bool BGRAConvolve2D(const unsigned char* sourceData,
int sourceByteRowStride,
bool sourceHasAlpha,
const SkConvolutionFilter1D& filterX,
const SkConvolutionFilter1D& filterY,
int outputByteRowStride,
unsigned char* output) {
int maxYFilterSize = filterY.maxFilter();
// The next row in the input that we will generate a horizontally
// convolved row for. If the filter doesn't start at the beginning of the
// image (this is the case when we are only resizing a subset), then we
// don't want to generate any output rows before that. Compute the starting
// row for convolution as the first pixel for the first vertical filter.
int filterOffset, filterLength;
const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
filterY.FilterForValue(0, &filterOffset, &filterLength);
int nextXRow = filterOffset;
// We loop over each row in the input doing a horizontal convolution. This
// will result in a horizontally convolved image. We write the results into
// a circular buffer of convolved rows and do vertical convolution as rows
// are available. This prevents us from having to store the entire
// intermediate image and helps cache coherency.
// We will need four extra rows to allow horizontal convolution could be done
// simultaneously. We also pad each row in row buffer to be aligned-up to
// 32 bytes.
// TODO(jiesun): We do not use aligned load from row buffer in vertical
// convolution pass yet. Somehow Windows does not like it.
int rowBufferWidth = (filterX.numValues() + 31) & ~0x1F;
int rowBufferHeight = maxYFilterSize +
(SkOpts::convolve_4_rows_horizontally != nullptr ? 4 : 0);
// check for too-big allocation requests : crbug.com/528628
{
int64_t size = sk_64_mul(rowBufferWidth, rowBufferHeight);
// need some limit, to avoid over-committing success from malloc, but then
// crashing when we try to actually use the memory.
// 100meg seems big enough to allow "normal" zoom factors and image sizes through
// while avoiding the crash seen by the bug (crbug.com/528628)
if (size > 100 * 1024 * 1024) {
// SkDebugf("BGRAConvolve2D: tmp allocation [%lld] too big\n", size);
return false;
}
}
CircularRowBuffer rowBuffer(rowBufferWidth,
rowBufferHeight,
filterOffset);
// Loop over every possible output row, processing just enough horizontal
// convolutions to run each subsequent vertical convolution.
SkASSERT(outputByteRowStride >= filterX.numValues() * 4);
int numOutputRows = filterY.numValues();
// We need to check which is the last line to convolve before we advance 4
// lines in one iteration.
int lastFilterOffset, lastFilterLength;
filterY.FilterForValue(numOutputRows - 1, &lastFilterOffset,
&lastFilterLength);
for (int outY = 0; outY < numOutputRows; outY++) {
filterValues = filterY.FilterForValue(outY,
&filterOffset, &filterLength);
// Generate output rows until we have enough to run the current filter.
while (nextXRow < filterOffset + filterLength) {
if (SkOpts::convolve_4_rows_horizontally != nullptr &&
nextXRow + 3 < lastFilterOffset + lastFilterLength) {
const unsigned char* src[4];
unsigned char* outRow[4];
for (int i = 0; i < 4; ++i) {
src[i] = &sourceData[(uint64_t)(nextXRow + i) * sourceByteRowStride];
outRow[i] = rowBuffer.advanceRow();
}
SkOpts::convolve_4_rows_horizontally(src, filterX, outRow, 4*rowBufferWidth);
nextXRow += 4;
} else {
SkOpts::convolve_horizontally(
&sourceData[(uint64_t)nextXRow * sourceByteRowStride],
filterX, rowBuffer.advanceRow(), sourceHasAlpha);
nextXRow++;
}
}
// Compute where in the output image this row of final data will go.
unsigned char* curOutputRow = &output[(uint64_t)outY * outputByteRowStride];
// Get the list of rows that the circular buffer has, in order.
int firstRowInCircularBuffer;
unsigned char* const* rowsToConvolve =
rowBuffer.GetRowAddresses(&firstRowInCircularBuffer);
// Now compute the start of the subset of those rows that the filter needs.
unsigned char* const* firstRowForFilter =
&rowsToConvolve[filterOffset - firstRowInCircularBuffer];
SkOpts::convolve_vertically(filterValues, filterLength,
firstRowForFilter,
filterX.numValues(), curOutputRow,
sourceHasAlpha);
}
return true;
}
void 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) {
int maxYFilterSize = filterY.maxFilter();
// The next row in the input that we will generate a horizontally
// convolved row for. If the filter doesn't start at the beginning of the
// image (this is the case when we are only resizing a subset), then we
// don't want to generate any output rows before that. Compute the starting
// row for convolution as the first pixel for the first vertical filter.
int filterOffset, filterLength;
const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
filterY.FilterForValue(0, &filterOffset, &filterLength);
int nextXRow = filterOffset;
// We loop over each row in the input doing a horizontal convolution. This
// will result in a horizontally convolved image. We write the results into
// a circular buffer of convolved rows and do vertical convolution as rows
// are available. This prevents us from having to store the entire
// intermediate image and helps cache coherency.
// We will need four extra rows to allow horizontal convolution could be done
// simultaneously. We also pad each row in row buffer to be aligned-up to
// 16 bytes.
// TODO(jiesun): We do not use aligned load from row buffer in vertical
// convolution pass yet. Somehow Windows does not like it.
int rowBufferWidth = (filterX.numValues() + 15) & ~0xF;
int rowBufferHeight = maxYFilterSize +
(convolveProcs.fConvolve4RowsHorizontally ? 4 : 0);
CircularRowBuffer rowBuffer(rowBufferWidth,
rowBufferHeight,
filterOffset);
// Loop over every possible output row, processing just enough horizontal
// convolutions to run each subsequent vertical convolution.
SkASSERT(outputByteRowStride >= filterX.numValues() * 4);
int numOutputRows = filterY.numValues();
// We need to check which is the last line to convolve before we advance 4
// lines in one iteration.
int lastFilterOffset, lastFilterLength;
// SSE2 can access up to 3 extra pixels past the end of the
// buffer. At the bottom of the image, we have to be careful
// not to access data past the end of the buffer. Normally
// we fall back to the C++ implementation for the last row.
// If the last row is less than 3 pixels wide, we may have to fall
// back to the C++ version for more rows. Compute how many
// rows we need to avoid the SSE implementation for here.
filterX.FilterForValue(filterX.numValues() - 1, &lastFilterOffset,
&lastFilterLength);
int avoidSimdRows = 1 + convolveProcs.fExtraHorizontalReads /
(lastFilterOffset + lastFilterLength);
filterY.FilterForValue(numOutputRows - 1, &lastFilterOffset,
&lastFilterLength);
for (int outY = 0; outY < numOutputRows; outY++) {
filterValues = filterY.FilterForValue(outY,
&filterOffset, &filterLength);
// Generate output rows until we have enough to run the current filter.
while (nextXRow < filterOffset + filterLength) {
if (convolveProcs.fConvolve4RowsHorizontally &&
nextXRow + 3 < lastFilterOffset + lastFilterLength -
avoidSimdRows) {
const unsigned char* src[4];
unsigned char* outRow[4];
for (int i = 0; i < 4; ++i) {
src[i] = &sourceData[(uint64_t)(nextXRow + i) * sourceByteRowStride];
outRow[i] = rowBuffer.advanceRow();
}
convolveProcs.fConvolve4RowsHorizontally(src, filterX, outRow, 4*rowBufferWidth);
nextXRow += 4;
} else {
// Check if we need to avoid SSE2 for this row.
if (convolveProcs.fConvolveHorizontally &&
nextXRow < lastFilterOffset + lastFilterLength -
avoidSimdRows) {
convolveProcs.fConvolveHorizontally(
&sourceData[(uint64_t)nextXRow * sourceByteRowStride],
filterX, rowBuffer.advanceRow(), sourceHasAlpha);
} else {
if (sourceHasAlpha) {
ConvolveHorizontallyAlpha(
&sourceData[(uint64_t)nextXRow * sourceByteRowStride],
filterX, rowBuffer.advanceRow());
} else {
ConvolveHorizontallyNoAlpha(
&sourceData[(uint64_t)nextXRow * sourceByteRowStride],
filterX, rowBuffer.advanceRow());
}
}
nextXRow++;
}
}
// Compute where in the output image this row of final data will go.
unsigned char* curOutputRow = &output[(uint64_t)outY * outputByteRowStride];
// Get the list of rows that the circular buffer has, in order.
int firstRowInCircularBuffer;
unsigned char* const* rowsToConvolve =
rowBuffer.GetRowAddresses(&firstRowInCircularBuffer);
// Now compute the start of the subset of those rows that the filter
// needs.
unsigned char* const* firstRowForFilter =
&rowsToConvolve[filterOffset - firstRowInCircularBuffer];
if (convolveProcs.fConvolveVertically) {
convolveProcs.fConvolveVertically(filterValues, filterLength,
firstRowForFilter,
filterX.numValues(), curOutputRow,
sourceHasAlpha);
} else {
ConvolveVertically(filterValues, filterLength,
firstRowForFilter,
filterX.numValues(), curOutputRow,
sourceHasAlpha);
}
}
}