[FFmpeg-devel] [PATCH] Set -mfpmath=sse for cpus that support it

[Måns Rullgård]

Vitor Sessak <vitor1001 at gmail.com> writes:

> M?ns Rullg?rd wrote:
>> Vitor Sessak <vitor1001 at gmail.com> writes:
>>
>>> $subj, 10% faster TwinVQ decoding.
>>>
>>> As a side note, this option is not enabled even when passing
>>> "-march=native -mtune=native" to GCC. A possible reason for it,
>>> quoting the GCC doc, is "The resulting code should be considerably
>>> faster [...] but may break some existing code that expects temporaries
>>> to be 80bit.".
>>>
>>> -Vitor
>>>
>>> Index: configure
>>> ===================================================================
>>> --- configure	(revision 19681)
>>> +++ configure	(working copy)
>>> @@ -1649,22 +1649,27 @@
>>>          cpuflags="-mcpu=cell"
>>>          enable ldbrx
>>>      ;;
>>> -    # targets that do NOT support conditional mov (cmov)
>>>      i[345]86|pentium|pentium-mmx|k6|k6-[23]|winchip-c6|winchip2|c3)
>>>          cpuflags="-march=$cpu"
>>>          disable cmov
>>>      ;;
>>> -    # targets that do support conditional mov (cmov)
>>> -    i686|pentiumpro|pentium[23]|pentium-m|athlon|athlon-tbird|athlon-4|athlon-[mx]p|athlon64|k8|opteron|athlon-fx|core2)
>>> +    i686|pentiumpro|pentium2|athlon|athlon-tbird)
>>>          cpuflags="-march=$cpu"
>>>          enable cmov
>>>          enable fast_cmov
>>>      ;;
>>> +    pentium3|pentium-m|athlon-4|athlon-[mx]p|athlon64|k8|opteron|athlon-fx|core2)
>>> +        cpuflags="-march=$cpu"
>>> +        enable cmov
>>> +        enable fast_cmov
>>> +        check_cflags -mfpmath=sse
>>> +    ;;
>>>      # targets that do support conditional mov but on which it's slow
>>>      pentium4|pentium4m|prescott|nocona)
>>>          cpuflags="-march=$cpu"
>>>          enable cmov
>>>          disable fast_cmov
>>> +        check_cflags -mfpmath=sse
>>>      ;;
>>>      sparc64)
>>>          cpuflags="-mcpu=v9"
>> You can't use check_cflags there.  Is there any reason to not attempt
>> using this flag with --cpu values?  I imagine a simple check_cflags
>> line together with the other x86 stuff (ebx checks etc) should be
>> enough.
>
> I can't just call "check_cflags -mfpmath=sse" for every x86 CPU
> because GCC does not fail with "-march=386 -mfpmath=sse", it gives a
> warning and uses 387 (so check_cflags would always add it).

Does any gcc version not have that flag?

-- 
M?ns Rullg?rd
mans at mansr.com