Optimizing bitblit.c / fb_pad_*
Subsystems (2)
Hi everybody,
could you please verify this change to bitblit.c.
Advantage: For drivers with accelerated blitting a significant performance boost.
The execution time for my test3 with cyblafb and ypanning drops from 0,870s (2.6.13-rc4 with the patch from Antonino) to 0,772s.
If you agree that this is the way to go, we should add some additional code for font resolutions different from 8x16.
Should we drop fb_pad_aligned() entirely from the kernel? As far as I see it is only used in cases that do not copy too much data. Replacing fb_pad_aligned at those places with proper inline code should also help to optimize performance.
cu,
Knut
drivers/video/console/bitblit.c +0 -0
@@ -113,12 +113,15 @@
unsigned int maxcnt = info->pixmap.size/cellsize; unsigned int scan_align = info->pixmap.scan_align - 1; unsigned int buf_align = info->pixmap.buf_align - 1; - unsigned int shift_low = 0, mod = vc->vc_font.width % 8; - unsigned int shift_high = 8, pitch, cnt, size, k; + unsigned int shift_low = 0; + unsigned int mod = vc->vc_font.width % 8; unsigned int idx = vc->vc_font.width >> 3; unsigned int attribute = get_attribute(info, scr_readw(s)); + unsigned int shift_high = 8; + unsigned int pitch, cnt, size, k; + unsigned int fast_8x16, i, j; struct fb_image image;
- u8 *src, *dst, *buf = NULL;
+ u8 *src, *dst, *dstp, *buf = NULL;
if (attribute) {
buf = kmalloc(cellsize, GFP_KERNEL);
@@ -134,6 +137,11 @@
image.height = vc->vc_font.height;
image.depth = 1;+ if (vc->vc_font.height == 16 && vc->vc_font.width == 8)
+ fast_8x16 = 1;
+ else
+ fast_8x16 = 0;
+
while (count) {
if (count > maxcnt)
cnt = k = maxcnt;
@@ -147,7 +155,8 @@
size &= ~buf_align;
dst = fb_get_buffer_offset(info, &info->pixmap, size);
image.data = dst;
- if (mod) {
+
+ if (!mod)
while (k--) {
src = vc->vc_font.data + (scr_readw(s++)&
charmask)*cellsize;
@@ -157,15 +166,23 @@
src = buf;
}- fb_pad_unaligned_buffer(dst, pitch, src, idx,
- image.height, shift_high,
- shift_low, mod);
- shift_low += mod;
- dst += (shift_low >= 8) ? width : width - 1;
- shift_low &= 7;
- shift_high = 8 - shift_low;
+ // Optimized for speed, memcpy() is too slow!
+ dstp = dst;
+ if (fast_8x16)
+ for (i = 16; i--; ) {
+ *dstp = src[15-i];
+ dstp += pitch;
+ }
+ else
+ for (i = image.height; i--; ) {
+ for (j = 0; j < idx; j++)
+ dstp[j] = src[j];
+ src += idx;
+ dstp += pitch;
+ }
+ dst += width;
}
- } else {
+ else
while (k--) {
src = vc->vc_font.data + (scr_readw(s++)&
charmask)*cellsize;
@@ -175,10 +192,15 @@
src = buf;
}- fb_pad_aligned_buffer(dst, pitch, src, idx, image.height);
- dst += width;
+ fb_pad_unaligned_buffer(dst, pitch, src, idx,
+ image.height, shift_high,
+ shift_low, mod);
+ shift_low += mod;
+ dst += (shift_low >= 8) ? width : width - 1;
+ shift_low &= 7;
+ shift_high = 8 - shift_low; } - } + info->fbops->fb_imageblit(info, &image); image.dx += cnt * vc->vc_font.width; count -= cnt;
-------------------------------------------------------
SF.Net email is sponsored by: Discover Easy Linux Migration Strategies from IBM. Find simple to follow Roadmaps, straightforward articles, informative Webcasts and more! Get everything you need to get up to speed, fast. http://ads.osdn.com/?ad_id=7477&alloc_id=16492&op=click
Quoted 75 lines · Knut Petersen
Hi everybody,
could you please verify this change to
bitblit.c.Advantage: For drivers with accelerated blitting a significant performance boost.
The execution time for my test3 with cyblafb and ypanning drops from 0,870s (2.6.13-rc4 with the patch from Antonino) to 0,772s.
If you agree that this is the way to go, we should add some additional code for font resolutions different from 8x16.
Should we drop
fb_pad_aligned()entirely from the kernel? As far as I see it is only
used in cases that do not copy too much data. Replacingfb_pad_alignedat those places with proper inline code should also help to optimize performance.cu,
Knut---
linux-2.6.13-rc4/drivers/video/console/bitblit.c2005-07-29
10:01:01.000000000 +0200
+++linux/drivers/video/console/bitblit.c2005-07-30
09:34:49.000000000 +0200(quoted diff) +13
-8@@ -113,12 +113,15 @@ unsigned int maxcnt = info->pixmap.size/cellsize; unsigned int scan_align = info->pixmap.scan_align - 1; unsigned int buf_align = info->pixmap.buf_align - 1; - unsigned int shift_low = 0, mod = vc->vc_font.width % 8; - unsigned int shift_high = 8, pitch, cnt, size, k; + unsigned int shift_low = 0; + unsigned int mod = vc->vc_font.width % 8; unsigned int idx = vc->vc_font.width >> 3; unsigned int attribute = get_attribute(info, scr_readw(s)); + unsigned int shift_high = 8; + unsigned int pitch, cnt, size, k; + unsigned int fast_8x16, i, j; struct fb_image image; - u8 *src, *dst, *buf = NULL; + u8 *src, *dst, *dstp, *buf = NULL; if (attribute) { buf = kmalloc(cellsize, GFP_KERNEL);@@ -134,6 +137,11 @@ image.height = vc->vc_font.height; image.depth = 1; + if (vc->vc_font.height == 16 && vc->vc_font.width == 8) + fast_8x16 = 1; + else + fast_8x16 = 0; + while (count) { if (count > maxcnt) cnt = k = maxcnt;@@ -147,7 +155,8 @@ size &= ~buf_align; dst = fb_get_buffer_offset(info, &info->pixmap, size); image.data = dst; - if (mod) { + + if (!mod) while (k--) { src = vc->vc_font.data + (scr_readw(s++)& charmask)*cellsize;@@ -157,15 +166,23 @@ src = buf; } - fb_pad_unaligned_buffer(dst, pitch, src, idx, - image.height, shift_high, - shift_low, mod); - shift_low += mod;
Review line 175
Quoted 10 lines
(quoted diff) +7
-3- dst += (shift_low >= 8) ? width : width - 1; - shift_low &= 7; - shift_high = 8 - shift_low; + // Optimized for speed, memcpy() is too slow! + dstp = dst; + if (fast_8x16) + for (i = 16; i--; ) { + *dstp = src[15-i]; + dstp += pitch; + }
why not make it to something like this?
for (i = cellsize; i--;) {
*dstp = src[cellsize-1-i];
dstp += pitch;
}This way, it will work with any fontsize, as long as the width is a 8-bit size-aligned. You can do the same with fb_pad_aligned_buffer(), so the rest can benefit too.
The above is probably a bit slower than your version, but it's a good compromise -- bit_putcs is ugly enough, the price for the optimization.
Tony
PS: Next time you submit a patch, add a Signed-off-line. See this document by akpm:
http://www.zip.com.au/~akpm/linux/patches/stuff/tpp.txt
-------------------------------------------------------
SF.Net email is sponsored by: Discover Easy Linux Migration Strategies from IBM. Find simple to follow Roadmaps, straightforward articles, informative Webcasts and more! Get everything you need to get up to speed, fast. http://ads.osdn.com/?ad_id=7477&alloc_id=16492&op=click
Hi Antonino,
Quoted 9 lines
Review
Quoted 5 lines
(quoted diff) +5
-0+ if (fast_8x16) + for (i = 16; i--; ) { + *dstp = src[15-i]; + dstp += pitch; + }why not make it to something like this?
Quoted 9 lines
for (i = cellsize; i--;) { *dstp = src[cellsize-1-i]; dstp += pitch; }This way, it will work with any fontsize, as long as the width is a 8-bit size-aligned. You can do the same with
fb_pad_aligned_buffer(), so the rest can benefit too.
fb_pad_unaligned() could be optimized localy, there are other cases of memcpy(), e.g. in softcursor, etc.
Well, this was nothing but a quick ´n dirty hack, I tried to find out if there is still potential for optimizations. If people agree that the gain in speed justifies it, I will try to look at those parts of the kernel _after_ I have submitted my cyberblade/i1 driver. I have to admit that I ´m not familar (enough) with the layers above the hardware driver level right now.
Quoted 1 line
PS: Next time you submit a patch, add a Signed-off-line.
ack. btw, should a new driver first be presented in [email protected] or lkml?
cu,
Knut
-------------------------------------------------------
SF.Net email is sponsored by: Discover Easy Linux Migration Strategies from IBM. Find simple to follow Roadmaps, straightforward articles, informative Webcasts and more! Get everything you need to get up to speed, fast. http://ads.osdn.com/?ad_id=7477&alloc_id=16492&op=click
Quoted 5 lines · Knut Petersen
Quoted 1 line
PS: Next time you submit a patch, add a Signed-off-line.
ack. btw, should a new driver first be presented in [email protected] or lkml?
Always include linux-fbdev-devel. But you may CC lkml.
Tony
-------------------------------------------------------
SF.Net email is sponsored by: Discover Easy Linux Migration Strategies from IBM. Find simple to follow Roadmaps, straightforward articles, informative Webcasts and more! Get everything you need to get up to speed, fast. http://ads.osdn.com/?ad_id=7477&alloc_id=16492&op=click
Quoted 16 lines
why not make it to something like this?
for (i = cellsize; i--;) { *dstp = src[cellsize-1-i]; dstp += pitch; }This way, it will work with any fontsize, as long as the width is a 8-bit size-aligned. You can do the same with
fb_pad_aligned_buffer(), so the rest can benefit too.The above is probably a bit slower than your version, but it's a good compromise --
bit_putcsis ugly enough, the price for the optimization.
Quoted 6 lines
Tony
PS: Next time you submit a patch, add a Signed-off-line. See this document by akpm:
If you embedded everything then there is no reason to do massivce copies. Right now this is not the highest priority. We have alot of other fixes that need to go in.
-------------------------------------------------------
SF.Net email is sponsored by: Discover Easy Linux Migration Strategies from IBM. Find simple to follow Roadmaps, straightforward articles, informative Webcasts and more! Get everything you need to get up to speed, fast. http://ads.osdn.com/?ad_id=7477&alloc_id=16492&op=click