NVIDIA CUDA (sequel) - PyCUDA Setup

Continuing the theme Configuring Ubuntu 8.10 for coding under NVIDIA CUDA , I decided to fix myself a workplace for using one of my favorite programming languages ​​- Python. To do this, there is an excellent set of PyCUDA libraries that allows you to implement all the delights of the NVIDIA CUDA API using a convenient and beautiful code. I will give an example of code from standard PyCUDA configuration:

import pycuda.autoinit
import pycuda.driver as drv
import numpy

mod = drv.SourceModule("""
__global__ void multiply_them(float *dest, float *a, float *b)
{
const int i = threadIdx.x;
dest[i] = a[i] * b[i];
}
""")

multiply_them = mod.get_function("multiply_them")

a = numpy.random.randn(400).astype(numpy.float32)
b = numpy.random.randn(400).astype(numpy.float32)

dest = numpy.zeros_like(a)
multiply_them(
drv.Out(dest), drv.In(a), drv.In(b),
block=(400,1,1))

print dest-a*b

Isn't it convenient and simple?

Let's start!

For normal PyCUDA operation, we will need:

- A machine with a UNIX-like system and access to the Internet.
- Nvidia CUDA toolkit. PyCuda was developed with an eye on version 2.0 beta, but for me it seems to work just fine with 2.1 beta.
- C ++ compiler, preferably gcc version not lower than 4.x.
- Installed Python interpreter, version 2.4 or newer.
')
Now install the latest version of the Boost C ++ library . Before installing it is worth checking whether the developer package for Python is installed, it will be called python-dev or python-devel . The installation is actually standard, just don’t forget to set the prefix - the directory for the collected inclusions. In this case, let it be $ HOME / pool .

$ tar xfj ~/downloads/boost_1_37_0.tar.bz2
$ cd boost_1_37_0
$ ./configure --prefix=$HOME/pool
$ make
$ make install

After the build (it should be successful) we check if the $ HOME / pool folder has appeared. If yes - then all the way, go ahead :)

Important note: If something like this appears at the end of the installation
...failed updating 30 targets...
...skipped 2 targets...
then check again if you have Python developer headers installed. If it was not possible to update only a few files (less than 5), then most likely everything will work fine, but for fidelity you can install libz-dev and libbz2-dev and build Boost again.

Now add the path to the newly compiled libraries to the environment variable:
$ export LD_LIBRARY_PATH=$HOME/pool/lib:${LD_LIBRARY_PATH}

If this is not done, then at the PyCUDA build stage, something like this will appear:
...blablabla...
ImportError: libboost_python-gcc42-mt-1_35.so.1.35.0:
cannot open shared object file: No such file or directory
...blablabla...

We download PyCUDA itself and go to the directory with it.
$ cd $HOME/downloads/pycuda-0.91.1

You need to install the numpy package. You can do this either by downloading it yourself from numpy.org and building it, or by using PyCUDA itself. Since we have already moved to the directory with the distribution, just type:
$ sudo python ez_setup.py // setup-tools
$ sudo easy_install numpy // numpy setup-tools

Order.

We collect PyCUDA.

It remains the last and most important thing - to install PyCUDA itself. The main difficulty lies in the fact that you need to specify a bunch of options during configuration.
$ ./configure \
--boost-inc-dir=$HOME/pool/include/boost-1_37 \
--boost-lib-dir=$HOME/pool/lib \
--boost-python-libname=boost_python-gcc41-mt \
--cuda-root=/where/ever/you/installed/cuda

And now - attention! If everything is clear with the first two parameters (explained above), then the last two need to be explained. boost_python-gcc41-mt is the name of the library in which gcc41 is the version of the compiler with which you compiled the boost . As I wrote in the previous article, when programming under CUDA, version 4.1 should usually be used, but if you did not set it to system default, then this is irrelevant in this case. You can check the version of the active compiler by typing
$ gcc -v

Well, in the last paragraph, as it is easy to guess, you need to replace / where / ever / you / installed / cuda with the directory with CUDA, that is, by default / usr / local / cuda .

By the way, before entering make install, you can double-check the entered addresses. They are all stored in the siteconf.py file, which is created during configuration. He looked like this before my build:

BOOST_INC_DIR = ['/ home / username / pool / include / boost-1_37']
BOOST_LIB_DIR = ['/ home / username / pool / lib']
BOOST_PYTHON_LIBNAME = ['boost_python-gcc41-mt']
CUDA_ROOT = '/ usr / local / cuda'
CUDADRV_LIB_DIR = []
CUDADRV_LIBNAME = ['cuda']
CXXFLAGS = []
LDFLAGS = []

If you want to reconfigure PyCUDA with new parameters, first delete this file. Or you can just make changes to it and type
$ sudo python setup.py build

Now you can safely compile:
$ sudo make install

If everything is done correctly, then there should be no errors during the assembly. Restart the computer, and for complete confidence you can go to the directory
/pycuda-0.91.1/test (or whatever version you have there) and run test_driver.py .

$ cd /pycuda-0.91.1/test
$ python test_driver.py

If the result is OK, then everything is in chocolate. If one or two errors (for some reason, I sometimes had glitches with memory) - just try a little later again.

Thank you for reading to the end, I will be glad to hear criticism :)

ADF : Everything was set up, including work in Eclipse. The problem was that the wrong rights to the cache folder were set up, the easiest way to do this:
$ sudo chmod 0777 -R /home/username/.pycuda-compiler-cache

Well, among other things, neither the interpreter nor the Eclipse itself should be launched from under the root. For other environment variables :) Thanks for the help in setting up a friend of Riz .