[ublas] Re: SIMD?

From: Tony Juricic (tonygeek_at_[hidden])
Date: 2005-03-14 19:17:57

Paul Baxter wrote:
> Personally, I can't help feeling a combination of expression 
> templates and SIMD instructions could be good for performance.
> 
> My initial comparisons of his valarray implementation against 
> IPP/MKL-BLAS and standard library alternatives were encouraging albeit 
> currently there is only limited type support on the PC.

Some here. I also used MacSTL with some encouraging results. Clearly, 
the fact that UBLAS implementation is in header files would not be a 
problem in principle. But I think that accomodating SIMD does affect the 
design of the library in the sense that appropriate macros would have to 
be defined to set 16-byte data alignment (for ex. on Intel platform, if 
SIMD is supported) and 'stabdard' ways for alternative SIMD 
implementation to substitute the existing one. This tends to make code 
more complex and 'dirty' with defines which, I assume, is the main 
reason why SIMD was not already seriously considered.

Tony