abinit/doc/developers/strain_nonlocal

	This directory contains documentation of the development of the
routines which implement the nonlocal pseudopotential operations in
response-function (RF) strain calculations.  These were written by D.
R. Hamann in 2002-2003, and the were first included in the 4.2 release.
While conceptually straightforward, the expressions to be evaluated are
highly complex, and the only practical way to develop the code was to
use a symbolic-manipulation program and automated post-processing.
These tools are collected here, and their structure and usage are
explained as an aid to future developers who embark on related
projects.


*********************** INTRODUCTION *********************************

	The underlying theoretical approach to treating strain within
density-functional perturbation theory is presented in strainpert.pdf,
which is a preprint that has been submitted to Phys. Rev. B.  It is
also posted on arXiv.org as cond-mat/0409269.  Before proceeding
further, it would be well to read this at least through Sec. III-E.
The equation numbers below refer to this preprint.

	The application of the nonlocal psp's in ground-state (GS)
calculations to find wavefunctions, energies, forces, and stresses was
developed by D. C. Allan.  He introduced the tensor formulation, Eq.
(51), and routines implementing Eq. (52) for nonlocal wave-function
operations, and the ground-state expectation values of Eq. (54) for
stresses.  His approach was extended to RF atomic-displacement
calculations by Allan and X. Gonze.  The necessary expressions were
derived and coded by hand in routines metcon, cont13, cont22, cont24,
etc. all in directory src/03nonlocal.  These are all called by nonlop,
which is used in both GS and RF calculations.  The new RF strain
routines were designed to parallel these existing routines as closely
as possible in their operations, arguments, and structures within which
they are called.  The new "machine-coded" routines are the following:

To evaluate the action of the first strain derivatives of the psp's on
wave functions, Eq. (54):

	metstr.F90

To evaluate the zero-order wave function expectation values of strain-
strain second derivatives, Eq. (55):

	contstr21.F90
	contstr22.F90
	contstr23.F90
	contstr24.F90
	contstr25a.F90
	contstr25.F90
	contstr26.F90

To evaluate the zero-order wave function expectation values of mixed
strain -- atomic-displacement second derivatives ("internal strain"),
Eq. (56):

	contistr01.F90
	contistr03.F90
	contistr10.F90
	contistr12.F90
	contistr21.F90
	contistr30.F90

	As discussed in Sec. III-E, the derivative nonlocal operators
couple "right (|K>)" and "left (<K'|)" tensors of different ranks.  The
"active" arguments aa and bb of all these routines are complex one-
dimensional arrays indexed by a dictionary order of the Eq. (51)
tensors, which have been summed over the wave-function c coefficients
in Eq. (33) with appropriate f form factors and their derivatives as in
Eqs. (48), (54-56).  In metstr, the bb argument is an output array of
the same structure to be used in forming the argument "vectout" of
nonlop.  Retaining the established calling structure, separate calling
statements are invoked for each particular right-to-left change in
tensor rank.  (These are noted in the comments in nonlop.)  For metstr,
the argument "iterm" determines the rank coupling.  However, the
second-derivative expressions are so complicated that each different
rank coupling has been coded in a separate routine in the contstr*,
contistr* series.


********************** MATHEMATICA PROGRAMS **************************

	The challenging task in creating these routines was to find the C
coefficients in Eqs. (54-56).  For each set of subscripts indicating a
particular pair of left and right tensor components, C is a different
polynomial in certain subsets of the 3X3 metric-tensor and metric-
tensor-derivative elements.  While hand computation of the many
thousands of required polynomials is essentially impossible, the
symbolic manipulation capabilities of Mathematica (henceforth "M") in
fact enable this to be carried out in just a few lines of "core" M
code.  The following four M programs are included in this directory:

	make_metstr3.m
	make_met2str.m
	make_metintstr.m
	make_metintstrv.m

These routines are all very similar.  I will discuss make_met2str.m in
detail, and note the differences for the others.  Some experience with
M must be assumed at this point, and it would probably be well to have
a reference at hand (eg. S. Wolfram, Mathematica, 3rd ed.,
(Wolfram Media, Champaign, IL, 1996).  A later edition should do.  One
need not necessarily know how to write programs in M to follow the
subsequent discussion -- experience using it interactively should
suffice.

	Make_met2str.m evaluates the C coefficients in Eq. (55), the
strain-strain second derivative term.  First, let's discuss the use of
this program.  From an interactive M window, one first loads the file,
executing the command

	<<make_met2str.m

This defines a function MakeMet2str[iterm,outfile], and one then
executes the following series of commands,

	MakeMet2str[1,"d2term1.out"]
	MakeMet2str[2,"d2term2.out"]
	...
	MakeMet2str[6,"d2term6.out"]

(where the outfile names are obviously arbitrary).  A few hundred lines
of d2term1.out are included in this directory to give an idea of what
the output looks like (there are 13,063 lines total!).

	Turning to the program itself, I will refer to line numbers:

Lines 1-23
These are a comment which was carried over from the first routine
of the series, metstr3.m, and does not accurately reflect what the
current routine does.  However, it gives the general idea.

Lines 25-28
	The function name, some M program conventions, and a list of
{private variables}.

Lines 35-165
	A dictionary of tensor products of k components, extending
Allan's scheme to (agh!) 7th rank.  This is a two-index M List.  The
indices following each line are comments for the programmer's benefit.

Lines 168-298
	The same thing for kp (<K'| in Eq. (55)).

Lines 306-308
	Definitions of the vectors k and kp as Lists of 3 components.

Lines 314-315
	The metric tensor m is defined as a 3X3 List of components, each
of which is a function of two strain variables s1 and s2.  There are
only 6 functions, not 9, building in the required symmetry of the
metric tensor.  The functions m11[s1,s2], etc. have not been defined.
Therefore, M will represent their derivatives symbolically, which is
what we want.

Lines 317,319
	These define dot products and squared magnitudes of our vectors
following Eq. (23).  Recall that k, kp are reduced wave vectors, which
is why the m comes in.

Line 376
	This List defines the modified Legendre polynomials in Eq. (49).

Line 326
	The start of a do loop over rank (the loop control is at its end
in M, line 369).  rank is the equivalent of the angular momentum l.

Lines 328-337
	This is a List of 6 possible polynomials "poly," each formed from
derivatives of the squared magnitudes of k and/or kp and/or of the
Legendre's, grouped so that each of the 6 represents a particular
increment between input and output tensor rank.  Each invocation of the
routine only uses one of these based on the "iterm" input variable for
reasons of code structure discussed in the introduction.  This is where
the formal derivatives of the undefined metric tensor component
functions are introduced.

Lines 339,340
	These 6-term lists specify the input-output rank couplings, once
again selected by the input iterm.

Lines 345-346
	This outputs some text and indices for the Fortran code.

Lines 348-349
	The number of elements of the input and output tensors for the
current rank and iterm.

Lines 354-357
	The start of a double do loop on input and output tensor indices,
and THE ONLY SIGNIFICANT COMPUTATION IN THE ENTIRE PROGRAM!  This is
where M does in a few lines (and a few to a few ten's of computer
minutes on a fairly old PIII machine) the work of a hundred graduate
students for a hundred years.  From the horrendously complex poly, the
"Coefficient" function extracts the coefficient of the product of the
specified pair of k, kp tensor elements.  It is smart enough to ignore
the order of the tensor components in poly, and probably does a pretty
good job of collecting terms and identifying common subexpressions.
Feeding the result to the "Simplify" function further cleans up and
shortens the expression, probably better than could be done by hand.

Lines 358-365
	The results are written to the output file.  Note that cm is the
C in Eq. (55), with the m, m' subscript indices.  The l subscript and
nu, nu' pair are respectively rank and iterm.  The superscript indices
belong to the metric-tensor derivative components of C, which will be
evaluated when the Fortran routines are executed.  The # symbol here and
elsewhere in the output is a flag to be used in post-processing.

Lines 367-369
	These are the controls at the ends of the triply-nested do loops.

	Differences in the other M routines are as follows.
make_metintstr.m evaluates C's in Eq. (56) for mixed strain -- atomic-
displacement derivatives.  This is similar to make_met2str.m in almost
all respects.  The difference is that the List poly is defined differently.
Here, we have only one strain derivative.  Since the atomic-displacement
derivative here is being taken with respect to a reduced atomic
coordinate, this derivative simply brings down a reduced k or kp
component (from the phase or "structure factor" exponential) without
any metric tensor components.  This and the rank-coupling Lists above
have 5 members in this case.  The C coefficients are output as members
of a 3-dimensional cm array here, their first index being the atomic-
displacement direction and the second and third m, m' as above.

	make_metintstrv.m is the routine that was actually used for the
internal strain routines.  The difference is explained in the comment
lines 31-36.

	make_metstr3.m is historically the first of this set.  It
evaluates C's in Eq. (54), the first strain derivative expression.  The
structure here is slightly different in that "rank" is an input
variable and "iterm" is an (unrolled) internal loop.  The 3 poly's are
defined on lines 333, 364, and 395.  Since there were some differences
in the text to be output for the Fortran code, it was easier to write
the "iterm" loop as unrolled.  The C coefficient cm is written as a 4-
index array here, since it will be saved in the Fortran routine metstr
and used for many executions.  After loading the file, MakeMetstr
should be executed with the input variable rank set to 1, 2, and 3 (p,
d, and f) and the resulting output files concatenated befor post-
processing.  It appears from the comments to have been necessary to do
rank = 0 by hand (something probably blew up).  This was a finite task.

************************* POST PROCESSING ****************************

	In case you haven't looked at the sample of d2term1.out, do so
now.  The M output for the formal derivatives of the undefined metric
tensor component functions, "Derivative(1,0)(m11)(s1,s1)," etc. is not
Fortran.  The shell scripts fort_fix3.scr, fort_2fix.scr, and
fort_fixis.scr post-process the raw M output for the C's in Eqs. (54),
(55), and (56) respectively.  There are basically two sections to each
script.  First, sed is invoked (with an in-line editing script) to
substitute Fortran-friendly (and much shorter) notations for the metric
tensor and metric-tensor derivative elements.  Each of these scripts
replaces the file on which it operates (which is specified as an inline
argument, eg. "./fort_2fix.scr d2term1.out"), so it is a good idea to
make a copy first.

	sed continues with some housekeeping.  Multiplication by 1 as an
operation is eliminated.  Where M has put floating-point number with no
decimal places, they are made into integers (eg. 6. -> 6).  Where there
are decimals, "d0" is appended to make sure we don't get single-
precision.  (Note that integers, of which there are lots, will not
interfere with double-precision evaluation of these expressions.)  The
one place we don't want integers is as divisors, so we restore double
precision (eg. /48. -> /48 -> /48.d0).

	Now we have cleaned-up, acceptable Fortran statements, but have
many short lines which do not have proper free-form F90 continuation.
The sed output is saved to a temporary file and further processed by
format_code.x, the executable compiled from format_code.c, to fix the
continuation and "compactify" the Fortran.  format_code.c simply copies
its standard input to its standard output one character at a time, but
deletes input newline characters. It counts and tests characters, and
inserts newlines, continuation &'s, and spaces to produce reasonably
uniform line lengths and follow Abinit indentation rules.  It breaks
lines after )'s or the last of several adjacent )'s in an attempt to
make the code at least somewhat readable, not that anyone would really
want to.  It also makes use of the "#" characters inserted in the M
programs to flag the start of each new "cm(i,j)=" statement.

	The post-processed result of the excerpt of d2term1.out is shown
in d2term1.fort, and starts at line 97 in contstr21.  This and the
other full routines start by computing the needed metric tensor
derivatives, evaluate all the necessary cm terms, save them for re-use
where possible, and do the actual operations on the aa and bb array
arguments at the very end.

	The post-processing produced nearly error-free code.  There was a
very small amount of hand-editing, no more than a few lines in each
routine at most to fix a few things that caused compiler errors.  While
the c program inserted a warning comment (for possible hand editing) in
places when the number of continuation lines exceeded the Fortran 90
Standard maximum, none of the compilers that have been used since this
has been part of Abinit seemed to mind, so these comments were deleted.

********************** INTERFACE WITH NONLOP *************************

	nonlop is a long and complicated routine, so a broad-brush
outline of what it does is in order.  Basically, it has two halves.  If
the input variable signs==1, only the first half is executed and
expectation values of the nonlocal psp's are output.  If signs=2, both
halves are executed; the first to evaluate the "right" side of the
separable nonlocal potential (or its derivatives) operating on the
input wave function, and the second half to deliver plane wave matrix
elements of the "left" side based on coefficients evaluated in the
first half.

	The computation of (form factors X tensors), (FFXT), acting on the
input wave function is done in the opernl* routines.  (These routines
all do the same thing, but are coded differently for better performance
with different computer architectures.)  The same routines also have an
"output" mode if needed in the second half of nonlop, but the discussion
below will focus on the "input" mode.

	The opernl3 routine has a loop over wave-vector blocks, and calls
the routine mkffkg3 inside this loop. mkffkg3 internally loops over the
wave vectors in the current block with index ipw.  The required
products of tensors, form factors, and form factor derivatives vary
widely depending on what is being done (the "choice" input variable),
and the number of projectors for each angular momentum for the current
type of atom.  The first thing mkffkg3 does is to compute the required
tensor products of components each wave vector in the block, the
kpgx(ipw,ii) array. This corresponds exactly to the tensor definitions
in make_met2str.m (regarding ii as a one-dimensional index of the tnk
List).  The remainder of the program packs the leading dimension of the
output array ffkg(iffkg,ipw) with FFXT's in an order determined by a
number of if's and do loops.  There is no simple mapping of the
implicit composite of indices into iffkg.  However, referring to the
tnk definition at the start of make_met2str.m, ffkg has sequential
groups of elements corresponding to the full range of j in each (*i,j*)
set.

	opernl3 forms dot products on the plane wave index of ffkg and
the wave function (including a phase factor), running over the entire
index range iffkg=1,nffkg without regard to the composite nature of the
iffkg index.  This is the most time-consuming part, and justifies the
arcane packing of ffkg.  The remainder of the routine PARTIALLY UNPACKS
the dot-product(iffkg) array into several output arrays, mimicking the
if and do structure of mkffkg to recreate the composite identity of the
iffkg index.  The reader is directed to the OUTPUT description in
opernl3 for details.  The lead index of these arrays generally
corresponds to a segment of 1,nffkg.

	Now we come to the rationale for the detailed discussion of the
last few paragraphs.  Segments of the output arrays of operln3 are the
aa, bb arguments of our new routines for strain-perturbation RF
calculations.  It was necessary to expand mkffkg3 to include the new
FFXT's needed for these calculations, and to expand the set of output
arrays and the unpacking part of opernl3 accordingly. The calling
statements for the strain routines must comply with the iffkg structure
(in its partially unpacked form) to pass the proper array segments as
aa, bb arguments.

	The other significant addition to implement the nonlocal strain
operations was the need to extend mkffnl, the routine which computes
the Eq. (48) form factors and their derivatives, to include the second
derivatives.  This is called by a number of routines prior to calling
nonlop.

	We have discussed the "input" actions of opernl3, which produce
expectation values.  The "output" mode producing wave functions utilizes
only one of the new routines, metstr, and further extensions of opernl3
(in the sign==-1, choice==3 section) to do the required "partial
repacking" into the full iffkg index range.

	The alternative routines opernl2 (calling mkffkg) and opernl4a,
opernl4b function in the same manner as opernl3 and mkffkg3, with
differences such as index ordering and loop unrolling.