Hi Michael,

2018-02-14 8:49 GMT+01:00 Michael Olbrich <m.olbrich@pengutronix.de>:
On Tue, Feb 13, 2018 at 05:12:14PM +0100, Guillermo Rodriguez Garcia wrote:
> 2018-02-13 16:37 GMT+01:00 Michael Olbrich <m.olbrich@pengutronix.de>:
> > On Tue, Feb 13, 2018 at 10:23:09AM +0100, Guillermo Rodriguez Garcia wrote:
> > > I need to use libjpeg-turbo for some apps that rely on specific
> > > libjpeg-turbo extensions that are not available in the standard (IJG's)
> > > libjpeg library. So I started to create a set of ptxdist rules for
> > > libjpeg-turbo.
> > >
> > > libjpeg-turbo is supposed to provide a drop-in replacement for the
> > standard
> > > IJG's libjpeg, and in fact many Linux distributions have already switched
> > > to it. So most packages that currently depend on LIBJPEG should be able
> > to
> > > use libjpeg-turbo as well without any changes.
> > >
> > > Example: gst-plugins-good1.in currently does the following:
> > >
> > > select LIBJPEG if GST_PLUGINS_GOOD1_JPEG
> > >
> > > However, gst-plugins-good1 should work just fine with libjpeg-turbo as
> > > well. In fact the upstream gst-plugins already switched to libjpeg-turbo
> > > some time ago (see the Cerbero scripts)
> > >
> > > So, what is the recommended way to approach this in ptxdist? I can
> > imagine
> > > some options:
> > >
> > > 1. Provide a separate set of .in/.make files for libjpeg-turbo and a
> > > separate LIBJPEG_TURBO package. So new packages that want to use
> > > libjpeg-turbo would need to explicitly select LIBJPEG_TURBO (instead of
> > > LIBJPEG). But this means that packages that currently use libjpeg would
> > not
> > > be able to benefit from the performance improvements in libjpeg-turbo.
> > Also
> > > I am not sure whether these two libraries (libjpeg and libjpeg-turbo) can
> > > actually coexist in the same system.
> >
> > No.
> >
> > > 2. Modify the LIBJPEG rules to provide an option to select between the
> > > standard (IJG) libjpeg, or libjpeg-turbo. This would be similar to what
> > is
> > > done in alsa.in/alsa.make to select between the "full" and "light"
> > versions
> > > of the library.
> >
> > Either this ...
> >
> > > 3. Other?
> >
> > ... or change the libjpeg rule to always build libjpeg-turbo instead.
> >
> > I would prefer to avoid an option. But I'm not sure
> > 1. if there are use-cases where libjpeg is preferable
> > 2. if there are packages that fail to build with libjpeg-turbo
> >
>
> Judging from the wide adoption of libjpeg-turbo by many popular Linux
> distros, I'd say that if we had to choose one, libjpeg-turbo may actually
> be the "safe" choice.
>
>
> >
> > I have some libjpeg-turbo stuff lying around here somewhere. It was just
> > never finished. I've just dumped it in my test BSP so see if anythings
> > fails to build. I'll report the result later.
> >
>
> Let me know if you need help; I am also working on this now (since I do
> have an app that explicitly needs libjpeg-turbo and doesn't work with the
> stock libjpeg)

My test build looks good. No failed packages. So I pushed what I have to
master. Please test it and compare it with what you have so far.

No build issues here and I can confirm that both gst-plugins-good1 (with the jpeg plugin) and imagemagick both work fine with libjpeg-turbo on my target platform.

Some comments:

- According to the documentation (https://github.com/libjpeg-turbo/libjpeg-turbo/blob/master/BUILDING.md), NASM or YASM are required if building x86 or x86-64 SIMD extensions. I have not verified this ,yself (I'm building for an ARM target) but since we are enabling SIMD, shouldn't the .in file select HOST_NASM? (ideally for ARCH_X86 only but it seems that this can't be checked in the .in file?)

- For ARM targets, --with-simd requires NEON support. Perhaps the .make file should do something like this:

# For ARM, enable SIMD extensions only if NEON is available
ifdef PTXCONF_ARCH_ARM_NEON
LIBJPEG_CONF_OPT += --with-simd
endif
# For architectures other than ARM, enable SIMD extensions unconditionally
ifndef PTXCONF_ARCH_ARM
LIBJPEG_CONF_OPT += --with-simd
endif


Best regards,

Guillermo Rodriguez Garcia
guille.rodriguez@gmail.com