Skip to content

Latest commit

 

History

History
112 lines (86 loc) · 5.76 KB

changes.md

File metadata and controls

112 lines (86 loc) · 5.76 KB
Raw

API Changes

Op Definition

  • Every Op class must declare a static constexpr NodeTypeInfo type_info{name, version} in the class definition and define it in the .cpp file. See any op definition for an example.
  • The boolean function is_type<T> is for testing if a node is the op T.
  • T as_type_ptr<T>() and T as_type<T>() will upcast Node to an explicit op class if it is of class T, or nullptr if it is not.

Backend library interface

  • Each backend BACKEND needs to define the macro ${BACKEND}_API appropriately to import symbols referenced from outside the library and to export them from within the library. See any of the ${backend}_backend_visibility.hpp files for an example.
  • The CMakeLists.txt file for a backend defines ${BACKEND}_BACKEND_DLL_EXPORTS. target_compile_definitions(${backend}_backend PRIVATE ${BACKEND}_BACKEND_DLL_EXPORTS)
  • Each backend must define a function ngraph_register_${backend}_backend that registers a backend constructor function and ensures that initializations are performed. ngraph/src/runtime/cpu/cpu_backend.cpp has an example that includes initializations. Remove the old backend constructor code.

Passes

  • LikeReplacement pass must be run by all transformers.
  • ngraph::pass::FusionType is now an enum class. Constant values defined by FusionType are created for backward compatibility and will be removed in future releases.

Nodes, Parameters

  • Nodes is now NodeVector
  • Parameters is now ParameterVector
  • NodeVector, ParameterVector, AxisVector, AxisSet, Shape, Stride, Coordinate, and CoordinateDiff are now classes, not type aliases.
  • PrimaryTensorView is now TensorView (and will merge into Tensor)
  • copy_with_new_args is protected; use copy_with_new_inputs which takes an OutputVector as an argument and preserves control dependencies.

Changes to ops

  • The namespace ngraph::op is only for actual ops. Helpers have been moved into ngraph::op::util:

    • BinaryElementwiseArithmetic
    • BinaryElementwiseComparison
    • BinaryElementwise
    • RequiresTensorViewArgs
    • UnaryElementwiseArithmetic
    • UnaryElementwise Ops defined outside of nGraph core will need to get the base class from ngraph::op::util and change the include file to #include "ngraph/ops/util/requires_tensor_view_args.hpp", etc.

    See any of the core ops for an example.

Changes to convolution and pooling ops

  • Backprop ops have been added for convolution ops.

  • The convolution and pooling ops have had several methods/fields renamed, to reflect a shift in terminology from "images" to "data". Generally this just means that you will have to s/image_batch/data_batch/ and s/image_dilation_strides/data_dilation_strides/.

  • The following functions have been removed:

    • AvgPool: get_channel_count get_input_image_physical_shape get_input_image_virtual_shape get_output_image_shape get_batch_size get_image_dimension_count
    • MaxPool: get_channel_count get_input_image_shape get_output_image_shape get_batch_size get_image_dimension_count
    • Convolution: get_input_channel_count get_output_channel_count get_input_image_physical_shape get_input_image_virtual_shape get_output_image_shape get_window_physical_shape get_window_virtual_shape get_batch_size get_image_dimension_count

    All of the above information can be inferred from the shapes and parameters of the op.

  • The AvgPool operator has a new attribute governing whether or not padding-region values are considered when computing a given window's average: include_padding_in_avg_computation. One of the class constructors adds this to the parameter list, and the others use a default value of false which matches the old behavior.

Negative convolution padding

Convolution now allows negative padding. This means that the padding_below and padding_above arguments now take type CoordinateDiff instead of Shape. CoordinateDiff is an alias for std::vector<std::ptrdiff_t>, which "is like size_t but is allowed to be negative". Callers may need to be adapted.

Changes to Concat op

  • get_concatenation_axis was renamed to get_axis. In order to provide backward compatibility get_concatenation_axis is now alis of get_axis method
  • set_concatenation_axis was renamed to set_axis. In order to provide backward compatibility set_concatenation_axis is now alis of set_axis method

Parameter and Function no longer take a type argument.

Changes to Tensor read and write methods

The read and write methods on ngraph::runtime::Tensor which take a tensor_offset as the second of three arguments have been deprecated. The replacement read and write methods take two arguments, the buffer pointer and the size. For any references to the deprecated methods remove the second argument, the tensor offset, to update to the new API. These old read/write methods have been decorated with deprecated warnings which may be enabled by setting -DNGRAPH_DEPRECATED_ENABLE=ON.

To update, remove the passed argument. For example,

// Old
make_shared<Parameter>(make_shared<descriptor::TensorViewType>(element::f32, Shape{2, 4}));
// New (remove TensorViewType)
make_shared<Parameter>(element::f32, Shape{2, 4});

// Old
make_shared<Function>(results, result_type, parameters);
// New
make_shared<Function>(results, parameters);

The runtime::Tensor methods to get_tensor<> and write(std::vector&) have been removed to the unit test directory under utils/test_tool.hpp read_vector and write_vector.

Changes to reshape op utils

Utility functions from src/ngraph/op/util/reshape.hpp, placed at namespace ngraph::op::util:

  • reshape
  • reorder_axes
  • transpose
  • flatten

Are moved to new location: src/ngraph/builder/reshape.hpp to namespace ngraph::builder.