
Rules_coding

ABINIT style for Fortran programming
(revised many times from the original draft in 1991 ... )

Copyright (C) 1991-2016 ABINIT group (DCA,XG,GMR,MM,HM,PT,MV)
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.txt .

=====================================================================

The following sections are covered :
0. Foreword.
1. Declarations.
2. Variables
3. File format
4. Constructs for flow control
5. Use of arrays
6. Coding practice
7. Exception handling, I/Os
8. To be avoided.
9. Use of BLAS and LAPACK subroutines.
10. Modules
11. Derived datatypes
12. Other topics
13. Useful links

=====================================================================

0. Foreword.

The ABINIT code should conform to the ANSI Fortran norm (Fortran 90).
This norm is abbreviated F90, while the
older norm, Fortran 77, is abbreviated F77.

The following set of rules complements the F90 standard.
The ABINIT code conforms to most of these additional rules already.
Each modification, or new routine is expected to adopt all
the standards, that will be also enforced gradually
in existing routines.

The developer is supposed to have a good knowledge of the F90
constructs. For an introduction to them, consult the
following URL :
... (to be filled)

The rules described afterwards do NOT apply to the routines
that are in the Lib_* directories.
Their source is voluntarily not modified (except recently, for OpenMP),
to keep coherence with the original versions, provided by somebody
who followed other coding rules.

Names having a special meaning in F90 are written here in uppercase.
They are usually NOT in uppercase in the code, however.

The basic utility of this document is to provide a
common style to all the routines. This is important,
in order for ABINIT developers
to understand quickly what a routine
does : they expect some information to be always present
in some part of each routine, or, in the
opposite sense, they should know where to place
each information (and not spend time to evaluate
where each information should be placed).
This document is also a way to enforce a
kind of "locality
of information" principle : all the information
needed to understand what a routine does should
be contained in the same routine, and if possible,
the information about a small part of the routine
should be contained within a screen page of this
part of the routine.

In general, the developer will conform to the style
by taking an existing routine as an example, and
use it to generate a new routine. However, it is
worth reading the present rules from time to time ...

Starting in January 2001, a working group
began to work on the update of these rules,
on the basis of a proposal by H. Matthis.
It became clear that the present document was already
very complex. It should be reordered, and also
reformatted using HTML and an index, so that the
reader can easily get the information he really needs.
This is to be done...
In the meantime, the result of the working group efforts
produced the new sections 10 (modules), 11 (derived datatypes),
as well as modifications (or addition) of the rules
 1.b, 1.c, 1.e, 1.f, 1.h, 2.a. 3.g, 3.n, 3.p, 3.q .
Modifications with respect to the rules for ABINITv3.0
are signaled thanks to the > sign.

Also, the new rules will be enforced gradually.
This means that not only the format, but also
the content of present document is IN A TRANSITIONAL state.


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

1. Declarations.

1.a. Use IMPLICIT NONE, forcing explicit declaration of all variables,
in all subroutines.
(enforced)

1.b. COMPLEX*16 or REAL*8 are not ANSI standards.
In the basis_defs module (see section 10.b), one defines
 the integer variable "dp":

INTEGER, PARAMETER :: dp = kind(1.0d0)

Accordingly, define floating point numbers as

REAL(dp) :: real_variable

(the double precision is standard, avoid single precision).
For the declaration of complex numbers, use two
real numbers
REAL(dp) :: complex_variable(2)
That is, place complex numbers in a single array with real parts in odd
array components and imaginary parts in even array components.
For historical reasons, using a COMPLEX subtype is not allowed
(sorry for this).

For an array of complex numbers, use
REAL(dp) :: array(2,size_array)

When setting values of constants, include "_dp" or "d0" on the end to
indicate double precision constants, e.g. 0.0_dp or 0.0d0 .

For integers, one has the choice between
INTEGER :: ii
and
INTEGER(i4b) :: ii
where i4b has been defined in the 'basis_defs' module.

1.c. Mention all the attributes of a variable in a single-line
declaration. This means that PARAMETER, SAVE, ALLOCATABLE, INTENT(IN),
... Should not be declared later. However, avoid the
DIMENSION attribute, and declare the dimension after each array.
This is to save numerous line in the declaration section.
The use of the double colon is the standard.
For example :
REAL(dp), PARAMETER :: aa=1.0d0, bb=1.0d0/3.0d0
INTEGER, ALLOCATABLE :: cc(:)

1.d. For character strings, use
CHARACTER*10 :: name(5)
Avoid the rather confusing
CHARACTER name(5)*10                 ! to be avoided
Eventually use
CHARACTER*(*) :: names(8)
for a string of characters whose length is known
only in the calling subroutine. It is better to avoid undefined length
strings but it is not always possible.
(enforced)

1.e. Logical variables are allowed, especially for very big arrays,
where the memory gain is important, or for temporary
logical arrays (work arrays). However, when a variable
has to describe a choice, use an integer instead of
a logical : this
paves the way to the introduction of further choices.
For logicals, one has the choice between
LOGICAL :: ii
and
LOGICAL(lgt) :: ii
where lgt has been defined in the 'basis_defs' module.
(enforced)

1.f. Never use automatic arrays, use allocatable arrays.
First, the handling of automatic and allocatable variables in memory
is different : automatic variables are placed in the stack, while
allocatable arrays are placed in the heap. The stack size is usually limited.
Second, this also allows the checking of the parts of the program
where memory is allocated, through the use of a
grep allocatable *     or     grep allocate *
command. In order for such a search to be efficient,
never continue an allocate or allocatable command on the next line.
(see also rule 5.f)
Pointers are allowed whenever a dynamically-allocated
multi-dimensional array has to be allocated in a subprogram and then
used and deallocated outside it.
Example :
REAL(dp),DIMENSION(:),POINTER :: array
nullify(array)
CALL init(array)
... (using array)
DEALLOCATE(array)

where the subroutine init contains

INTEGER :: size_array
REAL(dp),DIMENSION(:),POINTER :: array
size_array=...(some expression)
ALLOCATE(array(size_array))
... (initialize array)
END SUBROUTINE

There is a great advantage in using pointers as shown
in this example. It is no longer necessary to define
allocatable arrays in the main program itself  any array
should be allocated only when it is necessary,
and should be deallocated anywhere and as soon as
possible. The same construct allows to use arrays in derived
datatypes.

Also note that a pointer should be nullified as soon as possible.

(enforced)

1.g. Always declare explicitly in the subroutine
the dimension of a variable.
This is to allow to understand locally exactly what array operations
are doing. So, use
INTEGER :: zat(natom)
and avoid
INTEGER :: zat(*) 		! to be avoided
In general, avoid the declaration using asterisks.
(enforced)

1.h. Some compilers are very strict on the syntax (e.g. the PGI-HPF compiler
under Linux, on Pentium Pros), and do not allow for change of shape
between the calling routine and the called routine. Actually,
changing this characteristics is confusing.
For example :

PROGRAMME gnagna
REAL(dp) :: gprim(3,3)
CALL gnigni(gprim)
END

with

SUBROUTINE gnigni(gprim)
REAL(dp) :: gprim(9)
END SUBROUTINE

(the array gprim has not the same shape in gnagna and gnigni).
So, the rule to be enforced is : do not change the shape of the
arrays between calling and called routines. This, despite
the fact that only the address of the first element matters really.

Changing shape is not possible when using explicit
F90 interfaces of subprograms.
(enforced)

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

2. Variables

2.a. Never use a one-letter name for a variable, because they are
too difficult to find systematically. Even two-letter
variables (except duplicated letters, easy to find, or names
including j, k, q, x, or z, or some special variables
like "dp") should be avoided. So, use ii,jj,kk instead of i,j,k.
Note that F90 allows for names
longer than 6 characters, unlike F77.
(enforced)

2.b. Naming of looping variables (integers) :
	Use i... for the current (dummy) value of a given index
        Use n... for the number of a given quantity, in general
		also equal to its maximum (unlike in the old standard)
        Use m... for the maximum number of a given quantity,
		when it is not equal to the actual number of this
 		given quantity
For example: iatom, natom, matom, respectively.
(partially enforced)

2.c. For global variables, always use a single common name
throughout the entire code. Exception : utility programs should
use generic names for the arguments.
(partially enforced)

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

3. File format

3.a. Use the Fortran90 free format.
(132 column maximum, continuation sign & ... &, comment sign ! ,
no restriction on the use of the first columns)
Note that doubling the comment sign (!!) will be used to
allow processing by Robodoc, see rule 3.e .
(enforced)

3.b. Always use the continuation sign on both the continued line
(at its end), and on the continuation line (at its beginning).
(enforced)

3.c. Use indentation of two columns for each DO-loop or each
IF-block, or any other of structure.
(usually enforced)

3.d. Use indentation of 10 supplementary columns for the parts
of the code related to the parallel MPI version.
For example :
#          if defined MPI
           include 'mpif.h'
#          endif
Note that the # symbol, recognized by cpp, is in the first column.
This is because the cpp on the HP S-class machine does not accept
it if it is placed in another column.
(enforced)

3.e. The header of each routine must use a special format,
so that it can be processed by the Robodoc documentation
program (see http://www.xs4all.nl/~rfsber/Robo/robodoc.html
and http://www.xs4all.nl/~rfsber/Robo/manual.html).
This header of the file will have to mention :
- the name of the subroutine
- the function of the subroutine
- the copyright
- the complete set of input arguments with explanation
- the complete set of output arguments with explanation
- the complete set of arguments that are both input and output, with explanation
- the list of calling routines (produced automatically : do not worry about it).
- eventually some notes.
Then, the body of the routine will follow (with CPP F90 instructions,
see rule 3.f).
For Robodoc treating Fortran 90, there is a special meaning attached
to the lines of the source files that begin with a double bang :
!!
There are also keywords ("NAME","FUNCTION","COPYRIGHT","INPUTS","OUTPUT",
SIDE EFFECTS","NOTES","SOURCE" ...) to be used.
The header format is the most easily understood by a commented example ...


!!****f* ABINIT/name_of_subroutine
!! NAME
!! name_of_subroutine
!!
!! FUNCTION
!! explanation of "name_of_subroutine"
!! will be described here ...
!!
!! COPYRIGHT
!! Copyright (C) yyy1[-yyy2] ABINIT group (initials)
!! 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.txt .
!!
!! INPUTS
!!  arg1 = ...
!!  arg2 = ...
!!
!! OUTPUT
!!  arg3 = ...
!!  arg4 = ...
!!
!! SIDE EFFECTS
!! Input/Output   ! if this line is found in a routine, use "SIDE EFFECT".
!!  arg5 = ...
!!
!! PARENTS        ! This paragraph will be produced automatically
!!                ! It will mention the routines that call the present one.
!!
!! NOTES
!!  (miscellaneous note, warning, etc., if any)
!!
!! SOURCE


The string "name_of_subroutine" must be replaced by the name of the subroutine.
In the copyright notice, "yyy1" and "yyy2" must be replaced by years of modifications,
"initials" must be replaced by the initials of the contributors, for example
!! Copyright (C) 2007-2016 ABINIT group (DCA, XG).
"arg1=..." and similar strings must be replaced by the actual description
of input or output variables. They must appear in the same order in which
they appear as arguments to the subroutine. Note that the Robodoc keywords
are "INPUTS" (with a "S") and "OUTPUT" (without a "S").
If physical quantities are being passed
as arguments, give the appropriate dimension for the quantity
(e.g. Hartrees or Bohr). No work space private to each routine should be an argument
of the routine, use ALLOCATABLE, ALLOCATE, and DEALLOCATE inside
the subroutine.

After the line
!! SOURCE
follows the CPP and F90 lines.

(enforced)

3.f. After the header, and before the F90 declaration of variables,
one must mention, in order :
- eventually, global CPP definitions
- the subroutine name and arguments
- eventually, "include" files

As an example :

 #define FNLEN 132    /* defines the length of file name variables */
 subroutine name_of_subroutine (arg1,arg2,arg3,arg4,arg5)
 include 'file'

(enforced)

3.g. Every routine should then have the declaration IMPLICIT NONE.
This should then be followed by other declarations as follows.
*** Very important note : the order of declarations is automatically
enforced by the abirules script. The developer can perfectly
leave to the script the duty of ordering his new declarations,
at the time of the merge with the rest of the code ***
Note that within each list, the variable names should be
in alphabetic order.

A1. Arguments : scalars

INTENT(IN)
(1) integers
(2) double precision
(3) other variables in argument list
INTENT(OUT)
(4) integers
(5) double precision
(6) other variables in argument list
INTENT(INOUT)
(7) integers
(8) double precision
(9) other variables in argument list

A2. Arguments : arrays
 same as (1)-(9à) above

B1. Local variables (sequential) : scalars

INTEGER
(1) integer parameters
(2) integer with save attribute
(3) other integers
DOUBLE PRECISION
(4) double precision parameters
(5) double precision with save attribute
(6) other double precision

(7) other scalar variables

B2. Local variables (sequential) : arrays

INTEGER
(1) integer arrays
(2) allocatable integer arrays
(3) pointers to integer arrays
DOUBLE PRECISION
(4) real arrays
(5) allocatable real arrays
(6) pointers to real arrays

(7) other array variables

C1. Local variables (parallel case : MPI code section) : scalars
C2. Local variables (parallel case : MPI code section) : arrays
See the description of the sequential variables

NOTE that it is not possible to specify an INTENT
characteristics for POINTERs. However, they will be
placed in the correct section of the argument list.
See 3.q .
(the splitting based on INTENT is not yet enforced ;
apart this, enforced)

3.h. Do coding in lower case. Lower case is suggested because it looks
prettier. Use standard English grammar for comments (start sentences
with uppercase). The keywords DEBUG and ENDDEBUG, in uppercase,
should emphasize each part of the code, commented or uncommented,
that is written for the purpose of debugging.
(enforced)

3.i. Use comments frequently to explain manipulations and make
them in English.
(enforced)

3.j. Only permit a single "return" from any subroutine,
at the end of the subroutine, by using the
END SUBROUTINE statement (the old-fashioned RETURN statement
should be avoided). For program stopping, see the section
related to error handling.
(enforced)

3.k. For reason of readability and ease of maintenance (compilation time also),
files should not be too long.
In general, each subroutine will be contained in one file, whose
length should not exceed 850 lines. When a file exceeds 600 lines,
it is worthwhile to consider splitting it.
As of October 1998, about 13 out of 182 files exceeded 600 lines,
none of them beyond 850 lines.
It is also not worthwhile to write very short
subroutines. Indeed, the total number of subroutines should be kept
sufficiently low, in order to maintain them more easily. Also,
with a large number of routines, it becomes difficult to remember
what each routine is doing.
(should be enforced again for version 3.0 : there are presently
 more than 5 routines with more than 850 lines)

3.l. Equations mentioned in comment lines should be written
in Latex, for processing using Src2tex.
There are several ways to describe latex equations in the F90 comments
(these might be introduced by one "!" or two) :

1)  !! $ equation $
1') !  $ equation $
2)  !! $\displaystyle equation$
2') !  $\displaystyle equation$
3)  !!{{\ \begin{equation}
    !!  ... equation ...
    !!\end{equation} }}
3') !{{\ \begin{equation}
    !  ... equation ...
    !\end{equation} }}
4)  !!{{\ \begin{eqnarray}
    !!  ... equation ...
    !!\end{eqnarray} }}
4') !{{\ \begin{eqnarray}
    !  ... equation ...
    !\end{eqnarray} }}

For example :
!!  epsatm=$(4 \pi) \int_0^\infty [r^2 (V(r)+Zv/r) dr]$ (hartree)

If a long fraction sentence is found, "!! $\displaystyle equation $"
may be better for visualization.

In some cases, "!!\begin{equation} ...!!\end{equation}" or
"!!\begin{eqnarray} ...!!\end{eqnarray}" format might be better,
especially for very long equation lines.

3.m. Additional information about Src2tex and Robodoc.
As mentioned in rule 3.e , the content of the PARENT paragraph
will be created or updated automatically.
Also, the source files will be preprocessed automatically
before applying src2tex,
- the very first line will be added :
!{\src2tex{texfont=tt}}
- the very last line will also be added :
!!***
(usually enforced)

3.n. Place only one subroutine in each file.
This is to allow preprocessing for self-documentation
of sources. This rule is in conflict with the usual
use of modules, or chains of modules.
Resolving this conflict is the topics of current reflexion.


3.o. In order to ease the writing of the script that generates
"PARENT" routines before the self-documentation call to Robodoc,
it was needed to impose two additional rules :
- subroutines should be used instead of functions
- the instruction CALL should
always be the first of an instruction line. In particular,
the construct :

 IF(condition) CALL subroutine       !to be avoided

 should be avoided, and replaced by
 IF(condition) THEN
  CALL subroutine
 ENDIF
(enforced)

3.p. The file format might change, due to the recent
introduction of modules, pointers, interfaces,...
in ABINITv3.1 . This is a topics of current
reflexion. See HM2_document, section 2 of the
proposal by H. Matthis, in the present directory,
~abinit/doc/developers .

3.q. The INTENT must always be precised, except
for pointers (which have no intent) and subprogram
names given as formal arguments (rare).
Nevertheless pointers should be placed in list of
formal arguments according their 'quasi-intent'
in source code (read/modified/affected data).
On the other hand, a subprogram given as formal argument
behaves always like 'intent(in)'. In that case, the formal
argument consists of the explicit interface of the
subprogram.
A formal argument belonging to a structures type
must have an 'intent(in)' (respectively 'intent(out)')
if all fields of the structure are read-only
(resp. write-only). In all other cases, specify
an 'intent(inout)' and detail (with comments) the status
for each field.


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

4. Constructs for flow control

4.a. Do not use numeral labels. These can be avoided through the
use of the construct
DO ...
 ...
ENDDO
as well as the instruction (new to F90) CYCLE or EXIT,
or the constructs based on IF, CASE, WHERE or FORALL.
Note : Numeral labels can also be avoided for FORMAT statements by
placing the format descriptor in a character string, like in
WRITE (unit, '(i3)' ) nn
or
READ (texte, '(i6, f6.2)' ) nn, xx
A character string can be explicitly declared, then
initialized, if it is to be used more than once.
(enforced)

4.b. Literal labels or comment lines must be used for each
construct that is longer than one page of a screen (about 20 lines),
both at the beginning and at the end of the construct.
(enforced)

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

5. Use of arrays

5.a. Distinguish explicitly a vector operation from its
scalar counterpart, by specifying colons :
suppose
REAL(dp) :: aa(10), bb(10), cc(10)
then use
aa(:)=bb(:)+cc(:)
and not
aa=bb+cc			! to be avoided
However, this rule can lead to problems with the stack size, see rules 5c-e.
(enforced)

5.b. In order to improve the readability of the
code, use vector operations instead of
explicit DO-loop constructs. Use also the intrinsic
subroutines DOT_PRODUCT, or MATMUL, as well as
functions like SUM or MAXVAL. However, avoid using
segments as arguments. Use directly the array name : you
do not need to distinguish this operation from an hypothetical
scalar counterpart.
For reason of efficiency, very heavily used parts of
codes could be coded differently.
Use also BLAS and LAPACK routines for higher-level
constructs, although they
do not lead to as good a readability (see later).
(partially enforced)

5.c DO NOT pass as an argument of a subroutine, a segment
of an array. For example :
call routine(aa(1:naa))         ! to be avoided
Indeed, some compilers copy all the values of the segment into
a temporary variable. Thus, this is memory and CPU time consuming.
Moreover, the zone of memory that is used for that purpose can be limited
in size, and thus the code is constrained not by the full physical
memory of the machine, but by the size of that zone (stack size on the DECs).
For character strings, this restriction is not too important,
since character strings are rather short in any case, so the rule
is not to be followed for character strings. The same argument
applies when the size of the segment is know to be
limited in all cases (like in nonlop.f, setsym.f, xctotl.f or pseudopotential
routines for example).
(enforced, with the restrictions mentioned)

5.d Avoid the following construct, except when the segment size is
sufficiently small :
aa(:,ii)=aa(:,jj)
On the DEC machines, the compiler thinks it must first store
the segment aa(:,jj) in the stack, then copy the stack in aa(:,ii).
This not only doubles the time for copying, but can lead to stack size
problems. This likely happens each time the same array
appears both at the left-hand side and right-hand side of an assignment
expression.
However, there is no problem with
aa(:,ii)=bb(:,jj)
(enforced)

5.e Avoid the use of large segments in read or write operations :
read(6,*)aa(:,ii)
Again, on the DEC machines, the stack is used, and there are potential
problems with the declared stack size.
Use instead :
read(6,*)(aa(jj,ii),jj=jmin,jmax)
A full array, not presented as a segment, also has no problem :
read(6,*)bigarray
But avoid :
read(6,*)bigarray(:)
(enforced)

5.f For allocation and deallocation of arrays :
- never continue an allocate or allocatable command on the next line.
   (see rule 1.f)
(enforced)

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

6. Coding practice

6.a. Run innermost loop fastest in looping. This is simply
less CPU time consuming in Fortran.
(usually enforced)

6.b. Use generic function names instead of specializing to
double precision function calls for Fortran intrinsic functions.
E.g. use "cos" or "sin", not "dcos" or "dsin".
(enforced)

6.c. On many machines cache memory conflicts, which slow
down performance, are avoided by not using dimension of
arrays in multiples of 256. Thus it is recommended to
avoid array lengths which are multiples of 256. This
becomes especially important in the use of Fast Fourier
Transforms.
(usually enforced)

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

7. Exception handling and I/Os

7.a. The code issues four types of special messages :
ERROR, BUG, WARNING and COMMENT messages.
ERROR and BUG messages cause the code to stop,
immediately, or after a very small delay. An ERROR is attributed
to the user, while a BUG is attributed to the developer.
A WARNING message
indicates that something happened that is not as expected,
but this something is not so important as to make the code stop.
A COMMENT message gives some information to the user, about
something unusual. None of them should appear when the run
is completely normal, and runs as expected.
(enforced)

7.b. Some checking should be provided for input variables
for physical reasonableness or internal consistency.
This is intended to maintain a user-friendliness already present
in the code, and will greatly improve the possibility of widespread
use by new users.
(usually enforced)

7.c. Exception diagnostics from routines should be directed
to the standard output (std_out), and report :
- the name of the routine which is reporting the error;
- the sort of exception that is occurring (ERROR, BUG, WARNING or COMMENT);
- the present status;
- the expected status;
- in the case of an ERROR message or a BUG message,
what is wrong, like "The volume of the unit cell should not be zero";
- in the case of ERROR message, what action the user should take
(note that BUG messages are automatically appended with the directive
 "Action : contact ABINIT group").
Experience shows that the trouble spent putting good
descriptive error diagnostics in place is well repaid by
minimizing the time spent fixing trivial bugs.
(usually enforced)

7.d. Exiting should always be done through the LEAVE
subroutine (presently LEAVE_NEW), except for temporary debugging purposes.
The corresponding error
message must be printed through the wrtout routine, since
this routine can be used to modify the error message
in a standard way when needed.  See also the
rules_parallel file, since exiting is much more complicated
in the parallel case.
(exit through leave_new is enforced, systematic use of wrtout not yet)

7.e. Never write a character in the first column of the output file.
This is because the fldiff tool (see programmer_guide) uses the
signs in this column to perform comparison of test case output files
with reference ones.
(enforced)

7.f. When you output a message using the wrtout routine, do not
use the Fortran90 / (Slash) editing. It is not treated correctly.
Use instead  char(10), that will be handled in a special way
(note that the abbreviation 'ch10' for 'char(10)'
 is defined in basis_defs.f )
By contrast, do not use  char(10) when
you output something using the Fortran90 write intrinsic :
char(10) is the end-of-line sign ONLY under UNIX.
For example MacOS needs char(13) ... So, use the Fortran90 / (Slash)
editing in that case ... For example :
write(std_out,'(a,/,a,/)') ' hello',' world !'
(enforced)

7.g. When you write to the main output file, using the 'f'
format of Fortran (like in f12.4), in case there
is a possibility of different rounding on different
computers, add a small
number (on the order of 100000 times smaller than the
precision of the f format) to the number you want to write.
For example, be the number  numb=0.56250000000_dp ,
to be written using the format f11.3 :
 write(ab_out,'(f11.3)')numb      ! will give portability problems
On some computers, it will be written 0.562 , and on some others,
it will be written 0.563 . Instead, use :
 write(ab_out,'(f11.3)')numb+tol10
as tol10 is defined 0.0000000001_dp in defs_basis.f .
In this way, it will print as 0.563 always.

7.h. Use always the string "std_out" to write to the standard output file 
(the "log" file, with unit number 6). Use always the string "ab_out"
to write to the main output file (with unit number 7). Use alwas the string "std_err"
to write to the standard error file (with unit number 0). 
This is to allow correct redirection in case of massively 
parallel run (beyond 100 processors) : the writing
can be globally suppressed. Also, writing to the standard error file
should only be done when the code is about to stop 
gain to avoid problems with numerous write with massively parallel 
machines.
(enforced)

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

8. To be avoided.

8.a. Use <  instead of the old-fashioned .LT.   ,
     use <= instead of .LE.  ,
     use >  instead of .GT.  ,
     use >= instead of .GE.  ,
     use == instead of .EQ.  ,
     use /= instead of .NE.  ,
(enforced)

8.b. Avoid RETURN at the end of each subroutine, use only END SUBROUTINE
(this was not allowed by F77). Avoid RETURN at the end of each function,
use only END FUNCTION
(enforced)

8.c. Avoid DATA statements, initialize variables in their declaration line :
INTEGER :: nline = 6
(note : initialized variables get automatically the SAVE status)
(note : there are arrays constructors to initialize arrays in their
        declaration line )
(enforced)

8.d. Avoid the use of COMMON and EQUIVALENCE.
(enforced)

8.e. Do not use an isolated quote (') in a comment, like in the following :
! This isn't a good thing
Indeed, the CPP (preprocessor) does not treat properly such occurrences
on some machines. Instead replace it simply with
! This is not a good thing
or
! This isn t a good thing
Everybody will understand the latter, although grammatically incorrect.
(usually enforced)

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

9. Use of BLAS and LAPACK subroutines.

Note : BLAS and LAPACK subroutines are public domain subroutines, gathered
in libraries, that can be optimized by each hardware vendor on their
platform. BLAS means Basic Linear Algebra Subroutines. There are
different levels of BLAS :
- BLAS 1 : basic scalar and vector subroutines
- BLAS 2 : matrix-vector subroutines
- BLAS 3 : matrix-matrix subroutines
- Sparse BLAS : sparse vector operations
LAPACK is a set of subroutine that supersedes the old LINPACK and EISPACK
libraries, and addresses Linear equation solution and Eigensystem solution
(matrix inversion, linear least squares problems, orthogonal factorization,
ordinary eigenvalue problems, generalized eigenvalue problems,
singular value decomposition).
The Web addresses :
http://www.netlib.org/lapack/index.html
http://www.netlib.org/blas/index.html

9.a. For inversion or diagonalisation of small matrices
(however larger than 3*3), LAPACK should be used
preferentially to any other subroutines.
(usually enforced)

9.b. In general, the capabilities of BLAS subroutines can be obtained
by the array capabilities of F90. The corresponding code is usually
easier to read. For example,

REAL(dp) :: aa(npw),bb(npw)
CALL DCOPY(npw,aa,1,bb,1)

can be replaced by

REAL(dp) :: aa(npw),bb(npw)
bb(:)=aa(:)

So, the rule will be to use F90 instead of BLAS subroutines,
with some exceptions (see 9.c). In general, the following
can be avoided : DSCAL, DCOPY, DAXPY, DZERO, DSUM, DMAX ..., as well as
their corresponding COMPLEX variant.
(partially enforced)

9.c. There are exceptions :
- constructs that would still be easier to read with BLAS calls,
- constructs that are critical for speed, and faster with BLAS calls.
In the first class of routines, one find the scalar product between
complex numbers, in our particular representation by
REAL(dp) :: aa(2,size),bb(2,size)  arrays.
Indeed, the F90 function
DOT_PRODUCT(aa,bb) cannot be used, since it would not
lead to the evaluation of a
complex dot product, but a real dot product.
So, use
call DDOT(2*size,aa,1,bb,1)
(at present, no other BLAS routines than those called by LAPACK
 are used in ABINIT, so these rules are indications for the future)

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

10. Modules.

10.a. The architecture of ABINIT is modular, albeit
not principally relying on F90 modules in its present status.
Routines are partitioned in different subdirectories,
with a related hierarchy of usage. A more extensive use
of modules at that level is the topics of current
reflexion. See HM5_document, section 5 of the
proposal by H. Matthis, in the present directory,
~abinit/doc/developers .

10.b. The module named 'basis_defs' contains global variables
of ABINIT, and is used in every subroutine.
It contains the definitions of :
- subtypes (e.g. kind of floating numbers)
- I/O units
- selected fractionary real constants or trigonometric constants
- physical constants (conversion factors)
This module should also be mentioned in all block interface, e.g.:
interface
 subroutine jojo(npt,jiji)
  use basis_defs
  integer,  intent(in) :: npt
  real(dp) :: jiji
 end subroutine jojo
end interface
This is because dp must be defined before real(dp)

10.c. Modules are also useful for the definition
of F90 interfaces (for optional arguments),
or F90 derived datatypes (structured types).

10.d. Although they are allowed, no rules has been
set up for the use of optional arguments, and calls
based about keywords. This is a topics of
current reflexion. be cautious when you introduce
such a feature, and mention it to the ABINIT coordinator.



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

11. Derived datatypes

11.a. Derived datatypes should be declared in the
adequate module (see rule 10.c). These are powerful
F90 constructs, but the information about them is
not local to the subroutine where they are used, so
they should be introduced in a controlled way, in
order for the programmers to become sufficiently
easily familiarized with them: the introduction
of a new derived datatype must be made in agreement
with the coordinator of ABINIT.
The introduction of appropriate derived datatypes
in ABINIT is one of the central issues of v3.2 and
later versions.

11.b. Suffix a type identifier with '_type'.
For example :

TYPE ipe_type              ! information on program execution
 INTEGER :: nb_proc
 CHARACTER*63 :: fname
END TYPE ipe_type

which allows to define in subroutines :

TYPE(ipe_type) :: ipe

11.c. Pros and cons.
Grouping connected variables into structured types
is interesting for readability (it avoids too long
lists of formal arguments for instance) and may
facilitate programmation changes inside subprograms.
Imagine you want to add a new general information and
you want an access to this data in many subprograms :
if you insert a new field in 'ipe_type' structure,
it will not be necessary to change each subprogram
interface.

However, source code itself may become less readable.
Also, remember that the use of structured types is
never more efficient for CPU : complex declarations
should be avoided.


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

12. Other topics.


12.a. For the time being, pointers are only allowed
when an array has to be allocated in a subprogram, and
then used and deallocated outside it. See rule 1.f .
They should be nullified as soon as possible.

12.b. F95 constructs are not yet allowed. It has been
checked for ABINITv3.1 that not all compilers are
able to deal with them.

12.c. Object-oriented programming in F90 (definition
of operators acting on objects...) is the
subject of current reflexion. See for example the URL:
http://www.cs.rpi.edu/~szymansk/oof90.html

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

13. Useful links.

On the WEB page (http://www.cs.rpi.edu/~szymansk/oof90.html):
"Introduction to Object-Oriented Concepts using Fortran90" :
http://www.cs.rpi.edu/~szymansk/OOF90/Forum.html (shorter version)
http://www.cs.rpi.edu/~szymansk/OOF90/F90_Objects.html (PS file available)
(e.g. the data encapsulation is explained, taking a FFT routine
as such an example. This might be related to the port using fftw...)
The other PS file ("How to Express C++ Concepts in Fortran 90")
seems similar to the above notes.
Also, to the link "Gocha ! Click here for
Fortran90 bug bites" on this page. This page warns the usage of
the new functions in Fortran90.

"Compiler and tools tricks"
http://www.fortran-2000.com/ArnaudRecipes/CompilerTricks.html

"Arnaud's advice to develop and debug numerical applications"
http://www.fortran-2000.com/ArnaudRecipes/ArnaudAdvice.html

