CUTLASS results incorrect about cutlass

[kerrmudgeon]

Yes! We have root caused this problem and are preparing a patch coming soon.

Thank you for being patient.

[d4l3k]

@kerrmudgeon Any thoughts/eta for a fix?

[YSZhuoyang]

Same here.

Built successfully and ran tests with cutlass v1.1, CUDA V10.0.130, NV driver 410.66 on Ubuntu 18.04, gtx 960m GPU failed (and it seems all failed tests are sgemm... & dgemm...).

[kerrmudgeon]

Unfortunately, we believe this problem affects all GPUs in the Maxwell architecture. We have not yet root caused it.

Pascal, Volta, and Turing are okay.

Will report back when we have updates.

[d4l3k]

@kerrmudgeon Any updates?

[kerrmudgeon]

CUTLASS 1.2 mainly contributes optimizations for batched GEMM targeting Tensor Cores on Volta and Turing as well as optimizations for problems with small GEMM-M and GEMM-N dimensions.

We have not addressed the issue you and others have reported regarding CUTLASS on the Maxwell architecture, however. We are looking into it.

[azazhu]

hi Andrew, any updates?

[kerrmudgeon]

I have pushed a patch to the master branch. Can you verify that it resolves the correctness problem on Maxwell?

[YSZhuoyang]

Tried it again in the same environment and I saw all tests were passed.

[kerrmudgeon]

Good to hear.

I think this is safe to close. Feel free to re-open if you still have problems.