Skip to content
fomics edited this page Jul 13, 2015 · 3 revisions

PI: Matthias Liebendörfer

Institution: University of Basel

Application Area: Isotropic diffusion source approximation (IDSA) used to efficiently compute the very complicated transport of neutrinos through dense matter in a supernova explosion

Starting point: 90% of time spent in meanfreepath and uyupdate.

Experiences: Many failed attempts to inline internal routines. Routines needed to be manually integrated (introducing race conditions, later resolved). Managing host/device data transfers and split work into streams to hide memory updates. Struggling with remaining synchronisation even in the presence of the async() clause.

Comments: Would not have found solutions on our own; Hackathon with interdisciplinary aspect, enabled us to make quick progress

Future: port into the full code, overlap memory copies with user code to obtain larger speedup. Looking forward to making proposal at next allocation period.

Final Presentation