GPUPadOperand pass does not work for dynamic shapes

[nirvedhmeshram]

A fully dynamic mamtmul_transpose_b is lowered to the following dispatch with padding config by TileAndFuse.

Running the following command on it

iree-opt padding_failure.mlir --pass-pipeline="builtin.module(func.func(iree-codegen-gpu-pad-operands))" --debug

gives the following error

Query value bounds for: %44 = flow.dispatch.workload.ordinal %31, 1 : index (owner: flow.dispatch.workload.ordinal)
--> ValueBoundsOpInterface not implemented
[linalg-padding]: ----could not compute a bounding box for padding[linalg-padding]: --operand cannot be bound statically : %extracted_slice = tensor.extract_slice %52[%arg0, 0, 0, 0] [1, 16, %44, 4] [1, 1, 1, 1] : tensor<?x16x?x4xi8> to tensor<1x16x?x4xi8> -> FAIL

Here is the full log

Also here is the full dump from iree-compile.

[nirvedhmeshram]

Update on this, only reduction dim has the bounds issue, but since this is after distribute to workgroups bounds can be found on the parallel dims and this works (with a few modifications to the selection code) for static reduction dims