ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

GPU Package: Switching back to timer disabling with multiple MPI tasks per GPU. Logic added to prevent mem leak.

Browse Source
This commit is contained in:
W. Michael Brown
2022-09-28 21:02:16 -07:00
parent be98b5a168
commit 6e34d21b24
6 changed files with 55 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
lib/gpu/lal_device.cpp
View File

@ -265,15 +265,13 @@ int DeviceT::init_device(MPI_Comm world, MPI_Comm replica, const int ngpu,
// Time on the device only if 1 proc per gpu
_time_device=true;
#if 0
// XXX: the following setting triggers a memory leak with OpenCL and MPI
// setting _time_device=true for all processes doesn't seem to be a
// problem with either (no segfault, no (large) memory leak.
// thus keeping this disabled for now. may need to review later.
// 2018-07-23 <akohlmey@gmail.com>
// Previous source of OCL memory leak when time_device=false
// - Logic added to release OCL events when timers are not invoked
if (_procs_per_gpu>1)
_time_device=false;
#endif
if (!_time_device && _particle_split > 0)
gpu->configure_profiling(false);
// Set up a per device communicator
MPI_Comm_split(node_comm,my_gpu,0,&_comm_gpu);
@ -715,7 +713,9 @@ void DeviceT::estimate_gpu_overhead(const int kernel_calls,
dev_data_out[0].flush();
#endif
driver_time=MPI_Wtime()-driver_time;
double time=over_timer.seconds();
double time=0.0;
if (_time_device)
time=over_timer.seconds();
if (time_device()) {
for (int i=0; i<_data_in_estimate; i++)