accxx

A C++14 library for simplifying the use of accelerators (CUDA and OpenCL).

Motivation

This library is a collection of tools for integrating CUDA and OpenCL code as seemlessly as possible with modern C++ codebases.

Features

CUDA

• RAII device switch (accxx/cuda/cuda_device_guard.hpp)
• Checked CUDA calls (accxx/cuda/call.hpp) with associated error handlers (accxx/cuda/error_handler.hpp)
• Support for transforming CUDA error codes to std::error_code (accxx/cuda/system_error.hpp)

OpenCL

• Error code to string conversion (accxx/cl/cl_error_code.hpp)
• Checked OpenCL calls (accxx/cl/call.hpp) with associated error handlers (accxx/cl/error_handler.hpp)
• Support for transforming OpenCL error codes to std::error_code (accxx/cl/system_error.hpp)

Build

accxx has been tested with CMake 3.17, but theoretically 3.8 and higher is supported.

If you need to compile with CUDA, then you will need to have CUDA installed.

If you need to compile with OpenCL, then you will need OpenCL headers and an OpenCL library.

You can build with

$ git clone https://github.com/ipapadop/accxx
$ export ACCXX_SRC=`pwd`/accxx
$ mkdir accxx-build; cd accxx-build
$ cmake $ACCXX_SRC
$ make -j

If CUDA is not found, you can point it to the right CUDA installation with

$ cmake $ACCXX_SRC -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc

If your compiler is too new, the CUDA compiler, nvcc may not be able to compile CUDA programs. You can instruct nvcc to use your compiler of choice with (e.g., GCC 8 on Ubuntu)

$ cmake $ACCXX_SRC -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/g++-8

Install

Installation can be accomplished with

$ make documentation
$ make install

License

Copyright (c) 2019-2020 Yiannis Papadopoulos

Distributed under the terms of the MIT License.

(See accompanying file LICENSE or copy at http://opensource.org/licenses/MIT)

Related Work

• Your own error condition