xref: /petsc/src/vec/is/utils/psort.c (revision 3f02e49b19195914bf17f317a25cb39636853415)
1 #include <petsc/private/petscimpl.h>
2 #include <petscis.h> /*I "petscis.h" I*/
3 
4 /* This is the bitonic merge that works on non-power-of-2 sizes found at http://www.iti.fh-flensburg.de/lang/algorithmen/sortieren/bitonic/oddn.htm */
5 static PetscErrorCode PetscParallelSortInt_Bitonic_Merge(MPI_Comm comm, PetscMPIInt tag, PetscMPIInt rankStart, PetscMPIInt rankEnd, PetscMPIInt rank, PetscMPIInt n, PetscInt keys[], PetscInt buffer[], PetscBool forward)
6 {
7   PetscInt    diff;
8   PetscMPIInt split, mid, partner;
9 
10   PetscFunctionBegin;
11   diff = rankEnd - rankStart;
12   if (diff <= 0) PetscFunctionReturn(PETSC_SUCCESS);
13   if (diff == 1) {
14     if (forward) {
15       PetscCall(PetscSortInt(n, keys));
16     } else {
17       PetscCall(PetscSortReverseInt(n, keys));
18     }
19     PetscFunctionReturn(PETSC_SUCCESS);
20   }
21   split = 1;
22   while (2 * split < diff) split *= 2;
23   mid = rankStart + split;
24   if (rank < mid) {
25     partner = rank + split;
26   } else {
27     partner = rank - split;
28   }
29   if (partner < rankEnd) {
30     PetscMPIInt i;
31 
32     PetscCallMPI(MPI_Sendrecv(keys, n, MPIU_INT, partner, tag, buffer, n, MPIU_INT, partner, tag, comm, MPI_STATUS_IGNORE));
33     if ((rank < partner) == (forward == PETSC_TRUE)) {
34       for (i = 0; i < n; i++) keys[i] = (keys[i] <= buffer[i]) ? keys[i] : buffer[i];
35     } else {
36       for (i = 0; i < n; i++) keys[i] = (keys[i] > buffer[i]) ? keys[i] : buffer[i];
37     }
38   }
39   /* divide and conquer */
40   if (rank < mid) {
41     PetscCall(PetscParallelSortInt_Bitonic_Merge(comm, tag, rankStart, mid, rank, n, keys, buffer, forward));
42   } else {
43     PetscCall(PetscParallelSortInt_Bitonic_Merge(comm, tag, mid, rankEnd, rank, n, keys, buffer, forward));
44   }
45   PetscFunctionReturn(PETSC_SUCCESS);
46 }
47 
48 /* This is the bitonic sort that works on non-power-of-2 sizes found at http://www.iti.fh-flensburg.de/lang/algorithmen/sortieren/bitonic/oddn.htm */
49 static PetscErrorCode PetscParallelSortInt_Bitonic_Recursive(MPI_Comm comm, PetscMPIInt tag, PetscMPIInt rankStart, PetscMPIInt rankEnd, PetscMPIInt rank, PetscMPIInt n, PetscInt keys[], PetscInt buffer[], PetscBool forward)
50 {
51   PetscMPIInt diff, mid;
52 
53   PetscFunctionBegin;
54   diff = rankEnd - rankStart;
55   if (diff <= 0) PetscFunctionReturn(PETSC_SUCCESS);
56   if (diff == 1) {
57     if (forward) {
58       PetscCall(PetscSortInt(n, keys));
59     } else {
60       PetscCall(PetscSortReverseInt(n, keys));
61     }
62     PetscFunctionReturn(PETSC_SUCCESS);
63   }
64   mid = rankStart + diff / 2;
65   /* divide and conquer */
66   if (rank < mid) {
67     PetscCall(PetscParallelSortInt_Bitonic_Recursive(comm, tag, rankStart, mid, rank, n, keys, buffer, (PetscBool)!forward));
68   } else {
69     PetscCall(PetscParallelSortInt_Bitonic_Recursive(comm, tag, mid, rankEnd, rank, n, keys, buffer, forward));
70   }
71   /* bitonic merge */
72   PetscCall(PetscParallelSortInt_Bitonic_Merge(comm, tag, rankStart, rankEnd, rank, n, keys, buffer, forward));
73   PetscFunctionReturn(PETSC_SUCCESS);
74 }
75 
76 static PetscErrorCode PetscParallelSortInt_Bitonic(MPI_Comm comm, PetscInt n, PetscInt keys[])
77 {
78   PetscMPIInt size, rank, tag, mpin;
79   PetscInt   *buffer;
80 
81   PetscFunctionBegin;
82   PetscAssertPointer(keys, 3);
83   PetscCall(PetscCommGetNewTag(comm, &tag));
84   PetscCallMPI(MPI_Comm_size(comm, &size));
85   PetscCallMPI(MPI_Comm_rank(comm, &rank));
86   PetscCall(PetscMPIIntCast(n, &mpin));
87   PetscCall(PetscMalloc1(n, &buffer));
88   PetscCall(PetscParallelSortInt_Bitonic_Recursive(comm, tag, 0, size, rank, mpin, keys, buffer, PETSC_TRUE));
89   PetscCall(PetscFree(buffer));
90   PetscFunctionReturn(PETSC_SUCCESS);
91 }
92 
93 static PetscErrorCode PetscParallelSampleSelect(PetscLayout mapin, PetscLayout mapout, PetscInt keysin[], PetscInt *outpivots[])
94 {
95   PetscMPIInt  size, rank;
96   PetscInt    *pivots, *finalpivots, i;
97   PetscInt     non_empty, my_first, count;
98   PetscMPIInt *keys_per, max_keys_per;
99 
100   PetscFunctionBegin;
101   PetscCallMPI(MPI_Comm_size(mapin->comm, &size));
102   PetscCallMPI(MPI_Comm_rank(mapin->comm, &rank));
103 
104   /* Choose P - 1 pivots that would be ideal for the distribution on this process */
105   PetscCall(PetscMalloc1(size - 1, &pivots));
106   for (i = 0; i < size - 1; i++) {
107     PetscInt index;
108 
109     if (!mapin->n) {
110       /* if this rank is empty, put "infinity" in the list.  each process knows how many empty
111        * processes are in the layout, so those values will be ignored from the end of the sorted
112        * pivots */
113       pivots[i] = PETSC_INT_MAX;
114     } else {
115       /* match the distribution to the desired output described by mapout by getting the index
116        * that is approximately the appropriate fraction through the list */
117       index     = ((PetscReal)mapout->range[i + 1]) * ((PetscReal)mapin->n) / ((PetscReal)mapout->N);
118       index     = PetscMin(index, mapin->n - 1);
119       index     = PetscMax(index, 0);
120       pivots[i] = keysin[index];
121     }
122   }
123   /* sort the pivots in parallel */
124   PetscCall(PetscParallelSortInt_Bitonic(mapin->comm, size - 1, pivots));
125   if (PetscDefined(USE_DEBUG)) {
126     PetscBool sorted;
127 
128     PetscCall(PetscParallelSortedInt(mapin->comm, size - 1, pivots, &sorted));
129     PetscCheck(sorted, mapin->comm, PETSC_ERR_PLIB, "bitonic sort failed");
130   }
131 
132   /* if there are Z nonempty processes, we have (P - 1) * Z real pivots, and we want to select
133    * at indices Z - 1, 2*Z - 1, ... (P - 1) * Z - 1 */
134   non_empty = size;
135   for (i = 0; i < size; i++)
136     if (mapout->range[i] == mapout->range[i + 1]) non_empty--;
137   PetscCall(PetscCalloc1(size + 1, &keys_per));
138   my_first = -1;
139   if (non_empty) {
140     for (i = 0; i < size - 1; i++) {
141       size_t sample      = (i + 1) * non_empty - 1;
142       size_t sample_rank = sample / (size - 1);
143 
144       keys_per[sample_rank]++;
145       if (my_first < 0 && (PetscMPIInt)sample_rank == rank) my_first = (PetscInt)(sample - sample_rank * (size - 1));
146     }
147   }
148   for (i = 0, max_keys_per = 0; i < size; i++) max_keys_per = PetscMax(keys_per[i], max_keys_per);
149   PetscCall(PetscMalloc1(size * max_keys_per, &finalpivots));
150   /* now that we know how many pivots each process will provide, gather the selected pivots at the start of the array
151    * and allgather them */
152   for (i = 0; i < keys_per[rank]; i++) pivots[i] = pivots[my_first + i * non_empty];
153   for (i = keys_per[rank]; i < max_keys_per; i++) pivots[i] = PETSC_INT_MAX;
154   PetscCallMPI(MPI_Allgather(pivots, max_keys_per, MPIU_INT, finalpivots, max_keys_per, MPIU_INT, mapin->comm));
155   for (i = 0, count = 0; i < size; i++) {
156     PetscInt j;
157 
158     for (j = 0; j < max_keys_per; j++) {
159       if (j < keys_per[i]) finalpivots[count++] = finalpivots[i * max_keys_per + j];
160     }
161   }
162   *outpivots = finalpivots;
163   PetscCall(PetscFree(keys_per));
164   PetscCall(PetscFree(pivots));
165   PetscFunctionReturn(PETSC_SUCCESS);
166 }
167 
168 static PetscErrorCode PetscParallelRedistribute(PetscLayout map, PetscInt n, PetscInt arrayin[], PetscInt arrayout[])
169 {
170   PetscMPIInt  size, rank;
171   PetscInt     myOffset, nextOffset;
172   PetscCount   total;
173   PetscMPIInt  nfirst, nsecond;
174   PetscMPIInt  firsttag, secondtag;
175   MPI_Request  firstreqrcv;
176   MPI_Request *firstreqs;
177   MPI_Request *secondreqs;
178 
179   PetscFunctionBegin;
180   PetscCallMPI(MPI_Comm_size(map->comm, &size));
181   PetscCallMPI(MPI_Comm_rank(map->comm, &rank));
182   PetscCall(PetscCommGetNewTag(map->comm, &firsttag));
183   PetscCall(PetscCommGetNewTag(map->comm, &secondtag));
184   PetscCall(PetscMalloc2(size, &firstreqs, size, &secondreqs));
185   PetscCallMPI(MPI_Scan(&n, &nextOffset, 1, MPIU_INT, MPI_SUM, map->comm));
186   myOffset = nextOffset - n;
187   total    = map->range[rank + 1] - map->range[rank];
188   if (total > 0) PetscCallMPI(MPIU_Irecv(arrayout, total, MPIU_INT, MPI_ANY_SOURCE, firsttag, map->comm, &firstreqrcv));
189   nsecond = 0;
190   nfirst  = 0;
191   for (PetscMPIInt i = 0; i < size; i++) {
192     PetscInt itotal;
193     PetscInt overlap, oStart, oEnd;
194 
195     itotal = map->range[i + 1] - map->range[i];
196     if (itotal <= 0) continue;
197     oStart  = PetscMax(myOffset, map->range[i]);
198     oEnd    = PetscMin(nextOffset, map->range[i + 1]);
199     overlap = oEnd - oStart;
200     if (map->range[i] >= myOffset && map->range[i] < nextOffset) {
201       /* send first message */
202       PetscCallMPI(MPIU_Isend(&arrayin[map->range[i] - myOffset], overlap, MPIU_INT, i, firsttag, map->comm, &firstreqs[nfirst++]));
203     } else if (overlap > 0) {
204       /* send second message */
205       PetscCallMPI(MPIU_Isend(&arrayin[oStart - myOffset], overlap, MPIU_INT, i, secondtag, map->comm, &secondreqs[nsecond++]));
206     } else if (overlap == 0 && myOffset > map->range[i] && myOffset < map->range[i + 1]) {
207       /* send empty second message */
208       PetscCallMPI(MPIU_Isend(&arrayin[oStart - myOffset], 0, MPIU_INT, i, secondtag, map->comm, &secondreqs[nsecond++]));
209     }
210   }
211   if (total > 0) {
212     MPI_Status  status;
213     PetscMPIInt sender = -1;
214     PetscCount  filled = 0;
215 
216     PetscCallMPI(MPI_Wait(&firstreqrcv, &status));
217     sender = status.MPI_SOURCE;
218     PetscCallMPI(MPIU_Get_count(&status, MPIU_INT, &filled));
219     while (filled < total) {
220       PetscCount mfilled;
221 
222       sender++;
223       PetscCallMPI(MPIU_Recv(&arrayout[filled], total - filled, MPIU_INT, sender, secondtag, map->comm, &status));
224       PetscCallMPI(MPIU_Get_count(&status, MPIU_INT, &mfilled));
225       filled += mfilled;
226     }
227   }
228   PetscCallMPI(MPI_Waitall(nfirst, firstreqs, MPI_STATUSES_IGNORE));
229   PetscCallMPI(MPI_Waitall(nsecond, secondreqs, MPI_STATUSES_IGNORE));
230   PetscCall(PetscFree2(firstreqs, secondreqs));
231   PetscFunctionReturn(PETSC_SUCCESS);
232 }
233 
234 static PetscErrorCode PetscParallelSortInt_Samplesort(PetscLayout mapin, PetscLayout mapout, PetscInt keysin[], PetscInt keysout[])
235 {
236   PetscMPIInt  size, rank;
237   PetscInt    *pivots = NULL, *buffer;
238   PetscInt     j;
239   PetscMPIInt *keys_per_snd, *keys_per_rcv, *offsets_snd, *offsets_rcv, nrecv;
240 
241   PetscFunctionBegin;
242   PetscCallMPI(MPI_Comm_size(mapin->comm, &size));
243   PetscCallMPI(MPI_Comm_rank(mapin->comm, &rank));
244   PetscCall(PetscMalloc4(size, &keys_per_snd, size, &keys_per_rcv, size + 1, &offsets_snd, size + 1, &offsets_rcv));
245   /* sort locally */
246   PetscCall(PetscSortInt(mapin->n, keysin));
247   /* get P - 1 pivots */
248   PetscCall(PetscParallelSampleSelect(mapin, mapout, keysin, &pivots));
249   /* determine which entries in the sorted array go to which other processes based on the pivots */
250   j = 0;
251   for (PetscMPIInt i = 0; i < size - 1; i++) {
252     PetscInt prev = j;
253 
254     while ((j < mapin->n) && (keysin[j] < pivots[i])) j++;
255     PetscCall(PetscMPIIntCast(prev, &offsets_snd[i]));
256     PetscCall(PetscMPIIntCast(j - prev, &keys_per_snd[i]));
257   }
258   PetscCall(PetscMPIIntCast(j, &offsets_snd[size - 1]));
259   PetscCall(PetscMPIIntCast(mapin->n - j, &keys_per_snd[size - 1]));
260   PetscCall(PetscMPIIntCast(mapin->n, &offsets_snd[size]));
261   /* get the incoming sizes */
262   PetscCallMPI(MPI_Alltoall(keys_per_snd, 1, MPI_INT, keys_per_rcv, 1, MPI_INT, mapin->comm));
263   offsets_rcv[0] = 0;
264   for (PetscMPIInt i = 0; i < size; i++) offsets_rcv[i + 1] = offsets_rcv[i] + keys_per_rcv[i];
265   nrecv = offsets_rcv[size];
266   /* all to all exchange */
267   PetscCall(PetscMalloc1(nrecv, &buffer));
268   PetscCallMPI(MPI_Alltoallv(keysin, keys_per_snd, offsets_snd, MPIU_INT, buffer, keys_per_rcv, offsets_rcv, MPIU_INT, mapin->comm));
269   PetscCall(PetscFree(pivots));
270   PetscCall(PetscFree4(keys_per_snd, keys_per_rcv, offsets_snd, offsets_rcv));
271 
272   /* local sort */
273   PetscCall(PetscSortInt(nrecv, buffer));
274 #if defined(PETSC_USE_DEBUG)
275   {
276     PetscBool sorted;
277 
278     PetscCall(PetscParallelSortedInt(mapin->comm, nrecv, buffer, &sorted));
279     PetscCheck(sorted, mapin->comm, PETSC_ERR_PLIB, "samplesort (pre-redistribute) sort failed");
280   }
281 #endif
282 
283   /* redistribute to the desired order */
284   PetscCall(PetscParallelRedistribute(mapout, nrecv, buffer, keysout));
285   PetscCall(PetscFree(buffer));
286   PetscFunctionReturn(PETSC_SUCCESS);
287 }
288 
289 /*@
290   PetscParallelSortInt - Globally sort a distributed array of integers
291 
292   Collective
293 
294   Input Parameters:
295 + mapin  - `PetscLayout` describing the distribution of the input keys
296 . mapout - `PetscLayout` describing the desired distribution of the output keys
297 - keysin - the pre-sorted array of integers
298 
299   Output Parameter:
300 . keysout - the array in which the sorted integers will be stored.  If `mapin` == `mapout`, then `keysin` may be equal to `keysout`.
301 
302   Level: developer
303 
304   Notes:
305 
306   This implements a distributed samplesort, which, with local array sizes n_in and n_out,
307   global size N, and global number of MPI processes P, does\:
308 .vb
309   - sorts locally
310   - chooses pivots by sorting (in parallel) (P-1) pivot suggestions from each process using bitonic sort and allgathering a subset of (P-1) of those
311   - using to the pivots to repartition the keys by all-to-all exchange
312   - sorting the repartitioned keys locally (the array is now globally sorted, but does not match the mapout layout)
313   - redistributing to match the mapout layout
314 .ve
315 
316   If `keysin` != `keysout`, then `keysin` will not be changed during `PetscParallelSortInt()`.
317 
318 .seealso: `PetscSortInt()`, `PetscParallelSortedInt()`
319 @*/
320 PetscErrorCode PetscParallelSortInt(PetscLayout mapin, PetscLayout mapout, PetscInt keysin[], PetscInt keysout[])
321 {
322   PetscMPIInt size;
323   PetscMPIInt result;
324   PetscInt   *keysincopy = NULL;
325 
326   PetscFunctionBegin;
327   PetscAssertPointer(mapin, 1);
328   PetscAssertPointer(mapout, 2);
329   PetscCallMPI(MPI_Comm_compare(mapin->comm, mapout->comm, &result));
330   PetscCheck(result == MPI_IDENT || result == MPI_CONGRUENT, mapin->comm, PETSC_ERR_ARG_NOTSAMECOMM, "layouts are not on the same communicator");
331   PetscCall(PetscLayoutSetUp(mapin));
332   PetscCall(PetscLayoutSetUp(mapout));
333   if (mapin->n) PetscAssertPointer(keysin, 3);
334   if (mapout->n) PetscAssertPointer(keysout, 4);
335   PetscCheck(mapin->N == mapout->N, mapin->comm, PETSC_ERR_ARG_SIZ, "Input and output layouts have different global sizes (%" PetscInt_FMT " != %" PetscInt_FMT ")", mapin->N, mapout->N);
336   PetscCallMPI(MPI_Comm_size(mapin->comm, &size));
337   if (size == 1) {
338     if (keysout != keysin) PetscCall(PetscMemcpy(keysout, keysin, mapin->n * sizeof(PetscInt)));
339     PetscCall(PetscSortInt(mapout->n, keysout));
340     if (size == 1) PetscFunctionReturn(PETSC_SUCCESS);
341   }
342   if (keysout != keysin) {
343     PetscCall(PetscMalloc1(mapin->n, &keysincopy));
344     PetscCall(PetscMemcpy(keysincopy, keysin, mapin->n * sizeof(PetscInt)));
345     keysin = keysincopy;
346   }
347   PetscCall(PetscParallelSortInt_Samplesort(mapin, mapout, keysin, keysout));
348 #if defined(PETSC_USE_DEBUG)
349   {
350     PetscBool sorted;
351 
352     PetscCall(PetscParallelSortedInt(mapout->comm, mapout->n, keysout, &sorted));
353     PetscCheck(sorted, mapout->comm, PETSC_ERR_PLIB, "samplesort sort failed");
354   }
355 #endif
356   PetscCall(PetscFree(keysincopy));
357   PetscFunctionReturn(PETSC_SUCCESS);
358 }
359