Re: OSKI, an automatically tuned sparse kernel library.

[Richard Vuduc]

Andy Adler wrote:
> I've been trying to work on the octave sparse matrix multiply code;
> I have a patch for low density sparse matrices that should
> increase the performance by about 4x. When I get back from my
> vacation (end-July), I'll test and submit it.

I would be interested in hearing more about this too when you return.

> I wanted to try to compare this to the OSKI performance. OSKI
> provides a "make benchmarks" function with looks promising.
> Unfortunately, I couldn't find a way to compare to octave
> performance easily.
>
> 1. OSKI benchmarks creates its own test matrices. There doesn't
>    seem to be a way to enter or save these to be tested elsewhere.

There are some benchmark programs you can try running which 
accept Harwell-Boeing matrices as input. There is a very short 
description in Section 6.4 (p. 31) of the OSKI user's guide. The 
benchmark you want is probably "oskibench_autotune_Txy (where 
"xy" specifies a particular precision combination liked "id" for 
'int' indices and 'double' floating point values). This benchmark 
reads in a matrix pattern from a Harwell-Boeing file and reports 
performance for a particular workload. Run it without arguments 
to see its usage, or drop me a line when you return from vacation.

> 2. I wasn't able to find test code for sparse x sparse function.

Unfortunately, OSKI does not have a sparse mat x sparse mat function.

Best,
--rich