density_gpu c_cuda_gpu_recv_pi CUDA error SIGSEGV
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
BigDFT |
Invalid
|
Undecided
|
Unassigned |
Bug Description
This bug appears, in calculation with GPU accel and "free" boundary condition.
In the case of "periodic" boundary cond. , calculation works.
Try examples that come in distro a change boundary cond. in GPU-test.
When GPU disabled, both calculations PASS.
bigdft-1.7.6 on Fedora 21, AMD Athlon(tm) II X4, NVidia Tesla M2090, CUDA 7
enabled CUDA, MPI (OpenMPI 1.8.4)
Unix process (not MPI) 0 has GPU : 0
OK, all process has semaphores : 1081356
Check card on all nodes....
Logic CPU : 0 with real CPU 0
CUDA error: invalid argument, in file cudafct.cu:123
**** ERROR *** : c_cuda_gpu_recv_pi
Program received signal SIGSEGV: Segmentation fault - invalid memory reference.
Backtrace for this error:
#0 0x7F2D1E6CD517
#1 0x7F2D1E6CDB5E
#2 0x7F2D1D9B295F
#3 0x7F2D1DA19CEE
#4 0x9F8CA9 in local_network:
#5 0x9F8D95 in local_network:
#6 0x9F5C76 in sg_exec_all_streams at s_gpu_C.cpp:216
#7 0x643797 in local_partial_
#8 0x5C31EB in sumrho_ at sumrho.f90:134
#9 0x5AE225 in input_wf_diag_ at init.f90:1574
#10 0x5B2EF1 in input_wf_ at init.f90:2089
#11 0x5A1380 in cluster_ at cluster.f90:510
#12 0x4399B7 in __bigdft_
#13 0x40E371 in bigdft at BigDFT.f90:54
Segmentation fault (core dumped)
Hello,
the GPU acceleration is available within OpenCL acceleration.
The CUDA version of the convolutions (which were only available for Periodic BC) is deprecated and should not be used anymore.
It will be removed in future releases.
So the suggestion is to use OpenCL acceleration as it is more optimized and available for all BC.
Do not hesitate
Luigi