[CSC 435] Super Fast Code Results (Don't Believe Everything you See)

[Daryl Ebanks]

Andrew J. Pounds wrote:
> Okay -- before we all stand in awe of the code performance we saw from
> Long's DGEMM BLAS runs today (which a little off the cuff calculation
> in my head appeared to be in the hundreds of  gigaflops range) -- try
> the following test.
>
> Make ONE of the matrices used in the product calculation something
> other than the identity matrix and retry the calculation.  You should
> get something much more realistic (in the hundreds of  megaflops range).
> Can you tell me what is going on?
If changing a matrix to something other than the identity matrix
increases the time, perhaps the blas subroutine recognizes that the
multiplication is trivial in which case it may skip the multiplication
completely. That would explain the impossibly small times we were getting.
>
>