SCALAPACK 2.2.2
LAPACK: Linear Algebra PACKage
Loading...
Searching...
No Matches
PB_CInOutV.c
Go to the documentation of this file.
1/* ---------------------------------------------------------------------
2*
3* -- PBLAS auxiliary routine (version 2.0) --
4* University of Tennessee, Knoxville, Oak Ridge National Laboratory,
5* and University of California, Berkeley.
6* April 1, 1998
7*
8* ---------------------------------------------------------------------
9*/
10/*
11* Include files
12*/
13#include "../pblas.h"
14#include "../PBpblas.h"
15#include "../PBtools.h"
16#include "../PBblacs.h"
17#include "../PBblas.h"
18
19#ifdef __STDC__
20void PB_CInOutV( PBTYP_T * TYPE, char * ROWCOL, Int M, Int N, Int * DESCA,
21 Int K, char * BETA,
22 char * Y, Int IY, Int JY, Int * DESCY, char * YROC,
23 char * * TBETA, char * * YAPTR, Int * DYA,
24 Int * YAFREE, Int * YASUM, Int * YAPBY )
25#else
26void PB_CInOutV( TYPE, ROWCOL, M, N, DESCA, K,
27 BETA, Y, IY, JY, DESCY, YROC,
28 TBETA, YAPTR, DYA, YAFREE, YASUM, YAPBY )
29/*
30* .. Scalar Arguments ..
31*/
32 char * BETA, * ROWCOL, * * TBETA, * YROC;
33 Int * YAPBY, * YAFREE, IY, JY, K, M, N, * YASUM;
34 PBTYP_T * TYPE;
35/*
36* .. Array Arguments ..
37*/
38 Int * DESCA, * DESCY, * DYA;
39 char * Y, * * YAPTR;
40#endif
41{
42/*
43* Purpose
44* =======
45*
46* PB_CInOutV returns a pointer to an array that contains a one-dimen-
47* sional input/output subvector which is replicated over the rows or
48* columns of a submatrix described by DESCA. A subvector is specified
49* on input to this routine that is reused whenever possible. On return,
50* the subvector is specified by a pointer to some data, a descriptor
51* array describing its layout, a logical value indicating if this local
52* piece of data has been dynamically allocated by this function, a lo-
53* gical value specifying if sum reduction should occur, and finally a
54* logical value specifying if it is necessary to copy back the alloca-
55* ted data to the original data. This routine is specifically designed
56* for traditional Level 2 like PBLAS operations using an input/output
57* vector such as PxGEMV, PxSYMV ...
58*
59* Notes
60* =====
61*
62* A description vector is associated with each 2D block-cyclicly dis-
63* tributed matrix. This vector stores the information required to
64* establish the mapping between a matrix entry and its corresponding
65* process and memory location.
66*
67* In the following comments, the character _ should be read as
68* "of the distributed matrix". Let A be a generic term for any 2D
69* block cyclicly distributed matrix. Its description vector is DESC_A:
70*
71* NOTATION STORED IN EXPLANATION
72* ---------------- --------------- ------------------------------------
73* DTYPE_A (global) DESCA[ DTYPE_ ] The descriptor type.
74* CTXT_A (global) DESCA[ CTXT_ ] The BLACS context handle, indicating
75* the NPROW x NPCOL BLACS process grid
76* A is distributed over. The context
77* itself is global, but the handle
78* (the integer value) may vary.
79* M_A (global) DESCA[ M_ ] The number of rows in the distribu-
80* ted matrix A, M_A >= 0.
81* N_A (global) DESCA[ N_ ] The number of columns in the distri-
82* buted matrix A, N_A >= 0.
83* IMB_A (global) DESCA[ IMB_ ] The number of rows of the upper left
84* block of the matrix A, IMB_A > 0.
85* INB_A (global) DESCA[ INB_ ] The number of columns of the upper
86* left block of the matrix A,
87* INB_A > 0.
88* MB_A (global) DESCA[ MB_ ] The blocking factor used to distri-
89* bute the last M_A-IMB_A rows of A,
90* MB_A > 0.
91* NB_A (global) DESCA[ NB_ ] The blocking factor used to distri-
92* bute the last N_A-INB_A columns of
93* A, NB_A > 0.
94* RSRC_A (global) DESCA[ RSRC_ ] The process row over which the first
95* row of the matrix A is distributed,
96* NPROW > RSRC_A >= 0.
97* CSRC_A (global) DESCA[ CSRC_ ] The process column over which the
98* first column of A is distributed.
99* NPCOL > CSRC_A >= 0.
100* LLD_A (local) DESCA[ LLD_ ] The leading dimension of the local
101* array storing the local blocks of
102* the distributed matrix A,
103* IF( Lc( 1, N_A ) > 0 )
104* LLD_A >= MAX( 1, Lr( 1, M_A ) )
105* ELSE
106* LLD_A >= 1.
107*
108* Let K be the number of rows of a matrix A starting at the global in-
109* dex IA,i.e, A( IA:IA+K-1, : ). Lr( IA, K ) denotes the number of rows
110* that the process of row coordinate MYROW ( 0 <= MYROW < NPROW ) would
111* receive if these K rows were distributed over NPROW processes. If K
112* is the number of columns of a matrix A starting at the global index
113* JA, i.e, A( :, JA:JA+K-1, : ), Lc( JA, K ) denotes the number of co-
114* lumns that the process MYCOL ( 0 <= MYCOL < NPCOL ) would receive if
115* these K columns were distributed over NPCOL processes.
116*
117* The values of Lr() and Lc() may be determined via a call to the func-
118* tion PB_Cnumroc:
119* Lr( IA, K ) = PB_Cnumroc( K, IA, IMB_A, MB_A, MYROW, RSRC_A, NPROW )
120* Lc( JA, K ) = PB_Cnumroc( K, JA, INB_A, NB_A, MYCOL, CSRC_A, NPCOL )
121*
122* Arguments
123* =========
124*
125* TYPE (local input) pointer to a PBTYP_T structure
126* On entry, TYPE is a pointer to a structure of type PBTYP_T,
127* that contains type information (See pblas.h).
128*
129* ROWCOL (global input) pointer to CHAR
130* On entry, ROWCOL specifies if this routine should return a
131* row or column subvector replicated over the underlying subma-
132* trix as follows:
133* = 'R' or 'r': A row subvector is returned,
134* = 'C' or 'c': A column subvector is returned.
135*
136* M (global input) INTEGER
137* On entry, M specifies the number of rows of the underlying
138* submatrix described by DESCA. M must be at least zero.
139*
140* N (global input) INTEGER
141* On entry, N specifies the number of columns of the underlying
142* submatrix described by DESCA. N must be at least zero.
143*
144* DESCA (global and local input/output) INTEGER array
145* On entry, DESCA is an integer array of dimension DLEN_. This
146* is the array descriptor for the matrix A. EXCEPTIONALLY, THIS
147* INTERNAL ROUTINE MAY MODIFY DESCA IN ORDER TO MINIMIZE THE
148* AMOUNT OF DATA TO BE MOVED FOR THE VECTOR Y. SEE PxGEMV FOR
149* AN EXAMPLE.
150*
151* K (global input) INTEGER
152* On entry, K specifies the length of the non-distributed di-
153* mension of the subvector sub( Y ). K must be at least zero.
154*
155* BETA (global input) pointer to CHAR
156* On entry, BETA is a scalar the input subvector sub( Y ) must
157* be scaled by.
158*
159* Y (local input) pointer to CHAR
160* On entry, Y is an array of dimension (LLD_Y, Ky), where LLD_Y
161* is at least MAX( 1, Lr( K, IY ) ) when YROC is 'R' or 'r'
162* and MAX( 1, Lr( 1, IY+Ly-1 ) ) otherwise, and, Ky is at
163* least Lc( 1, JY+Ly-1 ) when YROC is 'R' or 'r' and
164* Lc( K, JY ) otherwise. Ly is N when ROWCOL is 'R' or 'r' and
165* M otherwise. Before entry, this array contains the local
166* entries of the matrix Y.
167*
168* IY (global input) INTEGER
169* On entry, IY specifies Y's global row index, which points to
170* the beginning of the submatrix sub( Y ).
171*
172* JY (global input) INTEGER
173* On entry, JY specifies Y's global column index, which points
174* to the beginning of the submatrix sub( Y ).
175*
176* DESCY (global and local input) INTEGER array
177* On entry, DESCY is an integer array of dimension DLEN_. This
178* is the array descriptor for the matrix Y.
179*
180* YROC (global input) pointer to CHAR
181* On entry, YROC specifies the orientation of the subvector
182* sub( Y ). When YROC is 'R' or 'r', sub( Y ) is a row vector,
183* and a column vector otherwise.
184*
185* TBETA (local output) pointer to pointer to CHAR
186* On exit, * TBETA is a scalar to be used locally to scale the
187* data pointed to by * YAPTR, in order to obtain the correct
188* result in the original data sub( Y ).
189*
190* YAPTR (local output) pointer to pointer to CHAR
191* On exit, * YAPTR is an array containing the same data as the
192* subvector sub( Y ) which is replicated over the rows or co-
193* lumns of the underlying matrix as specified by ROWCOL and
194* DESCA.
195*
196* DYA (global and local output) INTEGER array
197* On exit, DYA is a descriptor array of dimension DLEN_ descri-
198* bing the data layout of the data pointed to by * YAPTR.
199*
200* YAFREE (local output) INTEGER
201* On exit, YAFREE specifies if it was possible to reuse the
202* subvector sub( Y ), i.e., if some dynamic memory was alloca-
203* ted for the data pointed to by * YAPTR or not. When YAFREE is
204* zero, no dynamic memory was allocated. Otherwise, some dyna-
205* mic memory was allocated by this function that one MUST re-
206* lease as soon as possible.
207*
208* YASUM (global output) INTEGER
209* On exit, YASUM specifies if a global sum reduction should be
210* performed to obtain the correct sub( Y ). When YASUM is zero,
211* no reduction is to be performed, otherwise reduction should
212* occur.
213*
214* YAPBY (global output) INTEGER
215* On exit, YAPBY specifies if the data pointed to by * YAPTR
216* must be move back onto sub( Y ) to obtain the correct result.
217* When YAPBY is zero, no supplementary data movement is neces-
218* sary, otherwise a data redistribution should occur.
219*
220* -- Written on April 1, 1998 by
221* Antoine Petitet, University of Tennessee, Knoxville 37996, USA.
222*
223* ---------------------------------------------------------------------
224*/
225/*
226* .. Local Scalars ..
227*/
228 Int Acol, Aimb, Ainb, AisD, AisR, Amb, Amp, Anb, Anq, Arow, Ycol,
229 Yii, Yimb, Yimb1, Yinb, Yinb1, YisD, YisR, YisRow, Yjj, Yld,
230 Ymb, Ymp, Ynb, Ynq, Yrow, ctxt, izero=0, nprow, myrow, npcol,
231 mycol;
232/* ..
233* .. Executable Statements ..
234*
235*/
236/*
237* Initialize the output parameters to a default value
238*/
239 *YAFREE = 0;
240 *YASUM = 0;
241 *YAPBY = 0;
242 *YAPTR = NULL;
243 *TBETA = BETA;
244/*
245* Quick return if possible
246*/
247 if( ( M <= 0 ) || ( N <= 0 ) || ( K <= 0 ) )
248 {
249 if( Mupcase( ROWCOL[0] ) == CROW )
250 {
251 PB_Cdescset( DYA, K, N, 1, DESCA[INB_], 1, DESCA[NB_], DESCA[RSRC_],
252 DESCA[CSRC_], DESCA[CTXT_], 1 );
253 }
254 else
255 {
256 PB_Cdescset( DYA, M, K, DESCA[IMB_], 1, DESCA[MB_], 1, DESCA[RSRC_],
257 DESCA[CSRC_], DESCA[CTXT_], DESCA[LLD_] );
258 }
259 return;
260 }
261/*
262* Retrieve process grid information
263*/
264 Cblacs_gridinfo( ( ctxt = DESCY[CTXT_] ), &nprow, &npcol, &myrow, &mycol );
265/*
266* Retrieve sub( Y )'s local information: Yii, Yjj, Yrow, Ycol
267*/
268 Minfog2l( IY, JY, DESCY, nprow, npcol, myrow, mycol, Yii, Yjj, Yrow, Ycol );
269/*
270* Is sub( Y ) distributed or not, replicated or not ?
271*/
272 if( ( YisRow = ( Mupcase( YROC[0] ) == CROW ) ) != 0 )
273 {
274 YisD = ( ( Ycol >= 0 ) && ( npcol > 1 ) );
275 YisR = ( ( Yrow == -1 ) || ( nprow == 1 ) );
276 }
277 else
278 {
279 YisD = ( ( Yrow >= 0 ) && ( nprow > 1 ) );
280 YisR = ( ( Ycol == -1 ) || ( npcol == 1 ) );
281 }
282
283 Aimb = DESCA[IMB_ ]; Ainb = DESCA[INB_ ];
284 Amb = DESCA[MB_ ]; Anb = DESCA[NB_ ];
285 Arow = DESCA[RSRC_]; Acol = DESCA[CSRC_];
286
287 if( Mupcase( ROWCOL[0] ) == CROW )
288 {
289/*
290* Want a row vector
291*/
292 AisR = ( ( Arow < 0 ) || ( nprow == 1 ) );
293
294 if( YisRow )
295 {
296/*
297* It is possible to reuse sub( Y ) iff sub( Y ) is already a row vector.
298*/
299 AisD = ( ( Acol >= 0 ) && ( npcol > 1 ) );
300
301 Yinb = DESCY[INB_]; Ynb = DESCY[NB_];
302 Mfirstnb( Yinb1, N, JY, Yinb, Ynb );
303/*
304* sub( Y ) is aligned with A (reuse condition) iff both operands are not
305* distributed, or both of them are distributed and start in the same process
306* column and either N is smaller than the first blocksize of sub( Y ) and A,
307* or their column blocking factors match.
308*/
309 if( ( !AisD && !YisD ) ||
310 ( ( AisD && YisD ) &&
311 ( ( Acol == Ycol ) &&
312 ( ( ( Ainb >= N ) && ( Yinb1 >= N ) ) ||
313 ( ( Ainb == Yinb1 ) && ( Anb == Ynb ) ) ) ) ) )
314 {
315 Mnumroc( Ynq, N, 0, Yinb1, Ynb, mycol, Ycol, npcol );
316 Ymp = ( YisR ? K : ( ( myrow == Yrow ) ? K : 0 ) );
317
318 if( YisR )
319 {
320/*
321* If sub( Y ) is replicated, there is no need to move sub( Y ) after the
322* operation (*YAPBY = 0), and it can be reused.
323*/
324 *YAPBY = 0;
325 Yld = DESCY[ LLD_ ];
326 if( Ynq > 0 )
327 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
328
329 if( AisR )
330 {
331/*
332* If A is replicated as well, use BETA in every process row, and do not combine
333* the local results.
334*/
335 *TBETA = BETA;
336 *YASUM = 0;
337 }
338 else
339 {
340/*
341* Otherwise, use BETA in process row Arow and zero elsewhere. Reduce the local
342* result if there is more than one row in the process grid.
343*/
344 *TBETA = ( ( myrow == Arow ) ? BETA : TYPE->zero );
345 *YASUM = ( nprow > 1 );
346/*
347* If some process rows do not own any entries of A, better set sub( Y ) to zero
348* in those processes.
349*/
350 Mnumroc( Amp, M, 0, Aimb, Amb, myrow, Arow, nprow );
351 if( Amp <= 0 )
352 TYPE->Ftzscal( C2F_CHAR( ALL ), &K, &Ynq, &izero, *TBETA,
353 *YAPTR, &Yld );
354 }
355 }
356 else
357 {
358/*
359* sub( Y ) is not replicated, the descriptor of A may need to be modified ...
360*/
361 if( AisR )
362 {
363/*
364* If A is replicated, use only the copy in the process row where sub( Y )
365* resides -> modify DESCA !!!
366*/
367 *TBETA = BETA;
368 *YASUM = 0;
369 *YAPBY = 0;
370 Yld = DESCY[ LLD_ ];
371 DESCA[ IMB_ ] = M;
372 DESCA[ RSRC_ ] = Yrow;
373 if( ( Ynq > 0 ) && ( Ymp > 0 ) )
374 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
375 }
376 else
377 {
378 if( Mspan( M, 0, Aimb, Amb, Arow, nprow ) )
379 {
380/*
381* Otherwise, A is not replicated, let assume in addition that it spans more
382* than one process row
383*/
384 *YASUM = ( nprow > 1 );
385 *YAPBY = 0;
386
387 if( myrow == Yrow )
388 {
389/*
390* Reuse sub( Y ). If there is no entries of A in the process row where sub( Y )
391* resides, better scale it by BETA immediately.
392*/
393 *TBETA = BETA;
394 Yld = DESCY[ LLD_ ];
395 if( Ynq > 0 )
396 {
397 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
398 Mnumroc( Amp, M, 0, Aimb, Amb, myrow, Arow, nprow );
399 if( Amp <= 0 )
400 TYPE->Ftzscal( C2F_CHAR( ALL ), &K, &Ynq, &izero,
401 *TBETA, *YAPTR, &Yld );
402 }
403 }
404 else
405 {
406/*
407* Allocate space in the other process rows and initialize to zero.
408*/
409 *TBETA = TYPE->zero;
410 Yld = MAX( 1, K );
411 if( Ynq > 0 )
412 {
413 *YAPTR = PB_Cmalloc( K * Ynq * TYPE->size );
414 *YAFREE = 1;
415 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ),
416 &K, &Ynq, &izero, *TBETA, *TBETA,
417 *YAPTR, &Yld );
418 }
419 }
420 }
421 else
422 {
423/*
424* A spans only one process row
425*/
426 if( Yrow == Arow )
427 {
428/*
429* A and sub( Y ) resides in the same process row
430*/
431 *TBETA = BETA;
432 *YASUM = 0;
433 *YAPBY = 0;
434 Yld = DESCY[ LLD_ ];
435 if( ( myrow == Yrow ) && ( Ynq > 0 ) )
436 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
437 }
438 else
439 {
440/*
441* If sub( Y ) resides in another process row, then allocate zero-data in
442* process row where A resides, and set *YAPBY to 1, so that this data will be
443* added (moved) after the local operation has been performed.
444*/
445 *TBETA = TYPE->zero;
446 *YASUM = 0;
447 *YAPBY = 1;
448 Yrow = Arow;
449 Yld = MAX( 1, K );
450 if( myrow == Arow )
451 {
452 if( Ynq > 0 )
453 {
454 *YAPTR = PB_Cmalloc( K * Ynq * TYPE->size );
455 *YAFREE = 1;
456 TYPE->Ftzpad( C2F_CHAR( ALL ),
457 C2F_CHAR( NOCONJG ), &K, &Ynq,
458 &izero, *TBETA, *TBETA, *YAPTR,
459 &Yld );
460 }
461 }
462 }
463 }
464 }
465 }
466/*
467* Describe the resulting operand. Note that when reduction should occur, Yrow
468* contains the destination row. Assuming every process row needs the result,
469* Yrow is then -1.
470*/
471 MDescSet( DYA, K, N, K, Yinb1, 1, Ynb, Yrow, Ycol, ctxt, Yld );
472 return;
473 }
474 }
475/*
476* sub( Y ) cannot be reused, set TBETA to zero for the local operation, and
477* force YAPBY to 1 for the later update of sub( Y ).
478*/
479 *TBETA = TYPE->zero;
480 *YAPBY = 1;
481 Mnumroc( Anq, N, 0, Ainb, Anb, mycol, Acol, npcol );
482 Yld = MAX( 1, K );
483
484 if( YisR )
485 {
486/*
487* If sub( Y ) is replicated, allocate space in every process row owning some
488* columns of A and initialize it to zero. There may be some wasted space
489* (suppose A was residing in just one row), however, it is hoped that moving
490* back this data to sub( Y ) will then be cheaper ...
491*/
492 *YASUM = ( AisR ? 0 : ( nprow > 1 ) );
493 Arow = -1;
494 if( Anq > 0 )
495 {
496 *YAPTR = PB_Cmalloc( K * Anq * TYPE->size );
497 *YAFREE = 1;
498 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &K, &Anq,
499 &izero, *TBETA, *TBETA, *YAPTR, &Yld );
500 }
501 }
502 else
503 {
504/*
505* sub( Y ) resides within only one process row
506*/
507 if( AisR )
508 {
509/*
510* If A is replicated, then modify sub( A ) so that only one process row will
511* compute the result before moving it back to sub( Y ).
512*/
513 *YASUM = 0;
514 DESCA[ IMB_ ] = M;
515
516 if( YisRow )
517 {
518/*
519* Choose different process row than Yrow for better performance (more links)
520* of the later move-back phase.
521*/
522 DESCA[RSRC_] = Arow = MModSub1( Yrow, nprow );
523 }
524 else
525 {
526 DESCA[RSRC_] = Arow = 0;
527 }
528 if( ( myrow == Arow ) && ( Anq > 0 ) )
529 {
530 *YAPTR = PB_Cmalloc( K * Anq * TYPE->size );
531 *YAFREE = 1;
532 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &K, &Anq,
533 &izero, *TBETA, *TBETA, *YAPTR, &Yld );
534 }
535 }
536 else
537 {
538 if( Mspan( M, 0, Aimb, Amb, Arow, nprow ) )
539 {
540/*
541* If A is not replicated, and spans more than just one process row, then
542* allocate space in every process row and zero it.
543*/
544 *YASUM = ( nprow > 1 );
545 if( Anq > 0 )
546 {
547 *YAPTR = PB_Cmalloc( K * Anq * TYPE->size );
548 *YAFREE = 1;
549 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &K, &Anq,
550 &izero, *TBETA, *TBETA, *YAPTR, &K );
551 }
552 }
553 else
554 {
555/*
556* If A is not replicated, and spans only one process row, then allocate space
557* within that process row and zero it.
558*/
559 *YASUM = 0;
560 if( ( myrow == Arow ) && ( Anq > 0 ) )
561 {
562 *YAPTR = PB_Cmalloc( K * Anq * TYPE->size );
563 *YAFREE = 1;
564 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &K, &Anq,
565 &izero, *TBETA, *TBETA, *YAPTR, &K );
566 }
567 }
568 }
569 }
570/*
571* Describe the resulting operand. Note that when reduction should occur, Arow
572* contains the destination row. Assuming every process row needs the result,
573* Arow is then -1.
574*/
575 MDescSet( DYA, K, N, K, Ainb, 1, Anb, Arow, Acol, ctxt, Yld );
576 }
577 else
578 {
579/*
580* Want a column vector
581*/
582 AisR = ( ( Acol < 0 ) || ( npcol == 1 ) );
583
584 if( !YisRow )
585 {
586/*
587* It is possible to reuse sub( Y ) iff sub( Y ) is already a column vector.
588*/
589 AisD = ( ( Arow >= 0 ) && ( nprow > 1 ) );
590
591 Yimb = DESCY[IMB_]; Ymb = DESCY[MB_];
592 Mfirstnb( Yimb1, M, IY, Yimb, Ymb );
593/*
594* sub( Y ) is aligned with A (reuse condition) iff both operands are not
595* distributed, or both of them are distributed and start in the same process
596* row and either M is smaller than the first blocksize of sub( Y ) and A, or
597* their row blocking factors match.
598*/
599 if( ( !AisD && !YisD ) ||
600 ( ( AisD && YisD ) &&
601 ( ( Arow == Yrow ) &&
602 ( ( ( Aimb >= M ) && ( Yimb1 >= M ) ) ||
603 ( ( Aimb == Yimb1 ) && ( Amb == Ymb ) ) ) ) ) )
604 {
605 Mnumroc( Ymp, M, 0, Yimb1, Ymb, myrow, Yrow, nprow );
606 Ynq = ( YisR ? K : ( ( mycol == Ycol ) ? K : 0 ) );
607
608 if( YisR )
609 {
610/*
611* If sub( Y ) is replicated, there is no need to move sub( Y ) after the
612* operation (*YAPBY = 0), and it can be reused.
613*/
614 *YAPBY = 0;
615 Yld = DESCY[ LLD_ ];
616 if( Ymp > 0 )
617 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
618
619 if( AisR )
620 {
621/*
622* If A is replicated as well, use BETA in every process column, and do not
623* combine the local results.
624*/
625 *TBETA = BETA;
626 *YASUM = 0;
627 }
628 else
629 {
630/*
631* Otherwise, use BETA in process column Acol and zero elsewhere. Reduce the
632* local result if there is more than one column in the process grid.
633*/
634 *TBETA = ( ( mycol == Acol ) ? BETA : TYPE->zero );
635 *YASUM = ( npcol > 1 );
636/*
637* If some process columns do not own any entries of A, better set sub( Y ) to
638* zero in those processes.
639*/
640 Mnumroc( Anq, N, 0, Ainb, Anb, mycol, Acol, npcol );
641 if( Anq <= 0 )
642 TYPE->Ftzscal( C2F_CHAR( ALL ), &Ymp, &K, &izero, *TBETA,
643 *YAPTR, &Yld );
644 }
645 }
646 else
647 {
648/*
649* sub( Y ) is not replicated, the descriptor of A may need to be modified ...
650*/
651 if( AisR )
652 {
653/*
654* If A is replicated, use only the copy in the process column where sub( Y )
655* resides -> modify DESCA !!!
656*/
657 *TBETA = BETA;
658 *YASUM = 0;
659 *YAPBY = 0;
660 Yld = DESCY[ LLD_ ];
661 DESCA[ INB_ ] = N;
662 DESCA[ CSRC_ ] = Ycol;
663 if( ( Ymp > 0 ) && ( Ynq > 0 ) )
664 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
665 }
666 else
667 {
668 if( Mspan( N, 0, Ainb, Anb, Acol, npcol ) )
669 {
670/*
671* Otherwise, A is not replicated, let assume in addition that it spans more
672* than one process column
673*/
674 *YASUM = ( npcol > 1 );
675 *YAPBY = 0;
676
677 if( mycol == Ycol )
678 {
679/*
680* Reuse sub( Y ). If there is no entries of A in the process column where
681* sub( Y ) resides, better scale it by BETA immediately.
682*/
683 *TBETA = BETA;
684 Yld = DESCY[ LLD_ ];
685 if( Ymp > 0 )
686 {
687 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
688 Mnumroc( Anq, N, 0, Ainb, Anb, mycol, Acol, npcol );
689 if( Anq <= 0 )
690 TYPE->Ftzscal( C2F_CHAR( ALL ), &Ymp, &K, &izero,
691 *TBETA, *YAPTR, &Yld );
692 }
693 }
694 else
695 {
696/*
697* Allocate space in the other process columns and initialize to zero.
698*/
699 *TBETA = TYPE->zero;
700 Yld = MAX( 1, Ymp );
701 if( Ymp > 0 )
702 {
703 *YAPTR = PB_Cmalloc( Ymp * K * TYPE->size );
704 *YAFREE = 1;
705 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ),
706 &Ymp, &K, &izero, *TBETA, *TBETA,
707 *YAPTR, &Yld );
708 }
709 }
710 }
711 else
712 {
713/*
714* A spans only one process column
715*/
716 if( Ycol == Acol )
717 {
718/*
719* A and sub( Y ) resides in the same process column
720*/
721 *TBETA = BETA;
722 *YASUM = 0;
723 *YAPBY = 0;
724 Yld = DESCY[ LLD_ ];
725 if( ( mycol == Ycol ) && ( Ymp > 0 ) )
726 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
727 }
728 else
729 {
730/*
731* If sub( Y ) resides in another process column, then allocate zero-data in
732* process column where A resides, and set *YAPBY to 1, so that this data will
733* be added (moved) after the local operation has been performed.
734*/
735 *TBETA = TYPE->zero;
736 *YASUM = 0;
737 *YAPBY = 1;
738 Ycol = Acol;
739 Yld = MAX( 1, Ymp ) ;
740 if( mycol == Acol )
741 {
742 if( Ymp > 0 )
743 {
744 *YAPTR = PB_Cmalloc( Ymp * K * TYPE->size );
745 *YAFREE = 1;
746 TYPE->Ftzpad( C2F_CHAR( ALL ),
747 C2F_CHAR( NOCONJG ), &Ymp, &K,
748 &izero, *TBETA, *TBETA, *YAPTR,
749 &Yld );
750 }
751 }
752 }
753 }
754 }
755 }
756/*
757* Describe the resulting operand. Note that when reduction should occur, Ycol
758* contains the destination column. Assuming every process column needs the
759* result, Ycol is then -1.
760*/
761 MDescSet( DYA, M, K, Yimb1, K, Ymb, 1, Yrow, Ycol, ctxt, Yld );
762 return;
763 }
764 }
765/*
766* sub( Y ) cannot be reused, set TBETA to zero for the local operation, and
767* force YAPBY to 1 for the later update of sub( Y ).
768*/
769 *TBETA = TYPE->zero;
770 *YAPBY = 1;
771 Mnumroc( Amp, M, 0, Aimb, Amb, myrow, Arow, nprow );
772 Yld = MAX( 1, Amp );
773
774 if( YisR )
775 {
776/*
777* If sub( Y ) is replicated, allocate space in every process column owning some
778* rows of A and initialize it to zero. There may be some wasted space (suppose
779* A was residing in just one column), however, it is hoped that moving back
780* this data to sub( Y ) will then be cheaper ...
781*/
782 *YASUM = ( AisR ? 0 : ( npcol > 1 ) );
783 Acol = -1;
784 if( Amp > 0 )
785 {
786 *YAPTR = PB_Cmalloc( Amp * K * TYPE->size );
787 *YAFREE = 1;
788 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &Amp, &K,
789 &izero, *TBETA, *TBETA, *YAPTR, &Yld );
790 }
791 }
792 else
793 {
794/*
795* sub( Y ) resides within only one process column
796*/
797 if( AisR )
798 {
799/*
800* If A is replicated, then modify sub( A ) so that only one process column will
801* compute the result before moving it back to sub( Y ).
802*/
803 *YASUM = 0;
804 DESCA[ INB_ ] = N;
805
806 if( YisRow )
807 {
808 DESCA[ CSRC_ ] = Acol = 0;
809 }
810 else
811 {
812/*
813* Choose different process column than Ycol for better performance (more links)
814* of the later move-back phase.
815*/
816 DESCA[ CSRC_ ] = Acol = MModSub1( Ycol, npcol );
817 }
818 if( ( mycol == Acol ) && ( Amp > 0 ) )
819 {
820 *YAPTR = PB_Cmalloc( Amp * K * TYPE->size );
821 *YAFREE = 1;
822 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &Amp, &K,
823 &izero, *TBETA, *TBETA, *YAPTR, &Yld );
824 }
825 }
826 else
827 {
828 if( Mspan( N, 0, Ainb, Anb, Acol, npcol ) )
829 {
830/*
831* If A is not replicated, and spans more than just one process column, then
832* allocate space in every process column and zero it.
833*/
834 *YASUM = ( npcol > 1 );
835 if( Amp > 0 )
836 {
837 *YAPTR = PB_Cmalloc( Amp * K * TYPE->size );
838 *YAFREE = 1;
839 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &Amp, &K,
840 &izero, *TBETA, *TBETA, *YAPTR, &Yld );
841 }
842 }
843 else
844 {
845/*
846* If A is not replicated, and spans only one process column, then allocate
847* space within that process column and zero it.
848*/
849 *YASUM = 0;
850 if( ( mycol == Acol ) && ( Amp > 0 ) )
851 {
852 *YAPTR = PB_Cmalloc( Amp * K * TYPE->size );
853 *YAFREE = 1;
854 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &Amp, &K,
855 &izero, *TBETA, *TBETA, *YAPTR, &Yld );
856 }
857 }
858 }
859 }
860/*
861* Describe the resulting operand. Note that when reduction should occur, Acol
862* contains the destination column. Assuming every process column needs the
863* result, Acol is then -1.
864*/
865 MDescSet( DYA, M, K, Aimb, K, Amb, 1, Arow, Acol, ctxt, Yld );
866 }
867/*
868* End of PB_CInOutV
869*/
870}
#define Int
Definition Bconfig.h:22
#define C2F_CHAR(a)
Definition pblas.h:125
#define CROW
Definition PBblacs.h:21
void Cblacs_gridinfo()
#define ALL
Definition PBblas.h:50
#define NOCONJG
Definition PBblas.h:45
#define CTXT_
Definition PBtools.h:38
#define MAX(a_, b_)
Definition PBtools.h:77
#define MB_
Definition PBtools.h:43
char * PB_Cmalloc()
#define Mspan(n_, i_, inb_, nb_, srcproc_, nprocs_)
Definition PBtools.h:160
#define Mptr(a_, i_, j_, lda_, siz_)
Definition PBtools.h:132
#define LLD_
Definition PBtools.h:47
#define Mnumroc(np_, n_, i_, inb_, nb_, proc_, srcproc_, nprocs_)
Definition PBtools.h:222
#define MDescSet(desc, m, n, imb, inb, mb, nb, rsrc, csrc, ictxt, lld)
Definition PBtools.h:499
void PB_CInOutV()
#define RSRC_
Definition PBtools.h:45
void PB_Cdescset()
#define Mfirstnb(inbt_, n_, i_, inb_, nb_)
Definition PBtools.h:139
#define INB_
Definition PBtools.h:42
#define MModSub1(I, d)
Definition PBtools.h:105
#define Minfog2l(i_, j_, desc_, nr_, nc_, r_, c_, ii_, jj_, pr_, pc_)
Definition PBtools.h:428
#define CSRC_
Definition PBtools.h:46
#define IMB_
Definition PBtools.h:41
#define Mupcase(C)
Definition PBtools.h:83
#define NB_
Definition PBtools.h:44
#define TYPE
Definition clamov.c:7