Cuda dynamic parallelism

[jg piccinali]

Daint

• doc

Setup

• module swap PrgEnv-cray PrgEnv-gnu
• module swap gcc gcc/4.8.2
• module load craype-accel-nvidia35

Compile

• scorep --mpp=mpi --cuda cc -g -c mpic.c -o GNU_mpic.o
• scorep --mpp=mpi --cuda nvcc -G -arch=sm_35 -rdc=true -g -c mpicu.cu -o GNU_mpicu.o
• scorep --mpp=mpi --cuda nvcc -G -arch=sm_35 -dlink -g GNU_mpicu.o -o dlink_GNU_mpicu.o -lcudadevrt
• scorep --mpp=mpi --cuda cc -g GNU_mpic.o GNU_mpicu.o dlink_GNU_mpicu.o -lcudart -o GNU.SANTIS -lcudadevrt

Run

• export SCOREP_ENABLE_PROFILING=false
• export SCOREP_ENABLE_TRACING=true
• export SCOREP_CUDA_ENABLE=yes
• export SCOREP_CUDA_BUFFER=500000000000
• aprun -n 1 -N 1 -d 1 -j 1 ./GNU.SANTIS+sc142 4096

[Score-P] src/adapters/cuda/scorep_cupti4_activity.c:212: 
Warning: [CUPTI Activity] Dropped 8135 records. 
Current buffer size: 500000000000 bytes
To avoid dropping of records increase the buffer size!
Proposed minimum SCOREP_CUDA_BUFFER=500000715792
[Score-P] Please report this to [e-mail]. Thank you.
[Score-P] Try also to preserve any generated core dumps.