 |
Ginkgo
Generated from pipelines/1589998975 branch based on develop. Ginkgo version 1.10.0
A numerical linear algebra library targeting many-core architectures
|
|
Ginkgo
Generated from pipelines/1589998975 branch based on develop. Ginkgo version 1.10.0
A numerical linear algebra library targeting many-core architectures
|
We are glad that you are interested in contributing to Ginkgo. Please have a look at our coding guidelines before proposing a pull request.
GINKGO_DEVEL_TOOLS needs to be set to on to commit. This requires clang-format to be installed. See Automatic code formatting for more details. Once installed, you can run make format in your build/ folder to automatically format your modified files. As make format unstages your files post-formatting, you must stage the files again once you have verified that make format has done the appropriate formatting, before committing the files.Ginkgo is divided into a core module with common functionalities independent of the architecture, and several kernel modules (reference, omp, cuda, hip, dpcpp) which contain low-level computational routines for each supported architecture.
Some header files from the core module have to be extended to include special functionality for specific architectures. An example of this is core/base/math.hpp, which has a GPU counterpart in cuda/base/math.hpp. For such files you should always include the version from the module you are working on, and this file will internally include its core counterpart.
You can use and call functions of existing classes inside a kernel (that are defined and not just declared in a header file), however, you are not allowed to create new instances of a polymorphic class inside a kernel (or in general inside any kernel module like cuda/hip/omp/reference) as this creates circular dependencies between the core and the backend library. With this in mind, our CI contains a job which checks if such a circular dependency exists. These checks can be run manually using the -DGINKGO_CHECK_CIRCULAR_DEPS=ON option in the CMake configuration.
For example, when creating a new matrix class AB by combining existing classes A and B, the AB::apply() function composed of invocations to A::apply() and B::apply() can only be defined in the core module, it is not possible to create instances of A and B inside the AB kernel files. This is to avoid the aforementioned circular dependency issue. An example for such a class is the Hybrid matrix format, which uses the apply() of the Ell and Coo matrix formats. Nevertheless, it is possible to call the kernels themselves directly within the same executor. For example, cuda::dense::add_scaled() can be called from any other cuda kernel.
Ginkgo uses git, the distributed version control system to track code changes and coordinate work among its developers. A general guide to git can be found in its extensive documentation.
In Ginkgo, we prioritize keeping a clean history over accurate tracking of commits. git rebase is hence our command of choice to make sure that we have a nice and linear history, especially for pulling the latest changes from the develop branch. More importantly, rebasing upon develop is required before the commits of the PR are merged into the develop branch.
With software sustainability and maintainability in mind, it is important to write commit messages that are short, clear and informative. Ideally, this would be the format to prefer:
You can refer to this informative guide for more details.
Git has a nice feature where it allows you to add a co-author for your commit, if you would like to attribute credits for the changes made in the commit. This can be done by:
In the Ginkgo commit history, this is most common associated with suggested improvements from code reviews.
develop branch is the default branch to submit PR's to. From time to time, we merge the develop branch to the main branch and create tags on the main to create new releases of Ginkgo. Therefore, all pull requests must be merged into develop.WIP if you are still working on it, Ready for Review when it is ready for others to review it.ginkgo-project.READY TO MERGE. At this point the creator/assignee of the PR needs to verify that the branch is up to date with develop and rebase it on develop if it is not.Ginkgo uses pre-commit to automatically apply code formatting when committing changes to git. What formatting is applied is managed through ClangFormat with a custom .clang-format configuration file (mostly based on ClangFormat's Google style). Make sure you have pre-commit set up and running properly before committing anything that will end up in a pull request against ginkgo-project/ginkgo repository. In addition, you should never modify the .clang-format configuration file shipped with Ginkgo. E.g. if ClangFormat has trouble reading this file on your system, you should install a newer version of ClangFormat, and avoid commenting out parts of the configuration file.
ClangFormat is the primary tool that helps us achieve a uniform look of Ginkgo's codebase, while reducing the learning curve of potential contributors. However, ClangFormat configuration is not expressive enough to incorporate the entire coding style, so there are several additional rules that all contributed code should follow.
Note: To learn more about how ClangFormat will format your code, see existing files in Ginkgo, .clang-format configuration file shipped with Ginkgo, and ClangFormat's documentation.
Filenames use snake_case and use the following extensions:
.cpp.hpp.cu.cuh.hip.cpp.hip.hpp.hpp.inc.cmake.shNote: A C++ source/header file is considered a CUDA file if it contains CUDA code that is not guarded with #if guards that disable this code in non-CUDA compilers. I.e. if a file can be compiled by a general C++ compiler, it is not considered a CUDA file.
Macros (both object-like and function-like macros) use CAPITAL_CASE. They have to start with GKO_ to avoid name clashes (even if they are #undef-ed in the same file!).
Variables use snake_case.
Constants use snake_case.
Functions use snake_case.
Structures and classes which do not experience polymorphic behavior (i.e. do not contain virtual methods, nor members which experience polymorphic behavior) use snake_case.
All other structures and classes use CamelCase.
All structure / class members use the same naming scheme as they would if they were not members:
Additionally, non-public data members end with an underscore (_).
Namespaces use snake_case.
CamelCase, for example ValueType.snake_case, for example subwarp_size.Spaces and tabs are handled by ClangFormat, but blank lines are only partially handled (the current configuration doesn't allow for more than 2 blank lines). Thus, contributors should be aware of the following rules for blank lines:
struct x {
};
} // namespace foo
namespace bar {
namespace baz {
void f();
} // namespace baz
} // namespace bar
```
2. _exception_: in header files whose only purpose is to _declare_ a bunch
of functions (e.g. the `*_kernel.hpp` files) these declarations can be
separated by only 1 blank line (note: standard rules apply for all other
statements that might be present in that file)
3. _exception_: "related" statement can have 1 blank line between them.
"Related" is not a strictly defined adjective in this sense, but is in
general one of:
1. overload of a same function,
2. function / class template and it's specializations,
3. macro that modifies the meaning or adds functionality to the
previous / following statement.
However, simply calling function `f` from function `g` does not imply
that `f` and `g` are "related".
exception: there is no blank line between an access modifier (private, protected, public) and the following statement. example: ```c++ class foo { public: int get_x() const noexcept { return x_; }
int &get_x() noexcept { return x_; }
private: int x_; }; ```
The concrete ordering will be done by clang-format. Here are the rules that clang-format will follow. In general, all include statements should be present on the top of the file, ordered in the following groups, with one blank lines between each group:
core/foo/bar.hpp included in core/foo/bar.cpp, or in the unit testcore/test/foo/bar.cpp)vector)omp.h)papi.h)Example: A file core/base/my_file.cpp might have an include list like this:
This section presents the handling of the main header attributed to a file. For a given file, the main header is the header that contains the declarations of the functions, classes, etc., which are implemented in this file. In the previous example, this would be #include "ginkgo/core/base/my_file.hpp". The clang-format tool figures out the main header. The only intervention form a contributor is to always include the main header using "...".
Please note that this only applies to implementation files, so files ending in .cpp or .cu.
GINKGO_CHECK_CIRCULAR_DEPS enabled, this property is explicitly checked.iwyu (Include what you use) tool can be used to make sure that the headers are self-sufficient and that the compiled files ( .cu, .cpp, .hip.cpp ) include only what they use. A CI pipeline is available that runs with the iwyu tool. Please be aware that this tool can be incorrect in some cases.Single line statements should be avoided in all cases. Use of brackets is mandatory for all control flow constructs (e.g. if, for, while, ...).
C++ supports declaring / defining multiple variables using a single type-specifier. However, this is often very confusing as references and pointers exhibit strange behavior:
For this reason, always declare each variable on a separate line, with its own type-specifier.
All alignment in CMake files should use four spaces.
Macros in CMake do not have a scope. This means that any variable set in this macro will be available to the whole project. In contrast, functions in CMake have local scope and therefore all set variables are local only. In general, wrap all piece of algorithms using temporary variables in a function and use macros to propagate variables to the whole project.
All Ginkgo specific variables should be prefixed with a GINKGO_ and all functions by ginkgo_.
To facilitate easy development within Ginkgo and to encourage coders and scientists who do not want get bogged down by the details of the Ginkgo library, but rather focus on writing the algorithms and the kernels, Ginkgo provides the developers with a few helper scripts.
A create_new_algorithm.sh script is available for developers to facilitate easy addition of new algorithms. The options it provides can be queried with
The main objective of this script is to add files and boiler plate code for the new algorithm using a model and an instance of that model. For example, models can be any one of factorization, matrix, preconditioner or solver. For example to create a new solver named my_solver similar to gmres, you would set the ModelType to solver and set the ModelName to gmres. This would duplicate the core algorithm and kernels of the gmres algorithm and replace the naming to my_solver. Additionally, all the kernels of the new my_solver are marked as GKO_NOT_IMPLEMENTED. For easy navigation and .txt file is created in the folder where the script is run, which lists all the TODO's. These TODO's can also be found in the corresponding files.
We provide a cuda2hip script that converts cuda kernel code into hip kernel code. Internally, this script calls the hipify script provided by HIP, converting the CUDA syntax to HIP syntax. Additionally, it also automatically replaces the instances of CUDA with HIP as appropriate. Hence, this script can be called on a Ginkgo CUDA file. You can find this script in the dev_tools/scripts/ folder.
Ginkgo uses the GTest framework for the unit test framework within Ginkgo. Writing good tests are extremely important to verify the functionality of the new code and to make sure that none of the existing code has been broken.
TEST_FTYPED_TEST.ReferenceExecutor, are meant to be single threaded reference implementations. Therefore, tests for reference kernels need to be performed with data that can be as small as possible. For example, matrices lesser than 5x5 are acceptable. This allows the reviewers to verify the results for exactness with tools such as MATLAB.Documentation uses standard Doxygen.
Make use of @internal doxygen tag. This can be used for any comment which is not intended for users, but is useful to better understand a piece of code.
The documentation tags which use an additional name should be followed by two spaces in order to better distinguish the text from the doxygen tag. It is also possible to use a line break instead.
There are two main steps:
doc/ folder (you can copy it from the example most relevant to you) and adapt your example names and such, then you can modify the actual documentation.tooltip: A short description of the example.short-intro: The name of the example.results.dox: Run the example and write the output you get.kind: The kind of the example. For different kinds see the documentation. Examples can be of basic, techniques, logging, stopping_criteria or preconditioners. If your example does not fit any of these categories, feel free to create one.intro.dox: You write an explanation of your code with some introduction similar to what you see in an existing example most relevant to you.builds-on: You write the examples it builds on.doc/kind file in the example documentation.These are global objects and are shared inside the same translation unit. Therefore, whenever its state or formatting is changed (e.g. using std::hex or floating point formatting) inside library code, make sure to restore the state before returning the control to the user. See this stackoverflow question for examples on how to do it correctly. This is extremely important for header files.
By default, the -DCMAKE_CXX_FLAGS should be set to -Wpedantic to emit pedantic warnings by default. The CI system additionally also has a step where it compiles for pedantic warnings to be errors.
To facilitate finding circular dependencies issues (see Using library classes for more details), a CI step no-circular-deps was created. For more details on its usage, see this pipeline, where Ginkgo did not abide to this policy and PR #278 which fixed this. Note that doing so is not enough to guarantee with 100% accuracy that no circular dependency is present. For an example of such a case, take a look at this pipeline where one of the compiler setups detected an incorrect dependency of the cuda module (due to jacobi) on the core module.
1.8.16