Building conv and rnn from Deepbench_nvidia complains about cudnnSetRNNDescriptor_v6

[GaoHengSiang]

I am using system cuda (12.0), but cudnn is installed in a mamba env, so I ran the script below

# using system cuda, but cuda is not in /usr/local/cuda on this machine
export CUDA_INSTALL_PATH=/usr/lib/nvidia-cuda-toolkit
export PATH=$CUDA_ISNTALL_PATH/bin:$PATH
mamba activate accelsim
export CUDNN_PATH=$CONDA_PREFIX   # cuDNN from conda/mamba env

then under accelsim-framework/

$ source ./gpu-app-collection/src/setup_environment

the resulting message is as follows

make: Entering directory '/home/gaohengsiang/accelsim/accel-sim-framework/gpu-app-collection/src'
export BINDIR=/home/gaohengsiang/accelsim/accel-sim-framework/gpu-app-collection/src/..//bin/12.0; export BINSUBDIR=release; export ROOTOBJDIR=obj_12.0; make  GENCODE_SM10= GENCODE_SM13= GENCODE_SM20= GENCODE_SM30= GENCODE_SM35= GENCODE_SM37= GENCODE_SM50= GENCODE_SM20= CUBLAS_LIB=cublas_static CUDNN_LIB=cudnn_static -C cuda/DeepBench/code/nvidia
make[1]: Entering directory '/home/gaohengsiang/accelsim/accel-sim-framework/gpu-app-collection/src/cuda/DeepBench/code/nvidia'
mkdir -p bin 
mkdir -p bin
mkdir -p bin
/usr/lib/nvidia-cuda-toolkit/bin/nvcc gemm_bench.cu -DUSE_TENSOR_CORES=0 -DPAD_KERNELS=1 -o bin/gemm_bench -I ../kernels/ -I /usr/lib/nvidia-cuda-toolkit/include -L /usr/lib/nvidia-cuda-toolkit/lib64 -lcublas -L /usr/lib/nvidia-cuda-toolkit/lib64 -lcurand  -gencode=arch=compute_70,code=compute_70 -gencode=arch=compute_75,code=compute_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -std=c++14 --cudart shared -std=c++14
/usr/lib/nvidia-cuda-toolkit/bin/nvcc rnn_bench.cu -DUSE_TENSOR_CORES=0 -o bin/rnn_bench -I ../kernels/ -I /usr/lib/nvidia-cuda-toolkit/include -I /home/gaohengsiang/.local/share/mamba/envs/accelsim/include/ -L /home/gaohengsiang/.local/share/mamba/envs/accelsim/lib64/ -L /usr/lib/nvidia-cuda-toolkit/lib64 -lcurand -lcudnn  -gencode=arch=compute_70,code=compute_70 -gencode=arch=compute_75,code=compute_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -std=c++14 --cudart shared -std=c++14
/usr/lib/nvidia-cuda-toolkit/bin/nvcc conv_bench.cu -DUSE_TENSOR_CORES=0 -DPAD_KERNELS=1 -o bin/conv_bench -I ../kernels/ -I /usr/lib/nvidia-cuda-toolkit/include -I /home/gaohengsiang/.local/share/mamba/envs/accelsim/include/ -L /home/gaohengsiang/.local/share/mamba/envs/accelsim/lib64/ -L /usr/lib/nvidia-cuda-toolkit/lib64 -lcurand -lcudnn  -gencode=arch=compute_70,code=compute_70 -gencode=arch=compute_75,code=compute_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -std=c++14 --cudart shared -std=c++14
cudnn_helper.h(359): error: identifier "cudnnSetRNNDescriptor_v6" is undefined

cudnn_helper.h(359): error: identifier "cudnnSetRNNDescriptor_v6" is undefined

1 error detected in the compilation of "rnn_bench.cu".
1 error detected in the compilation of "conv_bench.cu".
make[1]: *** [Makefile:43: rnn] Error 2
make[1]: *** Waiting for unfinished jobs....
make[1]: *** [Makefile:38: conv] Error 2
/usr/lib/nvidia-cuda-toolkit/bin/nvcc gemm_bench.cu -DUSE_TENSOR_CORES=1 -DPAD_KERNELS=1 -o bin/gemm_bench-tencore -I ../kernels/ -I /usr/lib/nvidia-cuda-toolkit/include -L /usr/lib/nvidia-cuda-toolkit/lib64 -lcublas -L /usr/lib/nvidia-cuda-toolkit/lib64 -lcurand  -gencode=arch=compute_70,code=compute_70 -gencode=arch=compute_75,code=compute_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -std=c++14 --cudart shared 
make[1]: Leaving directory '/home/gaohengsiang/accelsim/accel-sim-framework/gpu-app-collection/src/cuda/DeepBench/code/nvidia'
make: *** [Makefile:115: Deepbench_nvidia] Error 2
make: Leaving directory '/home/gaohengsiang/accelsim/accel-sim-framework/gpu-app-collection/src'

gemm built successfully (both tencore and normal)

[JRPan]

We are not actively running Deepbench anymore. It's too old :)

But seems running a older cuDNN could fix this opencv/opencv#24983