J3/03-266r1
Date: 14 November 2003
To: J3
From: Aleksandar Donev
Subject: Post F2003: ELEMENTAL Procedures and Efficiency, and PURE
Reference: J3-007R3, J3/02-315
Response is in J3/03-277
______________________________________
Summary
______________________________________
I point to some problems with efficiently using ELEMENTAL procedures in
Fortran 2002 and suggest a compiling scheme that would fix this, along with
some needed backward-compatible modifications to the current draft that would
make this implementation more feasible. This was partly inspired by recent
posts/discussions with Van Snyder, as well as past personal frustrations with
elementals in Fortran 95. Problems also exist with PUREs, discussed later.
NOTE: There is an extended example at the end of this paper.
___________________
ELEMENTAL
___________________
Required changes to the way ELEMENTALs work include (in order of priority):
1. Remove the constraint prohibiting elementals as actual arguments and
procedure pointer targets, and,
2. Making sure that non-intrinsic elementals and non-elementals are not
mixable, i.e., an elemental procedure pointer, type-bound procedure or
procedure dummy argument should only be allowed to point, be bound to, or be
argument associated with a non-elemental.
3. Add the dummy argument attribute SCALAR (and possibly NONSCALAR) that
guarantee that the actual will be scalar.
Either
4a. Allow INTENT(OUT) SCALAR arguments, or even better,
4b. Remove the purity requirement on elementals and replace with the condition
that order of execution be unimportant to the user.
The problem described is part of a much more general issue of writing
TKR-generic procedures in Fortran (think templates in C++ or generic classes
in Eiffel), but no attempt is made here to solve the general problem (it is
coming shortly, after consultation with Van and others). Besides, it is too
late to finally introduce genericity properly in Fortran. But it is not too
late to fix elementals!
The proposed implementation is only a suggestion to vendors. I am aware that
it will not be high on their priority list. But at least we are making the
way for such efficient implementations of elementals later.
___________________
PURE
___________________
PURE procedures are another vastly underused F95 feature. They are a very nice
concept, but the basic problem is that the restrictions on PURE are too
strong. For example, one cannot even raise an error flag (think exceptions),
or even point a pointer to some global data (I am not saying modify the data,
but just point to the data). This is especially true because ELEMENTALs are
required to be PURE. The real requirement is that the order of execution of
ELEMENTALs not matter.
I make suggestions here on how to relax this for ELEMENTALs, but do not try to
fix PURE ones.
______________________________________
The Problem
______________________________________
There is a frequently appearing task in Fortran programming which still finds
no good solution: Making procedures that can be called with arrays of
arbitrary rank...efficiently. Assume for the moment that these routines can
be written, using array-syntax, in a way that is otherwise independent of the
rank. Elementals provide this, albeit only for pure operations that
completely disregard the possible "arrayness" of the arguments. More
importantly, there is an efficiency problem with elementals which has
unfortunately made them a very seldomly used array feature in Fortran 95
(with the exception of, of course, the intrinsics).
The problem with elementals is that to really have efficiency, in a lot of
cases, the loop over the array elements needs to go *inside* the elemental
procedure, rather then outside a procedure call to a scalar version of the
elemental. I will assume that inlining is not possible, and in the examples I
give this is the case, either because dynamic binding is used with OO
approaches, or because the user actually provides the elemental procedure.
Vendors typically implement the loop around the call, possibly since this is
easier. However, with a few modifications in the way ELEMENTAL procedures
work, a few deletions of restrictions on elementals which make no sense, and
a change in attitude from compiler writers, the usefulness of ELEMENTALs can
significantly be increased.
This issue is compounded with OO approaches, especially because of dynamic
dispatch, whose overhead can dramatically be reduced if the slot-table lookup
is done externally to the loop. Examples of the need for rank-indepedendent
procedures include random number generation (see comments on non-purity
below), as well as a whole collection of problems in which the user provides
a routine which operates on scalar arguments but in reallity needs to be
called on a grid of points {I attach at the end a (long) example of numerical
quadrature of a one-dimensional function Int[f(x),x=a..b]}.
A noticeable feature in both these examples, and in general a stumbling block
to putting the loop in the elemental, is the fact that some of the arguments,
typically the passed object in OOP solutions, are scalar. A commonly occuring
feature is also that the user can predict this ahead of time and would be
better off restricting this to be a scalar indeed. I suggest that a new
attribute for dummy arguments of ELEMENTAL procedures be added, say SCALAR,
which would state that this argument is not part of the "elementalness" of
the procedure.
I would really like to design a novel set of language features that deals with
this issue of genericity, i.e. writing code which applies to different TKR
actual arguments without any change. Kind of like templates in C++, but with
the added complexity of the rank. However, this is a major job. Therefore, I
resort to elementals, and suggest some modifications which would make them a
viable replacement.
______________________________________
The Solution
______________________________________
I give an example using random number generation below. Random number
generation is not pure, and the ordering of the generation does matter for
the final result. However, for the user the ordering does not matter
(otherwise he/she should not call the generator with an array argument but
put the look around the scalar version him/herself), just like with the
RANDOM_NUMBER intrinsic. I have assumed below that the restriction on SCALAR
arguments being intent(in) has been removed, although this issue is of course
separate from this proposal. For now, just take this as a very simple, yet
illustrative, example.
Here is how the declaration of an OO random number generator would look like
in my proposed scheme:
module Random_Number_Generation
type, extensible :: random_number_generator
integer :: seed=0
contains
procedure(next_number), pass(generator), deferred :: next_number
! Generates the next number in the pseudorandom stream
end type random_number_generator
abstract interface
elemental subroutine next_number(generator,number)
! Can be used to generate arrays of random numbers
class(random_number_generator), SCALAR, intent(inout) :: generator
real, intent(out) :: number
end function next_number
end abstract interface
end module Random_Number_Generation
For users that really care about efficiency, the above generator, without the
SCALAR, would not be a satisfactory solution in F2x, due to the above
efficiency concerns. The solution one would have to resort to, is a very ugly
one, in which there is a generic binding for next_number, which has to
overriden in every specific extension (implementation) of a random-number
generator. So now every time we extend random_number_generator we have to
write 8 (or maybe 9, with one special case for contiguous arrays) different
specific procedures, which will inline the random-number generator manually.
This produces long, obscure, macro-monsters of codes which sadly I have to
write almost every day. Notice that we cannot write these rank-specific
routines only once, and then reuse them in subsequent extensions, since we do
not yet have access to the scalar random-number generator at this API stage.
If instead, the compiler actually generated these 8/9 versions of the
elemental procedure for arguments of different rank automatically (which is
very simple), simplicity and efficiency would be combined in a very appealing
way. So an elemental procedure will in fact become something like a
dope-vector with 8/9 procedure pointers, one for scalars, one for ranks 1
through 7, and one for the case when all arrays involved are contiguous.
Elemental dummy procedure arguments, procedure pointers, and type-bound
elementals will in fact be implemented as such dope-vectors. Of course this
is just a suggestion and there other alternatives. It should be up to the
vendors to choose a strategy, and consistently use it in all cases elementals
are involved.
Some will raise the issue of what to do when some of the actual arguments for
one of the non-SCALAR arguments are also scalar (as they can be according to
the current rules for ELEMENTAL). In this case, the old strategy of simply
putting the loop around the scalar version will be satisfactory, and is still
available to the compiler. Or the compiler can simply replicate the scalar to
an array, depending on what it chooses to do. We could also add another dummy
argument attribute, say NONSCALAR, which will guarantee that the actual is an
array.
______________________________________
Required Changes and Edits
______________________________________
I do ask for some help to ensure that these are complete. I state the required
functionality first, and then the essential edits:
1. Remove the constraint prohibiting elementals as actual arguments and
procedure pointer targets, but make sure that
2. Non-intrinsic elementals and non-elementals should not be mixable, i.e., an
elemental procedure pointer, type-bound procedure or procedure dummy argument
should only be allowed to point, be bound to, or be argument associated with
a non-elemental.
Edits:
___________________
Procedure pointers:
143:28 Delete C728 and possibly replace with:
C728: If is a nonintrinsic elemental procedure, then
shall have an explicit interface and be elemental.
I do not think this is needed because of 144:16-20.
260: 31-32
Why is this there, and how do we need to change this?
___________________
Type-bound procedures and procedure components:
55:10+ Add
(2+1/2) If the inherited binding is elemental then the overriding binding
shall also be elemental.
___________________
Procedure dummy arguments:
263: 25 Remove C1229 and replace with
263: 30+
If procedure-name actual-arg is the name of a nonintrinsic elemental
procedure, the dummy argument corresponding to actual-arg shall have an
explicit interface and be elemental.
Is this enough and is there a better place to add this?
267: 25 Delete the parenthesized ending.
___________________
3. Add the dummy argument attribute SCALAR (and possibly NONSCALAR) that
guarantee that the actual will be scalar.
This will also be considered part of the procedure characteristics. I do not
give detailed edits here since this is subject to interpretation.
Either
4a. Allow INTENT(OUT) SCALAR arguments, or even better,
4b. Remove the purity requirement on elementals and replace with the condition
that order of execution be unimportant to the user.
The random-number generation example above illustrates the fact that the
requirements that the results of the execution of an elemental reference be
completely independent of the order of execution is too strong. Consider a
case where a user just wants an error flag raised if an error occurs during
the execution of an elemental reference. The user does *not* want an array of
flags (think exceptions). Solution 4a above provides an alternative.
______________________________________
Extended Example
______________________________________
Here is another example which attempts to illustrate the occurence of
non-inlinable ELEMENTALs in codes which require the user to provide some
procedure, which will be called on an array of non-zero rank, but the user
does not need to know this. It is definitely not a desirable thing to ask the
user to write 8 specific routines himself, as this is a hard and error-prone
task.
The example is one of numerical quadrature (integration) of a user-provided
function f(x) on the interval [a,b], i.e., the evaluation of a definite
integral (DI). An OOP solution is given. Here f(x) needs to be evaluated on a
whole grid of points (the integration mesh), which happens to be
two-dimensional because it represents some kind of hierarchical fancy mesh.
The code is not documented in detail. Feel free to ask me about it.
First we design an abstract specification for an OOP definite integrator:
module DI_Quadrature ! DI=Definite Integral
type, extensible :: di_problem
! The definite integral that is needed
! This will contain user-data upon extension
real :: x_lb=0.0, x_ub=1.0 ! Bounds of integral
...
contains
procedure(f), pass(di), deferred :: f
! User provided f(x)
...
end type di_problem
type, extensible :: di_integrator
! The actual numerical integrator
real :: global_error=0.0 ! Estimated error
...
contains
procedure(Integrate), pass(integrator), deferred :: Integrate
! The actual integrator--Euler, Runge-Kutta, etc.
...
end type di_problem
abstract interface
elemental function f(problem,x)
class(di_problem), intent(inout) :: problem
real, intent(in) :: x
real :: f
end function f
subroutine Integrate(problem,integrator)
class(di_problem), intent(in) :: problem
class(di_integrator), intent(inout) :: integrator
end subroutine Integrate
end interface
end module DI_Quadrature
Now we actually implement a specific fancy integration scheme:
module DI_Fancy_Quadrature
! Specific implementation: Fancy Fancy quadrature for DI's
use DI_Quadrature
type, extends(di_integrator) :: fancy_di_integrator
integer :: n_points=0 ! Number of needed points along x axis
real, dimension(:,:), allocatable :: x
! These are rank-2 for *this* integrator only
...
contains
procedure :: Integrate=>IntegrateFancy5
...
end type fancy_di_integrator
contains
subroutine IntegrateFancy5(problem,integrator)
class(di_problem), intent(in) :: problem
class(fancy_di_integrator), intent(inout) :: integrator
real, dimension(5,integrator%n_points), allocatable :: f_x
! f(x) on a fancy grid
...
f_x=problem%f(integrator%x) ! Call the elemental routine
! This will call the rank-2 specific ``elemental''
...
end subroutine
end module DI_Fancy_Quadrature
The user on the other hand writes a description of his integration problem (in
this case f(x)=a*x+x^b) for some user-specified a and b:
module My_DI
use DI_Quadrature
! We do *not* use DI_Fancy_Quadrature since this may not yet be written
type, extends(di_problem) :: my_di_problem
real :: a=1.0, b=1.0 ! *Toy* example
! In real life there would be something *complicated* here
! such as a Monte-Carlo simulator to obtain f(x)
contains
procedure :: f=>my_f
end type my_di_problem
contains
elemental function my_f(di_problem,x)
class(my_di_problem), intent(inout) :: my_di
real, intent(in) :: x
real :: f
f=my_di%a*x+x**(my_di%b) ! f(x)=ax+x^b
end function my_f
end module My_DI
And here is how this can actually be used in a program:
program DI_Example
use DI_Quadrature
use DI_Fancy_Quadrature
type(my_di_problem) :: problem
type(fancy_di_integrator) :: integrator
problem%a=2.0; problem%b=-2.0
// Make a specific DI
...
call integrator%integrate(problem)
! You can also try another integrator if you have one
...
end program DI_Example