Enabled again neigh no with tpa > 1 for 3-body gpu styles for backward compatibility, could be slower than neigh no tpa 1 in many cases

2017-07-22 22:57:37 -05:00
parent 3d1d0c58c7
commit a71f5a0c20
6 changed files with 349 additions and 217 deletions
--- a/lib/gpu/lal_base_three.cpp
+++ b/lib/gpu/lal_base_three.cpp
@ -73,7 +73,6 @@ int BaseThreeT::init_three(const int nlocal, const int nall,
  if (_threads_per_atom>1 && gpu_nbor==0) { // neigh no and tpa > 1
    nbor->packing(true);
    _nbor_data=&(nbor->dev_packed);
-    _threads_per_atom = 1; // enforce tpa = 1 for now
  } else  // neigh yes or tpa == 1
    _nbor_data=&(nbor->dev_nbor);
  if (_threads_per_atom*_threads_per_atom>device->warp_size())