A simple С++ library for building conveyor-like pipelines
This library allows users to easily create linear parallel pipelines.
For instance. imagine you have three functions f1, f2 and f3, and your result is calculated like
output = f3(f2(f1(input)));
So, if you continuously feed this pipeline with data, your timing looks like this:
f1: |XXXXXX|..............|XXXXXX|..............|XXXXXX|..............
f2: .......|XXXXXXXXXX|..........|XXXXXXXXXX|..........|XXXXXXXXXX|...
f3: ..................|XX|..................|XX|..................|XX|
An easy way o speed up the process is to execute the function chain in parallel, i.e. create several threads for each f3(f2(f1(...))) call. However, sometimes it is not possible, e.g. your functions depend on some hardware calculation modules with exclusive access or aren't reentrant.
Conveyourpp lib solves this problems by simplifying building of a parallel worker chain. For instance, the example above with conveyorpp (with zero-queue settings) looks like this:
f1: |XXXXXX||XXXXXX|...|XXXXXX|....|XXXXXX|....|XXXXXX|....
f2: .......|XXXXXXXXXX||XXXXXXXXXX||XXXXXXXXXX||XXXXXXXXXX|
f3: ..................|XX|........|XX|........|XX|.........
So if your calculation can be divided in several stages, you can utilize all the computational power you've got.
To install conveyourpp, clone this repository and perform the following actions:
mkdir build && cd build
cmake ..
make
Optionally you can install it via
sudo make install
Imagine you have three functions:
int f1(float);
double f2(int);
bool f3(double);
You can parallelize their execution by:
auto chainer = cnvpp::MakeChainer(f1, f2, f3);
This line of code creates execution chain f1->f2->f3, which is equivalent to f3(f2(f1(...)))
To feed this conveyor with input data, simply call:
int x = 42;
chainer.Call(x);
To obtain result of the computation, call
auto result = chainer.GetResult();
Method GetResult blocks until result is available. If the last function's return type is void,method GetResult is excluded from compilation.
Lambda-functions need to be processed with some special care. In particular, user has to manually describe their input and output types, e.g.
double multiplier = 2.5;
auto l1 = [=] (int x) { return (float)x * multiplier; };
auto l2 = [&] (float y) { return (size_t)(y / multiplier + 1); };
auto chainer = cnvpp::MakeChainer(
cnvpp::LambdaWrapper<float, int>(l1),
cnvpp::LambdaWrapper<size_t, float>(l2),
};
- Add more test and examples
- Add more documentation and diagrams
- Find a way to get rid of LambdaWrapper class and handle functions, class methods and lambdas in a unified way
- Make internal queuing mechanism more customizable (at the current moment each conveyor stage has internal queue with size 5)
- Add mechanisms for auto-balancing the load (e.g. user adds a sequence of actions to perform, and conveyorpp's runtime determines optimal amount of threads to use, not just execute each stage in a separate thread)