Week 12b: Faster and Faster

This past week we were able to improve upon our GPU processing code. We found that when converting framing from the CPU to the GPU it was significantly slower. This was due to a for loop within the data framing looping serially rather than in parallel. Our solution was to instead reshape the array for conversion, which runs in parallel on the GPU. Additionally, we were able to successfully implement multi-queuing by stacking multiple 3D arrays along the third axis. This led to an increase in FFT and framing speeds.

Previously, the CPU based processing code spent a majority of the time with framing and FFT. Now processed in parallel via GPU, time is more spent converting arrays. However, even with these added conversion times, there is still a significant improvement in speed. In terms of GPU, we found a framing speed improvement of around 40x from 0.022s to 0.0005s. In terms of FFT, we found an improvement of around 96x from 0.048s to 0.000496s. Overall, we have a new framerate ceiling of approximately 100fps from 14fps.

In addition, we deemed VisPy to be a promising alternative to Mayavi but not functionally applicable for our project’s purposes. Mayavi, while difficult to install, is still well established and functional. Visualization was not the focus of software development, as it was not necessarily a limitation. The main focus was on the processing of waveform data, which we dedicated our efforts towards and gained great success. 

In the conclusion of this project for this academic year, we have resolved any computational issues on the software end. In terms of suggestions for the future of this project, multi-queuing could potentially be refined, conversion times could possibly be reduced, and data acquisition through the system’s hardware could be improved.