[Mesa-dev] [PATCH v2 08/13] mesa/format_utils: Add a general format conversion function

[Brian Paul]

On 09/12/2014 08:09 AM, Brian Paul wrote:
> On 09/11/2014 04:58 PM, Jason Ekstrand wrote:
>>
>>
>> On Thu, Sep 11, 2014 at 3:53 PM, Dieter Nützel <Dieter at nuetzel-hh.de
>> <mailto:Dieter at nuetzel-hh.de>> wrote:
>>
>>     Am 12.09.2014 00:31, schrieb Jason Ekstrand:
>>
>>         On Thu, Sep 11, 2014 at 2:55 PM, Dieter Nützel
>>         <Dieter at nuetzel-hh.de <mailto:Dieter at nuetzel-hh.de>>
>>         wrote:
>>
>>             Am 15.08.2014 04:50, schrieb Jason Ekstrand:
>>
>>                 On Aug 14, 2014 7:13 PM, "Dieter Nützel"
>>                 <Dieter at nuetzel-hh.de <mailto:Dieter at nuetzel-hh.de>>
>>                 wrote:
>>
>>
>>                     Am 15.08.2014 02:36, schrieb Dave Airlie:
>>
>>                                 On 08/02/2014 02:11 PM, Jason Ekstrand
>>                                 wrote:
>>
>>
>>
>>                                     Most format conversion operations
>>                                     required by GL can be
>>
>>                 performed by
>>
>>                                     converting one channel at a time,
>>                                     shuffling the channels
>>
>>                 around, and
>>
>>                                     optionally filling missing channels
>>                                     with zeros and ones.
>>
>>                 This
>>                 adds a
>>
>>                                     function to do just that in a
>>                                     general, yet efficient, way.
>>
>>                                     v2:
>>                                     * Add better comments including full
>>                                     docs for functions
>>                                     * Don't use __typeof__
>>                                     * Use inline helpers instead of
>>                                     writing out conversions
>>
>>                 by
>>                 hand,
>>
>>                                     * Force full loop unrolling for
>>                                     better performance
>>
>>
>>
>>                         This file seems to anger gcc a lot.
>>
>>                         It seems to take upwards of a minute or two to
>>                         compile here.
>>
>>                         gcc 4.8.3 on 32-bit x86.
>>
>>                         Dave.
>>
>>
>>
>>                     For me (on our poor little Duron 1800/2 GB) it ran ~5
>>
>>                 minutes...
>>
>>
>>                     gcc 4.8.1 on 32-bit x86.
>>
>>
>>                 If we'd like, the way the macros are set up, it would be
>>                 easy to
>>                 change it so that we do less unrolling in the cases
>>                 where we are
>>                 actually doing substantial format conversion and
>>                 wouldn't notice
>>                 the
>>                 extra logic quite as much. I'll play with it a bit
>>                 tomorrow or
>>                 next
>>                 week and see how how much of a hit we would actually
>>                 take if we
>>                 unrolled a little less in places.
>>                 --Jason Ekstrand
>>
>>
>>             Ping.
>>
>>             In a second it took 11+ minutes , here...
>>
>>
>>         11 minutes! What system are you running?  and are you using
>> -03 or
>>         something?  Yes, we can do something to cut it down, but it will
>>         probably require a configure flag; the question is what flag.
>>
>>         --Jason
>>
>>
>>     See above, the old children's system... ;-)
>>     -O2 -m32 -march=athlon-mp -mtune=athlon-mp -m3dnow -msse -mmmx
>>     -mfpmath=sse,387 -pipe
>>
>>     Bad? - Worked for ages on AthlonMP....8-)
>>     Maybe it is bad on Duron (the MP thing, much smaller cache and
>>     better GCC), now.
>>
>>     Dieter
>>
>>
>> Yeah, my recommendation would be hacking the macros to not unroll and
>> keep the patch locally.  If you've got a better idea as to how to
>> organize the code so the compiler likes it, I'm open as long as we don't
>> loose performance.
>
> It looks like a release build with MSVC is taking quite a while to
> compile this file too (actually at link time when the optimizer kicks in).
>
> But even on my fast Linux system with gcc, the difference in compile
> time between -O0 and -O3 is pretty big (2 seconds vs. 1 minute, 3 seconds).
>
> I'm still prototyping something but it looks like breaking the top-level
> switch cases in _mesa_swizzle_and_convert() into separate functions
> reduces the time quite a bit.  Let me pursue that a bit further and see
> how it goes...

OK, I'm posting a couple patches:

mesa: break up _mesa_swizzle_and_convert() to reduce compile time

This reduces -O3 compile time with gcc to 1/4 of what it was.  Seems to 
reduce compile time with MSVC too, but I haven't really measured it. 
Dieter, can you test this patch on your system?


mesa: move i, j var decls into SWIZZLE_CONVERT_LOOP() macro

I think the optimizer can sometimes do a better job when loop variables 
are declared per-loop, rather than declared per-function.  But this 
patch increases the size of the .o file from 2556528 to 2933216 bytes 
(15%).  Jason, if you have a benchmark to measure the speed of this 
code, I'd be interested to know if this patch helps much.  Moving the 
declaration of 'j' inside the 's' loop makes it even bigger (3074472 bytes).

I haven't done a full piglit run on these changes yet.

-Brian