GitLab

Fixes #1957

Clamping was removed in #1853.

Turns out clamping as necessary to handle overflow/underflow cases. during downcasting, if it overflowed then without clamping it returned infinity.

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.

Changes to the way Pooling operation calculates pooling when there's padding. Old code would clip off any padding values before computing; for instance if an Average pooling window contained 0 1 2 where the 0 is padding, the result was 1.5 instead of 1.0. See Issue 1766

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

* Fix convert for the NaNs

* NaNs can't be compared, use std::isnan()

* formatting

* formatting

* formatting

* add extra tests

* initial

* Added tests and new functionality

* Update optimals handling

* Simplify conditionals

* Ref test, update docs

* Remove comment, suggestion unclear

---------
Co-authored-by: Umang Yadav <29876643+umangyadav@users.noreply.github.com>

Relies on Removed split_single_dyn_dim compile flag #1711
Exposes dynamic_dimension as a opaque object with dynamic_dimensions and optimals
Exposes ONNX dyn_input_dims and default_dyn_dim to run with dynamic batch
Updates api.py to be able to create objects from aggregate initialization (used for dynamic_dimension)
Uses offload copy for now

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

Adds a new GPU compiler pass split_single_dyn_dim that handles when one input parameter has a single non-fixed dynamic_dimension.
commonly occurs for dynamic batch or BERT sequence length
Splits the dynamic shape into several submodules will static input parameters to handle all of the cases in the dynamic_dimension range.
Essentially does what I manually did for the select_module verify tests
Adds a compile option split_single_dyn_dim that toggles the pass on/off. Defaults to false.
Updates verify_program.hpp and run_verify.cpp to allow for the tests to change the compile_options

Fixes #1595

Creates the select_module operator that selects one of the submodules passed to it to run based on the submodule parameters.  The submodule is selected by having the exact same static shapes for the arguments to select_module as the parameters in the submodule

Add dynamic shape support to slice operator.

First draft of this feature doesn't support ops slicing non-fixed, dynamic axes. Resulting shape in such cases is not guaranteed.* Also, onnx parsing doesn't support any arguments other than "axes".

* add dynamic shape support to concat operator.  Includes new op_shape_test and ref_ops_test cases

dyn shape support for Where operator.  Includes shape test, ref_ops test, onx_test. 

* Implement dynamic shapes for scatterND operators.

Dynamic shape support for gathernd op.

Dynamic shape support for gather op.

Extends ONNX Gemm parser to handle dynamic input shapes
Limits ONNX Gemm parsing to 2D input tensors for A and B inputs
As per the ONNX specifications
Changed Gemm ONNX tests to 2D input versions
Add onnx_verify tests for Gemm
Parsing ONNX Gemm links to more than one operator, checking that it produces the correct result

Extends reshape to handle the case of a single non-fixed dynamic_dimension

Extends pad operator to handle dynamic input shapes
Only handles computing the shape for adding constant padding to a dynamic shape
- adds the padding to the min, max, and opt values (unless opt is 0, where it keeps it 0)
- does not handle reflect padding with dynamic shapes

Implements dynamic shapes in reduce_op and all its child operator classes (reduce_max etc.)

No major changes required, use dyn_output and pass dynamic shape when calling compute_shape()
Adds dynamic shape tests

Extends the Argmax operator to handle dynamic input shapes.
Only shape function changes

Extends unsqueeze and squeeze to work for dynamic input shapes
Does not handle the steps parameter
Adds some additional negative axes shape tests

Fix problem with the contiguous operator constructing non-standard shape literals.  A non-standard literal will almost never be used, since a literal is known at compile time.  Added some comments on the intended behavior:

- literal{shape, vector} constructor with a non-standard shape is intended to keep the same ordering as the given vector. The data buffer will be populated such that when the non-standard indexing is used the original order is as given.
- literal{shape, argument} constructor directly copies the data buffer from the argument
- Changed non-standard literal fill() to use tensor_view iterators as it handles non-standard shapes now
- Changed the contiguous ref_ops_test to be more helpful

Extends the pooling operators for dynamic shape inputs

AveragePooling
GlobalAveragePooling
MaxPooling
GlobalMaxPooling
LpNormPooling
GlobalLpNormPooling
y.github.com>

Extends ref transpose operator for dynamic shapes
Make dynamic tests more consistent naming

Extends the ref contiguous operator to handle dynamic shapes
Updates the eliminate_contiguous pass to use the dyn_output struct

Updated Multibroadcast op to have a two input version for dynamic shapes
Current dynamic shape broadcasting logic
dynamic_dimensions must be the same or one of them is {1, 1, 0} or {1, 1, 1}
Works for dyn-dyn, dyn-static, and static-static shape combinations
Changed common.cpp for multibroadcasting for binary ops with dynamic shapes
Extended binary.hpp for dynamic shapes to test the new common.cpp stuff

Upgraded Dockerfiles and fixed tidy issues to make Ubuntu 20.04 and ROCm 5.3.0 the default 

updated GPU pad to now use JIT version.
added range functions for JIT kernels.

Refactor dynamic compute
- add a compute_output_shape object that implicitly converts to a new dyn_output or shape object
- dyn_output object can handle computing the static output shape of an operator given the input arguments shapes
  change an operator's compute function to argument compute(const dyn_output& dyn_out, std::vector<argument> args) to 
  use dyn_output object

Dynamic ref unary functions
-  Included these changes to have an example of the refactored dynamic compute being used
-  Changes to unary base class to handle dynamic shapes
-  Changed elu and leaky_relu to use unary base class and pointwise JIT

Removes use_dynamic_same_auto_pad
Change padding_mode to be used for dynamic padding
Move compute_padded_shape to pad_calc.cpp as it will be used in other dynamic padding cases
Fix same_lower compute_padded_shape bug and add a test.

Rewrites the TF batch norm like operators to other MIGX operators
Removes the code related to batch_norm_inference

Upgrade cppcheck to 2.9 

Using not and or improves readability. The cppcheck rule will help ensure we are doing it consistently.

Has NMS op output a dynamic shape (ONNX spec behavior)
Allows for dynamic input shape to NMS op