This is the successor to the original xdispatch project.
I started the original xdispatch project back in 2011 with an attempt to bring the then new concept of Grand Central Dispatch to more platforms than just iOS and macOS.
The project aimed to provide three deliverables:
- Porting the original libdispatch to Linux and Win32 by using as much of the original code as possible which also implied using ported versions of the libkqueue and libpthread_workqueue interfaces used by libdispatch originally.
- Offering an object oriented C++ interface to allow use of lambda functions which only started to get supported back then as a result of the new C++11 standard. This was in order to offer a cross platform alternative to the blocks feature added by Apple to their clang compiler.
- Providing a wrapper which tried to fit naturally into other Qt components like QtConcurrent.
These goals were accomplished with the latest feature version 0.7.2 published in January 2013 and development focusing on maintenance and bug fixes ever since.
Dial forward to June 2020 when I had gathered several lessons learned from using xdispatch in multiple projects from little university projects to full media servers sold on the app store:
-
The concept of queues really makes things easier not only for scaling on multiple cores but also for efficient serialization of acccess to resources like an OpenGL render context.
-
A lot of systems still have thread affinities assuming that certain tasks will always run on the same threads (again such as OpenGL render contexts).
The latter does not fit very well with a concept using a thread pool. Pooling of threads implies there must never be any assumptions on thread affinities.
This is when the idea of xdispatch v2 was born. It would continue to maintain close compatibility with the original xdispatch v1 C++ interface. However instead of trying to chase porting the latest features of the original libdispatch (and hence depending on faking Darwin specific interfaces) it would focus on providing a consistent interface for different thread pool concepts available on the respective platforms and programming environment.
This resulted in setting up the following rules:
- The xdipsatch2 core featureset must not depend on anything but a C++11 compliant compiler and STL.
- The entire interface is abstracted such that serial or parallel queues as well as timers and notifiers can be implemented by different backends all while maintaining feature consistency and allowing to mix xdispatch objects powered by different backends as required.
- Support a minimal, self hosted "naive" implementation as feature reference which might be less efficient but can always be supported within the limits set up in rule 1.
All further decisions are based on these rules. Backends need to implement the xdispatch::ibackend interface when they want to provide the full feature set on their own or may alternatively hook into the xdispatch::naive::ithread and xdispatch::naive::ithread_pool interfaces to augment the naive implementation with more efficient threading primitives.
The following backends have been implemented so far:
| Backend | Wrapped Library | Conditions |
|---|---|---|
| naive | std::thread | Always available |
| dispatch | libdispatch | dispatch.h header and interface found when building xdispatch2 |
| qt | Qt5 | Qt5 found when building xdispatch2 |
Numbers have been collected on an AMD Ryzen 3900x 12 core processor. They try to quantify the overhead induced by enqueuing and executing preallocated xdispatch::operation objects. The operation is dispatched 100 000 times using the respective primitive. Times describe the overhead for a single invocation. Sources can be found in tests/cxx_benchmark.cpp.
| Backend | Serial Queue | Parallel Queue | Group |
|---|---|---|---|
| naive | 187 nsec | 295 nsec | 644 nsec |
| dispatch | 585 nsec | 448 nsec | 780 nsec |
| qt | 544 nsec | 411 nsec | 693 nsec |
xdispatch2 is provided as conan v1 package through the registry at https://mirrors.emzeat.de.
Simply add this as a remote and install the package including all dependencies:
conan remote add emzeat https://codeberg.org/api/packages/emzeat/conan/
conan install --build=missing xdispatch/2.1.3@emzeat/ossAvailable options to configure enabled backends can be found by looking at the conanfile.py. Refer to the conan documentation for further details on consuming the package in your project.
As in v1 CMake is used as the build system wrapped with my mz-cmaketools for convenience.
These get pulled as git submodule so make sure to run git submodule update --init before trying to build.
Make sure you have cmake, g++ and ninja installed and available in your path. Use cmake to configure and build the project:
mkdir build/release
cd build/release
cmake -G Ninja -D CMAKE_BUILD_TYPE=Release -S ../.. -B .
cmake --build .In case you want to use Conan 1.x available in your path to pull dependencies run as:
MZ_USE_CONAN=true cmake -G Ninja -D CMAKE_BUILD_TYPE=Release -S ../.. -B .
cmake --build .The resulting binaries can be found in the bin directory of the build folder created above. You simply need to pull the include directory in your compiler's include path and link the xdispatch library with your program.
A (very) minimal example might look like this:
#include <xdispatch/dispatch>
#include <iostream>
int main(int, char**)
{
xdispatch::global_queue().async([]
{
std::cout << "PING" << std::endl;
xdispatch::main_queue().async([]{
std::cout << "PONG" << std::endl;
exit(0);
});
});
xdispatch::exec();
return 0;
}Have a look at the implemented tests for more elaborated examples.
We welcome any contributions.
We use GitHub to handle bug reports, feature requests and questions - please use "issues" for any of these. Code changes and documentation improvements get handled via pull requests.
In order to foster open exchange, all communication should happen using the English language.
Any changes to this repository get done via GitHub's standard workflow. See GitHub's First Contributions document for an introduction.
See the list of contributors generated by Github.
This project is licensed under the Apache 2.0 License.