[Pixman] [PATCH 0/4] ARM: NEON: Pipelined bilinear scanline functions

Taekyun Kim podain77 at gmail.com
Wed Sep 21 01:38:01 PDT 2011


From: Taekyun Kim <tkq.kim at samsung.com>

Hi, all

Bilinear functions in pixman-arm-neon-asm-bilinear.S have lots of room
for improvement. With some clean-up to bring tail/head scheme for them
and did instruction scheduling for mostly used over_8888_8888. Passed
make check.

I'm not sure that the scheduling is done nicely, but at least I've got
speed up on both cortex-a8 and a9 devices.

Performance of before/after on cortex-a8 @ 1GHz

<< 2000 x 2000 with scale factor close to 1.x >>
before : 39.71 Mpix/s
after  : 60.39 Mpix/s

Performance of before/after on cortex-a9 @ 1.2GHz

<< 2000 x 2000 with scale factor close to 1.x >>
before : 43.31 Mpix/s
after  : 65.83 Mpix/s

I will do optimization of other functions too based on the perfstat
results of popular cairo traces.

--
Best Regards,
Taekyun Kim

Taekyun Kim (4):
  ARM: NEON: Some cleanup of bilinear scanline functions
  ARM: NEON: Bilinear macro template for instruction scheduling
  ARM: NEON: Replace old bilinear scanline generator with new template
  ARM: NEON: Instruction scheduling of bilinear over_8888_8888

 pixman/pixman-arm-neon-asm-bilinear.S |  766 ++++++++++++++++++++++++++-------
 1 files changed, 605 insertions(+), 161 deletions(-)



More information about the Pixman mailing list