Closed fwinter closed 12 years ago
It's not a memory leak. Rerunning the blas tuning at the same local volume should fix this. See readme about this.
The need to retune for each volume is a real pain, but one that is about to disappear (by next week). See issue #11.
On Jan 31, 2012, at 10:05, Frank Winterreply@reply.github.com wrote:
During QUDA clover BiCGStab inversion the following happens:
BiCGstab: 250 iterations, r2 = 9.738363e-08 BiCGstab: 251 iterations, r2 = 1.270892e-07 QUDA error: (CUDA) too many resources requested for launch (node 1, blas_quda.cu:946 in copyCuda)
Setup as in issue #45. In short: Quda:master,QMP,cuda 4.0 Chroma:master QMP:OpenMPI 4 C2070 sharing 1 PCIe
true false BICGSTAB true DOUBLE HALF RECONS_12 RECONS_12 Any ideas ? --- Reply to this email directly or view it on GitHub: https://github.com/lattice/quda/issues/47
Given your silence on this, I'm closing this issue, as I think it's a non-issue.
Sorry, for the delay. "make tune" indeed solved the problem. Thanks!! Btw from the README I get the impression that tuning is something optional. However, good to know that it might cause problems not to run it.
During QUDA clover BiCGStab inversion the following happens:
BiCGstab: 250 iterations, r2 = 9.738363e-08 BiCGstab: 251 iterations, r2 = 1.270892e-07 QUDA error: (CUDA) too many resources requested for launch (node 1, blas_quda.cu:946 in copyCuda)
Setup as in issue #45. In short: Quda:master,QMP,cuda 4.0 Chroma:master QMP:OpenMPI 4 C2070 sharing 1 PCIe
AntiPeriodicT true
AsymmetricLinop false SolverType BICGSTAB Verbose true
CudaPrecision DOUBLE CudaSloppyPrecision HALF CudaReconstruct RECONS_12 CudaSloppyReconstruct RECONS_12
Any ideas ?