Release Notes for QUDA v0.5.1                           xx May 2013
-----------------------------

Overview:

QUDA is a library for performing calculations in lattice QCD on
graphics processing units (GPUs), leveraging NVIDIA's CUDA platform.
The current release includes optimized Dirac operators and solvers for
the following fermion actions:

* Wilson
* Clover-improved Wilson
* Twisted mass (including non-degenerate pairs)
* Improved staggered (asqtad or HISQ)
* Domain wall

Implementations of CG, multi-shift CG, BiCGstab, and DD-preconditioned
GCR are provided, including robust mixed-precision variants supporting
combinations of double, single, and half (16-bit "block floating
point") precision.  The library also includes routines for HISQ link
fattening and force terms for the HISQ fermion action and one-loop
improved Symanzik gauge action.  Use of many GPUs in parallel is
supported throughout, with communication handled by QMP or MPI.


Software Compatibility:

The library has been tested under Linux (CentOS 5.8 and Ubuntu 12.04)
using release 4.0, 4.1, 4.2, and 5.0 of the CUDA toolkit.  CUDA 3.x
and earlier are not supported.  The library also works under Mac OS X
10.6.8 ("Snow Leopard") and 10.7.3 ("Lion") on recent 64-bit
Intel-based Macs.

See also "Known Issues" below.


Hardware Compatibility:

For a list of supported devices, see

http://developer.nvidia.com/cuda-gpus

Before building the library, you should determine the "compute
capability" of your card, either from NVIDIA's documentation or by
running the deviceQuery example in the CUDA SDK, and pass the
appropriate value to QUDA's configure script.  For example, the Tesla
C2075 is listed on the above website as having compute capability 2.0,
and so to configure the library for this card, you'd run "configure
--enable-gpu-arch=sm_20 [other options]" before typing "make".

As of QUDA 0.4.0, only devices of compute capability 1.1 or greater
are supported.  See also "Known Issues" below.


Installation:

Installing the library involves running "configure" followed by
"make".  See "./configure --help" for a list of configure options.
At a minimum, you'll probably want to set the GPU architecture; see
"Hardware Compatibility" above.

Enabling multi-GPU support requires passing the --enable-multi-gpu
flag to configure, as well as --with-mpi=<PATH> and optionally
--with-qmp=<PATH>.  If the latter is given, QUDA will use QMP for
communications; otherwise, MPI will be called directly.  By default,
it is assumed that the MPI compiler wrappers are <MPI_PATH>/bin/mpicc
and <MPI_PATH>/bin/mpicxx for C and C++, respectively.  These choices
may be overriden by setting the CC and CXX variables on the command
line as follows:

./configure --enable-multi-gpu --with-mpi=<MPI_PATH> \
[--with-qmp=<QMP_PATH>] [OTHER_OPTIONS] CC=my_mpicc CXX=my_mpicxx

Finally, with some MPI implementations, executables compiled against
MPI will not run without "mpirun".  This has the side effect of
causing the configure script to believe that the compiler is failing
to produce a valid executable.  To skip these checks, one can trick
configure into thinking that it's cross-compiling by setting the
--build=none and --host=<HOST> flags.  For the latter,
"--host=x86_64-linux-gnu" should work on a 64-bit linux system.

By default only the QDP and MILC interfaces are enabled.  For
interfacing support with QDPJIT, BQCD or CPS; this should be enabled
at configure time with the appropriate flag, e.g.,
--enable-bqcd-interface.  To keep compilation time to a minimum it is
recommended to only enable those interfaces that are used by a given
application.  The QDP and MILC interfaces can be disabled with the,
e.g., --disable-milc-interface flag.

If Fortran interface support is desired, the F90 environment variable
should be set when configure is invoked, and "make fortran" must be
run explicitly, since the Fortran interface modules are not built by
default.

As examples, the scripts "configure.milc.titan" and
"configure.chroma.titan" are provided.  These configure QUDA for
expected use with MILC and Chroma, respectively, on Titan (the Tesla
K20X-powered Cray XK7 supercomputer at the Oak Ridge Leadership
Computing Facility).

Throughout the library, auto-tuning is used to select optimal launch
parameters for most performance-critical kernels.  This tuning
process takes some time and will generally slow things down the first
time a given kernel is called during a run.  To avoid this one-time
overhead in subsequent runs (using the same action, solver, lattice
volume, etc.), the optimal parameters are cached to disk.  For this
to work, the QUDA_RESOURCE_PATH environment variable must be set,
pointing to a writeable directory.  Note that since the tuned parameters
are hardware-specific, this "resource directory" should not be shared
between jobs running on different systems (e.g., two clusters
with different GPUs installed).  Attempting to use parameters tuned
for one card on a different card may lead to unexpected errors.


Using the Library:

Include the header file include/quda.h in your application, link
against lib/libquda.a, and study tests/invert_test.cpp (for Wilson,
clover, twisted-mass, or domain wall fermions) or
tests/staggered_invert_test.cpp (for asqtad/HISQ fermions) for
examples of the solver interface.  The various solver options are
enumerated in include/enum_quda.h.


Known Issues:

* Issues affecting devices with compute capability 1.1, 1.2, or 1.3
  (i.e., those predating the "Fermi" architecture):

  1. Building for these targets results in a large number of warnings
     that the compiler "Cannot tell what pointer points to, assuming
     global memory space."  The warnings are annoying but harmless.

  2. Computation of the Fermilab heavy-quark residual (and its use as
     a stopping condition in the solvers) is not supported on these
     devices.  Selecting it will result in a run-time error.

  3. It is anticipated that QUDA 0.6.0 will drop support for pre-Fermi
     cards completely.  QUDA 0.5.x will continue to be maintained as a
     separate branch (with bug fixes but minimal new features) for
     those needing this support.

* When used with drivers predating the CUDA 5.0 production release,
  Fermi-based GeForce cards suffer from occasional hangs when reading
  from double-precision textures, recoverable only with a soft reset.
  The recommended solution is to upgrade to the latest driver.
  Alternatively, such texture reads can be disabled (at the
  expense of some performance) by passing the
  --disable-fermi-double-tex flag to configure.

* For compatibility with CUDA, on 32-bit platforms the library is
  compiled with the GCC option -malign-double.  This differs from the
  GCC default and may affect the alignment of various structures,
  notably those of type QudaGaugeParam and QudaInvertParam, defined in
  quda.h.  Therefore, any code to be linked against QUDA should also
  be compiled with this option.

* The auto-tuner reports "0 Gflop/s" and "0 GB/s" for several of the
  Dslash kernels (visible if the verbosity is set to at least
  QUDA_SUMMARIZE), rather than the correct values.  This does not
  affect the tuning process or actual performance.

* Attempts to compile with GCC 4.7.x will fail due to bugs in the
  compiler.


Getting Help:

Please visit http://lattice.github.com/quda for contact information.
Bug reports are especially welcome.


Acknowledging QUDA:

If you find this software useful in your work, please cite:

M. A. Clark, R. Babich, K. Barros, R. Brower, and C. Rebbi, "Solving
Lattice QCD systems of equations using mixed precision solvers on GPUs,"
Comput. Phys. Commun. 181, 1517 (2010) [arXiv:0911.3191 [hep-lat]].

When taking advantage of multi-GPU support, please also cite:

R. Babich, M. A. Clark, B. Joo, G. Shi, R. C. Brower, and S. Gottlieb,
"Scaling lattice QCD beyond 100 GPUs," International Conference for
High Performance Computing, Networking, Storage and Analysis (SC),
2011 [arXiv:1109.2935 [hep-lat]].

Several other papers that might be of interest are listed at
http://lattice.github.com/quda .


Authors:

Ronald Babich (NVIDIA)
Kipton Barros (Los Alamos National Laboratory)
Richard Brower (Boston University)
Mike Clark (NVIDIA)
Justin Foley (University of Utah)
Joel Giedt (Rensselaer Polytechnic Institute)
Steven Gottlieb (Indiana University)
Balint Joo (Jefferson Laboratory)
Claudio Rebbi (Boston University)
Guochun Shi (NCSA)
Alexei Strelchenko (Fermi National Accelerator Laboratory)

Portions of this software were developed at the Innovative Systems Lab,
National Center for Supercomputing Applications
http://www.ncsa.uiuc.edu/AboutUs/Directorates/ISL.html

Development was supported in part by the U.S. Department of Energy
under grants DE-FC02-06ER41440, DE-FC02-06ER41449, and
DE-AC05-06OR23177; the National Science Foundation under grants
DGE-0221680, PHY-0427646, PHY-0835713, OCI-0946441, and OCI-1060067;
as well as the PRACE project funded in part by the EUs 7th Framework
Programme (FP7/2007-2013) under grants RI-211528 and FP7-261557.  Any
opinions, findings, and conclusions or recommendations expressed in
this material are those of the authors and do not necessarily reflect
the views of the Department of Energy, the National Science
Foundation, or the PRACE project.