Re: [eigen] compile time statistics

["Gael Guennebaud" <gael.guennebaud@xxxxxxxxx>]

arf, sorry.

for the cost model bench note that the number of loops depends on the
size of the matrices, so you cannot  really compare 2 "rows". Moreover
the number of loops for the scalar product is divided by two. The
purpose was to get a similar running time for all cases. That is weird
is that the scalar product is mush slower than the matrix sum, maybe
we could try a cwise product to see what happens....

gael.


On Tue, Jun 17, 2008 at 1:19 PM, Benoît Jacob <jacob@xxxxxxxxxxxxxxx> wrote:
> On Tuesday 17 June 2008 13:07:38 Gael Guennebaud wrote:
>> Hi,
>>
>> I attached
>
> you forgot to :)
>
> but thanks a lot, really looking forward to seeing your stats. Very useful.
>
> While I'm there : I'm currently running myself the eigencostmodel benchmark.
> I'm still totally puzzled by the fact that 8x4x8 is so fast, and generally
> that the bigger sizes are the faster ones, the only explanation I guess is
> vectorization, i'll retry with implicit vectorization disabled (i was already
> not using explicit vectorization).
>
> Benoit
>

Attachment: EigenCompilationTime_02.ods
Description: application/vnd.oasis.opendocument.spreadsheet