Nirvedh Meshram
Nirvedh Meshram
> > This is just merely based on the observation of this IR. Since we are adding zeros, can't we DCE the linalg.generic ? > > Good question. I honestly...
The problem is this dispatch ``` #map = affine_map (d0, d1, d2)> #map1 = affine_map (d0)> #map2 = affine_map (d0)> module { func.func @main_graph$async_dispatch_1_generic_288x28x28_f32(%arg0: tensor) -> tensor { %cst =...
I just tried #18474 and interestingly the failing dispatch passes but the next one fails on it, [here](https://gist.github.com/nirvedhmeshram/f0714f05c352dfb6601af135b587f07c) is a full dump This is the failing IR/pass ``` input_ir.mlir:15:13: error:...
We have the above issue fixed in #18657 and once #18474 also lands, this IR will work.
Yes #18474 has not landed yet. I will ping here when we can try it again.
Just FYI, the reason #18474 is not landed yet is there are numeric differences with it on some onnx models so we are making sure the numerics are correct before...
A continuation of this issue will be https://github.com/iree-org/iree/issues/18905
> I was able to solve this error by removing the input sizes and only using the input file, i.e. using `--input='@input.0.bin'` instead of `--input='[email protected]'`. It seems like GPU doesn't...
If this can be fixed in the front-end then thats great but I think we should be able to support this in the compiler, in that regard I think the...
We need to generate new linalg IR for this issue with https://github.com/llvm/torch-mlir/pull/3762 , since the ONXX IR has been already working when run directly from IREE