Please add cl-cuda

[takagi]

Cl-cuda is a library to use NVIDIA CUDA in Common Lisp programs.

https://github.com/takagi/cl-cuda

It depends on NVIDIA CUDA which is, you know, a parallel computing platform using GPUs.

[quicklisp]

Unfortunately, I don't have a system on which I can test this library. In addition, I'm reluctant to add projects that work only on a single implementation.

[takagi]

Unfortunately, I don't have a system on which I can test this library.

I guess that cl-gpu which is already included in quicklisp also needs CUDA enabled GPU as cl-cuda. Why is cl-gpu accepted?

In addition, I'm reluctant to add projects that work only on a single implementation.

Ok, I will make cl-cuda work on various implementations. The implementation-dependent thing is only a point executing a external program.

Thanks.

[quicklisp]

The cuda portion of cl-gpu is omitted. There are other systems in cl-gpu that do not require cuda.

[takagi]

I see. Then, if I would provide you a kind of CUDA environment, for example a CUDA enabled graphics card, could you accept cl-cuda?

[quicklisp]

I don't have a good way to merge that kind of info yet. I'll try to make it work better soon.

[takagi]

Ok. Please let me know after you will make it work!

[takagi]

Sorry for bothering you again. I want to ask you to add cl-cuda once more because I found this IRC log and guess you think a physical board is required for just building it.

http://ircbrowse.net/browse/lisp?events_page=244537

For just building cl-cuda, only CUDA SDK is required to be installed and no physical board or built-in GPU chip is needed. CUDA SDK can be downloaded from NVIDIA's website and is provided for Windows, Mac OSX and Linux.

Following shows a relation of some library operations and CUDA specific stuffs.

Ops.	CUDA SDK	a physical board
build	required	not
test	required	required
run	required	required

Here, build is (:ql :cl-cuda) and test is (:ql :cl-cuda-test). Run is a process of defining a kernel function, creating a CUDA context then launching the kernel function in the context. Since test, :cl-cuda-test, also includes this process, it requires a physical board.

Isn't it enough for accepting cl-cuda?

If not enough and requiring CUDA SDK on build is the problem, I can make cl-cuda not requiring it and compiled without any conditions raised on build, although I don't know it makes sense because this is almost the same case as any other CL systems that depend on other non-CL libraries, especially language bindings.

On the other hand, if the problem is requiring a physical board on test, :cl-cuda-test, I will think about another solution. For example I would host testing environment for myself and report its result to quicklisp, if possible.

Now, Common Lisp libraries not listed on Quicklisp is almost same as it does not exist. And cl-cuda's users strongly hope it accepted on Quicklisp.

[quicklisp]

Sorry for the delay.

I tried to get the SDK some time ago, but it did not run on my Debian system. I don't remember the reason. I will try to get it again and see if that helps me build cl-cuda.

[quicklisp]

I tried again. The deb approach seems to require a newer libc than Debian 7 provides.

Do you know how I can get it installed on Debian 7?

[takagi]

What is your gcc's version?

[quicklisp]

$ gcc --version
gcc (Debian 4.7.2-5) 4.7.2

[takagi]

gcc 4.6 might help you.

# apt-get install g++-4.6
# update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-4.6 46 --slave /usr/bin/g++ g++ /usr/bin/g++-4.6
# update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-4.7 47 --slave /usr/bin/g++ g++ /usr/bin/g++-4.7
# update-alternatives --config gcc

[takagi]

If not, I will also research about it.

[melisgl]

On debian testing (what will become debian 8) I simply install nvidia-cuda-toolkit.

[quicklisp]

@melisgl When I try that on testing, I get:

Package nvidia-cuda-toolkit is not available, but is referred to by another package.

Did I do something wrong?

[melisgl]

@quicklisp I think you need to add the non-free section to apt sources so it looks like this:

deb http://ftp.fi.debian.org/debian/ testing main non-free contrib

[quicklisp]

Thanks! I have made some progress.

@takagi Can you please add the :author, :license, and :description options to the systems of cl-cuda?

[takagi]

@quicklisp Yes. I will add the sections in a few days.

[takagi]

@quicklisp Now the systems of cl-cuda, cl-cuda and cl-cuda-interop, have :author, :license and :description options.

[mpsota]

@quicklisp Any progress on this ticket? I can confirm that one can build cl-cuda without CUDA enabled graphics card. All I needed under ubuntu is nvidia-cuda-toolkit.

[quicklisp]

Cool, it works now, and will be in the next release.