One more piece of the puzzle on the recent cuda_DPDP: I recompiled with
Scott's latest (15 feb) updates to kNTPKernels.h (using gcc 4.4.0).
Now I get a different error:
[case.nikkagpu]~/amber11/benchmarks/jac% ./bench.jac.cuda &
[1] 29813
[case.nikkagpu]~/amber11/benchmarks/jac% cudaMemcpyToSymbol: SetSim copy to
cSim failed all CUDA-capable devices are busy or unavailable
This happens if I set the GPU_ID to -1. Somehow, it can't choose which of 8
C2050's to run on. (This has not happened to me before).
If I explicitly choose a GPU_ID, SPDP works, DPDP fails:
[case.nikkagpu]~/amber11/test/cuda/jac% Run.jac 0 SPDP netcdf.mod
diffing jac.out.GPU_SPDP with jac.out
PASSED
==============================================================
[case.nikkagpu]~/amber11/test/cuda/jac% Run.jac 0 DPDP netcdf.mod
Error: invalid argument launching kernel kNLBuildNeighborList
Run.jac: Program error
(I'll send the files for a test case that seems to run slower with the new
code than the old in a separate email.)
...thx...dac
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Tue Feb 15 2011 - 17:30:02 PST
