Skip to content

CUDA Dialects#

The CUDA programming language is not formally specified. The "standard" is therefore approximately "whatever nvcc does". Although clang supports compiling CUDA, it supports a somewhat different dialect compared to nvcc.

You can read more about (some of) the specific differences in the LLVM manual page about it.

This leads to a problem: most CUDA code is written with nvcc in mind, but the only open-source compiler available with a CUDA frontend is clang. Many real-world CUDA programs cannot be successfully compiled with clang because they depend on nvcc's behaviour.

HIP experiences the same problem: the HIP compiler is based on LLVM, so HIP is closer to "LLVM-dialect CUDA" than it is to "nvcc-dialect CUDA". This causes some CUDA programs to fail in "interesting" ways when ported to HIP. It's not really the case that you can remap all the API calls to the HIP ones and expect it to work: nvcc-CUDA, and LLVM-CUDA/HIP have quite different C++ semantics.

SCALE resolves this issue by offering two compilers:

  • "nvcc": a clang frontend that replicates the behaviour of NVIDIA's nvcc, allowing existing CUDA programs to compile directly. This is similar to how LLVM achieves MSVC compatibility by providing clang-cl.
  • clang: providing clang's usual clang-dialect-CUDA support, with our opt-in language extensions.

Existing projects can be compiled without modification using the nvcc-equivalent compiler. Users of clang-dialect CUDA may use the provided clang compiler to compile for either platform.