PowerGridGnufft returns NaN with default settings on the 192x192x1x32 test dataset.

[acerjanic]

@siatjiasen reported that PowerGridGnufft did not return a correct image when run with the 192x192x1x32 test dataset. This sounds like a the data scaling issue that has popped up with the FFT step in the NUFFT operation.

I'll try to replicate this on CPU and GPU with the version of the code in the public repository.

[siatjiasen]

Hi Alex,

Really appreciate your kind help.

According to my test, the error "PowerGridGnufft did not return a correct image with the test data" might be related to the Ntimeseg parameter in config.xml. If the other parameters were kept as original, and set Ntimeseg = 1,2,3,4 gave a correct image result, while Ntimeseg >= 5 failed.

I got an academic version of PGI compiler and OpenACC, and were working on compiling GPU version of PowerGrid.

Thank you again.
Best wishes,
Jia Sen

[aihsani]

Has this issue been resolved? It still happens on the GPU about half of the time

[acerjanic]

This bug is still in the public version of the code. Internally, we have been working on a major refactor of the code and have fixed this bug. Unfortunately, the documentation for the internal version isn't ready yet, so that has been preventing us from releasing it.

We have moved to a containerized runtime (via NVIDIA-Docker) and to the HDF5 based ISMRMRD format for the data input. When we release, it will be a big change, so we want to make sure it's reasonably documented. Right now, I'm focusing on finishing up a manuscript for the project.

How soon do you need a fix?

[aihsani]

The sooner the better :) Is it possible to share privately? Or if you can share the fix for Nan would suffice too.