Gheorghe-Teodor Bercea

Results 15 comments of Gheorghe-Teodor Bercea

The true solution is to use `tensor` i.e. unknown shape and type. I'm not sure this is something that MLIR currently supports. Whichever type we end up using, shape inference...

@feiyuvl for hand written CUDA code, it might be easier to have a transformation from krnl to CUDA due to the fact that the number of operations in KRNL is...

Can we have a list of which ops have failed shape inference? Can we also have the data type which we are missing? Can you re-run the models again please?...

Do we have any updates on this? @tungld @chentong319

I have reopened this since Carlo reported the error above. The fix for this latest error is here: https://reviews.llvm.org/D138614

Thanks for the answer! :) My confusion stemmed from the fact that I was expecting each instruction that is emitted to be emitted with a valid type. In this case...

@jpienaar I managed to fix the error by converting my tensors into memrefs. The way I did this was by using the toy::CastOp. Is there a way to do this...

@shoshijak does NVCC not complain because it doesn't even attempt to unroll it or it actually unrolls it, have you checked the NVCC temp files to make sure NVCC actually...

@ppanchad-amd I am seeing this issue in latest ROCM so issue is still there.

Thank you for the answer @schung-amd The problem as I see it is actually not mn because mn is known at compile time because m and n are both known...