SCALAPACK 2.2.2
LAPACK: Linear Algebra PACKage
Loading...
Searching...
No Matches
PB_CInOutV2.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_CInOutV2( PBTYP_T * TYPE, char * CONJUG, char * ROWCOL, Int M,
21 Int N, Int KA, Int * DESCA, Int K, char * Y, Int IY,
22 Int JY, Int * DESCY, char * YROC, char * * YAPTR,
23 Int * DYA, Int * YAFREE, Int * YASUM, Int * YAPBY )
24#else
25void PB_CInOutV2( TYPE, CONJUG, ROWCOL, M, N, KA, DESCA, K, Y, IY, JY,
26 DESCY, YROC, YAPTR, DYA, YAFREE, YASUM, YAPBY )
27/*
28* .. Scalar Arguments ..
29*/
30 char * CONJUG, * ROWCOL, * YROC;
31 Int * YAPBY, * YAFREE, IY, JY, K, KA, M, N, * YASUM;
32 PBTYP_T * TYPE;
33/*
34* .. Array Arguments ..
35*/
36 Int * DESCA, * DESCY, * DYA;
37 char * Y, * * YAPTR;
38#endif
39{
40/*
41* Purpose
42* =======
43*
44* PB_CInOutV2 returns a pointer to an array that contains a one-dimen-
45* sional input/output subvector which is replicated over the rows or
46* columns of a submatrix described by DESCA. A subvector is specified
47* on input to this routine that is reused whenever possible. On return,
48* the subvector is specified by a pointer to some data, a descriptor
49* array describing its layout, a logical value indicating if this local
50* piece of data has been dynamically allocated by this function, a lo-
51* gical value specifying if sum reduction should occur, and finally a
52* logical value specifying if it is necessary to copy back the alloca-
53* ted data to the original data. This routine is specifically designed
54* for traditional Level 2 like PBLAS operations using an input/output
55* vector such as PxTRSV.
56*
57* Notes
58* =====
59*
60* A description vector is associated with each 2D block-cyclicly dis-
61* tributed matrix. This vector stores the information required to
62* establish the mapping between a matrix entry and its corresponding
63* process and memory location.
64*
65* In the following comments, the character _ should be read as
66* "of the distributed matrix". Let A be a generic term for any 2D
67* block cyclicly distributed matrix. Its description vector is DESC_A:
68*
69* NOTATION STORED IN EXPLANATION
70* ---------------- --------------- ------------------------------------
71* DTYPE_A (global) DESCA[ DTYPE_ ] The descriptor type.
72* CTXT_A (global) DESCA[ CTXT_ ] The BLACS context handle, indicating
73* the NPROW x NPCOL BLACS process grid
74* A is distributed over. The context
75* itself is global, but the handle
76* (the integer value) may vary.
77* M_A (global) DESCA[ M_ ] The number of rows in the distribu-
78* ted matrix A, M_A >= 0.
79* N_A (global) DESCA[ N_ ] The number of columns in the distri-
80* buted matrix A, N_A >= 0.
81* IMB_A (global) DESCA[ IMB_ ] The number of rows of the upper left
82* block of the matrix A, IMB_A > 0.
83* INB_A (global) DESCA[ INB_ ] The number of columns of the upper
84* left block of the matrix A,
85* INB_A > 0.
86* MB_A (global) DESCA[ MB_ ] The blocking factor used to distri-
87* bute the last M_A-IMB_A rows of A,
88* MB_A > 0.
89* NB_A (global) DESCA[ NB_ ] The blocking factor used to distri-
90* bute the last N_A-INB_A columns of
91* A, NB_A > 0.
92* RSRC_A (global) DESCA[ RSRC_ ] The process row over which the first
93* row of the matrix A is distributed,
94* NPROW > RSRC_A >= 0.
95* CSRC_A (global) DESCA[ CSRC_ ] The process column over which the
96* first column of A is distributed.
97* NPCOL > CSRC_A >= 0.
98* LLD_A (local) DESCA[ LLD_ ] The leading dimension of the local
99* array storing the local blocks of
100* the distributed matrix A,
101* IF( Lc( 1, N_A ) > 0 )
102* LLD_A >= MAX( 1, Lr( 1, M_A ) )
103* ELSE
104* LLD_A >= 1.
105*
106* Let K be the number of rows of a matrix A starting at the global in-
107* dex IA,i.e, A( IA:IA+K-1, : ). Lr( IA, K ) denotes the number of rows
108* that the process of row coordinate MYROW ( 0 <= MYROW < NPROW ) would
109* receive if these K rows were distributed over NPROW processes. If K
110* is the number of columns of a matrix A starting at the global index
111* JA, i.e, A( :, JA:JA+K-1, : ), Lc( JA, K ) denotes the number of co-
112* lumns that the process MYCOL ( 0 <= MYCOL < NPCOL ) would receive if
113* these K columns were distributed over NPCOL processes.
114*
115* The values of Lr() and Lc() may be determined via a call to the func-
116* tion PB_Cnumroc:
117* Lr( IA, K ) = PB_Cnumroc( K, IA, IMB_A, MB_A, MYROW, RSRC_A, NPROW )
118* Lc( JA, K ) = PB_Cnumroc( K, JA, INB_A, NB_A, MYCOL, CSRC_A, NPCOL )
119*
120* Arguments
121* =========
122*
123* TYPE (local input) pointer to a PBTYP_T structure
124* On entry, TYPE is a pointer to a structure of type PBTYP_T,
125* that contains type information (See pblas.h).
126*
127* CONJUG (global input) pointer to CHAR
128* On entry, CONJUG specifies if this routine should return
129* the conjugate subvector as follows:
130* = 'N' or 'n': The initial subvector is returned,
131* = 'Z' or 'z': The conjugate subvector is returned.
132*
133* ROWCOL (global input) pointer to CHAR
134* On entry, ROWCOL specifies if this routine should return a
135* row or column subvector replicated over the underlying subma-
136* trix as follows:
137* = 'R' or 'r': A row subvector is returned,
138* = 'C' or 'c': A column subvector is returned.
139*
140* M (global input) INTEGER
141* On entry, M specifies the number of rows of the underlying
142* submatrix described by DESCA. M must be at least zero.
143*
144* N (global input) INTEGER
145* On entry, N specifies the number of columns of the underlying
146* submatrix described by DESCA. N must be at least zero.
147*
148* KA (global input) INTEGER
149* On entry, KA specifies a global row index when ROWCOL is 'R'
150* or 'r' and a global column index otherwise. This index deter-
151* mines a process row or column in which the output subvector
152* contains a copy of the input subvector.
153*
154* DESCA (global and local input/output) INTEGER array
155* On entry, DESCA is an integer array of dimension DLEN_. This
156* is the array descriptor for the matrix A. EXCEPTIONALLY, THIS
157* INTERNAL ROUTINE MAY MODIFY DESCA IN ORDER TO MINIMIZE THE
158* AMOUNT OF DATA TO BE MOVED FOR THE VECTOR Y. SEE PxGEMV FOR
159* AN EXAMPLE.
160*
161* K (global input) INTEGER
162* On entry, K specifies the length of the non-distributed di-
163* mension of the subvector sub( Y ). K must be at least zero.
164*
165* Y (local input) pointer to CHAR
166* On entry, Y is an array of dimension (LLD_Y, Ky), where LLD_Y
167* is at least MAX( 1, Lr( K, IY ) ) when YROC is 'R' or 'r'
168* and MAX( 1, Lr( 1, IY+Ly-1 ) ) otherwise, and, Ky is at
169* least Lc( 1, JY+Ly-1 ) when YROC is 'R' or 'r' and
170* Lc( K, JY ) otherwise. Ly is N when ROWCOL is 'R' or 'r' and
171* M otherwise. Before entry, this array contains the local
172* entries of the matrix Y.
173*
174* IY (global input) INTEGER
175* On entry, IY specifies Y's global row index, which points to
176* the beginning of the submatrix sub( Y ).
177*
178* JY (global input) INTEGER
179* On entry, JY specifies Y's global column index, which points
180* to the beginning of the submatrix sub( Y ).
181*
182* DESCY (global and local input) INTEGER array
183* On entry, DESCY is an integer array of dimension DLEN_. This
184* is the array descriptor for the matrix Y.
185*
186* YROC (global input) pointer to CHAR
187* On entry, YROC specifies the orientation of the subvector
188* sub( Y ). When YROC is 'R' or 'r', sub( Y ) is a row vector,
189* and a column vector otherwise.
190*
191* YAPTR (local output) pointer to pointer to CHAR
192* On exit, * YAPTR is an array containing the same data as the
193* subvector sub( Y ) which is replicated over the rows or co-
194* lumns of the underlying matrix as specified by ROWCOL and
195* DESCA.
196*
197* DYA (global and local output) INTEGER array
198* On exit, DYA is a descriptor array of dimension DLEN_ descri-
199* bing the data layout of the data pointed to by * YAPTR.
200*
201* YAFREE (local output) INTEGER
202* On exit, YAFREE specifies if it was possible to reuse the
203* subvector sub( Y ), i.e., if some dynamic memory was alloca-
204* ted for the data pointed to by * YAPTR or not. When YAFREE is
205* zero, no dynamic memory was allocated. Otherwise, some dyna-
206* mic memory was allocated by this function that one MUST re-
207* lease as soon as possible.
208*
209* YASUM (global output) INTEGER
210* On exit, YASUM specifies if a global sum reduction should be
211* performed to obtain the correct sub( Y ). When YASUM is zero,
212* no reduction is to be performed, otherwise reduction should
213* occur.
214*
215* YAPBY (global output) INTEGER
216* On exit, YAPBY specifies if the data pointed to by * YAPTR
217* must be move back onto sub( Y ) to obtain the correct result.
218* When YAPBY is zero, no supplementary data movement is neces-
219* sary, otherwise a data redistribution should occur.
220*
221* -- Written on April 1, 1998 by
222* Antoine Petitet, University of Tennessee, Knoxville 37996, USA.
223*
224* ---------------------------------------------------------------------
225*/
226/*
227* .. Local Scalars ..
228*/
229 Int Acol, Acoldst, Aimb, Ainb, AisD, AisR, Amb, Amp, Anb, Anq,
230 Arow, Arowdst, Ycol, Yii, Yimb, Yimb1, Yinb, Yinb1, YisD,
231 YisR, YisRow, Yjj, Yld, Ymb, Ymp, Ynb, Ynq, Yrow, ctxt,
232 izero=0, nprow, myrow, npcol, mycol;
233/* ..
234* .. Executable Statements ..
235*
236*/
237/*
238* Initialize the output parameters to a default value
239*/
240 *YAFREE = 0;
241 *YASUM = 0;
242 *YAPBY = 0;
243 *YAPTR = NULL;
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* Figure out in which process row sub( Y ) or a copy of it should be found
295*/
296 Arowdst = PB_Cindxg2p( KA, Aimb, Amb, Arow, Arow, nprow );
297
298 if( YisRow && ( Mupcase( CONJUG[0] ) == CNOCONJG ) )
299 {
300/*
301* It is possible to reuse sub( Y ) iff sub( Y ) is already a row vector and
302* the data does not need to be conjugated.
303*/
304 AisD = ( ( Acol >= 0 ) && ( npcol > 1 ) );
305
306 Yinb = DESCY[INB_]; Ynb = DESCY[NB_];
307 Yinb1 = PB_Cfirstnb( N, JY, Yinb, Ynb );
308/*
309* sub( Y ) is aligned with A (reuse condition) iff both operands are not
310* distributed, or both of them are distributed and start in the same process
311* column and either N is smaller than the first blocksize of sub( Y ) and A,
312* or their column blocking factors match.
313*/
314 if( ( !AisD && !YisD ) ||
315 ( ( AisD && YisD ) &&
316 ( ( Acol == Ycol ) &&
317 ( ( ( Ainb >= N ) && ( Yinb1 >= N ) ) ||
318 ( ( Ainb == Yinb1 ) && ( Anb == Ynb ) ) ) ) ) )
319 {
320 Ynq = PB_Cnumroc( N, 0, Yinb1, Ynb, mycol, Ycol, npcol );
321 Ymp = ( YisR ? K : ( ( myrow == Yrow ) ? K : 0 ) );
322 Yld = MAX( 1, K );
323
324 if( YisR )
325 {
326/*
327* If sub( Y ) is replicated, there is no need to move sub( Y ) after the
328* operation (*YAPBY = 0), and it can be reused where needed and zeroed out
329* elsewhere.
330*/
331 *YASUM = ( AisR ? 0 : ( nprow > 1 ) );
332 *YAPBY = 0;
333 Yld = DESCY[ LLD_ ];
334 if( Ynq > 0 )
335 {
336 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
337 if( !AisR && ( myrow != Arowdst ) )
338 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &K,
339 &Ynq, &izero, TYPE->zero, TYPE->zero, *YAPTR,
340 &Yld );
341 }
342 }
343 else
344 {
345/*
346* sub( Y ) is not replicated, the descriptor of A may need to be modified ...
347*/
348 if( AisR )
349 {
350/*
351* If A is replicated, use only the copy in the process row where sub( Y )
352* resides -> modify DESCA !!!
353*/
354 *YASUM = 0;
355 *YAPBY = 0;
356 Yld = DESCY[ LLD_ ];
357 DESCA[ IMB_ ] = M;
358 DESCA[ RSRC_ ] = Yrow;
359 if( ( Ynq > 0 ) && ( Ymp > 0 ) )
360 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
361 }
362 else
363 {
364 if( PB_Cspan( M, 0, Aimb, Amb, Arow, nprow ) )
365 {
366/*
367* Otherwise, A is not replicated, let assume in addition that it spans more
368* than one process row.
369*/
370 *YASUM = ( nprow > 1 );
371 *YAPBY = 0;
372
373 if( myrow == Yrow )
374 {
375/*
376* If sub( Y ) is not in the desired process row, send it there and zero it.
377* Otherwise, reuse it.
378*/
379 Yld = DESCY[ LLD_ ];
380 if( Ynq > 0 )
381 {
382 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
383 if( Yrow != Arowdst )
384 {
385 TYPE->Cgesd2d( ctxt, K, Ynq, *YAPTR, Yld, Arowdst,
386 mycol );
387 TYPE->Ftzpad( C2F_CHAR( ALL ),
388 C2F_CHAR( NOCONJG ), &K, &Ynq,
389 &izero, TYPE->zero, TYPE->zero,
390 *YAPTR, &Yld );
391 }
392 }
393 }
394 else
395 {
396/*
397* Allocate space in the other process rows and initialize to zero. If sub( Y )
398* was not in the desired process row, receive it.
399*/
400 Yld = MAX( 1, K );
401 if( Ynq > 0 )
402 {
403 *YAPTR = PB_Cmalloc( K * Ynq * TYPE->size );
404 *YAFREE = 1;
405 if( ( Yrow != Arowdst ) && ( myrow == Arowdst ) )
406 TYPE->Cgerv2d( ctxt, K, Ynq, *YAPTR, Yld, Yrow,
407 mycol );
408 else
409 TYPE->Ftzpad( C2F_CHAR( ALL ),
410 C2F_CHAR( NOCONJG ), &K, &Ynq,
411 &izero, TYPE->zero, TYPE->zero,
412 *YAPTR, &Yld );
413 }
414 }
415 }
416 else
417 {
418/*
419* A spans only one process row
420*/
421 if( Yrow == Arow )
422 {
423/*
424* If A and sub( Y ) resides in the same process row, things are easy.
425*/
426 *YASUM = 0;
427 *YAPBY = 0;
428 Yld = DESCY[ LLD_ ];
429 if( ( myrow == Yrow ) && ( Ynq > 0 ) )
430 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
431 }
432 else
433 {
434/*
435* Otherwise, sub( Y ) resides in another process row, thus allocate zero-data
436* in process row where a copy of sub( Y ) is desired, and receive it. Set
437* *YAPBY to 1, so that this data will be added (moved) after the local
438* operation has been performed.
439*/
440 *YASUM = 0;
441 *YAPBY = 1;
442 if( Ynq > 0 )
443 {
444 if( myrow == Yrow )
445 {
446 Yld = DESCY[ LLD_ ];
447 TYPE->Cgesd2d( ctxt, K, Ynq, Mptr( Y, Yii, Yjj,
448 Yld, TYPE->size ), Yld, Arowdst,
449 mycol );
450 }
451 else if( myrow == Arowdst )
452 {
453 Yld = MAX( 1, K );
454 *YAPTR = PB_Cmalloc( K*Ynq*TYPE->size );
455 *YAFREE = 1;
456 TYPE->Cgerv2d( ctxt, K, Ynq, *YAPTR, Yld, Yrow,
457 mycol );
458 }
459 }
460 Yrow = Arowdst;
461 }
462 }
463 }
464 }
465/*
466* Describe the resulting operand. Note that when reduction should occur, Yrow
467* contains the destination row. Assuming every process row needs the result,
468* Yrow is then -1.
469*/
470 PB_Cdescset( DYA, K, N, K, Yinb1, 1, Ynb, Yrow, Ycol, ctxt, Yld );
471 return;
472 }
473 }
474/*
475* sub( Y ) cannot be reused, force YAPBY to 1 for the later update of sub( Y ).
476*/
477 *YAPBY = 1;
478 Anq = PB_Cnumroc( N, 0, Ainb, Anb, mycol, Acol, npcol );
479 Yld = MAX( 1, K );
480
481 if( YisR )
482 {
483/*
484* If sub( Y ) is replicated, allocate space in every process row owning some
485* columns of A and initialize it to zero only where needed. There may be some
486* wasted space (suppose A was residing in just one row), however, it is hoped
487* that moving back this data to sub( Y ) will then be cheaper ...
488*/
489 *YASUM = ( AisR ? 0 : ( nprow > 1 ) );
490 if( Anq > 0 )
491 {
492 *YAPTR = PB_Cmalloc( K * Anq * TYPE->size );
493 *YAFREE = 1;
494 if( ( Arowdst >= 0 ) && ( myrow != Arowdst ) )
495 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &K, &Anq,
496 &izero, TYPE->zero, TYPE->zero, *YAPTR, &Yld );
497 }
498 }
499 else
500 {
501/*
502* sub( Y ) resides in only one process row
503*/
504 if( AisR )
505 {
506/*
507* If A is replicated, then modify sub( A ) so that only one process row will
508* compute the result before moving it back to sub( Y ).
509*/
510 *YASUM = 0;
511 DESCA[ IMB_ ] = M;
512 if( YisRow )
513 {
514/*
515* Choose a different process row than Yrow for better performance (more links)
516* in the later move-back phase.
517*/
518 DESCA[RSRC_] = MModSub1( Yrow, nprow );
519 }
520 else
521 {
522 DESCA[RSRC_] = 0;
523 }
524 if( ( myrow == ( Arowdst = DESCA[RSRC_] ) ) && ( Anq > 0 ) )
525 {
526 *YAPTR = PB_Cmalloc( K * Anq * TYPE->size );
527 *YAFREE = 1;
528 }
529 }
530 else
531 {
532 if( PB_Cspan( M, 0, Aimb, Amb, Arow, nprow ) )
533 {
534/*
535* If A is not replicated, and spans more than just one process row, then
536* allocate space in every process row and zero it where needed.
537*/
538 *YASUM = ( nprow > 1 );
539 if( Anq > 0 )
540 {
541 *YAPTR = PB_Cmalloc( K * Anq * TYPE->size );
542 *YAFREE = 1;
543 if( myrow != Arowdst )
544 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &K,
545 &Anq, &izero, TYPE->zero, TYPE->zero, *YAPTR,
546 &Yld );
547 }
548 }
549 else
550 {
551/*
552* If A is not replicated, and spans only one process row, then allocate space
553* within that process row.
554*/
555 *YASUM = 0;
556 if( ( myrow == Arowdst ) && ( Anq > 0 ) )
557 {
558 *YAPTR = PB_Cmalloc( K * Anq * TYPE->size );
559 *YAFREE = 1;
560 }
561 }
562 }
563 }
564/*
565* Describe the resulting operand. Note that when reduction should occur,
566* Arowdst contains the destination row. Assuming every process row needs the
567* result, Arowdst is then -1.
568*/
569 PB_Cdescset( DYA, K, N, K, Ainb, 1, Anb, Arowdst, Acol, ctxt, Yld );
570/*
571* Move sub( Y ) in the desired processes and with the correct layout
572*/
573 if( YisRow )
574 {
575 PB_Cpaxpby( TYPE, CONJUG, K, N, TYPE->one, Y, IY, JY, DESCY, ROW,
576 TYPE->zero, *YAPTR, 0, 0, DYA, ROW );
577 }
578 else
579 {
580 PB_Cpaxpby( TYPE, CONJUG, N, K, TYPE->one, Y, IY, JY, DESCY, COLUMN,
581 TYPE->zero, *YAPTR, 0, 0, DYA, ROW );
582 }
583 }
584 else
585 {
586/*
587* Want a column vector with original data in col KA
588*/
589 AisR = ( ( Acol < 0 ) || ( npcol == 1 ) );
590/*
591* Figure out in which process column sub( Y ) or a copy of it should be found.
592*/
593 Acoldst = PB_Cindxg2p( KA, Ainb, Anb, Acol, Acol, npcol );
594
595 if( !( YisRow ) && ( Mupcase( CONJUG[0] ) == CNOCONJG ) )
596 {
597/*
598* It is possible to reuse sub( Y ) iff sub( Y ) is already a column vector and
599* the data does not need to be conjugated.
600*/
601 AisD = ( ( Arow >= 0 ) && ( nprow > 1 ) );
602
603 Yimb = DESCY[IMB_]; Ymb = DESCY[MB_];
604 Yimb1 = PB_Cfirstnb( M, IY, Yimb, Ymb );
605/*
606* sub( Y ) is aligned with A (reuse condition) iff both operands are not
607* distributed, or both of them are distributed and start in the same process
608* row and either M is smaller than the first blocksize of sub( Y ) and A, or
609* their row blocking factors match.
610*/
611 if( ( !AisD && !YisD ) ||
612 ( ( AisD && YisD ) &&
613 ( ( Arow == Yrow ) &&
614 ( ( ( Aimb >= M ) && ( Yimb1 >= M ) ) ||
615 ( ( Aimb == Yimb1 ) && ( Amb == Ymb ) ) ) ) ) )
616 {
617 Ymp = PB_Cnumroc( M, 0, Yimb1, Ymb, myrow, Yrow, nprow );
618 Ynq = ( YisR ? K : ( ( mycol == Ycol ) ? K : 0 ) );
619 Yld = MAX( 1, Ymp );
620
621 if( YisR )
622 {
623/*
624* If sub( Y ) is replicated, there is no need to move sub( Y ) after the
625* operation (*YAPBY = 0), and it can be reused where needed and zeroed out
626* elsewhere.
627*/
628 *YASUM = ( AisR ? 0 : ( npcol > 1 ) );
629 *YAPBY = 0;
630 Yld = DESCY[ LLD_ ];
631 if( Ymp > 0 )
632 {
633 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
634 if( !AisR && ( mycol != Acoldst ) )
635 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &Ymp,
636 &K, &izero, TYPE->zero, TYPE->zero, *YAPTR,
637 &Yld );
638 }
639 }
640 else
641 {
642/*
643* sub( Y ) is not replicated, the descriptor of A may need to be modified ...
644*/
645 if( AisR )
646 {
647/*
648* If A is replicated, use only the copy in the process column where sub( Y )
649* resides -> modify DESCA !!!
650*/
651 *YASUM = 0;
652 *YAPBY = 0;
653 Yld = DESCY[ LLD_ ];
654 DESCA[ INB_ ] = N;
655 DESCA[ CSRC_ ] = Ycol;
656 if( ( Ymp > 0 ) && ( Ynq > 0 ) )
657 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
658 }
659 else
660 {
661 if( PB_Cspan( N, 0, Ainb, Anb, Acol, npcol ) )
662 {
663/*
664* Otherwise, A is not replicated, let assume in addition that it spans more
665* than one process column.
666*/
667 *YASUM = ( npcol > 1 );
668 *YAPBY = 0;
669
670 if( mycol == Ycol )
671 {
672/*
673* If sub( Y ) is not in the desired process column, send it there and zero it.
674* Otherwise, reuse it.
675*/
676 Yld = DESCY[ LLD_ ];
677 if( Ymp > 0 )
678 {
679 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
680 if( Ycol != Acoldst )
681 {
682 TYPE->Cgesd2d( ctxt, Ymp, K, *YAPTR, Yld, myrow,
683 Acoldst );
684 TYPE->Ftzpad( C2F_CHAR( ALL ),
685 C2F_CHAR( NOCONJG ), &Ymp, &K,
686 &izero, TYPE->zero, TYPE->zero,
687 *YAPTR, &Yld );
688 }
689 }
690 }
691 else
692 {
693/*
694* Allocate space in the other process columns and initialize to zero. If
695* sub( Y ) was not in the desired process column, receive it.
696*/
697 Yld = MAX( 1, Ymp );
698 if( Ymp > 0 )
699 {
700 *YAPTR = PB_Cmalloc( Ymp * K * TYPE->size );
701 *YAFREE = 1;
702 if( ( Ycol != Acoldst ) && ( mycol == Acoldst ) )
703 TYPE->Cgerv2d( ctxt, Ymp, K, *YAPTR, Yld, myrow,
704 Ycol );
705 else
706 TYPE->Ftzpad( C2F_CHAR( ALL ),
707 C2F_CHAR( NOCONJG ), &Ymp, &K,
708 &izero, TYPE->zero, TYPE->zero,
709 *YAPTR, &Yld );
710 }
711 }
712 }
713 else
714 {
715/*
716* A spans only one process column
717*/
718 if( Ycol == Acol )
719 {
720/*
721* If A and sub( Y ) resides in the same process column, things are easy.
722*/
723 *YASUM = 0;
724 *YAPBY = 0;
725 Yld = DESCY[ LLD_ ];
726 if( ( mycol == Ycol ) && ( Ymp > 0 ) )
727 *YAPTR = Mptr( Y, Yii, Yjj, Yld, TYPE->size );
728 }
729 else
730 {
731/*
732* Otherwise, sub( Y ) resides in another process column, thus allocate
733* zero-data in process column where a copy of sub( Y ) is desired, and receive
734* it. Set *YAPBY to 1, so that this data will be added (moved) after the local
735* operation has been performed.
736*/
737 *YASUM = 0;
738 *YAPBY = 1;
739 if( Ymp > 0 )
740 {
741 if( mycol == Ycol )
742 {
743 Yld = DESCY[ LLD_ ];
744 TYPE->Cgesd2d( ctxt, Ymp, K, Mptr( Y, Yii, Yjj,
745 Yld, TYPE->size ), Yld, myrow,
746 Acoldst );
747 }
748 else if( mycol == Acoldst )
749 {
750 Yld = MAX( 1, Ymp ) ;
751 *YAPTR = PB_Cmalloc( Ymp * K * TYPE->size );
752 *YAFREE = 1;
753 TYPE->Cgerv2d( ctxt, Ymp, K, *YAPTR, Yld, myrow,
754 Ycol );
755 }
756 }
757 Ycol = Acoldst;
758 }
759 }
760 }
761 }
762/*
763* Describe the resulting operand. Note that when reduction should occur, Ycol
764* contains the destination column. Assuming every process column needs the
765* result, Ycol is then -1.
766*/
767 PB_Cdescset( DYA, M, K, Yimb1, K, Ymb, 1, Yrow, Ycol, ctxt, Yld );
768 return;
769 }
770 }
771/*
772* sub( Y ) cannot be reused, force YAPBY to 1 for the later update of sub( Y ).
773*/
774 *YAPBY = 1;
775 Amp = PB_Cnumroc( M, 0, Aimb, Amb, myrow, Arow, nprow );
776 Yld = MAX( 1, Amp );
777
778 if( YisR )
779 {
780/*
781* If sub( Y ) is replicated, allocate space in every process column owning some
782* columns of A and initialize it to zero only where needed. There may be some
783* wasted space (suppose A was residing in just one column), however, it is
784* hoped that moving back this data to sub( Y ) will then be cheaper ...
785*/
786 *YASUM = ( AisR ? 0 : ( npcol > 1 ) );
787 if( Amp > 0 )
788 {
789 *YAPTR = PB_Cmalloc( Amp * K * TYPE->size );
790 *YAFREE = 1;
791 if( ( Acoldst >= 0 ) && ( mycol != Acoldst ) )
792 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &Amp, &K,
793 &izero, TYPE->zero, TYPE->zero, *YAPTR, &Yld );
794 }
795 }
796 else
797 {
798/*
799* sub( Y ) resides in only one process column
800*/
801 if( AisR )
802 {
803/*
804* If A is replicated, then modify sub( A ) so that only one process column will
805* compute the result before moving it back to sub( Y ).
806*/
807 *YASUM = 0;
808 DESCA[ INB_ ] = N;
809 if( YisRow )
810 {
811 DESCA[ CSRC_ ] = 0;
812 }
813 else
814 {
815/*
816* Choose a different process column than Ycol for better performance (more
817* links) in the later move-back phase.
818*/
819 DESCA[ CSRC_ ] = MModSub1( Ycol, npcol );
820 }
821 if( ( mycol == ( Acoldst = DESCA[CSRC_] ) ) && ( Amp > 0 ) )
822 {
823 *YAPTR = PB_Cmalloc( Amp * K * TYPE->size );
824 *YAFREE = 1;
825 }
826 }
827 else
828 {
829 if( PB_Cspan( N, 0, Ainb, Anb, Acol, npcol ) )
830 {
831/*
832* If A is not replicated, and spans more than just one process column, then
833* allocate space in every process column and zero it where needed.
834*/
835 *YASUM = ( npcol > 1 );
836 if( Amp > 0 )
837 {
838 *YAPTR = PB_Cmalloc( Amp * K * TYPE->size );
839 *YAFREE = 1;
840 if( mycol != Acoldst )
841 TYPE->Ftzpad( C2F_CHAR( ALL ), C2F_CHAR( NOCONJG ), &Amp,
842 &K, &izero, TYPE->zero, TYPE->zero, *YAPTR,
843 &Yld );
844 }
845 }
846 else
847 {
848/*
849* If A is not replicated, and spans only one process column, then allocate
850* space within that process column.
851*/
852 *YASUM = 0;
853 if( ( mycol == Acoldst ) && ( Amp > 0 ) )
854 {
855 *YAPTR = PB_Cmalloc( Amp * K * TYPE->size );
856 *YAFREE = 1;
857 }
858 }
859 }
860 }
861/*
862* Describe the resulting operand. Note that when reduction should occur,
863* Acoldst contains the destination column. Assuming every process column needs
864* the result, Acoldst is then -1.
865*/
866 PB_Cdescset( DYA, M, K, Aimb, K, Amb, 1, Arow, Acoldst, ctxt, Yld );
867/*
868* Move sub( Y ) in the desired processes and with the correct layout
869*/
870 if( YisRow )
871 {
872 PB_Cpaxpby( TYPE, CONJUG, K, M, TYPE->one, Y, IY, JY, DESCY, ROW,
873 TYPE->zero, *YAPTR, 0, 0, DYA, COLUMN );
874 }
875 else
876 {
877 PB_Cpaxpby( TYPE, CONJUG, M, K, TYPE->one, Y, IY, JY, DESCY, COLUMN,
878 TYPE->zero, *YAPTR, 0, 0, DYA, COLUMN );
879 }
880 }
881/*
882* End of PB_CInOutV2
883*/
884}
#define Int
Definition Bconfig.h:22
#define C2F_CHAR(a)
Definition pblas.h:125
#define COLUMN
Definition PBblacs.h:45
#define CROW
Definition PBblacs.h:21
#define ROW
Definition PBblacs.h:46
void Cblacs_gridinfo()
#define ALL
Definition PBblas.h:50
#define CNOCONJG
Definition PBblas.h:19
#define NOCONJG
Definition PBblas.h:45
#define CTXT_
Definition PBtools.h:38
Int PB_Cfirstnb()
#define MAX(a_, b_)
Definition PBtools.h:77
#define MB_
Definition PBtools.h:43
char * PB_Cmalloc()
#define Mptr(a_, i_, j_, lda_, siz_)
Definition PBtools.h:132
#define LLD_
Definition PBtools.h:47
Int PB_Cnumroc()
void PB_CInOutV2()
#define RSRC_
Definition PBtools.h:45
void PB_Cdescset()
#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
Int PB_Cindxg2p()
#define Mupcase(C)
Definition PBtools.h:83
#define NB_
Definition PBtools.h:44
void PB_Cpaxpby()
Int PB_Cspan()
#define TYPE
Definition clamov.c:7