ONNX Runtime coding conventions and standards#

C++ Code Style#

Google style from https://google.github.io/styleguide/cppguide.html with a few minor alterations:

Max line length 120
- Aim for 80, but up to 120 is fine.
Exceptions
- Allowed to throw fatal errors that are expected to result in a top level handler catching them, logging them and terminating the program.
Non-const references
- Allowed
- Use a non-const reference for arguments that are modifiable but cannot be nullptr so the API clearly advertises the intent
- Const correctness and usage of smart pointers (shared_ptr and unique_ptr) is expected. A non-const reference equates to “this is a non-null object that you can change but are not being given ownership of”.
Prefer passing gsl::span<const T> by value (or std::span when supported) as input arguments when passing const references to containers with contiguous storage (like std::vector). This allows the function to be container independent, and the argument to represent arbitrary memory spans or sub-spans. The below examples allow the client code to use either std::vector or InlinedVector. An instance of a gsl::span would be created automatically.

/// Instead of
void foo(const std::vector<int64_t>&);

/// Use to pass any contiguous const container containing int64_t
// Now you can seamlessly pass either `std::vector`, `InlinedVector`, `std::array` or `gsl::span` as an argument.
void foo(gsl::span<const int64_t>);

// Example with pointer to const data. Instead of
void foo(const std::vector<const Node*>&);

// Use
void foo(gsl::span<const Node* const>);

Prefer returning gsl::span<const T> by value instead of a const reference to a contiguous member container. Prefer returning gsl::span instead of a pointer referring to a chunk of memory. The size is also included in the span. For example,

// Instead of
const std::vector<int64_t>& foo();

// Return a span by value
gsl::span<const int64_t> foo();

// Instead of
const int64_t* foo();

// Return a span by value
gsl::span<const int64_t> foo();

However, std::initializer_list<T> is not automatically convertible to a gsl::span<const T>. Use AsSpan({1, 2, 3}) defined at core/common/span_utils.h to convert std::initializer_list<T> to a span. You can also use std::array. For example,

// Original code
void foo(const std::vector<std::string>&);

foo({"abc", "dbf"}); // Works

// After refactoring to gsl::span it would no longer compile. Use AsSpan().
void foo(gsl::span<const std::string>);

foo(AsSpan<std::string>{"abc", "dbf"}); // Works

Prefer passing std::string_view by value instead of const std::string&. Make sure that the lifespan of a std::string instance ecplises the lifespan of the corresponding std::string_view instance.
using namespace permitted with limited scope
- Not allowing using namespace at all is overly restrictive. Follow the C++ Core Guidelines:
  - SF.6: Use using namespace directives for transition, for foundation libraries (such as std), or within a local scope (only)
  - SF.7: Don’t write using namespace at global scope in a header file

Containers to use#

Onnxruntime aims to reduce latency and latency variance by minimizing the amount of dynamic memory allocations.

The use of the following container typedefs to reduce memory allocations is required:
- Use TensorShapeVector typedef to build or modify shapes from core/framework/tensor_shape.h. It is based on a vector implementation that features small buffer optimization. Its small buffer size is the same to that of in TensorShape.
- Use InlinedVector<T> typedef instead of std::vector defined at core/common/inlined_containers_fwd.h. By default, it provides 64 bytes of inlined storage. You can customize inlined size with the second template non-type parameter N.
- Use InlinedHashSet<T> and InlinedHashMap<T> typedefs from core/common/inlined_containers.h. These are drop-in replacements for std::unordered_set/map that store their keys and values in one continuous buffer and reduce the number of allocations. They also do not allocate an end node when default constructed. Note, that these Hash containers do not provide pointer stability. std::map and std::set can often be replaced by hash containers as well.
- For the node based containers where pointer stability is required, use NodeHashSet and NodeHashMap. Although node based, they are more cache friendly.
- Use core/common/inlined_containers_fwd.h to forward declare any of the above container types.
- Consider using std::string_view for use in containers to reduce the number of allocations and avoid string duplication. Keep in mind that the lifespan of the objects being referred to must eclipse the lifespan of the corresponding std::string_view.
- We have selected to use Abseil library for the above typedefs. Abseil container documentation is here.
- Do not use Abseil library or absl namespace directly. We should be able to build Onnxruntime without Abseil.
- Use onnxruntime/tools/natvis/abseil-cpp.natvis for the above containers visualizations and debugging help in VS Studio and VS Code.
Prefer using reserve() and not resize() on vectors. resize() default constructs all the elements for the size which can be expensive/noticeable even if the type is trivial. Default values are rarely used in practice and it becomes a waste. Construction like std::vector<int>(10, 0) is the same as resize() and is potentially wasteful.
Use reserve() on hash containers and vectors. For example,

#include "core/common/inlined_containers.h"

void foo(gsl::span<const std::string> names) {
  // For local processing, names are still valid
  // use std::string_view to avoid duplicate memory allocations.
  // same code would work with std::unordered_set if built without Abseil
  InlinedHashSet<std::string_view> unique_names;
  unique_names.reserve(names.size());
  unique_names.insert(names.cbegin(), names.cend());
}

Other#

Qualify usages of auto with const, *, & and && where applicable to more clearly express the intent
When adding a new class, disable copy/assignment/move until you have a proven need for these capabilities. If a need arises, enable copy/assignment/move selectively, and when doing so validate that the implementation of the class supports what is being enabled.
- Use ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE initially
- See the other ORT_DISALLOW_* macros in microsoft/onnxruntime
Sometimes, std::unique_ptr might be considered for delayed or optional construction of objects or members of classes. Instead, use std::optional as appropriate to reduce the number of allocations.
Don’t use else after return. see: https://llvm.org/docs/CodingStandards.html#don-t-use-else-after-a-return
Don’t overuse std::shared_ptr. Use std::shared_ptr only if it’s not clear when and where the object will be de-allocated. See also: https://isocpp.github.io/CppCoreGuidelines/CppCoreGuidelines#Rf-shared_ptr
Avoid using the long type, which could be either 32 bits or 64 bits.
If there is a legitimate need to allocate objects on the heap, prefer using std::make_unique(). References for the reasoning:
Use SafeInt when calculating the size of memory to allocate to protect against overflow errors
- #include "core/common/safeint.h"
- search for SafeInt<size_t> in the code for examples
The following C++ warnings should never be disabled in onnxruntime VC++ projects(Required by Binskim).
1. 4018 ‘token’ : signed/unsigned mismatch
2. 4146 unary minus operator applied to unsigned type, result still unsigned
3. 4244 ‘argument’ : conversion from ‘type1’ to ‘type2’, possible loss of data. For example, casting a int64_t to size_t.
4. 4267 ‘var’ : conversion from ‘size_t’ to ‘type’, possible loss of data.
5. 4302 ‘conversion’ : truncation from ‘type 1’ to ‘type 2’
6. 4308 negative integral constant converted to unsigned type
7. 4532 ‘continue’ : jump out of __finally/finally block has undefined behavior during termination handling
8. 4533 initialization of ‘variable’ is skipped by ‘instruction’
9. 4700 uninitialized local variable ‘name’ used
10. 4789 buffer ‘identifier’ of size N bytes will be overrun; M bytes will be written starting at offset L
11. 4995 ‘function’: name was marked as #pragma deprecated
12. 4996 Your code uses a function, class member, variable, or typedef that’s marked deprecated

Clang-format#

Clang-format will handle automatically formatting code to these rules. There’s a Visual Studio plugin that can format on save at https://marketplace.visualstudio.com/items?itemName=LLVMExtensions.ClangFormat, or alternatively the latest versions of Visual Studio 2017 include clang-format support.

There is a .clang-format file in the root directory that has the max line length override and defaults to the google rules. This should be automatically discovered by the clang-format tools.

Code analysis#

Visual Studio Code Analysis with C++ Core guidelines rules enabled is configured to run on build for the onnxruntime_common, onnxruntime_graph and onnxruntime_util libraries. Updating the onnxruntime_framework and onnxruntime_provider libraries to enable Code Analysis and build warning free is pending.

Code changes should build with no Code Analysis warnings, however this is somewhat difficult to achieve consistently as the Code Analysis implementation is in fairly constant flux. Different minor releases may have less false positives (a build with the latest version may be warning free, and a build with an earlier version may not), or detect additional problems (an earlier version builds warning free and a later version doesn’t).

We use BinSkim Binary Analyzer to scan our binaries.

Unit Testing and Code Coverage#

There should be unit tests that cover the core functionality of the product, expected edge cases, and expected errors. Code coverage from these tests should aim at maintaining over 80% coverage.

All changes should be covered by new or existing unit tests.

In order to check that all the code you expect to be covered by testing is covered, run code coverage in Visual Studio using ‘Analyze Code Coverage’ under the Test menu.

There is a configuration file in onnxruntime/VSCodeCoverage.runsettings that can be used to configure code coverage so that it reports numbers for just the onnxruntime code. Select that file in Visual Studio via the Test menu: Test -> Test Settings -> Select Test Settings File.

Using Show Code Coverage Coloring will allow you to visually inspect which lines were hit by the tests. See https://docs.microsoft.com/en-us/visualstudio/test/using-code-coverage-to-determine-how-much-code-is-being-tested?view=vs-2017.

Python Code Style#

Follow the Black formatter’s coding style when possible. A maximum line length of 120 characters is allowed for consistency with the C++ code.

Please adhere to the PEP8 Style Guide. We use Google’s python style guide as the style guide which is an extension to PEP8.

Code can be validated with flake8 using the configuration file in the root directory called .flake8.

Use pyright, which is provided as a component of the pylance extension in VS Code for static type checking.

Auto-formatting is done with black and isort. The tools are configured in pyproject.toml. From anywhere in the repository, you can run

black .
isort .

to format Python files.

Use pydocstyle to lint documentation styles. pydocstyle is enabled in VS Code.

IDEs#

VS Code#

VS Code is automatically configured with workspace configurations.

For Python development is VS Code, read this tutorial for more information.

PyCharm#

Follow black’s documentation to set up the black formatter for PyCharm.

Testing#

We use the Python built-in unittest framework for creating unit tests and pytest to run them. Use pytest to create tests only when unittest does not fit the need.

Style#

Test the behavior, instead of the implementation. To make what a test is testing clear, the test methods should be named following the pattern test_<method or function name>_<expected behavior>_[when_<condition>].

e.g. test_method_x_raises_error_when_dims_is_not_a_sequence

Objective-C/C++ Code Style#

Please follow the Google Objective-C/C++ Style Guide.

Clang-format can be used to format Objective-C/C++ code. The .clang-format file is in the repository root directory.