GitLab

This is a companion PR to
https://github.com/ROCmSoftwarePlatform/rocMLIR/pull/1312 .
The updated commit hash points onto that PR branch, so coordinated
merges are advised.

With the above rocMLIR changes, the MLIR MIGraphX dialect now
represents both the dimensions and strides of tensors inside MLIR,
thus allowing NHWC convolutions to be correctly offloaded.

In this PR, we:
- Remove special handling for the case where non-standard shapes
become input to MLIR modules
- Fold broadcast and multibroadcast operations into the input size od
MLIR modules
- Update tests
- Add an extra TRACE_MLIR print to help debug crashes in the
high-level pipeline

The inaccuracy was caused by ONNX round requires nearest integer rounding for halway (0.5) cases.
std::round rounds away from zero, thus giving wrong results with halfway cases.
Replaced std::round with std::nearbyint which uses the correct rounding by default.

Fixes an issue that comes up for variable input slice with steps set manually in ONNX to default 1's.

Simplifies dimensions_of instructions to a literal when possible.
Intended to be used after the split_single_dyn_dim pass.

Since opset version 18, the Split operator allows splitting into unevenly sized outputs when the split input is not present, and a num_outputs attribute has been introduced.

 Add tracing to benchmark to show which kernels are running and the time of every kernel

Adds workarounds to avoid passing capture ops and scalar literals from quantization as arguments to ck_gemm.

* Disable -Wunsafe-buffer-usage when compiling gpu code