Building OpenCV code using Bazel

问题:

What is the best way to build C++ code that uses the OpenCV library using Bazel? I.e., what would the BUILD rules look like?

Bazel.io has docs for external dependencies but it's not very clear.

回答1:

There are a couple of options. The easiest way is probably to install locally in the way the OpenCV site recommends:

git clone https://github.com/Itseez/opencv.git
cd opencv/
mkdir build install
cd build
cmake -D CMAKE_BUILD_TYPE=Release -D CMAKE_INSTALL_PREFIX=/path/to/opencv/install ..
make install

Then add the following to your WORKSPACE file:

new_local_repository(
    name = "opencv",
    path = "/path/to/opencv/install",
    build_file = "opencv.BUILD",
)

Create opencv.BUILD in the same directory as WORKSPACE with the following:

cc_library(
    name = "opencv",
    srcs = glob(["lib/*.so*"]),
    hdrs = glob(["include/**/*.hpp"]),
    includes = ["include"],
    visibility = ["//visibility:public"], 
    linkstatic = 1,
)

Then your code can depend on @opencv//:opencv to link in the .so's under lib/ and reference the headers under include/.

However, this isn't very portable. If you want a portable solution (and you're feeling ambitious), you could add the OpenCV git repo to your workspace and download & build it. Something like:

# WORKSPACE
new_git_repository(
    name = "opencv",
    remote = "https://github.com/Itseez/opencv.git",
    build_file = "opencv.BUILD",
    tag = "3.1.0",
)

And make opencv.BUILD something like:

cc_library(
    name = "core",
    visibility = ["//visibility:public"],
    srcs = glob(["modules/core/src/**/*.cpp"]),
    hdrs = glob([
        "modules/core/src/**/*.hpp", 
        "modules/core/include/**/*.hpp"]
    ) + [":module-includes"],
)

genrule(
    name = "module-includes",
    cmd = "echo '#define HAVE_OPENCV_CORE' > $@",
    outs = ["opencv2/opencv_modules.hpp"],
)

...

Then your code could depend on more specific targets, e.g., @opencv//:core.

As a third option, you declare both cmake and OpenCV in your WORKSPACE file and use a genrule to run cmake on OpenCV from within Bazel.

回答2:

I succeed with @kristina's first option.

1. Install opencv:

   git clone https://github.com/Itseez/opencv.git
   
   cd opencv/
   
   mkdir build install
   
   cd build
   
   cmake -D CMAKE_BUILD_TYPE=Release -D CMAKE_INSTALL_PREFIX=/usr/local ..
   
   make install
   

2. Change WORKSPACE file (at tensorflow/WORKSPACE cloned from github)

   new_local_repository(
   
   name = "opencv",
   
   path = "/usr/local",
   
   build_file = "opencv.BUILD",
   
   )
   

3. Make opencv.BUILD file at the same place as WORKSPACE file:

   cc_library(
   
   name = "opencv",
   
   srcs = glob(["lib/*.so*"]),
   
   hdrs = glob(["include/**/*.hpp"]),
   
   includes = ["include"],
   
   visibility = ["//visibility:public"], 
   
   linkstatic = 1,
   
   )
   

4. You may have to config the opencv libs path:

a. Make sure you have /etc/ld.so.conf.d/opencv.conf file with content:

    /usr/local/lib

b. Run the command:

    sudo ldconfig -v

回答3:

This is what I did for OpenCV 2.4.13.2, core/ only. This approach goes from the opencv source, which is adapted from the accepted answer above by @kristina.

The first thing is to add the http_archive for the opencv 2.4 release:

# OpenCV 2.4.13.2
new_http_archive(
    name = "opencv2",
    url = "https://github.com/opencv/opencv/archive/2.4.13.2.zip",
    build_file = "third_party/opencv2.BUILD",
    strip_prefix = "opencv-2.4.13.2",
)

And then, add the file third_party/opencv2.BUILD as:

cc_library(
    name = "dynamicuda",
    hdrs = glob([
        "modules/dynamicuda/include/**/*.hpp",
    ]),
    includes = [
        "modules/dynamicuda/include"
    ],
)

cc_library(
    name = "core",
    visibility = ["//visibility:public"],
    srcs = glob(["modules/core/src/**/*.cpp"]),
    hdrs = glob([
        "modules/core/src/**/*.hpp",
        "modules/core/include/**/*.hpp",
     ]) + [
        ":module_includes",
        ":cvconfig",
        ":version_string",
    ],
    copts = [
        "-Imodules/dynamicuda/include",
    ],
    # Note that opencv core requires zlib and pthread to build.
    linkopts = ["-pthread", "-lz"],
    includes = [
        "modules/core/include",
    ],
    deps = [
        ":dynamicuda",
    ],
)

genrule(
    name = "module_includes",
    cmd = "echo '#define HAVE_OPENCV_CORE' > $@",
    outs = ["opencv2/opencv_modules.hpp"],
)

genrule(
    name = "cvconfig",
    outs = ["cvconfig.h"],
    cmd = """
cat > $@ <<"EOF"
// JPEG-2000
#define HAVE_JASPER

// IJG JPEG
#define HAVE_JPEG

// PNG
#define HAVE_PNG

// TIFF
#define HAVE_TIFF

// Compile for 'real' NVIDIA GPU architectures
#define CUDA_ARCH_BIN ""

// NVIDIA GPU features are used
#define CUDA_ARCH_FEATURES ""

// Compile for 'virtual' NVIDIA PTX architectures
#define CUDA_ARCH_PTX ""
EOF"""
)

genrule(
    name = "version_string",
    outs = ["version_string.inc"],
    cmd = """
cat > $@ <<"EOF"
"\\n"
)

Note that I did not put anything in the version_string.inc. It is just a C++ string literal which does not affect the functionality of OpenCV. If you are really interested in this file see this example.

After this you should be able to add target with dependencies on @opencv2//:core.