[FFmpeg-devel] [PATCH] Add check for Athlon64 and similar AMD processors with slow SSE2.

[Justin Ruggles]

On 02/02/2011 06:40 PM, M?ns Rullg?rd wrote:

> Justin Ruggles <justin.ruggles at gmail.com> writes:
> 
>> ---
>> This was ported from x264 so we need permission to relicense from
>> Loren, Jason, or whoever added this particular check in x264.
> 
> I doubt there's much copyright on checking a single bit, but it's
> always nice to ask first.
> 
>> Once this is added, I can go through libavcodec/x86 and test various
>> SSE2 functions on my Athlon64 to check if they're faster with other
>> versions.  We also have a couple functions now that are hackishly
>> handled by disabling SSE2 for all AMD processors by checking the
>> 3DNow flag.  Those are the ones I'll test first.
>>
>>  libavutil/x86/cpu.c |    7 +++++++
>>  1 files changed, 7 insertions(+), 0 deletions(-)
>>
>>
>> diff --git a/libavutil/x86/cpu.c b/libavutil/x86/cpu.c
>> index 4b6cb0d..7e847c1 100644
>> --- a/libavutil/x86/cpu.c
>> +++ b/libavutil/x86/cpu.c
>> @@ -109,6 +109,13 @@ int ff_get_cpu_flags_x86(void)
>>              rval |= AV_CPU_FLAG_MMX;
>>          if (ext_caps & (1<<22))
>>              rval |= AV_CPU_FLAG_MMX2;
>> +
>> +        /* Allow for selectively disabling SSE2 functions on AMD processors
>> +           with SSE2 support but not SSE4a.  This includes Athlon64,
>> +           some Opteron, and some Sempron processors.  MMX, SSE, or 3DNow! are
>> +           faster than SSE2 often enough to utilize this special-case flag. */
>> +        if (rval & AV_CPU_FLAG_SSE2 && !(ecx & 0x00000040))
>> +            rval |= AV_CPU_FLAG_SSE2SLOW;
>>      }
>>  
>>      if (!strncmp(vendor.c, "GenuineIntel", 12) &&
> 
> Why does this code only affect AMD processors?  I guess it has
> something to do with those magic bit checks before it...


That is a guess on my part.  I don't know much about cpuid.  I assume we
wouldn't check for 3DNow on a non-AMD processor, but that may be an
incorrect assumption.  x264 also checks the AMD vendor string in their
equivalent of this whole section, but FFmpeg does not.

-Justin