1.80 + bool transpose)
1.81 +{
1.82 + int diameter = leftRadius + rightRadius;
1.83 + int kernelSize = diameter + 1;
1.84 + int border = SkMin32(width, diameter);
1.85 + uint32_t scale = (1 << 24) / kernelSize;
1.86 + int new_width = width + SkMax32(leftRadius, rightRadius) * 2;
1.87 + int dst_x_stride = transpose ? height : 1;
1.88 + int dst_y_stride = transpose ? 1 : new_width;
1.89 + uint32_t half = 1 << 23;
1.90 + for (int y = 0; y < height; ++y) {
1.91 + uint32_t sum = 0;
1.92 + uint8_t* dptr = dst + y * dst_y_stride;
1.93 + const uint8_t* right = src + y * src_y_stride;
1.94 + const uint8_t* left = right;
1.95 + for (int x = 0; x < rightRadius - leftRadius; x++) {
1.96 + *dptr = 0;
1.97 + dptr += dst_x_stride;
1.98 + }
1.99 +#define LEFT_BORDER_ITER \
1.100 + sum += *right++; \
1.101 + *dptr = (sum * scale + half) >> 24; \
1.102 + dptr += dst_x_stride;
1.103 +
1.104 + int x = 0;
1.105 +#ifdef UNROLL_SEPARABLE_LOOPS
1.106 + for (; x < border - 16; x += 16) {
1.107 + LEFT_BORDER_ITER
1.108 + LEFT_BORDER_ITER
1.109 + LEFT_BORDER_ITER
1.110 + LEFT_BORDER_ITER
1.111 + LEFT_BORDER_ITER
1.112 + LEFT_BORDER_ITER
1.113 + LEFT_BORDER_ITER
1.114 + LEFT_BORDER_ITER
1.115 + LEFT_BORDER_ITER
1.116 + LEFT_BORDER_ITER
1.117 + LEFT_BORDER_ITER
1.118 + LEFT_BORDER_ITER
1.119 + LEFT_BORDER_ITER
1.120 + LEFT_BORDER_ITER
1.121 + LEFT_BORDER_ITER
1.122 + LEFT_BORDER_ITER
1.123 + }
1.124 +#endif
1.125 + for (; x < border; ++x) {
1.126 + LEFT_BORDER_ITER
1.127 + }
1.128 +#undef LEFT_BORDER_ITER
1.129 +#define TRIVIAL_ITER \
1.130 + *dptr = (sum * scale + half) >> 24; \
1.131 + dptr += dst_x_stride;
1.132 + x = width;
1.133 +#ifdef UNROLL_SEPARABLE_LOOPS
1.134 + for (; x < diameter - 16; x += 16) {
1.135 + TRIVIAL_ITER
1.136 + TRIVIAL_ITER
1.137 + TRIVIAL_ITER
1.138 + TRIVIAL_ITER
1.139 + TRIVIAL_ITER
1.140 + TRIVIAL_ITER
1.141 + TRIVIAL_ITER
1.142 + TRIVIAL_ITER
1.143 + TRIVIAL_ITER
1.144 + TRIVIAL_ITER
1.145 + TRIVIAL_ITER
1.146 + TRIVIAL_ITER
1.147 + TRIVIAL_ITER
1.148 + TRIVIAL_ITER
1.149 + TRIVIAL_ITER
1.150 + TRIVIAL_ITER
1.151 + }
1.152 +#endif
1.153 + for (; x < diameter; ++x) {
1.154 + TRIVIAL_ITER
1.155 + }
1.156 +#undef TRIVIAL_ITER
1.157 +#define CENTER_ITER \
1.158 + sum += *right++; \
1.159 + *dptr = (sum * scale + half) >> 24; \
1.160 + sum -= *left++; \
1.161 + dptr += dst_x_stride;
1.162 +
1.163 + x = diameter;
1.164 +#ifdef UNROLL_SEPARABLE_LOOPS
1.165 + for (; x < width - 16; x += 16) {
1.166 + CENTER_ITER
1.167 + CENTER_ITER
1.168 + CENTER_ITER
1.169 + CENTER_ITER
1.170 + CENTER_ITER
1.171 + CENTER_ITER
1.172 + CENTER_ITER
1.173 + CENTER_ITER
1.174 + CENTER_ITER
1.175 + CENTER_ITER
1.176 + CENTER_ITER
1.177 + CENTER_ITER
1.178 + CENTER_ITER
1.179 + CENTER_ITER
1.180 + CENTER_ITER
1.181 + CENTER_ITER
1.182 + }
1.183 +#endif
1.184 + for (; x < width; ++x) {
1.185 + CENTER_ITER
1.186 + }
1.187 +#undef CENTER_ITER
1.188 +#define RIGHT_BORDER_ITER \
1.189 + *dptr = (sum * scale + half) >> 24; \
1.190 + sum -= *left++; \
1.191 + dptr += dst_x_stride;
1.192 +
1.193 + x = 0;
1.194 +#ifdef UNROLL_SEPARABLE_LOOPS
1.195 + for (; x < border - 16; x += 16) {
1.196 + RIGHT_BORDER_ITER
1.197 + RIGHT_BORDER_ITER
1.198 + RIGHT_BORDER_ITER
1.199 + RIGHT_BORDER_ITER
1.200 + RIGHT_BORDER_ITER
1.201 + RIGHT_BORDER_ITER
1.202 + RIGHT_BORDER_ITER
1.203 + RIGHT_BORDER_ITER
1.204 + RIGHT_BORDER_ITER
1.205 + RIGHT_BORDER_ITER
1.206 + RIGHT_BORDER_ITER
1.207 + RIGHT_BORDER_ITER
1.208 + RIGHT_BORDER_ITER
1.209 + RIGHT_BORDER_ITER
1.210 + RIGHT_BORDER_ITER
1.211 + RIGHT_BORDER_ITER
1.212 + }
1.213 +#endif
1.214 + for (; x < border; ++x) {
1.215 + RIGHT_BORDER_ITER
1.216 + }
1.217 +#undef RIGHT_BORDER_ITER
1.218 + for (int x = 0; x < leftRadius - rightRadius; ++x) {
1.219 + *dptr = 0;
1.220 + dptr += dst_x_stride;
1.221 + }
1.222 + SkASSERT(sum == 0);
1.223 + }
1.224 + return new_width;
1.225 +}
1.226 +
1.227 +/**
1.228 + * This variant of the box blur handles blurring of non-integer radii. It
1.229 + * keeps two running sums: an outer sum for the rounded-up kernel radius, and
1.230 + * an inner sum for the rounded-down kernel radius. For each pixel, it linearly
1.231 + * interpolates between them. In float this would be:
1.232 + * outer_weight * outer_sum / kernelSize +
1.233 + * (1.0 - outer_weight) * innerSum / (kernelSize - 2)
1.234 + *
1.235 + * This is what the inner loop looks like before unrolling, and with the two
1.236 + * cases broken out separately (width < diameter, width >= diameter):
1.237 + *
1.238 + * if (width < diameter) {
1.239 + * for (int x = 0; x < width; x++) {
1.240 + * inner_sum = outer_sum;
1.241 + * outer_sum += *right++;
1.242 + * *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
1.243 + * dptr += dst_x_stride;
1.244 + * }
1.245 + * for (int x = width; x < diameter; ++x) {
1.246 + * *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
1.247 + * dptr += dst_x_stride;
1.248 + * }
1.249 + * for (int x = 0; x < width; x++) {
1.250 + * inner_sum = outer_sum - *left++;
1.251 + * *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
1.252 + * dptr += dst_x_stride;
1.253 + * outer_sum = inner_sum;
1.254 + * }
1.255 + * } else {
1.256 + * for (int x = 0; x < diameter; x++) {
1.257 + * inner_sum = outer_sum;
1.258 + * outer_sum += *right++;
1.259 + * *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
1.260 + * dptr += dst_x_stride;
1.261 + * }
1.262 + * for (int x = diameter; x < width; ++x) {
1.263 + * inner_sum = outer_sum - *left;
1.264 + * outer_sum += *right++;
1.265 + * *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
1.266 + * dptr += dst_x_stride;
1.267 + * outer_sum -= *left++;
1.268 + * }
1.269 + * for (int x = 0; x < diameter; x++) {
1.270 + * inner_sum = outer_sum - *left++;
1.271 + * *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
1.272 + * dptr += dst_x_stride;
1.273 + * outer_sum = inner_sum;
1.274 + * }
1.275 + * }
1.276 + * }
1.277 + * return new_width;
1.278 + */
1.279 +
1.280 +static int boxBlurInterp(const uint8_t* src, int src_y_stride, uint8_t* dst,
1.281 + int radius, int width, int height,
1.282 + bool transpose, uint8_t outer_weight)
1.283 +{
1.284 + int diameter = radius * 2;
1.285 + int kernelSize = diameter + 1;
1.286 + int border = SkMin32(width, diameter);
1.287 + int inner_weight = 255 - outer_weight;
1.288 + outer_weight += outer_weight >> 7;
1.289 + inner_weight += inner_weight >> 7;
1.290 + uint32_t outer_scale = (outer_weight << 16) / kernelSize;
1.291 + uint32_t inner_scale = (inner_weight << 16) / (kernelSize - 2);
1.292 + uint32_t half = 1 << 23;
1.293 + int new_width = width + diameter;
1.294 + int dst_x_stride = transpose ? height : 1;
1.295 + int dst_y_stride = transpose ? 1 : new_width;
1.296 + for (int y = 0; y < height; ++y) {
1.297 + uint32_t outer_sum = 0, inner_sum = 0;
1.298 + uint8_t* dptr = dst + y * dst_y_stride;
1.299 + const uint8_t* right = src + y * src_y_stride;
1.300 + const uint8_t* left = right;
1.301 + int x = 0;
1.302 +
1.303 +#define LEFT_BORDER_ITER \
1.304 + inner_sum = outer_sum; \
1.305 + outer_sum += *right++; \
1.306 + *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24; \
1.307 + dptr += dst_x_stride;
1.308 +
1.309 +#ifdef UNROLL_SEPARABLE_LOOPS
1.310 + for (;x < border - 16; x += 16) {
1.311 + LEFT_BORDER_ITER
1.312 + LEFT_BORDER_ITER
1.313 + LEFT_BORDER_ITER
1.314 + LEFT_BORDER_ITER
1.315 + LEFT_BORDER_ITER
1.316 + LEFT_BORDER_ITER
1.317 + LEFT_BORDER_ITER
1.318 + LEFT_BORDER_ITER
1.319 + LEFT_BORDER_ITER
1.320 + LEFT_BORDER_ITER
1.321 + LEFT_BORDER_ITER
1.322 + LEFT_BORDER_ITER
1.323 + LEFT_BORDER_ITER
1.324 + LEFT_BORDER_ITER
1.325 + LEFT_BORDER_ITER
1.326 + LEFT_BORDER_ITER
1.327 + }
1.328 +#endif
1.329 +
1.330 + for (;x < border; ++x) {
1.331 + LEFT_BORDER_ITER
1.332 + }
1.333 +#undef LEFT_BORDER_ITER
1.334 + for (int x = width; x < diameter; ++x) {
1.335 + *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24;
1.336 + dptr += dst_x_stride;
1.337 + }
1.338 + x = diameter;
1.339 +
1.340 +#define CENTER_ITER \
1.341 + inner_sum = outer_sum - *left; \
1.342 + outer_sum += *right++; \
1.343 + *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24; \
1.344 + dptr += dst_x_stride; \
1.345 + outer_sum -= *left++;
1.346 +
1.347 +#ifdef UNROLL_SEPARABLE_LOOPS
1.348 + for (; x < width - 16; x += 16) {
1.349 + CENTER_ITER
1.350 + CENTER_ITER
1.351 + CENTER_ITER
1.352 + CENTER_ITER
1.353 + CENTER_ITER
1.354 + CENTER_ITER
1.355 + CENTER_ITER
1.356 + CENTER_ITER
1.357 + CENTER_ITER
1.358 + CENTER_ITER
1.359 + CENTER_ITER
1.360 + CENTER_ITER
1.361 + CENTER_ITER
1.362 + CENTER_ITER
1.363 + CENTER_ITER
1.364 + CENTER_ITER
1.365 + }
1.366 +#endif
1.367 + for (; x < width; ++x) {
1.368 + CENTER_ITER
1.369 + }
1.370 +#undef CENTER_ITER
1.371 +
1.372 + #define RIGHT_BORDER_ITER \
1.373 + inner_sum = outer_sum - *left++; \
1.374 + *dptr = (outer_sum * outer_scale + inner_sum * inner_scale + half) >> 24; \
1.375 + dptr += dst_x_stride; \
1.376 + outer_sum = inner_sum;
1.377 +
1.378 + x = 0;
1.379 +#ifdef UNROLL_SEPARABLE_LOOPS
1.380 + for (; x < border - 16; x += 16) {
1.381 + RIGHT_BORDER_ITER
1.382 + RIGHT_BORDER_ITER
1.383 + RIGHT_BORDER_ITER
1.384 + RIGHT_BORDER_ITER
1.385 + RIGHT_BORDER_ITER
1.386 + RIGHT_BORDER_ITER
1.387 + RIGHT_BORDER_ITER
1.388 + RIGHT_BORDER_ITER
1.389 + RIGHT_BORDER_ITER
1.390 + RIGHT_BORDER_ITER
1.391 + RIGHT_BORDER_ITER
1.392 + RIGHT_BORDER_ITER
1.393 + RIGHT_BORDER_ITER
1.394 + RIGHT_BORDER_ITER
1.395 + RIGHT_BORDER_ITER
1.396 + RIGHT_BORDER_ITER
1.397 + }
1.398 +#endif
1.399 + for (; x < border; ++x) {
1.400 + RIGHT_BORDER_ITER
1.401 + }
1.402 +#undef RIGHT_BORDER_ITER
1.403 + SkASSERT(outer_sum == 0 && inner_sum == 0);
1.404 + }
1.405 + return new_width;
1.406 +}
1.407 +
1.408 +static void get_adjusted_radii(SkScalar passRadius, int *loRadius, int *hiRadius)
1.409 +{
1.410 + *loRadius = *hiRadius = SkScalarCeilToInt(passRadius);
1.411 + if (SkIntToScalar(*hiRadius) - passRadius > 0.5f) {
1.412 + *loRadius = *hiRadius - 1;
1.413 + }
1.414 +}
1.415 +
1.416 +#include "SkColorPriv.h"
1.417 +
1.418 +static void merge_src_with_blur(uint8_t dst[], int dstRB,
1.419 + const uint8_t src[], int srcRB,
1.420 + const uint8_t blur[], int blurRB,
1.421 + int sw, int sh) {
1.422 + dstRB -= sw;
1.423 + srcRB -= sw;
1.424 + blurRB -= sw;
1.425 + while (--sh >= 0) {
1.426 + for (int x = sw - 1; x >= 0; --x) {
1.427 + *dst = SkToU8(SkAlphaMul(*blur, SkAlpha255To256(*src)));
1.428 + dst += 1;
1.429 + src += 1;
1.430 + blur += 1;
1.431 + }
1.432 + dst += dstRB;
1.433 + src += srcRB;
1.434 + blur += blurRB;
1.435 + }
1.436 +}
1.437 +
1.438 +static void clamp_with_orig(uint8_t dst[], int dstRowBytes,
1.439 + const uint8_t src[], int srcRowBytes,
1.440 + int sw, int sh,
1.441 + SkBlurMask::Style style) {
1.442 + int x;
1.443 + while (--sh >= 0) {
1.444 + switch (style) {
1.445 + case SkBlurMask::kSolid_Style:
1.446 + for (x = sw - 1; x >= 0; --x) {
1.447 + int s = *src;
1.448 + int d = *dst;
1.449 + *dst = SkToU8(s + d - SkMulDiv255Round(s, d));
1.450 + dst += 1;
1.451 + src += 1;
1.452 + }
1.453 + break;
1.454 + case SkBlurMask::kOuter_Style:
1.455 + for (x = sw - 1; x >= 0; --x) {
1.456 + if (*src) {
1.457 + *dst = SkToU8(SkAlphaMul(*dst, SkAlpha255To256(255 - *src)));
1.458 + }
1.459 + dst += 1;
1.460 + src += 1;
1.461 + }
1.462 + break;
1.463 + default:
1.464 + SkDEBUGFAIL("Unexpected blur style here");
1.465 + break;
1.466 + }
1.467 + dst += dstRowBytes - sw;
1.468 + src += srcRowBytes - sw;
1.469 + }
1.470 +}
1.471 +
1.472 +///////////////////////////////////////////////////////////////////////////////
1.473 +
1.474 +// we use a local function to wrap the class static method to work around
1.475 +// a bug in gcc98
1.476 +void SkMask_FreeImage(uint8_t* image);
1.477 +void SkMask_FreeImage(uint8_t* image) {
1.478 + SkMask::FreeImage(image);
1.479 +}
1.480 +
1.481 +bool SkBlurMask::BoxBlur(SkMask* dst, const SkMask& src,
1.482 + SkScalar sigma, Style style, Quality quality,
1.483 + SkIPoint* margin) {
1.484 +
1.485 + if (src.fFormat != SkMask::kA8_Format) {
1.486 + return false;
1.487 + }
1.488 +
1.489 + // Force high quality off for small radii (performance)
1.490 + if (sigma <= SkIntToScalar(2)) {
1.491 + quality = kLow_Quality;
1.492 + }
1.493 +
1.494 + SkScalar passRadius;
1.495 + if (kHigh_Quality == quality) {
1.496 + // For the high quality path the 3 pass box blur kernel width is
1.497 + // 6*rad+1 while the full Gaussian width is 6*sigma.
1.498 + passRadius = sigma - (1/6.0f);
1.499 + } else {
1.500 + // For the low quality path we only attempt to cover 3*sigma of the
1.501 + // Gaussian blur area (1.5*sigma on each side). The single pass box
1.502 + // blur's kernel size is 2*rad+1.
1.503 + passRadius = 1.5f*sigma - 0.5f;
1.504 + }
1.505 +
1.506 + // highQuality: use three box blur passes as a cheap way
1.507 + // to approximate a Gaussian blur
1.508 + int passCount = (kHigh_Quality == quality) ? 3 : 1;
1.509 +
1.510 + int rx = SkScalarCeilToInt(passRadius);
1.511 + int outerWeight = 255 - SkScalarRoundToInt((SkIntToScalar(rx) - passRadius) * 255);
1.512 +
1.513 + SkASSERT(rx >= 0);
1.514 + SkASSERT((unsigned)outerWeight <= 255);
1.515 + if (rx <= 0) {
1.516 + return false;
1.517 + }
1.518 +
1.519 + int ry = rx; // only do square blur for now
1.520 +
1.521 + int padx = passCount * rx;
1.522 + int pady = passCount * ry;
1.523 +
1.524 + if (margin) {
1.525 + margin->set(padx, pady);
1.526 + }
1.527 + dst->fBounds.set(src.fBounds.fLeft - padx, src.fBounds.fTop - pady,
1.528 + src.fBounds.fRight + padx, src.fBounds.fBottom + pady);
1.529 +
1.530 + dst->fRowBytes = dst->fBounds.width();
1.531 + dst->fFormat = SkMask::kA8_Format;
1.532 + dst->fImage = NULL;
1.533 +
1.534 + if (src.fImage) {
1.535 + size_t dstSize = dst->computeImageSize();
1.536 + if (0 == dstSize) {
1.537 + return false; // too big to allocate, abort
1.538 + }
1.539 +
1.540 + int sw = src.fBounds.width();
1.541 + int sh = src.fBounds.height();
1.542 + const uint8_t* sp = src.fImage;
1.543 + uint8_t* dp = SkMask::AllocImage(dstSize);
1.544 + SkAutoTCallVProc<uint8_t, SkMask_FreeImage> autoCall(dp);
1.545 +
1.546 + // build the blurry destination
1.547 + SkAutoTMalloc<uint8_t> tmpBuffer(dstSize);
1.548 + uint8_t* tp = tmpBuffer.get();
1.549 + int w = sw, h = sh;
1.550 +
1.551 + if (outerWeight == 255) {
1.552 + int loRadius, hiRadius;
1.553 + get_adjusted_radii(passRadius, &loRadius, &hiRadius);
1.554 + if (kHigh_Quality == quality) {
1.555 + // Do three X blurs, with a transpose on the final one.
1.556 + w = boxBlur(sp, src.fRowBytes, tp, loRadius, hiRadius, w, h, false);
1.557 + w = boxBlur(tp, w, dp, hiRadius, loRadius, w, h, false);
1.558 + w = boxBlur(dp, w, tp, hiRadius, hiRadius, w, h, true);
1.559 + // Do three Y blurs, with a transpose on the final one.
1.560 + h = boxBlur(tp, h, dp, loRadius, hiRadius, h, w, false);
1.561 + h = boxBlur(dp, h, tp, hiRadius, loRadius, h, w, false);
1.562 + h = boxBlur(tp, h, dp, hiRadius, hiRadius, h, w, true);
1.563 + } else {
1.564 + w = boxBlur(sp, src.fRowBytes, tp, rx, rx, w, h, true);
1.565 + h = boxBlur(tp, h, dp, ry, ry, h, w, true);
1.566 + }
1.567 + } else {
1.568 + if (kHigh_Quality == quality) {
1.569 + // Do three X blurs, with a transpose on the final one.
1.570 + w = boxBlurInterp(sp, src.fRowBytes, tp, rx, w, h, false, outerWeight);
1.571 + w = boxBlurInterp(tp, w, dp, rx, w, h, false, outerWeight);
1.572 + w = boxBlurInterp(dp, w, tp, rx, w, h, true, outerWeight);
1.573 + // Do three Y blurs, with a transpose on the final one.
1.574 + h = boxBlurInterp(tp, h, dp, ry, h, w, false, outerWeight);
1.575 + h = boxBlurInterp(dp, h, tp, ry, h, w, false, outerWeight);
1.576 + h = boxBlurInterp(tp, h, dp, ry, h, w, true, outerWeight);
1.577 + } else {
1.578 + w = boxBlurInterp(sp, src.fRowBytes, tp, rx, w, h, true, outerWeight);
1.579 + h = boxBlurInterp(tp, h, dp, ry, h, w, true, outerWeight);
1.580 + }
1.581 + }
1.582 +
1.583 + dst->fImage = dp;
1.584 + // if need be, alloc the "real" dst (same size as src) and copy/merge
1.585 + // the blur into it (applying the src)
1.586 + if (style == kInner_Style) {
1.587 + // now we allocate the "real" dst, mirror the size of src
1.588 + size_t srcSize = src.computeImageSize();
1.589 + if (0 == srcSize) {
1.590 + return false; // too big to allocate, abort
1.591 + }
1.592 + dst->fImage = SkMask::AllocImage(srcSize);
1.593 + merge_src_with_blur(dst->fImage, src.fRowBytes,
1.594 + sp, src.fRowBytes,
1.595 + dp + passCount * (rx + ry * dst->fRowBytes),
1.596 + dst->fRowBytes, sw, sh);
1.597 + SkMask::FreeImage(dp);
1.598 + } else if (style != kNormal_Style) {
1.599 + clamp_with_orig(dp + passCount * (rx + ry * dst->fRowBytes),
1.600 + dst->fRowBytes, sp, src.fRowBytes, sw, sh, style);
1.601 + }
1.602 + (void)autoCall.detach();
1.603 + }
1.604 +
1.605 + if (style == kInner_Style) {
1.606 + dst->fBounds = src.fBounds; // restore trimmed bounds
1.607 + dst->fRowBytes = src.fRowBytes;
1.608 + }
1.609 +
1.610 + return true;
1.611 +}
1.612 +
1.613 +/* Convolving a box with itself three times results in a piecewise
1.614 + quadratic function:
1.615 +
1.616 + 0 x <= -1.5
1.617 + 9/8 + 3/2 x + 1/2 x^2 -1.5 < x <= -.5
1.618 + 3/4 - x^2 -.5 < x <= .5
1.619 + 9/8 - 3/2 x + 1/2 x^2 0.5 < x <= 1.5
1.620 + 0 1.5 < x
1.621 +
1.622 + Mathematica:
1.623 +
1.624 + g[x_] := Piecewise [ {
1.625 + {9/8 + 3/2 x + 1/2 x^2 , -1.5 < x <= -.5},
1.626 + {3/4 - x^2 , -.5 < x <= .5},
1.627 + {9/8 - 3/2 x + 1/2 x^2 , 0.5 < x <= 1.5}
1.628 + }, 0]
1.629 +
1.630 + To get the profile curve of the blurred step function at the rectangle
1.631 + edge, we evaluate the indefinite integral, which is piecewise cubic:
1.632 +
1.633 + 0 x <= -1.5
1.634 + 9/16 + 9/8 x + 3/4 x^2 + 1/6 x^3 -1.5 < x <= -0.5
1.635 + 1/2 + 3/4 x - 1/3 x^3 -.5 < x <= .5
1.636 + 7/16 + 9/8 x - 3/4 x^2 + 1/6 x^3 .5 < x <= 1.5
1.637 + 1 1.5 < x
1.638 +
1.639 + in Mathematica code:
1.640 +
1.641 + gi[x_] := Piecewise[ {
1.642 + { 0 , x <= -1.5 },
1.643 + { 9/16 + 9/8 x + 3/4 x^2 + 1/6 x^3, -1.5 < x <= -0.5 },
1.644 + { 1/2 + 3/4 x - 1/3 x^3 , -.5 < x <= .5},
1.645 + { 7/16 + 9/8 x - 3/4 x^2 + 1/6 x^3, .5 < x <= 1.5}
1.646 + },1]
1.647 +*/
1.648 +
1.649 +static float gaussianIntegral(float x) {
1.650 + if (x > 1.5f) {
1.651 + return 0.0f;
1.652 + }
1.653 + if (x < -1.5f) {
1.654 + return 1.0f;
1.655 + }
1.656 +
1.657 + float x2 = x*x;
1.658 + float x3 = x2*x;
1.659 +
1.660 + if ( x > 0.5f ) {
1.661 + return 0.5625f - (x3 / 6.0f - 3.0f * x2 * 0.25f + 1.125f * x);
1.662 + }
1.663 + if ( x > -0.5f ) {
1.664 + return 0.5f - (0.75f * x - x3 / 3.0f);
1.665 + }
1.666 + return 0.4375f + (-x3 / 6.0f - 3.0f * x2 * 0.25f - 1.125f * x);
1.667 +}
1.668 +
1.669 +/* ComputeBlurProfile allocates and fills in an array of floating
1.670 + point values between 0 and 255 for the profile signature of
1.671 + a blurred half-plane with the given blur radius. Since we're
1.672 + going to be doing screened multiplications (i.e., 1 - (1-x)(1-y))
1.673 + all the time, we actually fill in the profile pre-inverted
1.674 + (already done 255-x).
1.675 +
1.676 + It's the responsibility of the caller to delete the
1.677 + memory returned in profile_out.
1.678 +*/
1.679 +
1.680 +void SkBlurMask::ComputeBlurProfile(SkScalar sigma, uint8_t **profile_out) {
1.681 + int size = SkScalarCeilToInt(6*sigma);
1.682 +
1.683 + int center = size >> 1;
1.684 + uint8_t *profile = SkNEW_ARRAY(uint8_t, size);
1.685 +
1.686 + float invr = 1.f/(2*sigma);
1.687 +
1.688 + profile[0] = 255;
1.689 + for (int x = 1 ; x < size ; ++x) {
1.690 + float scaled_x = (center - x - .5f) * invr;
1.691 + float gi = gaussianIntegral(scaled_x);
1.692 + profile[x] = 255 - (uint8_t) (255.f * gi);
1.693 + }
1.694 +
1.695 + *profile_out = profile;
1.696 +}
1.697 +
1.698 +// TODO MAYBE: Maintain a profile cache to avoid recomputing this for
1.699 +// commonly used radii. Consider baking some of the most common blur radii
1.700 +// directly in as static data?
1.701 +
1.702 +// Implementation adapted from Michael Herf's approach:
1.703 +// http://stereopsis.com/shadowrect/
1.704 +
1.705 +uint8_t SkBlurMask::ProfileLookup(const uint8_t *profile, int loc, int blurred_width, int sharp_width) {
1.706 + int dx = SkAbs32(((loc << 1) + 1) - blurred_width) - sharp_width; // how far are we from the original edge?
1.707 + int ox = dx >> 1;
1.708 + if (ox < 0) {
1.709 + ox = 0;
1.710 + }
1.711 +
1.712 + return profile[ox];
1.713 +}
1.714 +
1.715 +void SkBlurMask::ComputeBlurredScanline(uint8_t *pixels, const uint8_t *profile,
1.716 + unsigned int width, SkScalar sigma) {
1.717 +
1.718 + unsigned int profile_size = SkScalarCeilToInt(6*sigma);
1.719 + SkAutoTMalloc<uint8_t> horizontalScanline(width);
1.720 +
1.721 + unsigned int sw = width - profile_size;
1.722 + // nearest odd number less than the profile size represents the center
1.723 + // of the (2x scaled) profile
1.724 + int center = ( profile_size & ~1 ) - 1;
1.725 +
1.726 + int w = sw - center;
1.727 +
1.728 + for (unsigned int x = 0 ; x < width ; ++x) {
1.729 + if (profile_size <= sw) {
1.730 + pixels[x] = ProfileLookup(profile, x, width, w);
1.731 + } else {
1.732 + float span = float(sw)/(2*sigma);
1.733 + float giX = 1.5f - (x+.5f)/(2*sigma);
1.734 + pixels[x] = (uint8_t) (255 * (gaussianIntegral(giX) - gaussianIntegral(giX + span)));
1.735 + }
1.736 + }
1.737 +}
1.738 +
1.739 +bool SkBlurMask::BlurRect(SkScalar sigma, SkMask *dst,
1.740 + const SkRect &src, Style style,
1.741 + SkIPoint *margin, SkMask::CreateMode createMode) {
1.742 + int profile_size = SkScalarCeilToInt(6*sigma);
1.743 +
1.744 + int pad = profile_size/2;
1.745 + if (margin) {
1.746 + margin->set( pad, pad );
1.747 + }
1.748 +
1.749 + dst->fBounds.set(SkScalarRoundToInt(src.fLeft - pad),
1.750 + SkScalarRoundToInt(src.fTop - pad),
1.751 + SkScalarRoundToInt(src.fRight + pad),
1.752 + SkScalarRoundToInt(src.fBottom + pad));
1.753 +
1.754 + dst->fRowBytes = dst->fBounds.width();
1.755 + dst->fFormat = SkMask::kA8_Format;
1.756 + dst->fImage = NULL;
1.757 +
1.758 + int sw = SkScalarFloorToInt(src.width());
1.759 + int sh = SkScalarFloorToInt(src.height());
1.760 +
1.761 + if (createMode == SkMask::kJustComputeBounds_CreateMode) {
1.762 + if (style == kInner_Style) {
1.763 + dst->fBounds.set(SkScalarRoundToInt(src.fLeft),
1.764 + SkScalarRoundToInt(src.fTop),
1.765 + SkScalarRoundToInt(src.fRight),
1.766 + SkScalarRoundToInt(src.fBottom)); // restore trimmed bounds
1.767 + dst->fRowBytes = sw;
1.768 + }
1.769 + return true;
1.770 + }
1.771 + uint8_t *profile = NULL;
1.772 +
1.773 + ComputeBlurProfile(sigma, &profile);
1.774 + SkAutoTDeleteArray<uint8_t> ada(profile);
1.775 +
1.776 + size_t dstSize = dst->computeImageSize();
1.777 + if (0 == dstSize) {
1.778 + return false; // too big to allocate, abort
1.779 + }
1.780 +
1.781 + uint8_t* dp = SkMask::AllocImage(dstSize);
1.782 +
1.783 + dst->fImage = dp;
1.784 +
1.785 + int dstHeight = dst->fBounds.height();
1.786 + int dstWidth = dst->fBounds.width();
1.787 +
1.788 + uint8_t *outptr = dp;
1.789 +
1.790 + SkAutoTMalloc<uint8_t> horizontalScanline(dstWidth);
1.791 + SkAutoTMalloc<uint8_t> verticalScanline(dstHeight);
1.792 +
1.793 + ComputeBlurredScanline(horizontalScanline, profile, dstWidth, sigma);
1.794 + ComputeBlurredScanline(verticalScanline, profile, dstHeight, sigma);
1.795 +
1.796 + for (int y = 0 ; y < dstHeight ; ++y) {
1.797 + for (int x = 0 ; x < dstWidth ; x++) {
1.798 + unsigned int maskval = SkMulDiv255Round(horizontalScanline[x], verticalScanline[y]);
1.799 + *(outptr++) = maskval;
1.800 + }
1.801 + }
1.802 +
1.803 + if (style == kInner_Style) {
1.804 + // now we allocate the "real" dst, mirror the size of src
1.805 + size_t srcSize = (size_t)(src.width() * src.height());
1.806 + if (0 == srcSize) {
1.807 + return false; // too big to allocate, abort
1.808 + }
1.809 + dst->fImage = SkMask::AllocImage(srcSize);
1.810 + for (int y = 0 ; y < sh ; y++) {
1.811 + uint8_t *blur_scanline = dp + (y+pad)*dstWidth + pad;
1.812 + uint8_t *inner_scanline = dst->fImage + y*sw;
1.813 + memcpy(inner_scanline, blur_scanline, sw);
1.814 + }
1.815 + SkMask::FreeImage(dp);
1.816 +
1.817 + dst->fBounds.set(SkScalarRoundToInt(src.fLeft),
1.818 + SkScalarRoundToInt(src.fTop),
1.819 + SkScalarRoundToInt(src.fRight),
1.820 + SkScalarRoundToInt(src.fBottom)); // restore trimmed bounds
1.821 + dst->fRowBytes = sw;
1.822 +
1.823 + } else if (style == kOuter_Style) {
1.824 + for (int y = pad ; y < dstHeight-pad ; y++) {
1.825 + uint8_t *dst_scanline = dp + y*dstWidth + pad;
1.826 + memset(dst_scanline, 0, sw);
1.827 + }
1.828 + } else if (style == kSolid_Style) {
1.829 + for (int y = pad ; y < dstHeight-pad ; y++) {
1.830 + uint8_t *dst_scanline = dp + y*dstWidth + pad;
1.831 + memset(dst_scanline, 0xff, sw);
1.832 + }
1.833 + }
1.834 + // normal and solid styles are the same for analytic rect blurs, so don't
1.835 + // need to handle solid specially.
1.836 +
1.837 + return true;
1.838 +}
1.839 +
1.840 +bool SkBlurMask::BlurRRect(SkScalar sigma, SkMask *dst,
1.841 + const SkRRect &src, Style style,
1.842 + SkIPoint *margin, SkMask::CreateMode createMode) {
1.843 + // Temporary for now -- always fail, should cause caller to fall back
1.844 + // to old path. Plumbing just to land API and parallelize effort.
1.845 +
1.846 + return false;
1.847 +}
1.848 +
1.849 +// The "simple" blur is a direct implementation of separable convolution with a discrete
1.850 +// gaussian kernel. It's "ground truth" in a sense; too slow to be used, but very
1.851 +// useful for correctness comparisons.
1.852 +
1.853 +bool SkBlurMask::BlurGroundTruth(SkScalar sigma, SkMask* dst, const SkMask& src,
1.854 + Style style, SkIPoint* margin) {
1.855 +
1.856 + if (src.fFormat != SkMask::kA8_Format) {
1.857 + return false;
1.858 + }
1.859 +
1.860 + float variance = sigma * sigma;
1.861 +
1.862 + int windowSize = SkScalarCeilToInt(sigma*6);
1.863 + // round window size up to nearest odd number
1.864 + windowSize |= 1;
1.865 +
1.866 + SkAutoTMalloc<float> gaussWindow(windowSize);
1.867 +
1.868 + int halfWindow = windowSize >> 1;
1.869 +
1.870 + gaussWindow[halfWindow] = 1;
1.871 +
1.872 + float windowSum = 1;
1.873 + for (int x = 1 ; x <= halfWindow ; ++x) {
1.874 + float gaussian = expf(-x*x / (2*variance));
1.875 + gaussWindow[halfWindow + x] = gaussWindow[halfWindow-x] = gaussian;
1.876 + windowSum += 2*gaussian;
1.877 + }
1.878 +
1.879 + // leave the filter un-normalized for now; we will divide by the normalization
1.880 + // sum later;
1.881 +
1.882 + int pad = halfWindow;
1.883 + if (margin) {
1.884 + margin->set( pad, pad );
1.885 + }
1.886 +
1.887 + dst->fBounds = src.fBounds;
1.888 + dst->fBounds.outset(pad, pad);
1.889 +
1.890 + dst->fRowBytes = dst->fBounds.width();
1.891 + dst->fFormat = SkMask::kA8_Format;
1.892 + dst->fImage = NULL;
1.893 +
1.894 + if (src.fImage) {
1.895 +
1.896 + size_t dstSize = dst->computeImageSize();
1.897 + if (0 == dstSize) {
1.898 + return false; // too big to allocate, abort
1.899 + }
1.900 +
1.901 + int srcWidth = src.fBounds.width();
1.902 + int srcHeight = src.fBounds.height();
1.903 + int dstWidth = dst->fBounds.width();
1.904 +
1.905 + const uint8_t* srcPixels = src.fImage;
1.906 + uint8_t* dstPixels = SkMask::AllocImage(dstSize);
1.907 + SkAutoTCallVProc<uint8_t, SkMask_FreeImage> autoCall(dstPixels);
1.908 +
1.909 + // do the actual blur. First, make a padded copy of the source.
1.910 + // use double pad so we never have to check if we're outside anything
1.911 +
1.912 + int padWidth = srcWidth + 4*pad;
1.913 + int padHeight = srcHeight;
1.914 + int padSize = padWidth * padHeight;
1.915 +
1.916 + SkAutoTMalloc<uint8_t> padPixels(padSize);
1.917 + memset(padPixels, 0, padSize);
1.918 +
1.919 + for (int y = 0 ; y < srcHeight; ++y) {
1.920 + uint8_t* padptr = padPixels + y * padWidth + 2*pad;
1.921 + const uint8_t* srcptr = srcPixels + y * srcWidth;
1.922 + memcpy(padptr, srcptr, srcWidth);
1.923 + }
1.924 +
1.925 + // blur in X, transposing the result into a temporary floating point buffer.
1.926 + // also double-pad the intermediate result so that the second blur doesn't
1.927 + // have to do extra conditionals.
1.928 +
1.929 + int tmpWidth = padHeight + 4*pad;
1.930 + int tmpHeight = padWidth - 2*pad;
1.931 + int tmpSize = tmpWidth * tmpHeight;
1.932 +
1.933 + SkAutoTMalloc<float> tmpImage(tmpSize);
1.934 + memset(tmpImage, 0, tmpSize*sizeof(tmpImage[0]));
1.935 +
1.936 + for (int y = 0 ; y < padHeight ; ++y) {
1.937 + uint8_t *srcScanline = padPixels + y*padWidth;
1.938 + for (int x = pad ; x < padWidth - pad ; ++x) {
1.939 + float *outPixel = tmpImage + (x-pad)*tmpWidth + y + 2*pad; // transposed output
1.940 + uint8_t *windowCenter = srcScanline + x;
1.941 + for (int i = -pad ; i <= pad ; ++i) {
1.942 + *outPixel += gaussWindow[pad+i]*windowCenter[i];
1.943 + }
1.944 + *outPixel /= windowSum;
1.945 + }
1.946 + }
1.947 +
1.948 + // blur in Y; now filling in the actual desired destination. We have to do
1.949 + // the transpose again; these transposes guarantee that we read memory in
1.950 + // linear order.
1.951 +
1.952 + for (int y = 0 ; y < tmpHeight ; ++y) {
1.953 + float *srcScanline = tmpImage + y*tmpWidth;
1.954 + for (int x = pad ; x < tmpWidth - pad ; ++x) {
1.955 + float *windowCenter = srcScanline + x;
1.956 + float finalValue = 0;
1.957 + for (int i = -pad ; i <= pad ; ++i) {
1.958 + finalValue += gaussWindow[pad+i]*windowCenter[i];
1.959 + }
1.960 + finalValue /= windowSum;
1.961 + uint8_t *outPixel = dstPixels + (x-pad)*dstWidth + y; // transposed output
1.962 + int integerPixel = int(finalValue + 0.5f);
1.963 + *outPixel = SkClampMax( SkClampPos(integerPixel), 255 );
1.964 + }
1.965 + }
1.966 +
1.967 + dst->fImage = dstPixels;
1.968 + // if need be, alloc the "real" dst (same size as src) and copy/merge
1.969 + // the blur into it (applying the src)
1.970 + if (style == kInner_Style) {
1.971 + // now we allocate the "real" dst, mirror the size of src
1.972 + size_t srcSize = src.computeImageSize();
1.973 + if (0 == srcSize) {
1.974 + return false; // too big to allocate, abort
1.975 + }
1.976 + dst->fImage = SkMask::AllocImage(srcSize);
1.977 + merge_src_with_blur(dst->fImage, src.fRowBytes,
1.978 + srcPixels, src.fRowBytes,
1.979 + dstPixels + pad*dst->fRowBytes + pad,
1.980 + dst->fRowBytes, srcWidth, srcHeight);
1.981 + SkMask::FreeImage(dstPixels);
1.982 + } else if (style != kNormal_Style) {
1.983 + clamp_with_orig(dstPixels + pad*dst->fRowBytes + pad,
1.984 + dst->fRowBytes, srcPixels, src.fRowBytes, srcWidth, srcHeight, style);
1.985 + }
1.986 + (void)autoCall.detach();
1.987 + }
1.988 +
1.989 + if (style == kInner_Style) {
1.990 + dst->fBounds = src.fBounds; // restore trimmed bounds
1.991 + dst->fRowBytes = src.fRowBytes;
1.992 + }
1.993 +
1.994 + return true;
1.995 +}
