655693 - ARM NEON optimizations for libpng

Status: RESOLVED WONTFIX

(Core :: Graphics: ImageLib, enhancement)

Description

[Siarhei Siamashka]

User-Agent:       Mozilla/5.0 (X11; U; Linux x86_64; en-US; rv:1.9.1.16) Gecko/20110323 Iceweasel/3.5.16 (like Firefox/3.5.16)
Build Identifier: 

PNG preprocessing filters consume roughly half of CPU time (the rest is spent in zlib). And the preprocessors code is a perfect target for SIMD optimizations.

Some ARM NEON optimizations seem to be available here (I have not tested them myself): http://groups.google.com/a/chromium.org/group/chromium-dev/browse_thread/thread/689cea4517c20bbc

Reproducible: Always

Comment 1

[Siarhei Siamashka]

One more link which suggests that upstreaming such optimizations might be difficult: https://groups.google.com/a/chromium.org/group/chromium-os-dev/msg/bcdffb62099ac0ee

There is also a related bug 428750 (MMX/SSE2 optimizations for libpng).

Comment 2

[Boris Zbarsky [:bzbarsky]]

Alon, do you want to take a look?

Comment 3

[Boris Zbarsky [:bzbarsky]]

Also Glenn, are you just planning to take something like this upstream?

Comment 4

[Glenn Randers-Pehrson]

I've asked the PNG group to have a look.

We are not fundamentally opposed to accepting assembler
code.  We supported MMX assembler for a few years but
had to eliminate it when we realized that we did not
have an acceptable open-source license for it and apparently
could not get one.

Glenn

Comment 5

[Siarhei Siamashka]

The NEON optimized paeth filter from that patch ('paeth_filter_3bpp_loop' function) seems to perform slightly slower than C code on ARM Cortex-A8 processor in my quick test, providing no expected performance improvement.

Comment 6

[Joe Drew (not getting mail)]

Given comment 5, we won't take those NEON optimizations; since that seems to be all this bug is about, let's close this.

Comment 7

[Siarhei Siamashka]

The Qualcomm code was indeed worthless, but there are useful ARM NEON optimizations introduced in libpng-1.5.7