search

N-BodyShop / changa

UIUC/PPL version of ChaNGa
http://hpcc.astro.washington.edu/tools/changa.html
GNU General Public License v2.0
42 stars 28 forks source link

problem with "gpu-local-tree-walk" enabled #131

Open darrenreed opened 2 years ago

darrenreed commented 2 years ago

The particle rung distribution from a GPU build is very different with and without "gpu-local-tree-walk" enabled. With that option disabled, the rung distribution matches that of a cpu build, as expected. Example rung distributions of step 1 of a ~4 billion particle run:


 CPU build:
 Rung distribution: ( 194114649, 3220278598, 561485123, 3709086, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ,) 

 GPU build, gpu-local-tree-walk disabled:
 Rung distribution: ( 194162039, 3220225643, 561490751, 3709023, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,) 

 GPU  build, gpu-local-tree-walk enabled:
 Rung distribution: ( 12915310, 167840400, 34995287, 27629871, 3716484141, 19722447, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,0, 0, 0,)
trquinn commented 1 year ago

I tried to reproduce this with the testcosmo. No luck: it seems to work OK with any of these builds.