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.
5 #include "SkConvolver.h"
11 // Converts the argument to an 8-bit unsigned value by clamping to the range
13 inline unsigned char ClampTo8(int a) {
14 if (static_cast<unsigned>(a) < 256) {
15 return a; // Avoid the extra check in the common case.
23 // Stores a list of rows in a circular buffer. The usage is you write into it
24 // by calling AdvanceRow. It will keep track of which row in the buffer it
25 // should use next, and the total number of rows added.
26 class CircularRowBuffer {
28 // The number of pixels in each row is given in |sourceRowPixelWidth|.
29 // The maximum number of rows needed in the buffer is |maxYFilterSize|
30 // (we only need to store enough rows for the biggest filter).
32 // We use the |firstInputRow| to compute the coordinates of all of the
33 // following rows returned by Advance().
34 CircularRowBuffer(int destRowPixelWidth, int maxYFilterSize,
36 : fRowByteWidth(destRowPixelWidth * 4),
37 fNumRows(maxYFilterSize),
39 fNextRowCoordinate(firstInputRow) {
40 fBuffer.reset(fRowByteWidth * maxYFilterSize);
41 fRowAddresses.reset(fNumRows);
44 // Moves to the next row in the buffer, returning a pointer to the beginning
46 unsigned char* advanceRow() {
47 unsigned char* row = &fBuffer[fNextRow * fRowByteWidth];
50 // Set the pointer to the next row to use, wrapping around if necessary.
52 if (fNextRow == fNumRows) {
58 // Returns a pointer to an "unrolled" array of rows. These rows will start
59 // at the y coordinate placed into |*firstRowIndex| and will continue in
60 // order for the maximum number of rows in this circular buffer.
62 // The |firstRowIndex_| may be negative. This means the circular buffer
63 // starts before the top of the image (it hasn't been filled yet).
64 unsigned char* const* GetRowAddresses(int* firstRowIndex) {
65 // Example for a 4-element circular buffer holding coords 6-9.
68 // Row 2 Coord 6 <- fNextRow = 2, fNextRowCoordinate = 10.
71 // The "next" row is also the first (lowest) coordinate. This computation
72 // may yield a negative value, but that's OK, the math will work out
73 // since the user of this buffer will compute the offset relative
74 // to the firstRowIndex and the negative rows will never be used.
75 *firstRowIndex = fNextRowCoordinate - fNumRows;
77 int curRow = fNextRow;
78 for (int i = 0; i < fNumRows; i++) {
79 fRowAddresses[i] = &fBuffer[curRow * fRowByteWidth];
81 // Advance to the next row, wrapping if necessary.
83 if (curRow == fNumRows) {
87 return &fRowAddresses[0];
91 // The buffer storing the rows. They are packed, each one fRowByteWidth.
92 SkTArray<unsigned char> fBuffer;
94 // Number of bytes per row in the |buffer|.
97 // The number of rows available in the buffer.
100 // The next row index we should write into. This wraps around as the
101 // circular buffer is used.
104 // The y coordinate of the |fNextRow|. This is incremented each time a
105 // new row is appended and does not wrap.
106 int fNextRowCoordinate;
108 // Buffer used by GetRowAddresses().
109 SkTArray<unsigned char*> fRowAddresses;
112 // Convolves horizontally along a single row. The row data is given in
113 // |srcData| and continues for the numValues() of the filter.
114 template<bool hasAlpha>
115 void ConvolveHorizontally(const unsigned char* srcData,
116 const SkConvolutionFilter1D& filter,
117 unsigned char* outRow) {
118 // Loop over each pixel on this row in the output image.
119 int numValues = filter.numValues();
120 for (int outX = 0; outX < numValues; outX++) {
121 // Get the filter that determines the current output pixel.
122 int filterOffset, filterLength;
123 const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
124 filter.FilterForValue(outX, &filterOffset, &filterLength);
126 // Compute the first pixel in this row that the filter affects. It will
127 // touch |filterLength| pixels (4 bytes each) after this.
128 const unsigned char* rowToFilter = &srcData[filterOffset * 4];
130 // Apply the filter to the row to get the destination pixel in |accum|.
132 for (int filterX = 0; filterX < filterLength; filterX++) {
133 SkConvolutionFilter1D::ConvolutionFixed curFilter = filterValues[filterX];
134 accum[0] += curFilter * rowToFilter[filterX * 4 + 0];
135 accum[1] += curFilter * rowToFilter[filterX * 4 + 1];
136 accum[2] += curFilter * rowToFilter[filterX * 4 + 2];
138 accum[3] += curFilter * rowToFilter[filterX * 4 + 3];
142 // Bring this value back in range. All of the filter scaling factors
143 // are in fixed point with kShiftBits bits of fractional part.
144 accum[0] >>= SkConvolutionFilter1D::kShiftBits;
145 accum[1] >>= SkConvolutionFilter1D::kShiftBits;
146 accum[2] >>= SkConvolutionFilter1D::kShiftBits;
148 accum[3] >>= SkConvolutionFilter1D::kShiftBits;
151 // Store the new pixel.
152 outRow[outX * 4 + 0] = ClampTo8(accum[0]);
153 outRow[outX * 4 + 1] = ClampTo8(accum[1]);
154 outRow[outX * 4 + 2] = ClampTo8(accum[2]);
156 outRow[outX * 4 + 3] = ClampTo8(accum[3]);
161 // There's a bug somewhere here with GCC autovectorization (-ftree-vectorize). We originally
162 // thought this was 32 bit only, but subsequent tests show that some 64 bit gcc compiles
165 // Dropping to -O2 disables -ftree-vectorize. GCC 4.6 needs noinline. http://skbug.com/2575
166 #if SK_HAS_ATTRIBUTE(optimize) && defined(SK_RELEASE)
167 #define SK_MAYBE_DISABLE_VECTORIZATION __attribute__((optimize("O2"), noinline))
169 #define SK_MAYBE_DISABLE_VECTORIZATION
172 SK_MAYBE_DISABLE_VECTORIZATION
173 static void ConvolveHorizontallyAlpha(const unsigned char* srcData,
174 const SkConvolutionFilter1D& filter,
175 unsigned char* outRow) {
176 return ConvolveHorizontally<true>(srcData, filter, outRow);
179 SK_MAYBE_DISABLE_VECTORIZATION
180 static void ConvolveHorizontallyNoAlpha(const unsigned char* srcData,
181 const SkConvolutionFilter1D& filter,
182 unsigned char* outRow) {
183 return ConvolveHorizontally<false>(srcData, filter, outRow);
186 #undef SK_MAYBE_DISABLE_VECTORIZATION
189 // Does vertical convolution to produce one output row. The filter values and
190 // length are given in the first two parameters. These are applied to each
191 // of the rows pointed to in the |sourceDataRows| array, with each row
192 // being |pixelWidth| wide.
194 // The output must have room for |pixelWidth * 4| bytes.
195 template<bool hasAlpha>
196 void ConvolveVertically(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
198 unsigned char* const* sourceDataRows,
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 // Compute the number of bytes over in each row that the current column
205 // we're convolving starts at. The pixel will cover the next 4 bytes.
206 int byteOffset = outX * 4;
208 // Apply the filter to one column of pixels.
210 for (int filterY = 0; filterY < filterLength; filterY++) {
211 SkConvolutionFilter1D::ConvolutionFixed curFilter = filterValues[filterY];
212 accum[0] += curFilter * sourceDataRows[filterY][byteOffset + 0];
213 accum[1] += curFilter * sourceDataRows[filterY][byteOffset + 1];
214 accum[2] += curFilter * sourceDataRows[filterY][byteOffset + 2];
216 accum[3] += curFilter * sourceDataRows[filterY][byteOffset + 3];
220 // Bring this value back in range. All of the filter scaling factors
221 // are in fixed point with kShiftBits bits of precision.
222 accum[0] >>= SkConvolutionFilter1D::kShiftBits;
223 accum[1] >>= SkConvolutionFilter1D::kShiftBits;
224 accum[2] >>= SkConvolutionFilter1D::kShiftBits;
226 accum[3] >>= SkConvolutionFilter1D::kShiftBits;
229 // Store the new pixel.
230 outRow[byteOffset + 0] = ClampTo8(accum[0]);
231 outRow[byteOffset + 1] = ClampTo8(accum[1]);
232 outRow[byteOffset + 2] = ClampTo8(accum[2]);
234 unsigned char alpha = ClampTo8(accum[3]);
236 // Make sure the alpha channel doesn't come out smaller than any of the
237 // color channels. We use premultipled alpha channels, so this should
238 // never happen, but rounding errors will cause this from time to time.
239 // These "impossible" colors will cause overflows (and hence random pixel
240 // values) when the resulting bitmap is drawn to the screen.
242 // We only need to do this when generating the final output row (here).
243 int maxColorChannel = SkTMax(outRow[byteOffset + 0],
244 SkTMax(outRow[byteOffset + 1],
245 outRow[byteOffset + 2]));
246 if (alpha < maxColorChannel) {
247 outRow[byteOffset + 3] = maxColorChannel;
249 outRow[byteOffset + 3] = alpha;
252 // No alpha channel, the image is opaque.
253 outRow[byteOffset + 3] = 0xff;
258 void ConvolveVertically(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
260 unsigned char* const* sourceDataRows,
262 unsigned char* outRow,
263 bool sourceHasAlpha) {
264 if (sourceHasAlpha) {
265 ConvolveVertically<true>(filterValues, filterLength,
266 sourceDataRows, pixelWidth,
269 ConvolveVertically<false>(filterValues, filterLength,
270 sourceDataRows, pixelWidth,
277 // SkConvolutionFilter1D ---------------------------------------------------------
279 SkConvolutionFilter1D::SkConvolutionFilter1D()
283 SkConvolutionFilter1D::~SkConvolutionFilter1D() {
286 void SkConvolutionFilter1D::AddFilter(int filterOffset,
287 const float* filterValues,
289 SkASSERT(filterLength > 0);
291 SkTArray<ConvolutionFixed> fixedValues;
292 fixedValues.reset(filterLength);
294 for (int i = 0; i < filterLength; ++i) {
295 fixedValues.push_back(FloatToFixed(filterValues[i]));
298 AddFilter(filterOffset, &fixedValues[0], filterLength);
301 void SkConvolutionFilter1D::AddFilter(int filterOffset,
302 const ConvolutionFixed* filterValues,
304 // It is common for leading/trailing filter values to be zeros. In such
305 // cases it is beneficial to only store the central factors.
306 // For a scaling to 1/4th in each dimension using a Lanczos-2 filter on
307 // a 1080p image this optimization gives a ~10% speed improvement.
308 int filterSize = filterLength;
309 int firstNonZero = 0;
310 while (firstNonZero < filterLength && filterValues[firstNonZero] == 0) {
314 if (firstNonZero < filterLength) {
315 // Here we have at least one non-zero factor.
316 int lastNonZero = filterLength - 1;
317 while (lastNonZero >= 0 && filterValues[lastNonZero] == 0) {
321 filterOffset += firstNonZero;
322 filterLength = lastNonZero + 1 - firstNonZero;
323 SkASSERT(filterLength > 0);
325 for (int i = firstNonZero; i <= lastNonZero; i++) {
326 fFilterValues.push_back(filterValues[i]);
329 // Here all the factors were zeroes.
333 FilterInstance instance;
335 // We pushed filterLength elements onto fFilterValues
336 instance.fDataLocation = (static_cast<int>(fFilterValues.count()) -
338 instance.fOffset = filterOffset;
339 instance.fTrimmedLength = filterLength;
340 instance.fLength = filterSize;
341 fFilters.push_back(instance);
343 fMaxFilter = SkTMax(fMaxFilter, filterLength);
346 const SkConvolutionFilter1D::ConvolutionFixed* SkConvolutionFilter1D::GetSingleFilter(
347 int* specifiedFilterlength,
349 int* filterLength) const {
350 const FilterInstance& filter = fFilters[0];
351 *filterOffset = filter.fOffset;
352 *filterLength = filter.fTrimmedLength;
353 *specifiedFilterlength = filter.fLength;
354 if (filter.fTrimmedLength == 0) {
358 return &fFilterValues[filter.fDataLocation];
361 void BGRAConvolve2D(const unsigned char* sourceData,
362 int sourceByteRowStride,
364 const SkConvolutionFilter1D& filterX,
365 const SkConvolutionFilter1D& filterY,
366 int outputByteRowStride,
367 unsigned char* output,
368 const SkConvolutionProcs& convolveProcs,
369 bool useSimdIfPossible) {
371 int maxYFilterSize = filterY.maxFilter();
373 // The next row in the input that we will generate a horizontally
374 // convolved row for. If the filter doesn't start at the beginning of the
375 // image (this is the case when we are only resizing a subset), then we
376 // don't want to generate any output rows before that. Compute the starting
377 // row for convolution as the first pixel for the first vertical filter.
378 int filterOffset, filterLength;
379 const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
380 filterY.FilterForValue(0, &filterOffset, &filterLength);
381 int nextXRow = filterOffset;
383 // We loop over each row in the input doing a horizontal convolution. This
384 // will result in a horizontally convolved image. We write the results into
385 // a circular buffer of convolved rows and do vertical convolution as rows
386 // are available. This prevents us from having to store the entire
387 // intermediate image and helps cache coherency.
388 // We will need four extra rows to allow horizontal convolution could be done
389 // simultaneously. We also pad each row in row buffer to be aligned-up to
391 // TODO(jiesun): We do not use aligned load from row buffer in vertical
392 // convolution pass yet. Somehow Windows does not like it.
393 int rowBufferWidth = (filterX.numValues() + 15) & ~0xF;
394 int rowBufferHeight = maxYFilterSize +
395 (convolveProcs.fConvolve4RowsHorizontally ? 4 : 0);
396 CircularRowBuffer rowBuffer(rowBufferWidth,
400 // Loop over every possible output row, processing just enough horizontal
401 // convolutions to run each subsequent vertical convolution.
402 SkASSERT(outputByteRowStride >= filterX.numValues() * 4);
403 int numOutputRows = filterY.numValues();
405 // We need to check which is the last line to convolve before we advance 4
406 // lines in one iteration.
407 int lastFilterOffset, lastFilterLength;
409 // SSE2 can access up to 3 extra pixels past the end of the
410 // buffer. At the bottom of the image, we have to be careful
411 // not to access data past the end of the buffer. Normally
412 // we fall back to the C++ implementation for the last row.
413 // If the last row is less than 3 pixels wide, we may have to fall
414 // back to the C++ version for more rows. Compute how many
415 // rows we need to avoid the SSE implementation for here.
416 filterX.FilterForValue(filterX.numValues() - 1, &lastFilterOffset,
418 int avoidSimdRows = 1 + convolveProcs.fExtraHorizontalReads /
419 (lastFilterOffset + lastFilterLength);
421 filterY.FilterForValue(numOutputRows - 1, &lastFilterOffset,
424 for (int outY = 0; outY < numOutputRows; outY++) {
425 filterValues = filterY.FilterForValue(outY,
426 &filterOffset, &filterLength);
428 // Generate output rows until we have enough to run the current filter.
429 while (nextXRow < filterOffset + filterLength) {
430 if (convolveProcs.fConvolve4RowsHorizontally &&
431 nextXRow + 3 < lastFilterOffset + lastFilterLength -
433 const unsigned char* src[4];
434 unsigned char* outRow[4];
435 for (int i = 0; i < 4; ++i) {
436 src[i] = &sourceData[(uint64_t)(nextXRow + i) * sourceByteRowStride];
437 outRow[i] = rowBuffer.advanceRow();
439 convolveProcs.fConvolve4RowsHorizontally(src, filterX, outRow);
442 // Check if we need to avoid SSE2 for this row.
443 if (convolveProcs.fConvolveHorizontally &&
444 nextXRow < lastFilterOffset + lastFilterLength -
446 convolveProcs.fConvolveHorizontally(
447 &sourceData[(uint64_t)nextXRow * sourceByteRowStride],
448 filterX, rowBuffer.advanceRow(), sourceHasAlpha);
450 if (sourceHasAlpha) {
451 ConvolveHorizontallyAlpha(
452 &sourceData[(uint64_t)nextXRow * sourceByteRowStride],
453 filterX, rowBuffer.advanceRow());
455 ConvolveHorizontallyNoAlpha(
456 &sourceData[(uint64_t)nextXRow * sourceByteRowStride],
457 filterX, rowBuffer.advanceRow());
464 // Compute where in the output image this row of final data will go.
465 unsigned char* curOutputRow = &output[(uint64_t)outY * outputByteRowStride];
467 // Get the list of rows that the circular buffer has, in order.
468 int firstRowInCircularBuffer;
469 unsigned char* const* rowsToConvolve =
470 rowBuffer.GetRowAddresses(&firstRowInCircularBuffer);
472 // Now compute the start of the subset of those rows that the filter
474 unsigned char* const* firstRowForFilter =
475 &rowsToConvolve[filterOffset - firstRowInCircularBuffer];
477 if (convolveProcs.fConvolveVertically) {
478 convolveProcs.fConvolveVertically(filterValues, filterLength,
480 filterX.numValues(), curOutputRow,
483 ConvolveVertically(filterValues, filterLength,
485 filterX.numValues(), curOutputRow,