swap.c

References to this file elsewhere.
1 #ifndef MS_SUA
2 # include <stdio.h>
3 #endif
4 #include <fcntl.h>
5 
6 #define STANDARD_ERROR 2
7 
8 #define STANDARD_OUTPUT 1
9 
10 #include "mpi.h"
11 #include "rsl_lite.h"
12 
13 #define  UP_EVEN(A)   ((A)+abs((A)%2))
14 #define  DOWN_EVEN(A) ((A) - abs((A)%2))
15 #define  UP_ODD(A)    ((A) + abs(((A)+1)%2))
16 #define  DOWN_ODD(A)  ((A) - abs(((A)+1)%2))
17 #define  MIN(A,B)     ((A)<(B)?(A):(B))
18 #define  MAX(A,B)     ((A)>(B)?(A):(B))
19 
20 static int *y_curs = NULL ;
21 static int *x_curs = NULL ;
22 static int *x_peermask = NULL ;
23 static int *nbytes = NULL ; 
24 static MPI_Request *x_recv = NULL , *x_send = NULL ;
25 
26 RSL_LITE_INIT_SWAP ( 
27                 int * Fcomm ,
28                 int * xy0 ,
29                 int * n3dR0, int *n2dR0, int * typesizeR0 , 
30                 int * n3dI0, int *n2dI0, int * typesizeI0 , 
31                 int * n3dD0, int *n2dD0, int * typesizeD0 , 
32                 int * n3dL0, int *n2dL0, int * typesizeL0 , 
33                 int * me0, int * np0 , int * np_x0 , int * np_y0 ,
34                 int * ids0 , int * ide0 , int * jds0 , int * jde0 , int * kds0 , int * kde0 ,
35                 int * ips0 , int * ipe0 , int * jps0 , int * jpe0 , int * kps0 , int * kpe0 )
36 {
37   int n3dR, n2dR, typesizeR ;
38   int n3dI, n2dI, typesizeI ;
39   int n3dD, n2dD, typesizeD ;
40   int n3dL, n2dL, typesizeL ;
41   int xy ;
42   int me, np, np_x, np_y ;
43   int ids , ide , jds , jde , kds , kde ;
44   int ips , ipe , jps , jpe , kps , kpe ;
45   int ips_send , ipe_send ;
46   int npts, i, ii, j, m, n, ps, pe, ops, ope ;
47   int Px, Py, P, coords[2] ;
48   int ips_swap, ipe_swap ;
49   MPI_Comm *comm, dummy_comm ;
50 
51   comm = &dummy_comm ;
52   *comm = MPI_Comm_f2c( *Fcomm ) ;
53 
54   xy = *xy0 ;
55   n3dR = *n3dR0 ; n2dR = *n2dR0 ; typesizeR = *typesizeR0 ;
56   n3dI = *n3dI0 ; n2dI = *n2dI0 ; typesizeI = *typesizeI0 ;
57   n3dD = *n3dD0 ; n2dD = *n2dD0 ; typesizeD = *typesizeD0 ;
58   n3dL = *n3dL0 ; n2dL = *n2dL0 ; typesizeL = *typesizeL0 ;
59   me = *me0 ; np = *np0 ; np_x = *np_x0 ; np_y = *np_y0 ;
60   ids = *ids0-1 ; ide = *ide0-1 ; jds = *jds0-1 ; jde = *jde0-1 ; kds = *kds0-1 ; kde = *kde0-1 ;
61   ips = *ips0-1 ; ipe = *ipe0-1 ; jps = *jps0-1 ; jpe = *jpe0-1 ; kps = *kps0-1 ; kpe = *kpe0-1 ;
62 
63   if ( nbytes == NULL ) nbytes = RSL_MALLOC ( int , np ) ;
64   if ( x_curs == NULL ) x_curs = RSL_MALLOC ( int , np ) ;
65   if ( x_peermask == NULL ) x_peermask = RSL_MALLOC ( int , np ) ;
66   if ( x_recv == NULL ) x_recv = RSL_MALLOC ( MPI_Request , np ) ;
67   if ( x_send == NULL ) x_send = RSL_MALLOC ( MPI_Request , np ) ;
68   for ( i = 0 ; i < np ; i++ ) { nbytes[i] = 0 ; x_curs[i] = 0 ; x_peermask[i] = 0 ; }
69 
70   if ( xy == 1 ) {   /* xy = 1, swap in X, otherwise Y */
71     n = (ide-ids+1)/4*2 ;
72     m = n*2 ;
73     ps = ips ;
74     pe = ipe ;
75     ops = jps ;
76     ope = jpe ;
77   } else {
78     n = (jde-jds+1)/4*2 ;
79     m = n*2 ;
80     ps = jps ;
81     pe = jpe ;
82     ops = ips ;
83     ope = ipe ;
84   }
85 
86   for ( i = UP_ODD( ps ) ; i <= MIN(pe,m) ; i+=2 ) {
87     ii = abs(i+n) % m ;
88     if ( xy == 1 ) {
89       TASK_FOR_POINT ( &ii , &jps , &ids, &ide , &jds, &jde , &np_x , &np_y , &Px, &Py ) ;
90       coords[1] = Px ; coords[0] = Py ;
91       MPI_Cart_rank( *comm, coords, &P ) ;
92     } else {
93       TASK_FOR_POINT ( &ips , &ii , &ids, &ide , &jds, &jde , &np_x , &np_y , &Px, &Py ) ;
94       coords[1] = Px ; coords[0] = Py ;
95       MPI_Cart_rank( *comm, coords, &P ) ;
96     }
97     nbytes[P] += typesizeR*(ope-ops+1)*(n3dR*(kpe-kps+1)+n2dR) +
98                  typesizeI*(ope-ops+1)*(n3dI*(kpe-kps+1)+n2dI) +
99                  typesizeD*(ope-ops+1)*(n3dD*(kpe-kps+1)+n2dD) +
100                  typesizeL*(ope-ops+1)*(n3dL*(kpe-kps+1)+n2dL) ;
101     x_peermask[P] = 1 ;
102   }
103 
104   for ( P = 0 ; P < np ; P++ ) {
105      if ( x_peermask[P] ) {
106        buffer_for_proc ( P , nbytes[P], RSL_RECVBUF ) ;
107        buffer_for_proc ( P , nbytes[P], RSL_SENDBUF ) ;
108      }
109   }
110 }
111 
112 RSL_LITE_PACK_SWAP ( int * Fcomm , char * buf , int * odd0 , int * typesize0 , int * xy0 , int * pu0 , char * memord , int * xstag0 ,
113            int *me0, int * np0 , int * np_x0 , int * np_y0 , 
114            int * ids0 , int * ide0 , int * jds0 , int * jde0 , int * kds0 , int * kde0 ,
115            int * ims0 , int * ime0 , int * jms0 , int * jme0 , int * kms0 , int * kme0 ,
116            int * ips0 , int * ipe0 , int * jps0 , int * jpe0 , int * kps0 , int * kpe0 )
117 {
118   int me, np, np_x, np_y ;
119   int odd , typesize ;
120   int ids , ide , jds , jde , kds , kde ;
121   int ims , ime , jms , jme , kms , kme ;
122   int ips , ipe , jps , jpe , kps , kpe ;
123   int xstag ;  /* 0 not stag, 1 stag */
124   int xy ;   /* y = 0 , x = 1 */
125   int pu ;   /* pack = 0 , unpack = 1 */
126   int i, ii, j, jj, m, n  ;
127   int ps, pe, ops, ope ;
128   register int k, t ;
129 #ifdef crayx1
130   register int i2,i3,i4,i_offset;
131 #endif
132   char *p ;
133   int da_buf ;
134   int Px, Py, P, coords[2] ;
135   int ierr = 0 ;
136   register int *pi, *qi ;
137   float f ;
138   MPI_Comm *comm, dummy_comm ;
139 
140   comm = &dummy_comm ;
141   *comm = MPI_Comm_f2c( *Fcomm ) ;
142 
143   me = *me0 ; np = *np0 ; np_x = *np_x0 ; np_y = *np_y0 ;
144   xstag = *xstag0 ;
145   odd = *odd0 ; typesize = *typesize0 ;
146   ids = *ids0-1 ; ide = *ide0-1 ; jds = *jds0-1 ; jde = *jde0-1 ; kds = *kds0-1 ; kde = *kde0-1 ;
147   ims = *ims0-1 ; ime = *ime0-1 ; jms = *jms0-1 ; jme = *jme0-1 ; kms = *kms0-1 ; kme = *kme0-1 ;
148   ips = *ips0-1 ; ipe = *ipe0-1 ; jps = *jps0-1 ; jpe = *jpe0-1 ; kps = *kps0-1 ; kpe = *kpe0-1 ;
149   xy = *xy0 ;
150   pu = *pu0 ;
151 
152 /* need to adapt for other memory orders */
153 #define RANGE(S1,E1,S2,E2,S3,E3,S4,E4) (((E1)-(S1)+1)*((E2)-(S2)+1)*(((E3)-(S3)+1)/2)*((E4)-(S4)+1))
154 #define IMAX(A) (((A)>ids)?(A):ids)
155 #define IMIN(A) (((A)<ide)?(A):ide)
156 #define JMAX(A) (((A)>jds)?(A):jds)
157 #define JMIN(A) (((A)<jde)?(A):jde)
158 
159   da_buf = ( pu == 0 ) ? RSL_SENDBUF : RSL_RECVBUF ;
160 
161 
162   if ( xy == 1 ) {   /* xy = 1, swap in X, otherwise Y */
163     n = (ide-ids+1)/4*2 ;
164     m = n*2 ;
165   } else {
166     n = (jde-jds+1)/4*2 ;
167     m = n*2 ;
168   }
169 
170   if ( np_x > 1 && xy == 1 ) {
171 
172     for ( i = UP_ODD(ips) ; i <= MIN(ipe,m) ; i+=2 ) {
173       ii = abs(i+n) % m ;
174       TASK_FOR_POINT ( &ii , &jps , &ids, &ide , &jds, &jde , &np_x , &np_y , &Px, &Py ) ;
175       coords[1] = Px ; coords[0] = Py ;
176       MPI_Cart_rank( *comm, coords, &P ) ;
177       p = buffer_for_proc( P , 0 , da_buf ) ;
178       if ( pu == 0 ) {
179 	if ( typesize == sizeof(int) ) {
180           for ( j = JMAX(jps) ; j <= JMIN(jpe) ; j++ ) {
181             for ( k = kps ; k <= kpe ; k++ ) {
182 	      pi = (int *)(p+x_curs[P]) ;
183 	      qi = (int *)((buf + typesize*( (i-ims) + (ime-ims+1)*(
184                                              (k-kms) + (j-jms)*(kme-kms+1))))) ;
185 	      *pi++ = *qi++ ;
186 	      x_curs[P] += typesize ;
187 	    }
188 	  }
189 	}
190 	else {
191           for ( j = JMAX(jps) ; j <= JMIN(jpe) ; j++ ) {
192             for ( k = kps ; k <= kpe ; k++ ) {
193               for ( t = 0 ; t < typesize ; t++ ) {
194                 *(p+x_curs[P]) = 
195                                *(buf + t + typesize*(
196                                       (i-ims) + (ime-ims+1)*(
197                                       (k-kms) + (j-jms)*(kme-kms+1))) ) ;
198                 x_curs[P]++ ;
199               }
200             }
201           }
202 	}
203       } else {
204 	if ( typesize == sizeof(int) ) {
205           for ( j = JMAX(jps) ; j <= JMIN(jpe) ; j++ ) {
206             for ( k = kps ; k <= kpe ; k++ ) {
207 	      pi = (int *)(p+x_curs[P]) ;
208 	      qi = (int *)((buf + typesize*( (i-ims) + (ime-ims+1)*(
209                                              (k-kms) + (j-jms)*(kme-kms+1))))) ;
210 	      *qi++ = *pi++ ;
211 	      x_curs[P] += typesize ;
212 	    }
213 	  }
214 	}
215 	else {
216           for ( j = JMAX(jps) ; j <= JMIN(jpe) ; j++ ) {
217             for ( k = kps ; k <= kpe ; k++ ) {
218               for ( t = 0 ; t < typesize ; t++ ) {
219                                *(buf + t + typesize*(
220                                       (i-ims) + (ime-ims+1)*(
221                                       (k-kms) + (j-jms)*(kme-kms+1))) ) =
222                 *(p+x_curs[P]) ;
223                 x_curs[P]++ ;
224               }
225             }
226           }
227         }
228       }
229     }
230   } else if ( np_y > 1 && xy == 0 ) {
231     for ( j = UP_ODD(jps) ; j <= MIN(jpe,m) ; j+=2 ) {
232       jj = abs(j+n) % m ;
233       TASK_FOR_POINT ( &ips , &jj , &ids, &ide , &jds, &jde , &np_x , &np_y , &Px, &Py ) ;
234       coords[1] = Px ; coords[0] = Py ;
235       MPI_Cart_rank( *comm, coords, &P ) ;
236       p = buffer_for_proc( P , 0 , da_buf ) ;
237       if ( pu == 0 ) {
238 	if ( typesize == sizeof(int) ) {
239           for ( i = IMAX(ips) ; i <= IMIN(ipe) ; i++ ) {
240             for ( k = kps ; k <= kpe ; k++ ) {
241 	      pi = (int *)(p+x_curs[P]) ;
242 	      qi = (int *)((buf + typesize*( (i-ims) + (ime-ims+1)*(
243                                              (k-kms) + (j-jms)*(kme-kms+1))))) ;
244 	      *pi++ = *qi++ ;
245 	      x_curs[P] += typesize ;
246 	    }
247 	  }
248 	}
249 	else {
250           for ( i = IMAX(ips) ; i <= IMIN(ipe) ; i++ ) {
251             for ( k = kps ; k <= kpe ; k++ ) {
252               for ( t = 0 ; t < typesize ; t++ ) {
253                 *(p+x_curs[P]) = 
254                                *(buf + t + typesize*(
255                                       (i-ims) + (ime-ims+1)*(
256                                       (k-kms) + (j-jms)*(kme-kms+1))) ) ;
257                 x_curs[P]++ ;
258               }
259             }
260           }
261 	}
262       } else {
263 	if ( typesize == sizeof(int) ) {
264           for ( i = IMAX(ips) ; i <= IMIN(ipe) ; i++ ) {
265             for ( k = kps ; k <= kpe ; k++ ) {
266 	      pi = (int *)(p+x_curs[P]) ;
267 	      qi = (int *)((buf + typesize*( (i-ims) + (ime-ims+1)*(
268                                              (k-kms) + (j-jms)*(kme-kms+1))))) ;
269 	      *qi++ = *pi++ ;
270 	      x_curs[P] += typesize ;
271 	    }
272 	  }
273 	}
274 	else {
275           for ( i = IMAX(ips) ; i <= IMIN(ipe) ; i++ ) {
276             for ( k = kps ; k <= kpe ; k++ ) {
277               for ( t = 0 ; t < typesize ; t++ ) {
278                                *(buf + t + typesize*(
279                                       (i-ims) + (ime-ims+1)*(
280                                       (k-kms) + (j-jms)*(kme-kms+1))) ) =
281                 *(p+x_curs[P]) ;
282                 x_curs[P]++ ;
283               }
284             }
285           }
286         }
287       }
288     }
289   }
290 }
291 
292 RSL_LITE_SWAP ( int * Fcomm0, int *me0, int * np0 , int * np_x0 , int * np_y0 )
293 {
294   int me, np, np_x, np_y ;
295   int yp, ym, xp, xm, nb ;
296   MPI_Status stat ;
297   MPI_Comm comm, *comm0, dummy_comm ;
298   int i, P ;
299 
300   comm0 = &dummy_comm ;
301   *comm0 = MPI_Comm_f2c( *Fcomm0 ) ;
302 #if 1
303 
304   comm = *comm0 ; me = *me0 ; np = *np0 ; np_x = *np_x0 ; np_y = *np_y0 ;
305 
306 /* fprintf(stderr,"RSL_LITE_SWAP\n") ; */
307 
308   for ( P = 0 ; P < np ; P++ ) {
309     if ( x_peermask[P] ) {
310       nb = buffer_size_for_proc( P, RSL_RECVBUF ) ;
311 /* fprintf(stderr,"posting irecv from %d, nb = %d\n",P,nb) ; */
312       MPI_Irecv ( buffer_for_proc( P, x_curs[P], RSL_RECVBUF ), nb, MPI_CHAR, P, me, comm, &(x_recv[P]) ) ;
313 /* fprintf(stderr,"sending to         %d, nb = %d\n",P,x_curs[P]) ; */
314       MPI_Isend ( buffer_for_proc( P, 0,         RSL_SENDBUF ), x_curs[P], MPI_CHAR, P, P, comm, &(x_send[P]) ) ;
315     }
316   }
317   for ( P = 0 ; P < np ; P++ ) {
318     if ( x_peermask[P] ) {
319       MPI_Wait( &x_recv[P], &stat ) ; 
320       MPI_Wait( &x_send[P], &stat ) ; 
321     }
322   }
323 #else 
324 # ifndef MS_SUA
325 fprintf(stderr,"RSL_LITE_SWAP disabled\n") ;
326 # endif
327 #endif
328   for ( i = 0 ; i < np ; i++ ) {  x_curs[i] = 0 ;  }
329 }
330