Version of cuDNN darknet

[saromano]

Hi,
I have a problem with cuDNN versions …
I’m working on a new system just fleshed out, I just downloaded the JetPack 4.5.1 version and loaded it on the microsd,I reinstalled OpenCV with CUDA ON.
This is the results:

Summary

Now I’m trying to compile yolov4 darknet, but when I run the command

cmake ..

I get this information

Summary

Darknet_VERSION: 0.2.5.4
CMake Warning at CMakeLists.txt:99 (message):
  vcpkg not found, toolchain not defined, system not win32 so build might
  fail


-- The C compiler identification is GNU 7.5.0
-- The CXX compiler identification is GNU 7.5.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Setting build type to 'Release' as none was specified.
-- Looking for a CUDA compiler
-- Looking for a CUDA compiler - /usr/local/cuda-10.2/bin/nvcc
-- The CUDA compiler identification is NVIDIA 10.2.89
-- Detecting CUDA compiler ABI info
-- Detecting CUDA compiler ABI info - done
-- Check for working CUDA compiler: /usr/local/cuda-10.2/bin/nvcc - skipped
-- Detecting CUDA compile features
-- Detecting CUDA compile features - done
-- Selected CMAKE_CUDA_ARCHITECTURES: 30
-- Your setup does not support half precision (it requires CUDA_ARCHITECTURES >= 70)
-- Looking for pthread.h
-- Looking for pthread.h - found
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Failed
-- Looking for pthread_create in pthreads
-- Looking for pthread_create in pthreads - not found
-- Looking for pthread_create in pthread
-- Looking for pthread_create in pthread - found
-- Found Threads: TRUE  
-- Found CUDA: /usr/local/cuda-10.2 (found suitable exact version "10.2") 
-- Found OpenCV: /usr/local (found version "4.5.0") 
-- Found Stb: /home/nvidia/darknet/3rdparty/stb/include  
-- Found OpenMP_C: -fopenmp (found version "4.5") 
-- Found OpenMP_CXX: -fopenmp (found version "4.5") 
-- Found OpenMP: TRUE (found version "4.5")  
-- Found CUDNN: /usr/include (found version "?") 
-- CMAKE_CUDA_FLAGS:  --compiler-options " -Wall -Wno-unused-result -Wno-unknown-pragmas -Wfatal-errors -Wno-deprecated-declarations -Wno-write-strings -DGPU -DCUDNN -DOPENCV -fPIC -fopenmp -Ofast " 
-- Looking for sgemm_
-- Looking for sgemm_ - not found
-- Looking for sgemm_
-- Looking for sgemm_ - found
-- Found BLAS: /usr/lib/aarch64-linux-gnu/libopenblas.so  
-- ZED SDK enabled
-- Configuring done
-- Generating done
-- Build files have been written to: /home/nvidia/darknet/build_release

what I don’t understand are

– Found CUDNN: /usr/include (found version “?”)

and

– Selected CMAKE_CUDA_ARCHITECTURES: 30
– Your setup does not support half precision (it requires CUDA_ARCHITECTURES >= 70)

how do i tell them where to find cuDNN? i know it is in this path: /usr/include/aarch64-linux-gnu

In .bashrc i added just:
export PATH="${PATH}:/usr/local/cuda-10.2/bin"

Thanks,
Salvatore

[AastaLLL]

Hi,

You can setup cmake variable via -D.
For example:

$ cmake -DCMAKE_CUDA_ARCHITECTURES=72 .

Thanks.

[saromano]

THANK YOU,
how do i tell cmake to use cuDNN files in /usr/include/aarch64-linux-gnu path? it finds some files in the /usr/include path

[AastaLLL]

Hi,

There are soft links of cuDNN in the /usr/include.
So the patch is fine.

The reason that cmake cannot find the cuDNN version is the information changed from cudnn.h into cudnn_version.h.
Please check the following page for the WAR:

cudnn8 version check fails

Thanks.