static uint16_t darken_modeproc16_255(SkPMColor src, uint16_t dst) {
SkASSERT(require_255(src));
unsigned r = SkFastMin32(SkPacked32ToR16(src), SkGetPackedR16(dst));
unsigned g = SkFastMin32(SkPacked32ToG16(src), SkGetPackedG16(dst));
unsigned b = SkFastMin32(SkPacked32ToB16(src), SkGetPackedB16(dst));
return SkPackRGB16(r, g, b);
}
/**
* This is the path for apply_kernel_interp() to be taken when the kernel
* is wider than the source image.
*/
static void kernel_interp_clamped(uint8_t dst[], int rx, int ry,
const uint32_t sum[], int sw, int sh, U8CPU outerWeight) {
SkASSERT(2*rx > sw);
int innerWeight = 255 - outerWeight;
// round these guys up if they're bigger than 127
outerWeight += outerWeight >> 7;
innerWeight += innerWeight >> 7;
uint32_t outerScale = (outerWeight << 16) / ((2*rx + 1)*(2*ry + 1));
uint32_t innerScale = (innerWeight << 16) / ((2*rx - 1)*(2*ry - 1));
int sumStride = sw + 1;
int dw = sw + 2*rx;
int dh = sh + 2*ry;
int prev_y = -2*ry;
int next_y = 1;
for (int y = 0; y < dh; ++y) {
int py = SkClampPos(prev_y) * sumStride;
int ny = SkFastMin32(next_y, sh) * sumStride;
int ipy = SkClampPos(prev_y + 1) * sumStride;
int iny = SkClampMax(next_y - 1, sh) * sumStride;
int prev_x = -2*rx;
int next_x = 1;
for (int x = 0; x < dw; ++x) {
int px = SkClampPos(prev_x);
int nx = SkFastMin32(next_x, sw);
int ipx = SkClampPos(prev_x + 1);
int inx = SkClampMax(next_x - 1, sw);
uint32_t outerSum = sum[px+py] + sum[nx+ny]
- sum[nx+py] - sum[px+ny];
uint32_t innerSum = sum[ipx+ipy] + sum[inx+iny]
- sum[inx+ipy] - sum[ipx+iny];
*dst++ = SkToU8((outerSum * outerScale
+ innerSum * innerScale) >> 24);
prev_x += 1;
next_x += 1;
}
prev_y += 1;
next_y += 1;
}
}
static SkPMColor lighten_modeproc(SkPMColor src, SkPMColor dst) {
unsigned sa = SkGetPackedA32(src);
unsigned da = SkGetPackedA32(dst);
unsigned src_scale = SkAlpha255To256(255 - sa);
unsigned dst_scale = SkAlpha255To256(255 - da);
unsigned ra = sa + da - SkAlphaMulAlpha(sa, da);
unsigned rr = lighten_p(SkGetPackedR32(src), SkGetPackedR32(dst),
src_scale, dst_scale);
unsigned rg = lighten_p(SkGetPackedG32(src), SkGetPackedG32(dst),
src_scale, dst_scale);
unsigned rb = lighten_p(SkGetPackedB32(src), SkGetPackedB32(dst),
src_scale, dst_scale);
return SkPackARGB32(ra, SkFastMin32(rr, ra),
SkFastMin32(rg, ra), SkFastMin32(rb, ra));
}
/**
* This is the path for apply_kernel() to be taken when the kernel
* is wider than the source image.
*/
static void kernel_clamped(uint8_t dst[], int rx, int ry, const uint32_t sum[],
int sw, int sh) {
SkASSERT(2*rx > sw);
uint32_t scale = (1 << 24) / ((2*rx + 1)*(2*ry + 1));
int sumStride = sw + 1;
int dw = sw + 2*rx;
int dh = sh + 2*ry;
int prev_y = -2*ry;
int next_y = 1;
for (int y = 0; y < dh; ++y) {
int py = SkClampPos(prev_y) * sumStride;
int ny = SkFastMin32(next_y, sh) * sumStride;
int prev_x = -2*rx;
int next_x = 1;
for (int x = 0; x < dw; ++x) {
int px = SkClampPos(prev_x);
int nx = SkFastMin32(next_x, sw);
// TODO: should we be adding 1/2 (1 << 23) to round to the
// nearest integer here?
uint32_t tmp = sum[px+py] + sum[nx+ny] - sum[nx+py] - sum[px+ny];
*dst++ = SkToU8(tmp * scale >> 24);
prev_x += 1;
next_x += 1;
}
prev_y += 1;
next_y += 1;
}
}
/**
* This is the path for apply_kernel() to be taken when the kernel
* is wider than the source image.
*/
static void kernel_clamped(uint8_t dst[], int rx, int ry, const uint32_t sum[],
int sw, int sh) {
SkASSERT(2*rx > sw);
uint32_t scale = (1 << 24) / ((2*rx + 1)*(2*ry + 1));
int sumStride = sw + 1;
int dw = sw + 2*rx;
int dh = sh + 2*ry;
int prev_y = -2*ry;
int next_y = 1;
for (int y = 0; y < dh; y++) {
int py = SkClampPos(prev_y) * sumStride;
int ny = SkFastMin32(next_y, sh) * sumStride;
int prev_x = -2*rx;
int next_x = 1;
for (int x = 0; x < dw; x++) {
int px = SkClampPos(prev_x);
int nx = SkFastMin32(next_x, sw);
uint32_t tmp = sum[px+py] + sum[nx+ny] - sum[nx+py] - sum[px+ny];
*dst++ = SkToU8(tmp * scale >> 24);
prev_x += 1;
next_x += 1;
}
prev_y += 1;
next_y += 1;
}
}
void SkScalerContext::getImage(const SkGlyph& origGlyph) {
const SkGlyph* glyph = &origGlyph;
SkGlyph tmpGlyph;
if (fMaskFilter) { // restore the prefilter bounds
tmpGlyph.init(origGlyph.fID);
// need the original bounds, sans our maskfilter
SkMaskFilter* mf = fMaskFilter;
fMaskFilter = NULL; // temp disable
this->getMetrics(&tmpGlyph);
fMaskFilter = mf; // restore
tmpGlyph.fImage = origGlyph.fImage;
// we need the prefilter bounds to be <= filter bounds
SkASSERT(tmpGlyph.fWidth <= origGlyph.fWidth);
SkASSERT(tmpGlyph.fHeight <= origGlyph.fHeight);
glyph = &tmpGlyph;
}
if (fRec.fFrameWidth > 0 || fPathEffect != NULL || fRasterizer != NULL) {
SkPath devPath, fillPath;
SkMatrix fillToDevMatrix;
this->internalGetPath(*glyph, &fillPath, &devPath, &fillToDevMatrix);
if (fRasterizer) {
SkMask mask;
glyph->toMask(&mask);
mask.fFormat = SkMask::kA8_Format;
sk_bzero(glyph->fImage, mask.computeImageSize());
if (!fRasterizer->rasterize(fillPath, fillToDevMatrix, NULL,
fMaskFilter, &mask,
SkMask::kJustRenderImage_CreateMode)) {
return;
}
} else {
SkBitmap bm;
SkBitmap::Config config;
SkMatrix matrix;
SkRegion clip;
SkPaint paint;
SkDraw draw;
if (SkMask::kA8_Format == fRec.fMaskFormat) {
config = SkBitmap::kA8_Config;
paint.setAntiAlias(true);
} else {
SkASSERT(SkMask::kBW_Format == fRec.fMaskFormat);
config = SkBitmap::kA1_Config;
paint.setAntiAlias(false);
}
clip.setRect(0, 0, glyph->fWidth, glyph->fHeight);
matrix.setTranslate(-SkIntToScalar(glyph->fLeft),
-SkIntToScalar(glyph->fTop));
bm.setConfig(config, glyph->fWidth, glyph->fHeight,
glyph->rowBytes());
bm.setPixels(glyph->fImage);
sk_bzero(glyph->fImage, bm.height() * bm.rowBytes());
draw.fClip = &clip;
draw.fMatrix = &matrix;
draw.fBitmap = &bm;
draw.fBounder = NULL;
draw.drawPath(devPath, paint);
}
} else {
this->getGlyphContext(*glyph)->generateImage(*glyph);
}
if (fMaskFilter) {
SkMask srcM, dstM;
SkMatrix matrix;
// the src glyph image shouldn't be 3D
SkASSERT(SkMask::k3D_Format != glyph->fMaskFormat);
glyph->toMask(&srcM);
fRec.getMatrixFrom2x2(&matrix);
if (fMaskFilter->filterMask(&dstM, srcM, matrix, NULL)) {
int width = SkFastMin32(origGlyph.fWidth, dstM.fBounds.width());
int height = SkFastMin32(origGlyph.fHeight, dstM.fBounds.height());
int dstRB = origGlyph.rowBytes();
int srcRB = dstM.fRowBytes;
const uint8_t* src = (const uint8_t*)dstM.fImage;
uint8_t* dst = (uint8_t*)origGlyph.fImage;
if (SkMask::k3D_Format == dstM.fFormat) {
// we have to copy 3 times as much
height *= 3;
}
// clean out our glyph, since it may be larger than dstM
//sk_bzero(dst, height * dstRB);
while (--height >= 0) {
memcpy(dst, src, width);
src += srcRB;
dst += dstRB;
}
SkMask::FreeImage(dstM.fImage);
}
}
// check to see if we should filter the alpha channel
if (NULL == fMaskFilter &&
fRec.fMaskFormat != SkMask::kBW_Format &&
fRec.fMaskFormat != SkMask::kLCD16_Format &&
fRec.fMaskFormat != SkMask::kLCD32_Format &&
(fRec.fFlags & (kGammaForBlack_Flag | kGammaForWhite_Flag)) != 0)
{
const uint8_t* table = (fRec.fFlags & kGammaForBlack_Flag) ? gBlackGammaTable : gWhiteGammaTable;
if (NULL != table) {
uint8_t* dst = (uint8_t*)origGlyph.fImage;
unsigned rowBytes = origGlyph.rowBytes();
for (int y = origGlyph.fHeight - 1; y >= 0; --y) {
for (int x = origGlyph.fWidth - 1; x >= 0; --x) {
dst[x] = table[dst[x]];
}
dst += rowBytes;
}
}
}
}
void SkScalerContext::getImage(const SkGlyph& origGlyph) {
const SkGlyph* glyph = &origGlyph;
SkGlyph tmpGlyph;
if (fMaskFilter) { // restore the prefilter bounds
tmpGlyph.init(origGlyph.fID);
// need the original bounds, sans our maskfilter
SkMaskFilter* mf = fMaskFilter;
fMaskFilter = NULL; // temp disable
this->getMetrics(&tmpGlyph);
fMaskFilter = mf; // restore
tmpGlyph.fImage = origGlyph.fImage;
// we need the prefilter bounds to be <= filter bounds
SkASSERT(tmpGlyph.fWidth <= origGlyph.fWidth);
SkASSERT(tmpGlyph.fHeight <= origGlyph.fHeight);
glyph = &tmpGlyph;
}
if (fGenerateImageFromPath) {
SkPath devPath, fillPath;
SkMatrix fillToDevMatrix;
SkMask mask;
this->internalGetPath(*glyph, &fillPath, &devPath, &fillToDevMatrix);
glyph->toMask(&mask);
if (fRasterizer) {
mask.fFormat = SkMask::kA8_Format;
sk_bzero(glyph->fImage, mask.computeImageSize());
if (!fRasterizer->rasterize(fillPath, fillToDevMatrix, NULL,
fMaskFilter, &mask,
SkMask::kJustRenderImage_CreateMode)) {
return;
}
} else {
generateMask(mask, devPath);
}
} else {
this->getGlyphContext(*glyph)->generateImage(*glyph);
}
if (fMaskFilter) {
SkMask srcM, dstM;
SkMatrix matrix;
// the src glyph image shouldn't be 3D
SkASSERT(SkMask::k3D_Format != glyph->fMaskFormat);
glyph->toMask(&srcM);
fRec.getMatrixFrom2x2(&matrix);
if (fMaskFilter->filterMask(&dstM, srcM, matrix, NULL)) {
int width = SkFastMin32(origGlyph.fWidth, dstM.fBounds.width());
int height = SkFastMin32(origGlyph.fHeight, dstM.fBounds.height());
int dstRB = origGlyph.rowBytes();
int srcRB = dstM.fRowBytes;
const uint8_t* src = (const uint8_t*)dstM.fImage;
uint8_t* dst = (uint8_t*)origGlyph.fImage;
if (SkMask::k3D_Format == dstM.fFormat) {
// we have to copy 3 times as much
height *= 3;
}
// clean out our glyph, since it may be larger than dstM
//sk_bzero(dst, height * dstRB);
while (--height >= 0) {
memcpy(dst, src, width);
src += srcRB;
dst += dstRB;
}
SkMask::FreeImage(dstM.fImage);
}
}
}
void SkScalerContext::getImage(const SkGlyph& origGlyph) {
const SkGlyph* glyph = &origGlyph;
SkGlyph tmpGlyph;
// in case we need to call generateImage on a mask-format that is different
// (i.e. larger) than what our caller allocated by looking at origGlyph.
SkAutoMalloc tmpGlyphImageStorage;
// If we are going to draw-from-path, then we cannot generate color, since
// the path only makes a mask. This case should have been caught up in
// generateMetrics().
SkASSERT(!fGenerateImageFromPath ||
SkMask::kARGB32_Format != origGlyph.fMaskFormat);
if (fMaskFilter) { // restore the prefilter bounds
tmpGlyph.initGlyphIdFrom(origGlyph);
// need the original bounds, sans our maskfilter
SkMaskFilter* mf = fMaskFilter;
fMaskFilter = nullptr; // temp disable
this->getMetrics(&tmpGlyph);
fMaskFilter = mf; // restore
// we need the prefilter bounds to be <= filter bounds
SkASSERT(tmpGlyph.fWidth <= origGlyph.fWidth);
SkASSERT(tmpGlyph.fHeight <= origGlyph.fHeight);
if (tmpGlyph.fMaskFormat == origGlyph.fMaskFormat) {
tmpGlyph.fImage = origGlyph.fImage;
} else {
tmpGlyphImageStorage.reset(tmpGlyph.computeImageSize());
tmpGlyph.fImage = tmpGlyphImageStorage.get();
}
glyph = &tmpGlyph;
}
if (fGenerateImageFromPath) {
SkPath devPath, fillPath;
SkMatrix fillToDevMatrix;
SkMask mask;
this->internalGetPath(*glyph, &fillPath, &devPath, &fillToDevMatrix);
glyph->toMask(&mask);
if (fRasterizer) {
mask.fFormat = SkMask::kA8_Format;
sk_bzero(glyph->fImage, mask.computeImageSize());
if (!fRasterizer->rasterize(fillPath, fillToDevMatrix, nullptr,
fMaskFilter, &mask,
SkMask::kJustRenderImage_CreateMode)) {
return;
}
if (fPreBlend.isApplicable()) {
applyLUTToA8Mask(mask, fPreBlend.fG);
}
} else {
SkASSERT(SkMask::kARGB32_Format != mask.fFormat);
generateMask(mask, devPath, fPreBlend);
}
} else {
generateImage(*glyph);
}
if (fMaskFilter) {
SkMask srcM, dstM;
SkMatrix matrix;
// the src glyph image shouldn't be 3D
SkASSERT(SkMask::k3D_Format != glyph->fMaskFormat);
SkAutoSMalloc<32*32> a8storage;
glyph->toMask(&srcM);
if (SkMask::kARGB32_Format == srcM.fFormat) {
// now we need to extract the alpha-channel from the glyph's image
// and copy it into a temp buffer, and then point srcM at that temp.
srcM.fFormat = SkMask::kA8_Format;
srcM.fRowBytes = SkAlign4(srcM.fBounds.width());
size_t size = srcM.computeImageSize();
a8storage.reset(size);
srcM.fImage = (uint8_t*)a8storage.get();
extract_alpha(srcM,
(const SkPMColor*)glyph->fImage, glyph->rowBytes());
}
fRec.getMatrixFrom2x2(&matrix);
if (fMaskFilter->filterMask(&dstM, srcM, matrix, nullptr)) {
int width = SkFastMin32(origGlyph.fWidth, dstM.fBounds.width());
int height = SkFastMin32(origGlyph.fHeight, dstM.fBounds.height());
int dstRB = origGlyph.rowBytes();
int srcRB = dstM.fRowBytes;
const uint8_t* src = (const uint8_t*)dstM.fImage;
uint8_t* dst = (uint8_t*)origGlyph.fImage;
if (SkMask::k3D_Format == dstM.fFormat) {
// we have to copy 3 times as much
height *= 3;
}
// clean out our glyph, since it may be larger than dstM
//sk_bzero(dst, height * dstRB);
while (--height >= 0) {
memcpy(dst, src, width);
src += srcRB;
dst += dstRB;
}
SkMask::FreeImage(dstM.fImage);
if (fPreBlendForFilter.isApplicable()) {
applyLUTToA8Mask(srcM, fPreBlendForFilter.fG);
}
}
}
}
/**
* sw and sh are the width and height of the src. Since the sum buffer
* matches that, but has an extra row and col at the beginning (with zeros),
* we can just use sw and sh as our "max" values for pinning coordinates
* when sampling into sum[][]
*
* The inner loop is conceptually simple; we break it into several variants
* to improve performance. Here's the original version:
for (int x = 0; x < dw; ++x) {
int px = SkClampPos(prev_x);
int nx = SkFastMin32(next_x, sw);
int ipx = SkClampPos(prev_x + 1);
int inx = SkClampMax(next_x - 1, sw);
uint32_t outerSum = sum[px+py] + sum[nx+ny]
- sum[nx+py] - sum[px+ny];
uint32_t innerSum = sum[ipx+ipy] + sum[inx+iny]
- sum[inx+ipy] - sum[ipx+iny];
*dst++ = SkToU8((outerSum * outerScale
+ innerSum * innerScale) >> 24);
prev_x += 1;
next_x += 1;
}
* The sections are:
* left-hand section, where prev_x is clamped to 0
* center section, where neither prev_x nor next_x is clamped
* right-hand section, where next_x is clamped to sw
* On some operating systems, the center section is unrolled for additional
* speedup.
*/
static void apply_kernel_interp(uint8_t dst[], int rx, int ry,
const uint32_t sum[], int sw, int sh, U8CPU outerWeight) {
SkASSERT(rx > 0 && ry > 0);
SkASSERT(outerWeight <= 255);
if (2*rx > sw) {
kernel_interp_clamped(dst, rx, ry, sum, sw, sh, outerWeight);
return;
}
int innerWeight = 255 - outerWeight;
// round these guys up if they're bigger than 127
outerWeight += outerWeight >> 7;
innerWeight += innerWeight >> 7;
uint32_t outerScale = (outerWeight << 16) / ((2*rx + 1)*(2*ry + 1));
uint32_t innerScale = (innerWeight << 16) / ((2*rx - 1)*(2*ry - 1));
int sumStride = sw + 1;
int dw = sw + 2*rx;
int dh = sh + 2*ry;
int prev_y = -2*ry;
int next_y = 1;
SkASSERT(2*rx <= dw - 2*rx);
for (int y = 0; y < dh; ++y) {
int py = SkClampPos(prev_y) * sumStride;
int ny = SkFastMin32(next_y, sh) * sumStride;
int ipy = SkClampPos(prev_y + 1) * sumStride;
int iny = SkClampMax(next_y - 1, sh) * sumStride;
int prev_x = -2*rx;
int next_x = 1;
int x = 0;
for (; x < 2*rx; ++x) {
SkASSERT(prev_x < 0);
SkASSERT(next_x <= sw);
int px = 0;
int nx = next_x;
int ipx = 0;
int inx = next_x - 1;
uint32_t outerSum = sum[px+py] + sum[nx+ny]
- sum[nx+py] - sum[px+ny];
uint32_t innerSum = sum[ipx+ipy] + sum[inx+iny]
- sum[inx+ipy] - sum[ipx+iny];
*dst++ = SkToU8((outerSum * outerScale
+ innerSum * innerScale) >> 24);
prev_x += 1;
next_x += 1;
}
int i0 = prev_x + py;
int i1 = next_x + ny;
int i2 = next_x + py;
int i3 = prev_x + ny;
int i4 = prev_x + 1 + ipy;
int i5 = next_x - 1 + iny;
int i6 = next_x - 1 + ipy;
int i7 = prev_x + 1 + iny;
#if UNROLL_KERNEL_LOOP
for (; x < dw - 2*rx - 4; x += 4) {
SkASSERT(prev_x >= 0);
SkASSERT(next_x <= sw);
uint32_t outerSum = sum[i0++] + sum[i1++] - sum[i2++] - sum[i3++];
uint32_t innerSum = sum[i4++] + sum[i5++] - sum[i6++] - sum[i7++];
*dst++ = SkToU8((outerSum * outerScale
+ innerSum * innerScale) >> 24);
outerSum = sum[i0++] + sum[i1++] - sum[i2++] - sum[i3++];
innerSum = sum[i4++] + sum[i5++] - sum[i6++] - sum[i7++];
*dst++ = SkToU8((outerSum * outerScale
+ innerSum * innerScale) >> 24);
outerSum = sum[i0++] + sum[i1++] - sum[i2++] - sum[i3++];
innerSum = sum[i4++] + sum[i5++] - sum[i6++] - sum[i7++];
*dst++ = SkToU8((outerSum * outerScale
+ innerSum * innerScale) >> 24);
outerSum = sum[i0++] + sum[i1++] - sum[i2++] - sum[i3++];
innerSum = sum[i4++] + sum[i5++] - sum[i6++] - sum[i7++];
*dst++ = SkToU8((outerSum * outerScale
+ innerSum * innerScale) >> 24);
prev_x += 4;
next_x += 4;
}
#endif
for (; x < dw - 2*rx; ++x) {
SkASSERT(prev_x >= 0);
SkASSERT(next_x <= sw);
uint32_t outerSum = sum[i0++] + sum[i1++] - sum[i2++] - sum[i3++];
uint32_t innerSum = sum[i4++] + sum[i5++] - sum[i6++] - sum[i7++];
*dst++ = SkToU8((outerSum * outerScale
+ innerSum * innerScale) >> 24);
prev_x += 1;
next_x += 1;
}
for (; x < dw; ++x) {
SkASSERT(prev_x >= 0);
SkASSERT(next_x > sw);
int px = prev_x;
int nx = sw;
int ipx = prev_x + 1;
int inx = sw;
uint32_t outerSum = sum[px+py] + sum[nx+ny]
- sum[nx+py] - sum[px+ny];
uint32_t innerSum = sum[ipx+ipy] + sum[inx+iny]
- sum[inx+ipy] - sum[ipx+iny];
*dst++ = SkToU8((outerSum * outerScale
+ innerSum * innerScale) >> 24);
prev_x += 1;
next_x += 1;
}
prev_y += 1;
next_y += 1;
}
}
/**
* sw and sh are the width and height of the src. Since the sum buffer
* matches that, but has an extra row and col at the beginning (with zeros),
* we can just use sw and sh as our "max" values for pinning coordinates
* when sampling into sum[][]
*
* The inner loop is conceptually simple; we break it into several sections
* to improve performance. Here's the original version:
for (int x = 0; x < dw; ++x) {
int px = SkClampPos(prev_x);
int nx = SkFastMin32(next_x, sw);
uint32_t tmp = sum[px+py] + sum[nx+ny] - sum[nx+py] - sum[px+ny];
*dst++ = SkToU8(tmp * scale >> 24);
prev_x += 1;
next_x += 1;
}
* The sections are:
* left-hand section, where prev_x is clamped to 0
* center section, where neither prev_x nor next_x is clamped
* right-hand section, where next_x is clamped to sw
* On some operating systems, the center section is unrolled for additional
* speedup.
*/
static void apply_kernel(uint8_t dst[], int rx, int ry, const uint32_t sum[],
int sw, int sh) {
if (2*rx > sw) {
kernel_clamped(dst, rx, ry, sum, sw, sh);
return;
}
uint32_t scale = (1 << 24) / ((2*rx + 1)*(2*ry + 1));
int sumStride = sw + 1;
int dw = sw + 2*rx;
int dh = sh + 2*ry;
int prev_y = -2*ry;
int next_y = 1;
SkASSERT(2*rx <= dw - 2*rx);
for (int y = 0; y < dh; ++y) {
int py = SkClampPos(prev_y) * sumStride;
int ny = SkFastMin32(next_y, sh) * sumStride;
int prev_x = -2*rx;
int next_x = 1;
int x = 0;
for (; x < 2*rx; ++x) {
SkASSERT(prev_x <= 0);
SkASSERT(next_x <= sw);
int px = 0;
int nx = next_x;
uint32_t tmp = sum[px+py] + sum[nx+ny] - sum[nx+py] - sum[px+ny];
*dst++ = SkToU8(tmp * scale >> 24);
prev_x += 1;
next_x += 1;
}
int i0 = prev_x + py;
int i1 = next_x + ny;
int i2 = next_x + py;
int i3 = prev_x + ny;
#if UNROLL_KERNEL_LOOP
for (; x < dw - 2*rx - 4; x += 4) {
SkASSERT(prev_x >= 0);
SkASSERT(next_x <= sw);
uint32_t tmp = sum[i0++] + sum[i1++] - sum[i2++] - sum[i3++];
*dst++ = SkToU8(tmp * scale >> 24);
tmp = sum[i0++] + sum[i1++] - sum[i2++] - sum[i3++];
*dst++ = SkToU8(tmp * scale >> 24);
tmp = sum[i0++] + sum[i1++] - sum[i2++] - sum[i3++];
*dst++ = SkToU8(tmp * scale >> 24);
tmp = sum[i0++] + sum[i1++] - sum[i2++] - sum[i3++];
*dst++ = SkToU8(tmp * scale >> 24);
prev_x += 4;
next_x += 4;
}
#endif
for (; x < dw - 2*rx; ++x) {
SkASSERT(prev_x >= 0);
SkASSERT(next_x <= sw);
uint32_t tmp = sum[i0++] + sum[i1++] - sum[i2++] - sum[i3++];
*dst++ = SkToU8(tmp * scale >> 24);
prev_x += 1;
next_x += 1;
}
for (; x < dw; ++x) {
SkASSERT(prev_x >= 0);
SkASSERT(next_x > sw);
int px = prev_x;
int nx = sw;
uint32_t tmp = sum[px+py] + sum[nx+ny] - sum[nx+py] - sum[px+ny];
*dst++ = SkToU8(tmp * scale >> 24);
prev_x += 1;
next_x += 1;
}
prev_y += 1;
next_y += 1;
}
}