mozilla
/
pjs
зеркало из https://github.com/mozilla/pjs.git
1
0
Форкнуть 0
Код Задачи Пакеты Проекты Релизы Вики Активность

Bug 399488. Faster gaussian blur. r=longsonr@gmail.com, sr=roc@ocallahan.org, a1.9=beltzner@mozilla.com

This commit is contained in:
jwatt%jwatt.org 2008-03-17 15:06:08 +00:00
Родитель 54606fd9fd
Коммит 9130ee5b59
1 изменённых файлов: 98 добавлений и 61 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

159
content/svg/content/src/nsSVGFilters.cpp
Просмотреть файл

@ -572,13 +572,15 @@ private:
nsresult GetDXY(PRUint32 *aDX, PRUint32 *aDY, const nsSVGFilterInstance& aInstance); nsresult GetDXY(PRUint32 *aDX, PRUint32 *aDY, const nsSVGFilterInstance& aInstance);
void InflateRectForBlur(nsRect* aRect, const nsSVGFilterInstance& aInstance); void InflateRectForBlur(nsRect* aRect, const nsSVGFilterInstance& aInstance);
PRUint8 *SetupPredivide(PRUint32 size) const;
void BoxBlurH(PRUint8 *aInput, PRUint8 *aOutput, void BoxBlurH(PRUint8 *aInput, PRUint8 *aOutput,
PRInt32 aStride, const nsRect& aRegion, PRInt32 aStride, const nsRect& aRegion,
PRUint32 leftLobe, PRUint32 rightLobe); PRUint32 leftLobe, PRUint32 rightLobe, const PRUint8 *prediv);
void BoxBlurV(PRUint8 *aInput, PRUint8 *aOutput, void BoxBlurV(PRUint8 *aInput, PRUint8 *aOutput,
PRInt32 aStride, const nsRect& aRegion, PRInt32 aStride, const nsRect& aRegion,
PRUint32 topLobe, PRUint32 bottomLobe); PRUint32 topLobe, PRUint32 bottomLobe, const PRUint8 *prediv);
void GaussianBlur(PRUint8 *aInput, PRUint8 *aOutput, void GaussianBlur(PRUint8 *aInput, PRUint8 *aOutput,
nsSVGFilterResource *aFilterResource, nsSVGFilterResource *aFilterResource,
@ -684,39 +686,41 @@ nsSVGFEGaussianBlurElement::ParseAttribute(PRInt32 aNameSpaceID, nsIAtom* aName,
void void
nsSVGFEGaussianBlurElement::BoxBlurH(PRUint8 *aInput, PRUint8 *aOutput, nsSVGFEGaussianBlurElement::BoxBlurH(PRUint8 *aInput, PRUint8 *aOutput,
PRInt32 aStride, const nsRect &aRegion, PRInt32 aStride, const nsRect &aRegion,
PRUint32 leftLobe, PRUint32 rightLobe) PRUint32 leftLobe, PRUint32 rightLobe,
const PRUint8 *prediv)
{ {
PRInt32 boxSize = leftLobe + rightLobe + 1; PRInt32 boxSize = leftLobe + rightLobe + 1;
PRInt32 posStart = aRegion.x - leftLobe;
for (PRInt32 y = aRegion.y; y < aRegion.YMost(); y++) { for (PRInt32 y = aRegion.y; y < aRegion.YMost(); y++) {
PRUint32 sums[4] = {0, 0, 0, 0}; PRUint32 sums[4] = {0, 0, 0, 0};
for (PRInt32 i=0; i < boxSize; i++) { PRInt32 lineIndex = aStride * y;
PRInt32 pos = aRegion.x - leftLobe + i; for (PRInt32 i = 0; i < boxSize; i++) {
PRInt32 pos = posStart + i;
pos = PR_MAX(pos, aRegion.x); pos = PR_MAX(pos, aRegion.x);
pos = PR_MIN(pos, aRegion.XMost() - 1); pos = PR_MIN(pos, aRegion.XMost() - 1);
sums[0] += aInput[aStride*y + 4*pos ]; PRInt32 index = lineIndex + (pos << 2);
sums[1] += aInput[aStride*y + 4*pos + 1]; sums[0] += aInput[index ];
sums[2] += aInput[aStride*y + 4*pos + 2]; sums[1] += aInput[index + 1];
sums[3] += aInput[aStride*y + 4*pos + 3]; sums[2] += aInput[index + 2];
sums[3] += aInput[index + 3];
} }
for (PRInt32 x = aRegion.x; x < aRegion.XMost(); x++) { for (PRInt32 x = aRegion.x; x < aRegion.XMost(); x++) {
PRInt32 index = lineIndex + (x << 2);
aOutput[index ] = prediv[sums[0]];
aOutput[index + 1] = prediv[sums[1]];
aOutput[index + 2] = prediv[sums[2]];
aOutput[index + 3] = prediv[sums[3]];
PRInt32 tmp = x - leftLobe; PRInt32 tmp = x - leftLobe;
PRInt32 last = PR_MAX(tmp, aRegion.x); PRInt32 last = PR_MAX(tmp, aRegion.x);
PRInt32 next = PR_MIN(tmp + boxSize, aRegion.XMost() - 1); PRInt32 next = PR_MIN(tmp + boxSize, aRegion.XMost() - 1);
PRInt32 index2 = lineIndex + (next << 2);
aOutput[aStride*y + 4*x ] = sums[0]/boxSize; PRInt32 index3 = lineIndex + (last << 2);
aOutput[aStride*y + 4*x + 1] = sums[1]/boxSize; sums[0] += aInput[index2 ] - aInput[index3 ];
aOutput[aStride*y + 4*x + 2] = sums[2]/boxSize; sums[1] += aInput[index2 + 1] - aInput[index3 + 1];
aOutput[aStride*y + 4*x + 3] = sums[3]/boxSize; sums[2] += aInput[index2 + 2] - aInput[index3 + 2];
sums[3] += aInput[index2 + 3] - aInput[index3 + 3];
sums[0] += aInput[aStride*y + 4*next ] -
aInput[aStride*y + 4*last ];
sums[1] += aInput[aStride*y + 4*next + 1] -
aInput[aStride*y + 4*last + 1];
sums[2] += aInput[aStride*y + 4*next + 2] -
aInput[aStride*y + 4*last + 2];
sums[3] += aInput[aStride*y + 4*next + 3] -
aInput[aStride*y + 4*last + 3];
} }
} }
} }
@ -724,43 +728,56 @@ nsSVGFEGaussianBlurElement::BoxBlurH(PRUint8 *aInput, PRUint8 *aOutput,
void void
nsSVGFEGaussianBlurElement::BoxBlurV(PRUint8 *aInput, PRUint8 *aOutput, nsSVGFEGaussianBlurElement::BoxBlurV(PRUint8 *aInput, PRUint8 *aOutput,
PRInt32 aStride, const nsRect &aRegion, PRInt32 aStride, const nsRect &aRegion,
PRUint32 topLobe, PRUint32 bottomLobe) PRUint32 topLobe, PRUint32 bottomLobe,
const PRUint8 *prediv)
{ {
PRInt32 boxSize = topLobe + bottomLobe + 1; PRInt32 boxSize = topLobe + bottomLobe + 1;
PRInt32 posStart = aRegion.y - topLobe;
for (PRInt32 x = aRegion.x; x < aRegion.XMost(); x++) { for (PRInt32 x = aRegion.x; x < aRegion.XMost(); x++) {
PRUint32 sums[4] = {0, 0, 0, 0}; PRUint32 sums[4] = {0, 0, 0, 0};
for (PRInt32 i=0; i < boxSize; i++) { PRInt32 fourX = x << 2;
PRInt32 pos = aRegion.y - topLobe + i; for (PRInt32 i = 0; i < boxSize; i++) {
PRInt32 pos = posStart + i;
pos = PR_MAX(pos, aRegion.y); pos = PR_MAX(pos, aRegion.y);
pos = PR_MIN(pos, aRegion.YMost() - 1); pos = PR_MIN(pos, aRegion.YMost() - 1);
sums[0] += aInput[aStride*pos + 4*x ]; PRInt32 index = aStride * pos + fourX;
sums[1] += aInput[aStride*pos + 4*x + 1]; sums[0] += aInput[index ];
sums[2] += aInput[aStride*pos + 4*x + 2]; sums[1] += aInput[index + 1];
sums[3] += aInput[aStride*pos + 4*x + 3]; sums[2] += aInput[index + 2];
sums[3] += aInput[index + 3];
} }
for (PRInt32 y = aRegion.y; y < aRegion.YMost(); y++) { for (PRInt32 y = aRegion.y; y < aRegion.YMost(); y++) {
PRInt32 index = aStride * y + fourX;
aOutput[index ] = prediv[sums[0]];
aOutput[index + 1] = prediv[sums[1]];
aOutput[index + 2] = prediv[sums[2]];
aOutput[index + 3] = prediv[sums[3]];
PRInt32 tmp = y - topLobe; PRInt32 tmp = y - topLobe;
PRInt32 last = PR_MAX(tmp, aRegion.y); PRInt32 last = PR_MAX(tmp, aRegion.y);
PRInt32 next = PR_MIN(tmp + boxSize, aRegion.YMost() - 1); PRInt32 next = PR_MIN(tmp + boxSize, aRegion.YMost() - 1);
PRInt32 index2 = aStride * next + fourX;
aOutput[aStride*y + 4*x ] = sums[0]/boxSize; PRInt32 index3 = aStride * last + fourX;
aOutput[aStride*y + 4*x + 1] = sums[1]/boxSize; sums[0] += aInput[index2 ] - aInput[index3 ];
aOutput[aStride*y + 4*x + 2] = sums[2]/boxSize; sums[1] += aInput[index2 + 1] - aInput[index3 + 1];
aOutput[aStride*y + 4*x + 3] = sums[3]/boxSize; sums[2] += aInput[index2 + 2] - aInput[index3 + 2];
sums[3] += aInput[index2 + 3] - aInput[index3 + 3];
sums[0] += aInput[aStride*next + 4*x ] -
aInput[aStride*last + 4*x ];
sums[1] += aInput[aStride*next + 4*x + 1] -
aInput[aStride*last + 4*x + 1];
sums[2] += aInput[aStride*next + 4*x + 2] -
aInput[aStride*last + 4*x + 2];
sums[3] += aInput[aStride*next + 4*x + 3] -
aInput[aStride*last + 4*x + 3];
} }
} }
} }
PRUint8 *
nsSVGFEGaussianBlurElement::SetupPredivide(PRUint32 size) const
{
PRUint8 *tmp = new PRUint8[size * 256];
if (tmp) {
for (PRUint32 i = 0; i < 256; i++)
memset(tmp + i * size, i, size);
}
return tmp;
}
nsresult nsresult
nsSVGFEGaussianBlurElement::GetDXY(PRUint32 *aDX, PRUint32 *aDY, nsSVGFEGaussianBlurElement::GetDXY(PRUint32 *aDX, PRUint32 *aDY,
const nsSVGFilterInstance& aInstance) const nsSVGFilterInstance& aInstance)
@ -791,8 +808,12 @@ nsSVGFEGaussianBlurElement::GaussianBlur(PRUint8 *aInput, PRUint8 *aOutput,
nsSVGFilterResource *aFilterResource, nsSVGFilterResource *aFilterResource,
PRUint32 aDX, PRUint32 aDY) PRUint32 aDX, PRUint32 aDY)
{ {
PRUint8 *tmp = static_cast<PRUint8*> NS_ASSERTION(aDX > 0 && aDY > 0, "Invalid stdDeviation!");
(calloc(aFilterResource->GetDataSize(), 1));
nsAutoArrayPtr<PRUint8> tmp(new PRUint8[aFilterResource->GetDataSize()]);
if (!tmp)
return;
memset(tmp, 0, aFilterResource->GetDataSize());
nsRect rect = aFilterResource->GetSurfaceRect(); nsRect rect = aFilterResource->GetSurfaceRect();
#ifdef DEBUG_tor #ifdef DEBUG_tor
fprintf(stderr, "FILTER GAUSS rect: %d,%d %dx%d\n", fprintf(stderr, "FILTER GAUSS rect: %d,%d %dx%d\n",
@ -803,37 +824,53 @@ nsSVGFEGaussianBlurElement::GaussianBlur(PRUint8 *aInput, PRUint8 *aOutput,
if (aDX & 1) { if (aDX & 1) {
// odd // odd
BoxBlurH(aInput, tmp, stride, rect, aDX/2, aDX/2); nsAutoArrayPtr<PRUint8> prediv(SetupPredivide(2 * (aDX / 2) + 1));
BoxBlurH(tmp, aOutput, stride, rect, aDX/2, aDX/2); if (!prediv)
BoxBlurH(aOutput, tmp, stride, rect, aDX/2, aDX/2); return;
BoxBlurH(aInput, tmp, stride, rect, aDX/2, aDX/2, prediv);
BoxBlurH(tmp, aOutput, stride, rect, aDX/2, aDX/2, prediv);
BoxBlurH(aOutput, tmp, stride, rect, aDX/2, aDX/2, prediv);
} else { } else {
// even // even
if (aDX == 0) { if (aDX == 0) {
aFilterResource->CopyImageSubregion(tmp, aInput); aFilterResource->CopyImageSubregion(tmp, aInput);
} else { } else {
BoxBlurH(aInput, tmp, stride, rect, aDX/2, aDX/2 - 1); nsAutoArrayPtr<PRUint8> prediv(SetupPredivide(2 * (aDX / 2) + 1));
BoxBlurH(tmp, aOutput, stride, rect, aDX/2 - 1, aDX/2); nsAutoArrayPtr<PRUint8> prediv2(SetupPredivide(2 * (aDX / 2)));
BoxBlurH(aOutput, tmp, stride, rect, aDX/2, aDX/2); if (!prediv || !prediv2)
return;
BoxBlurH(aInput, tmp, stride, rect, aDX/2, aDX/2 - 1, prediv2);
BoxBlurH(tmp, aOutput, stride, rect, aDX/2 - 1, aDX/2, prediv2);
BoxBlurH(aOutput, tmp, stride, rect, aDX/2, aDX/2, prediv);
} }
} }
if (aDY & 1) { if (aDY & 1) {
// odd // odd
BoxBlurV(tmp, aOutput, stride, rect, aDY/2, aDY/2); nsAutoArrayPtr<PRUint8> prediv(SetupPredivide(2 * (aDY / 2) + 1));
BoxBlurV(aOutput, tmp, stride, rect, aDY/2, aDY/2); if (!prediv)
BoxBlurV(tmp, aOutput, stride, rect, aDY/2, aDY/2); return;
BoxBlurV(tmp, aOutput, stride, rect, aDY/2, aDY/2, prediv);
BoxBlurV(aOutput, tmp, stride, rect, aDY/2, aDY/2, prediv);
BoxBlurV(tmp, aOutput, stride, rect, aDY/2, aDY/2, prediv);
} else { } else {
// even // even
if (aDY == 0) { if (aDY == 0) {
aFilterResource->CopyImageSubregion(aOutput, tmp); aFilterResource->CopyImageSubregion(aOutput, tmp);
} else { } else {
BoxBlurV(tmp, aOutput, stride, rect, aDY/2, aDY/2 - 1); nsAutoArrayPtr<PRUint8> prediv(SetupPredivide(2 * (aDY / 2) + 1));
BoxBlurV(aOutput, tmp, stride, rect, aDY/2 - 1, aDY/2); nsAutoArrayPtr<PRUint8> prediv2(SetupPredivide(2 * (aDY / 2)));
BoxBlurV(tmp, aOutput, stride, rect, aDY/2, aDY/2); if (!prediv || !prediv2)
return;
BoxBlurV(tmp, aOutput, stride, rect, aDY/2, aDY/2 - 1, prediv2);
BoxBlurV(aOutput, tmp, stride, rect, aDY/2 - 1, aDY/2, prediv2);
BoxBlurV(tmp, aOutput, stride, rect, aDY/2, aDY/2, prediv);
} }
} }
free(tmp);
} }
nsresult nsresult
@ -4307,7 +4344,7 @@ nsSVGFEConvolveMatrixElement::Filter(nsSVGFilterInstance *instance)
if (orderX > NS_SVG_OFFSCREEN_MAX_DIMENSION || if (orderX > NS_SVG_OFFSCREEN_MAX_DIMENSION ||
orderY > NS_SVG_OFFSCREEN_MAX_DIMENSION) orderY > NS_SVG_OFFSCREEN_MAX_DIMENSION)
return NS_ERROR_FAILURE; return NS_ERROR_FAILURE;
nsAutoPtr<float> kernel(new float[orderX * orderY]); nsAutoArrayPtr<float> kernel(new float[orderX * orderY]);
if (!kernel) if (!kernel)
return NS_ERROR_FAILURE; return NS_ERROR_FAILURE;
for (PRUint32 i = 0; i < num; i++) { for (PRUint32 i = 0; i < num; i++) {