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 "SkConvolver.h"
6 #include "SkTArray.h"
7
8 namespace {
9
10 // Converts the argument to an 8-bit unsigned value by clamping to the range
11 // 0-255.
ClampTo8(int a)12 inline unsigned char ClampTo8(int a) {
13 if (static_cast<unsigned>(a) < 256) {
14 return a; // Avoid the extra check in the common case.
15 }
16 if (a < 0) {
17 return 0;
18 }
19 return 255;
20 }
21
22 // Stores a list of rows in a circular buffer. The usage is you write into it
23 // by calling AdvanceRow. It will keep track of which row in the buffer it
24 // should use next, and the total number of rows added.
25 class CircularRowBuffer {
26 public:
27 // The number of pixels in each row is given in |sourceRowPixelWidth|.
28 // The maximum number of rows needed in the buffer is |maxYFilterSize|
29 // (we only need to store enough rows for the biggest filter).
30 //
31 // We use the |firstInputRow| to compute the coordinates of all of the
32 // following rows returned by Advance().
CircularRowBuffer(int destRowPixelWidth,int maxYFilterSize,int firstInputRow)33 CircularRowBuffer(int destRowPixelWidth, int maxYFilterSize,
34 int firstInputRow)
35 : fRowByteWidth(destRowPixelWidth * 4),
36 fNumRows(maxYFilterSize),
37 fNextRow(0),
38 fNextRowCoordinate(firstInputRow) {
39 fBuffer.reset(fRowByteWidth * maxYFilterSize);
40 fRowAddresses.reset(fNumRows);
41 }
42
43 // Moves to the next row in the buffer, returning a pointer to the beginning
44 // of it.
advanceRow()45 unsigned char* advanceRow() {
46 unsigned char* row = &fBuffer[fNextRow * fRowByteWidth];
47 fNextRowCoordinate++;
48
49 // Set the pointer to the next row to use, wrapping around if necessary.
50 fNextRow++;
51 if (fNextRow == fNumRows) {
52 fNextRow = 0;
53 }
54 return row;
55 }
56
57 // Returns a pointer to an "unrolled" array of rows. These rows will start
58 // at the y coordinate placed into |*firstRowIndex| and will continue in
59 // order for the maximum number of rows in this circular buffer.
60 //
61 // The |firstRowIndex_| may be negative. This means the circular buffer
62 // starts before the top of the image (it hasn't been filled yet).
GetRowAddresses(int * firstRowIndex)63 unsigned char* const* GetRowAddresses(int* firstRowIndex) {
64 // Example for a 4-element circular buffer holding coords 6-9.
65 // Row 0 Coord 8
66 // Row 1 Coord 9
67 // Row 2 Coord 6 <- fNextRow = 2, fNextRowCoordinate = 10.
68 // Row 3 Coord 7
69 //
70 // The "next" row is also the first (lowest) coordinate. This computation
71 // may yield a negative value, but that's OK, the math will work out
72 // since the user of this buffer will compute the offset relative
73 // to the firstRowIndex and the negative rows will never be used.
74 *firstRowIndex = fNextRowCoordinate - fNumRows;
75
76 int curRow = fNextRow;
77 for (int i = 0; i < fNumRows; i++) {
78 fRowAddresses[i] = &fBuffer[curRow * fRowByteWidth];
79
80 // Advance to the next row, wrapping if necessary.
81 curRow++;
82 if (curRow == fNumRows) {
83 curRow = 0;
84 }
85 }
86 return &fRowAddresses[0];
87 }
88
89 private:
90 // The buffer storing the rows. They are packed, each one fRowByteWidth.
91 SkTArray<unsigned char> fBuffer;
92
93 // Number of bytes per row in the |buffer|.
94 int fRowByteWidth;
95
96 // The number of rows available in the buffer.
97 int fNumRows;
98
99 // The next row index we should write into. This wraps around as the
100 // circular buffer is used.
101 int fNextRow;
102
103 // The y coordinate of the |fNextRow|. This is incremented each time a
104 // new row is appended and does not wrap.
105 int fNextRowCoordinate;
106
107 // Buffer used by GetRowAddresses().
108 SkTArray<unsigned char*> fRowAddresses;
109 };
110
111 // Convolves horizontally along a single row. The row data is given in
112 // |srcData| and continues for the numValues() of the filter.
113 template<bool hasAlpha>
ConvolveHorizontally(const unsigned char * srcData,const SkConvolutionFilter1D & filter,unsigned char * outRow)114 void ConvolveHorizontally(const unsigned char* srcData,
115 const SkConvolutionFilter1D& filter,
116 unsigned char* outRow) {
117 // Loop over each pixel on this row in the output image.
118 int numValues = filter.numValues();
119 for (int outX = 0; outX < numValues; outX++) {
120 // Get the filter that determines the current output pixel.
121 int filterOffset, filterLength;
122 const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
123 filter.FilterForValue(outX, &filterOffset, &filterLength);
124
125 // Compute the first pixel in this row that the filter affects. It will
126 // touch |filterLength| pixels (4 bytes each) after this.
127 const unsigned char* rowToFilter = &srcData[filterOffset * 4];
128
129 // Apply the filter to the row to get the destination pixel in |accum|.
130 int accum[4] = {0};
131 for (int filterX = 0; filterX < filterLength; filterX++) {
132 SkConvolutionFilter1D::ConvolutionFixed curFilter = filterValues[filterX];
133 accum[0] += curFilter * rowToFilter[filterX * 4 + 0];
134 accum[1] += curFilter * rowToFilter[filterX * 4 + 1];
135 accum[2] += curFilter * rowToFilter[filterX * 4 + 2];
136 if (hasAlpha) {
137 accum[3] += curFilter * rowToFilter[filterX * 4 + 3];
138 }
139 }
140
141 // Bring this value back in range. All of the filter scaling factors
142 // are in fixed point with kShiftBits bits of fractional part.
143 accum[0] >>= SkConvolutionFilter1D::kShiftBits;
144 accum[1] >>= SkConvolutionFilter1D::kShiftBits;
145 accum[2] >>= SkConvolutionFilter1D::kShiftBits;
146 if (hasAlpha) {
147 accum[3] >>= SkConvolutionFilter1D::kShiftBits;
148 }
149
150 // Store the new pixel.
151 outRow[outX * 4 + 0] = ClampTo8(accum[0]);
152 outRow[outX * 4 + 1] = ClampTo8(accum[1]);
153 outRow[outX * 4 + 2] = ClampTo8(accum[2]);
154 if (hasAlpha) {
155 outRow[outX * 4 + 3] = ClampTo8(accum[3]);
156 }
157 }
158 }
159
160 // There's a bug somewhere here with GCC autovectorization (-ftree-vectorize). We originally
161 // thought this was 32 bit only, but subsequent tests show that some 64 bit gcc compiles
162 // suffer here too.
163 //
164 // Dropping to -O2 disables -ftree-vectorize. GCC 4.6 needs noinline. https://bug.skia.org/2575
165 #if SK_HAS_ATTRIBUTE(optimize) && defined(SK_RELEASE)
166 #define SK_MAYBE_DISABLE_VECTORIZATION __attribute__((optimize("O2"), noinline))
167 #else
168 #define SK_MAYBE_DISABLE_VECTORIZATION
169 #endif
170
171 SK_MAYBE_DISABLE_VECTORIZATION
ConvolveHorizontallyAlpha(const unsigned char * srcData,const SkConvolutionFilter1D & filter,unsigned char * outRow)172 static void ConvolveHorizontallyAlpha(const unsigned char* srcData,
173 const SkConvolutionFilter1D& filter,
174 unsigned char* outRow) {
175 return ConvolveHorizontally<true>(srcData, filter, outRow);
176 }
177
178 SK_MAYBE_DISABLE_VECTORIZATION
ConvolveHorizontallyNoAlpha(const unsigned char * srcData,const SkConvolutionFilter1D & filter,unsigned char * outRow)179 static void ConvolveHorizontallyNoAlpha(const unsigned char* srcData,
180 const SkConvolutionFilter1D& filter,
181 unsigned char* outRow) {
182 return ConvolveHorizontally<false>(srcData, filter, outRow);
183 }
184
185 #undef SK_MAYBE_DISABLE_VECTORIZATION
186
187
188 // Does vertical convolution to produce one output row. The filter values and
189 // length are given in the first two parameters. These are applied to each
190 // of the rows pointed to in the |sourceDataRows| array, with each row
191 // being |pixelWidth| wide.
192 //
193 // The output must have room for |pixelWidth * 4| bytes.
194 template<bool hasAlpha>
ConvolveVertically(const SkConvolutionFilter1D::ConvolutionFixed * filterValues,int filterLength,unsigned char * const * sourceDataRows,int pixelWidth,unsigned char * outRow)195 void ConvolveVertically(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
196 int filterLength,
197 unsigned char* const* sourceDataRows,
198 int pixelWidth,
199 unsigned char* outRow) {
200 // We go through each column in the output and do a vertical convolution,
201 // generating one output pixel each time.
202 for (int outX = 0; outX < pixelWidth; outX++) {
203 // Compute the number of bytes over in each row that the current column
204 // we're convolving starts at. The pixel will cover the next 4 bytes.
205 int byteOffset = outX * 4;
206
207 // Apply the filter to one column of pixels.
208 int accum[4] = {0};
209 for (int filterY = 0; filterY < filterLength; filterY++) {
210 SkConvolutionFilter1D::ConvolutionFixed curFilter = filterValues[filterY];
211 accum[0] += curFilter * sourceDataRows[filterY][byteOffset + 0];
212 accum[1] += curFilter * sourceDataRows[filterY][byteOffset + 1];
213 accum[2] += curFilter * sourceDataRows[filterY][byteOffset + 2];
214 if (hasAlpha) {
215 accum[3] += curFilter * sourceDataRows[filterY][byteOffset + 3];
216 }
217 }
218
219 // Bring this value back in range. All of the filter scaling factors
220 // are in fixed point with kShiftBits bits of precision.
221 accum[0] >>= SkConvolutionFilter1D::kShiftBits;
222 accum[1] >>= SkConvolutionFilter1D::kShiftBits;
223 accum[2] >>= SkConvolutionFilter1D::kShiftBits;
224 if (hasAlpha) {
225 accum[3] >>= SkConvolutionFilter1D::kShiftBits;
226 }
227
228 // Store the new pixel.
229 outRow[byteOffset + 0] = ClampTo8(accum[0]);
230 outRow[byteOffset + 1] = ClampTo8(accum[1]);
231 outRow[byteOffset + 2] = ClampTo8(accum[2]);
232 if (hasAlpha) {
233 unsigned char alpha = ClampTo8(accum[3]);
234
235 // Make sure the alpha channel doesn't come out smaller than any of the
236 // color channels. We use premultipled alpha channels, so this should
237 // never happen, but rounding errors will cause this from time to time.
238 // These "impossible" colors will cause overflows (and hence random pixel
239 // values) when the resulting bitmap is drawn to the screen.
240 //
241 // We only need to do this when generating the final output row (here).
242 int maxColorChannel = SkTMax(outRow[byteOffset + 0],
243 SkTMax(outRow[byteOffset + 1],
244 outRow[byteOffset + 2]));
245 if (alpha < maxColorChannel) {
246 outRow[byteOffset + 3] = maxColorChannel;
247 } else {
248 outRow[byteOffset + 3] = alpha;
249 }
250 } else {
251 // No alpha channel, the image is opaque.
252 outRow[byteOffset + 3] = 0xff;
253 }
254 }
255 }
256
ConvolveVertically(const SkConvolutionFilter1D::ConvolutionFixed * filterValues,int filterLength,unsigned char * const * sourceDataRows,int pixelWidth,unsigned char * outRow,bool sourceHasAlpha)257 void ConvolveVertically(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
258 int filterLength,
259 unsigned char* const* sourceDataRows,
260 int pixelWidth,
261 unsigned char* outRow,
262 bool sourceHasAlpha) {
263 if (sourceHasAlpha) {
264 ConvolveVertically<true>(filterValues, filterLength,
265 sourceDataRows, pixelWidth,
266 outRow);
267 } else {
268 ConvolveVertically<false>(filterValues, filterLength,
269 sourceDataRows, pixelWidth,
270 outRow);
271 }
272 }
273
274 } // namespace
275
276 // SkConvolutionFilter1D ---------------------------------------------------------
277
SkConvolutionFilter1D()278 SkConvolutionFilter1D::SkConvolutionFilter1D()
279 : fMaxFilter(0) {
280 }
281
~SkConvolutionFilter1D()282 SkConvolutionFilter1D::~SkConvolutionFilter1D() {
283 }
284
AddFilter(int filterOffset,const ConvolutionFixed * filterValues,int filterLength)285 void SkConvolutionFilter1D::AddFilter(int filterOffset,
286 const ConvolutionFixed* filterValues,
287 int filterLength) {
288 // It is common for leading/trailing filter values to be zeros. In such
289 // cases it is beneficial to only store the central factors.
290 // For a scaling to 1/4th in each dimension using a Lanczos-2 filter on
291 // a 1080p image this optimization gives a ~10% speed improvement.
292 int filterSize = filterLength;
293 int firstNonZero = 0;
294 while (firstNonZero < filterLength && filterValues[firstNonZero] == 0) {
295 firstNonZero++;
296 }
297
298 if (firstNonZero < filterLength) {
299 // Here we have at least one non-zero factor.
300 int lastNonZero = filterLength - 1;
301 while (lastNonZero >= 0 && filterValues[lastNonZero] == 0) {
302 lastNonZero--;
303 }
304
305 filterOffset += firstNonZero;
306 filterLength = lastNonZero + 1 - firstNonZero;
307 SkASSERT(filterLength > 0);
308
309 fFilterValues.append(filterLength, &filterValues[firstNonZero]);
310 } else {
311 // Here all the factors were zeroes.
312 filterLength = 0;
313 }
314
315 FilterInstance instance;
316
317 // We pushed filterLength elements onto fFilterValues
318 instance.fDataLocation = (static_cast<int>(fFilterValues.count()) -
319 filterLength);
320 instance.fOffset = filterOffset;
321 instance.fTrimmedLength = filterLength;
322 instance.fLength = filterSize;
323 fFilters.push(instance);
324
325 fMaxFilter = SkTMax(fMaxFilter, filterLength);
326 }
327
GetSingleFilter(int * specifiedFilterlength,int * filterOffset,int * filterLength) const328 const SkConvolutionFilter1D::ConvolutionFixed* SkConvolutionFilter1D::GetSingleFilter(
329 int* specifiedFilterlength,
330 int* filterOffset,
331 int* filterLength) const {
332 const FilterInstance& filter = fFilters[0];
333 *filterOffset = filter.fOffset;
334 *filterLength = filter.fTrimmedLength;
335 *specifiedFilterlength = filter.fLength;
336 if (filter.fTrimmedLength == 0) {
337 return nullptr;
338 }
339
340 return &fFilterValues[filter.fDataLocation];
341 }
342
BGRAConvolve2D(const unsigned char * sourceData,int sourceByteRowStride,bool sourceHasAlpha,const SkConvolutionFilter1D & filterX,const SkConvolutionFilter1D & filterY,int outputByteRowStride,unsigned char * output,const SkConvolutionProcs & convolveProcs,bool useSimdIfPossible)343 bool BGRAConvolve2D(const unsigned char* sourceData,
344 int sourceByteRowStride,
345 bool sourceHasAlpha,
346 const SkConvolutionFilter1D& filterX,
347 const SkConvolutionFilter1D& filterY,
348 int outputByteRowStride,
349 unsigned char* output,
350 const SkConvolutionProcs& convolveProcs,
351 bool useSimdIfPossible) {
352
353 int maxYFilterSize = filterY.maxFilter();
354
355 // The next row in the input that we will generate a horizontally
356 // convolved row for. If the filter doesn't start at the beginning of the
357 // image (this is the case when we are only resizing a subset), then we
358 // don't want to generate any output rows before that. Compute the starting
359 // row for convolution as the first pixel for the first vertical filter.
360 int filterOffset, filterLength;
361 const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
362 filterY.FilterForValue(0, &filterOffset, &filterLength);
363 int nextXRow = filterOffset;
364
365 // We loop over each row in the input doing a horizontal convolution. This
366 // will result in a horizontally convolved image. We write the results into
367 // a circular buffer of convolved rows and do vertical convolution as rows
368 // are available. This prevents us from having to store the entire
369 // intermediate image and helps cache coherency.
370 // We will need four extra rows to allow horizontal convolution could be done
371 // simultaneously. We also pad each row in row buffer to be aligned-up to
372 // 16 bytes.
373 // TODO(jiesun): We do not use aligned load from row buffer in vertical
374 // convolution pass yet. Somehow Windows does not like it.
375 int rowBufferWidth = (filterX.numValues() + 15) & ~0xF;
376 int rowBufferHeight = maxYFilterSize +
377 (convolveProcs.fConvolve4RowsHorizontally ? 4 : 0);
378
379 // check for too-big allocation requests : crbug.com/528628
380 {
381 int64_t size = sk_64_mul(rowBufferWidth, rowBufferHeight);
382 // need some limit, to avoid over-committing success from malloc, but then
383 // crashing when we try to actually use the memory.
384 // 100meg seems big enough to allow "normal" zoom factors and image sizes through
385 // while avoiding the crash seen by the bug (crbug.com/528628)
386 if (size > 100 * 1024 * 1024) {
387 // SkDebugf("BGRAConvolve2D: tmp allocation [%lld] too big\n", size);
388 return false;
389 }
390 }
391
392 CircularRowBuffer rowBuffer(rowBufferWidth,
393 rowBufferHeight,
394 filterOffset);
395
396 // Loop over every possible output row, processing just enough horizontal
397 // convolutions to run each subsequent vertical convolution.
398 SkASSERT(outputByteRowStride >= filterX.numValues() * 4);
399 int numOutputRows = filterY.numValues();
400
401 // We need to check which is the last line to convolve before we advance 4
402 // lines in one iteration.
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,
418 &lastFilterLength);
419
420 for (int outY = 0; outY < numOutputRows; outY++) {
421 filterValues = filterY.FilterForValue(outY,
422 &filterOffset, &filterLength);
423
424 // Generate output rows until we have enough to run the current filter.
425 while (nextXRow < filterOffset + filterLength) {
426 if (convolveProcs.fConvolve4RowsHorizontally &&
427 nextXRow + 3 < lastFilterOffset + lastFilterLength -
428 avoidSimdRows) {
429 const unsigned char* src[4];
430 unsigned char* outRow[4];
431 for (int i = 0; i < 4; ++i) {
432 src[i] = &sourceData[(uint64_t)(nextXRow + i) * sourceByteRowStride];
433 outRow[i] = rowBuffer.advanceRow();
434 }
435 convolveProcs.fConvolve4RowsHorizontally(src, filterX, outRow, 4*rowBufferWidth);
436 nextXRow += 4;
437 } else {
438 // Check if we need to avoid SSE2 for this row.
439 if (convolveProcs.fConvolveHorizontally &&
440 nextXRow < lastFilterOffset + lastFilterLength -
441 avoidSimdRows) {
442 convolveProcs.fConvolveHorizontally(
443 &sourceData[(uint64_t)nextXRow * sourceByteRowStride],
444 filterX, rowBuffer.advanceRow(), sourceHasAlpha);
445 } else {
446 if (sourceHasAlpha) {
447 ConvolveHorizontallyAlpha(
448 &sourceData[(uint64_t)nextXRow * sourceByteRowStride],
449 filterX, rowBuffer.advanceRow());
450 } else {
451 ConvolveHorizontallyNoAlpha(
452 &sourceData[(uint64_t)nextXRow * sourceByteRowStride],
453 filterX, rowBuffer.advanceRow());
454 }
455 }
456 nextXRow++;
457 }
458 }
459
460 // Compute where in the output image this row of final data will go.
461 unsigned char* curOutputRow = &output[(uint64_t)outY * outputByteRowStride];
462
463 // Get the list of rows that the circular buffer has, in order.
464 int firstRowInCircularBuffer;
465 unsigned char* const* rowsToConvolve =
466 rowBuffer.GetRowAddresses(&firstRowInCircularBuffer);
467
468 // Now compute the start of the subset of those rows that the filter
469 // needs.
470 unsigned char* const* firstRowForFilter =
471 &rowsToConvolve[filterOffset - firstRowInCircularBuffer];
472
473 if (convolveProcs.fConvolveVertically) {
474 convolveProcs.fConvolveVertically(filterValues, filterLength,
475 firstRowForFilter,
476 filterX.numValues(), curOutputRow,
477 sourceHasAlpha);
478 } else {
479 ConvolveVertically(filterValues, filterLength,
480 firstRowForFilter,
481 filterX.numValues(), curOutputRow,
482 sourceHasAlpha);
483 }
484 }
485 return true;
486 }
487