ABINIT Discussion Forums

Hi, everyone.
I am a new user of abinit, and I have some questions on the parameters in the input files.
I guess that the maxProcessors=npband*npfft*npkpt*npspinor. But I don't know why the GPU version work only when the npfft equal to 1.

--- !ERROR
message: |
When use_gpu_cuda is on, the number of FFT processors, npfft, must be 1
However, npfft=3
src_file: mpi_setup.F90
src_line: 640

And I want to ask you whether there are effects on the output result when I change the npfft to 1 in order to make the GPU version work.

Hi

The GPU are very good for FFT calculations therefore we impose for the GPU version to put all the FFT (for eache band) to be on one GPU which means npfft = 1.This is to increase performance on GPU.

Except for numerical noise, the result with npfft =1 et npfft != 1 and with or without GPU should be the same.

Cheers

But I don't know why the GPU version work only when the npfft equal to 1.

Because FFTs in parallel with MPI and multiple GPUs are not supported.
For the time being, one has to use npfft = 1 in the input file.

And I want to ask you whether there are effects on the output result when I change the npfft to 1 in order to make the GPU version work.

No, npfft specifies the number of processors used to perform the FFT.
Except for minor differences due to roundoff errors and MPI operations, there's should be no noticeable
difference in the *final* results if you change npfft.
The number of iterations and the intermediate results, on the other hand, could change if you play with npband, npfft ...