1INDEXING(1) User Contributed Perl Documentation INDEXING(1)
2
3
4
6 PDL::Indexing - how to index piddles.
7
9 This manpage should serve as a first tutorial on the indexing and
10 threading features of PDL.
11
12 This manpage is still in alpha development and not yet complete. "Meta"
13 comments that point out deficiencies/omissions of this document will be
14 surrounded by square brackets ([]), e.g. [ Hopefully I will be able to
15 remove this paragraph at some time in the future ]. Furthermore, it is
16 possible that there are errors in the code examples. Please report any
17 errors to Christian Soeller (c.soeller@auckland.ac.nz).
18
19 Still to be done are (please bear with us and/or ask on the mailing
20 list, see PDL::FAQ):
21
22 · document perl level threading
23
24 · threadids
25
26 · update and correct description of slice
27
28 · new functions in slice.pd (affine, lag, splitdim)
29
30 · reworking of paragraph on explicit threading
31
33 A lot of the flexibility and power of PDL relies on the indexing and
34 looping features of the perl extension. Indexing allows access to the
35 data of a pdl object in a very flexible way. Threading provides effi‐
36 cient implicit looping functionality (since the loops are implemented
37 as optimized C code).
38
39 Pdl objects (later often called "pdls") are perl objects that represent
40 multidimensional arrays and operations on those. In contrast to simple
41 perl @x style lists the array data is compactly stored in a single
42 block of memory thus taking up a lot less memory and enabling use of
43 fast C code to implement operations (e.g. addition, etc) on pdls.
44
45 pdls can have children
46
47 Central to many of the indexing capabilities of PDL are the relation of
48 "parent" and "child" between pdls. Many of the indexing commands create
49 a new pdl from an existing pdl. The new pdl is the "child" and the old
50 one is the "parent". The data of the new pdl is defined by a transfor‐
51 mation that specifies how to generate (compute) its data from the par‐
52 ent's data. The relation between the child pdl and its parent are often
53 bidirectional, meaning that changes in the child's data are propagated
54 back to the parent. (Note: You see, we are aiming in our terminology
55 already towards the new dataflow features. The kind of dataflow that is
56 used by the indexing commands (about which you will learn in a minute)
57 is always in operation, not only when you have explicitly switched on
58 dataflow in your pdl by saying "$a->doflow". For further information
59 about data flow check the dataflow manpage.)
60
61 Another way to interpret the pdls created by our indexing commands is
62 to view them as a kind of intelligent pointer that points back to some
63 portion or all of its parent's data. Therefore, it is not surprising
64 that the parent's data (or a portion of it) changes when manipulated
65 through this "pointer". After these introductory remarks that hopefully
66 prepared you for what is coming (rather than confuse you too much) we
67 are going to dive right in and start with a description of the indexing
68 commands and some typical examples how they might be used in PDL pro‐
69 grams. We will further illustrate the pointer/dataflow analogies in the
70 context of some of the examples later on.
71
72 There are two different implementations of this ``smart pointer'' rela‐
73 tionship: the first one, which is a little slower but works for any
74 transformation is simply to do the transformation forwards and back‐
75 wards as necessary. The other is to consider the child piddle a ``vir‐
76 tual'' piddle, which only stores a pointer to the parent and access
77 information so that routines which use the child piddle actually
78 directly access the data in the parent. If the virtual piddle is given
79 to a routine which cannot use it, PDL transparently physicalizes the
80 virtual piddle before letting the routine use it.
81
82 Currently (1.94_01) all transformations which are ``affine'', i.e. the
83 indices of the data item in the parent piddle are determined by a lin‐
84 ear transformation (+ constant) from the indices of the child piddle
85 result in virtual piddles. All other indexing routines (e.g.
86 "->index(...)") result in physical piddles. All routines compiled by
87 PP can accept affine piddles (except those routines that pass pointers
88 to external library functions).
89
90 Note that whether something is affine or not does not affect the seman‐
91 tics of what you do in any way: both
92
93 $a->index(...) .= 5;
94 $a->slice(...) .= 5;
95
96 change the data in $a. The affinity does, however, have a significant
97 impact on memory usage and performance.
98
99 Slicing pdls
100
101 Probably the most important application of the concept of parent/child
102 pdls is the representation of rectangular slices of a physical pdl by a
103 virtual pdl. Having talked long enough about concepts let's get more
104 specific. Suppose we are working with a 2D pdl representing a 5x5 image
105 (its unusually small so that we can print it without filling several
106 screens full of digits ;).
107
108 perldl> $im = sequence(5,5)
109 perldl> p $im
110
111 [
112 [ 0 1 2 3 4]
113 [ 5 6 7 8 9]
114 [10 11 12 13 14]
115 [15 16 17 18 19]
116 [20 21 22 23 24]
117 ]
118
119 perldl> help vars
120 PDL variables in package main::
121
122 Name Type Dimension Flow State Mem
123 ----------------------------------------------------------------
124 $im Double D [5,5] P 0.20Kb
125
126 [ here it might be appropriate to quickly talk about the "help vars"
127 command that provides information about pdls in the interactive
128 "perldl" shell that comes with pdl. ]
129
130 Now suppose we want to create a 1-D pdl that just references one line
131 of the image, say line 2; or a pdl that represents all even lines of
132 the image (imagine we have to deal with even and odd frames of an
133 interlaced image due to some peculiar behaviour of our frame grabber).
134 As another frequent application of slices we might want to create a pdl
135 that represents a rectangular region of the image with top and bottom
136 reversed. All these effects (and many more) can be easily achieved with
137 the powerful slice function:
138
139 perldl> $line = $im->slice(':,(2)')
140 perldl> $even = $im->slice(':,1:-1:2')
141 perldl> $area = $im->slice('3:4,3:1')
142 perldl> help vars # or just PDL->vars
143 PDL variables in package main::
144
145 Name Type Dimension Flow State Mem
146 ----------------------------------------------------------------
147 $even Double D [5,2] -C 0.00Kb
148 $im Double D [5,5] P 0.20Kb
149 $line Double D [5] -C 0.00Kb
150 $area Double D [2,3] -C 0.00Kb
151
152 All three "child" pdls are children of $im or in the other (largely
153 equivalent) interpretation pointers to data of $im. Operations on
154 those virtual pdls access only those portions of the data as specified
155 by the argument to slice. So we can just print line 2:
156
157 perldl> p $line
158 [10 11 12 13 14]
159
160 Also note the difference in the "Flow State" of $area above and below:
161
162 perldl> p $area
163 perldl> help $area
164 This variable is Double D [2,3] VC 0.00Kb
165
166 The following demonstrates that $im and $line really behave as you
167 would exspect from a pointer-like object (or in the dataflow picture:
168 the changes in $line's data are propagated back to $im):
169
170 perldl> $im++
171 perldl> p $line
172 [11 12 13 14 15]
173 perldl> $line += 2
174 perldl> p $im
175
176 [
177 [ 1 2 3 4 5]
178 [ 6 7 8 9 10]
179 [13 14 15 16 17]
180 [16 17 18 19 20]
181 [21 22 23 24 25]
182 ]
183
184 Note how assignment operations on the child virtual pdls change the
185 parent physical pdl and vice versa (however, the basic "=" assignment
186 doesn't, use ".=" to obtain that effect. See below for the reasons).
187 The virtual child pdls are something like "live links" to the "origi‐
188 nal" parent pdl. As previously said, they can be thought of to work
189 similiar to a C-pointer. But in contrast to a C-pointer they carry a
190 lot more information. Firstly, they specify the structure of the data
191 they represent (the dimensionality of the new pdl) and secondly, spec‐
192 ify how to create this structure from its parents data (the way this
193 works is buried in the internals of PDL and not important for you to
194 know anyway (unless you want to hack the core in the future or would
195 like to become a PDL guru in general (for a definition of this strange
196 creature see PDL::Internals)).
197
198 The previous examples have demonstrated typical usage of the slice
199 function. Since the slicing functionality is so important here is an
200 explanation of the syntax for the string argument to slice:
201
202 $vpdl = $a->slice('ind0,ind1...')
203
204 where "ind0" specifies what to do with index No 0 of the pdl $a, etc.
205 Each element of the comma separated list can have one of the following
206 forms:
207
208 ':' Use the whole dimension
209
210 'n' Use only index "n". The dimension of this index in the resulting
211 virtual pdl is 1. An example involving those first two index for‐
212 mats:
213
214 perldl> $column = $im->slice('2,:')
215 perldl> $row = $im->slice(':,0')
216 perldl> p $column
217
218 [
219 [ 3]
220 [ 8]
221 [15]
222 [18]
223 [23]
224 ]
225
226 perldl> p $row
227
228 [
229 [1 2 3 4 5]
230 ]
231
232 perldl> help $column
233 This variable is Double D [1,5] VC 0.00Kb
234
235 perldl> help $row
236 This variable is Double D [5,1] VC 0.00Kb
237
238 '(n)' Use only index "n". This dimension is removed from the resulting
239 pdl (relying on the fact that a dimension of size 1 can always be
240 removed). The distinction between this case and the previous one
241 becomes important in assignments where left and right hand side
242 have to have appropriate dimensions.
243
244 perldl> $line = $im->slice(':,(0)')
245 perldl> help $line
246 This variable is Double D [5] -C 0.00Kb
247
248 perldl> p $line
249 [1 2 3 4 5]
250
251 Spot the difference to the previous example?
252
253 'n1:n2' or 'n1:n2:n3'
254 Take the range of indices from "n1" to "n2" or (second form) take
255 the range of indices from "n1" to "n2" with step "n3". An example
256 for the use of this format is the previous definition of the
257 subimage composed of even lines.
258
259 perldl> $even = $im->slice(':,1:-1:2')
260
261 This example also demonstrates that negative indices work like
262 they do for normal perl style arrays by counting backwards from
263 the end of the dimension. If "n2" is smaller than "n1" (in the
264 example -1 is equivalent to index 4) the elements in the virtual
265 pdl are effectively reverted with respect to its parent.
266
267 '*[n]'
268 Add a dummy dimension. The size of this dimension will be 1 by
269 default or equal to "n" if the optional numerical argument is
270 given.
271
272 Now, this is really something a bit strange on first sight. What
273 is a dummy dimension? A dummy dimension inserts a dimension where
274 there wasn't one before. How is that done ? Well, in the case of
275 the new dimension having size 1 it can be easily explained by the
276 way in which you can identify a vector (with "m" elements) with
277 an "(1,m)" or "(m,1)" matrix. The same holds obviously for higher
278 dimensional objects. More interesting is the case of a dummy
279 dimensions of size greater than one (e.g. "slice('*5,:')"). This
280 works in the same way as a call to the dummy function creates a
281 new dummy dimension. So read on and check its explanation below.
282
283 '([n1:n2[:n3]]=i)'
284 [Not yet implemented ??????] With an argument like this you make
285 generalised diagonals. The diagonal will be dimension no. "i" of
286 the new output pdl and (if optional part in brackets specified)
287 will extend along the range of indices specified of the respec‐
288 tive parent pdl's dimension. In general an argument like this
289 only makes sense if there are other arguments like this in the
290 same call to slice. The part in brackets is optional for this
291 type of argument. All arguments of this type that specify the
292 same target dimension "i" have to relate to the same number of
293 indices in their parent dimension. The best way to explain it is
294 probably to give an example, here we make a pdl that refers to
295 the elements along the space diagonal of its parent pdl (a cube):
296
297 $cube = zeroes(5,5,5);
298 $sdiag = $cube->slice('(=0),(=0),(=0)');
299
300 The above command creates a virtual pdl that represents the diag‐
301 onal along the parents' dimension no. 0, 1 and 2 and makes its
302 dimension 0 (the only dimension) of it. You use the extended syn‐
303 tax if the dimension sizes of the parent dimensions you want to
304 build the diagonal from have different sizes or you want to
305 reverse the sequence of elements in the diagonal, e.g.
306
307 $rect = zeroes(12,3,5,6,2);
308 $vpdl = $rect->slice('2:7,(0:1=1),(4),(5:4=1),(=1)');
309
310 So the elements of $vpdl will then be related to those of its
311 parent in way we can express as:
312
313 vpdl(i,j) = rect(i+2,j,4,5-j,j) 0<=i<5, 0<=j<2
314
315 [ work in the new index function: "$b = $a->index($c);" ???? ]
316
317 There are different kinds of assignments in PDL
318
319 The previous examples have already shown that virtual pdls can be used
320 to operate on or access portions of data of a parent pdl. They can also
321 be used as lvalues in assignments (as the use of "++" in some of the
322 examples above has already demonstrated). For explicit assignments to
323 the data represented by a virtual pdl you have to use the overloaded
324 ".=" operator (which in this context we call propagated assignment).
325 Why can't you use the normal assignment operator "="?
326
327 Well, you definitely still can use the '=' operator but it wouldn't do
328 what you want. This is due to the fact that the '=' operator cannot be
329 overloaded in the same way as other assignment operators. If we tried
330 to use '=' to try to assign data to a portion of a physical pdl through
331 a virtual pdl we wouldn't achieve the desired effect (instead the vari‐
332 able representing the virtual pdl (a reference to a blessed thingy)
333 would after the assignment just contain the reference to another
334 blessed thingy which would behave to future assignments as a "physical"
335 copy of the original rvalue [this is actually not yet clear and subject
336 of discussions in the PDL developers mailing list]. In that sense it
337 would break the connection of the pdl to the parent [ isn't this behav‐
338 iour in a sense the opposite of what happens in dataflow, where ".="
339 breaks the connection to the parent? ].
340
341 E.g.
342
343 perldl> $line = $im->slice(':,(2)')
344 perldl> $line = zeroes(5);
345 perldl> $line++;
346 perldl> p $im
347
348 [
349 [ 1 2 3 4 5]
350 [ 6 7 8 9 10]
351 [13 14 15 16 17]
352 [16 17 18 19 20]
353 [21 22 23 24 25]
354 ]
355
356 perldl> p $line
357 [1 1 1 1 1]
358
359 But using ".="
360
361 perldl> $line = $im->slice(':,(2)')
362 perldl> $line .= zeroes(5)
363 perldl> $line++
364 perldl> p $im
365
366 [
367 [ 1 2 3 4 5]
368 [ 6 7 8 9 10]
369 [ 1 1 1 1 1]
370 [16 17 18 19 20]
371 [21 22 23 24 25]
372 ]
373
374 perldl> print $line
375 [1 1 1 1 1]
376
377 Also, you can substitute
378
379 perldl> $line .= 0;
380
381 for the assignment above (the zero is converted to a scalar piddle,
382 with no dimensions so it can be assigned to any piddle).
383
384 Related to the assignment feature is a little trap for the unwary:
385 since perl currently does not allow subroutines to return lvalues the
386 following shortcut of the above is flagged as a compile time error:
387
388 perldl> $im->slice(':,(2)') .= zeroes(5)->xvals->float
389
390 instead you have to say something like
391
392 perldl> ($pdl = $im->slice(':,(2)')) .= zeroes(5)->xvals->float
393
394 We hope that future versions of perl will allow the simpler syntax
395 (i.e. allow subroutines to return lvalues). [Note: perl v5.6.0 does
396 allow this, but it is an experimental feature. However, early reports
397 suggest it works in simple situations]
398
399 Note that there can be a problem with assignments like this when lvalue
400 and rvalue pdls refer to overlapping portions of data in the parent
401 pdl:
402
403 # revert the elements of the first line of $a
404 ($tmp = $a->slice(':,(1)')) .= $a->slice('-1:0,(1)');
405
406 Currently, the parent data on the right side of the assignments is not
407 copied before the (internal) assignment loop proceeds. Therefore, the
408 outcome of this assignment will depend on the sequence in which ele‐
409 ments are assigned and almost certainly not do what you wanted. So the
410 semantics are currently undefined for now and liable to change anytime.
411 To obtain the desired behaviour, use
412
413 ($tmp = $a->slice(':,(1)')) .= $a->slice('-1:0,(1)')->copy;
414
415 which makes a physical copy of the slice or
416
417 ($tmp = $a->slice(':,(1)')) .= $a->slice('-1:0,(1)')->sever;
418
419 which returns the same slice but severs the connection of the slice to
420 its parent.
421
422 Other functions that manipulate dimensions
423
424 Having talked extensively about the slice function it should be noted
425 that this is not the only PDL indexing function. There are additional
426 indexing functions which are also useful (especially in the context of
427 threading which we will talk about later). Here are a list and some
428 examples how to use them.
429
430 "dummy"
431 inserts a dummy dimension of the size you specify (default 1) at
432 the chosen location. You can't wait to hear how that is achieved?
433 Well, all elements with index "(X,x,Y)" ("0<=x<size_of_dummy_dim")
434 just map to the element with index "(X,Y)" of the parent pdl (where
435 "X" and "Y" refer to the group of indices before and after the
436 location where the dummy dimension was inserted.)
437
438 This example calculates the x coordinate of the centroid of an
439 image (later we will learn that we didn't actually need the dummy
440 dimension thanks to the magic of implicit threading; but using
441 dummy dimensions the code would also work in a threadless world;
442 though once you have worked with PDL threads you wouldn't want to
443 live without them again).
444
445 # centroid
446 ($xd,$yd) = $im->dims;
447 $xc = sum($im*xvals(zeroes($xd))->dummy(1,$yd))/sum($im);
448
449 Let's explain how that works in a little more detail. First, the
450 product:
451
452 $xvs = xvals(zeroes($xd));
453 print $xvs->dummy(1,$yd); # repeat the line $yd times
454 $prod = $im*xvs->dummy(1,$yd); # form the pixelwise product with
455 # the repeated line of x-values
456
457 The rest is then summing the results of the pixelwise product
458 together and normalising with the sum of all pixel values in the
459 original image thereby calculating the x-coordinate of the "center
460 of mass" of the image (interpreting pixel values as local mass)
461 which is known as the centroid of an image.
462
463 Next is a (from the point of view of memory consumption) very cheap
464 conversion from greyscale to RGB, i.e. every pixel holds now a
465 triple of values instead of a scalar. The three values in the
466 triple are, fortunately, all the same for a grey image, so that our
467 trick works well in that it maps all the three members of the
468 triple to the same source element:
469
470 # a cheap greyscale to RGB conversion
471 $rgb = $grey->dummy(0,3)
472
473 Unfortunately this trick cannot be used to convert your old B/W
474 photos to color ones in the way you'd like. :(
475
476 Note that the memory usage of piddles with dummy dimensions is
477 especially sensitive to the internal representation. If the piddle
478 can be represented as a virtual affine (``vaffine'') piddle, only
479 the control structures are stored. But if $b in
480
481 $a = zeroes(10000);
482 $b = $a->dummy(1,10000);
483
484 is made physical by some routine, you will find that the memory
485 usage of your program has suddenly grown by 100Mb.
486
487 "diagonal"
488 replaces two dimensions (which have to be of equal size) by one
489 dimension that references all the elements along the "diagonal"
490 along those two dimensions. Here, we have two examples which should
491 appear familiar to anyone who has ever done some linear algebra.
492 Firstly, make a unity matrix:
493
494 # unity matrix
495 $e = zeroes(float, 3, 3); # make everything zero
496 ($tmp = $e->diagonal(0,1)) .= 1; # set the elements along the diagonal to 1
497 print $e;
498
499 Or the other diagonal:
500
501 ($tmp = $e->slice(':-1:0')->diagonal(0,1)) .= 2;
502 print $e;
503
504 (Did you notice how we used the slice function to revert the
505 sequence of lines before setting the diagonal of the new child,
506 thereby setting the cross diagonal of the parent ?) Or a mapping
507 from the space of diagonal matrices to the field over which the
508 matrices are defined, the trace of a matrix:
509
510 # trace of a matrix
511 $trace = sum($mat->diagonal(0,1)); # sum all the diagonal elements
512
513 "xchg" and "mv"
514 xchg exchanges or "transposes" the two specified dimensions. A
515 straightforward example:
516
517 # transpose a matrix (without explicitly reshuffling data and
518 # making a copy)
519 $prod = $a x $a->xchg(0,1);
520
521 $prod should now be pretty close to the unity matrix if $a is an
522 orthogonal matrix. Often "xchg" will be used in the context of
523 threading but more about that later.
524
525 mv works in a similar fashion. It moves a dimension (specified by
526 its number in the parent) to a new position in the new child pdl:
527
528 $b = $a->mv(4,0); # make the 5th dimension of $a the first in the
529 # new child $b
530
531 The difference between "xchg" and "mv" is that "xchg" only changes
532 the position of two dimensions with each other, whereas "mv"
533 inserts the first dimension to the place of second, moving the
534 other dimensions around accordingly.
535
536 "clump"
537 collapses several dimensions into one. Its only argument specifies
538 how many dimensions of the source pdl should be collapsed (starting
539 from the first). An (admittedly unrealistic) example is a 3D pdl
540 which holds data from a stack of image files that you have just
541 read in. However, the data from each image really represents a 1D
542 time series and has only been arranged that way because it was dig‐
543 itized with a frame grabber. So to have it again as an array of
544 time sequences you say
545
546 perldl> $seqs = $stack->clump(2)
547 perldl> help vars
548 PDL variables in package main::
549
550 Name Type Dimension Flow State Mem
551 ----------------------------------------------------------------
552 $seqs Double D [8000,50] -C 0.00Kb
553 $stack Double D [100,80,50] P 3.05Mb
554
555 Unrealistic as it may seem, our confocal microscope software writes
556 data (sometimes) this way. But more often you use clump to achieve
557 a certain effect when using implicit or explicit threading.
558
559 Calls to indexing functions can be chained
560
561 As you might have noticed in some of the examples above calls to the
562 indexing functions can be nicely chained since all of these functions
563 return a newly created child object. However, when doing extensive
564 index manipulations in a chain be sure to keep track of what you are
565 doing, e.g.
566
567 $a->xchg(0,1)->mv(0,4)
568
569 moves the dimension 1 of $a to position 4 since when the second command
570 is executed the original dimension 1 has been moved to position 0 of
571 the new child that calls the "mv" function. I think you get the idea
572 (in spite of my convoluted explanations).
573
574 Propagated assignments ('.=') and dummy dimensions
575
576 A sublety related to indexing is the assignment to pdls containing
577 dummy dimensions of size greater than 1. These assignments (using ".=")
578 are forbidden since several elements of the lvalue pdl point to the
579 same element of the parent. As a consequence the value of those parent
580 elements are potentially ambiguous and would depend on the sequence in
581 which the implementation makes the assignments to elements. Therefore,
582 an assignment like this:
583
584 $a = pdl [1,2,3];
585 $b = $a->dummy(1,4);
586 $b .= yvals(zeroes(3,4));
587
588 can produce unexpected results and the results are explicitly undefined
589 by PDL because when PDL gets parallel computing features, the current
590 result may well change.
591
592 From the point of view of dataflow the introduction of greater-size-
593 than-one dummy dimensions is regarded as an irreversible transformation
594 (similar to the terminology in thermodynamics) which precludes backward
595 propagation of assignment to a parent (which you had explicitly
596 requested using the ".=" assignment). A similar problem to watch out
597 for occurs in the context of threading where sometimes dummy dimensions
598 are created implicitly during the thread loop (see below).
599
600 Reasons for the parent/child (or "pointer") concept
601
602 [ this will have to wait a bit ]
603
604 XXXXX being memory efficient
605 XXXXX in the context of threading
606 XXXXX very flexible and powerful way of accessing portions of pdl data
607 (in much more general way than sec, etc allow)
608 XXXXX efficient implementation
609 XXXXX difference to section/at, etc.
610
611 How to make things physical again
612
613 [ XXXXX fill in later when everything has settled a bit more ]
614
615 ** When needed (xsub routine interfacing C lib function)
616 ** How achieved (->physical)
617 ** How to test (isphysical (explain how it works currently))
618 ** ->copy and ->sever
619
621 In the previous paragraph on indexing we have already mentioned the
622 term occasionally but now its really time to talk explicitly about
623 "threading" with pdls. The term threading has many different meanings
624 in different fields of computing. Within the framework of PDL it could
625 probably be loosely defined as an implicit looping facility. It is
626 implicit because you don't specify anything like enclosing for-loops
627 but rather the loops are automatically (or 'magically') generated by
628 PDL based on the dimensions of the pdls involved. This should give you
629 a first idea why the index/dimension manipulating functions you have
630 met in the previous paragraphs are especially important and useful in
631 the context of threading. The other ingredient for threading (apart
632 from the pdls involved) is a function that is threading aware (gener‐
633 ally, these are PDL::PP compiled functions) and that the pdls are
634 "threaded" over. So much about the terminology and now let's try to
635 shed some light on what it all means.
636
637 Implicit threading - a first example
638
639 There are two slightly different variants of threading. We start with
640 what we call "implicit threading". Let's pick a practical example that
641 involves looping of a function over many elements of a pdl. Suppose we
642 have an RGB image that we want to convert to greyscale. The RGB image
643 is represented by a 3-dim pdl "im(3,x,y)" where the first dimension
644 contains the three color components of each pixel and "x" and "y" are
645 width and height of the image, respectively. Next we need to specify
646 how to convert a color-triple at a given pixel into a greyvalue (to be
647 a realistic example it should represent the relative intensity with
648 which our color insensitive eye cells would detect that color to
649 achieve what we would call a natural conversion from color to
650 greyscale). An approximation that works quite well is to compute the
651 grey intensity from each RGB triplet (r,g,b) as a weighted sum
652
653 greyvalue = 77/256*r + 150/256*g + 29/256*b =
654 inner([77,150,29]/256, [r,g,b])
655
656 where the last form indicates that we can write this as an inner prod‐
657 uct of the 3-vector comprising the weights for red, green and blue com‐
658 ponents with the 3-vector containing the color components. Tradition‐
659 ally, we might have written a function like the following to process
660 the whole image:
661
662 my @dims=$im->dims;
663 # here normally check that first dim has correct size (3), etc
664 $grey=zeroes(@dims[1,2]); # make the pdl for the resulting grey image
665 $w = pdl [77,150,29] / 256; # the vector of weights
666 for ($j=0;$j<dims[2];$j++) {
667 for ($i=0;$i<dims[1];$i++) {
668 # compute the pixel value
669 $tmp = inner($w,$im->slice(':,(i),(j)'));
670 set($grey,$i,$j,$tmp); # and set it in the greyscale image
671 }
672 }
673
674 Now we write the same using threading (noting that "inner" is a thread‐
675 ing aware function defined in the PDL::Primitive package)
676
677 $grey = inner($im,pdl([77,150,29]/256));
678
679 We have ended up with a one-liner that automatically creates the pdl
680 $grey with the right number and size of dimensions and performs the
681 loops automatically (these loops are implemented as fast C code in the
682 internals of PDL). Well, we still owe you an explanation how this
683 'magic' is achieved.
684
685 How does the example work ?
686
687 The first thing to note is that every function that is threading aware
688 (these are without exception functions compiled from concise descrip‐
689 tions by PDL::PP, later just called PP-functions) expects a defined
690 (minimum) number of dimensions (we call them core dimensions) from each
691 of its pdl arguments. The inner function expects two one-dimensional
692 (input) parameters from which it calculates a zero-dimensional (output)
693 parameter. We write that symbolically as "inner((n),(n),[o]())" and
694 call it "inner"'s signature, where n represents the size of that dimen‐
695 sion. n being equal in the first and second parameter means that those
696 dimensions have to be of equal size in any call. As a different example
697 take the outer product which takes two 1D vectors to generate a 2D
698 matrix, symbolically written as "outer((n),(m),[o](n,m))". The "[o]" in
699 both examples indicates that this (here third) argument is an output
700 argument. In the latter example the dimensions of first and second
701 argument don't have to agree but you see how they determine the size of
702 the two dimensions of the output pdl.
703
704 Here is the point when threading finally enters the game. If you call
705 PP-functions with pdls that have more than the required core dimensions
706 the first dimensions of the pdl arguments are used as the core dimen‐
707 sions and the additional extra dimensions are threaded over. Let us
708 demonstrate this first with our example above
709
710 $grey = inner($im,$w); # w is the weight vector from above
711
712 In this case $w is 1D and so supplied just the core dimension, $im is
713 3D, more specifically "(3,x,y)". The first dimension (of size 3) is the
714 required core dimension that matches (as required by inner) the first
715 (and only) dimension of $w. The second dimension is the first thread
716 dimension (of size "x") and the third is here the second thread dimen‐
717 sion (of size "y"). The output pdl is automatically created (as
718 requested by setting $grey to "null" prior to invocation). The output
719 dimensions are obtained by appending the loop dimensions (here "(x,y)")
720 to the core output dimensions (here 0D) to yield the final dimensions
721 of the autocreated pdl (here "0D+2D=2D" to yield a 2D output of size
722 "(x,y)").
723
724 So the above command calls the core functioniality that computes the
725 inner product of two 1D vectors "x*y" times with $w and all 1D slices
726 of the form "(':,(i),(j)')" of $im and sets the respective elements of
727 the output pdl "$grey(i,j)" to the result of each computation. We could
728 write that symbolically as
729
730 $grey(0,0) = f($w,$im(:,(0),(0)))
731 $grey(1,0) = f($w,$im(:,(1),(0)))
732 .
733 .
734 .
735 $grey(x-2,y-1) = f($w,$im(:,(x-2),(y-1)))
736 $grey(x-1,y-1) = f($w,$im(:,(x-1),(y-1)))
737
738 But this is done automatically by PDL without writing any explicit perl
739 loops. We see that the command really creates an output pdl with the
740 right dimensions and sets the elements indeed to the result of the com‐
741 putation for each pixel of the input image.
742
743 When even more pdls and extra dimensions are involved things get a bit
744 more complicated. We will first give the general rules how the thread
745 dimensions depend on the dimensions of input pdls enabling you to fig‐
746 ure out the dimensionality of an autocreated output pdl (for any given
747 set of input pdls and core dimensions of the PP-function in question).
748 The general rules will most likely appear a bit confusing on first
749 sight so that we'll set out to illustrate the usage with a set of fur‐
750 ther examples (which will hopefully also demonstrate that there are
751 indeed many practical situations where threading comes in extremly
752 handy).
753
754 A call for coding discipline
755
756 Before we point out the other technical details of threading, please
757 note this call for programming discipline when using threading:
758
759 In order to preserve human readability, PLEASE comment any nontrivial
760 expression in your code involving threading. Most importantly, for any
761 subroutine, include information at the beginning about what you expect
762 the dimensions to represent (or ranges of dimensions).
763
764 As a warning, look at this undocumented function and try to guess what
765 might be going on:
766
767 sub lookup {
768 my ($im,$palette) = @_;
769 my $res;
770 index($palette->xchg(0,1),
771 $im->long->dummy(0,($palette->dim)[0]),
772 ($res=null));
773 return $res;
774 }
775
776 Would you agree that it might be difficult to figure out expected
777 dimensions, purpose of the routine, etc ? (If you want to find out
778 what this piece of code does, see below)
779
780 How to figure out the loop dimensions
781
782 There are a couple of rules that allow you to figure out number and
783 size of loop dimensions (and if the size of your input pdls comply with
784 the threading rules). Dimensions of any pdl argument are broken down
785 into two groups in the following: Core dimensions (as defined by the
786 PP-function, see Appendix B for a list of PDL primitives) and extra
787 dimensions which comprises all remaining dimensions of that pdl. For
788 example calling a function "func" with the signature
789 "func((n,m),[o](n))" with a pdl "a(2,4,7,1,3)" as "f($a,($o = null))"
790 results in the semantic splitting of a's dimensions into: core dimen‐
791 sions "(2,4)" and extra dimensions "(7,1,3)".
792
793 R0 Core dimensions are identified with the first N dimensions of the
794 respective pdl argument (and are required). Any further dimen‐
795 sions are extra dimensions and used to determine the loop dimen‐
796 sions.
797
798 R1 The number of (implicit) loop dimensions is equal to the maximal
799 number of extra dimensions taken over the set of pdl arguments.
800
801 R2 The size of each of the loop dimensions is derived from the size
802 of the respective dimensions of the pdl arguments. The size of a
803 loop dimension is given by the maximal size found in any of the
804 pdls having this extra dimension.
805
806 R3 For all pdls that have a given extra dimension the size must be
807 equal to the size of the loop dimension (as determined by the
808 previous rule) or 1; otherwise you raise a runtime exception. If
809 the size of the extra dimension in a pdl is one it is implicitly
810 treated as a dummy dimension of size equal to that loop dim size
811 when performing the thread loop.
812
813 R4 If a pdl doesn't have a loop dimension, in the thread loop this
814 pdl is treated as if having a dummy dimension of size equal to
815 the size of that loop dimension.
816
817 R5 If output autocreation is used (by setting the relevant pdl to
818 "PDL->null" before invocation) the number of dimensions of the
819 created pdl is equal to the sum of the number of core output
820 dimensions + number of loop dimensions. The size of the core out‐
821 put dimensions is derived from the relevant dimension of input
822 pdls (as specified in the function definition) and the sizes of
823 the other dimensions are equal to the size of the loop dimension
824 it is derived from. The automatically created pdl will be physi‐
825 cal (unless dataflow is in operation).
826
827 In this context, note that you can run into the problem with assignment
828 to pdls containing greater-than-one dummy dimensions (see above).
829 Although your output pdl(s) didn't contain any dummy dimensions in the
830 first place they may end up with implicitly created dummy dimensions
831 according to R4.
832
833 As an example, suppose we have a (here unspecified) PP-function with
834 the signature:
835
836 func((m,n),(m,n,o),(m),[o](m,o))
837
838 and you call it with 3 pdls "a(5,3,10,11)", "b(5,3,2,10,1,12)", and
839 "c(5,1,11,12)" as
840
841 func($a,$b,$c,($d=null))
842
843 then the number of loop dimensions is 3 (by "R0+R1" from $b and $c)
844 with sizes "(10,11,12)" (by R2); the two output core dimensions are
845 "(5,2)" (from the signature of func) resulting in a 5-dimensional out‐
846 put pdl $c of size "(5,2,10,11,12)" (see R5) and (the automatically
847 created) $d is derived from "($a,$b,$c)" in a way that can be expressed
848 in pdl pseudo-code as
849
850 $d(:,:,i,j,k) .= func($a(:,:,i,j),$b(:,:,:,i,0,k),$c(:,0,j,k))
851 with 0<=i<10, 0<=j<=11, 0<=k<12
852
853 If we analyze the color to greyscale conversion again with these rules
854 in mind we note another great advantage of implicit threading. We can
855 call the conversion with a pdl representing a pixel (im(3)), a line of
856 rgb pixels ("im(3,x)"), a proper color image ("im(3,x,y)") or a whole
857 stack of RGB images ("im(3,x,y,z)"). As long as $im is of the form
858 "(3,...)" the automatically created output pdl will contain the right
859 number of dimensions and contain the intensity data as we exspect it
860 since the loops have been implicitly performed thanks to implicit
861 threading. You can easily convince yourself that calling with a color
862 pixel $grey is 0D, with a line it turns out 1D grey(x), with an image
863 we get "grey(x,y)" and finally we get a converted image stack
864 "grey(x,y,z)".
865
866 Let's fill these general rules with some more life by going through a
867 couple of further examples. The reader may try to figure out equivalent
868 formulations with explicit for-looping and compare the flexibility of
869 those routines using implicit threading to the explicit formulation.
870 Furthermore, especially when using several thread dimensions it is a
871 useful exercise to check the relative speed by doing some benchmark
872 tests (which we still have to do).
873
874 First in the row is a slightly reworked centroid example, now coded
875 with threading in mind.
876
877 # threaded mult to calculate centroid coords, works for stacks as well
878 $xc = sumover(($im*xvals(($im->dims)[0]))->clump(2)) /
879 sumover($im->clump(2));
880
881 Let's analyse what's going on step by step. First the product:
882
883 $prod = $im*xvals(zeroes(($im->dims)[0]))
884
885 This will actually work for $im being one, two, three, and higher
886 dimensional. If $im is one-dimensional it's just an ordinary product
887 (in the sense that every element of $im is multiplied with the respec‐
888 tive element of "xvals(...)"), if $im has more dimensions further
889 threading is done by adding appropriate dummy dimensions to
890 "xvals(...)" according to R4. More importantly, the two sumover oper‐
891 ations show a first example of how to make use of the dimension manipu‐
892 lating commands. A quick look at sumover's signature will remind you
893 that it will only "gobble up" the first dimension of a given input pdl.
894 But what if we want to really compute the sum over all elements of the
895 first two dimensions? Well, nothing keeps us from passing a virtual pdl
896 into sumover which in this case is formed by clumping the first two
897 dimensions of the "parent pdl" into one. From the point of view of the
898 parent pdl the sum is now computed over the first two dimensions, just
899 as we wanted, though sumover has just done the job as specified by its
900 signature. Got it ?
901
902 Another little finesse of writing the code like that: we intentionally
903 used "sumover($pdl->clump(2))" instead of "sum($pdl)" so that we can
904 either pass just an image "(x,y)" or a stack of images "(x,y,t)" into
905 this routine and get either just one x-coordiante or a vector of
906 x-coordinates (of size t) in return.
907
908 Another set of common operations are what one could call "projection
909 operations". These operations take a N-D pdl as input and return a
910 (N-1)-D "projected" pdl. These operations are often performed with
911 functions like sumover, prodover, minimum and maximum. Using again
912 images as examples we might want to calculate the maximum pixel value
913 for each line of an image or image stack. We know how to do that
914
915 # maxima of lines (as function of line number and time)
916 maximum($stack,($ret=null));
917
918 But what if you want to calculate maxima per column when implicit
919 threading always applies the core functionality to the first dimension
920 and threads over all others? How can we achieve that instead the core
921 functionality is applied to the second dimension and threading is done
922 over the others. Can you guess it? Yes, we make a virtual pdl that has
923 the second dimension of the "parent pdl" as its first dimension using
924 the "mv" command.
925
926 # maxima of columns (as function of column number and time)
927 maximum($stack->mv(0,1),($ret=null));
928
929 and calculating all the sums of sub-slices over the third dimension is
930 now almost too easy
931
932 # sums of pixles in time (assuming time is the third dim)
933 sumover($stack->mv(0,2),($ret=null));
934
935 Finally, if you want to apply the operation to all elements (like max
936 over all elements or sum over all elements) regardless of the dimen‐
937 sions of the pdl in question "clump" comes in handy. As an example look
938 at the definition of "sum" (as defined in "Basic.pm"):
939
940 sub sum {
941 PDL::Primitive::sumover($name->clump(-1),($tmp=null));
942 return $tmp->at(); # return a perl number, not a 0D pdl
943 }
944
945 We have already mentioned that all basic operations support threading
946 and assignment is no exception. So here are a couple of threaded
947 assignments
948
949 perldl> $im = zeroes(byte, 10,20)
950 perldl> $line = exp(-rvals(10)**2/9)
951 # threaded assignment
952 perldl> $im .= $line # set every line of $im to $line
953 perldl> $im2 .= 5 # set every element of $im2 to 5
954
955 By now you probably see how it works and what it does, don't you?
956
957 To finish the examples in this paragraph here is a function to create
958 an RGB image from what is called a palette image. The palette image
959 consists of two parts: an image of indices into a color lookup table
960 and the color lookup table itself. [ describe how it works ] We are
961 going to use a PP-function we haven't encoutered yet in the previous
962 examples. It is the aptly named index function, signature
963 "((n),(),[o]())" (see Appendix B) with the core functionality that
964 "index(pdl (0,2,4,5),2,($ret=null))" will return the element with index
965 2 of the first input pdl. In this case, $ret will contain the value 4.
966 So here is the example:
967
968 # a threaded index lookup to generate an RGB, or RGBA or YMCK image
969 # from a palette image (represented by a lookup table $palette and
970 # an color-index image $im)
971 # you can say just dummy(0) since the rules of threading make it fit
972 perldl> index($palette->xchg(0,1),
973 $im->long->dummy(0,($palette->dim)[0]),
974 ($res=null));
975
976 Let's go through it and explain the steps involved. Assuming we are
977 dealing with an RGB lookup-table $palette is of size "(3,x)". First we
978 exchange the dimensions of the palette so that looping is done over the
979 first dimension of $palette (of size 3 that represent r, g, and b com‐
980 ponents). Now looking at $im, we add a dummy dimension of size equal to
981 the length of the number of components (in the case we are discussing
982 here we could have just used the number 3 since we have 3 color compo‐
983 nents). We can use a dummy dimension since for red, green and blue
984 color components we use the same index from the original image, e.g.
985 assuming a certain pixel of $im had the value 4 then the lookup should
986 produce the triple
987
988 [palette(0,4),palette(1,4),palette(2,4)]
989
990 for the new red, green and blue components of the output image. Hope‐
991 fully by now you have some sort of idea what the above piece of code is
992 supposed to do (it is often actually quite complicated to describe in
993 detail how a piece of threading code works; just go ahead and experi‐
994 ment a bit to get a better feeling for it).
995
996 If you have read the threading rules carefully, then you might have
997 noticed that we didn't have to explicitely state the size of the dummy
998 dimension that we created for $im; when we create it with size 1 (the
999 default) the rules of threading make it automatically fit to the
1000 desired size (by rule R3, in our example the size would be 3 assuming a
1001 palette of size "(3,x)"). Since situations like this do occur often in
1002 practice this is actually why rule R3 has been introduced (the part
1003 that makes dimensions of size 1 fit to the thread loop dim size). So we
1004 can just say
1005
1006 perldl> index($palette->xchg(0,1),$im->long->dummy(0),($res=null));
1007
1008 Again, you can convince yourself that this routine will create the
1009 right output if called with a pixel ($im is 0D), a line ($im is 1D), an
1010 image ($im is 2D), ..., an RGB lookup table (palette is "(3,x)") and
1011 RGBA lookup table (palette is "(4,x)", see e.g. OpenGL). This flexibil‐
1012 ity is achieved by the rules of threading which are made to do the
1013 right thing in most situations.
1014
1015 To wrap it all up once again, the general idea is as follows. If you
1016 want to achieve looping over certain dimensions and have the core func‐
1017 tionality applied to another specified set of dimensions you use the
1018 dimension manipulating commands to create a (or several) virtual pdl(s)
1019 so that from the point of view of the parent pdl(s) you get what you
1020 want (always having the signature of the function in question and R1-R5
1021 in mind!). Easy, isn't it ?
1022
1023 Output autocreation and PP-function calling conventions
1024
1025 At this point we have to divert to some technical detail that has to do
1026 with the general calling conventions of PP-functions and the automatic
1027 creation of output arguments. Basically, there are two ways of invok‐
1028 ing pdl routines, namely
1029
1030 $result = func($a,$b);
1031
1032 and
1033
1034 func($a,$b,$result);
1035
1036 If you are only using implicit threading then the output variable can
1037 be automatically created by PDL. You flag that to the PP-function by
1038 setting the output argument to a special kind of pdl that is returned
1039 from a call to the function "PDL->null" that returns an essentially
1040 "empty" pdl (for those interested in details there is a flag in the C
1041 pdl structure for this). The dimensions of the created pdl are deter‐
1042 mined by the rules of implicit threading: the first dimensions are the
1043 core output dimensions to which the threading dimensions are appended
1044 (which are in turn determined by the dimensions of the input pdls as
1045 described above). So you can say
1046
1047 func($a,$b,($result=PDL->null));
1048
1049 or
1050
1051 $result = func($a,$b)
1052
1053 which are exactly equivalent.
1054
1055 Be warned that you can not use output autocreation when using explicit
1056 threading (for reasons explained in the following section on explicit
1057 threading, the second variant of threading).
1058
1059 In "tight" loops you probably want to avoid the implicit creation of a
1060 temporary pdl in each step of the loop that comes along with the "func‐
1061 tional" style but rather say
1062
1063 # create output pdl of appropriate size only at first invocation
1064 $result = null;
1065 for (0...$n) {
1066 func($a,$b,$result); # in all but the first invocation $result
1067 func2($b); # is defined and has the right size to
1068 # take the output provided $b's dims don't change
1069 twiddle($result,$a); # do something from $result to $a for iteration
1070 }
1071
1072 The take-home message of this section once more: be aware of the limi‐
1073 tation on output creation when using explicit threading.
1074
1075 Explicit threading
1076
1077 Having so far only talked about the first flavour of threading it is
1078 now about time to introduce the second variant. Instead of shuffling
1079 around dimensions all the time and relying on the rules of implicit
1080 threading to get it all right you sometimes might want to specify in a
1081 more explicit way how to perform the thread loop. It is probably not
1082 too surprising that this variant of the game is called explicit thread‐
1083 ing. Now, before we create the wrong impression: it is not either
1084 implicit or explicit; the two flavours do mix. But more about that
1085 later.
1086
1087 The two most used functions with explicit threading are thread and
1088 unthread. We start with an example that illustrates typical usage of
1089 the former:
1090
1091 [ # ** this is the worst possible example to start with ]
1092 # but can be used to show that $mat += $line is different from
1093 # $mat->thread(0) += $line
1094 # explicit threading to add a vector to each column of a matrix
1095 perldl> $mat = zeroes(4,3)
1096 perldl> $line = pdl (3.1416,2,-2)
1097 perldl> ($tmp = $mat->thread(0)) += $line
1098
1099 In this example, "$mat->thread(0)" tells PDL that you want the second
1100 dimension of this pdl to be threaded over first leading to a thread
1101 loop that can be expressed as
1102
1103 for (j=0; j<3; j++) {
1104 for (i=0; i<4; i++) {
1105 mat(i,j) += src(j);
1106 }
1107 }
1108
1109 "thread" takes a list of numbers as arguments which explicitly specify
1110 which dimensions to thread over first. With the introduction of
1111 explicit threading the dimensions of a pdl are conceptually split into
1112 three different groups the latter two of which we have already encoun‐
1113 tered: thread dimensions, core dimensions and extra dimensions.
1114
1115 Conceptually, it is best to think of those dimensions of a pdl that
1116 have been specified in a call to "thread" as being taken away from the
1117 set of normal dimensions and put on a separate stack. So assuming we
1118 have a pdl "a(4,7,2,8)" saying
1119
1120 $b = $a->thread(2,1)
1121
1122 creates a new virtual pdl of dimension "b(4,8)" (which we call the
1123 remaining dims) that also has 2 thread dimensions of size "(2,7)". For
1124 the purposes of this document we write that symbolically as
1125 "b(4,8){2,7}". An important difference to the previous examples where
1126 only implicit threading was used is the fact that the core dimensions
1127 are matched against the remaining dimensions which are not necessarily
1128 the first dimensions of the pdl. We will now specify how the presence
1129 of thread dimensions changes the rules R1-R5 for threadloops (which
1130 apply to the special case where none of the pdl arguments has any
1131 thread dimensions).
1132
1133 T0 Core dimensions are matched against the first n remaining dimen‐
1134 sions of the pdl argument (note the difference to R1). Any further
1135 remaining dimensions are extra dimensions and are used to determine
1136 the implicit loop dimensions.
1137
1138 T1a The number of implicit loop dimensions is equal to the maximal num‐
1139 ber of extra dimensions taken over the set of pdl arguments.
1140
1141 T1b The number of explicit loop dimensions is equal to the maximal num‐
1142 ber of thread dimensions taken over the set of pdl arguments.
1143
1144 T1c The total number of loop dimensions is equal to the sum of explicit
1145 loop dimensions and implicit loop dimensions. In the thread loop,
1146 explicit loop dimensions are threaded over first followed by
1147 implicit loop dimensions.
1148
1149 T2 The size of each of the loop dimensions is derived from the size of
1150 the respective dimensions of the pdl arguments. It is given by the
1151 maximal size found in any pdls having this thread dimension (for
1152 explicit loop dimensions) or extra dimension (for implicit loop
1153 dimensions).
1154
1155 T3 This rule applies to any explicit loop dimension as well as any
1156 implicit loop dimension. For all pdls that have a given
1157 thread/extra dimension the size must be equal to the size of the
1158 respective explicit/implicit loop dimension or 1; otherwise you
1159 raise a runtime exception. If the size of a thread/extra dimension
1160 of a pdl is one it is implicitly treated as a dummy dimension of
1161 size equal to the explicit/implicit loop dimension.
1162
1163 T4 If a pdl doesn't have a thread/extra dimension that corresponds to
1164 an explicit/implicit loop dimension, in the thread loop this pdl is
1165 treated as if having a dummy dimension of size equal to the size of
1166 that loop dimension.
1167
1168 T4a All pdls that do have thread dimensions must have the same number
1169 of thread dimensions.
1170
1171 T5 Output autocreation cannot be used if any of the pdl arguments has
1172 any thread dimensions. Otherwise R5 applies.
1173
1174 The same restrictions apply with regard to implicit dummy dimensions
1175 (created by application of T4) as already mentioned in the section on
1176 implicit threading: if any of the output pdls has an (explicit or
1177 implicitly created) greater-than-one dummy dimension a runtime excep‐
1178 tion will be raised.
1179
1180 Let us demonstrate these rules at work in a generic case. Suppose we
1181 have a (here unspecified) PP-function with the signature:
1182
1183 func((m,n),(m),(),[o](m))
1184
1185 and you call it with 3 pdls "a(5,3,10,11)", "b(3,5,10,1,12)", "c(10)"
1186 and an output pdl "d(3,11,5,10,12)" (which can here not be automati‐
1187 cally created) as
1188
1189 func($a->thread(1,3),$b->thread(0,3),$c,$d->thread(0,1))
1190
1191 From the signature of func and the above call the pdls split into the
1192 following groups of core, extra and thread dimensions (written in the
1193 form "pdl(core dims){thread dims}[extra dims]"):
1194
1195 a(5,10){3,11}[] b(5){3,1}[10,12] c(){}[10] d(5){3,11}[10,12]
1196
1197 With this to help us along (it is in general helpful to write the argu‐
1198 ments down like this when you start playing with threading and want to
1199 keep track of what is going on) we further deduce that the number of
1200 explicit loop dimensions is 2 (by T1b from $a and $b) with sizes
1201 "(3,11)" (by T2); 2 implicit loop dimensions (by T1a from $b and $d) of
1202 size "(10,12)" (by T2) and the elements of are computed from the input
1203 pdls in a way that can be expressed in pdl pseudo-code as
1204
1205 for (l=0;l<12;l++)
1206 for (k=0;k<10;k++)
1207 for (j=0;j<11;j++) effect of treating it as dummy dim (index j)
1208 for (i=0;i<3;i++) ⎪
1209 d(i,j,:,k,l) = func(a(:,i,:,j),b(i,:,k,0,l),c(k))
1210
1211 Uhhmpf, this example was really not easy in terms of bookeeping. It
1212 serves mostly as an example how to figure out what's going on when you
1213 encounter a complicated looking expression. But now it is really time
1214 to show that threading is useful by giving some more of our so called
1215 "practical" examples.
1216
1217 [ The following examples will need some additional explanations in the
1218 future. For the moment please try to live with the comments in the code
1219 fragments. ]
1220
1221 Example 1:
1222
1223 *** inverse of matrix represented by eigvecs and eigvals
1224 ** given a symmetrical matrix M = A^T x diag(lambda_i) x A
1225 ** => inverse M^-1 = A^T x diag(1/lambda_i) x A
1226 ** first $tmp = diag(1/lambda_i)*A
1227 ** then A^T * $tmp by threaded inner product
1228 # index handling so that matrices print correct under pdl
1229 $inv .= $evecs*0; # just copy to get appropriately sized output
1230 $tmp .= $evecs; # initialise, no backpropagation
1231 ($tmp2 = $tmp->thread(0)) /= $evals; # threaded division
1232 # and now a matrix multiplication in disguise
1233 PDL::Primitive::inner($evecs->xchg(0,1)->thread(-1,1),
1234 $tmp->thread(0,-1),
1235 $inv->thread(0,1));
1236 # alternative for matrix mult using implicit threading,
1237 # first xchg only for transpose
1238 PDL::Primitive::inner($evecs->xchg(0,1)->dummy(1),
1239 $tmp->xchg(0,1)->dummy(2),
1240 ($inv=null));
1241
1242 Example 2:
1243
1244 # outer product by threaded multiplication
1245 # stress that we need to do it with explicit call to my_biop1
1246 # when using explicit threading
1247 $res=zeroes(($a->dims)[0],($b->dims)[0]);
1248 my_biop1($a->thread(0,-1),$b->thread(-1,0),$res->(0,1),"*");
1249 # similiar thing by implicit threading with autocreated pdl
1250 $res = $a->dummy(1) * $b->dummy(0);
1251
1252 Example 3:
1253
1254 # different use of thread and unthread to shuffle a number of
1255 # dimensions in one go without lots of calls to ->xchg and ->mv
1256
1257 # use thread/unthread to shuffle dimensions around
1258 # just try it out and compare the child pdl with its parent
1259 $trans = $a->thread(4,1,0,3,2)->unthread;
1260
1261 Example 4:
1262
1263 # calculate a couple of bounding boxes
1264 # $bb will hold BB as [xmin,xmax],[ymin,ymax],[zmin,zmax]
1265 # we use again thread and unthread to shuffle dimensions around
1266 perldl> $bb = zeroes(double, 2,3 );
1267 perldl> minimum($vertices->thread(0)->clump->unthread(1),
1268 $bb->slice('(0),:'));
1269 perldl> maximum($vertices->thread(0)->clump->unthread(1),
1270 $bb->slice('(1),:'));
1271
1272 Example 5:
1273
1274 # calculate a self-ratioed (i.e. self normalized) sequence of images
1275 # uses explicit threading and an implicitly threaded division
1276 $stack = read_image_stack();
1277 # calculate the average (per pixel average) of the first $n+1 images
1278 $aver = zeroes([stack->dims]->[0,1]); # make the output pdl
1279 sumover($stack->slice(":,:,0:$n")->thread(0,1),$aver);
1280 $aver /= ($n+1);
1281 $stack /= $aver; # normalize the stack by doing a threaded divison
1282 # implicit versus explicit
1283 # alternatively calculate $aver with implicit threading and autocreation
1284 sumover($stack->slice(":,:,0:$n")->mv(2,0),($aver=null));
1285 $aver /= ($n+1);
1286 #
1287
1288 Implicit versus explicit threading
1289
1290 In this paragraph we are going to illustrate when explicit threading is
1291 preferrable over implicit threading and vice versa. But then again,
1292 this is probably not the best way of putting the case since you already
1293 know: the two flavours do mix. So, it's more about how to get the best
1294 of both worlds and, anyway, in the best of perl traditions: TIMTOWTDI !
1295
1296 [ Sorry, this still has to be filled in in a later release; either
1297 refer to above examples or choose some new ones ]
1298
1299 Finally, this may be a good place to justify all the technical detail
1300 we have been going on about for a couple of pages: why threading ?
1301
1302 Well, code that uses threading should be (considerably) faster than
1303 code that uses explicit for-loops (or similar perl constructs) to
1304 achieve the same functionality. Especially on supercomputers (with vec‐
1305 tor computing facilities/parallel processing) PDL threading will be
1306 implemented in a way that takes advantage of the additional facilities
1307 of these machines. Furthermore, it is a conceptually simply construct
1308 (though technical details might get involved at times) and can greatly
1309 reduce the syntactical complexity of PDL code (but keep the admonition
1310 for documentation in mind). Once you are comfortable with the threading
1311 way of thinking (and coding) it shouldn't be too difficult to under‐
1312 stand code that somebody else has written than (provided he gave you an
1313 idea what exspected input dimensions are, etc.). As a general tip to
1314 increase the performance of your code: if you have to introduce a loop
1315 into your code try to reformulate the problem so that you can use
1316 threading to perform the loop (as with anything there are exceptions to
1317 this rule of thumb; but the authors of this document tend to think that
1318 these are rare cases ;).
1319
1321 An easy way to define functions that are aware of indexing and thread‐
1322 ing (and the universe and everything)
1323
1324 PDL:PP is part of the PDL distribution. It is used to generate func‐
1325 tions that are aware of indexing and threading rules from very concise
1326 descriptions. It can be useful for you if you want to write your own
1327 functions or if you want to interface functions from an external
1328 library so that they support indexing and threading (and mabe dataflow
1329 as well, see PDL::Dataflow). For further details check PDL::PP.
1330
1332 Affine transformations - a special class of simple and powerful trans‐
1333 formations
1334
1335 [ This is also something to be added in future releases. Do we already
1336 have the general make_affine routine in PDL ? It is possible that we
1337 will reference another appropriate manpage from here ]
1338
1340 signatures of standard PDL::PP compiled functions
1341
1342 A selection of signatures of PDL primitives to show how many dimensions
1343 PP compiled functions gobble up (and therefore you can figure out what
1344 will be threaded over). Most of those functions are the basic ones
1345 defined in "primitive.pd"
1346
1347 # functions in primitive.pd
1348 #
1349 sumover ((n),[o]())
1350 prodover ((n),[o]())
1351 axisvalues ((n)) inplace
1352 inner ((n),(n),[o]())
1353 outer ((n),(m),[o](n,m))
1354 innerwt ((n),(n),(n),[o]())
1355 inner2 ((m),(m,n),(n),[o]())
1356 inner2t ((j,n),(n,m),(m,k),[o]())
1357 index (1D,0D,[o])
1358 minimum (1D,[o])
1359 maximum (1D,[o])
1360 wstat ((n),(n),(),[o],())
1361 assgn ((),())
1362
1363 # basic operations
1364 binary operations ((),(),[o]())
1365 unary operations ((),[o]())
1366
1368 Copyright (C) 1997 Christian Soeller (c.soeller@auckland.ac.nz) & Tuo‐
1369 mas J. Lukka (lukka@fas.harvard.edu). All rights reserved. Although
1370 destined for release as a man page with the standard PDL distribution,
1371 it is not public domain. Permission is granted to freely distribute
1372 verbatim copies of this document provided that no modifications outside
1373 of formatting be made, and that this notice remain intact. You are
1374 permitted and encouraged to use its code and derivatives thereof in
1375 your own source code for fun or for profit as you see fit.
1376
1377
1378
1379perl v5.8.8 2003-05-21 INDEXING(1)