superpoca / caffe-cudnn8

Caffe with cuDNN8

The original version caffe can only work with cuDNN7.x, but since CUDA11.x, only cudnn8 is supported, so I made some changes in code to make caffe able to work with CUDA11.x + cuDNN8.x.

I have made code changes in fo cudnn8:

   cmake/Cuda.cmake
   src/caffe/layers/cudnn_ndconv_layer.cu
   src/caffe/layers/cudnn_conv_layer.cpp
   src/caffe/layers/cudnn_deconv_layer.cpp

This version code can directly work with cuDNN8.x, If you do want to use caffe with cudnn7.x(original caffe), you need to make these changes in the above files:

1) Open the file “cmake/Cuda.cmake”.  replace "cudnn_version.h" with "cudnn.h" by commenting/uncommenting the lines where they are.
2) In cudnn_ndconv_layer.cu, cudnn_conv_layer.cpp and cudnn_deconv_layer.cpp  change all "if CUDNN_VERSION_MIN(8, 0, 0)  // 0" to "#if 0 // CUDNN_VERSION_MIN(8, 0, 0)".

How to use

Step 1 Install dependencies

sudo apt-get install libprotobuf-dev libleveldb-dev  libblas-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler
sudo apt-get install --no-install-recommends libboost-all-dev
sudo apt-get install libopenblas-dev liblapack-dev libatlas-base-dev
sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev
sudo apt-get install git cmake build-essential
sudo pip install graphviz	

Step 2 download caffe-cudnn8 source code

git clone https://github.com/Jeremy-J-J/caffe-cudnn8.git

Step 3 Build settings

Although I provide the Makefile.config and Makefile that can directly work, but i think you also need to know how it works.
(1) Create a file named "Makefile.config"
	cp Makefile.config.example Makefile.config

(2) Set the CUDA ARCH
	According to the actual to set.
	CUDA_ARCH := -gencode arch=compute_75,code=sm_75
	
(3) Set use Opencv3
	OPENCV_VERSION := 3
	
(4) Set CUDA path
	CUDA_DIR := /usr/local/cuda
	
(5) Set Python Include
	PYTHON_INCLUDE := /root/anaconda3/include \
					  /root/anaconda3/include/python3.7m \
					  /root/anaconda3/lib/python3.7/site-packages/numpy/core/include

(6) Set Python Library
	PYTHON_LIB := /root/anaconda3/lib
	
(7) Set the other Include and Library
	INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial /home/0_env/opencv/include
	LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/x86_64-linux-gnu/hdf5/serial /home/0_env/opencv/lib

                    

	LIBRARIES += glog gflags protobuf leveldb snappy \
				 lmdb boost_system hdf5_hl hdf5 m\
				 opencv_core opencv_highgui opencv_imgproc opencv_imgcodecs

(8) Modify the Makefile
	# After LIBRARIES add opencv_imgcodecs
	LIBRARIES += opencv_core opencv_highgui opencv_imgproc opencv_imgcodecs

Step 4 Build

make all -j32
make test -j32
make runtest -j32
make pycaffe

Step 5 Configure environment

vim ~/.bashrc
export PYTHONPATH=$PYTHONPATH:/home/Jeremy-J-J/caffe/python

source ~/.bashrc