Re: [PATCH 1/2] fbdev: Improve performance of sys_fillrect()
Hi Thomas,
On Thu, Feb 17, 2022 at 11:34:04AM +0100, Thomas Zimmermann wrote:
> Improve the performance of sys_fillrect() by using word-aligned
> 32/64-bit mov instructions. While the code tried to implement this,
> the compiler failed to create fast instructions. The resulting
> binary instructions were even slower than cfb_fillrect(), which
> uses the same algorithm, but operates on I/O memory.
>
> A microbenchmark measures the average number of CPU cycles
> for sys_fillrect() after a stabilizing period of a few minutes
> (i7-4790, FullHD, simpledrm, kernel with debugging). The value
> for CFB is given as a reference.
>
> sys_fillrect(), new: 26586 cycles
> sys_fillrect(), old: 166603 cycles
> cfb_fillrect(): 41012 cycles
>
> In the optimized case, sys_fillrect() is now ~6x faster than before
> and ~1.5x faster than the CFB implementation.
>
> Signed-off-by: Thomas Zimmermann
Nice optimization.
Reviewed-by: Sam Ravnborg
> ---
> drivers/video/fbdev/core/sysfillrect.c | 16 +++-
> 1 file changed, 3 insertions(+), 13 deletions(-)
>
> diff --git a/drivers/video/fbdev/core/sysfillrect.c
> b/drivers/video/fbdev/core/sysfillrect.c
> index 33ee3d34f9d2..bcdcaeae6538 100644
> --- a/drivers/video/fbdev/core/sysfillrect.c
> +++ b/drivers/video/fbdev/core/sysfillrect.c
> @@ -50,19 +50,9 @@ bitfill_aligned(struct fb_info *p, unsigned long *dst, int
> dst_idx,
>
> /* Main chunk */
> n /= bits;
> - while (n >= 8) {
> - *dst++ = pat;
> - *dst++ = pat;
> - *dst++ = pat;
> - *dst++ = pat;
> - *dst++ = pat;
> - *dst++ = pat;
> - *dst++ = pat;
> - *dst++ = pat;
> - n -= 8;
> - }
> - while (n--)
> - *dst++ = pat;
> + memset_l(dst, pat, n);
> + dst += n;
> +
> /* Trailing bits */
> if (last)
> *dst = comp(pat, *dst, last);
> --
> 2.34.1
Re: [PATCH 1/2] fbdev: Improve performance of sys_fillrect()
Hello Thomas,
On 2/17/22 11:34, Thomas Zimmermann wrote:
> Improve the performance of sys_fillrect() by using word-aligned
> 32/64-bit mov instructions. While the code tried to implement this,
> the compiler failed to create fast instructions. The resulting
> binary instructions were even slower than cfb_fillrect(), which
> uses the same algorithm, but operates on I/O memory.
>
> A microbenchmark measures the average number of CPU cycles
> for sys_fillrect() after a stabilizing period of a few minutes
> (i7-4790, FullHD, simpledrm, kernel with debugging). The value
> for CFB is given as a reference.
>
> sys_fillrect(), new: 26586 cycles
> sys_fillrect(), old: 166603 cycles
> cfb_fillrect(): 41012 cycles
>
> In the optimized case, sys_fillrect() is now ~6x faster than before
> and ~1.5x faster than the CFB implementation.
>
Wow, that's a big speedup!
> Signed-off-by: Thomas Zimmermann
> ---
> drivers/video/fbdev/core/sysfillrect.c | 16 +++-
> 1 file changed, 3 insertions(+), 13 deletions(-)
>
> diff --git a/drivers/video/fbdev/core/sysfillrect.c
> b/drivers/video/fbdev/core/sysfillrect.c
> index 33ee3d34f9d2..bcdcaeae6538 100644
> --- a/drivers/video/fbdev/core/sysfillrect.c
> +++ b/drivers/video/fbdev/core/sysfillrect.c
> @@ -50,19 +50,9 @@ bitfill_aligned(struct fb_info *p, unsigned long *dst, int
> dst_idx,
>
> /* Main chunk */
> n /= bits;
> - while (n >= 8) {
> - *dst++ = pat;
> - *dst++ = pat;
> - *dst++ = pat;
> - *dst++ = pat;
> - *dst++ = pat;
> - *dst++ = pat;
> - *dst++ = pat;
> - *dst++ = pat;
> - n -= 8;
> - }
> - while (n--)
> - *dst++ = pat;
> + memset_l(dst, pat, n);
> + dst += n;
> +
Also the code is much more simpler / easy to read now. Amazing patch.
Reviewed-by: Javier Martinez Canillas
Best regards,
--
Javier Martinez Canillas
Linux Engineering
Red Hat
[PATCH 1/2] fbdev: Improve performance of sys_fillrect()
Improve the performance of sys_fillrect() by using word-aligned
32/64-bit mov instructions. While the code tried to implement this,
the compiler failed to create fast instructions. The resulting
binary instructions were even slower than cfb_fillrect(), which
uses the same algorithm, but operates on I/O memory.
A microbenchmark measures the average number of CPU cycles
for sys_fillrect() after a stabilizing period of a few minutes
(i7-4790, FullHD, simpledrm, kernel with debugging). The value
for CFB is given as a reference.
sys_fillrect(), new: 26586 cycles
sys_fillrect(), old: 166603 cycles
cfb_fillrect(): 41012 cycles
In the optimized case, sys_fillrect() is now ~6x faster than before
and ~1.5x faster than the CFB implementation.
Signed-off-by: Thomas Zimmermann
---
drivers/video/fbdev/core/sysfillrect.c | 16 +++-
1 file changed, 3 insertions(+), 13 deletions(-)
diff --git a/drivers/video/fbdev/core/sysfillrect.c
b/drivers/video/fbdev/core/sysfillrect.c
index 33ee3d34f9d2..bcdcaeae6538 100644
--- a/drivers/video/fbdev/core/sysfillrect.c
+++ b/drivers/video/fbdev/core/sysfillrect.c
@@ -50,19 +50,9 @@ bitfill_aligned(struct fb_info *p, unsigned long *dst, int
dst_idx,
/* Main chunk */
n /= bits;
- while (n >= 8) {
- *dst++ = pat;
- *dst++ = pat;
- *dst++ = pat;
- *dst++ = pat;
- *dst++ = pat;
- *dst++ = pat;
- *dst++ = pat;
- *dst++ = pat;
- n -= 8;
- }
- while (n--)
- *dst++ = pat;
+ memset_l(dst, pat, n);
+ dst += n;
+
/* Trailing bits */
if (last)
*dst = comp(pat, *dst, last);
--
2.34.1
