Intermittent bifrost.linalg test failures

[jaycedowell]

Occasionally we see test failures on the self-hosted bifrost.linalg suite. Now that I'm looking for one to point to I cannot find one.

[jaycedowell]

Here's one: #167 (comment)

[jaycedowell]

I wonder if this is somehow related to #210. The only places where BF_STATUS_UNSUPPORTED_SHAPE can be thrown from a LinAlg call are in linalg_kernels.cu:

• bf_cherk_N
• bf_cgemm_TN_smallM_staticN_v2
• bf_cgemm_TN_smallM

These are all kind of trivial though. It's mostly value checking for the matrix shape. There are a couple of comparisons of the batch size with the texture memory size that can also throw this. It would be nice to know exactly which BF_STATUS_UNSUPPORTED_SHAPE we are hitting.