GitLab

Introduce dilations attribute to pooling operators reference implementation.

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.

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.

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

Add support for the QLinearMatMul onnx operator

Makes ConstantOfShape work with a variable dimensions input

* 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>

Add parsing support for value_float, value_floats, value_int, value_ints attributes
Disable failing tests
Resolves Test failures due to IndexError: _Map_base::at migraphx-benchmark/AMDMIGraphX#76

Adds support for dynamic input shape in pooling operator along with auto-padding. This combination requires that the padding (and therefore the output shape) can't be computed until runtime.

Allows slice to work with variable starts, ends, and axes input
Outputs a dynamic shape even with a static shape data input when the starts and ends are variable

* add quant_dot fusion, clip literal opt

Fixes #1746

BatchNorm only has x as the runtime input parameter for the following equation. All the other parameters are compile-time constants and related operations can be const-folded before quantizing to fp16 to preserve precision.

Renames deconvolution -> convolution_backwards to be more consistent with the literature
Note: this is not the cross-correlation operator (which is the adjoint of convolution). This is technically a standard convolution operator combined with an upsampling operator rather than a downsampling operator.
Adds unit tests for the padding, strides, dilations, and other op attributes.
Throws on auto_pad attribute since it has not been implemented
Previously it read the attribute and set it but then did nothing with it
Extended for dynamic shapes
Does not support using asymmetric padding (padding_L != padding_R) and output_shape with dynamic shapes.

Updates ref version of clip to work with dynamic shapes
Encountered in agentmodel

* fix parse_instancenorm to create broadcast and multibroadcast instructions with two dynamic shape arguments instead of 1.  Their make_op() functions don't support dynamic shapes when called with one input.  This caused an error when parsing an ONNX 3duunet model

* Use add_common_op() to create multibroadcast op.

* add verification and parsing test for instance_norm with dynamic input.  Parse test doesn't pass.

* fix for test; still doesn't pass

* another fix for test; still doesn't pass

* work in progress, instance_norm_dyn_batch_test works but instance_norm_test doesn't

* fix onnx instancenorm tests to match parser changes.  Passes all check tests

* Updated comments explaining usage of add_common_op()

* hand-merged conflicts with develop

* fix instance_norm_half_test after merge

* add Onnx test instance_norm_dyn_batch_half_test

* add shape test cases broadcast_1in_dyn_error and multibroadcast_1in_dyn_error_0

By converting to fp32 : fp16 3d-unet model accuracy comes out the same as FP32 accuracy.

By using reduce_sum method on Fp16 : accuracy comes out ~0.9% lower compared to fp32 while keeping entire model in fp16.

Fixes the above behavior
This needs to be changed to allow for setting static shapes with map_dyn_input_dims since you cannot also use map_input_dims

Makes the optimals into a std::set<std::size_t>
Changes shape object functions to handle the opts change
Changes to convolution, flatten, pooling, and convolution in that they no longer calculate the output optimal dimensions. Instead returns empty opts. Will need to change this in the future if we want to support dynamic shapes fully.
Many changes to tests and shape calls with respect to the new optimals