I am trying to use CMake to compile a library using cuda.
I want cuda to be optional, i.e. I want my library to compile if nvcc is not available on the system.
Here is how I am checking if cuda is available in my CMake:
include(CheckLanguage)
check_language(CUDA)
if (CMAKE_CUDA_COMPILER)
enable_language(CUDA)
else()
message(WARNING "CUDA not found: GPU features won't be available.")
endif ()
Here are the environment variables I am setting:
$> env | grep CXX
CUDAHOSTCXX=/usr/bin/g++-6
CUDACXX=/usr/bin/nvcc
CXX=/usr/bin/g++
Here is the output of the cmake command:
$> cmake ..
-- The CXX compiler identification is GNU 7.4.0
-- Check for working CXX compiler: /usr/bin/g++
-- Check for working CXX compiler: /usr/bin/g++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Looking for a CUDA compiler
-- Looking for a CUDA compiler - /usr/bin/nvcc
-- The CUDA compiler identification is NVIDIA 9.1.85
-- Check for working CUDA compiler: /usr/bin/nvcc
-- Check for working CUDA compiler: /usr/bin/nvcc -- works
-- Detecting CUDA compiler ABI info
-- Detecting CUDA compiler ABI info - done
-- Configuring done
CMake Error: Error required internal CMake variable not set, cmake may not be built correctly.
Missing variable is:
CMAKE_CUDA_COMPILE_WHOLE_COMPILATION
-- Generating done
CMake Generate step failed. Build files cannot be regenerated correctly.
I am not finding anything usefull conserning CMAKE_CUDA_COMPILE_WHOLE_COMPILATION.
How can I make this cmake works? What did I missed?
I just fingure out what the problem was...
I was calling a function
function(enable_cuda_if_available)
# Here CUDA is properly found and variable are correctly set
include(CheckLanguage)
check_language(CUDA)
if (CMAKE_CUDA_COMPILER)
enable_language(CUDA)
else ()
message(WARNING "CUDA not found: GPU features won't be available.")
endif ()
endfunction()
enable_cuda_if_available()
# Here CUDA's variables have not been forwarded to the parent scope and leading to an error
I still don't know how to automatically forward all variables set by enable_cuda_if_available to the parent scope
Related
I am not very familiar with installing package by conda. I install the OpenMPI by conda install -c conda-forge/label/gcc7 openmpi. After that I write the following CMakeLists.txt file
cmake_minimum_required(VERSION 3.13 FATAL_ERROR)
project(CUDA_MPI LANGUAGES CXX CUDA)
find_package(CUDA REQUIRED)
find_package(MPI REQUIRED)
add_library(Timer STATIC timer.h
timer.cpp)
add_library(Coll STATIC collectives.h
collectives.cu)
add_executable(CUDA_MPI ${CMAKE_CURRENT_SOURCE_DIR}/test/test.cpp)
include_directories(/usr/lib/x86_64-linux-gnu/openmpi/include)
target_link_libraries(CUDA_MPI PUBLIC Timer Coll MPI::MPI_CXX ${CUDA_LIBRARIES})
My configuration log is as the following. It seems that CMake tool can find the library of OpenMPI but it cannot find the head file.
-- The CXX compiler identification is GNU 7.3.0
-- The CUDA compiler identification is NVIDIA 10.0.130
-- Check for working CXX compiler: /home/szhangcj/.conda/envs/distributed/bin/x86_64-conda_cos6-linux-gnu-c++
-- Check for working CXX compiler: /home/szhangcj/.conda/envs/distributed/bin/x86_64-conda_cos6-linux-gnu-c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Check for working CUDA compiler: /usr/local/cuda/bin/nvcc
-- Check for working CUDA compiler: /usr/local/cuda/bin/nvcc -- works
-- Detecting CUDA compiler ABI info
-- Detecting CUDA compiler ABI info - done
-- Looking for C++ include pthread.h
-- Looking for C++ include pthread.h - found
-- Looking for pthread_create
-- Looking for pthread_create - not found
-- 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 (found version "10.0")
-- Found MPI_CXX: /home/szhangcj/.conda/envs/distributed/lib/libmpi_cxx.so (found version "3.1")
-- Found MPI: TRUE (found version "3.1")
-- Configuring done
-- Generating done
-- Build files have been written to: /home/szhangcj/HPC/MPI/baidu-allreduce/build
The error message from the generating stage is as the following.
Scanning dependencies of target Timer
[ 14%] Building CXX object CMakeFiles/Timer.dir/timer.cpp.o
[ 28%] Linking CXX static library libTimer.a
[ 28%] Built target Timer
Scanning dependencies of target Coll
[ 42%] Building CUDA object CMakeFiles/Coll.dir/collectives.cu.o
/home/szhangcj/HPC/MPI/baidu-allreduce/collectives.cu:9:10: fatal error: mpi.h: No such file or directory
#include <mpi.h>
^~~~~~~
compilation terminated.
CMakeFiles/Coll.dir/build.make:62: recipe for target 'CMakeFiles/Coll.dir/collectives.cu.o' failed
make[2]: *** [CMakeFiles/Coll.dir/collectives.cu.o] Error 1
CMakeFiles/Makefile2:147: recipe for target 'CMakeFiles/Coll.dir/all' failed
make[1]: *** [CMakeFiles/Coll.dir/all] Error 2
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2
I have checked the installation directory and I found that there is no include directory under the openmpi folder. How can I fix this?
What's more, on another machine, I repeat the same process to install the Openmpi by using conda. After that, I even cannot find the MPI library.
This question already has answers here:
Listing include_directories in CMake
(3 answers)
Closed 3 years ago.
Since in CMAKE 3.10, CUDA macro is supported by default (https://cmake.org/cmake/help/latest/module/FindCUDA.html).
But I can't find the variable CUDA_INCLUDE_DIRS
cmake_minimum_required(VERSION 3.10 FATAL_ERROR)
project(cmake_and_cuda LANGUAGES CXX CUDA)
message(${CUDA_INCLUDE_DIRS})
the errors are
-- The CXX compiler identification is GNU 5.4.0
-- The CUDA compiler identification is NVIDIA 10.0.130
-- Check for working CXX compiler: /usr/bin/c++
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Check for working CUDA compiler: /usr/local/cuda/bin/nvcc
-- Check for working CUDA compiler: /usr/local/cuda/bin/nvcc -- works
-- Detecting CUDA compiler ABI info
-- Detecting CUDA compiler ABI info - done
CMake Error at CMakeLists.txt:7 (message):
message called with incorrect number of arguments
-- Configuring incomplete, errors occurred!
See also "/home/tumh/code-samples/posts/cmake/build/CMakeFiles/CMakeOutput.log".
Any idea?
Just use the ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}.
In all .cu src files, there is no need to manually add
include_directories(${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}),
because CMake will do it for you.
But for .cpp src files, if you need to include <cuda_runtime.h>, you must manually add
include_directories(${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
in your CMakeLists.txt.
I have this following (minimal) CMakeLists.txt supposed to find GLib via pkg-config and add the libs to the foo target:
cmake_minimum_required(VERSION 2.10 FATAL_ERROR)
project(foo)
find_package(PkgConfig REQUIRED)
pkg_check_modules(GLib REQUIRED glib-2.0)
add_executable(foo foo.cpp)
message(WARNING "libs:" ${GLIB_LIBRARIES})
message(WARNING "includes:" ${GLIB_INCLUDE_DIRS})
target_link_libraries(foo PUBLIC ${GLIB_LIBRARIES})
target_include_directories(foo PUBLIC ${GLIB_INCLUDE_DIRS})
No matter what I try, I get (note the Found glib-2.0 part):
-- The C compiler identification is GNU 7.3.0
-- The CXX compiler identification is GNU 7.3.0
-- Check for working C compiler: /usr/bin/cc
-- Check for working C compiler: /usr/bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compile features
-- Detecting C compile features - done
-- Check for working CXX compiler: /usr/bin/c++
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found PkgConfig: /usr/bin/pkg-config (found version "0.29.1")
-- Checking for module 'glib-2.0'
-- Found glib-2.0, version 2.56.1
CMake Warning at CMakeLists.txt:6 (message):
libs:
CMake Warning at CMakeLists.txt:7 (message):
includes:
CMake Error at CMakeLists.txt:9 (target_include_directories):
target_include_directories called with invalid arguments
-- Configuring incomplete, errors occurred!
See also "/tmp/aaa/CMakeFiles/CMakeOutput.log".
and I cannot see, reading the CMake reference what arguments are invalid (note: this question is different from cmake target_include_directories called with invalid arguments). I also looked at CMake's FindPkgConfig documentation which gives glib as an example and I am not able to reproduce it (${GLIB_VERSION}). I tried GLIB_ and GLIB2_ prefixes and all I get is empty strings.
The messages show the variables are empty, though pkg-config reports values correctly:
$ pkg-config glib-2.0 --cflags --libs
-I/usr/include/glib-2.0 -I/usr/lib/x86_64-linux-gnu/glib-2.0/include -lglib-2.0
The CMake version I have is 2.10.
Can someone shed light on the issue?
I think your problem is the casing of GLIB_INCLUDE_DIRS and GLIB_LIBRARIES. They should be GLib_INCLUDE_DIRS and GLib_LIBRARIES since you specified "GLib" as the first argument of pkg_check_modules. I'm guessing that the target_include_directories() doesn't like not getting any arguments after the
PUBLIC (although it works for me in cmake 3.5).
I am trying to compile a library with both, C++ and CUDA source files. I am using GNU make with CMake. My compiler of choice is clang, since CUDA only supports gcc up to version 5 and Debian 9 has gcc 6 as its oldest version and I have to use software that is provided by the Debian 9 or 10 repositories.
CMake version is 3.9.0
clang version is 3.8.1
cc and c++ in /usr/bin correctly link to clang and clang++ ẃhich also link to the correct files.
Unfortunately the initial checks of CMake for CUDA fail although everything, as far as I can see, seems to be set up correctly. It looks like the arguments aren't passed correctly to the CUDA compiler.
This is a part of my project's main CMake file:
cmake_minimum_required (VERSION 3.9.0 FATAL_ERROR)
project (dev)
find_package(CUDA REQUIRED)
set(CUDA_HOST_COMPILATION_CPP ON)
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -Wno-deprecated-gpu-targets -ccbin clang-3.8")
The library's CMake file looks like this:
cmake_minimum_required(VERSION 3.9.0 FATAL_ERROR)
project (libname LANGUAGES CXX CUDA)
file(GLOB SOURCES "*.cu" "*.cpp")
add_library(libname ${SOURCES})
set_target_properties(libname PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
set_target_properties(libname PROPERTIES POSITION_INDEPENDENT_CODE ON)
This is the output of CMake:
-- The C compiler identification is Clang 3.8.1
-- The CXX compiler identification is Clang 3.8.1
-- Check for working C compiler: /usr/bin/cc
-- Check for working C compiler: /usr/bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compile features
-- Detecting C compile features - done
-- Check for working CXX compiler: /usr/bin/c++
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Looking for pthread.h
-- Looking for pthread.h - found
-- Looking for pthread_create
-- Looking for pthread_create - not found
-- 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 (found version "8.0")
-- The CUDA compiler identification is unknown
-- Check for working CUDA compiler: /usr/bin/nvcc
-- Check for working CUDA compiler: /usr/bin/nvcc -- broken
The CUDA/nvcc tests of CMake fails with the following errors:
Change Dir: /home/user/projects/hamonIC-linux-experimental/current_state/working_copy/code/build/CMakeFiles/CMakeTmp
Run Build Command:"/usr/bin/make" "cmTC_6d1a9/fast"
/usr/bin/make -f CMakeFiles/cmTC_6d1a9.dir/build.make
CMakeFiles/cmTC_6d1a9.dir/build
make[1]: Entering directory
'/home/user/projects/hamonIC-linux-experimental/current_state/working_copy/code/build/CMakeFiles/CMakeTmp'
Building CUDA object CMakeFiles/cmTC_6d1a9.dir/main.cu.o
/usr/bin/nvcc -x cu -c
/home/user/projects/hamonIC-linux-experimental/current_state/working_copy/code/build/CMakeFiles/CMakeTmp/main.cu
-o CMakeFiles/cmTC_6d1a9.dir/main.cu.o
nvcc warning : The 'compute_20', 'sm_20', and 'sm_21' architectures are
deprecated, and may be removed in a future release (Use
-Wno-deprecated-gpu-targets to suppress warning).
ERROR: No supported gcc/g++ host compiler found, but clang-3.8 is
available.
Use 'nvcc -ccbin clang-3.8' to use that instead.
CMakeFiles/cmTC_6d1a9.dir/build.make:65: recipe for target
'CMakeFiles/cmTC_6d1a9.dir/main.cu.o' failed
make[1]: *** [CMakeFiles/cmTC_6d1a9.dir/main.cu.o] Error 1
make[1]: Leaving directory
'/home/user/projects/hamonIC-linux-experimental/current_state/working_copy/code/build/CMakeFiles/CMakeTmp'
Makefile:126: recipe for target 'cmTC_6d1a9/fast' failed
make: *** [cmTC_6d1a9/fast] Error 2
CMake will not be able to correctly generate this project.
Call Stack (most recent call first):
hic_iplibrary/source/ciccommon/CMakeLists.txt:7 (project)
The CUDA NVCC flags in CMake have to be semicolon delimited and not whitespace delimited.
Change the flags in your CMakeLists.txt to use:
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS};-Wno-deprecated-gpu-targets;-ccbin=clang-3.8")
and that should allow CMake to pass your flags to NVCC.
For reasons that I cannot completely understand, Cmake awlays chooses the GNU compiler toolset when compiling software.
My enviroment looks like this:
which cc
/opt/cray/xt-asyncpe/4.9/bin/cc
which CC
/opt/cray/xt-asyncpe/4.9/bin/CC
echo $CC
/opt/cray/xt-asyncpe/4.9/bin/cc
echo $CXX
/opt/cray/xt-asyncpe/4.9/bin/CC
but when I use cmake I get this
Using existing /opt/cmake/2.8.4/bin/cmake
-- The C compiler identification is GNU
-- The CXX compiler identification is GNU
-- Check for working C compiler: /usr/bin/gcc
-- Check for working C compiler: /usr/bin/gcc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
And it builds all the software with g++ commands. Why is this going on? How does one set the compiler?
You can also set the env vars CC and CXX much like autotools.
CC=cc CXX=CC cmake ...
Make sure you start with an empty build tree.
I'm not sure why CMake favours GCC.
However, to set the compilers, use:
cmake <path to CMakeLists.txt> -DCMAKE_C_COMPILER=/opt/cray/xt-asyncpe/4.9/bin/cc -DCMAKE_CXX_COMPILER=/opt/cray/xt-asyncpe/4.9/bin/CC
These values will be cached, so subsequent runs of CMake (if required) can simply be invoked by:
cmake .