hobold (hobold@vectorizer.org) on 8/19/10 wrote:
---------------------------
>Steve Underwood (steveu@coppice.org) on 8/18/10 wrote:
>---------------------------
>[...]
>>It took a lot of time for people to get the best out of hand shuffling
>>things with SSSE3, and the next generation core made this complexity something that
>>needs to be ripped out of the code. AAAHHHHH!
>>
>
>http://www.khronos.org/developers/library/2010_siggraph_bof_opencl/OpenCL-BOF-Intel-SIGGRAPH-Jul10.pdf
>
>Granted, this approach will not give you the equivalent of fully hand tuned performance,
>but it can come reasonably close. And better yet, it abstracts from the gazillion
>existing SSE variants and the upcoming AVX and Larrabee ISA variations. I wonder
>if that will end up being a more significant advantage than the abstraction from
>completely different hardware types by different vendors.

So OpenCl becomes the new Java, sort of? I like the idea. Not being stuck to a single ISA or memory model and have great performance at the same time is.. great. However OS kernel is still going to be as ISA specific as before and same with compiler, so ISA still matters a lot. Still OpenCL might make it easier to embrace new ideas in future massively parallel compute engines.