Hello,
I had an issue that has been mentioned before at:
http://archive.ambermd.org/201307/0514.html
This occurs when running a minimisation for an 80,000 atom system with cut
offs over a certain size (10.0 and 11.0 work fine, but 12.0 and larger do
not). The error is reported as GpuBuffer::Download failed unspecified
launch error (i.e. probably a CUDA segfault).
I ran pmemd.cuda through cuda-memcheck and it highlighted calls to
kNLBuildNeighborListOrthogonal16_kernel. All of the errors were of the
following form:
========= Invalid __global__ write of size 4
========= at 0x00000e38 in kNLBuildNeighborListOrthogonal16_kernel(void)
========= by thread (527,0,0) in block (0,0,0)
========= Address 0x70eaede3c is out of bounds
The threads in question were threads (512-527,0,0) in block (0,0,0) and
(192-207,0,0) in block (7,0,0). I tried looking at kBNL.h, but I don't
really know CUDA well enough to work out where the indexing might have gone
awry.
If it helps, I can upload input files as well.
The code was built yesterday with all the available bugfixes using CUDA
5.5, ifort and the Intel MKL.
Thanks,
Kyle
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Aug 07 2013 - 10:30:03 PDT
