Re: [gentoo-soc] Benchmarking suite - Report 7

[Andrea Arteaga]

> How about FFTs rather than DFTs?

"A fast Fourier transform (FFT) is an efficient algorithm to compute
the discrete Fourier transform (DFT) and its inverse" [0]
The tests compute the DFT by using FFTW, which of course implements
the FFT algorithms.
Or did you mean something else?


> Is there any intent that it be able to automatically select the best
> parameters in addition to just graphing the results? From looking at
> some of those graphs, it seems like if you applied some smoothing, you
> could pick the peaks out fairly easily.

It would actually quite easy to find the best implementation for a
specific operation and matrix size. But...
Have a look of my BLAS results [1]: you can notice that eigen is very
slow for some opeartion like axpy, while it is very fast with matrix
multiplications or system solution; the gotoblas and openblas
implementations are usually the slowest with small matrices, but are
really fast with big matrices; atlas has swinging performances (see
matrix-matrix multiply), making it sometimes quite fast, sometimes
very slow.

Bref: there are quite a few parameters that the user would insert in
order to decide automatically the best suited implementation. Big
matrices or tiny ones? Just Level 2 and 3 operations or Level 1, too?
Constancy or peak performance? Constructing such an automatic "filter"
would require a lot of work and would probabily be difficult making it
so good as the human eye is... I could maybe create kinda score
system, and give the user an hint of the most performing
implementation, but having a 2-minutes look at the graphical results
would always be the best thing to do. The only drawback of a manual
decision is that only a few (6-7) implementation can be shown at the
same time in order to keep the plots readable.

Thanks for the comments.
Cheers
Andrea

[0] http://en.wikipedia.org/wiki/Fast_Fourier_transform
[1] http://www.phys.ethz.ch/~arteagaa/soc/blas/