non homegenous GEMM_MULTITHREAD_THRESHOLD usage

[jeromerobert]

In gemv.c GEMM_MULTITHREAD_THRESHOLD is squared so it's homogenous to a matrix size. In gemm.c it's not squared so it's homogeneous to a number of floating point operations

So there is no chance that changing GEMM_MULTITHREAD_THRESHOLD in Makefiles.rules would give good results for both gemm and gemv.

Shouldn't gemv.c be fixed so that is use GEMM_MULTITHREAD_THRESHOLD as a number of floating point operations ?

A possibly related bug is #216 .

[jeromerobert]

As gemv.c, trsm.c also consider GEMM_MULTITHREAD_THRESHOLD as a matrix size.

[xianyi]

@jeromerobert , I think GEMM_MULTITHREAD_THRESHOLD of gemv.c is also a number of floating point operations since gemv is O(n^2).

Or, different threshold for different functions?

[jeromerobert]

The gemv.c code is:

if ( MNK <= (24.0 * 24.0  * (double) (GEMM_MULTITHREAD_THRESHOLD*GEMM_MULTITHREAD_THRESHOLD) )  )

MNK is the number of floating point operation. But we compare to a_const * GEMM_MULTITHREAD_THRESHOLD * GEMM_MULTITHREAD_THRESHOLD while we should, IMHO, compare to a_const * GEMM_MULTITHREAD_THRESHOLD, as done in gemm.

Or, different threshold for different functions?

I'm not sure that's a good idea because that will make OpenBLAS very difficult to calibrate and maintain.

[jeromerobert]

I think we need something like:

if(MNK <= 2304 * GEMM_MULTITHREAD_THRESHOLD)

which is equivalent when GEMM_MULTITHREAD_THRESHOLD = 4.

[jeromerobert]

@wernsaar Could we have your feeling, as you are the author of gemv and trsm thresholding ?

[wernsaar]

Yes, that's right.

On 01/17/2016 07:02 PM, Jerome Robert wrote:

@wernsaar https://github.com/wernsaar Could we have your feeling, as
you are the author of |gemv| and |trsm| thresholding ?

—
Reply to this email directly or view it on GitHub
#742 (comment).