J3/14-211r1
To: J3
From: Malcolm Cohen
Subject: REDUCE
Date: 2014 June 27
1. Introduction
The current definition of REDUCE is unclear and needs improvement.
Furthermore, in light of discussion some changes to the specifications and
syntax are warranted.
2. Deficiencies of the current definition
That the initial order of the sequence is intended to be "array element
order" is implied but not specified.
That the application of the function (to adjacent elements of the sequence)
should be with the argument order being the same as the sequence order is
implied
"OPERATION(x,y) for adjacent x and y in the sequence"
but not rigorously specified.
It might be thought useful to explain the consequences of a non-associative
operation in slightly more detail than a simple recommendation.
We could add something to Annex A, but since this already has the enormous
catchall
- the values returned by some intrinsic functions (13);
and indeed hardly any processor-dependent functions are individually
listed, that would appear to be unnecessary.
3. Discussion of ORDERED
It has been proposed that there be an ORDERED argument that specifies that
the operation "be applied to the elements of the sequence in order". Taken
at face value, this would prohibit application of the associative law, i.e.
for 4 elements a1 to a4 would require computation of
((a1 op a2) op a3) op a4)
and not permit
(a1 op a2) op (a3 op a4)
However, the proposal states
"The requirement for commutativity is removed when ORDERED is
absent or present with the value FALSE."
This is puzzling since there is no requirement for commutativity in the
current REDUCE specification/syntax/text. Perhaps the intended meaning was
that the requirement that *the processor* not "commute" the arguments be
lifted. This provides little benefit since computing a3 op a4 is not more
onerous than computing a4 op a3. Yes, there can potentially be some small
benefit in a parallel execution environment [such as OpenMP] in the case
where the cost of computing the operation is highly variable, but since
enabling that involves higher overheads this only applies to a tiny
minority of cases; this does not justify the extra complication to the
standard.
Furthermore, in discussion it was denied that ORDERED=.TRUE. was intended
to prohibit re-associating the operations. If that is indeed the case, it
would imply that ORDERED would be a spectacularly poor choice of keyword.
To sum up: there are three cases
(1) strict in-order evaluation: in this case the REDUCE intrinsic cannot
run any faster than a trivial user function; the only benefit is that
of clarity/convenience: no need to write a simple function.
(2) evaluation applying the associative law; that is, re-associating the
operand sequence is permitted. This is the model envisaged by the
current text. It gains nearly all of the possible performance
benefits. The operation "should" be associative otherwise the result
will be indeterminate in some way.
(3) evaluation applying both the associative and commutative laws: that is,
both re-associating and argument swapping are permitted. This has
little if any performance benefit over case (2). The operation
"should" be both commutative and associative for this to work.
It is clear that the default should be case (2), as this provides the
maximum benefit for a very reasonable requirement: that the operation
"should" be mathematically associative - in practice, what this means is
that any non-associativity should be unimportant to the user.
If the convenience factor of allowing case (1) is wanted, then indeed an
ORDERED argument would be in order.
There really does appear to be little point to case (3), which is unsafe
for a default and unlikely to give any significant benefit. It could
easily enough be provided by a compiler option such as "-Ounsafe" for the
few people who might want to do it. If we really do want to add this mode,
a separate COMMUTE argument (COMMUTE=.TRUE. to permit argument swapping)
would be in order.
Potential straw votes:
sv Add ORDERED= argument for strict evaluation. Default ORDERED=.FALSE..
result 8-3-4, ORDERED= added.
sv Add COMMUTE= argument for (3) above. Default COMMUTE=.FALSE..
result 1-10-5, COMMUTE= not added.
3. Revised Specifications
(Changes marked with **)
Add an intrinsic function that reduces an array by a user-defined
operation.
Similarly to the other reduction intrinsic functions,
(a) reduction of a single dimension shall be provided, and
(b) use of a logical mask array shall be supported.
The array to be reduced may be of any intrinsic or derived type. The
result for reduction of a single element shall be that element. The result
for reduction of zero elements (the "identity" for the operation) shall be
specifiable.
** Reduction of a zero-sized array with no identify specification shall
** cause error termination.
The user-defined operation should be mathematically associative but need
not be computationally associative. In order to facilitate reduction of
arrays of such things as quaternions and matrices, commutativity will not
be required. When more than two elements are being reduced, the operation
may be associatively applied to elements and intermediate results in any
order that does not commute (swap) operands.
** It shall be possible to request strict in-order evaluation.
4. Syntax
Same as before, plus
** Optional ORDERED argument, default .FALSE.,
** for strict in-order evaluation.
5. Edits to 14-007r1
[383:34] 13.7.140 REDUCE, heading,
replace "IDENTITY" by "IDENTITY, ORDERED" twice.
[384:13] 13.7.140 REDUCE, p5 Result Value, Case (i),
After "is the result of an iterative process."
Insert "The initial order of the sequence is array element order."
{Specify initial sequence precisely.}
[384:14-15] Ditto,
After "for adjacent x and y in the sequence,"
Insert "with x immediately preceding y,".
{Specify the obvious.}
[384:15] Ditto, p5 Result Value, Case (i),
After "x and y with r" insert
"; if ORDERED is present with the value true, x and y shall be the first
two elements of the sequence".
{Strict left-to-right evaluation for ORDERED=.TRUE..}
[384:17] Ditto,
Change "and is processor dependent otherwise"
To "and otherwise, error termination is initiated".
{Specify error termination on missing IDENTITY and zero-sized array.}
[384:33+] 13.7.140 REDUCE, after p6 Examples, insert note
"NOTE 13.nn
If OPERATION is not computationally associative, REDUCE without
ORDERED=.TRUE. with the same argument values might not always produce
the same result, as the processor can apply the associative law to the
evaluation."
===END===