961759 - Add NEON optimizations for Moz2D filters

Status: ASSIGNED

(Core :: Graphics, defect)

Comment 1

[Markus Stange [:mstange]]

Attachment: wip

Work left to do:
 - cleanup
 - Maybe add an alternative implementation for color matrix that maps better to NEON and does less things that are only map well to SSE2
 - Find out why fopen("/proc/cpuinfo", "r") fails on my unrooted Nexus 4 with Android 4.4.2 - it only fails when called from Factory.cpp / FactorHasNEON.c, but it works when called from pixman-arm.c, and I don't know why
 - Remove as much of the ApplyComposition code as possible and use DrawTargetCairo instead, which should already have fast paths for this
 - For every NEON implementation, verify that it's actually faster than using the scalar code

Comment 2

[Markus Stange [:mstange]]

Attachment: part 1: Add Factory::HasNeon()

Comment 3

[Markus Stange [:mstange]]

Attachment: part 2: Refactor color matrix filter

Comment 4

[Markus Stange [:mstange]]

Attachment: part 3: Add NEON support

The B2G ICS Emulator failures of feDisplacementMap-alpha-01.svg and feTurbulence-offset.svg are slightly worrying, but I haven't been able to reproduce them on my Flame. While unlikely, I hope they're caused by an emulator bug. https://tbpl.mozilla.org/?tree=Try&rev=5ef38ec96df7

Comment 5

[Markus Stange [:mstange]]

Comment on attachment 8498675 [details] [diff] [review]
part 3: Add NEON support

Review of attachment 8498675 [details] [diff] [review]:
-----------------------------------------------------------------

::: gfx/2d/FilterProcessing.cpp
@@ +52,1 @@
>    }

oops, will fix this

Comment 6

[Bas Schouten (:bas.schouten)]

Comment on attachment 8498675 [details] [diff] [review]
part 3: Add NEON support

Review of attachment 8498675 [details] [diff] [review]:
-----------------------------------------------------------------

This code is fairly hard to review, make sure it's well tested! Since with the current patch it's never used ;-).

::: gfx/2d/FilterProcessing.cpp
@@ +52,1 @@
>    }

good ;) I suspect this will make Neon usage really really rare ;-) *starts looking for a processor with SSE2 and Neon support*