[BUG] `matmul` do not support int32 tensors. about matx HOT 4 CLOSED

Hi @AtomicVar, thanks for the report. You're right cuBLAS doesn't support it. The reason we disabled CUTLASS is because of the compile times taking extremely long in the unit tests. Even though the functionality and speed were both adequate, the compiles were taking over 40 minutes. We can add a message for this saying it's an unsupported type.

from matx.

@AtomicVar I've submitted MR #540 to resolve this

from matx.

@AtomicVar according to the cublas team we're not meeting the requirements here: https://docs.nvidia.com/cuda/cublas/index.html#cublasltmatmul

Namely A must be transposed.

If that still works for you, we can add support for it when those requirements are met.

from matx.

@cliffburdick Actually I can use float matmul instead of int matmul to accomplish the same task. So it won't be a problem if int is not supported. I just think the unsupported conditions should be documented and we need to throw and print clear error messages.

from matx.