By: Axel (no.delete@this.spam.org), January 21, 2011 6:51 am
Room: Moderated Discussions
Eric Bron (eric.bron@zvisuelREMOVE.com) on 1/20/11 wrote:
---------------------------
>>Biggest advantage of AVX is that it supplies 3operand instruction formats for *all*
>>previous SSE1,2,3,4.1/.2/AES instructions.
>
>from hands on experiments the speedup after recompiling legacy source for AVX-128
>give you less than 5% speedup (vs SSE on the same machine), the code is 10% - 15%
>more compact but you mostly remove register to register moves which were fast, the
>bottlenecks (loads/stores, branch misses, etc.) are exactly the same, so I'm quite
>sure nobody will endure the burden to have an AVX path just for this 5% speedup
>
Hm ok, thanks for your empiric data, definitly better than my guesses ;-)
However, is it really a burden to have that AVX path ? +5% "just" for recompiliation would be quite nice, imo. Or did you encounter several (severe) problems ?
---------------------------
>>Biggest advantage of AVX is that it supplies 3operand instruction formats for *all*
>>previous SSE1,2,3,4.1/.2/AES instructions.
>
>from hands on experiments the speedup after recompiling legacy source for AVX-128
>give you less than 5% speedup (vs SSE on the same machine), the code is 10% - 15%
>more compact but you mostly remove register to register moves which were fast, the
>bottlenecks (loads/stores, branch misses, etc.) are exactly the same, so I'm quite
>sure nobody will endure the burden to have an AVX path just for this 5% speedup
>
Hm ok, thanks for your empiric data, definitly better than my guesses ;-)
However, is it really a burden to have that AVX path ? +5% "just" for recompiliation would be quite nice, imo. Or did you encounter several (severe) problems ?
