Merge pull request #14660 from YashasSamaga:dnn-cuda-build

add cuDNN dependency and setup build for cuda4dnn (#14660)

* update cmake for cuda4dnn

- Adds FindCUDNN
- Adds new options:
   * WITH_CUDA
   * OPENCV_DNN_CUDA
- Adds CUDA4DNN preprocessor symbol for the DNN module

* FIX: append EXCLUDE_CUDA instead of overwrite

* remove cuDNN dependency for user apps

* fix unused variable warning
This commit is contained in:
Yashas Samaga B L 2019-06-02 11:47:15 +00:00 committed by Alexander Alekhin
parent 3ccb1d67c2
commit ae279966c2
13 changed files with 228 additions and 11 deletions

View File

@ -249,6 +249,9 @@ OCV_OPTION(WITH_CUFFT "Include NVidia Cuda Fast Fourier Transform (FFT) library
OCV_OPTION(WITH_CUBLAS "Include NVidia Cuda Basic Linear Algebra Subprograms (BLAS) library support" WITH_CUDA
VISIBLE_IF WITH_CUDA
VERIFY HAVE_CUBLAS)
OCV_OPTION(WITH_CUDNN "Include NVIDIA CUDA Deep Neural Network (cuDNN) library support" WITH_CUDA
VISIBLE_IF WITH_CUDA
VERIFY HAVE_CUDNN)
OCV_OPTION(WITH_NVCUVID "Include NVidia Video Decoding library support" WITH_CUDA
VISIBLE_IF WITH_CUDA
VERIFY HAVE_NVCUVID)
@ -918,6 +921,9 @@ if(HAVE_CUDA)
if(HAVE_CUBLAS)
set(OPENCV_LINKER_LIBS ${OPENCV_LINKER_LIBS} ${CUDA_cublas_LIBRARY})
endif()
if(HAVE_CUDNN)
set(OPENCV_LINKER_LIBS ${OPENCV_LINKER_LIBS} ${CUDNN_LIBRARIES})
endif()
if(HAVE_CUFFT)
set(OPENCV_LINKER_LIBS ${OPENCV_LINKER_LIBS} ${CUDA_cufft_LIBRARY})
endif()
@ -1573,6 +1579,11 @@ if(WITH_CUDA OR HAVE_CUDA)
status(" NVIDIA GPU arch:" ${OPENCV_CUDA_ARCH_BIN})
status(" NVIDIA PTX archs:" ${OPENCV_CUDA_ARCH_PTX})
endif()
endif()
if(WITH_CUDNN OR HAVE_CUDNN)
status("")
status(" cuDNN:" HAVE_CUDNN THEN "YES (ver ${CUDNN_VERSION})" ELSE NO)
endif()
if(WITH_VULKAN OR HAVE_VULKAN)

105
cmake/FindCUDNN.cmake Normal file
View File

@ -0,0 +1,105 @@
# template taken from https://cmake.org/cmake/help/v3.14/manual/cmake-developer.7.html
# Distributed under the OSI-approved BSD 3-Clause License. See accompanying
# file Copyright.txt or https://cmake.org/licensing for details.
#[=======================================================================[.rst:
FindCUDNN
---------
Finds the cuDNN library.
Requires:
^^^^^^^^^
find_cuda_helper_libs from FindCUDA.cmake
i.e. CUDA module should be found using FindCUDA.cmake before attempting to find cuDNN
Result Variables
^^^^^^^^^^^^^^^^
This will define the following variables:
``CUDNN_FOUND``
``CUDNN_INCLUDE_DIRS`` location of cudnn.h
``CUDNN_LIBRARIES`` location of cudnn library
Cache Variables
^^^^^^^^^^^^^^^
The following cache variables will be set if cuDNN was found. They may also be set on failure.
``CUDNN_LIBRARY``
``CUDNN_INCLUDE_DIR``
``CUDNN_VERSION``
``CUDNN_VERSION_MAJOR`` INTERNAL
``CUDNN_VERSION_MINOR`` INTERNAL
``CUDNN_VERSION_PATCH`` INTERNAL
#]=======================================================================]
# find the library
if(CUDA_FOUND)
find_cuda_helper_libs(cudnn)
set(CUDNN_LIBRARY ${CUDA_cudnn_LIBRARY} CACHE FILEPATH "location of the cuDNN library")
unset(CUDA_cudnn_LIBRARY CACHE)
endif()
# find the include
if(CUDNN_LIBRARY)
find_path(CUDNN_INCLUDE_DIR
cudnn.h
PATHS ${CUDA_TOOLKIT_INCLUDE}
DOC "location of cudnn.h"
NO_DEFAULT_PATH
)
if(NOT CUDNN_INCLUDE_DIR)
find_path(CUDNN_INCLUDE_DIR
cudnn.h
DOC "location of cudnn.h"
)
endif()
endif()
# extract version from the include
if(CUDNN_INCLUDE_DIR)
file(READ "${CUDNN_INCLUDE_DIR}/cudnn.h" CUDNN_H_CONTENTS)
string(REGEX MATCH "define CUDNN_MAJOR ([0-9]+)" _ "${CUDNN_H_CONTENTS}")
set(CUDNN_MAJOR_VERSION ${CMAKE_MATCH_1} CACHE INTERNAL "")
string(REGEX MATCH "define CUDNN_MINOR ([0-9]+)" _ "${CUDNN_H_CONTENTS}")
set(CUDNN_MINOR_VERSION ${CMAKE_MATCH_1} CACHE INTERNAL "")
string(REGEX MATCH "define CUDNN_PATCHLEVEL ([0-9]+)" _ "${CUDNN_H_CONTENTS}")
set(CUDNN_PATCH_VERSION ${CMAKE_MATCH_1} CACHE INTERNAL "")
set(CUDNN_VERSION
"${CUDNN_MAJOR_VERSION}.${CUDNN_MINOR_VERSION}.${CUDNN_PATCH_VERSION}"
CACHE
STRING
"cuDNN version"
)
unset(CUDNN_H_CONTENTS)
endif()
include(FindPackageHandleStandardArgs)
find_package_handle_standard_args(CUDNN
FOUND_VAR CUDNN_FOUND
REQUIRED_VARS
CUDNN_LIBRARY
CUDNN_INCLUDE_DIR
VERSION_VAR CUDNN_VERSION
)
if(CUDNN_FOUND)
set(CUDNN_LIBRARIES ${CUDNN_LIBRARY})
set(CUDNN_INCLUDE_DIRS ${CUDNN_INCLUDE_DIR})
endif()
mark_as_advanced(
CUDNN_LIBRARY
CUDNN_INCLUDE_DIR
CUDNN_VERSION
)

View File

@ -37,6 +37,16 @@ if(CUDA_FOUND)
set(HAVE_CUBLAS 1)
endif()
if(WITH_CUDNN)
set(CMAKE_MODULE_PATH "${OpenCV_SOURCE_DIR}/cmake" ${CMAKE_MODULE_PATH})
find_host_package(CUDNN "${MIN_VER_CUDNN}")
list(REMOVE_AT CMAKE_MODULE_PATH 0)
if(CUDNN_FOUND)
set(HAVE_CUDNN 1)
endif()
endif()
if(WITH_NVCUVID)
find_cuda_helper_libs(nvcuvid)
if(WIN32)
@ -293,6 +303,13 @@ if(HAVE_CUDA)
endforeach()
endif()
if(HAVE_CUDNN)
foreach(p ${CUDNN_LIBRARIES})
get_filename_component(_tmp ${p} PATH)
list(APPEND CUDA_LIBS_PATH ${_tmp})
endforeach()
endif()
if(HAVE_CUFFT)
foreach(p ${CUDA_cufft_LIBRARY})
get_filename_component(_tmp ${p} PATH)
@ -311,7 +328,10 @@ if(HAVE_CUDA)
set(CUDA_cublas_LIBRARY_ABS ${CUDA_cublas_LIBRARY})
ocv_convert_to_lib_name(CUDA_cublas_LIBRARY ${CUDA_cublas_LIBRARY})
endif()
if(HAVE_CUDNN)
set(CUDNN_LIBRARIES_ABS ${CUDNN_LIBRARIES})
ocv_convert_to_lib_name(CUDNN_LIBRARIES ${CUDNN_LIBRARIES})
endif()
if(HAVE_CUFFT)
set(CUDA_cufft_LIBRARY_ABS ${CUDA_cufft_LIBRARY})
ocv_convert_to_lib_name(CUDA_cufft_LIBRARY ${CUDA_cufft_LIBRARY})

View File

@ -1,5 +1,6 @@
set(MIN_VER_CMAKE 3.5.1)
set(MIN_VER_CUDA 6.5)
set(MIN_VER_CUDNN 6)
set(MIN_VER_PYTHON2 2.7)
set(MIN_VER_PYTHON3 3.2)
set(MIN_VER_ZLIB 1.2.3)

View File

@ -33,7 +33,7 @@
# The verbose template for OpenCV module:
#
# ocv_add_module(modname <dependencies>)
# ocv_glob_module_sources(([EXCLUDE_CUDA] <extra sources&headers>)
# ocv_glob_module_sources(([EXCLUDE_OPENCL] [EXCLUDE_CUDA] <extra sources&headers>)
# or glob them manually and ocv_set_module_sources(...)
# ocv_module_include_directories(<extra include directories>)
# ocv_create_module()

View File

@ -99,10 +99,15 @@ if(HAVE_CUDA)
set(CPACK_DEB_libs_PACKAGE_DEPENDS "${CPACK_DEB_libs_PACKAGE_DEPENDS}, cuda-cufft-${cuda_version_suffix}")
set(CPACK_DEB_dev_PACKAGE_DEPENDS "${CPACK_DEB_dev_PACKAGE_DEPENDS}, cuda-cufft-dev-${cuda_version_suffix}")
endif()
if(HAVE_HAVE_CUBLAS)
if(HAVE_CUBLAS)
set(CPACK_DEB_libs_PACKAGE_DEPENDS "${CPACK_DEB_libs_PACKAGE_DEPENDS}, cuda-cublas-${cuda_version_suffix}")
set(CPACK_DEB_dev_PACKAGE_DEPENDS "${CPACK_DEB_dev_PACKAGE_DEPENDS}, cuda-cublas-dev-${cuda_version_suffix}")
endif()
if(HAVE_CUDNN)
# TODO
#ex: libcudnn7_7.5.1.10-1+cuda10.1_amd64
#ex: libcudnn7-dev_7.5.1.10-1+cuda10.1_amd64
endif()
endif()
set(CPACK_COMPONENT_dev_DEPENDS libs)
endif()

View File

@ -6,6 +6,9 @@ set(OpenCV_USE_CUBLAS "@HAVE_CUBLAS@")
set(OpenCV_USE_CUFFT "@HAVE_CUFFT@")
set(OpenCV_USE_NVCUVID "@HAVE_NVCUVID@")
set(OpenCV_CUDNN_VERSION "@CUDNN_VERSION@")
set(OpenCV_USE_CUDNN "@HAVE_CUDNN@")
if(NOT CUDA_FOUND)
find_host_package(CUDA ${OpenCV_CUDA_VERSION} EXACT REQUIRED)
else()

View File

@ -34,11 +34,14 @@
/* Cocoa API */
#cmakedefine HAVE_COCOA
/* NVIDIA CUDA Runtime API*/
#cmakedefine HAVE_CUDA
/* NVIDIA CUDA Basic Linear Algebra Subprograms (BLAS) API*/
#cmakedefine HAVE_CUBLAS
/* NVIDIA CUDA Runtime API*/
#cmakedefine HAVE_CUDA
/* NVIDIA CUDA Deep Neural Network (cuDNN) API*/
#cmakedefine HAVE_CUDNN
/* NVIDIA CUDA Fast Fourier Transform (FFT) API*/
#cmakedefine HAVE_CUFFT

View File

@ -16,10 +16,16 @@ ocv_option(OPENCV_DNN_OPENCL "Build with OpenCL support" HAVE_OPENCL AND NOT APP
if(OPENCV_DNN_OPENCL AND HAVE_OPENCL)
add_definitions(-DCV_OCL4DNN=1)
else()
ocv_cmake_hook_append(INIT_MODULE_SOURCES_opencv_dnn "${CMAKE_CURRENT_LIST_DIR}/cmake/hooks/INIT_MODULE_SOURCES_opencv_dnn.cmake")
endif()
ocv_option(OPENCV_DNN_CUDA "Build with CUDA support" HAVE_CUDA AND HAVE_CUBLAS AND HAVE_CUDNN)
if(OPENCV_DNN_CUDA AND HAVE_CUDA AND HAVE_CUBLAS AND HAVE_CUDNN)
add_definitions(-DCV_CUDA4DNN=1)
endif()
ocv_cmake_hook_append(INIT_MODULE_SOURCES_opencv_dnn "${CMAKE_CURRENT_LIST_DIR}/cmake/hooks/INIT_MODULE_SOURCES_opencv_dnn.cmake")
if(MSVC)
add_definitions( -D_CRT_SECURE_NO_WARNINGS=1 )
ocv_warnings_disable(CMAKE_CXX_FLAGS /wd4244 /wd4267 /wd4018 /wd4355 /wd4800 /wd4251 /wd4996 /wd4146
@ -73,12 +79,19 @@ endif()
set(include_dirs ${fw_inc})
set(sources_options "")
set(libs libprotobuf ${LAPACK_LIBRARIES})
if(OPENCV_DNN_OPENCL AND HAVE_OPENCL)
list(APPEND include_dirs ${OPENCL_INCLUDE_DIRS})
else()
set(sources_options EXCLUDE_OPENCL)
endif()
if(OPENCV_DNN_CUDA AND HAVE_CUDA AND HAVE_CUBLAS AND HAVE_CUDNN)
list(APPEND include_dirs ${CUDA_TOOLKIT_INCLUDE} ${CUDNN_INCLUDE_DIRS})
else()
set(sources_options ${sources_options} EXCLUDE_CUDA)
endif()
ocv_module_include_directories(${include_dirs})
if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
ocv_append_source_files_cxx_compiler_options(fw_srcs "-Wno-suggest-override") # GCC

View File

@ -1,3 +1,11 @@
message(STATUS "opencv_dnn: filter out ocl4dnn source code")
ocv_list_filterout(OPENCV_MODULE_${the_module}_SOURCES "/ocl4dnn/")
ocv_list_filterout(OPENCV_MODULE_${the_module}_HEADERS "/ocl4dnn/")
if(NOT (OPENCV_DNN_OPENCL AND HAVE_OPENCL))
message(STATUS "opencv_dnn: filter out ocl4dnn source code")
ocv_list_filterout(OPENCV_MODULE_${the_module}_SOURCES "/ocl4dnn/")
ocv_list_filterout(OPENCV_MODULE_${the_module}_HEADERS "/ocl4dnn/")
endif()
if(NOT (OPENCV_DNN_CUDA AND HAVE_CUDA AND HAVE_CUBLAS AND HAVE_CUDNN))
message(STATUS "opencv_dnn: filter out cuda4dnn source code")
ocv_list_filterout(OPENCV_MODULE_${the_module}_SOURCES "/cuda4dnn/")
ocv_list_filterout(OPENCV_MODULE_${the_module}_HEADERS "/cuda4dnn/")
endif()

View File

@ -0,0 +1,18 @@
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
// this file is a stub and will be removed once actual code is added
#include "../precomp.hpp"
#include <cuda_runtime.h>
#ifndef HAVE_CUDA
# error "CUDA files should not be compiled if CUDA was not enabled"
#endif
__global__ void cuda4dnn_build_test_kernel(float* addr) {
int idx = threadIdx.x;
addr[idx] = 0.0;
}

View File

@ -0,0 +1,18 @@
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
// this file is a stub and will be removed once actual code is added
#include "../precomp.hpp"
#ifndef HAVE_CUDA
# error "CUDA4DNN should be enabled iff CUDA and cuDNN were found"
#endif
#include <cudnn.h>
void cuda4dnn_build_test_func() {
auto ver = cudnnGetVersion();
CV_UNUSED(ver);
}

View File

@ -53,10 +53,22 @@
#else
#undef HAVE_OPENCL
#endif
#ifndef CV_CUDA4DNN
#define CV_CUDA4DNN 0
#endif
#if CV_CUDA4DNN
#ifndef HAVE_CUDA
#error "Configuration error: re-run CMake from clean build directory"
#endif
#else
#undef HAVE_CUDA
#endif
#include <opencv2/core/ocl.hpp>
#include <opencv2/core/opencl/ocl_defs.hpp>
#include <opencv2/core/utils/trace.hpp>
#include <opencv2/dnn.hpp>
#include <opencv2/dnn/all_layers.hpp>