In the most common use case, CUDA kernels are called using a triple chevron syntax <<<>>> that specifies the kernel launch configuration (i.e. number of threads, thread blocks, shared memory use and stream identifier), in addition to the arguments to the kernel.
Because the triple chevron syntax is CUDA-specific, and therefore can only be handled by the NVIDIA compiler driver (nvcc), any translation unit that calls a CUDA kernel using this syntax must be saved as a CUDA file (*.cu). Thus, calling CUDA kernels 'pollutes' C/C++ translation units.
This is at best a minor annoyance. A more serious problem is that, on some platforms, dependent on the version of the CUDA runtime supported or installed, the nvcc compiler driver does not invoke the latest C/C++ compiler in order to 'handle' non-CUDA code in a CUDA file. For example, CUDA 8 on the Linux platform nvcc calls/requires the now-ancient gcc version 5. This means that the use of C++14 and C++17 is impossible in CUDA files, with CUDA 8. (CUDA 9.2 however supports gcc variants up through 7.3.1. The CUDA 9.2 programming guide and nvcc manual also specifically identify support for C++14. Thanks to Robert Crovella for this clarification.)
InvokeKernel is a simple C++ template for invoking CUDA kernels in a way that solves the 'pollution' problem; so that C++ translation units that call CUDA kernels do not need to be saved as CUDA files, and any limitations in C++ support (by the version of gcc required by the version of CUDA installed) is localized to the kernels.
In contrast to the version of the talk presented at Meeting C++ London (see slides), the latest version of this code template introduces a pair of macros KERNEL_REFERENCES and DEFINE_KERNEL (see Kernels.h) uses xmacros so that users do not have to modify the files InvokeKernel.h and InvokeKernel.cu at multiple points every time a new kernel is introduced (see shotgun surgery).
- Build library libCudaKernels by running make command
- Build main.cpp and link with the library e.g.
g++ main.cpp -L./ -lCudaKernels -L/usr/local/cuda-8.0/lib64 -lcuda -lcudart -o prog
MIT License