Summary: | sci-libs/vtk-9.3.0[cuda] configuration fails: Error detecting architecture flags for CUDA. Please set VTKm_CUDA_Architecture manually. | ||
---|---|---|---|
Product: | Gentoo Linux | Reporter: | Miroslav Šulc <fordfrog> |
Component: | Current packages | Assignee: | Paul Zander <negril.nx+gentoo> |
Status: | CONFIRMED --- | ||
Severity: | normal | CC: | anton.gubarkov, negril.nx+gentoo, proxy-maint, sci |
Priority: | Normal | ||
Version: | unspecified | ||
Hardware: | All | ||
OS: | Linux | ||
Whiteboard: | |||
Package list: | Runtime testing required: | --- | |
Attachments: | build.ong |
Description
Miroslav Šulc
2024-04-14 19:56:07 UTC
VTK-9.3 switched to enable_language support for CUDA. That brought some changes. VTK_CUDA_ARCH is no longer used and would be VTKm_CUDA_Architecture now. As stated in the die message you posted. But with enable_language support you can now use CUDAARCHS to set the architecture. This will also set VTKm_CUDA_Architecture. See https://cmake.org/cmake/help/latest/envvar/CUDAARCHS.html and https://cmake.org/cmake/help/latest/prop_tgt/CUDA_ARCHITECTURES.html. In the ebuild is code that auto detects the present architectures and tries to set CUDAARCHS if not already set. The problem now is that with the changes from nvidia-drivers-550 and the whole trying to access /proc/self/task, that detection broke. See: > trying to determine host CUDAARCHS > __nvcc_device_query failed to call cudaLoader::cuInit(0) with error 0x130 (CUDA_ERROR_OPERATING_SYSTEM) I haven't gotten around fixing that yet because it should be done in the cuda.eclass together with a lot of other changes. The easiest fix for you now is setting CUDAARCHS="89". The good thing is, Ada cards are supported now. thanks! putting CUDAARCHS="89" to /etc/portage/make.conf makes the configuration phase pass. |