[Openal-devel] [Fwd: [PATCH] MMX routines for mixing audio (linux) (v2)]

[James Tomaschke]

>>>I am a bit confused: Does the compiled code actually work? The
>>>pointers in the
>>>code are 32bit wide, but on x86_64 they are 64bit wide, so how can
>>>this work
>>>if you link it to 64biot code? I think _alHaveMMX() might work, but
>>>alMMXmemcpy()?
>>> 
>>>
>>>      
>>>
>>You would think it would, oh and it does segfault.  I am working on a
>>testcase for that specific routine, it's going to require "yasm -m amd64
>>-f elf" to assemble so if you can work that into the configure.ac I'll
>>try and patch the asm.
>>    
>>
>
>Here, I did a quick hacking, and I guess it should work.
>  
>
That'll run the assembler, and it's pretty much what I got.

Have you checked the MPlayer sources?  It does all this fast memcpy 
stuff as well as detection of processer during config/compile and 
runtime.  It also supports SSE2/MMX2/3DNow/MMX/fallback, in that order 
(fastest first).  They used a simple if(gCpuCaps.hasSSE2) { ... } else 
..., but I would have setup function pointers to the proper fast_memcpy 
routines on startup and get rid of the memory access/test overhead.

Their assembly is done through inline gas in libvo/aclib_template.c, but 
can be converted to NASM/YASM.

/j

-- 
James Tomaschke
Snr. Software Engineer
InfoSemiconductor, Inc.