CUDA routines not tested

[fgerick]

Github actions don't have GPU access, need e.g. BuildKite.

[AshtonSBradley]

Thought I would mention this error trying to run the readme example

julia> cfg_gpu = SHTnsCfg(64; shtype=SHTns.QuickInit(;gpu=true))
  cuda GPU #0 "NVIDIA TITAN V" found (warp size = 32, compute capabilities = 7.0, CU=80, max_threads=163840, 11.8 GB, L2 cache=4.5MB).
!!! Use theta-contiguous FFT on GPU !!!
=> Using VkFFT v1.2.33
vkfft init FAILED with error code 4001
CUDA ERROR sht_gpu.cu:247 : an illegal memory access was encountered!

ERROR cuModuleLoadDataEx failed with error 700
SHTnsCfg{Real, SHTns.Orthonormal, SHTns.QuickInit}
lmax: 64, mmax: 64, mres: 1, nlat: 68, nphi: 140

on this

┌ Info: System information:
│ CUDA runtime 12.9, artifact installation
│ CUDA driver 12.3
│ NVIDIA driver 545.23.8
│ 
│ CUDA libraries: 
│ - CUBLAS: 12.9.0
│ - CURAND: 10.3.10
│ - CUFFT: 11.4.0
│ - CUSOLVER: 11.7.4
│ - CUSPARSE: 12.5.9
│ - CUPTI: 2025.2.0 (API 27.0.0)
│ - NVML: 12.0.0+545.23.8
│ 
│ Julia packages: 
│ - CUDA: 5.8.2
│ - CUDA_Driver_jll: 0.13.0+0
│ - CUDA_Runtime_jll: 0.17.0+0
│ 
│ Toolchain:
│ - Julia: 1.11.5
│ - LLVM: 16.0.6
│ 
│ 1 device:
└   0: NVIDIA TITAN V (sm_70, 5.255 GiB / 12.000 GiB available)

[fgerick]

Hmm not sure what is happening here, it works fine on the V100 and A100 GPUs that I can access. @nschaeff perhaps you understand what happens in the c-library/VKFFT? Perhaps what I have done in the SHTns_jll build script #remove mtune and gencode flags, replace by nvcc -arch=all (good?) does not produce valid binaries for all NVIDIA GPUs?

One thing you could try @AshtonSBradley would be to compile SHTns on your computer yourself using the right target for the GPU and system CUDA, and check if that solves the issue. You can link SHTns.jl to the custom build library see first page of the documentation.

Depending on the outcome we could adjust the SHTns_jll build script.