Swarm

Swarm is a multithreading library with no other dependancies than STL. It allows you to execute functions accross multiple threads.

Usage

The function to execute with Swarm need to have this signature : void(uint32_t, uint32_t)

The first argument is the executing thread's id and the second is the number of threads currently working on the function.

When a function to execute is sent to Swarm, a WorkGroup object is returned. In order to wait for the funtion to end, waitExecutionDone has to be called on the WorkGroup. This also allows you to run two groups on different functions asynchronously.

uint32_t thread_count(4);
WorkGroup group_1 = swarm.execute(function_1, 2);
WorkGroup group_2 = swarm.execute(function_2, 2);

// function_1 and function_2 are being processed at the same time

group_1.waitExecutionDone();
group_2.waitExecutionDone();

Example

1. Include header

#include "swarm.hpp"

2. Create concurrent function

#include <vector>

const size_t vec_size(100000000);

std::vector<float> v1(vec_size);
std::vector<float> v2(vec_size);
std::vector<float> v3(vec_size);

void job(uint32_t worker_id, uint32_t worker_count)
{
    // Number of values for which the thread is responsible
    const uint32_t step = vec_size / worker_count;
    // First value for the thread
    const uint32_t start_index = worker_id * step;
    // Last value
    const uint32_t end_index = (worker_id < worker_count - 1) ? start_index + step : vec_size - 1;
    
    // The actuat loop
    for (uint32_t i(start_index); i < end_index; ++i)
    {
        vec3[i] = vec1[i] + vec2[i]
    }
}

3. Create swarm and execute job

// Create a Swarm object with 16 thread
const uint32_t thread_count(16);
swrm::Swarm swarm(thread_count);
swrm::WorkGroup group = swarm.execute(job, thread_count);

Putting all together

#include "swarm.hpp"
#include <vector>

const size_t vec_size(100000000);

std::vector<float> v1(vec_size);
std::vector<float> v2(vec_size);
std::vector<float> v3(vec_size);

void job(uint32_t worker_id, uint32_t worker_count)
{
    // Number of values for which the thread is responsible
    const uint32_t step = vec_size / worker_count;
    // First value for the thread
    const uint32_t start_index = worker_id * step;
    // Last value
    const uint32_t end_index = (worker_id == worker_count - 1) ? start_index + step : vec_size - 1;
  
    // The actuat loop
    for (uint32_t i(start_index); i < end_index; ++i)
    {
        v3[i] = v1[i] + v2[i]
    }
}


int main()
{
    // Create a Swarm object with 16 thread
    const uint32_t thread_count(16);
    swrm::Swarm swarm(thread_count);
    // Start parallel job
    swrm::WorkGroup group = swarm.execute(job);
    // Wait for the job to terminate
    group.waitExecutionDone();
  
    return 0;
}

It is also possible to use a lambda instead of a pointer to a function

swrm::WorkGroup group = swarm.execute([&](uint32_t thread_id, uint32_t group_size){
    /* 
        The code here
    */
});

Performance

1D naive mean filter on 100.000.000 values with a kernel width of 64 on an i9-9900K

Thread count	Execution time	Speed Up
1	3.1 seconds	x1
16	0.28 seconds	x11

johnBuffer/Swarm