Run coredumped

[AlvinChen13]

My system is installed 4 P100 GPUs and CUDA 8.0. For NCCL, it runs well. And I compile the benchmark by 'make CUDA_PATH=/usr/local/cuda CUDNN_PATH=/usr/local/cuda MPI_PATH=/home/userid/ompi NCCL_PATH=/home/userid/weike/nccl/ ARCH=sm_61'

Anyone can help me the coredump?

userid@ubuntu-WK-4xP100:~/weike/DeepBench/code$ source /.bashrc
userid@ubuntu-WK-4xP100:/weike/DeepBench/code$ bin/gemm_bench
Times

m       n      k      a_t     b_t      time (usec) 

main: #1.
main: #2.
main: #3.
terminate called after throwing an instance of 'thrust::system::system_error'
what(): function_attributes(): after cudaFuncGetAttributes: invalid device function
Aborted (core dumped)
userid@ubuntu-WK-4xP100:~/weike/DeepBench/code$

[AlvinChen13]

BTW, cudnn is v5.1

[sharannarang]

@AlvinChen13 , The recommend architecture for P100 is sm_60. Some of the other Pascal GPUs use sm_61. This page provides more details:

https://developer.nvidia.com/cuda-gpus

The GEMM benchmark works on a P100 with sm_60. Can you try running it again with the correct architecture?

[AlvinChen13]

Thanks, fixed the issue.