BLIS Test Failure in BlueGene/Q about blis HOT 11 CLOSED

I checked the bgq micro-kernels, and they do in fact honor both row and column strides. However, they do not implement the "beta == 0" case in a nan/inf-safe manner. The "beta == 0" case is used by the operations you quote, so my best guess is that this is the problem.

from blis.

Dominic: Is it just the complex cases that are broken?

Field: Yeah, I never went back and fixed the 'beta == 0' case when that
became a requirement. I should do that, but since the single-precision
kernels are just the reference ones, that isn't the only problem that we're
seeing here.

What are the tests named like blis__4mh__ ? Are they the only ones that
fail?

-Tyler

On Mon, Aug 31, 2015 at 8:56 AM, Field G. Van Zee [email protected]
wrote:

I checked the bgq micro-kernels, and they do in fact honor both row and
column strides. However, they do not implement the "beta == 0" case in a
nan/inf-safe manner. The "beta == 0" case is used by the operations you
quote, so my best guess is that this is the problem.

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

from blis.

Tyler: Hmm, now that I think about it, my theory about "beta == 0" would not explain why the cgemm4mh is failing. I would be willing to debug this, but unless/until we reproduce it on an Intel architecture, I can't really justify making it a priority. (I don't have access to, much less experience with, bgq hardware.)

from blis.

@fgvanzee I can get you access easily. You just have to say you want it :-)

from blis.

Tyler: apparently all failure cases are complex cases.Thanks!

Field: we can also arrange an account if you need to access our system to debug BLIS.

from blis.

Hi is there any update on this thread?

from blis.

Anyone is trying to fix this problem?

from blis.

@fgvanzee Is this the same thing as issue number #51 ?

from blis.

@tlrmchlsmth Probably. However, the vulnerability to nan's/inf's due to not handling "beta == 0" still needs to get fixed eventually, even if it is not causing the symptoms originally reported by Dominic, so I figured we should keep the issue open.

from blis.

@tlrmchlsmth, according to #51, setting compiler flag to -O0 should fix its problem, but it does not work for the problem on BGQ. (for version 0.2.0)
blis_cgemm3m1_hh_ccc 400 400 400 0.000 4.31e-07 PASS
blis_cgemm4mh_nn_ccc 100 100 100 0.000 4.21e-04 FAILURE
blis_cgemm4mh_nn_ccc 200 200 200 0.000 1.34e-04 FAILURE
blis_cgemm4mh_nn_ccc 300 300 300 0.000 7.35e-05 MARGINAL
blis_cgemm4mh_nn_ccc 400 400 400 0.000 5.06e-05 MARGINAL
blis_cgemm4mh_nc_ccc 100 100 100 0.000 3.29e-04 FAILURE
blis_cgemm4mh_nc_ccc 200 200 200 0.000 1.50e-04 FAILURE
blis_cgemm4mh_nc_ccc 300 300 300 0.000 7.29e-05 MARGINAL
blis_cgemm4mh_nc_ccc 400 400 400 0.000 4.83e-05 MARGINAL
blis_cgemm4mh_nt_ccc 100 100 100 0.000 3.95e-04 FAILURE
blis_cgemm4mh_nt_ccc 200 200 200 0.000 1.42e-04 FAILURE
blis_cgemm4mh_nt_ccc 300 300 300 0.000 7.96e-05 MARGINAL
blis_cgemm4mh_nt_ccc 400 400 400 0.000 4.71e-05 MARGINAL
blis_cgemm4mh_nh_ccc 100 100 100 0.000 3.93e-04 FAILURE
blis_cgemm4mh_nh_ccc 200 200 200 0.000 1.41e-04 FAILURE
blis_cgemm4mh_nh_ccc 300 300 300 0.000 7.86e-05 MARGINAL

from blis.

I am closing this issue because ALCF Blue Gene/Q systems have been retired and thus friends of BLIS like me lack the ability to reproduce user errors.

from blis.