User Tools

Site Tools


cudnn

cudnn

NVidia cuDNN is installed with Cuda 11.5. Example:

pinnacle-l1:rfeynman:$ git clone https://github.com/Hardware-Alchemy/cuDNN-sample
pinnacle-l1:rfeynman:$ cd cuDNN-sample/cudnn_samples_v7/conv_sample
pinnacle-l1:rfeynman:$ srun --nodes=1 --ntasks-per-node=1 --cpus-per-task=32 \
  --partition gpu06 --qos gpu --time=6:00:00 --pty /bin/bash
c1715:rfeynman:/cuDNN-sample/cudnn_samples_v7/conv_sample$ module load gcc/9.3.1 cuda/11.5
c1715:rfeynman:/cuDNN-sample/cudnn_samples_v7/conv_sample$ GENCODE_FLAGS=\
"-gencode arch=compute_72,code=sm_72" make
Makefile:170: "print CUDA version  11050 "
Makefile:171: "print sms new  30 35 50 53 60 61 70 72 "
g++ -I/share/apps/cuda/cuda-11.5/include   -o fp16_emu.o -c fp16_emu.cpp
g++ -I/share/apps/cuda/cuda-11.5/include   -o conv_sample.o -c conv_sample.cpp
/share/apps/cuda/cuda-11.5/bin/nvcc -ccbin g++ -m64 -gencode arch=compute_70,code=sm_70 \
-o conv_sample fp16_dev.o fp16_emu.o conv_sample.o  -lcudart -lcublas -lcudnn -lstdc++ -lm
c1715:rfeynman:/cuDNN-sample/cudnn_samples_v7/conv_sample$ ./conv_sample
Testing single precision
Testing conv
^^^^ CUDA : elapsed = 0.484545 sec,  
Test PASSED
Testing half precision (math in single precision)
Testing conv
^^^^ CUDA : elapsed = 5.48363e-05 sec,  
Test PASSED
c1715:rfeynman:/cuDNN-sample/cudnn_samples_v7/conv_sample $
cudnn.txt · Last modified: 2022/02/08 18:40 by root