infinite loop in TransposeEliminationPass with dynamic input

[zezhang]

Following IR causes infinite loop in the TransposeEliminationPass

#map = affine_map<(d0, d1, d2) -> (d0, d2, d1)>
module {
  func.func @test(%arg0: tensor<?x80x80xf32>) -> tensor<?x80x80xf32> {
    %cst_f32 = tensorrt.constant dense<1.000000e+00> : tensor<1x1x1xf32>
    %1 = tensorrt.transpose {permutation = #map} %arg0 : tensor<?x80x80xf32> to tensor<?x80x80xf32>
    %2 = tensorrt.element_wise <kSUB>(%cst_f32, %1 : tensor<1x1x1xf32>, tensor<?x80x80xf32>) -> tensor<?x80x80xf32>
    return %2 : tensor<?x80x80xf32>
  }
}

More details: I think the issue comes from PushdownTransposeEwise patten (may exist in other similar patterns as well). Could be something wrong with pushDownTransposePrecondition when calculating memoryCost with dynamic inputs?

[christopherbate]

Thanks for reporting this. It looks like guess is correct, it doesn't handle dynamic shapes correctly. I've tested a simple fix internally and should have it sync'd up here by end of week.

[christopherbate]

Fixed in #371