From the master code，it seems only tensorop is supported？ Is there any plan to support simt opcode? For model for which precision is import, we could not use tensor cores.
You are right we only support tensorcore for now. SIMT support can be easily added, but there is no plan to add such feature.
Is there any doc or any guidelines for tile descriptions of simt? I try to set different shapes, some of them will trigger compile assert errors.
The best source for now are the simt kernel generators in cutlass: https://github.com/NVIDIA/cutlass/blob/1604ebaf10189c3558e5182990e8d5dc41adc1dd/tools/library/scripts/generator.py#L3631
1 Like