GitLab

Add min and max reduction modes
Implement duplicate index handling
Add corresponding reference, onnx, and verify tests.

Introduce dilations attribute to pooling operators reference implementation.

Handles all 4 Fp8 dtypes listed here : https://onnx.ai/onnx/technical/float8.html
Follows saturation/clipping logic from table there as well : https://onnx.ai/onnx/technical/float8.html#cast
Only adding fp8e4m3fnuz in MIGraphX IR for now.

Reworked the simplify_qdq pass to support:

Per-axis quantization (ie. allow 1D scales and zero points)
Allow broadcast and transpose ops between dq and quant_op

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.

* 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>

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

New compiler pass that simplifies dynamic shapes related operators to their static versions if possible
Will normally be used after a split_single_dyn_dim pass

* Rectify flipped coordinate_transformation_mode logic in ROIAlign
* Handle both opset 10 and 16 versions
* Fix version check and clang tidy warning
Co-authored-by: Dino Musić <dino.music@htecgroup.com>

Implements a fill operator that sets the values in an output buffer to a given value
Will be used when parsing ONNX ConstantOfShape
Can also be used when a buffer needs to be filled with a value that is determined at runtime

New op that populates a shape with random numbers with a uniform distribution. The rand_uniform op. can implement the Onnx RandomUniform instruction, and can also create the random number sequence necessary to implement Multinomial. (At this time, our Onnx Multinomial parsing generates a random sequence of numbers when parsing as a workaround, so that the resulting program uses the same "random" set every time.)

Arguments: shape, seed. Shape is required; can be static or dynamic. Seed is still optional in this version. If it's not given at inference time, use the value in the creation attribute seed. Update: deleted A boolean use_auto_seed causes any given seed to be ignored.

Makes a version of allocate that takes in dimensions and allocates a buffer
Going to create a simplify_dynamic_ops compiler pass that will use the use_shape_attr flag
The ONNX op ConstantOfShape needs the buffer to be filled with a specific value, so going to make another PR for that and a fill operator

Reduce memory footprint by std::move of temporary (potentially very large) containers.
Minor cleanup for performance optimization: e.g. of Index() calculation -- which can get repeated millions of times in large tensors/vectors in a single Visit.