Implement MLIR enablement heuristic #2296
Description
Per discussion with @pfultz2 , rocMLIR-based kernel generation will be enabled in the 6.0 release in the following circumstances:
- Any convolution fusion
- Any i8 gemm fusion
- All Navi3x standalone convolutions
- With a flag that is off by default, fp* gemm fusions
Except:
- 3x3 winnograd convolutions fusions (except on Navi)
- K > 2048 on gemm (as for CK)
- This should be K relative to G * M * N, but if that can't be implemented by release, we'll go for the simple filter
#2274 implements part of this heuristic, but is incomplete (and, for example, doesn't have the "don't skip winnogradable convolutions on Navi" part)
@causten @jerryyin for awareness
If someone could at in Andy (I think) who has the performance measurements scripts for more comments, that'd be helpful