Actual source code: vscatsf.c
petsc-3.12.5 2020-03-29
1: #include <petsc/private/vecscatterimpl.h>
2: #include <petsc/private/sfimpl.h>
3: #include <../src/vec/is/sf/impls/basic/sfbasic.h> /* for VecScatterRemap_SF */
4: #include <../src/vec/is/sf/impls/basic/sfpack.h>
6: #if defined(PETSC_HAVE_CUDA)
7: #include <../src/vec/vec/impls/seq/seqcuda/cudavecimpl.h>
8: #endif
10: typedef struct {
11: PetscSF sf; /* the whole scatter, including local and remote */
12: PetscSF lsf; /* the local part of the scatter, used for SCATTER_LOCAL */
13: PetscInt bs; /* block size */
14: MPI_Datatype unit; /* one unit = bs PetscScalars */
15: } VecScatter_SF;
17: static PetscErrorCode VecScatterBegin_SF(VecScatter vscat,Vec x,Vec y,InsertMode addv,ScatterMode mode)
18: {
19: VecScatter_SF *data=(VecScatter_SF*)vscat->data;
20: PetscSF sf;
21: MPI_Op mop=MPI_OP_NULL;
22: PetscMPIInt size;
26: if (x != y) {VecLockReadPush(x);}
28: if (use_gpu_aware_mpi) {
29: VecGetArrayReadInPlace(x,&vscat->xdata);
30: } else {
31: #if defined(PETSC_HAVE_CUDA)
32: PetscBool is_cudatype = PETSC_FALSE;
33: PetscObjectTypeCompareAny((PetscObject)x,&is_cudatype,VECSEQCUDA,VECMPICUDA,VECCUDA,"");
34: if (is_cudatype) {
35: VecCUDAAllocateCheckHost(x);
36: if (x->offloadmask == PETSC_OFFLOAD_GPU) {
37: if (x->spptr && vscat->spptr) {VecCUDACopyFromGPUSome_Public(x,(PetscCUDAIndices)vscat->spptr,mode);}
38: else {VecCUDACopyFromGPU(x);}
39: }
40: vscat->xdata = *((PetscScalar**)x->data);
41: } else
42: #endif
43: {VecGetArrayRead(x,&vscat->xdata);}
44: }
46: if (x != y) {
47: if (use_gpu_aware_mpi) {VecGetArrayInPlace(y,&vscat->ydata);}
48: else {VecGetArray(y,&vscat->ydata);}
49: } else vscat->ydata = (PetscScalar *)vscat->xdata;
50: VecLockWriteSet_Private(y,PETSC_TRUE);
52: /* SCATTER_LOCAL indicates ignoring inter-process communication */
53: MPI_Comm_size(PetscObjectComm((PetscObject)data->sf),&size);
54: if ((mode & SCATTER_LOCAL) && size > 1) { /* Lazy creation of data->lsf since SCATTER_LOCAL is uncommon */
55: if (!data->lsf) {PetscSFCreateLocalSF_Private(data->sf,&data->lsf);}
56: sf = data->lsf;
57: } else {
58: sf = data->sf;
59: }
61: if (addv == INSERT_VALUES) mop = MPIU_REPLACE;
62: else if (addv == ADD_VALUES) mop = MPIU_SUM; /* Petsc defines its own MPI datatype and SUM operation for __float128 etc. */
63: else if (addv == MAX_VALUES) mop = MPIU_MAX;
64: else if (addv == MIN_VALUES) mop = MPIU_MIN;
65: else SETERRQ1(PetscObjectComm((PetscObject)sf),PETSC_ERR_SUP,"Unsupported InsertMode %D in VecScatterBegin/End",addv);
67: if (mode & SCATTER_REVERSE) { /* REVERSE indicates leaves to root scatter. Note that x and y are swapped in input */
68: PetscSFReduceBegin(sf,data->unit,vscat->xdata,vscat->ydata,mop);
69: } else { /* FORWARD indicates x to y scatter, where x is root and y is leaf */
70: PetscSFBcastAndOpBegin(sf,data->unit,vscat->xdata,vscat->ydata,mop);
71: }
72: return(0);
73: }
75: static PetscErrorCode VecScatterEnd_SF(VecScatter vscat,Vec x,Vec y,InsertMode addv,ScatterMode mode)
76: {
77: VecScatter_SF *data=(VecScatter_SF*)vscat->data;
78: PetscSF sf;
79: MPI_Op mop=MPI_OP_NULL;
80: PetscMPIInt size;
84: /* SCATTER_LOCAL indicates ignoring inter-process communication */
85: MPI_Comm_size(PetscObjectComm((PetscObject)data->sf),&size);
86: sf = ((mode & SCATTER_LOCAL) && size > 1) ? data->lsf : data->sf;
88: if (addv == INSERT_VALUES) mop = MPIU_REPLACE;
89: else if (addv == ADD_VALUES) mop = MPIU_SUM;
90: else if (addv == MAX_VALUES) mop = MPIU_MAX;
91: else if (addv == MIN_VALUES) mop = MPIU_MIN;
92: else SETERRQ1(PetscObjectComm((PetscObject)sf),PETSC_ERR_SUP,"Unsupported InsertMode %D in VecScatterBegin/End",addv);
94: if (mode & SCATTER_REVERSE) { /* reverse scatter sends leaves to roots. Note that x and y are swapped in input */
95: PetscSFReduceEnd(sf,data->unit,vscat->xdata,vscat->ydata,mop);
96: } else { /* forward scatter sends roots to leaves, i.e., x to y */
97: PetscSFBcastAndOpEnd(sf,data->unit,vscat->xdata,vscat->ydata,mop);
98: }
100: if (x != y) {
101: if (use_gpu_aware_mpi) {VecRestoreArrayReadInPlace(x,&vscat->xdata);}
102: else {VecRestoreArrayRead(x,&vscat->xdata);}
103: VecLockReadPop(x);
104: }
106: if (use_gpu_aware_mpi) {VecRestoreArrayInPlace(y,&vscat->ydata);}
107: else {VecRestoreArray(y,&vscat->ydata);}
108: VecLockWriteSet_Private(y,PETSC_FALSE);
110: return(0);
111: }
113: static PetscErrorCode VecScatterCopy_SF(VecScatter vscat,VecScatter ctx)
114: {
115: VecScatter_SF *data=(VecScatter_SF*)vscat->data,*out;
119: PetscMemcpy(ctx->ops,vscat->ops,sizeof(vscat->ops));
120: PetscNewLog(ctx,&out);
121: PetscSFDuplicate(data->sf,PETSCSF_DUPLICATE_GRAPH,&out->sf);
122: PetscSFSetUp(out->sf);
123: /* Do not copy lsf. Build it on demand since it is rarely used */
125: out->bs = data->bs;
126: if (out->bs > 1) {
127: MPI_Type_dup(data->unit,&out->unit); /* Since oldtype is committed, so is newtype, according to MPI */
128: } else {
129: out->unit = MPIU_SCALAR;
130: }
131: ctx->data = (void*)out;
132: return(0);
133: }
135: static PetscErrorCode VecScatterDestroy_SF(VecScatter vscat)
136: {
137: VecScatter_SF *data = (VecScatter_SF *)vscat->data;
141: PetscSFDestroy(&data->sf);
142: PetscSFDestroy(&data->lsf);
143: if (data->bs > 1) {MPI_Type_free(&data->unit);}
144: PetscFree(vscat->data);
145: return(0);
146: }
148: static PetscErrorCode VecScatterView_SF(VecScatter vscat,PetscViewer viewer)
149: {
150: VecScatter_SF *data = (VecScatter_SF *)vscat->data;
154: PetscSFView(data->sf,viewer);
155: return(0);
156: }
158: /* VecScatterRemap provides a light way to slightly modify a VecScatter. Suppose the input vscat scatters
159: x[i] to y[j], tomap gives a plan to change vscat to scatter x[tomap[i]] to y[j]. Note that in SF,
160: x is roots. That means we need to change incoming stuffs such as bas->irootloc[].
161: */
162: static PetscErrorCode VecScatterRemap_SF(VecScatter vscat,const PetscInt *tomap,const PetscInt *frommap)
163: {
164: VecScatter_SF *data = (VecScatter_SF *)vscat->data;
165: PetscSF sf = data->sf;
166: PetscInt i,bs = data->bs;
167: PetscMPIInt size;
168: PetscBool ident = PETSC_TRUE,isbasic,isneighbor;
169: PetscSFType type;
170: PetscSF_Basic *bas = NULL;
174: /* check if it is an identity map. If it is, do nothing */
175: if (tomap) {
176: for (i=0; i<sf->nroots*bs; i++) {if (i != tomap[i]) {ident = PETSC_FALSE; break; } }
177: if (ident) return(0);
178: }
179: if (frommap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unable to remap the FROM in scatters yet");
180: if (!tomap) return(0);
182: MPI_Comm_size(PetscObjectComm((PetscObject)data->sf),&size);
184: /* Since the indices changed, we must also update the local SF. But we do not do it since
185: lsf is rarely used. We just destroy lsf and rebuild it on demand from updated data->sf.
186: */
187: if (data->lsf) {PetscSFDestroy(&data->lsf);}
189: PetscSFGetType(sf,&type);
190: PetscObjectTypeCompare((PetscObject)sf,PETSCSFBASIC,&isbasic);
191: PetscObjectTypeCompare((PetscObject)sf,PETSCSFNEIGHBOR,&isneighbor);
192: if (!isbasic && !isneighbor) SETERRQ1(PetscObjectComm((PetscObject)sf),PETSC_ERR_SUP,"VecScatterRemap on SF type %s is not supported",type);
194: PetscSFSetUp(sf); /* to bulid sf->irootloc if SetUp is not yet called */
196: /* Root indices are going to be remapped. This is tricky for SF. Root indices are used in sf->rremote,
197: sf->remote and bas->irootloc. The latter one is cheap to remap, but the former two are not.
198: To remap them, we have to do a bcast from roots to leaves, to let leaves know their updated roots.
199: Since VecScatterRemap is supposed to be a cheap routine to adapt a vecscatter by only changing where
200: x[] data is taken, we do not remap sf->rremote, sf->remote. The consequence is that operations
201: accessing them (such as PetscSFCompose) may get stale info. Considering VecScatter does not need
202: that complicated SF operations, we do not remap sf->rremote, sf->remote, instead we destroy them
203: so that code accessing them (if any) will crash (instead of get silent errors). Note that BcastAndOp/Reduce,
204: which are used by VecScatter and only rely on bas->irootloc, are updated and correct.
205: */
206: sf->remote = NULL;
207: PetscFree(sf->remote_alloc);
208: /* Not easy to free sf->rremote since it was allocated with PetscMalloc4(), so just give it crazy values */
209: for (i=0; i<sf->roffset[sf->nranks]; i++) sf->rremote[i] = PETSC_MIN_INT;
211: /* Indices in tomap[] are for each indivisual vector entry. But indices in sf are for each
212: block in the vector. So before the remapping, we have to expand indices in sf by bs, and
213: after the remapping, we have to shrink them back.
214: */
215: bas = (PetscSF_Basic*)sf->data;
216: for (i=0; i<bas->ioffset[bas->niranks]; i++) bas->irootloc[i] = tomap[bas->irootloc[i]*bs]/bs;
217: #if defined(PETSC_HAVE_CUDA)
218: /* Free the irootloc copy on device. We allocate a new copy and get the updated value on demand. See PetscSFGetRootIndicesWithMemType_Basic() */
219: if (bas->irootloc_d) {cudaError_t err = cudaFree(bas->irootloc_d);CHKERRCUDA(err);bas->irootloc_d=NULL;}
220: #endif
221: /* Destroy and then rebuild root packing optimizations since indices are changed */
222: PetscSFPackDestroyOptimizations_Basic(sf);
223: PetscSFPackSetupOptimizations_Basic(sf);
224: return(0);
225: }
227: static PetscErrorCode VecScatterGetRemoteCount_SF(VecScatter vscat,PetscBool send,PetscInt *num_procs,PetscInt *num_entries)
228: {
229: PetscErrorCode ierr;
230: VecScatter_SF *data = (VecScatter_SF *)vscat->data;
231: PetscSF sf = data->sf;
232: PetscInt nranks,remote_start;
233: PetscMPIInt rank;
234: const PetscInt *offset;
235: const PetscMPIInt *ranks;
238: MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);
240: /* This routine is mainly used for MatMult's Mvctx. In Mvctx, we scatter an MPI vector x to a sequential vector lvec.
241: Remember x is roots and lvec is leaves. 'send' means roots to leaves communication. If 'send' is true, we need to
242: get info about which ranks this processor needs to send to. In other words, we need to call PetscSFGetLeafRanks().
243: If send is false, we do the opposite, calling PetscSFGetRootRanks().
244: */
245: if (send) {PetscSFGetLeafRanks(sf,&nranks,&ranks,&offset,NULL);}
246: else {PetscSFGetRootRanks(sf,&nranks,&ranks,&offset,NULL,NULL);}
247: if (nranks) {
248: remote_start = (rank == ranks[0])? 1 : 0;
249: if (num_procs) *num_procs = nranks - remote_start;
250: if (num_entries) *num_entries = offset[nranks] - offset[remote_start];
251: } else {
252: if (num_procs) *num_procs = 0;
253: if (num_entries) *num_entries = 0;
254: }
255: return(0);
256: }
258: static PetscErrorCode VecScatterGetRemote_SF(VecScatter vscat,PetscBool send,PetscInt *n,const PetscInt **starts,const PetscInt **indices,const PetscMPIInt **procs,PetscInt *bs)
259: {
260: PetscErrorCode ierr;
261: VecScatter_SF *data = (VecScatter_SF *)vscat->data;
262: PetscSF sf = data->sf;
263: PetscInt nranks,remote_start;
264: PetscMPIInt rank;
265: const PetscInt *offset,*location;
266: const PetscMPIInt *ranks;
269: MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);
271: if (send) {PetscSFGetLeafRanks(sf,&nranks,&ranks,&offset,&location);}
272: else {PetscSFGetRootRanks(sf,&nranks,&ranks,&offset,&location,NULL);}
274: if (nranks) {
275: remote_start = (rank == ranks[0])? 1 : 0;
276: if (n) *n = nranks - remote_start;
277: if (starts) *starts = &offset[remote_start];
278: if (indices) *indices = location; /* not &location[offset[remote_start]]. Starts[0] may point to the middle of indices[] */
279: if (procs) *procs = &ranks[remote_start];
280: } else {
281: if (n) *n = 0;
282: if (starts) *starts = NULL;
283: if (indices) *indices = NULL;
284: if (procs) *procs = NULL;
285: }
287: if (bs) *bs = 1;
288: return(0);
289: }
291: static PetscErrorCode VecScatterGetRemoteOrdered_SF(VecScatter vscat,PetscBool send,PetscInt *n,const PetscInt **starts,const PetscInt **indices,const PetscMPIInt **procs,PetscInt *bs)
292: {
296: VecScatterGetRemote_SF(vscat,send,n,starts,indices,procs,bs);
297: return(0);
298: }
300: static PetscErrorCode VecScatterRestoreRemote_SF(VecScatter vscat,PetscBool send,PetscInt *n,const PetscInt **starts,const PetscInt **indices,const PetscMPIInt **procs,PetscInt *bs)
301: {
303: if (starts) *starts = NULL;
304: if (indices) *indices = NULL;
305: if (procs) *procs = NULL;
306: return(0);
307: }
309: static PetscErrorCode VecScatterRestoreRemoteOrdered_SF(VecScatter vscat,PetscBool send,PetscInt *n,const PetscInt **starts,const PetscInt **indices,const PetscMPIInt **procs,PetscInt *bs)
310: {
313: VecScatterRestoreRemote_SF(vscat,send,n,starts,indices,procs,bs);
314: return(0);
315: }
317: typedef enum {IS_INVALID, IS_GENERAL, IS_BLOCK, IS_STRIDE} ISTypeID;
319: PETSC_STATIC_INLINE PetscErrorCode ISGetTypeID_Private(IS is,ISTypeID *id)
320: {
322: PetscBool same;
325: *id = IS_INVALID;
326: PetscObjectTypeCompare((PetscObject)is,ISGENERAL,&same);
327: if (same) {*id = IS_GENERAL; goto functionend;}
328: PetscObjectTypeCompare((PetscObject)is,ISBLOCK,&same);
329: if (same) {*id = IS_BLOCK; goto functionend;}
330: PetscObjectTypeCompare((PetscObject)is,ISSTRIDE,&same);
331: if (same) {*id = IS_STRIDE; goto functionend;}
332: functionend:
333: return(0);
334: }
336: static PetscErrorCode VecScatterSetUp_SF(VecScatter vscat)
337: {
339: VecScatter_SF *data;
340: MPI_Comm xcomm,ycomm,bigcomm;
341: Vec x=vscat->from_v,y=vscat->to_v,xx,yy;
342: IS ix=vscat->from_is,iy=vscat->to_is,ixx,iyy;
343: PetscMPIInt xcommsize,ycommsize,rank;
344: PetscInt i,n,N,nroots,nleaves,*ilocal,xstart,ystart,ixsize,iysize,xlen,ylen;
345: const PetscInt *xindices,*yindices;
346: PetscSFNode *iremote;
347: PetscLayout xlayout,ylayout;
348: ISTypeID ixid,iyid;
349: PetscInt bs,bsx,bsy,min,max,m[2],ixfirst,ixstep,iyfirst,iystep;
350: PetscBool can_do_block_opt=PETSC_FALSE;
353: PetscNewLog(vscat,&data);
354: data->bs = 1; /* default, no blocking */
355: data->unit = MPIU_SCALAR;
357: /*
358: Let P and S stand for parallel and sequential vectors respectively. There are four combinations of vecscatters: PtoP, PtoS,
359: StoP and StoS. The assumption of VecScatterCreate(Vec x,IS ix,Vec y,IS iy,VecScatter *newctx) is: if x is parallel, then ix
360: contains global indices of x. If x is sequential, ix contains local indices of x. Similarily for y and iy.
362: SF builds around concepts of local leaves and remote roots. We treat source vector x as roots and destination vector y as
363: leaves. A PtoS scatter can be naturally mapped to SF. We transform PtoP and StoP to PtoS, and treat StoS as trivial PtoS.
364: */
365: PetscObjectGetComm((PetscObject)x,&xcomm);
366: PetscObjectGetComm((PetscObject)y,&ycomm);
367: MPI_Comm_size(xcomm,&xcommsize);
368: MPI_Comm_size(ycomm,&ycommsize);
370: /* NULL ix or iy in VecScatterCreate(x,ix,y,iy,newctx) has special meaning. Recover them for these cases */
371: if (!ix) {
372: if (xcommsize > 1 && ycommsize == 1) { /* PtoS: null ix means the whole x will be scattered to each seq y */
373: VecGetSize(x,&N);
374: ISCreateStride(PETSC_COMM_SELF,N,0,1,&ix);
375: } else { /* PtoP, StoP or StoS: null ix means the whole local part of x will be scattered */
376: VecGetLocalSize(x,&n);
377: VecGetOwnershipRange(x,&xstart,NULL);
378: ISCreateStride(PETSC_COMM_SELF,n,xstart,1,&ix);
379: }
380: }
382: if (!iy) {
383: if (xcommsize == 1 && ycommsize > 1) { /* StoP: null iy means the whole y will be scattered to from each seq x */
384: VecGetSize(y,&N);
385: ISCreateStride(PETSC_COMM_SELF,N,0,1,&iy);
386: } else { /* PtoP, StoP or StoS: null iy means the whole local part of y will be scattered to */
387: VecGetLocalSize(y,&n);
388: VecGetOwnershipRange(y,&ystart,NULL);
389: ISCreateStride(PETSC_COMM_SELF,n,ystart,1,&iy);
390: }
391: }
393: /* Do error checking immediately after we have non-empty ix, iy */
394: ISGetLocalSize(ix,&ixsize);
395: ISGetLocalSize(iy,&iysize);
396: VecGetSize(x,&xlen);
397: VecGetSize(y,&ylen);
398: if (ixsize != iysize) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Scatter sizes of ix and iy don't match locally");
399: ISGetMinMax(ix,&min,&max);
400: if (min < 0 || max >= xlen) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Scatter indices in ix are out of range");
401: ISGetMinMax(iy,&min,&max);
402: if (min < 0 || max >= ylen) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Scatter indices in iy are out of range");
404: /* Extract info about ix, iy for further test */
405: ISGetTypeID_Private(ix,&ixid);
406: ISGetTypeID_Private(iy,&iyid);
407: if (ixid == IS_BLOCK) {ISGetBlockSize(ix,&bsx);}
408: else if (ixid == IS_STRIDE) {ISStrideGetInfo(ix,&ixfirst,&ixstep);}
410: if ( iyid == IS_BLOCK) {ISGetBlockSize(iy,&bsy);}
411: else if (iyid == IS_STRIDE) {ISStrideGetInfo(iy,&iyfirst,&iystep);}
413: /* Check if a PtoS is special ToAll/ToZero scatters, which can be results of VecScatterCreateToAll/Zero.
414: ToAll means a whole MPI vector is copied to a seq vector on every process. ToZero means a whole MPI
415: vector is copied to a seq vector on rank 0 and other processes do nothing(i.e.,they input empty ix,iy).
417: We can optimize these scatters with MPI collectives. We can also avoid costly analysis used for general scatters.
418: */
419: if (xcommsize > 1 && ycommsize == 1) { /* Ranks do not diverge at this if-test */
420: PetscInt pattern[2] = {0, 0}; /* A boolean array with pattern[0] for allgather-like (ToAll) and pattern[1] for gather-like (ToZero) */
421: PetscLayout map;
423: MPI_Comm_rank(xcomm,&rank);
424: VecGetLayout(x,&map);
425: if (!rank) {
426: if (ixid == IS_STRIDE && iyid == IS_STRIDE && ixsize == xlen && ixfirst == 0 && ixstep == 1 && iyfirst == 0 && iystep == 1) {
427: /* Rank 0 scatters the whole mpi x to seq y, so it is either a ToAll or a ToZero candidate in its view */
428: pattern[0] = pattern[1] = 1;
429: }
430: } else {
431: if (ixid == IS_STRIDE && iyid == IS_STRIDE && ixsize == xlen && ixfirst == 0 && ixstep == 1 && iyfirst == 0 && iystep == 1) {
432: /* Other ranks also scatter the whole mpi x to seq y, so it is a ToAll candidate in their view */
433: pattern[0] = 1;
434: } else if (ixsize == 0) {
435: /* Other ranks do nothing, so it is a ToZero candiate */
436: pattern[1] = 1;
437: }
438: }
440: /* One stone (the expensive allreduce) two birds: pattern[] tells if it is ToAll or ToZero */
441: MPIU_Allreduce(MPI_IN_PLACE,pattern,2,MPIU_INT,MPI_LAND,xcomm);
443: if (pattern[0] || pattern[1]) {
444: PetscSFCreate(xcomm,&data->sf);
445: PetscSFSetGraphWithPattern(data->sf,map,pattern[0] ? PETSCSF_PATTERN_ALLGATHER : PETSCSF_PATTERN_GATHER);
446: goto functionend; /* No further analysis needed. What a big win! */
447: }
448: }
450: /* Continue ...
451: Do block optimization by taking advantage of high level info available in ix, iy.
452: The block optimization is valid when all of the following conditions are met:
453: 1) ix, iy are blocked or can be blocked (i.e., strided with step=1);
454: 2) ix, iy have the same block size;
455: 3) all processors agree on one block size;
456: 4) no blocks span more than one process;
457: */
458: bigcomm = (xcommsize == 1) ? ycomm : xcomm;
460: /* Processors could go through different path in this if-else test */
461: m[0] = m[1] = PETSC_MPI_INT_MIN;
462: if (ixid == IS_BLOCK && iyid == IS_BLOCK) {
463: m[0] = PetscMax(bsx,bsy);
464: m[1] = -PetscMin(bsx,bsy);
465: } else if (ixid == IS_BLOCK && iyid == IS_STRIDE && iystep==1 && iyfirst%bsx==0) {
466: m[0] = bsx;
467: m[1] = -bsx;
468: } else if (ixid == IS_STRIDE && iyid == IS_BLOCK && ixstep==1 && ixfirst%bsy==0) {
469: m[0] = bsy;
470: m[1] = -bsy;
471: }
472: /* Get max and min of bsx,bsy over all processes in one allreduce */
473: MPIU_Allreduce(MPI_IN_PLACE,m,2,MPIU_INT,MPI_MAX,bigcomm);
474: max = m[0]; min = -m[1];
476: /* Since we used allreduce above, all ranks will have the same min and max. min==max
477: implies all ranks have the same bs. Do further test to see if local vectors are dividable
478: by bs on ALL ranks. If they are, we are ensured that no blocks span more than one processor.
479: */
480: if (min == max && min > 1) {
481: VecGetLocalSize(x,&xlen);
482: VecGetLocalSize(y,&ylen);
483: m[0] = xlen%min;
484: m[1] = ylen%min;
485: MPIU_Allreduce(MPI_IN_PLACE,m,2,MPIU_INT,MPI_LOR,bigcomm);
486: if (!m[0] && !m[1]) can_do_block_opt = PETSC_TRUE;
487: }
489: /* If can_do_block_opt, then shrink x, y, ix and iy by bs to get xx, yy, ixx and iyy, whose indices
490: and layout are actually used in building SF. Suppose blocked ix representing {0,1,2,6,7,8} has
491: indices {0,2} and bs=3, then ixx = {0,2}; suppose strided iy={3,4,5,6,7,8}, then iyy={1,2}.
493: xx is a little special. If x is seq, then xx is the concatenation of seq x's on ycomm. In this way,
494: we can treat PtoP and StoP uniformly as PtoS.
495: */
496: if (can_do_block_opt) {
497: const PetscInt *indices;
499: data->bs = bs = min;
500: MPI_Type_contiguous(bs,MPIU_SCALAR,&data->unit);
501: MPI_Type_commit(&data->unit);
503: /* Shrink x and ix */
504: VecCreateMPIWithArray(bigcomm,1,xlen/bs,PETSC_DECIDE,NULL,&xx); /* We only care xx's layout */
505: if (ixid == IS_BLOCK) {
506: ISBlockGetIndices(ix,&indices);
507: ISBlockGetLocalSize(ix,&ixsize);
508: ISCreateGeneral(PETSC_COMM_SELF,ixsize,indices,PETSC_COPY_VALUES,&ixx);
509: ISBlockRestoreIndices(ix,&indices);
510: } else { /* ixid == IS_STRIDE */
511: ISGetLocalSize(ix,&ixsize);
512: ISCreateStride(PETSC_COMM_SELF,ixsize/bs,ixfirst/bs,1,&ixx);
513: }
515: /* Shrink y and iy */
516: VecCreateMPIWithArray(ycomm,1,ylen/bs,PETSC_DECIDE,NULL,&yy);
517: if (iyid == IS_BLOCK) {
518: ISBlockGetIndices(iy,&indices);
519: ISBlockGetLocalSize(iy,&iysize);
520: ISCreateGeneral(PETSC_COMM_SELF,iysize,indices,PETSC_COPY_VALUES,&iyy);
521: ISBlockRestoreIndices(iy,&indices);
522: } else { /* iyid == IS_STRIDE */
523: ISGetLocalSize(iy,&iysize);
524: ISCreateStride(PETSC_COMM_SELF,iysize/bs,iyfirst/bs,1,&iyy);
525: }
526: } else {
527: ixx = ix;
528: iyy = iy;
529: yy = y;
530: if (xcommsize == 1) {VecCreateMPIWithArray(bigcomm,1,xlen,PETSC_DECIDE,NULL,&xx);} else xx = x;
531: }
533: /* Now it is ready to build SF with preprocessed (xx, yy) and (ixx, iyy) */
534: ISGetIndices(ixx,&xindices);
535: ISGetIndices(iyy,&yindices);
536: VecGetLayout(xx,&xlayout);
538: if (ycommsize > 1) {
539: /* PtoP or StoP */
541: /* Below is a piece of complex code with a very simple goal: move global index pairs (xindices[i], yindices[i]),
542: to owner process of yindices[i] according to ylayout, i = 0..n.
544: I did it through a temp sf, but later I thought the old design was inefficient and also distorted log view.
545: We want to mape one VecScatterCreate() call to one PetscSFCreate() call. The old design mapped to three
546: PetscSFCreate() calls. This code is on critical path of VecScatterSetUp and is used by every VecScatterCreate.
547: So I commented it out and did another optimized implementation. The commented code is left here for reference.
548: */
549: #if 0
550: const PetscInt *degree;
551: PetscSF tmpsf;
552: PetscInt inedges=0,*leafdata,*rootdata;
554: VecGetOwnershipRange(xx,&xstart,NULL);
555: VecGetLayout(yy,&ylayout);
556: VecGetOwnershipRange(yy,&ystart,NULL);
558: VecGetLocalSize(yy,&nroots);
559: ISGetLocalSize(iyy,&nleaves);
560: PetscMalloc2(nleaves,&iremote,nleaves*2,&leafdata);
562: for (i=0; i<nleaves; i++) {
563: PetscLayoutFindOwnerIndex(ylayout,yindices[i],&iremote[i].rank,&iremote[i].index);
564: leafdata[2*i] = yindices[i];
565: leafdata[2*i+1] = (xcommsize > 1)? xindices[i] : xindices[i] + xstart;
566: }
568: PetscSFCreate(ycomm,&tmpsf);
569: PetscSFSetGraph(tmpsf,nroots,nleaves,NULL,PETSC_USE_POINTER,iremote,PETSC_USE_POINTER);
571: PetscSFComputeDegreeBegin(tmpsf,°ree);
572: PetscSFComputeDegreeEnd(tmpsf,°ree);
574: for (i=0; i<nroots; i++) inedges += degree[i];
575: PetscMalloc1(inedges*2,&rootdata);
576: PetscSFGatherBegin(tmpsf,MPIU_2INT,leafdata,rootdata);
577: PetscSFGatherEnd(tmpsf,MPIU_2INT,leafdata,rootdata);
579: PetscFree2(iremote,leafdata);
580: PetscSFDestroy(&tmpsf);
582: /* rootdata contains global index pairs (i, j). j's are owned by the current process, but i's can point to anywhere.
583: We convert j to local, and convert i to (rank, index). In the end, we get an PtoS suitable for building SF.
584: */
585: nleaves = inedges;
586: VecGetLocalSize(xx,&nroots);
587: PetscMalloc1(nleaves,&ilocal);
588: PetscMalloc1(nleaves,&iremote);
590: for (i=0; i<inedges; i++) {
591: ilocal[i] = rootdata[2*i] - ystart; /* covert y's global index to local index */
592: PetscLayoutFindOwnerIndex(xlayout,rootdata[2*i+1],&iremote[i].rank,&iremote[i].index); /* convert x's global index to (rank, index) */
593: }
594: PetscFree(rootdata);
595: #else
596: PetscInt j,k,n,disp,rlentotal,*sstart,*xindices_sorted,*yindices_sorted;
597: const PetscInt *yrange;
598: PetscMPIInt nsend,nrecv,nreq,count,yrank,*slens,*rlens,*sendto,*recvfrom,tag1,tag2;
599: PetscInt *rxindices,*ryindices;
600: MPI_Request *reqs,*sreqs,*rreqs;
602: /* Sorting makes code simpler, faster and also helps getting rid of many O(P) arrays, which hurt scalability at large scale
603: yindices_sorted - sorted yindices
604: xindices_sorted - xindices sorted along with yindces
605: */
606: ISGetLocalSize(ixx,&n); /*ixx, iyy have the same local size */
607: PetscMalloc2(n,&xindices_sorted,n,&yindices_sorted);
608: PetscArraycpy(xindices_sorted,xindices,n);
609: PetscArraycpy(yindices_sorted,yindices,n);
610: PetscSortIntWithArray(n,yindices_sorted,xindices_sorted);
611: VecGetOwnershipRange(xx,&xstart,NULL);
612: if (xcommsize == 1) {for (i=0; i<n; i++) xindices_sorted[i] += xstart;} /* Convert to global indices */
614: /*=============================================================================
615: Calculate info about messages I need to send
616: =============================================================================*/
617: /* nsend - number of non-empty messages to send
618: sendto - [nsend] ranks I will send messages to
619: sstart - [nsend+1] sstart[i] is the start index in xsindices_sorted[] I send to rank sendto[i]
620: slens - [ycommsize] I want to send slens[i] entries to rank i.
621: */
622: VecGetLayout(yy,&ylayout);
623: PetscLayoutGetRanges(ylayout,&yrange);
624: PetscCalloc1(ycommsize,&slens); /* The only O(P) array in this algorithm */
626: i = j = nsend = 0;
627: while (i < n) {
628: if (yindices_sorted[i] >= yrange[j+1]) { /* If i-th index is out of rank j's bound */
629: do {j++;} while (yindices_sorted[i] >= yrange[j+1] && j < ycommsize); /* Increase j until i-th index falls in rank j's bound */
630: if (j == ycommsize) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Index %D not owned by any process, upper bound %D",yindices_sorted[i],yrange[ycommsize]);
631: }
632: i++;
633: if (!slens[j]++) nsend++;
634: }
636: PetscMalloc2(nsend+1,&sstart,nsend,&sendto);
638: sstart[0] = 0;
639: for (i=j=0; i<ycommsize; i++) {
640: if (slens[i]) {
641: sendto[j] = (PetscMPIInt)i;
642: sstart[j+1] = sstart[j] + slens[i];
643: j++;
644: }
645: }
647: /*=============================================================================
648: Calculate the reverse info about messages I will recv
649: =============================================================================*/
650: /* nrecv - number of messages I will recv
651: recvfrom - [nrecv] ranks I recv from
652: rlens - [nrecv] I will recv rlens[i] entries from rank recvfrom[i]
653: rlentotal - sum of rlens[]
654: rxindices - [rlentotal] recv buffer for xindices_sorted
655: ryindices - [rlentotal] recv buffer for yindices_sorted
656: */
657: PetscGatherNumberOfMessages(ycomm,NULL,slens,&nrecv);
658: PetscGatherMessageLengths(ycomm,nsend,nrecv,slens,&recvfrom,&rlens);
659: PetscFree(slens); /* Free the O(P) array ASAP */
660: rlentotal = 0; for (i=0; i<nrecv; i++) rlentotal += rlens[i];
662: /*=============================================================================
663: Communicate with processors in recvfrom[] to populate rxindices and ryindices
664: ============================================================================*/
665: PetscCommGetNewTag(ycomm,&tag1);
666: PetscCommGetNewTag(ycomm,&tag2);
667: PetscMalloc2(rlentotal,&rxindices,rlentotal,&ryindices);
668: PetscMPIIntCast((nsend+nrecv)*2,&nreq);
669: PetscMalloc1(nreq,&reqs);
670: sreqs = reqs;
671: rreqs = reqs + nsend*2;
673: for (i=disp=0; i<nrecv; i++) {
674: count = rlens[i];
675: MPI_Irecv(rxindices+disp,count,MPIU_INT,recvfrom[i],tag1,ycomm,rreqs+i);
676: MPI_Irecv(ryindices+disp,count,MPIU_INT,recvfrom[i],tag2,ycomm,rreqs+nrecv+i);
677: disp += rlens[i];
678: }
680: for (i=0; i<nsend; i++) {
681: PetscMPIIntCast(sstart[i+1]-sstart[i],&count);
682: MPI_Isend(xindices_sorted+sstart[i],count,MPIU_INT,sendto[i],tag1,ycomm,sreqs+i);
683: MPI_Isend(yindices_sorted+sstart[i],count,MPIU_INT,sendto[i],tag2,ycomm,sreqs+nsend+i);
684: }
685: MPI_Waitall(nreq,reqs,MPI_STATUS_IGNORE);
687: /* Transform VecScatter into SF */
688: nleaves = rlentotal;
689: PetscMalloc1(nleaves,&ilocal);
690: PetscMalloc1(nleaves,&iremote);
691: MPI_Comm_rank(ycomm,&yrank);
692: for (i=disp=0; i<nrecv; i++) {
693: for (j=0; j<rlens[i]; j++) {
694: k = disp + j; /* k-th index pair */
695: ilocal[k] = ryindices[k] - yrange[yrank]; /* Convert y's global index to local index */
696: PetscLayoutFindOwnerIndex(xlayout,rxindices[k],&iremote[k].rank,&iremote[k].index); /* Convert x's global index to (rank, index) */
697: }
698: disp += rlens[i];
699: }
701: PetscFree2(sstart,sendto);
702: PetscFree(slens);
703: PetscFree(rlens);
704: PetscFree(recvfrom);
705: PetscFree(reqs);
706: PetscFree2(rxindices,ryindices);
707: PetscFree2(xindices_sorted,yindices_sorted);
708: #endif
709: } else {
710: /* PtoS or StoS */
711: ISGetLocalSize(iyy,&nleaves);
712: PetscMalloc1(nleaves,&ilocal);
713: PetscMalloc1(nleaves,&iremote);
714: PetscArraycpy(ilocal,yindices,nleaves);
715: for (i=0; i<nleaves; i++) {PetscLayoutFindOwnerIndex(xlayout,xindices[i],&iremote[i].rank,&iremote[i].index);}
716: }
718: /* MUST build SF on xx's comm, which is not necessarily identical to yy's comm.
719: In SF's view, xx contains the roots (i.e., the remote) and iremote[].rank are ranks in xx's comm.
720: yy contains leaves, which are local and can be thought as part of PETSC_COMM_SELF. */
721: PetscSFCreate(PetscObjectComm((PetscObject)xx),&data->sf);
722: PetscSFSetFromOptions(data->sf);
723: VecGetLocalSize(xx,&nroots);
724: PetscSFSetGraph(data->sf,nroots,nleaves,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER); /* Give ilocal/iremote to petsc and no need to free them here */
726: /* Free memory no longer needed */
727: ISRestoreIndices(ixx,&xindices);
728: ISRestoreIndices(iyy,&yindices);
729: if (can_do_block_opt) {
730: VecDestroy(&xx);
731: VecDestroy(&yy);
732: ISDestroy(&ixx);
733: ISDestroy(&iyy);
734: } else if (xcommsize == 1) {
735: VecDestroy(&xx);
736: }
738: functionend:
739: if (!vscat->from_is) {ISDestroy(&ix);}
740: if (!vscat->to_is ) {ISDestroy(&iy);}
742: /* vecscatter uses eager setup */
743: PetscSFSetUp(data->sf);
745: vscat->data = (void*)data;
746: vscat->ops->begin = VecScatterBegin_SF;
747: vscat->ops->end = VecScatterEnd_SF;
748: vscat->ops->remap = VecScatterRemap_SF;
749: vscat->ops->copy = VecScatterCopy_SF;
750: vscat->ops->destroy = VecScatterDestroy_SF;
751: vscat->ops->view = VecScatterView_SF;
752: vscat->ops->getremotecount = VecScatterGetRemoteCount_SF;
753: vscat->ops->getremote = VecScatterGetRemote_SF;
754: vscat->ops->getremoteordered = VecScatterGetRemoteOrdered_SF;
755: vscat->ops->restoreremote = VecScatterRestoreRemote_SF;
756: vscat->ops->restoreremoteordered = VecScatterRestoreRemoteOrdered_SF;
757: return(0);
758: }
760: PetscErrorCode VecScatterCreate_SF(VecScatter ctx)
761: {
765: ctx->ops->setup = VecScatterSetUp_SF;
766: PetscObjectChangeTypeName((PetscObject)ctx,VECSCATTERSF);
767: PetscInfo(ctx,"Using StarForest for vector scatter\n");
768: return(0);
769: }