How can I add the cusparse library from CUDA in a CMakeLists.txt-file, such that the nvcc compiler includes it automatically with -lcusparse? I already added the line
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-lcusparse)
in CMakeLists.txt with no success. It looks like I'm missing something, because Nsight throws the error
undefined reference to 'cusparseDestroyMatDescr'.
Although when I exclude this line where cusparseDestroyMatDescr is called via commenting it, the Nsight project builds with no error, even with these three lines of code included
cusparseStatus_t status;
cusparseHandle_t handle=0;
cusparseMatDescr_t descr=0;
So it looks like it knows what cusparseStatus_t and so on is, but it does not know what cusparseDestroyMatDescr is.
What do I miss?
The correct way in CMake to link a library is using
target_link_libraries( target library ).
If you use FindCUDA to locate the CUDA installation, the variable CUDA_cusparse_LIBRARY will be defined. Thus, all you need to do is
target_link_libraries( target ${CUDA_cusparse_LIBRARY} )
I recommend to use the CMake CUDAToolkit package, which is available with CMake 3.17 and newer:
find_package(CUDAToolkit REQUIRED)
...
target_link_libraries(target CUDA::cusparse)
Related
I compile cocos2d-x(version 3.6) using visual studio 2015, the error occurred, saying:
fatal error C1189: #error: Macro definition of snprintf conflicts with Standard Library function declaration
Almost the same question like this link
here
I try to follow the first answer and then search most results on cocos forum but also failed, I'm noob and really have no idea now ..
And here it's my source code where defined snprintf on header file stdio.h
#if defined snprintf
// This definition of snprintf will generate "warning C4005: 'snprintf': macro
// redefinition" with a subsequent line indicating where the previous definition
// of snprintf was. This makes it easier to find where snprintf was defined.
#pragma warning(push, 1)
#pragma warning(1: 4005)
#define snprintf Do not define snprintf as a macro
#pragma warning(pop)
#error Macro definition of snprintf conflicts with Standard Library function declaration
#endif
Could someone help me .. thanks!
I am getting the same error trying to build libsndfile-1. I solved it by building using VS2013 instead of VS2015. (I think it should be possible to simply install VS2013 Build Tools and build from VS2015).
edit: to install the VS2013 build toolset, run the VS2015 installer and select 'Windows 8.1 and Windows Phone 8.0/8.1 Tools'
I am trying to compile a CUDA 5.5 application on nsight with ubuntu 12.04
At first I was getting an issue about missing header files such as #include <helper_cuda_drvapi.h>
To fix this I added the path /usr/include/samples/common/inc to my includes list.
This solved the missing header file issue but caused a new issue.
when trying to compile the program on nsight I get the following errors
/usr/local/cuda-5.5/samples/common/inc/helper_cuda_drvapi.h:278: undefined reference to cuInit'
/usr/local/cuda-5.5/samples/common/inc/helper_cuda_drvapi.h:279: undefined reference tocuDeviceGetCount'
/usr/local/cuda-5.5/samples/common/inc/helper_cuda_drvapi.h:290: undefined reference to cuDeviceGetName'
/usr/local/cuda-5.5/samples/common/inc/helper_cuda_drvapi.h:291: undefined reference tocuDeviceComputeCapability'
/usr/local/cuda-5.5/samples/common/inc/helper_cuda_drvapi.h:294: undefined reference to cuDeviceGetAttribute'
/usr/local/cuda-5.5/samples/common/inc/helper_cuda_drvapi.h:327: undefined reference tocuDeviceGetAttribute'
/usr/local/cuda-5.5/samples/common/inc/helper_cuda_drvapi.h:330: undefined reference to cuDeviceGetAttribute'
/usr/local/cuda-5.5/samples/common/inc/helper_cuda_drvapi.h:333: undefined reference tocuDeviceComputeCapability'
/usr/local/cuda-5.5/samples/common/inc/helper_cuda_drvapi.h:336: undefined reference to `cuDeviceGetAttribute'
any suggestions?
Thanks in advance
*****************UPDATE************
What it basically comes down to is I am trying to compile the "CUDA Video Decoder GL API" sample program on linux and it is not working because of some error with the header files. Does anyone know why this is?
UPDATE
The undefined references are to CUDA driver API methods. helper_cuda_drvapi.h has the following comment near the top:
Helper functions for CUDA Driver API error handling (make sure that CUDA_H is included in your projects)
So, in your .cu and .cpp files, before the #include <helper_cuda_drvapi.h>, include cuda.h:
#include "cuda.h"
#include <helper_cuda_drvapi.h>
See this question for more information about the CUDA headers.
You need to manually link with libcuda (Nsight projects use Runtime API)
To link with this library:
Go to Properties for your project, open General / Path and Symbols
On the Libraries tab add cuda (without prefix or suffix - in theory should make your project more clossplatform. You may also want to check "Add to all configurations" when adding the library - otherwise it will be for your current build configuration (e.g. "Debug" or "Release" only).
Update: Project settings screenshot:
I have a project that builds a library and I want to add some cuda support to it.
The structure is:
|Basedir
|_subdir1
|_subdir2
The basic structure of the CMakeLists.txt files: (subdir2 is not important).
in Basedir:
cmake_minimum_required(VERSION 2.6)
PROJECT(myproject)
find_package(CUDA)
INCLUDE_DIRECTORIES(${MYPROJECT_SOURCE_DIR})
ADD_SUBDIRECTORY(subdir1)
ADD_SUBDIRECTORY(subdir2)
in subdir1:
ADD_LIBRARY(mylib shared
file1.cpp
file2.cpp
file3.cpp
)
INSTALL(
TARGETS mylib
DESTINATION lib
PERMISSIONS
OWNER_READ OWNER_WRITE OWNER_EXECUTE
GROUP_READ GROUP_EXECUTE
WORLD_READ WORLD_EXECUTE
)
FILE(GLOB_RECURSE HEADERS RELATIVE ${MYPROJECT_SOURCE_DIR}/myproject *.h)
FOREACH(HEADER ${HEADERS})
STRING(REGEX MATCH "(.*)[/\\]" DIR ${HEADER})
INSTALL(FILES ${HEADER} DESTINATION include/myproject/${DIR})
ENDFOREACH(HEADER)
I actually don't really know how to put the cuda-support into it. I want to replace file2.cpp with file2.cu and I did that, but it didn't build the .cu file, only the cpp files.
Do I have to add CUDA_ADD_EXECUTABLE() to include any cuda-files? How will I then link it to the other files?
I tried adding the following to the CMakeLists.txt in subdir1:
CUDA_ADD_EXECUTABLE(cuda file2.cu OPTIONS -arch sm_20)
That will compile the file but build an executable cuda. How do I link it to mylib?
Just with?:
TARGET_LINK_LIBRARIES(cuda mylib)
I have to admit that I'm not experienced in cmake, but I guess you figured that.
You can use CUDA_ADD_LIBRARY for mylib project. It works as CUDA_ADD_EXECUTABLE but for libraries.
CUDA_ADD_LIBRARY(mylib SHARED
file1.cpp
file2.cu
file3.cpp
OPTIONS -arch sm_20
)
TARGET_LINK_LIBRARIES(mylib ${CUDA_LIBRARIES})
I need to add the csv module in one module compute.ml of an OCAML package.(see discussion ocaml hash from mysql)
Do you know of a simple way to amend the makefile (it is a complex document in my case and I don't think that I can fully rewrite it or switch to ocamlfind install) to enable a compilation without "Unbound value Csv.load" error messages?
ie can I include a new library inside an existing stable Ocaml package?
modified module : compute.ml,
modification;
let data = Csv.load ("foo.csv")
....
error message during compilation :
camlp5r ../wserver/pa_macro5.cmo -DUNIX -o compute.ppo compute.ml
ocamlopt.opt -warn-error A -I ../wserver -I ../sally -I +camlp5 -c -impl compute.ppo
File "compute.ml", line 110, characters 13-21:
Error: Unbound value Csv.load
Thanks for help
You have to tell ocamlopt where to look for csv.cm{i,x} files with the appropriate -I option, like the ones you already have at the end of your command line.
If the library lies in a subdirectory of OCaml's standard library (as given by ocamlc -where), you can use
-I +csv_dir, as is done for camlp5 in your example. Otherwise, you'll have to provide the full path.
Note that this is not the only modification that you'll have to do: you'll also have to add csv.cmx (or .cmxa) on the command line that perform the final link of your application/library.
I am trying to set up Point Cloud Library trunk build with CUDA options enabled.
I believe I have installed CUDA correctly, following these instructions.
In the cmake options for the PCL build, some options are unrecognised:
Is there something I can manually set CUDA_SDK_ROOT_DIR to? Likewise for the other unfound options.
CUDA_SDK_ROOT_DIR should be set to the direction in which you installed the NVIDIA's GPU Computing SDK. The GPU Computing SDK is downloadable from the same page at NVIDIA where you downloaded CUDA. By default, this SDK will install to $HOME/NVIDIA_GPU_Computing_SDK. Set it appropriately and then rerun cmake.
Edit:
The CUDA_SDK_ROOT_DIR variable is actually looking for the sub-directory beneath $HOME/NVIDIA_GPU_Computing_SDK that contains the version of CUDA you're using. For me, this is $HOME/NVIDIA_GPU_Computing_SDK/CUDA/v4.1.
The source code for FindCUDA.cmake gives some hints on how this path is found:
########################
# Look for the SDK stuff. As of CUDA 3.0 NVSDKCUDA_ROOT has been replaced with
# NVSDKCOMPUTE_ROOT with the old CUDA C contents moved into the C subdirectory
find_path(CUDA_SDK_ROOT_DIR common/inc/cutil.h
"$ENV{NVSDKCOMPUTE_ROOT}/C"
"$ENV{NVSDKCUDA_ROOT}"
"[HKEY_LOCAL_MACHINE\\SOFTWARE\\NVIDIA Corporation\\Installed Products\\NVIDIA SDK 10\\Compute;InstallDir]"
"/Developer/GPU\ Computing/C"
)
I.e. check that NVSDKCOMPUTE_ROOT or NVSDKCUDA_ROOT environment variables are set correctly.
On a Linux machine,..
Add "$ENV{HOME}/NVIDIA_GPU_Computing_SDK/C" to the 'find_path' options in FindCUDA.cmake module: (usr/share/cmake-2.8/Modules/FindCUDA.cmake)
########################
# Look for the SDK stuff. As of CUDA 3.0 NVSDKCUDA_ROOT has been replaced with
# NVSDKCOMPUTE_ROOT with the old CUDA C contents moved into the C subdirectory
find_path(CUDA_SDK_ROOT_DIR common/inc/cutil.h
"$ENV{HOME}/NVIDIA_GPU_Computing_SDK/C"
"$ENV{NVSDKCOMPUTE_ROOT}/C"
"$ENV{NVSDKCUDA_ROOT}"
"[HKEY_LOCAL_MACHINE\\SOFTWARE\\NVIDIA Corporation\\Installed Products\\NVIDIA SDK 10\\Compute;InstallDir]"
"/Developer/GPU\ Computing/C"
)
cmake now finds my 4.0 SDK automatically.
But my build still fails to find cutil.h, even though it is there. $HOME/NVIDIA_GPU_Computing_SDK/C/common/inc/cutil.h. I had to add an include flag to the project to get it to finally work. CUDA_NVCC_FLAGS : -I/home/bill/NVIDIA_GPU_Computing_SDK/C/common/inc
Note: -I/$HOME/NVIDIA_GPU_Computing_SDK/C/common/inc does NOT work. (The env $HOME is set correctly.)