Re: [eigen] Use of streaming loads in eigen

[ Thread Index | Date Index | More Archives ]

On Fri, Mar 27, 2009 at 2:57 PM, Rohit Garg <rpg.314@xxxxxxxxx> wrote:
> In the PacketMath.h file, the only eigen function to send the sse
> registers back to memory is the _mm_store_ intrinsic. Has anybody
> looked at using the _mm_stream intrinsics instead. They should help
> atleast in BLAS 1 and BLAS 2 codes where the code is fundamentally
> bandwidth bound. For a pseudo benchmark. look here
> Even for BLAS 3 it should only help. However, for small sized stuff,
> it will definitely not be helpful as one will want to have the item in
> the cache as long as possible as it doesn't pollute the cache because
> of it's small size.

that's exactly why we cannot use it in Eigen: we cannot know if the
data are going to reused soon or not, and even in the best case
scenario the improvement would be pretty low...

> --
> Rohit Garg
> Senior Undergraduate
> Department of Physics
> Indian Institute of Technology
> Bombay

Mail converted by MHonArc 2.6.19+