*Installation Notes and Requirements:*

To use the *cudaBayesreg* package on your system you need the following
installed:

   1.

      CUDA-capable GPU

   2.

      CUDA software, which includes the *nvcc* (release 3.0-beta1)
      NVIDIA Cuda Compiler driver (available at no cost from
      http://www.nvidia.com/cuda)

   3.

      A supported version of Linux with g++ GNU compiler; (gcc/g++
      version 3.4 or 4.3 is required, since nvcc is currently not fully
      compatible with gcc v. 4.4).

*Notes*:

   1.

      A general purpose C compiler is needed by nvcc. On Linux
      platforms, the compiler is assumed to be ‘gcc’ , or ‘g++’ for
      linking. In order to use *gcc version 3.4,* for instance, rename
      it gcc34, and make a link in directory “*/tmp*” to it, e.g. “ln -s
      /usr/bin/gcc34 gcc”. The package's Makefiles expects this
      configuration to be in place (it uses the nvcc flag “*-ccbin /tmp*”).

   2.

      In some Linux distributions (e.g. Fedora), to install “gcc34” and
      “g++34” you just need to install the GNU compatibility packages
      “compat-gcc-34” and “compat-gcc-34-c++”.

   3.

      To reconfigure the Makefiles with a different “-ccbin” flag, edit
      the “configure.ac” file and run “autoconf; configure” in the shell.

   4.

      The package's host C++ code uses the shared version of the newmat
      matrix library (http://www.robertnz.net/cpp_site.html). The
      Makefile builds *libnewmat.so* and install it in src/newmat11.

   5.

      In *cudaMultireg.cu* three constants control the amount of CUDA
      allocated space: REGDIM=4096, OBSDIM=128, and XLIM=5. These
      constants define the maximum number of parallel regressions, the
      maximum length of voxel time-series, and the maximum number of
      regression variables, respectively. If these values do not suit
      your needs modify the values, and recompile the code.







