[Shootout-list] Proposal: revised matrix multiplication test
Einar Karttunen
ekarttun@cs.helsinki.fi
Thu, 4 Nov 2004 00:41:38 +0200
On 03.11 21:41, Aaron Denney wrote:
> N is given on the command line, and the matrix is given on standard
> input, in the form dimension D on the first line, and then D lines
> each having D numbers being the elements of the matrix in row-major
> order.
Please make the matrix constant, but require the programs to
work with any other matrix too.
> To avoid differences in handling overflow between implementations, I
> would suggest using a signed permutation matrix[1] to test, though the code
> must, of course, work on any square matrix.
>
> Rationale:
> The current tests relies on doing the same work multiple times, in order
> to get the times high enough to get reliable measurements. Rather
> than this, I would prefer that all the work actually be necessary.
agreed.
> [1] each row and column is zero, except for one entry which may be
> either plus or minus one. There are D!*2^D different ones.
We need to make a dense matrix representation explicit then,
otherwise people will use sparse solutions.
- Einar Karttunen