Project

General

Profile

Bug #3071

complex.nbnxn-ljpme-LB-geometric failing with OpenCL timing debug mode assertions

Added by Szilárd Páll 3 months ago.

Status:
Accepted
Priority:
Normal
Assignee:
Szilárd Páll
Category:
mdrun
Target version:
2020
Affected version - extra info:
Affected version:
git master
Difficulty:
uncategorized
Close

Description

Intermittent error likely caused by empty domains getting their timer region left open -- needs further investigation.

Triggered by pre-submit config:

gcc-9 openmp simd=avx2_256 gpuhw=amd opencl-1.2 clFFT-2.14 hwloc libhwloc-1.11.2 cmake-3.14.5 host=bs_gpu01,bs_gpu01 (Aug 30, 2019 4:37:37 PM)


20 steps,      0.0 ps.

-------------------------------------------------------
Program:     gmx mdrun, version 2020-dev-20190830-c5d6111
Source file: src/gromacs/gpu_utils/gpuregiontimer.h (line 94)
Function:    GpuRegionTimerWrapper<GpuRegionTimerImpl>::openTimingRegion(CommandStream) [with GpuRegionTimerImpl = GpuRegionTimerImpl; CommandStream = _cl_command_queue*]::<lambda()>
MPI rank:    1 (out of 2)

Assertion failed:
Condition: debugState_ == TimerState::Idle
GPU timer should be idle, but is stopped.

For more information and tips for troubleshooting, please check the GROMACS
website at http://www.gromacs.org/Documentation/Errors

Also available in: Atom PDF