451621 - Pixman optimizations for arm

Status: RESOLVED FIXED

Description

[Jeff Muizelaar [:jrmuizel]]

Created attachment 334945 [details] [diff] [review]
go faster

Make a pixman faster on arm. Also adds a faster rectilinear nearest neighbour image scaling.

Comment 1

[Jeff Muizelaar [:jrmuizel]]

Created attachment 334966 [details] [diff] [review]
go faster w/ better configury

Detect if the compiler supports ARM simd instructions

Comment 2

[Stuart Parmenter]

+ * Copyright © 208 Mozilla Corporation
that was a long time ago

Comment 3

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

It's well-aged!

Comment 4

[Jeff Muizelaar [:jrmuizel]]

Created attachment 336679 [details] [diff] [review]
arm pixman patch update

Update pixman patch to match what was submitted for upstream inclusion

Comment 5

[Jeff Muizelaar [:jrmuizel]]

Created attachment 336912 [details] [diff] [review]
arm pixman v2

Fix the configury by adding HAVE_ARM_SIMD to autoconf.mk.in

Comment 6

[Jeff Muizelaar [:jrmuizel]]

Created attachment 336920 [details] [diff] [review]
arm pixman v3

use one fewer registers in fbCompositeSolidMask_nx8x8888arm allowing compiling with worser compilers.

Comment 7

[Jeff Muizelaar [:jrmuizel]]

Created attachment 336971 [details] [diff] [review]
arm pixman v4

The last patch was broken and wrong, this patch fixes it.

Comment 8

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Created attachment 337251 [details] [diff] [review]
additional patch

Additional patch -- not ARM specific, but it's related to our arm work.  Just sticking it here so that it's in bugzilla, both of these should hopefully end up in pixman upstream shortly.  I will probably commit these both to our own repo before that happens, though, since I think we're waiting on 0.12.0 to be released before putting in the arm code.

Comment 9

[Guillaume Letellier]

Hi,
A colleague of mine had a brief look at Jeff's patch and has a few comments:
1. In the first while() loop of fbCompositeSrcAdd_8000x8000arm you can use UQADD8, it doesn't matter that you just want the bottom lane. Similarly in the third loop.

2. There are many AND operations with %[component_mask] to extract bytes. These can be replaced by UXTB16, saving the mask register.

3. There are mask operations with %[component_mask] followed by accumulates.
These can be replaced by the combined mask and accumulate UXTAB16.

Hope that helps.

Comment 10

[Jeff Muizelaar [:jrmuizel]]

Created attachment 338653 [details] [diff] [review]
arm pixman v5

This version addresses Guillaume's comments (Thanks!) and squeezes some more performance out. e.g. over8888x8888 is about 10-11% faster.

Comment 11

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Pulled this in along with a few more pixman updates to 0.12.0.

19351   5d807b616378   2008-09-17 14:15 -0700   vladimir
  b=451621; push new pixman with arm fast-paths; r=me