1 // Copyright (c) 2011 The Chromium Authors. All rights reserved. 2 // Use of this source code is governed by a BSD-style license that can be 3 // found in the LICENSE file. 4 5 #include <algorithm> 6 7 #include "base/logging.h" 8 #include "skia/ext/convolver.h" 9 #include "skia/ext/convolver_SSE2.h" 10 #include "skia/ext/convolver_mips_dspr2.h" 11 #include "third_party/skia/include/core/SkSize.h" 12 #include "third_party/skia/include/core/SkTypes.h" 13 14 namespace skia { 15 16 namespace { 17 18 // Converts the argument to an 8-bit unsigned value by clamping to the range 19 // 0-255. 20 inline unsigned char ClampTo8(int a) { 21 if (static_cast<unsigned>(a) < 256) 22 return a; // Avoid the extra check in the common case. 23 if (a < 0) 24 return 0; 25 return 255; 26 } 27 28 // Takes the value produced by accumulating element-wise product of image with 29 // a kernel and brings it back into range. 30 // All of the filter scaling factors are in fixed point with kShiftBits bits of 31 // fractional part. 32 inline unsigned char BringBackTo8(int a, bool take_absolute) { 33 a >>= ConvolutionFilter1D::kShiftBits; 34 if (take_absolute) 35 a = std::abs(a); 36 return ClampTo8(a); 37 } 38 39 // Stores a list of rows in a circular buffer. The usage is you write into it 40 // by calling AdvanceRow. It will keep track of which row in the buffer it 41 // should use next, and the total number of rows added. 42 class CircularRowBuffer { 43 public: 44 // The number of pixels in each row is given in |source_row_pixel_width|. 45 // The maximum number of rows needed in the buffer is |max_y_filter_size| 46 // (we only need to store enough rows for the biggest filter). 47 // 48 // We use the |first_input_row| to compute the coordinates of all of the 49 // following rows returned by Advance(). 50 CircularRowBuffer(int dest_row_pixel_width, int max_y_filter_size, 51 int first_input_row) 52 : row_byte_width_(dest_row_pixel_width * 4), 53 num_rows_(max_y_filter_size), 54 next_row_(0), 55 next_row_coordinate_(first_input_row) { 56 buffer_.resize(row_byte_width_ * max_y_filter_size); 57 row_addresses_.resize(num_rows_); 58 } 59 60 // Moves to the next row in the buffer, returning a pointer to the beginning 61 // of it. 62 unsigned char* AdvanceRow() { 63 unsigned char* row = &buffer_[next_row_ * row_byte_width_]; 64 next_row_coordinate_++; 65 66 // Set the pointer to the next row to use, wrapping around if necessary. 67 next_row_++; 68 if (next_row_ == num_rows_) 69 next_row_ = 0; 70 return row; 71 } 72 73 // Returns a pointer to an "unrolled" array of rows. These rows will start 74 // at the y coordinate placed into |*first_row_index| and will continue in 75 // order for the maximum number of rows in this circular buffer. 76 // 77 // The |first_row_index_| may be negative. This means the circular buffer 78 // starts before the top of the image (it hasn't been filled yet). 79 unsigned char* const* GetRowAddresses(int* first_row_index) { 80 // Example for a 4-element circular buffer holding coords 6-9. 81 // Row 0 Coord 8 82 // Row 1 Coord 9 83 // Row 2 Coord 6 <- next_row_ = 2, next_row_coordinate_ = 10. 84 // Row 3 Coord 7 85 // 86 // The "next" row is also the first (lowest) coordinate. This computation 87 // may yield a negative value, but that's OK, the math will work out 88 // since the user of this buffer will compute the offset relative 89 // to the first_row_index and the negative rows will never be used. 90 *first_row_index = next_row_coordinate_ - num_rows_; 91 92 int cur_row = next_row_; 93 for (int i = 0; i < num_rows_; i++) { 94 row_addresses_[i] = &buffer_[cur_row * row_byte_width_]; 95 96 // Advance to the next row, wrapping if necessary. 97 cur_row++; 98 if (cur_row == num_rows_) 99 cur_row = 0; 100 } 101 return &row_addresses_[0]; 102 } 103 104 private: 105 // The buffer storing the rows. They are packed, each one row_byte_width_. 106 std::vector<unsigned char> buffer_; 107 108 // Number of bytes per row in the |buffer_|. 109 int row_byte_width_; 110 111 // The number of rows available in the buffer. 112 int num_rows_; 113 114 // The next row index we should write into. This wraps around as the 115 // circular buffer is used. 116 int next_row_; 117 118 // The y coordinate of the |next_row_|. This is incremented each time a 119 // new row is appended and does not wrap. 120 int next_row_coordinate_; 121 122 // Buffer used by GetRowAddresses(). 123 std::vector<unsigned char*> row_addresses_; 124 }; 125 126 // Convolves horizontally along a single row. The row data is given in 127 // |src_data| and continues for the num_values() of the filter. 128 template<bool has_alpha> 129 void ConvolveHorizontally(const unsigned char* src_data, 130 const ConvolutionFilter1D& filter, 131 unsigned char* out_row) { 132 // Loop over each pixel on this row in the output image. 133 int num_values = filter.num_values(); 134 for (int out_x = 0; out_x < num_values; out_x++) { 135 // Get the filter that determines the current output pixel. 136 int filter_offset, filter_length; 137 const ConvolutionFilter1D::Fixed* filter_values = 138 filter.FilterForValue(out_x, &filter_offset, &filter_length); 139 140 // Compute the first pixel in this row that the filter affects. It will 141 // touch |filter_length| pixels (4 bytes each) after this. 142 const unsigned char* row_to_filter = &src_data[filter_offset * 4]; 143 144 // Apply the filter to the row to get the destination pixel in |accum|. 145 int accum[4] = {0}; 146 for (int filter_x = 0; filter_x < filter_length; filter_x++) { 147 ConvolutionFilter1D::Fixed cur_filter = filter_values[filter_x]; 148 accum[0] += cur_filter * row_to_filter[filter_x * 4 + 0]; 149 accum[1] += cur_filter * row_to_filter[filter_x * 4 + 1]; 150 accum[2] += cur_filter * row_to_filter[filter_x * 4 + 2]; 151 if (has_alpha) 152 accum[3] += cur_filter * row_to_filter[filter_x * 4 + 3]; 153 } 154 155 // Bring this value back in range. All of the filter scaling factors 156 // are in fixed point with kShiftBits bits of fractional part. 157 accum[0] >>= ConvolutionFilter1D::kShiftBits; 158 accum[1] >>= ConvolutionFilter1D::kShiftBits; 159 accum[2] >>= ConvolutionFilter1D::kShiftBits; 160 if (has_alpha) 161 accum[3] >>= ConvolutionFilter1D::kShiftBits; 162 163 // Store the new pixel. 164 out_row[out_x * 4 + 0] = ClampTo8(accum[0]); 165 out_row[out_x * 4 + 1] = ClampTo8(accum[1]); 166 out_row[out_x * 4 + 2] = ClampTo8(accum[2]); 167 if (has_alpha) 168 out_row[out_x * 4 + 3] = ClampTo8(accum[3]); 169 } 170 } 171 172 // Does vertical convolution to produce one output row. The filter values and 173 // length are given in the first two parameters. These are applied to each 174 // of the rows pointed to in the |source_data_rows| array, with each row 175 // being |pixel_width| wide. 176 // 177 // The output must have room for |pixel_width * 4| bytes. 178 template<bool has_alpha> 179 void ConvolveVertically(const ConvolutionFilter1D::Fixed* filter_values, 180 int filter_length, 181 unsigned char* const* source_data_rows, 182 int pixel_width, 183 unsigned char* out_row) { 184 // We go through each column in the output and do a vertical convolution, 185 // generating one output pixel each time. 186 for (int out_x = 0; out_x < pixel_width; out_x++) { 187 // Compute the number of bytes over in each row that the current column 188 // we're convolving starts at. The pixel will cover the next 4 bytes. 189 int byte_offset = out_x * 4; 190 191 // Apply the filter to one column of pixels. 192 int accum[4] = {0}; 193 for (int filter_y = 0; filter_y < filter_length; filter_y++) { 194 ConvolutionFilter1D::Fixed cur_filter = filter_values[filter_y]; 195 accum[0] += cur_filter * source_data_rows[filter_y][byte_offset + 0]; 196 accum[1] += cur_filter * source_data_rows[filter_y][byte_offset + 1]; 197 accum[2] += cur_filter * source_data_rows[filter_y][byte_offset + 2]; 198 if (has_alpha) 199 accum[3] += cur_filter * source_data_rows[filter_y][byte_offset + 3]; 200 } 201 202 // Bring this value back in range. All of the filter scaling factors 203 // are in fixed point with kShiftBits bits of precision. 204 accum[0] >>= ConvolutionFilter1D::kShiftBits; 205 accum[1] >>= ConvolutionFilter1D::kShiftBits; 206 accum[2] >>= ConvolutionFilter1D::kShiftBits; 207 if (has_alpha) 208 accum[3] >>= ConvolutionFilter1D::kShiftBits; 209 210 // Store the new pixel. 211 out_row[byte_offset + 0] = ClampTo8(accum[0]); 212 out_row[byte_offset + 1] = ClampTo8(accum[1]); 213 out_row[byte_offset + 2] = ClampTo8(accum[2]); 214 if (has_alpha) { 215 unsigned char alpha = ClampTo8(accum[3]); 216 217 // Make sure the alpha channel doesn't come out smaller than any of the 218 // color channels. We use premultipled alpha channels, so this should 219 // never happen, but rounding errors will cause this from time to time. 220 // These "impossible" colors will cause overflows (and hence random pixel 221 // values) when the resulting bitmap is drawn to the screen. 222 // 223 // We only need to do this when generating the final output row (here). 224 int max_color_channel = std::max(out_row[byte_offset + 0], 225 std::max(out_row[byte_offset + 1], out_row[byte_offset + 2])); 226 if (alpha < max_color_channel) 227 out_row[byte_offset + 3] = max_color_channel; 228 else 229 out_row[byte_offset + 3] = alpha; 230 } else { 231 // No alpha channel, the image is opaque. 232 out_row[byte_offset + 3] = 0xff; 233 } 234 } 235 } 236 237 void ConvolveVertically(const ConvolutionFilter1D::Fixed* filter_values, 238 int filter_length, 239 unsigned char* const* source_data_rows, 240 int pixel_width, 241 unsigned char* out_row, 242 bool source_has_alpha) { 243 if (source_has_alpha) { 244 ConvolveVertically<true>(filter_values, filter_length, 245 source_data_rows, 246 pixel_width, 247 out_row); 248 } else { 249 ConvolveVertically<false>(filter_values, filter_length, 250 source_data_rows, 251 pixel_width, 252 out_row); 253 } 254 } 255 256 } // namespace 257 258 // ConvolutionFilter1D --------------------------------------------------------- 259 260 ConvolutionFilter1D::ConvolutionFilter1D() 261 : max_filter_(0) { 262 } 263 264 ConvolutionFilter1D::~ConvolutionFilter1D() { 265 } 266 267 void ConvolutionFilter1D::AddFilter(int filter_offset, 268 const float* filter_values, 269 int filter_length) { 270 SkASSERT(filter_length > 0); 271 272 std::vector<Fixed> fixed_values; 273 fixed_values.reserve(filter_length); 274 275 for (int i = 0; i < filter_length; ++i) 276 fixed_values.push_back(FloatToFixed(filter_values[i])); 277 278 AddFilter(filter_offset, &fixed_values[0], filter_length); 279 } 280 281 void ConvolutionFilter1D::AddFilter(int filter_offset, 282 const Fixed* filter_values, 283 int filter_length) { 284 // It is common for leading/trailing filter values to be zeros. In such 285 // cases it is beneficial to only store the central factors. 286 // For a scaling to 1/4th in each dimension using a Lanczos-2 filter on 287 // a 1080p image this optimization gives a ~10% speed improvement. 288 int filter_size = filter_length; 289 int first_non_zero = 0; 290 while (first_non_zero < filter_length && filter_values[first_non_zero] == 0) 291 first_non_zero++; 292 293 if (first_non_zero < filter_length) { 294 // Here we have at least one non-zero factor. 295 int last_non_zero = filter_length - 1; 296 while (last_non_zero >= 0 && filter_values[last_non_zero] == 0) 297 last_non_zero--; 298 299 filter_offset += first_non_zero; 300 filter_length = last_non_zero + 1 - first_non_zero; 301 SkASSERT(filter_length > 0); 302 303 for (int i = first_non_zero; i <= last_non_zero; i++) 304 filter_values_.push_back(filter_values[i]); 305 } else { 306 // Here all the factors were zeroes. 307 filter_length = 0; 308 } 309 310 FilterInstance instance; 311 312 // We pushed filter_length elements onto filter_values_ 313 instance.data_location = (static_cast<int>(filter_values_.size()) - 314 filter_length); 315 instance.offset = filter_offset; 316 instance.trimmed_length = filter_length; 317 instance.length = filter_size; 318 filters_.push_back(instance); 319 320 max_filter_ = std::max(max_filter_, filter_length); 321 } 322 323 const ConvolutionFilter1D::Fixed* ConvolutionFilter1D::GetSingleFilter( 324 int* specified_filter_length, 325 int* filter_offset, 326 int* filter_length) const { 327 const FilterInstance& filter = filters_[0]; 328 *filter_offset = filter.offset; 329 *filter_length = filter.trimmed_length; 330 *specified_filter_length = filter.length; 331 if (filter.trimmed_length == 0) 332 return NULL; 333 334 return &filter_values_[filter.data_location]; 335 } 336 337 typedef void (*ConvolveVertically_pointer)( 338 const ConvolutionFilter1D::Fixed* filter_values, 339 int filter_length, 340 unsigned char* const* source_data_rows, 341 int pixel_width, 342 unsigned char* out_row, 343 bool has_alpha); 344 typedef void (*Convolve4RowsHorizontally_pointer)( 345 const unsigned char* src_data[4], 346 const ConvolutionFilter1D& filter, 347 unsigned char* out_row[4]); 348 typedef void (*ConvolveHorizontally_pointer)( 349 const unsigned char* src_data, 350 const ConvolutionFilter1D& filter, 351 unsigned char* out_row, 352 bool has_alpha); 353 354 struct ConvolveProcs { 355 // This is how many extra pixels may be read by the 356 // conolve*horizontally functions. 357 int extra_horizontal_reads; 358 ConvolveVertically_pointer convolve_vertically; 359 Convolve4RowsHorizontally_pointer convolve_4rows_horizontally; 360 ConvolveHorizontally_pointer convolve_horizontally; 361 }; 362 363 void SetupSIMD(ConvolveProcs *procs) { 364 #ifdef SIMD_SSE2 365 procs->extra_horizontal_reads = 3; 366 procs->convolve_vertically = &ConvolveVertically_SSE2; 367 procs->convolve_4rows_horizontally = &Convolve4RowsHorizontally_SSE2; 368 procs->convolve_horizontally = &ConvolveHorizontally_SSE2; 369 #elif defined SIMD_MIPS_DSPR2 370 procs->extra_horizontal_reads = 3; 371 procs->convolve_vertically = &ConvolveVertically_mips_dspr2; 372 procs->convolve_horizontally = &ConvolveHorizontally_mips_dspr2; 373 #endif 374 } 375 376 void BGRAConvolve2D(const unsigned char* source_data, 377 int source_byte_row_stride, 378 bool source_has_alpha, 379 const ConvolutionFilter1D& filter_x, 380 const ConvolutionFilter1D& filter_y, 381 int output_byte_row_stride, 382 unsigned char* output, 383 bool use_simd_if_possible) { 384 ConvolveProcs simd; 385 simd.extra_horizontal_reads = 0; 386 simd.convolve_vertically = NULL; 387 simd.convolve_4rows_horizontally = NULL; 388 simd.convolve_horizontally = NULL; 389 if (use_simd_if_possible) { 390 SetupSIMD(&simd); 391 } 392 393 int max_y_filter_size = filter_y.max_filter(); 394 395 // The next row in the input that we will generate a horizontally 396 // convolved row for. If the filter doesn't start at the beginning of the 397 // image (this is the case when we are only resizing a subset), then we 398 // don't want to generate any output rows before that. Compute the starting 399 // row for convolution as the first pixel for the first vertical filter. 400 int filter_offset, filter_length; 401 const ConvolutionFilter1D::Fixed* filter_values = 402 filter_y.FilterForValue(0, &filter_offset, &filter_length); 403 int next_x_row = filter_offset; 404 405 // We loop over each row in the input doing a horizontal convolution. This 406 // will result in a horizontally convolved image. We write the results into 407 // a circular buffer of convolved rows and do vertical convolution as rows 408 // are available. This prevents us from having to store the entire 409 // intermediate image and helps cache coherency. 410 // We will need four extra rows to allow horizontal convolution could be done 411 // simultaneously. We also padding each row in row buffer to be aligned-up to 412 // 16 bytes. 413 // TODO(jiesun): We do not use aligned load from row buffer in vertical 414 // convolution pass yet. Somehow Windows does not like it. 415 int row_buffer_width = (filter_x.num_values() + 15) & ~0xF; 416 int row_buffer_height = max_y_filter_size + 417 (simd.convolve_4rows_horizontally ? 4 : 0); 418 CircularRowBuffer row_buffer(row_buffer_width, 419 row_buffer_height, 420 filter_offset); 421 422 // Loop over every possible output row, processing just enough horizontal 423 // convolutions to run each subsequent vertical convolution. 424 SkASSERT(output_byte_row_stride >= filter_x.num_values() * 4); 425 int num_output_rows = filter_y.num_values(); 426 427 // We need to check which is the last line to convolve before we advance 4 428 // lines in one iteration. 429 int last_filter_offset, last_filter_length; 430 431 // SSE2 can access up to 3 extra pixels past the end of the 432 // buffer. At the bottom of the image, we have to be careful 433 // not to access data past the end of the buffer. Normally 434 // we fall back to the C++ implementation for the last row. 435 // If the last row is less than 3 pixels wide, we may have to fall 436 // back to the C++ version for more rows. Compute how many 437 // rows we need to avoid the SSE implementation for here. 438 filter_x.FilterForValue(filter_x.num_values() - 1, &last_filter_offset, 439 &last_filter_length); 440 int avoid_simd_rows = 1 + simd.extra_horizontal_reads / 441 (last_filter_offset + last_filter_length); 442 443 filter_y.FilterForValue(num_output_rows - 1, &last_filter_offset, 444 &last_filter_length); 445 446 for (int out_y = 0; out_y < num_output_rows; out_y++) { 447 filter_values = filter_y.FilterForValue(out_y, 448 &filter_offset, &filter_length); 449 450 // Generate output rows until we have enough to run the current filter. 451 while (next_x_row < filter_offset + filter_length) { 452 if (simd.convolve_4rows_horizontally && 453 next_x_row + 3 < last_filter_offset + last_filter_length - 454 avoid_simd_rows) { 455 const unsigned char* src[4]; 456 unsigned char* out_row[4]; 457 for (int i = 0; i < 4; ++i) { 458 src[i] = &source_data[(next_x_row + i) * source_byte_row_stride]; 459 out_row[i] = row_buffer.AdvanceRow(); 460 } 461 simd.convolve_4rows_horizontally(src, filter_x, out_row); 462 next_x_row += 4; 463 } else { 464 // Check if we need to avoid SSE2 for this row. 465 if (simd.convolve_horizontally && 466 next_x_row < last_filter_offset + last_filter_length - 467 avoid_simd_rows) { 468 simd.convolve_horizontally( 469 &source_data[next_x_row * source_byte_row_stride], 470 filter_x, row_buffer.AdvanceRow(), source_has_alpha); 471 } else { 472 if (source_has_alpha) { 473 ConvolveHorizontally<true>( 474 &source_data[next_x_row * source_byte_row_stride], 475 filter_x, row_buffer.AdvanceRow()); 476 } else { 477 ConvolveHorizontally<false>( 478 &source_data[next_x_row * source_byte_row_stride], 479 filter_x, row_buffer.AdvanceRow()); 480 } 481 } 482 next_x_row++; 483 } 484 } 485 486 // Compute where in the output image this row of final data will go. 487 unsigned char* cur_output_row = &output[out_y * output_byte_row_stride]; 488 489 // Get the list of rows that the circular buffer has, in order. 490 int first_row_in_circular_buffer; 491 unsigned char* const* rows_to_convolve = 492 row_buffer.GetRowAddresses(&first_row_in_circular_buffer); 493 494 // Now compute the start of the subset of those rows that the filter 495 // needs. 496 unsigned char* const* first_row_for_filter = 497 &rows_to_convolve[filter_offset - first_row_in_circular_buffer]; 498 499 if (simd.convolve_vertically) { 500 simd.convolve_vertically(filter_values, filter_length, 501 first_row_for_filter, 502 filter_x.num_values(), cur_output_row, 503 source_has_alpha); 504 } else { 505 ConvolveVertically(filter_values, filter_length, 506 first_row_for_filter, 507 filter_x.num_values(), cur_output_row, 508 source_has_alpha); 509 } 510 } 511 } 512 513 void SingleChannelConvolveX1D(const unsigned char* source_data, 514 int source_byte_row_stride, 515 int input_channel_index, 516 int input_channel_count, 517 const ConvolutionFilter1D& filter, 518 const SkISize& image_size, 519 unsigned char* output, 520 int output_byte_row_stride, 521 int output_channel_index, 522 int output_channel_count, 523 bool absolute_values) { 524 int filter_offset, filter_length, filter_size; 525 // Very much unlike BGRAConvolve2D, here we expect to have the same filter 526 // for all pixels. 527 const ConvolutionFilter1D::Fixed* filter_values = 528 filter.GetSingleFilter(&filter_size, &filter_offset, &filter_length); 529 530 if (filter_values == NULL || image_size.width() < filter_size) { 531 NOTREACHED(); 532 return; 533 } 534 535 int centrepoint = filter_length / 2; 536 if (filter_size - filter_offset != 2 * filter_offset) { 537 // This means the original filter was not symmetrical AND 538 // got clipped from one side more than from the other. 539 centrepoint = filter_size / 2 - filter_offset; 540 } 541 542 const unsigned char* source_data_row = source_data; 543 unsigned char* output_row = output; 544 545 for (int r = 0; r < image_size.height(); ++r) { 546 unsigned char* target_byte = output_row + output_channel_index; 547 // Process the lead part, padding image to the left with the first pixel. 548 int c = 0; 549 for (; c < centrepoint; ++c, target_byte += output_channel_count) { 550 int accval = 0; 551 int i = 0; 552 int pixel_byte_index = input_channel_index; 553 for (; i < centrepoint - c; ++i) // Padding part. 554 accval += filter_values[i] * source_data_row[pixel_byte_index]; 555 556 for (; i < filter_length; ++i, pixel_byte_index += input_channel_count) 557 accval += filter_values[i] * source_data_row[pixel_byte_index]; 558 559 *target_byte = BringBackTo8(accval, absolute_values); 560 } 561 562 // Now for the main event. 563 for (; c < image_size.width() - centrepoint; 564 ++c, target_byte += output_channel_count) { 565 int accval = 0; 566 int pixel_byte_index = (c - centrepoint) * input_channel_count + 567 input_channel_index; 568 569 for (int i = 0; i < filter_length; 570 ++i, pixel_byte_index += input_channel_count) { 571 accval += filter_values[i] * source_data_row[pixel_byte_index]; 572 } 573 574 *target_byte = BringBackTo8(accval, absolute_values); 575 } 576 577 for (; c < image_size.width(); ++c, target_byte += output_channel_count) { 578 int accval = 0; 579 int overlap_taps = image_size.width() - c + centrepoint; 580 int pixel_byte_index = (c - centrepoint) * input_channel_count + 581 input_channel_index; 582 int i = 0; 583 for (; i < overlap_taps - 1; ++i, pixel_byte_index += input_channel_count) 584 accval += filter_values[i] * source_data_row[pixel_byte_index]; 585 586 for (; i < filter_length; ++i) 587 accval += filter_values[i] * source_data_row[pixel_byte_index]; 588 589 *target_byte = BringBackTo8(accval, absolute_values); 590 } 591 592 source_data_row += source_byte_row_stride; 593 output_row += output_byte_row_stride; 594 } 595 } 596 597 void SingleChannelConvolveY1D(const unsigned char* source_data, 598 int source_byte_row_stride, 599 int input_channel_index, 600 int input_channel_count, 601 const ConvolutionFilter1D& filter, 602 const SkISize& image_size, 603 unsigned char* output, 604 int output_byte_row_stride, 605 int output_channel_index, 606 int output_channel_count, 607 bool absolute_values) { 608 int filter_offset, filter_length, filter_size; 609 // Very much unlike BGRAConvolve2D, here we expect to have the same filter 610 // for all pixels. 611 const ConvolutionFilter1D::Fixed* filter_values = 612 filter.GetSingleFilter(&filter_size, &filter_offset, &filter_length); 613 614 if (filter_values == NULL || image_size.height() < filter_size) { 615 NOTREACHED(); 616 return; 617 } 618 619 int centrepoint = filter_length / 2; 620 if (filter_size - filter_offset != 2 * filter_offset) { 621 // This means the original filter was not symmetrical AND 622 // got clipped from one side more than from the other. 623 centrepoint = filter_size / 2 - filter_offset; 624 } 625 626 for (int c = 0; c < image_size.width(); ++c) { 627 unsigned char* target_byte = output + c * output_channel_count + 628 output_channel_index; 629 int r = 0; 630 631 for (; r < centrepoint; ++r, target_byte += output_byte_row_stride) { 632 int accval = 0; 633 int i = 0; 634 int pixel_byte_index = c * input_channel_count + input_channel_index; 635 636 for (; i < centrepoint - r; ++i) // Padding part. 637 accval += filter_values[i] * source_data[pixel_byte_index]; 638 639 for (; i < filter_length; ++i, pixel_byte_index += source_byte_row_stride) 640 accval += filter_values[i] * source_data[pixel_byte_index]; 641 642 *target_byte = BringBackTo8(accval, absolute_values); 643 } 644 645 for (; r < image_size.height() - centrepoint; 646 ++r, target_byte += output_byte_row_stride) { 647 int accval = 0; 648 int pixel_byte_index = (r - centrepoint) * source_byte_row_stride + 649 c * input_channel_count + input_channel_index; 650 for (int i = 0; i < filter_length; 651 ++i, pixel_byte_index += source_byte_row_stride) { 652 accval += filter_values[i] * source_data[pixel_byte_index]; 653 } 654 655 *target_byte = BringBackTo8(accval, absolute_values); 656 } 657 658 for (; r < image_size.height(); 659 ++r, target_byte += output_byte_row_stride) { 660 int accval = 0; 661 int overlap_taps = image_size.height() - r + centrepoint; 662 int pixel_byte_index = (r - centrepoint) * source_byte_row_stride + 663 c * input_channel_count + input_channel_index; 664 int i = 0; 665 for (; i < overlap_taps - 1; 666 ++i, pixel_byte_index += source_byte_row_stride) { 667 accval += filter_values[i] * source_data[pixel_byte_index]; 668 } 669 670 for (; i < filter_length; ++i) 671 accval += filter_values[i] * source_data[pixel_byte_index]; 672 673 *target_byte = BringBackTo8(accval, absolute_values); 674 } 675 } 676 } 677 678 void SetUpGaussianConvolutionKernel(ConvolutionFilter1D* filter, 679 float kernel_sigma, 680 bool derivative) { 681 DCHECK(filter != NULL); 682 DCHECK_GT(kernel_sigma, 0.0); 683 const int tail_length = static_cast<int>(4.0f * kernel_sigma + 0.5f); 684 const int kernel_size = tail_length * 2 + 1; 685 const float sigmasq = kernel_sigma * kernel_sigma; 686 std::vector<float> kernel_weights(kernel_size, 0.0); 687 float kernel_sum = 1.0f; 688 689 kernel_weights[tail_length] = 1.0f; 690 691 for (int ii = 1; ii <= tail_length; ++ii) { 692 float v = std::exp(-0.5f * ii * ii / sigmasq); 693 kernel_weights[tail_length + ii] = v; 694 kernel_weights[tail_length - ii] = v; 695 kernel_sum += 2.0f * v; 696 } 697 698 for (int i = 0; i < kernel_size; ++i) 699 kernel_weights[i] /= kernel_sum; 700 701 if (derivative) { 702 kernel_weights[tail_length] = 0.0; 703 for (int ii = 1; ii <= tail_length; ++ii) { 704 float v = sigmasq * kernel_weights[tail_length + ii] / ii; 705 kernel_weights[tail_length + ii] = v; 706 kernel_weights[tail_length - ii] = -v; 707 } 708 } 709 710 filter->AddFilter(0, &kernel_weights[0], kernel_weights.size()); 711 } 712 713 } // namespace skia 714