coppolachan / rhmc-on-gpu

Rational Hybrid Monte Carlo simulation of QCD with staggered fermions on GPU

Rational Hybrid Montecarlo with non improved staggered fermions

---

Developers and mantainers
Claudio Bonati (C++ framework, CUDA kernels) ([email protected])
Guido Cossu (CUDA kernels) ([email protected])

History:
Version 0.1 9/2011 CUDA kernels

---

1. Features

Performs Rational Hybrid Monte Carlo simulation of QCD with N_f flavors of staggered (non improved) fermions. Except some minor functions, all the calculations are performed using NVIDIA gpus, implementing a CUDA version of all the relevant kernels (Dirac operator, Dirac operator inversion, gauge and momenta update).

See the following reference for more details:
'QCD simulations with staggered fermions on GPUs'
Arxiv Paper 1106.5673

2. Compilation

The typical compilation follows the pattern:

./configure && make

this will compile a code to run in a single core CPU.

3. Enabling CUDA routines

In order to enable CUDA routines type the following command (also look at the various options in configure --help)

./configure --enable-cuda=ARCH

where ARCH should be substituted by one of the following strings "sm_XX" in order to choose the target NVIDIA architecture

• sm_10 ISA_1, Basic features
• sm_11 atomic memory operations on global memory
• sm_12 atomic memory operations on shared memory, vote instructions
• sm_13 double precision floating point support (Tesla Cards)
• sm_20 Fermi architecture support

please refer to the original NVIDIA documentation for more details on which cards support the several architectures.

4. Issues

Some newer compilers, like GNU gcc 4.5, are incompatible with the NVIDIA nvcc compiler (CUDA 4.0). In order to solve this problem you should install a different version of the compiler, e.g. gcc 4.4, in the directory PATH_TO_COMPILER and then configure with the following command

./configure --enable-cuda=ARCH --with-cuda_comp=PATH_TO_COMPILER CXX=PATH_TO_COMPILER/g++ CC=PATH_TO_COMPILER/gcc

Also, if the mandatory library MPFR (for Multiple Precision in rational approximations) is installed in a non standard location (not /usr/lib) use the flag --with-mpfr=PATH where PATH is the location of the library in your system

5. Known bugs

• Some older versions of autoreconf fail to build a working configure file. Usually the one shipped is working.

If you find any bug please report any to the developers on GitHub issues:

RHMC on GPU issues