vexx_gamma_gpu band group
In vexx_gamma and vexx_gamma_gpu, the input argument psi has the dimension of (lda*npol,m). I believe the 2nd dimension should have been max_ibands for the band group parallelization to work. In the CPU case this is probably harmless. In the GPU case, however, the code can segfault at ALLOCATE(psi_d, source=psi).
To reproduce, run the following calculation using 6 GPUs and 6 band groups.
There is no problem in vexx_k and vexx_k_gpu. psi is declared as (lda*npol,max_ibands) in the former and (:,:) in the latter.
A few GPU arrays in vexx_gamma_gpu are allocated but never used, notably hpsi_d. They have been cleaned up to save GPU memory.