On 12/27/10 5:38 PM, Richard Hennessy wrote:
> How hard to just change matrix multiplication to use threads? The
> parallelism here is so obvious.
In another message you said you were doing convolution via matrix
multiplication. In that case, why not use FFTs to do the convolution.
That would speed things up much more than threads would. Your O(n^3)
matrix multiplies would still be O(n^3), but if you use FFTs it would be
O(n*log(n)).
How big are your matrices?
It would probably be fairly easy to make matrix multiplication
threaded. But a GOOD implementation would be hard.
Ray, who's never really implemented threaded anything