kevmtan / electrocuda Goto Github PK

The CUDA implementation of filtiflt is slower than the Matlab/CPU implementation (for zero-phase high-pass filtering)

Hi Kevin,
I went through the Wiki of Single subject pipeline for intracranial EEG.
Based on your later reference to the steps in Source separation in the Optional steps for gamma or HFB bands, I guess those steps in the Source separation should be named '7-12' instead of the current '6-11'.
It seems I could not submit a pull request in the Wiki, thus I'm opening an issue to remind you of that. : )

Best,
Xiaoyu

A call to the function will only utilize ~7 threads

NITRC references all the neuroscience tools used in Research. NITRC is a non-profit NIH-sponsored website. This will help increase your tool (and associated paper) visibility.

https://www.nitrc.org/

AMICA is better than the Infomax algorithm used in CUDAICA; the former is noise-resistant & quasi-nonstationary (can use multiple mixture models with adaptive likelihood across timeframes)... AMICA is currently Intel CPU-only & much slower than CUDAICA

Compile AMICA's fortran source with CUDA's Fortran compiler... sounds simple but it's not

Required Fortran edits:

• Switch all calls to Intel MKL libraries/functions to their CUDA-optimized alternatives (function names are different)
• Add logic for ascertaining available RAM & VRAM, edit existing data chunking logic to account for both (e.g. differentiate GPU-constrained machines)
• Convert all data to single-precision (Jason: 'ok for theaters multiplication') except during precision-critical operations (follow CUDAICA as guide for implementing this)

Alternate Matlab-CUDA implementation:

• See Jason's e-mail with matlab implementation of AMICA
• Add logic for data chunking using gpuArray calls
• Add logic for normalization & integration of weights across data chunks
• Keep as Matlab code or use Matlab CUDA compiler?

Lots of work but keep in back burner

Currently you have to save data to disk before ThunderSVM can read it into GPU. Horrible performance during heavy I/O tasks like hyperparameter optimization