[Ffmpeg-devel] mmx/sse optimizations

[Luca Barbato]

Rich Felker wrote:
>>
>>	SIMD code is a different beast than normal ASM. I think this tool 
>>	has its place. Perhaps another idea would be a tool that enables the user 
>>to describe some kind of general algorithm with a set of SIMD-like 
>>primitives.

...

>> Then the tool decides what the equivalent code would look 
>>like using different SIMD sets (MMX, SSE2, AltiVec) AND can output the 
>>correct ASM syntax for use in gcc and MSVC.
> 
> 
> Hmm, sounds a lot like vector intrinsics. IMO they're a trap that lead
> to people making simd code that's 5-25% slower than handwritten asm.

Prove it on non flawed SIMD and non flawed CC. I'm quite sure nobody 
will get a schedule right for each cpu. The problem is that gcc + 
mmx_intrinsics == boom ...

lu

-- 

Luca Barbato

Gentoo/linux Developer		Gentoo/PPC Operational Leader
http://dev.gentoo.org/~lu_zero