Use FusedLocation to hold layer metadata

[yizhuoz004]

No description provided.

[yizhuoz004]

@pranavm-nvidia Need your advise on this issue. Previously the stderr looks like this:

ITensor::getDimensions: Error Code 4: Shape Error (reshape changes volume. Reshaping [2,2] to [3,3]. In operator() at optimizer/common/shape/shapeContext.cpp:4727)
loc("%t2,%t3;;<out>;;%t4"): error: ranks of MLIR Tensor is 2 while TRT ITensor has rank -1
loc("%t2,%t3;;<out>;;%t4"): error: 'tensorrt.shuffle' op failed to encode operation
loc("%t1,%t0;;<out>;;%t2"): error: failed to encode block
loc("%t2,%t3;;<out>;;%t4"): error: failed to translate function 'tensorrt_cluster' to a TensorRT engine

But now we changed to fusedLocation, and the stderr became:

ITensor::getDimensions: Error Code 4: Shape Error (reshape changes volume. Reshaping [2,2] to [3,3]. In operator() at optimizer/common/shape/shapeContext.cpp:4727)
loc(fused<"<TraceOp: %t4 = reshape(%t2 : tensor<?x?xf32:gpu:0>, %t3 : tensor<2xi32:gpu:0>) : tensor<?x?xf32:gpu:0>, Stack Info:   --> /tripy/test_tril.py:7 in <module>()  >, ">["%t2,%t3;;<out>;;%t4"]): error: ranks of MLIR Tensor is 2 while TRT ITensor has rank -1
loc(fused<"<TraceOp: %t4 = reshape(%t2 : tensor<?x?xf32:gpu:0>, %t3 : tensor<2xi32:gpu:0>) : tensor<?x?xf32:gpu:0>, Stack Info:   --> /tripy/test_tril.py:7 in <module>()  >, ">["%t2,%t3;;<out>;;%t4"]): error: 'tensorrt.shuffle' op failed to encode operation
loc(fused<"<TraceOp: %t2 = broadcast(%t1 : tensor<f32:gpu:0>, %t0 : tensor<2xi32:gpu:0>) : tensor<?x?xf32:gpu:0>, Stack Info:   --> /tripy/nvtripy/frontend/ops/ones.py:53 in ones() --> /tripy/test_tril.py:6 in <module>()  >, ">["%t1,%t0;;<out>;;%t2"]): error: failed to encode block
loc(fused<"<TraceOp: %t4 = reshape(%t2 : tensor<?x?xf32:gpu:0>, %t3 : tensor<2xi32:gpu:0>) : tensor<?x?xf32:gpu:0>, Stack Info:   --> /tripy/test_tril.py:7 in <module>()  >, ">["%t2,%t3;;<out>;;%t4"]): error: failed to translate function 'tensorrt_cluster' to a TensorRT engine

We can extract the original ["%t2,%t3;;<out>;;%t4"] part from the new stderr, or do you think there's a new way as there's already some stack info in the error message?

[pranavm-nvidia]

@yizhuoz004 I think you can just update the regexs here: https://github.com/NVIDIA/TensorRT-Incubator/blob/main/tripy/nvtripy/backend/mlir/utils.py#L132. We probably don't want to include the stack info in this part of the message - we should keep our current logic of swapping out the location information with the tensor name.