However, I noticed that none of the generated solutions are using the WMMA instructions. My architecture does have HasWMMA=1.
I confirmed that kernel["EnableMatrixInstructions'] is False in the kernel writer code... so that's not good - the problem is probably up the stack from there.
I thought I could add EnableMatrixInstructions to the rocblas_sgemm_example.yaml but I get an error Tensile::FATAL: Invalid parameter name: EnableMatrixInstruction
So:
is it possible to enable and disable matrix instructions through the yaml? If so how?
if not, is there another way to force this setting on or off?
I used the the example at the bottom of https://github.com/ROCm/Tensile/wiki#quick-example-ubuntu to generate a set of solutions.
However, I noticed that none of the generated solutions are using the WMMA instructions. My architecture does have HasWMMA=1.
I confirmed that kernel["EnableMatrixInstructions'] is False in the kernel writer code... so that's not good - the problem is probably up the stack from there.
I thought I could add EnableMatrixInstructions to the rocblas_sgemm_example.yaml but I get an error Tensile::FATAL: Invalid parameter name: EnableMatrixInstruction
So:
Thanks!