fbdev: Improve performance of sys_imageblit()
authorThomas Zimmermann <tzimmermann@suse.de>
Wed, 23 Feb 2022 19:38:01 +0000 (20:38 +0100)
committerThomas Zimmermann <tzimmermann@suse.de>
Wed, 2 Mar 2022 19:20:46 +0000 (20:20 +0100)
commit6f29e04938bf509fccfad490a74284cf158891ce
tree29ee3fde14a288d51e34cf9bdbd03702bee641fa
parent7dbc515f5ca4b7867e34c0c4379591cb8b47d64f
fbdev: Improve performance of sys_imageblit()

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.

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().

v2:
* move switch out of inner loop (Gerd)
* remove test for alignment of dst1 (Sam)

Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Javier Martinez Canillas <javierm@redhat.com>
Acked-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20220223193804.18636-3-tzimmermann@suse.de
drivers/video/fbdev/core/sysimgblt.c