Add Neon-based memset32/16 for ARM

Add versions of memset32 and memset16 that utilize the Neon registers for ARM processors.

[pgalizia, 2010-06-02 13:37:32]

Updated headers to Apache license

[pgalizia, 2010-06-02 13:38:19]

Added Neon-register optimized versions for ARM.

[agl, 2010-06-03 21:32:07]

I don't read ARM assembly so I'll pass this by someone who might. Otherwise,
I'll trust that it works and thus I think it's good to go.

You've signed either the individual or corporate copyright agreement, right?

[pgalizia, 2010-06-03 22:51:49]

On 2010/06/03 21:32:07, agl wrote:
> I don't read ARM assembly so I'll pass this by someone who might. Otherwise,
> I'll trust that it works and thus I think it's good to go.
> 
> You've signed either the individual or corporate copyright agreement, right?

Corporate copyright agreement.  Thanks for passing that code on.

[agl, 2010-06-04 14:48:09]

landed as Skia r573

[Sledge, 2010-07-21 14:30:33]

On 2010/06/04 14:48:09, agl wrote:
> landed as Skia r573

Hmmm. Looking at the memset32 code here, it's running slower for me on a
Versatile Express board than the normal portable code. pgalizia: what hardware
did you test on please?

I've also found some some normal optimised ARM code (non-neon) which runs a LOT
faster for me. See

http://www.netmite.com/android/mydroid/1.6/system/core/libcutils/memset32.S

for the code. Might be worth pulling into upstream skia?

[agl, 2010-07-21 15:43:58]

On Wed, Jul 21, 2010 at 10:30 AM,  <EinvalSledge@googlemail.com> wrote:
> Hmmm. Looking at the memset32 code here, it's running slower for me on a
> Versatile Express board than the normal portable code. pgalizia: what
> hardware did you test on please?
>
> I've also found some some normal optimised ARM code (non-neon) which
> runs a LOT faster for me. See
>
> http://www.netmite.com/android/mydroid/1.6/system/core/libcutils/memset32.S
>
> for the code. Might be worth pulling into upstream skia?

I pretty much have to depend on others to tell me what the best code
for ARM is. I don't have an ARM build nor ARM hardware.

You're also welcome to submit patches.


AGL

[reed, 2010-07-22 12:18:23]

I suggest we submit both version to the tree: neon and non-neon-ARM,
in separate files. That way a given platform can just modify their
Makefile to pull in which ever one is deemed fastest for them.

On Wed, Jul 21, 2010 at 11:43 AM, Adam Langley <agl@chromium.org> wrote:
> On Wed, Jul 21, 2010 at 10:30 AM,  <EinvalSledge@googlemail.com> wrote:
>> Hmmm. Looking at the memset32 code here, it's running slower for me on a
>> Versatile Express board than the normal portable code. pgalizia: what
>> hardware did you test on please?
>>
>> I've also found some some normal optimised ARM code (non-neon) which
>> runs a LOT faster for me. See
>>
>> http://www.netmite.com/android/mydroid/1.6/system/core/libcutils/memset32.S
>>
>> for the code. Might be worth pulling into upstream skia?
>
> I pretty much have to depend on others to tell me what the best code
> for ARM is. I don't have an ARM build nor ARM hardware.
>
> You're also welcome to submit patches.
>
>
> AGL
>

[pgalizia, 2010-07-22 14:17:31]

That works, considering that different ARM systems may have different
optimal configurations.

We were testing against Qualcomm Snapdragon 8650 for that particular update.
It'll be interesting to see how that other version works on that platform as
well.

Thanks,
Pat

> -----Original Message-----
> From: Mike Reed [mailto:reed@android.com]
> Sent: Thursday, July 22, 2010 8:18 AM
> To: Adam Langley
> Cc: pgalizia@codeaurora.org; EinvalSledge@googlemail.com;
> reply@codereview.appspotmail.com
> Subject: Re: Add Neon-based memset32/16 for ARM (issue1157045)
> 
> I suggest we submit both version to the tree: neon and non-neon-ARM,
> in separate files. That way a given platform can just modify their
> Makefile to pull in which ever one is deemed fastest for them.
> 
> On Wed, Jul 21, 2010 at 11:43 AM, Adam Langley <agl@chromium.org>
> wrote:
> > On Wed, Jul 21, 2010 at 10:30 AM,  <EinvalSledge@googlemail.com>
> wrote:
> >> Hmmm. Looking at the memset32 code here, it's running slower for me
> on a
> >> Versatile Express board than the normal portable code. pgalizia:
> what
> >> hardware did you test on please?
> >>
> >> I've also found some some normal optimised ARM code (non-neon) which
> >> runs a LOT faster for me. See
> >>
> >>
> http://www.netmite.com/android/mydroid/1.6/system/core/libcutils/memset
> 32.S
> >>
> >> for the code. Might be worth pulling into upstream skia?
> >
> > I pretty much have to depend on others to tell me what the best code
> > for ARM is. I don't have an ARM build nor ARM hardware.
> >
> > You're also welcome to submit patches.
> >
> >
> > AGL
> >

[Sledge, 2010-08-11 15:14:22]

New patch submitted as issue 1973042:

http://codereview.appspot.com/1973042/show