| File: | root/firefox-clang/obj-x86_64-pc-linux-gnu/gfx/2d/./../../../gfx/2d/SkConvolver.cpp |
| Warning: | line 597, column 50 Value stored to 'filterValues' during its initialization is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | // Copyright (c) 2011-2016 Google Inc. |
| 2 | // Use of this source code is governed by a BSD-style license that can be |
| 3 | // found in the gfx/skia/LICENSE file. |
| 4 | |
| 5 | #include "SkConvolver.h" |
| 6 | |
| 7 | #include <algorithm> |
| 8 | |
| 9 | #ifdef USE_SSE21 |
| 10 | # include "mozilla/SSE.h" |
| 11 | #endif |
| 12 | |
| 13 | #ifdef USE_NEON |
| 14 | # include "mozilla/arm.h" |
| 15 | #endif |
| 16 | |
| 17 | namespace skia { |
| 18 | |
| 19 | using mozilla::gfx::BytesPerPixel; |
| 20 | using mozilla::gfx::IsOpaque; |
| 21 | using mozilla::gfx::SurfaceFormat; |
| 22 | |
| 23 | // Converts the argument to an 8-bit unsigned value by clamping to the range |
| 24 | // 0-255. |
| 25 | static inline unsigned char ClampTo8(int a) { |
| 26 | if (static_cast<unsigned>(a) < 256) { |
| 27 | return a; // Avoid the extra check in the common case. |
| 28 | } |
| 29 | if (a < 0) { |
| 30 | return 0; |
| 31 | } |
| 32 | return 255; |
| 33 | } |
| 34 | |
| 35 | // Convolves horizontally along a single row. The row data is given in |
| 36 | // |srcData| and continues for the numValues() of the filter. |
| 37 | template <bool hasAlpha> |
| 38 | void ConvolveHorizontally(const unsigned char* srcData, |
| 39 | const SkConvolutionFilter1D& filter, |
| 40 | unsigned char* outRow) { |
| 41 | // Loop over each pixel on this row in the output image. |
| 42 | int numValues = filter.numValues(); |
| 43 | for (int outX = 0; outX < numValues; outX++) { |
| 44 | // Get the filter that determines the current output pixel. |
| 45 | int filterOffset, filterLength; |
| 46 | const SkConvolutionFilter1D::ConvolutionFixed* filterValues = |
| 47 | filter.FilterForValue(outX, &filterOffset, &filterLength); |
| 48 | |
| 49 | // Compute the first pixel in this row that the filter affects. It will |
| 50 | // touch |filterLength| pixels (4 bytes each) after this. |
| 51 | const unsigned char* rowToFilter = &srcData[filterOffset * 4]; |
| 52 | |
| 53 | // Apply the filter to the row to get the destination pixel in |accum|. |
| 54 | int accum[4] = {0}; |
| 55 | for (int filterX = 0; filterX < filterLength; filterX++) { |
| 56 | SkConvolutionFilter1D::ConvolutionFixed curFilter = filterValues[filterX]; |
| 57 | accum[0] += curFilter * rowToFilter[filterX * 4 + 0]; |
| 58 | accum[1] += curFilter * rowToFilter[filterX * 4 + 1]; |
| 59 | accum[2] += curFilter * rowToFilter[filterX * 4 + 2]; |
| 60 | if (hasAlpha) { |
| 61 | accum[3] += curFilter * rowToFilter[filterX * 4 + 3]; |
| 62 | } |
| 63 | } |
| 64 | |
| 65 | // Bring this value back in range. All of the filter scaling factors |
| 66 | // are in fixed point with kShiftBits bits of fractional part. |
| 67 | // Add rounding bias before truncating to avoid systematic darkening. |
| 68 | constexpr int kRound = 1 << (SkConvolutionFilter1D::kShiftBits - 1); |
| 69 | accum[0] = (accum[0] + kRound) >> SkConvolutionFilter1D::kShiftBits; |
| 70 | accum[1] = (accum[1] + kRound) >> SkConvolutionFilter1D::kShiftBits; |
| 71 | accum[2] = (accum[2] + kRound) >> SkConvolutionFilter1D::kShiftBits; |
| 72 | |
| 73 | if (hasAlpha) { |
| 74 | accum[3] = (accum[3] + kRound) >> SkConvolutionFilter1D::kShiftBits; |
| 75 | } |
| 76 | |
| 77 | // Store the new pixel. |
| 78 | outRow[outX * 4 + 0] = ClampTo8(accum[0]); |
| 79 | outRow[outX * 4 + 1] = ClampTo8(accum[1]); |
| 80 | outRow[outX * 4 + 2] = ClampTo8(accum[2]); |
| 81 | if (hasAlpha) { |
| 82 | outRow[outX * 4 + 3] = ClampTo8(accum[3]); |
| 83 | } |
| 84 | } |
| 85 | } |
| 86 | |
| 87 | // Does vertical convolution to produce one output row. The filter values and |
| 88 | // length are given in the first two parameters. These are applied to each |
| 89 | // of the rows pointed to in the |sourceDataRows| array, with each row |
| 90 | // being |pixelWidth| wide. |
| 91 | // |
| 92 | // The output must have room for |pixelWidth * 4| bytes. |
| 93 | template <bool hasAlpha> |
| 94 | void ConvolveVertically( |
| 95 | const SkConvolutionFilter1D::ConvolutionFixed* filterValues, |
| 96 | int filterLength, unsigned char* const* sourceDataRows, int pixelWidth, |
| 97 | unsigned char* outRow) { |
| 98 | // We go through each column in the output and do a vertical convolution, |
| 99 | // generating one output pixel each time. |
| 100 | for (int outX = 0; outX < pixelWidth; outX++) { |
| 101 | // Compute the number of bytes over in each row that the current column |
| 102 | // we're convolving starts at. The pixel will cover the next 4 bytes. |
| 103 | int byteOffset = outX * 4; |
| 104 | |
| 105 | // Apply the filter to one column of pixels. |
| 106 | int accum[4] = {0}; |
| 107 | for (int filterY = 0; filterY < filterLength; filterY++) { |
| 108 | SkConvolutionFilter1D::ConvolutionFixed curFilter = filterValues[filterY]; |
| 109 | accum[0] += curFilter * sourceDataRows[filterY][byteOffset + 0]; |
| 110 | accum[1] += curFilter * sourceDataRows[filterY][byteOffset + 1]; |
| 111 | accum[2] += curFilter * sourceDataRows[filterY][byteOffset + 2]; |
| 112 | if (hasAlpha) { |
| 113 | accum[3] += curFilter * sourceDataRows[filterY][byteOffset + 3]; |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | // Bring this value back in range. All of the filter scaling factors |
| 118 | // are in fixed point with kShiftBits bits of precision. |
| 119 | constexpr int kRound = 1 << (SkConvolutionFilter1D::kShiftBits - 1); |
| 120 | accum[0] = (accum[0] + kRound) >> SkConvolutionFilter1D::kShiftBits; |
| 121 | accum[1] = (accum[1] + kRound) >> SkConvolutionFilter1D::kShiftBits; |
| 122 | accum[2] = (accum[2] + kRound) >> SkConvolutionFilter1D::kShiftBits; |
| 123 | if (hasAlpha) { |
| 124 | accum[3] = (accum[3] + kRound) >> SkConvolutionFilter1D::kShiftBits; |
| 125 | } |
| 126 | |
| 127 | // Store the new pixel. |
| 128 | outRow[byteOffset + 0] = ClampTo8(accum[0]); |
| 129 | outRow[byteOffset + 1] = ClampTo8(accum[1]); |
| 130 | outRow[byteOffset + 2] = ClampTo8(accum[2]); |
| 131 | |
| 132 | if (hasAlpha) { |
| 133 | unsigned char alpha = ClampTo8(accum[3]); |
| 134 | |
| 135 | // Make sure the alpha channel doesn't come out smaller than any of the |
| 136 | // color channels. We use premultipled alpha channels, so this should |
| 137 | // never happen, but rounding errors will cause this from time to time. |
| 138 | // These "impossible" colors will cause overflows (and hence random pixel |
| 139 | // values) when the resulting bitmap is drawn to the screen. |
| 140 | // |
| 141 | // We only need to do this when generating the final output row (here). |
| 142 | int maxColorChannel = |
| 143 | std::max(outRow[byteOffset + 0], |
| 144 | std::max(outRow[byteOffset + 1], outRow[byteOffset + 2])); |
| 145 | if (alpha < maxColorChannel) { |
| 146 | outRow[byteOffset + 3] = maxColorChannel; |
| 147 | } else { |
| 148 | outRow[byteOffset + 3] = alpha; |
| 149 | } |
| 150 | } else { |
| 151 | // No alpha channel, the image is opaque. |
| 152 | outRow[byteOffset + 3] = 0xff; |
| 153 | } |
| 154 | } |
| 155 | } |
| 156 | |
| 157 | // Convolves horizontally along a single row. The row data is given in |
| 158 | // |srcData| and continues for the numValues() of the filter. |
| 159 | void ConvolveHorizontallyA8(const unsigned char* srcData, |
| 160 | const SkConvolutionFilter1D& filter, |
| 161 | unsigned char* outRow) { |
| 162 | // Loop over each pixel on this row in the output image. |
| 163 | int numValues = filter.numValues(); |
| 164 | for (int outX = 0; outX < numValues; outX++) { |
| 165 | // Get the filter that determines the current output pixel. |
| 166 | int filterOffset, filterLength; |
| 167 | const SkConvolutionFilter1D::ConvolutionFixed* filterValues = |
| 168 | filter.FilterForValue(outX, &filterOffset, &filterLength); |
| 169 | |
| 170 | // Compute the first pixel in this row that the filter affects. It will |
| 171 | // touch |filterLength| pixels (4 bytes each) after this. |
| 172 | const unsigned char* rowToFilter = &srcData[filterOffset]; |
| 173 | |
| 174 | // Apply the filter to the row to get the destination pixel in |accum|. |
| 175 | int accum = 0; |
| 176 | for (int filterX = 0; filterX < filterLength; filterX++) { |
| 177 | SkConvolutionFilter1D::ConvolutionFixed curFilter = filterValues[filterX]; |
| 178 | accum += curFilter * rowToFilter[filterX]; |
| 179 | } |
| 180 | |
| 181 | // Bring this value back in range. All of the filter scaling factors |
| 182 | // are in fixed point with kShiftBits bits of fractional part. |
| 183 | constexpr int kRound = 1 << (SkConvolutionFilter1D::kShiftBits - 1); |
| 184 | accum = (accum + kRound) >> SkConvolutionFilter1D::kShiftBits; |
| 185 | |
| 186 | // Store the new pixel. |
| 187 | outRow[outX] = ClampTo8(accum); |
| 188 | } |
| 189 | } |
| 190 | |
| 191 | // Does vertical convolution to produce one output row. The filter values and |
| 192 | // length are given in the first two parameters. These are applied to each |
| 193 | // of the rows pointed to in the |sourceDataRows| array, with each row |
| 194 | // being |pixelWidth| wide. |
| 195 | // |
| 196 | // The output must have room for |pixelWidth| bytes. |
| 197 | void ConvolveVerticallyA8( |
| 198 | const SkConvolutionFilter1D::ConvolutionFixed* filterValues, |
| 199 | int filterLength, unsigned char* const* sourceDataRows, int pixelWidth, |
| 200 | unsigned char* outRow) { |
| 201 | // We go through each column in the output and do a vertical convolution, |
| 202 | // generating one output pixel each time. |
| 203 | for (int outX = 0; outX < pixelWidth; outX++) { |
| 204 | // Apply the filter to one column of pixels. |
| 205 | int accum = 0; |
| 206 | for (int filterY = 0; filterY < filterLength; filterY++) { |
| 207 | SkConvolutionFilter1D::ConvolutionFixed curFilter = filterValues[filterY]; |
| 208 | accum += curFilter * sourceDataRows[filterY][outX]; |
| 209 | } |
| 210 | |
| 211 | // Bring this value back in range. All of the filter scaling factors |
| 212 | // are in fixed point with kShiftBits bits of precision. |
| 213 | constexpr int kRound = 1 << (SkConvolutionFilter1D::kShiftBits - 1); |
| 214 | accum = (accum + kRound) >> SkConvolutionFilter1D::kShiftBits; |
| 215 | |
| 216 | // Store the new pixel. |
| 217 | outRow[outX] = ClampTo8(accum); |
| 218 | } |
| 219 | } |
| 220 | |
| 221 | #ifdef USE_SSE21 |
| 222 | void convolve_vertically_avx2(const int16_t* filter, int filterLen, |
| 223 | uint8_t* const* srcRows, int width, uint8_t* out, |
| 224 | bool hasAlpha); |
| 225 | void convolve_horizontally_avx2(const unsigned char* srcData, |
| 226 | const SkConvolutionFilter1D& filter, |
| 227 | unsigned char* outRow, bool hasAlpha); |
| 228 | void convolve_horizontally_sse2(const unsigned char* srcData, |
| 229 | const SkConvolutionFilter1D& filter, |
| 230 | unsigned char* outRow, bool hasAlpha); |
| 231 | void convolve_vertically_sse2(const int16_t* filter, int filterLen, |
| 232 | uint8_t* const* srcRows, int width, uint8_t* out, |
| 233 | bool hasAlpha); |
| 234 | #elif defined(USE_NEON) |
| 235 | void convolve_horizontally_neon(const unsigned char* srcData, |
| 236 | const SkConvolutionFilter1D& filter, |
| 237 | unsigned char* outRow, bool hasAlpha); |
| 238 | void convolve_vertically_neon(const int16_t* filter, int filterLen, |
| 239 | uint8_t* const* srcRows, int width, uint8_t* out, |
| 240 | bool hasAlpha); |
| 241 | #endif |
| 242 | |
| 243 | void convolve_horizontally(const unsigned char* srcData, |
| 244 | const SkConvolutionFilter1D& filter, |
| 245 | unsigned char* outRow, SurfaceFormat format) { |
| 246 | if (format == SurfaceFormat::A8) { |
| 247 | ConvolveHorizontallyA8(srcData, filter, outRow); |
| 248 | return; |
| 249 | } |
| 250 | |
| 251 | bool hasAlpha = !IsOpaque(format); |
| 252 | #ifdef USE_SSE21 |
| 253 | if (mozilla::supports_avx2()) { |
| 254 | convolve_horizontally_avx2(srcData, filter, outRow, hasAlpha); |
| 255 | return; |
| 256 | } |
| 257 | if (mozilla::supports_sse2()) { |
| 258 | convolve_horizontally_sse2(srcData, filter, outRow, hasAlpha); |
| 259 | return; |
| 260 | } |
| 261 | #elif defined(USE_NEON) |
| 262 | if (mozilla::supports_neon()) { |
| 263 | convolve_horizontally_neon(srcData, filter, outRow, hasAlpha); |
| 264 | return; |
| 265 | } |
| 266 | #endif |
| 267 | if (hasAlpha) { |
| 268 | ConvolveHorizontally<true>(srcData, filter, outRow); |
| 269 | } else { |
| 270 | ConvolveHorizontally<false>(srcData, filter, outRow); |
| 271 | } |
| 272 | } |
| 273 | |
| 274 | void convolve_vertically( |
| 275 | const SkConvolutionFilter1D::ConvolutionFixed* filterValues, |
| 276 | int filterLength, unsigned char* const* sourceDataRows, int pixelWidth, |
| 277 | unsigned char* outRow, SurfaceFormat format) { |
| 278 | if (format == SurfaceFormat::A8) { |
| 279 | ConvolveVerticallyA8(filterValues, filterLength, sourceDataRows, pixelWidth, |
| 280 | outRow); |
| 281 | return; |
| 282 | } |
| 283 | |
| 284 | bool hasAlpha = !IsOpaque(format); |
| 285 | #ifdef USE_SSE21 |
| 286 | if (mozilla::supports_avx2()) { |
| 287 | convolve_vertically_avx2(filterValues, filterLength, sourceDataRows, |
| 288 | pixelWidth, outRow, hasAlpha); |
| 289 | return; |
| 290 | } |
| 291 | if (mozilla::supports_sse2()) { |
| 292 | convolve_vertically_sse2(filterValues, filterLength, sourceDataRows, |
| 293 | pixelWidth, outRow, hasAlpha); |
| 294 | return; |
| 295 | } |
| 296 | #elif defined(USE_NEON) |
| 297 | if (mozilla::supports_neon()) { |
| 298 | convolve_vertically_neon(filterValues, filterLength, sourceDataRows, |
| 299 | pixelWidth, outRow, hasAlpha); |
| 300 | return; |
| 301 | } |
| 302 | #endif |
| 303 | if (hasAlpha) { |
| 304 | ConvolveVertically<true>(filterValues, filterLength, sourceDataRows, |
| 305 | pixelWidth, outRow); |
| 306 | } else { |
| 307 | ConvolveVertically<false>(filterValues, filterLength, sourceDataRows, |
| 308 | pixelWidth, outRow); |
| 309 | } |
| 310 | } |
| 311 | |
| 312 | // Stores a list of rows in a circular buffer. The usage is you write into it |
| 313 | // by calling AdvanceRow. It will keep track of which row in the buffer it |
| 314 | // should use next, and the total number of rows added. |
| 315 | class CircularRowBuffer { |
| 316 | public: |
| 317 | // The number of pixels in each row is given in |sourceRowPixelWidth|. |
| 318 | // The maximum number of rows needed in the buffer is |maxYFilterSize| |
| 319 | // (we only need to store enough rows for the biggest filter). |
| 320 | // |
| 321 | // We use the |firstInputRow| to compute the coordinates of all of the |
| 322 | // following rows returned by Advance(). |
| 323 | CircularRowBuffer(int destRowPixelWidth, int maxYFilterSize, |
| 324 | int firstInputRow) |
| 325 | : fRowByteWidth(destRowPixelWidth * 4), |
| 326 | fNumRows(maxYFilterSize), |
| 327 | fNextRow(0), |
| 328 | fNextRowCoordinate(firstInputRow) {} |
| 329 | |
| 330 | bool AllocBuffer() { |
| 331 | return fBuffer.resize(fRowByteWidth * fNumRows) && |
| 332 | fRowAddresses.resize(fNumRows); |
| 333 | } |
| 334 | |
| 335 | // Moves to the next row in the buffer, returning a pointer to the beginning |
| 336 | // of it. |
| 337 | unsigned char* advanceRow() { |
| 338 | unsigned char* row = &fBuffer[fNextRow * fRowByteWidth]; |
| 339 | fNextRowCoordinate++; |
| 340 | |
| 341 | // Set the pointer to the next row to use, wrapping around if necessary. |
| 342 | fNextRow++; |
| 343 | if (fNextRow == fNumRows) { |
| 344 | fNextRow = 0; |
| 345 | } |
| 346 | return row; |
| 347 | } |
| 348 | |
| 349 | // Returns a pointer to an "unrolled" array of rows. These rows will start |
| 350 | // at the y coordinate placed into |*firstRowIndex| and will continue in |
| 351 | // order for the maximum number of rows in this circular buffer. |
| 352 | // |
| 353 | // The |firstRowIndex_| may be negative. This means the circular buffer |
| 354 | // starts before the top of the image (it hasn't been filled yet). |
| 355 | unsigned char* const* GetRowAddresses(int* firstRowIndex) { |
| 356 | // Example for a 4-element circular buffer holding coords 6-9. |
| 357 | // Row 0 Coord 8 |
| 358 | // Row 1 Coord 9 |
| 359 | // Row 2 Coord 6 <- fNextRow = 2, fNextRowCoordinate = 10. |
| 360 | // Row 3 Coord 7 |
| 361 | // |
| 362 | // The "next" row is also the first (lowest) coordinate. This computation |
| 363 | // may yield a negative value, but that's OK, the math will work out |
| 364 | // since the user of this buffer will compute the offset relative |
| 365 | // to the firstRowIndex and the negative rows will never be used. |
| 366 | *firstRowIndex = fNextRowCoordinate - fNumRows; |
| 367 | |
| 368 | int curRow = fNextRow; |
| 369 | for (int i = 0; i < fNumRows; i++) { |
| 370 | fRowAddresses[i] = &fBuffer[curRow * fRowByteWidth]; |
| 371 | |
| 372 | // Advance to the next row, wrapping if necessary. |
| 373 | curRow++; |
| 374 | if (curRow == fNumRows) { |
| 375 | curRow = 0; |
| 376 | } |
| 377 | } |
| 378 | return &fRowAddresses[0]; |
| 379 | } |
| 380 | |
| 381 | private: |
| 382 | // The buffer storing the rows. They are packed, each one fRowByteWidth. |
| 383 | mozilla::Vector<unsigned char> fBuffer; |
| 384 | |
| 385 | // Number of bytes per row in the |buffer|. |
| 386 | int fRowByteWidth; |
| 387 | |
| 388 | // The number of rows available in the buffer. |
| 389 | int fNumRows; |
| 390 | |
| 391 | // The next row index we should write into. This wraps around as the |
| 392 | // circular buffer is used. |
| 393 | int fNextRow; |
| 394 | |
| 395 | // The y coordinate of the |fNextRow|. This is incremented each time a |
| 396 | // new row is appended and does not wrap. |
| 397 | int fNextRowCoordinate; |
| 398 | |
| 399 | // Buffer used by GetRowAddresses(). |
| 400 | mozilla::Vector<unsigned char*> fRowAddresses; |
| 401 | }; |
| 402 | |
| 403 | SkConvolutionFilter1D::SkConvolutionFilter1D() : fMaxFilter(0) {} |
| 404 | |
| 405 | bool SkConvolutionFilter1D::AddFilter(int filterOffset, |
| 406 | const ConvolutionFixed* filterValues, |
| 407 | int filterLength) { |
| 408 | // It is common for leading/trailing filter values to be zeros. In such |
| 409 | // cases it is beneficial to only store the central factors. |
| 410 | // For a scaling to 1/4th in each dimension using a Lanczos-2 filter on |
| 411 | // a 1080p image this optimization gives a ~10% speed improvement. |
| 412 | int filterSize = filterLength; |
| 413 | int firstNonZero = 0; |
| 414 | while (firstNonZero < filterLength && filterValues[firstNonZero] == 0) { |
| 415 | firstNonZero++; |
| 416 | } |
| 417 | |
| 418 | if (firstNonZero < filterLength) { |
| 419 | // Here we have at least one non-zero factor. |
| 420 | int lastNonZero = filterLength - 1; |
| 421 | while (lastNonZero >= 0 && filterValues[lastNonZero] == 0) { |
| 422 | lastNonZero--; |
| 423 | } |
| 424 | |
| 425 | filterOffset += firstNonZero; |
| 426 | filterLength = lastNonZero + 1 - firstNonZero; |
| 427 | MOZ_ASSERT(filterLength > 0)do { static_assert( mozilla::detail::AssertionConditionType< decltype(filterLength > 0)>::isValid, "invalid assertion condition" ); if ((__builtin_expect(!!(!(!!(filterLength > 0))), 0))) { do { } while (false); MOZ_ReportAssertionFailure("filterLength > 0" , "./../../../gfx/2d/SkConvolver.cpp", 427); AnnotateMozCrashReason ("MOZ_ASSERT" "(" "filterLength > 0" ")"); do { MOZ_CrashSequence (__null, 427); __attribute__((nomerge)) ::abort(); } while (false ); } } while (false); |
| 428 | |
| 429 | if (!fFilterValues.append(&filterValues[firstNonZero], filterLength)) { |
| 430 | return false; |
| 431 | } |
| 432 | } else { |
| 433 | // Here all the factors were zeroes. |
| 434 | filterLength = 0; |
| 435 | } |
| 436 | |
| 437 | FilterInstance instance = { |
| 438 | // We pushed filterLength elements onto fFilterValues |
| 439 | int(fFilterValues.length()) - filterLength, filterOffset, filterLength, |
| 440 | filterSize}; |
| 441 | if (!fFilters.append(instance)) { |
| 442 | if (filterLength > 0) { |
| 443 | fFilterValues.shrinkBy(filterLength); |
| 444 | } |
| 445 | return false; |
| 446 | } |
| 447 | |
| 448 | fMaxFilter = std::max(fMaxFilter, filterLength); |
| 449 | return true; |
| 450 | } |
| 451 | |
| 452 | bool SkConvolutionFilter1D::ComputeFilterValues( |
| 453 | const SkBitmapFilter& aBitmapFilter, int32_t aSrcSize, int32_t aDstSize) { |
| 454 | // When we're doing a magnification, the scale will be larger than one. This |
| 455 | // means the destination pixels are much smaller than the source pixels, and |
| 456 | // that the range covered by the filter won't necessarily cover any source |
| 457 | // pixel boundaries. Therefore, we use these clamped values (max of 1) for |
| 458 | // some computations. |
| 459 | float scale = float(aDstSize) / float(aSrcSize); |
| 460 | float clampedScale = std::min(1.0f, scale); |
| 461 | // This is how many source pixels from the center we need to count |
| 462 | // to support the filtering function. |
| 463 | float srcSupport = aBitmapFilter.width() / clampedScale; |
| 464 | float invScale = 1.0f / scale; |
| 465 | |
| 466 | mozilla::Vector<float, 64> filterValues; |
| 467 | mozilla::Vector<ConvolutionFixed, 64> fixedFilterValues; |
| 468 | |
| 469 | // Loop over all pixels in the output range. We will generate one set of |
| 470 | // filter values for each one. Those values will tell us how to blend the |
| 471 | // source pixels to compute the destination pixel. |
| 472 | |
| 473 | // This value is computed based on how SkTDArray::resizeStorageToAtLeast works |
| 474 | // in order to ensure that it does not overflow or assert. That functions |
| 475 | // computes |
| 476 | // n+4 + (n+4)/4 |
| 477 | // and we want to to fit in a 32 bit signed int. Equating that to 2^31-1 and |
| 478 | // solving n gives n = (2^31-6)*4/5 = 1717986913.6 |
| 479 | const int32_t maxToPassToReserveAdditional = 1717986913; |
| 480 | |
| 481 | int32_t filterValueCount = int32_t(ceilf(aDstSize * srcSupport * 2)); |
| 482 | if (aDstSize > maxToPassToReserveAdditional || filterValueCount < 0 || |
| 483 | filterValueCount > maxToPassToReserveAdditional || |
| 484 | !reserveAdditional(aDstSize, filterValueCount)) { |
| 485 | return false; |
| 486 | } |
| 487 | size_t oldFiltersLength = fFilters.length(); |
| 488 | size_t oldFilterValuesLength = fFilterValues.length(); |
| 489 | int oldMaxFilter = fMaxFilter; |
| 490 | for (int32_t destI = 0; destI < aDstSize; destI++) { |
| 491 | // This is the pixel in the source directly under the pixel in the dest. |
| 492 | // Note that we base computations on the "center" of the pixels. To see |
| 493 | // why, observe that the destination pixel at coordinates (0, 0) in a 5.0x |
| 494 | // downscale should "cover" the pixels around the pixel with *its center* |
| 495 | // at coordinates (2.5, 2.5) in the source, not those around (0, 0). |
| 496 | // Hence we need to scale coordinates (0.5, 0.5), not (0, 0). |
| 497 | float srcPixel = (static_cast<float>(destI) + 0.5f) * invScale; |
| 498 | |
| 499 | // Compute the (inclusive) range of source pixels the filter covers. |
| 500 | // Clamp in the integer domain to avoid float rounding imprecision with |
| 501 | // values near int32 extremes. |
| 502 | int32_t srcBegin = |
| 503 | int32_t(std::clamp(int64_t(floorf(srcPixel - srcSupport)), int64_t(0), |
| 504 | int64_t(aSrcSize) - 1)); |
| 505 | int32_t srcEnd = int32_t(std::clamp(int64_t(ceilf(srcPixel + srcSupport)), |
| 506 | int64_t(0), int64_t(aSrcSize) - 1)); |
| 507 | |
| 508 | // Compute the unnormalized filter value at each location of the source |
| 509 | // it covers. |
| 510 | |
| 511 | // Sum of the filter values for normalizing. |
| 512 | // Distance from the center of the filter, this is the filter coordinate |
| 513 | // in source space. We also need to consider the center of the pixel |
| 514 | // when comparing distance against 'srcPixel'. In the 5x downscale |
| 515 | // example used above the distance from the center of the filter to |
| 516 | // the pixel with coordinates (2, 2) should be 0, because its center |
| 517 | // is at (2.5, 2.5). |
| 518 | int32_t filterCount = srcEnd - srcBegin + 1; |
| 519 | if (filterCount <= 0 || !filterValues.resize(filterCount) || |
| 520 | !fixedFilterValues.resize(filterCount)) { |
| 521 | return false; |
| 522 | } |
| 523 | |
| 524 | float destFilterDist = |
| 525 | (static_cast<float>(srcBegin) + 0.5f - srcPixel) * clampedScale; |
| 526 | float filterSum = 0.0f; |
| 527 | for (int32_t index = 0; index < filterCount; index++) { |
| 528 | float filterValue = aBitmapFilter.evaluate(destFilterDist); |
| 529 | filterValues[index] = filterValue; |
| 530 | filterSum += filterValue; |
| 531 | destFilterDist += clampedScale; |
| 532 | } |
| 533 | |
| 534 | // The filter must be normalized so that we don't affect the brightness of |
| 535 | // the image. Convert to normalized fixed point. |
| 536 | ConvolutionFixed fixedSum = 0; |
| 537 | float invFilterSum = 1.0f / filterSum; |
| 538 | for (int32_t fixedI = 0; fixedI < filterCount; fixedI++) { |
| 539 | ConvolutionFixed curFixed = ToFixed(filterValues[fixedI] * invFilterSum); |
| 540 | fixedSum += curFixed; |
| 541 | fixedFilterValues[fixedI] = curFixed; |
| 542 | } |
| 543 | |
| 544 | // The conversion to fixed point will leave some rounding errors, which |
| 545 | // we add back in to avoid affecting the brightness of the image. We |
| 546 | // arbitrarily add this to the center of the filter array (this won't always |
| 547 | // be the center of the filter function since it could get clipped on the |
| 548 | // edges, but it doesn't matter enough to worry about that case). |
| 549 | ConvolutionFixed leftovers = ToFixed(1) - fixedSum; |
| 550 | fixedFilterValues[filterCount / 2] += leftovers; |
| 551 | |
| 552 | if (!AddFilter(srcBegin, fixedFilterValues.begin(), filterCount)) { |
| 553 | fFilters.shrinkTo(oldFiltersLength); |
| 554 | fFilterValues.shrinkTo(oldFilterValuesLength); |
| 555 | fMaxFilter = oldMaxFilter; |
| 556 | return false; |
| 557 | } |
| 558 | } |
| 559 | |
| 560 | return maxFilter() > 0 && numValues() == aDstSize; |
| 561 | } |
| 562 | |
| 563 | // Does a two-dimensional convolution on the given source image. |
| 564 | // |
| 565 | // It is assumed the source pixel offsets referenced in the input filters |
| 566 | // reference only valid pixels, so the source image size is not required. Each |
| 567 | // row of the source image starts |sourceByteRowStride| after the previous |
| 568 | // one (this allows you to have rows with some padding at the end). |
| 569 | // |
| 570 | // The result will be put into the given output buffer. The destination image |
| 571 | // size will be xfilter.numValues() * yfilter.numValues() pixels. It will be |
| 572 | // in rows of exactly xfilter.numValues() * 4 bytes. |
| 573 | // |
| 574 | // |sourceHasAlpha| is a hint that allows us to avoid doing computations on |
| 575 | // the alpha channel if the image is opaque. If you don't know, set this to |
| 576 | // true and it will work properly, but setting this to false will be a few |
| 577 | // percent faster if you know the image is opaque. |
| 578 | // |
| 579 | // The layout in memory is assumed to be 4-bytes per pixel in B-G-R-A order |
| 580 | // (this is ARGB when loaded into 32-bit words on a little-endian machine). |
| 581 | /** |
| 582 | * Returns false if it was unable to perform the convolution/rescale. in which |
| 583 | * case the output buffer is assumed to be undefined. |
| 584 | */ |
| 585 | bool BGRAConvolve2D(const unsigned char* sourceData, int sourceByteRowStride, |
| 586 | SurfaceFormat format, const SkConvolutionFilter1D& filterX, |
| 587 | const SkConvolutionFilter1D& filterY, |
| 588 | int outputByteRowStride, unsigned char* output) { |
| 589 | int maxYFilterSize = filterY.maxFilter(); |
| 590 | |
| 591 | // The next row in the input that we will generate a horizontally |
| 592 | // convolved row for. If the filter doesn't start at the beginning of the |
| 593 | // image (this is the case when we are only resizing a subset), then we |
| 594 | // don't want to generate any output rows before that. Compute the starting |
| 595 | // row for convolution as the first pixel for the first vertical filter. |
| 596 | int filterOffset = 0, filterLength = 0; |
| 597 | const SkConvolutionFilter1D::ConvolutionFixed* filterValues = |
Value stored to 'filterValues' during its initialization is never read | |
| 598 | filterY.FilterForValue(0, &filterOffset, &filterLength); |
| 599 | int nextXRow = filterOffset; |
| 600 | |
| 601 | // We loop over each row in the input doing a horizontal convolution. This |
| 602 | // will result in a horizontally convolved image. We write the results into |
| 603 | // a circular buffer of convolved rows and do vertical convolution as rows |
| 604 | // are available. This prevents us from having to store the entire |
| 605 | // intermediate image and helps cache coherency. |
| 606 | // We will need four extra rows to allow horizontal convolution could be done |
| 607 | // simultaneously. We also pad each row in row buffer to be aligned-up to |
| 608 | // 32 bytes. |
| 609 | // TODO(jiesun): We do not use aligned load from row buffer in vertical |
| 610 | // convolution pass yet. Somehow Windows does not like it. |
| 611 | int rowBufferWidth = (filterX.numValues() + 31) & ~0x1F; |
| 612 | int rowBufferHeight = maxYFilterSize; |
| 613 | |
| 614 | // check for too-big allocation requests : crbug.com/528628 |
| 615 | { |
| 616 | int64_t size = int64_t(rowBufferWidth) * int64_t(rowBufferHeight); |
| 617 | // need some limit, to avoid over-committing success from malloc, but then |
| 618 | // crashing when we try to actually use the memory. |
| 619 | // 100meg seems big enough to allow "normal" zoom factors and image sizes |
| 620 | // through while avoiding the crash seen by the bug (crbug.com/528628) |
| 621 | if (size > 100 * 1024 * 1024) { |
| 622 | // printf_stderr("BGRAConvolve2D: tmp allocation [%lld] too |
| 623 | // big\n", size); |
| 624 | return false; |
| 625 | } |
| 626 | } |
| 627 | |
| 628 | CircularRowBuffer rowBuffer(rowBufferWidth, rowBufferHeight, filterOffset); |
| 629 | if (!rowBuffer.AllocBuffer()) { |
| 630 | return false; |
| 631 | } |
| 632 | |
| 633 | // Loop over every possible output row, processing just enough horizontal |
| 634 | // convolutions to run each subsequent vertical convolution. |
| 635 | MOZ_ASSERT(outputByteRowStride >=do { static_assert( mozilla::detail::AssertionConditionType< decltype(outputByteRowStride >= filterX.numValues() * BytesPerPixel (format))>::isValid, "invalid assertion condition"); if (( __builtin_expect(!!(!(!!(outputByteRowStride >= filterX.numValues () * BytesPerPixel(format)))), 0))) { do { } while (false); MOZ_ReportAssertionFailure ("outputByteRowStride >= filterX.numValues() * BytesPerPixel(format)" , "./../../../gfx/2d/SkConvolver.cpp", 636); AnnotateMozCrashReason ("MOZ_ASSERT" "(" "outputByteRowStride >= filterX.numValues() * BytesPerPixel(format)" ")"); do { MOZ_CrashSequence(__null, 636); __attribute__((nomerge )) ::abort(); } while (false); } } while (false) |
| 636 | filterX.numValues() * BytesPerPixel(format))do { static_assert( mozilla::detail::AssertionConditionType< decltype(outputByteRowStride >= filterX.numValues() * BytesPerPixel (format))>::isValid, "invalid assertion condition"); if (( __builtin_expect(!!(!(!!(outputByteRowStride >= filterX.numValues () * BytesPerPixel(format)))), 0))) { do { } while (false); MOZ_ReportAssertionFailure ("outputByteRowStride >= filterX.numValues() * BytesPerPixel(format)" , "./../../../gfx/2d/SkConvolver.cpp", 636); AnnotateMozCrashReason ("MOZ_ASSERT" "(" "outputByteRowStride >= filterX.numValues() * BytesPerPixel(format)" ")"); do { MOZ_CrashSequence(__null, 636); __attribute__((nomerge )) ::abort(); } while (false); } } while (false); |
| 637 | int numOutputRows = filterY.numValues(); |
| 638 | |
| 639 | // We need to check which is the last line to convolve before we advance 4 |
| 640 | // lines in one iteration. |
| 641 | int lastFilterOffset, lastFilterLength; |
| 642 | filterY.FilterForValue(numOutputRows - 1, &lastFilterOffset, |
| 643 | &lastFilterLength); |
| 644 | |
| 645 | for (int outY = 0; outY < numOutputRows; outY++) { |
| 646 | filterValues = filterY.FilterForValue(outY, &filterOffset, &filterLength); |
| 647 | |
| 648 | // Generate output rows until we have enough to run the current filter. |
| 649 | while (nextXRow < filterOffset + filterLength) { |
| 650 | convolve_horizontally( |
| 651 | &sourceData[(uint64_t)nextXRow * sourceByteRowStride], filterX, |
| 652 | rowBuffer.advanceRow(), format); |
| 653 | nextXRow++; |
| 654 | } |
| 655 | |
| 656 | // Compute where in the output image this row of final data will go. |
| 657 | unsigned char* curOutputRow = &output[(uint64_t)outY * outputByteRowStride]; |
| 658 | |
| 659 | // Get the list of rows that the circular buffer has, in order. |
| 660 | int firstRowInCircularBuffer; |
| 661 | unsigned char* const* rowsToConvolve = |
| 662 | rowBuffer.GetRowAddresses(&firstRowInCircularBuffer); |
| 663 | |
| 664 | // Now compute the start of the subset of those rows that the filter needs. |
| 665 | unsigned char* const* firstRowForFilter = |
| 666 | &rowsToConvolve[filterOffset - firstRowInCircularBuffer]; |
| 667 | |
| 668 | convolve_vertically(filterValues, filterLength, firstRowForFilter, |
| 669 | filterX.numValues(), curOutputRow, format); |
| 670 | } |
| 671 | return true; |
| 672 | } |
| 673 | |
| 674 | } // namespace skia |