Paul Fultz II
Paul Fultz II
> Fix P2 : We need a another dump /w MIGRAPHX_TRACE_MLIR=1 before it is benchmarked along with the solution. I dont think we can dump MLIR after its been compiled....
program.cpp is target-independent and shouldn't be checking for target-dependent operators. Also such a calculation is operator-specific and should reside with the operator. The `operation` class should have a method to...
Reopening since the PR wasnt finished.
> We won't hit the verify "accuracy" bug for the quantized resnet50 model with the change from MLIR. If we dont hit an accuracy error then we shouldn't remove it.
Will this handle cases like this? ``` ┌─┐ │a│ └┬┘ ┌▽──┐ │b │ └┬─┬┘ │┌▽┐ ││c│ │└┬┘ │┌▽─┐ ││d │ │└┬─┘ ┌▽─▽┐ │e │ └───┘ ``` Especially if `c` or...
> I don't think that will be handled well even with existing logic on develop branch or with this PR. > > It is implictly assumed that only ONE of...
> If b + c is matched and fused then it would work as expected. b would be part of multiple outputs and then later consumed by e . This...
> the question is whether it ever makes sense to define a shape as both standard and broadcast Yes it does make sense. "Standard" shape basically means that the offset...
> Took some time but replicated the fail on rocm-rome-6. Examining the output now. Output file: So the failure is in fuse_ops, which means we have a bug somewhere where...
Overall, I think this looks good, but need another review from Alan or Shiv, and to fix the convert issue with MLIR.