|
From: Julian S. <js...@ac...> - 2012-08-23 18:21:10
|
> Other than aesthetics, is there a technical reason for wanting this? > (For example, do you think the performance would be improved for > 128-bit stores?) True, partly for aesthetics, but also partly because as you say, it would probably improve performance for 128/256 bit stores. Anyway, that's something for the future; not now. > Also, do you want my initial implementation to include V256 -- which, > as far as I can tell, does not yet exist as a typedef -- or is V128 > enough? I'm unclear what you mean. Do you mean should you introduce a V256 type, or do 256 bit returns? Ideally both. With 128 bit returns only, we will have to split 256s into two 128s, which still leaves the original style overrun problems in cases where vectorising compilers (icc) generates AVX vector loops. J |