CUDA Development Guide

Generally, since CUDA can be C or C++ flavored, follow either C Development Guide or dev/cxx-guide, as appropriate, with the additions below. If something below contradicts the C/C++ style guides, go with the below for a CUDA project.

Coding Style

Files

• All source files should have the exact license text (e.g., GNU GPLv3), or an abbreviated version and a pointer to full license text (e.g., Copyright Foo Corp blah blah blah. See LICENSE.md for details).

• For C/C++ headers, use a .h/.hpp extension, respectively.

• For C/C++ sources, use a .c/.cpp extension, respectively for improved compilation speed.

• Use CUDA source files (.cu) only for code blocks containing device code (i.e., at least 1 __device__ or __global__ definition)

• Use empty __host__ and __device__ definition guards in function headers to make them portable for builds without CUDA support.

• Do not declare __global__ functions in C/C++ header or source files.

• Declare kernel calling functions in C/C++ headers, and encapsulate pointers to device and host memory locations.

Naming

• All global variables should be prefixed with g_.

• All device-local variables should be prefixed with d_.