Open main menu

CDOT Wiki β

Changes

Ghost Cells

274 bytes added, 02:14, 7 April 2019
Result
==== Result ====
 
'''POST Presentation Results''' Contrary to the presentation's conclusion, the ghost-cell method proved to be more effective with some changes to the logic than simpler global-memory-based counterpart. It does require some preparation in the host machine. The gain is small.
 
[[File:optimized.png|center|frame|GPU highlights. para-ghost-pre-co2, which implements Ghost Cell + Prefetch + Coaleased memory + logic change, is slightly faster than simpler Prefetch+Coaleased memory that uses Global Memory. Both methods are superior than calling the conditional-less kernel 1000 times over PCIe.]]
[[File:all.png|center|frame|Using UsinGPU highlights. Ghost Cell + Prefetch + Coaleased memory + logic change is slightly faster than simpler Prefetch+Coaleased memory that uses Global Memoryg GPU significantly improved Calculation Time over the CPU counterparts.]] """POST Presentation Results""" Contrary to the presentation's conclusion, the ghost-cell method proved to be more effective with some changes to the logic.
70
edits