search

nodrogluap / OpenDBA

GPU-accelerated Dynamic Time Warp (DTW) Barycenter Averaging
Other
64 stars 13 forks source link

Slow segmentation when thousand of input files. #6

Closed nodrogluap closed 4 years ago

nodrogluap commented 4 years ago

CUDA managed memory is being assigned for all input sequences at once, which leads to memory/swap trashing and major slowdown. Code change needed to chunk up the input sequences so all data in a chunk being worked on actively is likely fitting in the GPU RAM.

nodrogluap commented 4 years ago

Removed managed memory for segmentation buffer, which seems to help a lot.