Skip to content

HIP-gpu-streamed FFTs & p[c/r]egterg(w scalapack) & sum_band_gamma

Fabrizio Ferrari requested to merge fabrizio22/q-e:fftstreams_aligned2 into develop_omp5

Authors: @bellenlau and @fabrizio22
Main thing - streamed FFTs: the streamed FFTs algorithm has been ported on AMD GPUs with HIP by following the scheme previously adopted for Nvidia GPUs with CUDA.
NOTE: to compile PWscf with the HIP-streamed FFTs enabled it is necessary to add the -D__OMP_MANY_FFT flag (DFLAGS) in the make.inc and to separately compile the HIP-kernels file first. In the q-e directory:
q-e$ ./build_hipker.sh
q-e$ make pw
Parallel diag: while waiting for the final version of the gpu API solvers we have ported the Davidson driver for CPU-distributed diagonalization (with CPU backend scalapack).
sum_band_gamma: the main parts of sum_band_gamma have been ported, as it was previously done with the k-point version.

Merge request reports