GitLab

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.

Add support for the QLinearMatMul onnx operator

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.

* Fix scatter operator for nonstandard shapes

remove standard() shape check for scatter inputs.

* Add nostandard input tests for scatter

---------
Co-authored-by: Chris Austen <causten@users.noreply.github.com>

Makes ConstantOfShape work with a variable dimensions input

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

MIGraphX verification by default uses normalized RMS error as the basis for the verification.  This change adds some logic to allow migraphx to do "np.allclose" type of elementwise verification using atol and rtol.

Commit also includes changes to consistently pass "gold" or "expected" results as the second argument for "verify_range()" calls.  Default RMS tolerance inside driver is set to 0.001 which IMO is high for FP32 compared to what we had earlier. Need better defaults

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.