lammps/lib/gpu

/* ----------------------------------------------------------------------
   LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
   http://lammps.sandia.gov, Sandia National Laboratories
   Steve Plimpton, sjplimp@sandia.gov

   Copyright (2003) Sandia Corporation.  Under the terms of Contract
   DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
   certain rights in this software.  This software is distributed under 
   the GNU General Public License.

   See the README file in the top-level LAMMPS directory.
------------------------------------------------------------------------- */

/* ----------------------------------------------------------------------
   Contributing authors: Mike Brown (SNL), wmbrown@sandia.gov
                         Peng Wang (Nvidia), penwang@nvidia.com
                         Paul Crozier (SNL), pscrozi@sandia.gov
------------------------------------------------------------------------- */

                          GENERAL NOTES
                          
This library, libgpu.a, provides routines for GPU acceleration
of LAMMPS pair styles.  Currently, only CUDA enabled GPUs are
supported.  Compilation of this library requires installing the CUDA
GPU driver and CUDA toolkit for your operating system. In addition to
the LAMMPS library, the binary nvc_get_devices will also be
built. This can be used to query the names and properties of GPU
devices on your system.

NOTE: Installation of the CUDA SDK is not required.

Current pair styles supporting GPU acceleration:

  1. lj/cut/gpu
  2. gayberne/gpu

                     MULTIPLE LAMMPS PROCESSES
                     
When using GPU acceleration, you are restricted to one physical GPU
per LAMMPS process. This can be multiple GPUs on a single node or
across multiple nodes. Intructions on GPU assignment can be found in
the LAMMPS documentation.

                            SPEEDUPS

The speedups that can be obtained using this library are highly
dependent on the GPU architecture and the computational expense of the
pair potential. When comparing a single precision Tesla C1060 run to a
serial Intel Xeon 5140 2.33 GHz serial run, the speedup is ~4.42x for
lj/cut with a cutoff of 2.5. For gayberne with a cutoff of 7, the
speedup is >103x for 8000 particles. The speedup will improve with an
increase in the number of particles or an increase in the cutoff.

                    BUILDING AND PRECISION MODES

To build, edit the CUDA_CPP, CUDA_ARCH, CUDA_PREC, and CUDA_LINK files for
your machine. Type make. Additionally, the GPU package must be installed and
compiled for LAMMPS. The library supports 3 precision modes as determined by 
the CUDA_PREC variable:

  CUDA_PREC = -D_SINGLE_SINGLE  # Single precision for all calculations
  CUDA_PREC = -D_DOUBLE_DOUBLE  # Double precision for all calculations
  CUDA_PREC = -D_SINGLE_DOUBLE  # Accumulation of forces, etc. in double

NOTE: For the lj/cut pair style, only single precision will be used, even 
      if double precision is specified.
  
NOTE: Double precision is only supported on certain GPUS (with
      compute capability>=1.3).
      
NOTE: For Tesla and other graphics cards with compute capability>=1.3,
      make sure that -arch=sm_13 is set on the CUDA_ARCH line.

NOTE: The gayberne/gpu pair style will only be installed if the ASPHERE
      package has been installed before installing the GPU package in LAMMPS.
      
                           GPU MEMORY

Upon initialization of the gayberne/gpu pair style, the library will reserve
memory for 64K atoms per GPU or 70% of each cards GPU memory, whichever value
is limiting.  The value of 70% can be changed by editing the
PERCENT_GPU_MEMORY definition in the source file. For gayberne/gpu, the value
of 64K cannot be increased and is the maximum number of atoms allowed per
GPU. Using the 'neigh_modify one' modifier in your LAMMPS input script
can help to increase maximum number of atoms per GPU for cards with
limited memory.