Skip to content

vtkm does not compile for CUDA architectures < 60 (Solution included)

Problem

I found a compile error in:

  • vtkm 2.0.0
  • on Windows MSVC,
  • Using CUDA 11.8, and
  • CMAKE_CUDA_ARCHITECTURES "all-major"

Looking at the source code, the problem is also in master, but I haven't tested it on master, because compilation it takes forever. The problem should be reproducible when compiling with CUDA architecture 'sm_35' though.

The compile fails in Atomic.h, because atomicAdd() is not natively supported for double in cuda architectures < 60.

Solution

The fallback implementation is actually in the source file Atomic.h, but it is named wrongly.

Since there are no usages of vtkmAtomicAdd() in the whole code base, I assume it should be named AtomicAddImpl() and was just a mistake during refactoring. The same applies to the implementation for vtkm::Float32 directly above, which should be renamed from vtkmAtomicAddImpl() to AtomicAddImpl().

Renaming both of the implementations to AtomicAddImpl() lets me compile on above config using "all-major", which for me goes down to 'sm_35'.