489148 - Implement ARM NEON optimized colorspace conversion for JPEG decoding

Status: RESOLVED DUPLICATE of bug 698519

(Core :: Graphics: ImageLib, enhancement)

Description

[Ilpo Ruotsalainen]

User-Agent:       Opera/9.63 (X11; Linux i686; U; en) Presto/2.1.1
Build Identifier: 

Implementing the YCC->ARGB colorspace conversion with ARM NEON instructions can speed up decoding JPEG images a lot.

Reproducible: Always

Comment 1

[Ilpo Ruotsalainen]

Attachment: WIP patch

This patch is by no means complete, the routine itself works (accuracy is slightly worse than the C code, as with the SSE2 version), but there's no NEON detection around. The assembler is optimized for the SnapDragon CPU but is very fast on Cortex-A8 also.

Benchmark results:
without patch:  93 runs, min/max/avg/stddev: 368.70/410.75/399.10/9.29
with patch:    643 runs, min/max/avg/stddev: 314.13/356.16/345.72/5.94

Would love some feedback on the detection part, is runtime detection wanted?

Comment 2

[Jeff Muizelaar [:jrmuizel]]

We're going to have some runtime detection code for pixman. See bug 482676. I'm not sure the best way to share it, but it seems like it would be valuable.

Comment 3

[Justin Lebar (not reading bugmail)]

I'm trying to get all (or at least more of) our CPU/compiler feature detection into xpcom/glue/SSE.h (which may need a name change).  Perhaps we could put whatever detection we need in there, too.

Ilpo, are you still interested in pushing this patch forward?  If not, I might pick it up.

Comment 4

[Virtual_ManPL [:Virtual] 🇵🇱 - (please needinfo? me - so I will see your comment/reply/question/etc.)]

What are we doing with this bug since bug #573948 landed ?

Comment 5

[Justin Lebar (not reading bugmail)]

(In reply to comment #4)
> What are we doing with this bug since bug #573948 landed ?

bug 584652 is one way forward.  If we write NEON methods for libjpeg-turbo, we could integrate colorspace conversions into those methods.

Comment 6

[Makoto Kato [:m_kato]]

(In reply to Justin Lebar [:jlebar] from comment #5)
> (In reply to comment #4)
> > What are we doing with this bug since bug #573948 landed ?
> 
> bug 584652 is one way forward.  If we write NEON methods for libjpeg-turbo,
> we could integrate colorspace conversions into those methods.

libjepg-turbo 1.1.90 (1.2.x or later) has several NEON optimization. Performance data is https://wiki.linaro.org/TomGall/LibJpegTurbo.

Comment 7

[Jeff Muizelaar [:jrmuizel]]

Recent libjpeg turbo has landed.