On 11/2/2012 3:10 PM, Jens Mueller wrote:
I see. Thanks for clarifying.
If I want fast vector operations I have to use core.simd. The built-in
vector operations won't fit the bill.
I think a better quote would be "If i want HARDWARE vector operations..."; this is not automatically faster by nature, it requires strict self-control in terms of application, and very careful attention if you want your code to be portable.
At the moment, yes.
However, Manu is working on developing a higher order layer.
I have a fork; some people are using it already. It still needs a lot of work though; some compilers missing parts, platforms not supported.
That said, it's not an effort to address D's natural vector syntax, the key goal is to provide a hardware SIMD API that is as orthogonal as possible and portable (with confidence it will run reasonably well).
I wonder if druntime could be enhanced to use the SIMD stuff though in the functions that perform the natural vector operations, might offer some nice little boosts.