Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(233)

Side by Side Diff: src/core/SkConvolver.cpp

Issue 2481733003: Make SSE2/Neon convolution functions not to read extra bytes (Closed)
Patch Set: Change macros to functions Created 4 years, 1 month ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View unified diff | Download patch
« no previous file with comments | « src/core/SkConvolver.h ('k') | src/opts/SkBitmapFilter_opts_SSE2.cpp » ('j') | no next file with comments »
Toggle Intra-line Diffs ('i') | Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
OLDNEW
1 // Copyright (c) 2011 The Chromium Authors. All rights reserved. 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 2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file. 3 // found in the LICENSE file.
4 4
5 #include "SkConvolver.h" 5 #include "SkConvolver.h"
6 #include "SkTArray.h" 6 #include "SkTArray.h"
7 7
8 namespace { 8 namespace {
9 9
10 // Converts the argument to an 8-bit unsigned value by clamping to the range 10 // Converts the argument to an 8-bit unsigned value by clamping to the range
(...skipping 383 matching lines...) Expand 10 before | Expand all | Expand 10 after
394 filterOffset); 394 filterOffset);
395 395
396 // Loop over every possible output row, processing just enough horizontal 396 // Loop over every possible output row, processing just enough horizontal
397 // convolutions to run each subsequent vertical convolution. 397 // convolutions to run each subsequent vertical convolution.
398 SkASSERT(outputByteRowStride >= filterX.numValues() * 4); 398 SkASSERT(outputByteRowStride >= filterX.numValues() * 4);
399 int numOutputRows = filterY.numValues(); 399 int numOutputRows = filterY.numValues();
400 400
401 // We need to check which is the last line to convolve before we advance 4 401 // We need to check which is the last line to convolve before we advance 4
402 // lines in one iteration. 402 // lines in one iteration.
403 int lastFilterOffset, lastFilterLength; 403 int lastFilterOffset, lastFilterLength;
404
405 // SSE2 can access up to 3 extra pixels past the end of the
406 // buffer. At the bottom of the image, we have to be careful
407 // not to access data past the end of the buffer. Normally
408 // we fall back to the C++ implementation for the last row.
409 // If the last row is less than 3 pixels wide, we may have to fall
410 // back to the C++ version for more rows. Compute how many
411 // rows we need to avoid the SSE implementation for here.
412 filterX.FilterForValue(filterX.numValues() - 1, &lastFilterOffset,
413 &lastFilterLength);
414 int avoidSimdRows = 1 + convolveProcs.fExtraHorizontalReads /
415 (lastFilterOffset + lastFilterLength);
416
417 filterY.FilterForValue(numOutputRows - 1, &lastFilterOffset, 404 filterY.FilterForValue(numOutputRows - 1, &lastFilterOffset,
418 &lastFilterLength); 405 &lastFilterLength);
419 406
420 for (int outY = 0; outY < numOutputRows; outY++) { 407 for (int outY = 0; outY < numOutputRows; outY++) {
421 filterValues = filterY.FilterForValue(outY, 408 filterValues = filterY.FilterForValue(outY,
422 &filterOffset, &filterLength); 409 &filterOffset, &filterLength);
423 410
424 // Generate output rows until we have enough to run the current filter. 411 // Generate output rows until we have enough to run the current filter.
425 while (nextXRow < filterOffset + filterLength) { 412 while (nextXRow < filterOffset + filterLength) {
426 if (convolveProcs.fConvolve4RowsHorizontally && 413 if (convolveProcs.fConvolve4RowsHorizontally &&
427 nextXRow + 3 < lastFilterOffset + lastFilterLength - 414 nextXRow + 3 < lastFilterOffset + lastFilterLength) {
428 avoidSimdRows) {
429 const unsigned char* src[4]; 415 const unsigned char* src[4];
430 unsigned char* outRow[4]; 416 unsigned char* outRow[4];
431 for (int i = 0; i < 4; ++i) { 417 for (int i = 0; i < 4; ++i) {
432 src[i] = &sourceData[(uint64_t)(nextXRow + i) * sourceByteRo wStride]; 418 src[i] = &sourceData[(uint64_t)(nextXRow + i) * sourceByteRo wStride];
433 outRow[i] = rowBuffer.advanceRow(); 419 outRow[i] = rowBuffer.advanceRow();
434 } 420 }
435 convolveProcs.fConvolve4RowsHorizontally(src, filterX, outRow, 4 *rowBufferWidth); 421 convolveProcs.fConvolve4RowsHorizontally(src, filterX, outRow, 4 *rowBufferWidth);
436 nextXRow += 4; 422 nextXRow += 4;
437 } else { 423 } else {
438 // Check if we need to avoid SSE2 for this row. 424 if (convolveProcs.fConvolveHorizontally) {
439 if (convolveProcs.fConvolveHorizontally &&
440 nextXRow < lastFilterOffset + lastFilterLength -
441 avoidSimdRows) {
442 convolveProcs.fConvolveHorizontally( 425 convolveProcs.fConvolveHorizontally(
443 &sourceData[(uint64_t)nextXRow * sourceByteRowStride], 426 &sourceData[(uint64_t)nextXRow * sourceByteRowStride],
444 filterX, rowBuffer.advanceRow(), sourceHasAlpha); 427 filterX, rowBuffer.advanceRow(), sourceHasAlpha);
445 } else { 428 } else {
446 if (sourceHasAlpha) { 429 if (sourceHasAlpha) {
447 ConvolveHorizontallyAlpha( 430 ConvolveHorizontallyAlpha(
448 &sourceData[(uint64_t)nextXRow * sourceByteRowStride ], 431 &sourceData[(uint64_t)nextXRow * sourceByteRowStride ],
449 filterX, rowBuffer.advanceRow()); 432 filterX, rowBuffer.advanceRow());
450 } else { 433 } else {
451 ConvolveHorizontallyNoAlpha( 434 ConvolveHorizontallyNoAlpha(
(...skipping 25 matching lines...) Expand all
477 sourceHasAlpha); 460 sourceHasAlpha);
478 } else { 461 } else {
479 ConvolveVertically(filterValues, filterLength, 462 ConvolveVertically(filterValues, filterLength,
480 firstRowForFilter, 463 firstRowForFilter,
481 filterX.numValues(), curOutputRow, 464 filterX.numValues(), curOutputRow,
482 sourceHasAlpha); 465 sourceHasAlpha);
483 } 466 }
484 } 467 }
485 return true; 468 return true;
486 } 469 }
OLDNEW
« no previous file with comments | « src/core/SkConvolver.h ('k') | src/opts/SkBitmapFilter_opts_SSE2.cpp » ('j') | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698