Minimum Compiler Version Update #5285
Description
After the 3.7 release, Kokkos is requiring C++17, which necessitates an upgrade in minimum compiler versions.
Furthermore, we are planning to move HIP out of Experimental in the release after 3.7 and start more stable support for ROCM versions. The question arises what that means for minimum versions Kokkos should support.
Here is the proposal we are working with right now, however we are considering lifting the version a bit higher than what we could get away with. The main reason to do so is avoiding known bugs, as well as making our burden for testing/support a bit easier.
| Compiler Line | Current Minimum | Initial Proposal | Considering | Decision |
|---|---|---|---|---|
| GCC | 5.3 | 7.3 | 8.2 | 8.2 |
| Clang | 4.0 | 8.0 | 11.1 | 8.0 |
| Clang as CUDA compiler | 10.0 | 10.0 | 11.1 | 10.0 |
Intel icpc |
17.0 | 19.1 | 2019.5 (aka icpc --version 19.0.5) |
2019.5 |
CUDA nvcc |
9.2.88 | 11.0 | 11.x | 11.0 |
| Minimum CUDA for use with Clang as CUDA compiler | 10.0.1 | 10.0.1 | -- | 10.0.1 |
| ROCM | 4.5.0 | 5.2.0 | -- | 5.2.0 |
IntelLLVM (CPU) icpx |
2021.1.1 | 2021.1.1 | -- | 2021.1.1 |
| IntelLLVM (SYCL) | 2022.0.0 | 2022.2.0 | -- | 2022.2.0 |
NVHPC nvc++ |
22.3 | 22.3 | -- | 22.3 |
| MSVC | 19.29 | |||
| IBM XL | Not supported | |||
| PGI | now via NVC++ |
Couple notes:
- Clang as CUDA compiler requires at least 10.x possibly 11.x we test 11.0 currently.
- NVC++ as CUDA compiler is not officially supported yet, we expect official support later this year
- The OpenMPTarget backend has separate requirements for newer compilers
- The upcoming OpenACC backend may have separate requirements for newer compilers