Culib - library to work with CUDA, using STL-like abstract types and algorithms.

Library provides:

• device_pointer class template abstraction, represents device memory address. Such a pointer can be dereferenced on host side - proxy object is returned.

• generic copy algorithm to copy data in any direction, supports copy in multiple threads.

• multithreading subsystem, including pool of reusable page-locked memory blocks and thread-pool to be used by copy.

• custom memcpy implementation that utilizes AVX can be used if standart implementation is too slow.

• cuda_storage class template abstraction represents device memory block and provides STL-like interface.

• device and page-locked memory allocators.

• CUDA helpers like: cuda_error_check, cuda_set_device, cuda_timer, ...

Including into project

Cmake add_subdirectory(...) command can be used to bring culib into your project:

cmake_minimum_required(VERSION 3.5)
project(my_project)
add_subdirectory(path_to_culib_dir culib)
add_executable(my_target)
target_link_libraries(my_target PRIVATE culib)
...

Usage

#include <iostream>
#include "cuda_storage.hpp"

int main(int argc, const char* argv[]){

    const auto n = 1<<20;
    //select device to allocate memory block
    culib::cuda_set_device(0);

    //initialization
    //from init list
    culib::cuda_storage<double> stor0{1,2,3,4,5,6,7,8,9,10};
    std::cout<<std::endl<<*stor0.begin();   //1
    std::cout<<std::endl<<*(stor0.begin()+3);   //4
    *stor0.begin() = 1.1;
    std::cout<<std::endl<<*stor0.begin();   //1.1

    //from iterators range
    std::vector<double> vec(n,0);
    std::iota(vec.begin(),vec.end(),0.5);
    culib::cuda_storage<double> stor1(vec.begin(),vec.end());
    std::cout<<std::endl<<*(stor1.begin()+1234);    //1234.5

    //from size and value
    culib::cuda_storage<double> stor2(n);
    std::cout<<std::endl<<*(stor2.begin()+1234);    //any
    culib::cuda_storage<double> stor3(n,2.2);
    std::cout<<std::endl<<*(stor3.begin()+1234);    //2.2

    //copy
    //device to host
    std::vector<double> vec1(stor1.size(),0.0);
    copy(stor1.begin(),stor1.end(),vec1.begin());
    std::cout<<std::endl<<*(vec1.begin()+1234); //1234.5
    //host to device
    copy(vec1.begin(),vec1.end(),stor2.begin());
    std::cout<<std::endl<<*(stor2.begin()+1234);    //1234.5
    //device to device
    copy(stor2.begin(),stor2.end(),stor3.begin());
    std::cout<<std::endl<<*(stor3.begin()+1234);    //1234.5

    //fill
    fill(stor3.begin(),stor3.begin()+100,0.0);
    std::cout<<std::endl<<*stor3.begin();   //0.0

    //clone
    auto stor2_copy = stor2.clone();
    //clone to peer
    culib::cuda_set_device(1);
    auto stor2_copy_peer = stor2.clone();

    //clear - deallocates device memory block
    stor2_copy.clear();

    //print
    std::cout<<std::endl<<stor0;    //[10 {1.1 2 3 4 5 6 7 8 9 10}]

    //call CUDA API
    using culib::cuda_assert;
    cuda_error_check(cudaMemset(stor2.begin(),0,stor2.size()*sizeof(double)));
    std::cout<<std::endl<<*(stor2.begin()+1234);    //0

    return 0;
}

Build tests and benchmarks

Catch framework is used for testing.

To build and run tests:

cmake -B build_dir -DBUILD_TEST=ON
cmake --build build_dir
build_dir/test/Test

To build and run benchmarks:

cmake -B build_dir -DBUILD_BENCHMARK=ON
cmake --build build_dir
build_dir/benchmark/Benchmark

License

This software is licensed under the BSL 1.0.