Reorder the separable blur passes into XXX/YYY, with an image transpose on the last pass of each group. This results in continuguous memory reads in all passes, giving a 22% speedup on theverge.skp over the previous separable implementation, and a 30%-50% improvement over the existing implementation (depending on platform).

Review URL: https://codereview.appspot.com/6851053

git-svn-id: http://skia.googlecode.com/svn/trunk@6445 2bbb7eff-a529-9590-31e7-b0007b416f81
This commit is contained in:
senorblanco@chromium.org 2012-11-15 20:27:35 +00:00
parent 0561a3ccd2
commit 908276b396

View File

@ -12,80 +12,50 @@
#include "SkTemplates.h"
#include "SkEndian.h"
static int boxBlurX(const uint8_t* src, int src_row_bytes,
uint8_t* dst, int dst_row_bytes,
int radius, int width, int height)
/**
* This function performs a box blur in X, of the given radius. If the
* "transpose" parameter is true, it will transpose the pixels on write,
* such that X and Y are swapped. Reads are always performed from contiguous
* memory in X, for speed. The destination buffer (dst) must be at least
* (width + radius * 2) * height bytes in size.
*/
static int boxBlur(const uint8_t* src, int src_y_stride, uint8_t* dst,
int radius, int width, int height, bool transpose)
{
int kernelSize = radius * 2 + 1;
int border = SkMin32(width, radius * 2);
uint32_t scale = (1 << 24) / kernelSize;
int new_width = width + radius * 2;
int dst_x_stride = transpose ? height : 1;
int dst_y_stride = transpose ? 1 : new_width;
for (int y = 0; y < height; ++y) {
int sum = 0;
uint8_t* dptr = dst + y * dst_row_bytes;
const uint8_t* sptr = src + y * src_row_bytes - radius;
uint8_t* dptr = dst + y * dst_y_stride;
const uint8_t* right = src + y * src_y_stride;
const uint8_t* left = right;
for (int x = 0; x < border; ++x) {
sum += *(sptr + radius);
*dptr++ = (sum * scale) >> 24;
sptr++;
sum += *right++;
*dptr = (sum * scale) >> 24;
dptr += dst_x_stride;
}
for (int x = width; x < radius * 2; ++x) {
*dptr++ = (sum * scale) >> 24;
sptr++;
*dptr = (sum * scale) >> 24;
dptr += dst_x_stride;
}
for (int x = radius * 2; x < width; ++x) {
sum += *(sptr + radius);
*dptr++ = (sum * scale) >> 24;
sum -= *(sptr - radius);
sptr++;
sum += *right++;
*dptr = (sum * scale) >> 24;
sum -= *left++;
dptr += dst_x_stride;
}
for (int x = 0; x < border; ++x) {
*dptr++ = (sum * scale) >> 24;
sum -= *(sptr - radius);
sptr++;
*dptr = (sum * scale) >> 24;
sum -= *left++;
dptr += dst_x_stride;
}
SkASSERT(sum == 0);
}
return width + radius * 2;
}
static int boxBlurY(const uint8_t* src, int src_row_bytes,
uint8_t* dst, int dst_row_bytes,
int radius, int width, int height)
{
int kernelSize = radius * 2 + 1;
uint32_t scale = (1 << 24) / kernelSize;
int border = SkMin32(height, radius * 2);
for (int x = 0; x < width; ++x) {
int sum = 0;
uint8_t* dptr = dst + x;
const uint8_t* sptr = src + x - radius * src_row_bytes;
for (int y = 0; y < border; ++y) {
sum += *(sptr + radius * src_row_bytes);
*dptr = (sum * scale) >> 24;
sptr += src_row_bytes;
dptr += dst_row_bytes;
}
for (int y = height; y < radius * 2; ++y) {
*dptr = (sum * scale) >> 24;
sptr += src_row_bytes;
dptr += dst_row_bytes;
}
for (int y = radius * 2; y < height; ++y) {
sum += *(sptr + radius * src_row_bytes);
*dptr = (sum * scale) >> 24;
sum -= *(sptr - radius * src_row_bytes);
sptr += src_row_bytes;
dptr += dst_row_bytes;
}
for (int y = 0; y < border; ++y) {
*dptr = (sum * scale) >> 24;
sum -= *(sptr - radius * src_row_bytes);
sptr += src_row_bytes;
dptr += dst_row_bytes;
}
SkASSERT(sum == 0);
}
return height + radius * 2;
return new_width;
}
// Unrolling the integer blur kernel seems to give us a ~15% speedup on Windows,
@ -683,13 +653,18 @@ bool SkBlurMask::Blur(SkMask* dst, const SkMask& src,
uint8_t* tp = tmpBuffer.get();
int w = sw, h = sh;
w = boxBlurX(sp, src.fRowBytes, tp, dst->fRowBytes, rx, w, h);
h = boxBlurY(tp, dst->fRowBytes, dp, dst->fRowBytes, ry, w, h);
if (quality == kHigh_Quality) {
w = boxBlurX(dp, dst->fRowBytes, tp, dst->fRowBytes, rx, w, h);
h = boxBlurY(tp, dst->fRowBytes, dp, dst->fRowBytes, ry, w, h);
w = boxBlurX(dp, dst->fRowBytes, tp, dst->fRowBytes, rx, w, h);
h = boxBlurY(tp, dst->fRowBytes, dp, dst->fRowBytes, ry, w, h);
// Do three X blurs, with a transpose on the final one.
w = boxBlur(sp, src.fRowBytes, tp, rx, w, h, false);
w = boxBlur(tp, w, dp, rx, w, h, false);
w = boxBlur(dp, w, tp, rx, w, h, true);
// Do three Y blurs, with a transpose on the final one.
h = boxBlur(tp, h, dp, ry, h, w, false);
h = boxBlur(dp, h, tp, ry, h, w, false);
h = boxBlur(tp, h, dp, ry, h, w, true);
} else {
w = boxBlur(sp, src.fRowBytes, tp, rx, w, h, true);
h = boxBlur(tp, h, dp, ry, h, w, true);
}
} else {
const size_t storageW = sw + 2 * (passCount - 1) * rx + 1;