I have Nvidea SDK 7.5. I want to compile OpenMP for CUDA. On step "Clone, build & install OpenMP target backends" I get:
Yet I see no libcuda.lib at all in my SDK:
So I wonder what is difference between lcuda and lcudart and if there is any where one can find lcuda in CUDA 7.5 SDK?
The l you are discussing in "lcuda" and "lcudart" is actually part of the compiler switch. The library is cuda or libcuda, and cudart or libcudart.
On linux, I don't think these libraries end in .lib either. They should end in .so That's pretty evident in your printout - take a look.
Anyway, libcuda.so is installed by the GPU driver installer, not the toolkit.
(except possibly for stubs, which you should not use) libcuda.so will not be found in the toolkit location but somewhere else where the driver puts it. This particular location will vary depending on the specific linux distro.
So I suggest using find or a similar linux command to locate it on your machine. It may be in /usr/lib or /usr/lib64 or someplace like that.
If you can't find it at all, then the likely explanation is that you have not installed (properly) the GPU driver on your machine.
In a nutshell, libcuda.so provides access to the CUDA driver API, whereas libcudart.so provides access to the CUDA runtime API.
By the way, in the future, please post text output, not pictures, for actual text output from your console.
Related
I'm currently working on a Linux project. This project needs to run under every Linux distribution (without installing any package/libraries/others for the clients) and it's a bit hard to do it well.
I already tried to do it myself, see this, i have also tried to use CDE but it didn't work well since i got an error with some distribution. For example:
Ubuntu 8.04: Impossible to read the header ELF
Debian 7.8: version of GLIBC_2.14 not found
So, i would like to know if there is a way to get a package of my program who can run under every Linux distribution.
Thanks
Edit: I would like to avoid the static compilation, since my program is pretty big.
There are big differences between linux distributions, especially version of libraries and package management system.
The only way how to do it is to build/compile your project against all libraries you need to use statically, and distribute them with your project.
For example skype and ejabberd do it this way.
I've a problem with a library developed in Visual C++ with Cuda.
I want to use my dll on different computers, all with NVIDIA drivers, but it depends
by CUDART32_42_9.DLL and CUBLAS32_43_9.DLL; so, depending on 32_42_9, it doesn't work with any different cuda version: neither inferiror (that is reasonable) nor superior (that is my problem).
Have you got any suggestion about how to solve the problem? Do I have to deploy my library with CUDART32_42_9.DLL and CUBLAS32_43_9.DL attached?
Thank you!
It is the responsibility of the application developer to redistribute the necessary CUDA libraries with the application. The EULA.txt document in the CUDA Toolkit doc directory has a section Attachment A that defines Redistributable Software. Redistributable software includes but is not limited to cudart* and cublas*. See EULA.txt for more details.
The end user will need to download a compliant version of the NVIDIA graphics driver for their platform.
I have developed a small application in Qt Creator on Ubuntu 12.04 which I want should run on any other linux distro (mostly different versions of CentOS and ubuntu), just like any portable application on windows does.
I want to be able to simply share the binary file of the Application, and run the application.
I am able to successfully do this in windows, by just building the project in QT Creator and then putting the required libraries in the Application directory and then transfering them to other windows systems.
I searched all over and found out that I should be trying to build the project using LSB(Linux Standard Base) Compatibility, so that it runs on other linux distros. Is that the right way to do this?
I am very new to Qt and also to Linux (dont know much of Shell Scripting).
Thus, I dont know how I should proceed to make the Application LSB Compliant.
I have refered to, the following links:
Distributing Qt-based binaries on Linux and
Deploying Qt applications on Linux but have not beem able to understand what I am suposed to do.
I also found this question here which states a very similar situation as mine, but because I am a novice, I dont know how I should do this.
Moreover, considering that the first two articles were written 6 years back, shouldn't there be a simpler way to deploy Qt apps on the linux platform now?
I also saw something about static linking, is that the way to go?
Isn't there a way by which all of this can be done through Qt Creator itself?
If there is no hope of creating a portable Qt Application for Linux, then is there a way, say a shell script or something that would combine all the steps required to compile the Qt project on another computer and run it. Say, download Qt-SDK if not present, run qmake and make and then the newly compiled application, if not already there, so that the user can run the program just by running one script.
Your problem here is not the Linux Standard Base, but rather the presence or not of the specific version of Qt you need (or a later one).
Exactly like on a Windows machine, a user may have any of Qt installed, or they may not have it at all. On Windows it is easier to check for the presence of a certain version of Qt than it is on Linux, thus it is easier to write install tools that automate the experience.
To solve your problem there are a few ways:
Inform the user that your program requires a certain version of Qt or higher, and let the user handle the problem
Learn how to create packages for every distribution you want to target and create specific packages
Use a program like 0Install or Elf Statifier to create a package/executable containing all the necessary libraries.
The latter is similar to what many Windows and Mac programs do (they include every library they need within the installer), but it is not the preferred way on Linux, which relies heavily on shared libraries.
Making a binary application compatible with any other Linux distro is practically impossible since you will never know in advance which libraries are available in distro X, or what version of that library is available. Even among a single distro (e.g. Ubuntu), binary application are almost never backward-compatible, since anything built on Ubuntu 12.04 will have dependencies on versions libraries which are installed on that version of Ubuntu, and trying to run that binary on Ubuntu 10.04 will most probably fail simply because it doesn't have a recent enough version of glibc or some other necessary library.
However, the idea can be much more implementable if you limit yourself to a finite list of distros and versions of those distros. You can then know which libraries are available for those distros, and aim for the lowest common denominator. I used to maintain a binary application which had to support several distros (Ubuntu, Fedora, OpenSUSE, SLED, Mandriva), and the way I would do it is install the oldest distro I was targeting on my build machine. That way, the binary application would be linked to the oldest versions of the libraries available on those distros. Unless there's a new major version of such a library (which happens quite rarely, and even then, distros usually distribute the previous major version for a while for compatibility purposes), your compiled binary will then be compatible with all your targeted distros.
Therefore, the quick piece of advice I would give for your situation, use the oldest LTS version of Ubuntu which is still supported (10.04 at the moment) for your development, and you should be pretty safe for most recent popular distros. For the application you already developped on Ubuntu 12.04, you should have no problem simply recompiling the same source on 10.04. Understand that you will never however achieve 100% compatibility with a compiled C++ Qt application.
If Qt is not all that important to you, you could use a higher-level or interpreted language such as Python, Java, Perl or Ruby. With such languages, you can usually count on the language implementation already being installed on the target distro.
Deploy an application in Linux is a nightmare, luckily there are some solutions. Check this projects to build a portable binary with all their dependencies bundled:
http://statifier.sourceforge.net/statifier/main.html
http://www.magicermine.com/index.html
http://www.pgbovine.net/cde.html
Another solution is make a portable 0install package:
http://0install.net/
I recomend this solution. Personally I have been problems with the 3 first packagers.
I'd like to set up a cross-compilation environment on a Ubuntu 9.10 box. From the documents I've read so far (these ones, for example) this involves compiling the toolchain of the target platforms.
My question is: how do you determine the required version of each of the packages in the toolchain for a specific target platform? Is there any rule of thumb I can follow?
This is a list found in one of the websites linked above:
binutils-2.16.1.tar.bz2
linux-2.6.20.1.tar.bz2
glibc-2.5.tar.bz2
glibc-linuxthreads-2.5.tar.bz2
gcc-core-4.2.0.tar.bz2
gcc-g++-4.2.0.tar.bz2
But suppose I want to generate executables for standard Ubuntu 8.04 and CentOS 5.3 boxes. What are the necessary packages?
My primary need is to avoid errors like "/usr/lib/libstdc++.so.6: version `GLIBCXX_3.4.11' not found" in the customers' machines but in the future I want to deal with different architectures as well.
It is generally a good idea to build a cross-toolchain that uses the same version of libc (and other libraries) found on the target system. This is especially important in the case of libraries that use versioned symbols or you could wind up with errors like "/usr/lib/libstdc++.so.6: version 'GLIBCXX_3.4.11' not found".
Same Architecture
For generating executables for standard Ubuntu 8.04 and CentOS 5.3 systems, you could install the distributions in virtual machines and do the necessary compilation from within the virtual machine to guarantee the resulting binaries are compatible with the library versions from each distribution.
Another option would be to setup chroot build environments instead of virtual machines for the target distributions.
You could also build toolchains targeted at different environments (different library versions) and build under your Ubuntu 9.10 environment without using virtual machines or chroot environments. I have used Dan Kegel's crosstool for creating such cross-toolchains.
Different Architecture
As I noted in my answer to a another cross-compiler question, I used Dan Kegel's crosstool for creating my arm cross-toolchain.
It appears it may be slightly out of date, but there is a matrix of build results for various architectures to help determine a suitable combination of gcc, glibc, binutils, and linux kernel headers.
Required Package Versions
In my experience, there really isn't a rule of thumb. Not all combinations of gcc, binutils, glibc, and linux headers will build successfully. Even if the build completes, some level of testing is necessary to validate the build's success. This is sometimes done by compiling the Linux kernel with your new cross-toolchain. Depending on the target system and architecture, some patching of the source may be necessary to produce a successful build.
Since you are setting up this cross-compilation environment on Ubuntu 9.10, you might want to look into the dpkg-cross package.
Compiling for other Linux distributions is easiest by installing them in virtual machines (apt-get install kvm) and then doing the compilation from within. You can also script them to do it automatically. Building a cross-compiler and providing the exact same versions of all libraries and such, as the other Linux distro does, is nearly impossible.
My question is: how do you determine
the required version of each of the
packages in the toolchain for a
specific target platform?
...
binutils-2.16.1.tar.bz2
gcc-core-4.2.0.tar.bz2
gcc-g++-4.2.0.tar.bz2
Generally pick the latest stable: these only affect your local toolchain, not runtime.
linux-2.6.20.1.tar.bz2
You don't need this. (For targeting embedded platforms you might use it.)
glibc-2.5.tar.bz2
glibc-linuxthreads-2.5.tar.bz2
You don't need these. I.e. you should not download them or build them; you should link against the versions from the oldest distro you want to support.
Is there any
rule of thumb I can follow?
But suppose I want to generate
executables for standard Ubuntu 8.04
and CentOS 5.3 boxes. What are the
necessary packages?
You survey the distros you want to target, find the lowest common denominator versions of
of libc, libstdc++, pthreads, and any other shared library you will link with, then copy these libs and corresponding headers from the box that has these LCD versions to your toolchain.
[edit] I should clarify, you really want to get all the dependent libs from a single system. Picking and choosing the LCD of each file version from different distributions is a recipe for a quick trip to dependency hell.
Depending on your target platforms, have you considered using Optware?
I'm currently working on getting Mono and Moonlight built for my Palm Pre using the cross-compilation toolchain (and the Optware makefiles handle the majority of dependencies already).
Is there anyway to write dlls in linux?
Do I have to install windows to write dlls in linux? Right now one of my courses requires me to write a dll for this.
You should take a look into 'shared libraries'
http://www.linux.org/docs/ldp/howto/Program-Library-HOWTO/shared-libraries.html
Lots of folks are getting near the right answer but not providing it: gcc can generate win32 PE/COFF files without problem, and of course can always build as a cross compiler on any platform it can target. The binutils port targets windows .exe and .dll files natively, and there's a "dlltool" utility for handling the edge cases where Unix and Windows linkage metaphors are different.
Additionally, the "mingw32" project provides a set of link libraries and header files for building C applications against the win32 API. These likewise install just fine on any Unix.
Here's a site I turned up after a quick google with instructions for building the toolchain.
Not really. Building any kind of executable intended for OS "A" while using OS "B" is a process commonly known as cross-compilation. In this partciluar case, you would need a cross-compiler running on Linux, but targetting Windows. I don't know any vendor selling such a product.