Searched hist:"29 ec485eb72075150292bd9d6291eab6d473a1fc" (Results 1 – 1 of 1) sorted by relevance
| /libCEED/backends/cuda/ |
| H A D | ceed-cuda-compile.cpp | 29ec485eb72075150292bd9d6291eab6d473a1fc Wed May 22 21:00:00 UTC 2024 Jed Brown <jed@jedbrown.org> backends/cuda: NVRTC compile to CUBIN when supported (resolve #1587)
This allows using a newer CUDA runtime with an older driver, and seems
to have no downsides.
NVRTC can generate cubins directly starting with CUDA 11.1. [...]
NVRTC used to support only virtual architectures through the option
-arch, since it was only emitting PTX. It will now support actual
architectures as well to emit SASS. The interface is augmented to
retrieve either the PTX or cubin if an actual architecture is
specified.
https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#dynamic-code-generation
|