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 tzimmermann@suse.de

Nice optimization. Reviewed-by: Sam Ravnborg sam@ravnborg.org

---

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

Hi Sam

Am 18.02.22 um 11:14 schrieb Sam Ravnborg:

Hi Thomas,

On Thu, Feb 17, 2022 at 11:34:05AM +0100, Thomas Zimmermann wrote:

...

Improve the performance of sys_imageblit() by manually unrolling the inner blitting loop and moving some invariants out. The compiler failed to do this automatically. The resulting binary code was even slower than the cfb_imageblit() helper, which uses the same algorithm, but operates on I/O memory.

It would be super to have the same optimization done to cfb_imageblit(), to prevent that the two codebases diverge more than necessary. Also I think cfb_ version would also see a performance gain from this.

Yes, I can do that.

The actual implementation looks good. So with or without the extra un-rolling the patch is: Acked-by: Sam Ravnborg sam@ravnborg.org

One small nit belwo.

Sam

...

A microbenchmark measures the average number of CPU cycles for sys_imageblit() 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_imageblit(), new: 25934 cycles sys_imageblit(), old: 35944 cycles cfb_imageblit(): 30566 cycles

In the optimized case, sys_imageblit() is now ~30% faster than before and ~20% faster than cfb_imageblit().

Signed-off-by: Thomas Zimmermann tzimmermann@suse.de

drivers/video/fbdev/core/sysimgblt.c | 51 +++++++++++++++++++++------- 1 file changed, 39 insertions(+), 12 deletions(-)

diff --git a/drivers/video/fbdev/core/sysimgblt.c b/drivers/video/fbdev/core/sysimgblt.c index a4d05b1b17d7..d70d65af6fcb 100644 --- a/drivers/video/fbdev/core/sysimgblt.c +++ b/drivers/video/fbdev/core/sysimgblt.c @@ -188,23 +188,32 @@ static void fast_imageblit(const struct fb_image *image, struct fb_info *p, { u32 fgx = fgcolor, bgx = bgcolor, bpp = p->var.bits_per_pixel; u32 ppw = 32/bpp, spitch = (image->width + 7)/8;

• u32 bit_mask, end_mask, eorx, shift;

• u32 bit_mask, eorx; const char *s = image->data, *src; u32 *dst;

• const u32 *tab = NULL;
• int i, j, k;

• const u32 *tab;
• size_t tablen;
• u32 colortab[16];
• int i, j, k, jdecr;
• if ((uintptr_t)dst1 % 8)

• return;
  

This check is new - and should not trigger ever. Maybe add an unlikely and a WARN_ON_ONCE()?

I think I can remove this test. It was supposed to tell the compiler that dst1 is 8-aligned, but I don't think it worked.

Best regards Thomas

...

switch (bpp) { case 8: tab = fb_be_math(p) ? cfb_tab8_be : cfb_tab8_le;

• tablen = 16;
  

  break; case 16: tab = fb_be_math(p) ? cfb_tab16_be : cfb_tab16_le;

• tablen = 4;
  

  break; case 32:

• default: tab = cfb_tab32;

• tablen = 2;
  

  break;

• default:

• return;
  

  }

  for (i = ppw-1; i--; ) {

@@ -217,19 +226,37 @@ static void fast_imageblit(const struct fb_image *image, struct fb_info *p, bit_mask = (1 << ppw) - 1; eorx = fgx ^ bgx; k = image->width/ppw;

• jdecr = 8 / ppw;
• for (i = 0; i < tablen; ++i)

• colortab[i] = (tab[i] & eorx) ^ bgx;
  

This code could have been embedded with the switch (bpp) { That would have made some sense I think. But both ways works, so this was just a small observation.

Sam