GitLab

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.

 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

This will enable MLIR by default for these cases:

Any convolution fusion
Any int8 gemm fusion
All Navi3 standalone convolutions
With a flag(ie MIGRAPHX_ENABLE_MLIR) to enable MLIR for floating-point gemm fusions
Except:

3x3 winnograd convolutions fusions (except on Navi)
K > 2048 on gemm (as CK)
Also there is MIGRAPHX_DISABLE_MLIR to disable MLIR completely.

Currently, we see MLIR partition candidates recieving non-standard shape due to not fusing in squeeze, flatten and unsqueeze ops. These ops could be canonicalized to reshape without introducing additional ops as long as MLIR backend is concerned.

* some skeleton code to handle attention patterns

add flags for ck, Enable CK with hipRTC.  CK can be used with the MIGRAPHX_ENABLE_CK=1 and MIGRAPHX_TUNE_CK=1

Enable MLIR performance enhancements with MIGRAPHX_ENABLE_MLIR=1

Modify reshapes to use reshape_lazy for aliasing and then reshape for a reshape copy operation to eliminate contiguous

Currently, the is_mlir_conv predicate wasn't being used when
offloading standalone convolutions to MLIR on Navi3x, which caused
failures relating to being unable to construct the MLIR program when a
3D convlolution was passed in.

This commit amends the standalone lowering to use said predicate, as
well as to include quant_convolution and quant_dot into the set of
operations that get a standalone lowering.