1 /* ---------------------------------------------------------------------
2 *
3 *  -- PBLAS 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__
pdasum_(int * N,double * ASUM,double * X,int * IX,int * JX,int * DESCX,int * INCX)20 void pdasum_( int * N, double * ASUM,
21               double * X, int * IX, int * JX, int * DESCX, int * INCX )
22 #else
23 void pdasum_( N, ASUM, X, IX, JX, DESCX, INCX )
24 /*
25 *  .. Scalar Arguments ..
26 */
27    int            * INCX, * IX, * JX, * N;
28    double         * ASUM;
29 /*
30 *  .. Array Arguments ..
31 */
32    int            * DESCX;
33    double         * X;
34 #endif
35 {
36 /*
37 *  Purpose
38 *  =======
39 *
40 *  PDASUM returns the sum of absolute values of the entries of a subvec-
41 *  tor sub( X ),
42 *
43 *  where
44 *
45 *     sub( X ) denotes X(IX,JX:JX+N-1) if INCX = M_X,
46 *                      X(IX:IX+N-1,JX) if INCX = 1 and INCX <> M_X.
47 *
48 *  Notes
49 *  =====
50 *
51 *  A description  vector  is associated with each 2D block-cyclicly dis-
52 *  tributed matrix.  This  vector  stores  the  information  required to
53 *  establish the  mapping  between a  matrix entry and its corresponding
54 *  process and memory location.
55 *
56 *  In  the  following  comments,   the character _  should  be  read  as
57 *  "of  the  distributed  matrix".  Let  A  be a generic term for any 2D
58 *  block cyclicly distributed matrix.  Its description vector is DESC_A:
59 *
60 *  NOTATION         STORED IN       EXPLANATION
61 *  ---------------- --------------- ------------------------------------
62 *  DTYPE_A (global) DESCA[ DTYPE_ ] The descriptor type.
63 *  CTXT_A  (global) DESCA[ CTXT_  ] The BLACS context handle, indicating
64 *                                   the NPROW x NPCOL BLACS process grid
65 *                                   A  is  distributed over. The context
66 *                                   itself  is  global,  but  the handle
67 *                                   (the integer value) may vary.
68 *  M_A     (global) DESCA[ M_     ] The  number of rows in the distribu-
69 *                                   ted matrix A, M_A >= 0.
70 *  N_A     (global) DESCA[ N_     ] The number of columns in the distri-
71 *                                   buted matrix A, N_A >= 0.
72 *  IMB_A   (global) DESCA[ IMB_   ] The number of rows of the upper left
73 *                                   block of the matrix A, IMB_A > 0.
74 *  INB_A   (global) DESCA[ INB_   ] The  number  of columns of the upper
75 *                                   left   block   of   the  matrix   A,
76 *                                   INB_A > 0.
77 *  MB_A    (global) DESCA[ MB_    ] The blocking factor used to  distri-
78 *                                   bute the last  M_A-IMB_A  rows of A,
79 *                                   MB_A > 0.
80 *  NB_A    (global) DESCA[ NB_    ] The blocking factor used to  distri-
81 *                                   bute the last  N_A-INB_A  columns of
82 *                                   A, NB_A > 0.
83 *  RSRC_A  (global) DESCA[ RSRC_  ] The process row over which the first
84 *                                   row of the matrix  A is distributed,
85 *                                   NPROW > RSRC_A >= 0.
86 *  CSRC_A  (global) DESCA[ CSRC_  ] The  process column  over  which the
87 *                                   first column of  A  is  distributed.
88 *                                   NPCOL > CSRC_A >= 0.
89 *  LLD_A   (local)  DESCA[ LLD_   ] The  leading dimension  of the local
90 *                                   array  storing  the  local blocks of
91 *                                   the distributed matrix A,
92 *                                   IF( Lc( 1, N_A ) > 0 )
93 *                                      LLD_A >= MAX( 1, Lr( 1, M_A ) )
94 *                                   ELSE
95 *                                      LLD_A >= 1.
96 *
97 *  Let K be the number of  rows of a matrix A starting at the global in-
98 *  dex IA,i.e, A( IA:IA+K-1, : ). Lr( IA, K ) denotes the number of rows
99 *  that the process of row coordinate MYROW ( 0 <= MYROW < NPROW ) would
100 *  receive if these K rows were distributed over NPROW processes.  If  K
101 *  is the number of columns of a matrix  A  starting at the global index
102 *  JA, i.e, A( :, JA:JA+K-1, : ), Lc( JA, K ) denotes the number  of co-
103 *  lumns that the process MYCOL ( 0 <= MYCOL < NPCOL ) would  receive if
104 *  these K columns were distributed over NPCOL processes.
105 *
106 *  The values of Lr() and Lc() may be determined via a call to the func-
107 *  tion PB_Cnumroc:
108 *  Lr( IA, K ) = PB_Cnumroc( K, IA, IMB_A, MB_A, MYROW, RSRC_A, NPROW )
109 *  Lc( JA, K ) = PB_Cnumroc( K, JA, INB_A, NB_A, MYCOL, CSRC_A, NPCOL )
110 *
111 *  Arguments
112 *  =========
113 *
114 *  N       (global input) INTEGER
115 *          On entry,  N  specifies the length of the subvector sub( X ).
116 *          N must be at least zero.
117 *
118 *  ASUM    (local output) DOUBLE PRECISION
119 *          On exit,  ASUM  specifies  the sum of absolute values of  the
120 *          subvector  sub( X )  only in its scope (See below for further
121 *          details).
122 *
123 *  X       (local input) DOUBLE PRECISION array
124 *          On entry, X is an array of dimension (LLD_X, Kx), where LLD_X
125 *          is   at  least  MAX( 1, Lr( 1, IX ) )  when  INCX = M_X   and
126 *          MAX( 1, Lr( 1, IX+N-1 ) )  otherwise,  and,  Kx  is  at least
127 *          Lc( 1, JX+N-1 )  when  INCX = M_X  and Lc( 1, JX ) otherwise.
128 *          Before  entry,  this array  contains the local entries of the
129 *          matrix X.
130 *
131 *  IX      (global input) INTEGER
132 *          On entry, IX  specifies X's global row index, which points to
133 *          the beginning of the submatrix sub( X ).
134 *
135 *  JX      (global input) INTEGER
136 *          On entry, JX  specifies X's global column index, which points
137 *          to the beginning of the submatrix sub( X ).
138 *
139 *  DESCX   (global and local input) INTEGER array
140 *          On entry, DESCX  is an integer array of dimension DLEN_. This
141 *          is the array descriptor for the matrix X.
142 *
143 *  INCX    (global input) INTEGER
144 *          On entry,  INCX   specifies  the  global  increment  for  the
145 *          elements of  X.  Only two values of  INCX   are  supported in
146 *          this version, namely 1 and M_X. INCX  must not be zero.
147 *
148 *  Further Details
149 *  ===============
150 *
151 *  When  the  result  of  a vector-oriented PBLAS call is a scalar, this
152 *  scalar  is set only within the process scope which owns the vector(s)
153 *  being operated on. Let sub( X ) be a generic term for the input  vec-
154 *  tor(s). Then, the processes owning the correct the answer is determi-
155 *  ned as follows:  if  an  operation involves more than one vector, the
156 *  processes receiving the result will be the union of the following set
157 *  of processes for each vector:
158 *
159 *  If N = 1, M_X = 1 and INCX = 1,  then  one cannot determine if a pro-
160 *  cess  row  or  process column owns the vector operand, therefore only
161 *  the process owning sub( X ) receives the correct result;
162 *
163 *  If  INCX = M_X, then sub( X )  is a vector distributed over a process
164 *  row. Each process in this row receives the result;
165 *
166 *  If  INCX = 1, then  sub( X )  is  a vector distributed over a process
167 *  column. Each process in this column receives the result;
168 *
169 *  -- Written on April 1, 1998 by
170 *     Antoine Petitet, University of Tennessee, Knoxville 37996, USA.
171 *
172 *  ---------------------------------------------------------------------
173 */
174 /*
175 *  .. Local Scalars ..
176 */
177    char           top;
178    int            Xcol, Xi, Xii, Xj, Xjj, Xld, Xnp, Xnq, Xrow, ctxt, info,
179                   mycol, myrow, npcol, nprow;
180 /*
181 *  .. Local Arrays ..
182 */
183    int            Xd[DLEN_];
184 /* ..
185 *  .. Executable Statements ..
186 *
187 */
188    PB_CargFtoC( *IX, *JX, DESCX, &Xi, &Xj, Xd );
189 #ifndef NO_ARGCHK
190 /*
191 *  Test the input parameters
192 */
193    Cblacs_gridinfo( ( ctxt = Xd[CTXT_] ), &nprow, &npcol, &myrow, &mycol );
194    if( !( info = ( ( nprow == -1 ) ? -( 601 + CTXT_ ) : 0 ) ) )
195       PB_Cchkvec( ctxt, "PDASUM", "X", *N, 1, Xi, Xj, Xd, *INCX, 6, &info );
196    if( info ) { PB_Cabort( ctxt, "PDASUM", info ); return; }
197 #endif
198 /*
199 *  Initialize ASUM
200 */
201    *ASUM = ZERO;
202 /*
203 *  Quick return if possible
204 */
205    if( *N == 0 ) return;
206 /*
207 *  Retrieve process grid information
208 */
209 #ifdef NO_ARGCHK
210    Cblacs_gridinfo( ( ctxt = Xd[CTXT_] ), &nprow, &npcol, &myrow, &mycol );
211 #endif
212 /*
213 *  Retrieve sub( X )'s local information: Xii, Xjj, Xrow, Xcol
214 */
215    PB_Cinfog2l( Xi, Xj, Xd, nprow, npcol, myrow, mycol, &Xii, &Xjj,
216                 &Xrow, &Xcol );
217 /*
218 *  Handle degenerate case separately, sub( X )'s scope is just one process
219 */
220    if( ( *N == 1 ) && ( *INCX == 1 ) && ( Xd[M_] == 1 ) )
221    {
222 /*
223 *  Make sure I own some data and compute ASUM
224 */
225       if( ( ( myrow == Xrow ) || ( Xrow < 0 ) ) &&
226           ( ( mycol == Xcol ) || ( Xcol < 0 ) ) )
227       {
228          *ASUM = ABS( X[Xii+Xjj*Xd[LLD_]] );
229       }
230       return;
231    }
232    else if( *INCX == Xd[M_] )
233    {
234 /*
235 *  sub( X ) resides in (a) process row(s)
236 */
237       if( ( myrow == Xrow ) || ( Xrow < 0 ) )
238       {
239 /*
240 *  Make sure I own some data and compute the local sum
241 */
242          Xnq = PB_Cnumroc( *N, Xj, Xd[INB_], Xd[NB_], mycol, Xd[CSRC_], npcol );
243          if( Xnq > 0 )
244          {
245             Xld = Xd[LLD_];
246             dvasum_( &Xnq, ((char *) ASUM), ((char *)( X+(Xii+Xjj*Xld) )),
247                      &Xld );
248          }
249 /*
250 *  If Xnq <= 0, ASUM is zero (see initialization above)
251 */
252          if( ( npcol > 1 ) && ( Xcol >= 0 ) )
253          {
254 /*
255 *  Combine the local results if npcol > 1 and Xcol >= 0, i.e sub( X ) is
256 *  distributed.
257 */
258             top = *PB_Ctop( &ctxt, COMBINE, ROW, TOP_GET );
259             Cdgsum2d( ctxt, ROW, &top, 1, 1, ((char *)ASUM), 1, -1,
260                       mycol );
261          }
262       }
263       return;
264    }
265    else
266    {
267 /*
268 *  sub( X ) resides in (a) process column(s)
269 */
270       if( ( mycol == Xcol ) || ( Xcol < 0 ) )
271       {
272 /*
273 *  Make sure I own some data and compute the local sum
274 */
275          Xnp = PB_Cnumroc( *N, Xi, Xd[IMB_], Xd[MB_], myrow, Xd[RSRC_], nprow );
276          if( Xnp > 0 )
277          {
278             dvasum_( &Xnp, ((char *) ASUM),
279                      ((char *)( X+(Xii+Xjj*Xd[LLD_]) )), INCX );
280          }
281 /*
282 *  If Xnp <= 0, ASUM is zero (see initialization above)
283 */
284          if( ( nprow > 1 ) && ( Xrow >= 0 ) )
285          {
286 /*
287 *  Combine the local results if nprow > 1 and Xrow >= 0, i.e sub( X ) is
288 *  distributed.
289 */
290             top = *PB_Ctop( &ctxt, COMBINE, COLUMN, TOP_GET );
291             Cdgsum2d( ctxt, COLUMN, &top, 1, 1, ((char *)ASUM), 1, -1,
292                       mycol );
293          }
294       }
295       return;
296    }
297 /*
298 *  End of PDASUM
299 */
300 }
301