Copyright (C) 1999-2001 ABINIT group (XG,DCA)
This file is distributed under the terms of the GNU General Public License, see ~abinit/COPYING or http://www.gnu.org/copyleft/gpl.txt .
For the initials of contributors, see ~abinit/doc/developers/contributors .

Features_v2.3

Description of the major features of the version 2.3
of the ABINIT code.

Content :
0. Related documentation.
1. Available physical properties.
2. Speed and memory.
3. The user's point of view.
4. The programmer's point of view.

---------------------------


0. Related documentation
------------------------

The reader might consult the latest version of the 'context' file
for the description of the ABINIT project and its history.
The latest version of the 'planning' file will give an idea of
future developments.
The different versions of the 'release_notes' files will allow to
see the actual development of the project since version 1.5,
released in August 1998.
Then, there are also the 'new_user_guide' and 'abinis_help' files,
for accurate descriptions of the code and its use.


1. Available physical properties
________________________________

A. Computation of the total energy of an assembly of nuclei and electrons
placed in a repeated cell.

A.1. The computation is done using plane waves and pseudopotentials.

A.2. The total energy computation is done according to the Density Functional
Theory. Most of the important local approximations (LDA) are
available, including the Perdew-Zunger one. Two different local spin density
(LSD) are available, including the Perdew Wang 92, and one due to M. Teter.
The Perdew-Burke-Ernzerhof GGA (spin unpolarized as well as polarized)
is also available.

A.3. Self-consistent calculations will generate the DFT ground-state,
with associated energy and density. Afterwards, a non-self-consistent
calculation might generate eigenenergies at a large number
of k-points, for band structures. The smeared Density-Of-States is available.

A.4. The program admits many different types of pseudopotentials.
There are two complete sets of pseudopotentials available for
the whole periodic table, one of the Troullier-Martins type,
one of the Goedecker type (this one include the spin-orbit coupling).
Four codes are available to generate
new pseudopotentials when needed. Two of them are able to generate
pseudopotentials with a core hole, in order to compute core-level shifts.
Two of them are able to generate GGA pseudopotentials.
No ultra-soft pseudopotential can be used.

A.5. Metallic as well as insulating systems can be treated. Schemes for
determination of the occupation number include the Fermi broadening,
the Gaussian broadening, the Gaussian-Hermite broadening, as well as
the modifications proposed by Marzari.

A.6. The cell may be orthogonal or non-orthogonal.
Any kind of symmetries and corresponding sets of k-point can be input,
and taken into account in the computation.

A.7. The electronic system may be computed in the
spin-unpolarized or spin-polarized
case, with the possibility to impose occupation numbers of majority and
minority spins, and the spins of the starting configuration.

A.8. The total energy and electronic structure can be provided
with the spin-orbit coupling included. Spin-orbit coupling
is not yet available for forces, stresses and response functions.

A.9. The decomposition of energy in its different component (local potential,
XC, hartree ...) is provided.


B. Derivatives of the total energy.

B.1. Hellman-Feynman forces are computed from an analytical formula,
and corresponds exactly to the limit of finite differences of energy
for infinitesimally small atomic displacements when the ground-state
calculation is at convergence. This feature is available for all the
cases where the total energy can be computed, except spin-orbit.
A correction for non-converged
cases allows to get accurate forces with less converged wavefunctions
than without it. The decomposition of the forces in their different
components can be provided.

B.2. Stress can also be computed. This feature is available for all
the cases where the total energy can be computed (except spin-orbit).
A facility for correcting the computed stress by the Pulay stress
is provided. The decomposition of the stresses in their different
components can be provided. A smearing scheme applied to the kinetic
energy allows to get smooth energy curves as a function of
lattice parameters and angles.

B.3. Accurate responses to atomic displacements and
homogeneous electric fields
are available, and allows to compute the interatomic force constants,
the Born effective charges, the dielectric constant, the phonon band
structure. Thermodynamical properties, like the free energy,
the heat capacity or the entropy, can also be computed, in the
quasi-harmonic approximation. Not yet available for spin-polarized,
spin-orbit, or GGA.

B.4. Approximate susceptibility matrix (zero frequency, q=0) and
dielectric matrix can be computed, thanks to a sum over states.

B.5. Excited states of atoms and molecules (spin-singlet as well
as spin-triplet) can be computed within TDDFT.


C. Displacement of atoms, and changes of call parameters.

C.1. Different algorithms (Broyden; modified Broyden; Verlet with sudden
stop of atoms) allows to find the equilibrium configuration
of the nuclei, for which the forces vanish. The cell parameters
can also be optimized concurently with the atomic positions.
Specified lattice parameters, or angles, or atomic positions,
can be kept fixed if needed.

C.2. Two molecular dynamics algorithm (Numerov or Verlet)
allow to perform simulations in real
(simulated) time. The displacement of atoms may be computed according to
Newton's law, or by adding a friction force to it.
Nose-Hoover thermostat is available with Verlet algorithm.

C.3. The code can provide an automatic analysis of bond lengths and angles,
and the atomic coordinates in a format suitable for vizualisation with XMOL.


D. Graphical tools.

D.1. A special part in the tutorial (see later) indicates how
to generate properly formatted data for the visualisation of :
- the band structure (visualisation thanks to XMGR)
- total energies vs different parameters (also using XMGR)
- the charge density (3D isosurfaces)
 (the cut3d postprocessor must be used, followed by matlab)
The cut3d postprocessor also allows to prepare 2D charge density plots.



2. Speed and memory.
____________________

A. Speed in the sequential version

A.1. Depending on the number of atoms, there are two regimes in the
code : at low number of atoms and electrons,
the CPU time is dominated by Fast Fourier
Transforms with an average scaling O(N^2 log N) where N is some
number characteristics of the size of the system (atoms, electrons);
at large number of atoms and electrons, the CPU time is dominated
by non-local operator aplication and orthogonalisation, with
an average scaling O(N^3).

A.2. The complex-to-complex Fast Fourier Transform routine for application
of the Hamiltonian has been highly optimized, and take into account
the fact that the wavefunction do not fill the reciprocal space FFT box.
Library FFTs are also available, but they are found to be slower than the
present FFT routine, developed starting from a routine
provided by S. Goedecker.
A real-to-complex FFT is used for treating potential and densities
of the ground state, since they are real.
For selected k-points, invariant under time-reversal symmetry,
( (0 0 0), (1/2 0 0), (0 1/2 0), (0 0 1/2), (1/2 1/2 0) ... (1/2 1/2 1/2) ),
the number of planewave explicitly treated is divided by two. A
real-to-complex FFT is used then.

A.3. The non-local potential is applied in reciprocal space. It has been
optimized carefully, although there is still some speed-up to be
coded when the k-point is invariant under time-reversal.
The orthogonalisation procedure can be done twice per loop or only once.

A.4. At the level of the generation of electronic eigenfunctions, an efficient
band-by-band preconditioned conjugate-gradient algorithm is used,
in its non-self-consistent version.

A.5. At the level of the self-consistency loop, an efficient
potential-based preconditioned conjugate-gradient algorithm is used.
Simple mixing is also available, as well as the Anderson algorithm.
Preconditioning of this algorithm is achieved through a model
dielectric function, or through an approximate dielectric matrix.


B. Speed in the parallel version

B.1. For ground-state calculations, the code has been parallelized
on the k-points, and on the FFT grid and plane wave coefficients.
For the k-point parallelisation (using MPI), the communication
load is generally
very small. This allows it to be used on a cluster of workstations.
However, the number of nodes that can be used in parallel might
be small, and depends strongly on the physics of the problem.
The FFT grid parallelisation (using OpenMP) works only
for SMP machines, and is still to be
optimized.

B.2. For response calculations, the code has been parallelized
on both k-points and bands, as well as
on the FFT grid and plane wave coefficients.
For the k-points and bands parallelisation,
the communication load is rather
small also, and, unlike for the GS calculations, the number
of nodes that can be used in parallel will be large,
nearly independently of the physics of the problem.
The FFT grid parallelisation (using OpenMP) works only
for SMP machines, and is still to be
optimized.

B.3. A careful study of the speed-up should still be done,
in both the GS and RF cases.


C. Memory.

C.1. The requirements of the different conjugate gradient algorithms
on memory are relatively low, especially when the number of atoms
is large. Optionally, it is even possible to use disk space
to save memory, at the expense of computing time.
In particular, when the number of k points is large, they can
be stored in memory one at a time. Phase factors in the application of
the non-local operator can also be recomputed at each application,
in order to save memory. For k-points that are invariant under
time-reversal symmetry, the storage required for wavefunctions
is half the storage for other k-point.


3. The user's point of view.
____________________________

A. The Web site.

A.1. A Web site can be accessed. The complete sources (and all the tests)
of the ABINIT package are available there.
Executables for many different platforms are also available,
in specific packages that also include
the 'Infos' directory are also available.
Installation notes, current features of ABINIT,
the tutorial, on-line help can be vizualized directly from the web.

A.2. Also available from the Web site :
- the pseudopotentials
- some utilities (including cut3d, a density analyser),
- three mailing lists
(one for the developpers, one for the 'normal' users, one
for the advisory commitee).
- the ABINIT bibliography database, that contains references of papers
in which ABINIT or one of its predecessors have been used.


B. Portability.

B.1. The ABINIT package has been installed successfully on the
following different platforms :

- IBM RS6000 (models : 590, 3CT, nighthawk) based on Power 2 and 3+ processors.
- PC/Linux based on PPro, PII or PIII processors,
 pghpf compiler or fujitsu compiler.
- HP/SPP1600, HP/S-class, HP/N-class based on the HP 7200, 8000
 and 8500 processors.
- DEC alpha workstations under OSF, based on EV56, EV6 or EV67.
- SGI Origin 2000
- CRAY T3E
- FUJITSU VPP-700
- Sun ultrasparc II
- NEC
- PC under Windows
- Mac

B.2. In particular, the parallel version is available on clusters of
Intel/Linux, DEC or IBM workstations, as well as on
CRAY T3E, SGI Origin 2000, HP/SPP1600, HP/S-class, HP/N-class,
FUJITSU VPP-700 machines.

B.2. Installation is made thanks to a sophisticated (but robust) suite of
makefiles and scripts, and use a file preprocessor.
Thanks to these, all machine-dependent parameters
are grouped in one single short file for each machine.
The parallel and sequential
version of the code, as well as the different versions for the different
machines, are prepared on-the-fly, by this suite of makefiles and scripts,
so that there is only one unique source code.

B.3. Binaries for different machines can be managed in the
same main directory, as they might be placed automatically in different
sub-directories.


C. Running jobs, input and output files.

C.1. The input variables are gathered in one unique file,
read by a text processing facility build in the code.
Many defaults values are provided, so that
the input file can be kept rather short.

C.2. Many different stopping criteria allow the user to target the accuracy
he or she wants to obtain.

C.3. The outputs are provided to one main file and one auxiliary file,
as well as different specialized files (for density, potential,
wavefunctions, ...) .
The main file is shorter than the auxiliary file, and well formatted,
while all important results are gathered
there. It can be used for archival purposes. The auxiliary (log) file
will contain all exception messages.

C.4. Exception handling is provided through four different types
of messages : COMMENT, WARNING, BUG and ERROR. In each case, the
accurate meaning of the exception is described, as well as the
eventual action to be taken by the user.

C.5. Statistics of foreseen memory and disk usage is printed at
the beginning of the run. Statistics of CPU time usage
is printed at the end of the run.
(These features must still be modified for RF case)

C.6. There is a facility to stop the run in a clean way at any time.
The user may specify a cpu time limit, after which the job must end
smoothly.

C.7. A status file, updated very frequently, gives an on-the-fly
report of progress of the current run.

C.8. The code can handle multiple datasets contained in the input
file, where generic input variables valid for all datasets
can be defined. These calculations for different dataset
can be chained, so that in one run, many complex tasks can be accomplished.
This allows easy convergence studies.



D. Documentation.

D.1. A new_user_guide and a few rather detailed help files
(abinis_help, respfn_help, ifc_help, merge_help, newsp_help) are available.

D.2. A tutorial is available. It starts with the computation of
different properties of the H2 molecule, describes convergence studies,
then focuses on bulk Silicon, and ends with the study of Al surface.

D.3. Many test cases are provided, and can help the user in setting up a run.

D.4. An ABINIT bibliography database, that contains references of papers
in which ABINIT or one of its predecessors is available on the Web site.


E. Generation of the k-points, geometries, and starting wavefunctions

E.1. The code can automatically generate symmetries from the primitive cell
and the position of atoms. In this case, it identifies
automatically the Bravais lattice and point group.
Alternatively, it can start from the
symmetries and generate the atomic positions from the irreducible
set. Also, a database of the 230 spatial groups of symmetry
is built inside ABINIT.
The generation of special k point sets (Monkhorst-Pack sets)
and band structure k points can also be done directly inside ABINIT.
A list of interesting k point sets, can be generated automatically,
including a measure of their accuracy in term of integration
within the Brillouin Zone.

E.2. A geometry builder is available inside the code. It can take a group
of atoms, rotate it, translate it and repeat it, then create vacancies.

E.3. A utility for generating wavefunctions with new characteristic
(cut-off, k-point) from already existing wavefunctions with
different characteristics is available (newsp).


F. Automatic determination of input parameters.

F.1. Many defaults are provided.

F.2. The FFT grid parameters can be automatically generated
from the cut-off energy and geometry of the system.

F.3. The number of bands and starting occupation numbers can be automatically
generated from the input set of atoms.

F.4. There is a database of atomic masses.



4. The programmer's point of view.
__________________________________

A. The code will distributed without charge under the
GNU General Public Licence (GPL). This
will garantee that the future modifications of the code
stay available to the developpers and users for free.
This Copyright is often referred to as a "Copyleft".

B. The code is written in clean Fortran90. Strict programming rules
have been followed. These are documented. Comments are numerous, and
all in english.

C. Quick, fully automatic, testing of the code is available in the makefile,
giving diagnostics on the validity of computed energy, forces, stresses
and eigenvalues for five typical cases.

D. More extensive testing is provided in five batteries of tests
(altogether more than 150 different runs), with automatic comparison
with results of preceding versions. A specialized diff script (called
'fldiff') has been written in order to ease the diagnostic on the
suite of tests. In addition to tests of the correctness of the execution,
for the sequential or parallel version of the main code, as well as some
utilities, there are automatic diagnostics of the speed of crucial
routines, and the response to a load of up to 4 instances of the
main code, running concurrently.

E. Debugging facilities are provided inside the code, and can be directly
accessed from the input file. The compilation can be done in both
debugging or normal mode : the C-preprocessed files are either kept
or removed automatically.
<\pre>