18cd53115SBarry Smith 220c24465SJunchao Zhang #include "petscsf.h" 340e23c03SJunchao Zhang #include <../src/vec/is/sf/impls/basic/sfbasic.h> 4cd620004SJunchao Zhang #include <../src/vec/is/sf/impls/basic/sfpack.h> 5b23bfdefSJunchao Zhang 640e23c03SJunchao Zhang /*===================================================================================*/ 740e23c03SJunchao Zhang /* SF public interface implementations */ 840e23c03SJunchao Zhang /*===================================================================================*/ 940e23c03SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFSetUp_Basic(PetscSF sf) 1095fce210SBarry Smith { 1195fce210SBarry Smith PetscErrorCode ierr; 12b23bfdefSJunchao Zhang PetscSF_Basic *bas = (PetscSF_Basic*)sf->data; 1395fce210SBarry Smith PetscInt *rlengths,*ilengths,i; 1440e23c03SJunchao Zhang PetscMPIInt rank,niranks,*iranks,tag; 1595fce210SBarry Smith MPI_Comm comm; 16b5a8e515SJed Brown MPI_Group group; 1740e23c03SJunchao Zhang MPI_Request *rootreqs,*leafreqs; 1895fce210SBarry Smith 1995fce210SBarry Smith PetscFunctionBegin; 20*ffc4695bSBarry Smith ierr = MPI_Comm_group(PETSC_COMM_SELF,&group);CHKERRMPI(ierr); 21b5a8e515SJed Brown ierr = PetscSFSetUpRanks(sf,group);CHKERRQ(ierr); 22*ffc4695bSBarry Smith ierr = MPI_Group_free(&group);CHKERRMPI(ierr); 2395fce210SBarry Smith ierr = PetscObjectGetComm((PetscObject)sf,&comm);CHKERRQ(ierr); 2440e23c03SJunchao Zhang ierr = PetscObjectGetNewTag((PetscObject)sf,&tag);CHKERRQ(ierr); 25*ffc4695bSBarry Smith ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 2695fce210SBarry Smith /* 2795fce210SBarry Smith * Inform roots about how many leaves and from which ranks 2895fce210SBarry Smith */ 29785e854fSJed Brown ierr = PetscMalloc1(sf->nranks,&rlengths);CHKERRQ(ierr); 30cd620004SJunchao Zhang /* Determine number, sending ranks and length of incoming */ 3195fce210SBarry Smith for (i=0; i<sf->nranks; i++) { 3295fce210SBarry Smith rlengths[i] = sf->roffset[i+1] - sf->roffset[i]; /* Number of roots referenced by my leaves; for rank sf->ranks[i] */ 3395fce210SBarry Smith } 3440e23c03SJunchao Zhang ierr = PetscCommBuildTwoSided(comm,1,MPIU_INT,sf->nranks-sf->ndranks,sf->ranks+sf->ndranks,rlengths+sf->ndranks,&niranks,&iranks,(void**)&ilengths);CHKERRQ(ierr); 35c943f53fSJed Brown 360b899082SJunchao Zhang /* Sort iranks. See use of VecScatterGetRemoteOrdered_Private() in MatGetBrowsOfAoCols_MPIAIJ() on why. 370b899082SJunchao Zhang We could sort ranks there at the price of allocating extra working arrays. Presumably, niranks is 380b899082SJunchao Zhang small and the sorting is cheap. 390b899082SJunchao Zhang */ 400b899082SJunchao Zhang ierr = PetscSortMPIIntWithIntArray(niranks,iranks,ilengths);CHKERRQ(ierr); 410b899082SJunchao Zhang 42c943f53fSJed Brown /* Partition into distinguished and non-distinguished incoming ranks */ 43c943f53fSJed Brown bas->ndiranks = sf->ndranks; 44c943f53fSJed Brown bas->niranks = bas->ndiranks + niranks; 45c943f53fSJed Brown ierr = PetscMalloc2(bas->niranks,&bas->iranks,bas->niranks+1,&bas->ioffset);CHKERRQ(ierr); 46c943f53fSJed Brown bas->ioffset[0] = 0; 47c943f53fSJed Brown for (i=0; i<bas->ndiranks; i++) { 48c943f53fSJed Brown bas->iranks[i] = sf->ranks[i]; 49c943f53fSJed Brown bas->ioffset[i+1] = bas->ioffset[i] + rlengths[i]; 50c943f53fSJed Brown } 5140e23c03SJunchao Zhang if (bas->ndiranks > 1 || (bas->ndiranks == 1 && bas->iranks[0] != rank)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Broken setup for shared ranks"); 5240e23c03SJunchao Zhang for (; i<bas->niranks; i++) { 53c943f53fSJed Brown bas->iranks[i] = iranks[i-bas->ndiranks]; 54c943f53fSJed Brown bas->ioffset[i+1] = bas->ioffset[i] + ilengths[i-bas->ndiranks]; 55c943f53fSJed Brown } 56c943f53fSJed Brown bas->itotal = bas->ioffset[i]; 5795fce210SBarry Smith ierr = PetscFree(rlengths);CHKERRQ(ierr); 58c943f53fSJed Brown ierr = PetscFree(iranks);CHKERRQ(ierr); 59c943f53fSJed Brown ierr = PetscFree(ilengths);CHKERRQ(ierr); 6095fce210SBarry Smith 6195fce210SBarry Smith /* Send leaf identities to roots */ 62c943f53fSJed Brown ierr = PetscMalloc1(bas->itotal,&bas->irootloc);CHKERRQ(ierr); 6340e23c03SJunchao Zhang ierr = PetscMalloc2(bas->niranks-bas->ndiranks,&rootreqs,sf->nranks-sf->ndranks,&leafreqs);CHKERRQ(ierr); 6440e23c03SJunchao Zhang for (i=bas->ndiranks; i<bas->niranks; i++) { 65*ffc4695bSBarry Smith ierr = MPI_Irecv(bas->irootloc+bas->ioffset[i],bas->ioffset[i+1]-bas->ioffset[i],MPIU_INT,bas->iranks[i],tag,comm,&rootreqs[i-bas->ndiranks]);CHKERRMPI(ierr); 6640e23c03SJunchao Zhang } 6740e23c03SJunchao Zhang for (i=0; i<sf->nranks; i++) { 6895fce210SBarry Smith PetscMPIInt npoints; 6995fce210SBarry Smith ierr = PetscMPIIntCast(sf->roffset[i+1] - sf->roffset[i],&npoints);CHKERRQ(ierr); 7040e23c03SJunchao Zhang if (i < sf->ndranks) { 7140e23c03SJunchao Zhang if (sf->ranks[i] != rank) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot interpret distinguished leaf rank"); 7240e23c03SJunchao Zhang if (bas->iranks[0] != rank) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot interpret distinguished root rank"); 7340e23c03SJunchao Zhang if (npoints != bas->ioffset[1]-bas->ioffset[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Distinguished rank exchange has mismatched lengths"); 7440e23c03SJunchao Zhang ierr = PetscArraycpy(bas->irootloc+bas->ioffset[0],sf->rremote+sf->roffset[i],npoints);CHKERRQ(ierr); 75c943f53fSJed Brown continue; 76c943f53fSJed Brown } 77*ffc4695bSBarry Smith ierr = MPI_Isend(sf->rremote+sf->roffset[i],npoints,MPIU_INT,sf->ranks[i],tag,comm,&leafreqs[i-sf->ndranks]);CHKERRMPI(ierr); 78bf39f1bfSJed Brown } 79*ffc4695bSBarry Smith ierr = MPI_Waitall(bas->niranks-bas->ndiranks,rootreqs,MPI_STATUSES_IGNORE);CHKERRMPI(ierr); 80*ffc4695bSBarry Smith ierr = MPI_Waitall(sf->nranks-sf->ndranks,leafreqs,MPI_STATUSES_IGNORE);CHKERRMPI(ierr); 8140e23c03SJunchao Zhang ierr = PetscFree2(rootreqs,leafreqs);CHKERRQ(ierr); 8295fce210SBarry Smith 83cd620004SJunchao Zhang sf->nleafreqs = sf->nranks - sf->ndranks; 84cd620004SJunchao Zhang bas->nrootreqs = bas->niranks - bas->ndiranks; 85cd620004SJunchao Zhang sf->persistent = PETSC_TRUE; 86eb02082bSJunchao Zhang 87cd620004SJunchao Zhang /* Setup fields related to packing */ 88cd620004SJunchao Zhang ierr = PetscSFSetUpPackFields(sf);CHKERRQ(ierr); 8995fce210SBarry Smith PetscFunctionReturn(0); 9095fce210SBarry Smith } 9195fce210SBarry Smith 9240e23c03SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFReset_Basic(PetscSF sf) 9395fce210SBarry Smith { 9495fce210SBarry Smith PetscErrorCode ierr; 95cd620004SJunchao Zhang PetscSF_Basic *bas = (PetscSF_Basic*)sf->data; 9695fce210SBarry Smith 9795fce210SBarry Smith PetscFunctionBegin; 9829046d53SLisandro Dalcin if (bas->inuse) SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_ARG_WRONGSTATE,"Outstanding operation has not been completed"); 99c943f53fSJed Brown ierr = PetscFree2(bas->iranks,bas->ioffset);CHKERRQ(ierr); 100c943f53fSJed Brown ierr = PetscFree(bas->irootloc);CHKERRQ(ierr); 1017fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_DEVICE) 10220c24465SJunchao Zhang for (PetscInt i=0; i<2; i++) {ierr = PetscSFFree(sf,PETSC_MEMTYPE_DEVICE,bas->irootloc_d[i]);CHKERRQ(ierr);} 103eb02082bSJunchao Zhang #endif 104cd620004SJunchao Zhang ierr = PetscSFLinkDestroy(sf,&bas->avail);CHKERRQ(ierr); 105cd620004SJunchao Zhang ierr = PetscSFResetPackFields(sf);CHKERRQ(ierr); 10695fce210SBarry Smith PetscFunctionReturn(0); 10795fce210SBarry Smith } 10895fce210SBarry Smith 10940e23c03SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFDestroy_Basic(PetscSF sf) 11095fce210SBarry Smith { 11195fce210SBarry Smith PetscErrorCode ierr; 11295fce210SBarry Smith 11395fce210SBarry Smith PetscFunctionBegin; 114f6d956f6SStefano Zampini ierr = PetscSFReset_Basic(sf);CHKERRQ(ierr); 11595fce210SBarry Smith ierr = PetscFree(sf->data);CHKERRQ(ierr); 11695fce210SBarry Smith PetscFunctionReturn(0); 11795fce210SBarry Smith } 11895fce210SBarry Smith 11940e23c03SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFView_Basic(PetscSF sf,PetscViewer viewer) 12095fce210SBarry Smith { 12195fce210SBarry Smith PetscErrorCode ierr; 12295fce210SBarry Smith PetscBool iascii; 12395fce210SBarry Smith 12495fce210SBarry Smith PetscFunctionBegin; 12595fce210SBarry Smith ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 126b23bfdefSJunchao Zhang if (iascii) {ierr = PetscViewerASCIIPrintf(viewer," sort=%s\n",sf->rankorder ? "rank-order" : "unordered");CHKERRQ(ierr);} 12795fce210SBarry Smith PetscFunctionReturn(0); 12895fce210SBarry Smith } 12995fce210SBarry Smith 130eb02082bSJunchao Zhang static PetscErrorCode PetscSFBcastAndOpBegin_Basic(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,const void *rootdata,PetscMemType leafmtype,void *leafdata,MPI_Op op) 13195fce210SBarry Smith { 13295fce210SBarry Smith PetscErrorCode ierr; 133cd620004SJunchao Zhang PetscSFLink link = NULL; 134851d6770SJunchao Zhang MPI_Request *rootreqs = NULL,*leafreqs = NULL; 135cd620004SJunchao Zhang PetscSF_Basic *bas = (PetscSF_Basic*)sf->data; 13695fce210SBarry Smith 13795fce210SBarry Smith PetscFunctionBegin; 138cd620004SJunchao Zhang /* Create a communication link, which provides buffers & MPI requests etc */ 139cd620004SJunchao Zhang ierr = PetscSFLinkCreate(sf,unit,rootmtype,rootdata,leafmtype,leafdata,op,PETSCSF_BCAST,&link);CHKERRQ(ierr); 140cd620004SJunchao Zhang /* Get MPI requests from the link. We do not need buffers explicitly since we use persistent MPI */ 141cd620004SJunchao Zhang ierr = PetscSFLinkGetMPIBuffersAndRequests(sf,link,PETSCSF_ROOT2LEAF,NULL,NULL,&rootreqs,&leafreqs);CHKERRQ(ierr); 142cd620004SJunchao Zhang /* Post Irecv for remote */ 143*ffc4695bSBarry Smith ierr = MPI_Startall_irecv(sf->leafbuflen[PETSCSF_REMOTE],unit,sf->nleafreqs,leafreqs);CHKERRMPI(ierr); 144cd620004SJunchao Zhang /* Pack rootdata and do Isend for remote */ 145cd620004SJunchao Zhang ierr = PetscSFLinkPackRootData(sf,link,PETSCSF_REMOTE,rootdata);CHKERRQ(ierr); 146*ffc4695bSBarry Smith ierr = MPI_Startall_isend(bas->rootbuflen[PETSCSF_REMOTE],unit,bas->nrootreqs,rootreqs);CHKERRMPI(ierr); 147cd620004SJunchao Zhang /* Do local BcastAndOp, which overlaps with the irecv/isend above */ 148cd620004SJunchao Zhang ierr = PetscSFLinkBcastAndOpLocal(sf,link,rootdata,leafdata,op);CHKERRQ(ierr); 14995fce210SBarry Smith PetscFunctionReturn(0); 15095fce210SBarry Smith } 15195fce210SBarry Smith 15200816365SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFBcastAndOpEnd_Basic(PetscSF sf,MPI_Datatype unit,const void *rootdata,void *leafdata,MPI_Op op) 15395fce210SBarry Smith { 15495fce210SBarry Smith PetscErrorCode ierr; 155cd620004SJunchao Zhang PetscSFLink link = NULL; 15695fce210SBarry Smith 15795fce210SBarry Smith PetscFunctionBegin; 158cd620004SJunchao Zhang /* Retrieve the link used in XxxBegin() with root/leafdata as key */ 159cd620004SJunchao Zhang ierr = PetscSFLinkGetInUse(sf,unit,rootdata,leafdata,PETSC_OWN_POINTER,&link);CHKERRQ(ierr); 160cd620004SJunchao Zhang /* Wait for the completion of mpi */ 161cd620004SJunchao Zhang ierr = PetscSFLinkMPIWaitall(sf,link,PETSCSF_ROOT2LEAF);CHKERRQ(ierr); 162cd620004SJunchao Zhang /* Unpack leafdata and reclaim the link */ 163cd620004SJunchao Zhang ierr = PetscSFLinkUnpackLeafData(sf,link,PETSCSF_REMOTE,leafdata,op);CHKERRQ(ierr); 164cd620004SJunchao Zhang ierr = PetscSFLinkReclaim(sf,&link);CHKERRQ(ierr); 165cd620004SJunchao Zhang PetscFunctionReturn(0); 166cd620004SJunchao Zhang } 167cd620004SJunchao Zhang 168cd620004SJunchao Zhang /* Shared by ReduceBegin and FetchAndOpBegin */ 169cd620004SJunchao Zhang PETSC_STATIC_INLINE PetscErrorCode PetscSFLeafToRootBegin_Basic(PetscSF sf,MPI_Datatype unit,PetscMemType leafmtype,const void *leafdata,PetscMemType rootmtype,void *rootdata,MPI_Op op,PetscSFOperation sfop,PetscSFLink *out) 170cd620004SJunchao Zhang { 171cd620004SJunchao Zhang PetscErrorCode ierr; 172cd620004SJunchao Zhang PetscSFLink link; 173cd620004SJunchao Zhang PetscSF_Basic *bas = (PetscSF_Basic*)sf->data; 174cd620004SJunchao Zhang MPI_Request *rootreqs = NULL,*leafreqs = NULL; 175cd620004SJunchao Zhang 176cd620004SJunchao Zhang PetscFunctionBegin; 177cd620004SJunchao Zhang ierr = PetscSFLinkCreate(sf,unit,rootmtype,rootdata,leafmtype,leafdata,op,sfop,&link);CHKERRQ(ierr); 178cd620004SJunchao Zhang ierr = PetscSFLinkGetMPIBuffersAndRequests(sf,link,PETSCSF_LEAF2ROOT,NULL,NULL,&rootreqs,&leafreqs);CHKERRQ(ierr); 179*ffc4695bSBarry Smith ierr = MPI_Startall_irecv(bas->rootbuflen[PETSCSF_REMOTE],unit,bas->nrootreqs,rootreqs);CHKERRMPI(ierr); 180cd620004SJunchao Zhang ierr = PetscSFLinkPackLeafData(sf,link,PETSCSF_REMOTE,leafdata);CHKERRQ(ierr); 181*ffc4695bSBarry Smith ierr = MPI_Startall_isend(sf->leafbuflen[PETSCSF_REMOTE],unit,sf->nleafreqs,leafreqs);CHKERRMPI(ierr); 182cd620004SJunchao Zhang *out = link; 18395fce210SBarry Smith PetscFunctionReturn(0); 18495fce210SBarry Smith } 18595fce210SBarry Smith 18695fce210SBarry Smith /* leaf -> root with reduction */ 187eb02082bSJunchao Zhang static PetscErrorCode PetscSFReduceBegin_Basic(PetscSF sf,MPI_Datatype unit,PetscMemType leafmtype,const void *leafdata,PetscMemType rootmtype,void *rootdata,MPI_Op op) 18895fce210SBarry Smith { 18995fce210SBarry Smith PetscErrorCode ierr; 190cd620004SJunchao Zhang PetscSFLink link = NULL; 19195fce210SBarry Smith 19295fce210SBarry Smith PetscFunctionBegin; 193cd620004SJunchao Zhang ierr = PetscSFLeafToRootBegin_Basic(sf,unit,leafmtype,leafdata,rootmtype,rootdata,op,PETSCSF_REDUCE,&link);CHKERRQ(ierr); 194cd620004SJunchao Zhang ierr = PetscSFLinkReduceLocal(sf,link,leafdata,rootdata,op);CHKERRQ(ierr); 19595fce210SBarry Smith PetscFunctionReturn(0); 19695fce210SBarry Smith } 19795fce210SBarry Smith 19800816365SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFReduceEnd_Basic(PetscSF sf,MPI_Datatype unit,const void *leafdata,void *rootdata,MPI_Op op) 19995fce210SBarry Smith { 20095fce210SBarry Smith PetscErrorCode ierr; 201cd620004SJunchao Zhang PetscSFLink link = NULL; 20295fce210SBarry Smith 20395fce210SBarry Smith PetscFunctionBegin; 204cd620004SJunchao Zhang ierr = PetscSFLinkGetInUse(sf,unit,rootdata,leafdata,PETSC_OWN_POINTER,&link);CHKERRQ(ierr); 205cd620004SJunchao Zhang ierr = PetscSFLinkMPIWaitall(sf,link,PETSCSF_LEAF2ROOT);CHKERRQ(ierr); 206cd620004SJunchao Zhang ierr = PetscSFLinkUnpackRootData(sf,link,PETSCSF_REMOTE,rootdata,op);CHKERRQ(ierr); 207cd620004SJunchao Zhang ierr = PetscSFLinkReclaim(sf,&link);CHKERRQ(ierr); 20895fce210SBarry Smith PetscFunctionReturn(0); 20995fce210SBarry Smith } 21095fce210SBarry Smith 211eb02082bSJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFFetchAndOpBegin_Basic(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,void *rootdata,PetscMemType leafmtype,const void *leafdata,void *leafupdate,MPI_Op op) 21295fce210SBarry Smith { 21395fce210SBarry Smith PetscErrorCode ierr; 214cd620004SJunchao Zhang PetscSFLink link = NULL; 21595fce210SBarry Smith 21695fce210SBarry Smith PetscFunctionBegin; 217cd620004SJunchao Zhang ierr = PetscSFLeafToRootBegin_Basic(sf,unit,leafmtype,leafdata,rootmtype,rootdata,op,PETSCSF_FETCH,&link);CHKERRQ(ierr); 218cd620004SJunchao Zhang ierr = PetscSFLinkFetchAndOpLocal(sf,link,rootdata,leafdata,leafupdate,op);CHKERRQ(ierr); 21995fce210SBarry Smith PetscFunctionReturn(0); 22095fce210SBarry Smith } 22195fce210SBarry Smith 22200816365SJunchao Zhang static PetscErrorCode PetscSFFetchAndOpEnd_Basic(PetscSF sf,MPI_Datatype unit,void *rootdata,const void *leafdata,void *leafupdate,MPI_Op op) 22395fce210SBarry Smith { 22495fce210SBarry Smith PetscErrorCode ierr; 225cd620004SJunchao Zhang PetscSFLink link = NULL; 226851d6770SJunchao Zhang MPI_Request *rootreqs = NULL,*leafreqs = NULL; 227cd620004SJunchao Zhang PetscSF_Basic *bas = (PetscSF_Basic*)sf->data; 22895fce210SBarry Smith 22995fce210SBarry Smith PetscFunctionBegin; 230cd620004SJunchao Zhang ierr = PetscSFLinkGetInUse(sf,unit,rootdata,leafdata,PETSC_OWN_POINTER,&link);CHKERRQ(ierr); 23195fce210SBarry Smith /* This implementation could be changed to unpack as receives arrive, at the cost of non-determinism */ 232cd620004SJunchao Zhang ierr = PetscSFLinkMPIWaitall(sf,link,PETSCSF_LEAF2ROOT);CHKERRQ(ierr); 233cd620004SJunchao Zhang /* Do fetch-and-op, the (remote) update results are in rootbuf */ 234cd620004SJunchao Zhang ierr = PetscSFLinkFetchRootData(sf,link,PETSCSF_REMOTE,rootdata,op);CHKERRQ(ierr); 23540e23c03SJunchao Zhang 236cd620004SJunchao Zhang /* Bcast rootbuf to leafupdate */ 237cd620004SJunchao Zhang ierr = PetscSFLinkGetMPIBuffersAndRequests(sf,link,PETSCSF_ROOT2LEAF,NULL,NULL,&rootreqs,&leafreqs);CHKERRQ(ierr); 238cd620004SJunchao Zhang /* Post leaf receives and root sends */ 239*ffc4695bSBarry Smith ierr = MPI_Startall_irecv(sf->leafbuflen[PETSCSF_REMOTE],unit,sf->nleafreqs,leafreqs);CHKERRMPI(ierr); 240*ffc4695bSBarry Smith ierr = MPI_Startall_isend(bas->rootbuflen[PETSCSF_REMOTE],unit,bas->nrootreqs,rootreqs);CHKERRMPI(ierr); 241b23bfdefSJunchao Zhang /* Unpack and insert fetched data into leaves */ 242cd620004SJunchao Zhang ierr = PetscSFLinkMPIWaitall(sf,link,PETSCSF_ROOT2LEAF);CHKERRQ(ierr); 243cd620004SJunchao Zhang ierr = PetscSFLinkUnpackLeafData(sf,link,PETSCSF_REMOTE,leafupdate,MPIU_REPLACE);CHKERRQ(ierr); 244cd620004SJunchao Zhang ierr = PetscSFLinkReclaim(sf,&link);CHKERRQ(ierr); 24595fce210SBarry Smith PetscFunctionReturn(0); 24695fce210SBarry Smith } 24795fce210SBarry Smith 24840e23c03SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFGetLeafRanks_Basic(PetscSF sf,PetscInt *niranks,const PetscMPIInt **iranks,const PetscInt **ioffset,const PetscInt **irootloc) 2498750ddebSJunchao Zhang { 2508750ddebSJunchao Zhang PetscSF_Basic *bas = (PetscSF_Basic*)sf->data; 2518750ddebSJunchao Zhang 2528750ddebSJunchao Zhang PetscFunctionBegin; 2538750ddebSJunchao Zhang if (niranks) *niranks = bas->niranks; 2548750ddebSJunchao Zhang if (iranks) *iranks = bas->iranks; 2558750ddebSJunchao Zhang if (ioffset) *ioffset = bas->ioffset; 2568750ddebSJunchao Zhang if (irootloc) *irootloc = bas->irootloc; 2578750ddebSJunchao Zhang PetscFunctionReturn(0); 2588750ddebSJunchao Zhang } 2598750ddebSJunchao Zhang 260f659e5c7SJunchao Zhang /* An optimized PetscSFCreateEmbeddedSF. We aggresively make use of the established communication on sf. 261f659e5c7SJunchao Zhang We need one bcast on sf, and no communication anymore to build the embedded sf. Note that selected[] 262f659e5c7SJunchao Zhang was sorted before calling the routine. 263f659e5c7SJunchao Zhang */ 264f659e5c7SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFCreateEmbeddedSF_Basic(PetscSF sf,PetscInt nselected,const PetscInt *selected,PetscSF *newsf) 265f659e5c7SJunchao Zhang { 266f659e5c7SJunchao Zhang PetscSF esf; 267cd620004SJunchao Zhang PetscInt esf_nranks,esf_ndranks,*esf_roffset,*esf_rmine,*esf_rremote; 268cd620004SJunchao Zhang PetscInt i,j,p,q,nroots,esf_nleaves,*new_ilocal,nranks,ndranks,niranks,ndiranks,minleaf,maxleaf,maxlocal; 269cd620004SJunchao Zhang char *rootdata,*leafdata,*leafmem; /* Only stores 0 or 1, so we can save memory with char */ 270f659e5c7SJunchao Zhang PetscMPIInt *esf_ranks; 271f659e5c7SJunchao Zhang const PetscMPIInt *ranks,*iranks; 272cd620004SJunchao Zhang const PetscInt *roffset,*rmine,*rremote,*ioffset,*irootloc; 273f659e5c7SJunchao Zhang PetscBool connected; 274f659e5c7SJunchao Zhang PetscSFNode *new_iremote; 275f659e5c7SJunchao Zhang PetscSF_Basic *bas; 276f659e5c7SJunchao Zhang PetscErrorCode ierr; 277f659e5c7SJunchao Zhang 278f659e5c7SJunchao Zhang PetscFunctionBegin; 279f659e5c7SJunchao Zhang ierr = PetscSFCreate(PetscObjectComm((PetscObject)sf),&esf);CHKERRQ(ierr); 28020c24465SJunchao Zhang ierr = PetscSFSetFromOptions(esf);CHKERRQ(ierr); 281f659e5c7SJunchao Zhang ierr = PetscSFSetType(esf,PETSCSFBASIC);CHKERRQ(ierr); /* This optimized routine can only create a basic sf */ 282f659e5c7SJunchao Zhang 283cd620004SJunchao Zhang /* Find out which leaves are still connected to roots in the embedded sf by doing a Bcast */ 284f659e5c7SJunchao Zhang ierr = PetscSFGetGraph(sf,&nroots,NULL,NULL,NULL);CHKERRQ(ierr); 285f659e5c7SJunchao Zhang ierr = PetscSFGetLeafRange(sf,&minleaf,&maxleaf);CHKERRQ(ierr); 286cd620004SJunchao Zhang maxlocal = maxleaf - minleaf + 1; 287cd620004SJunchao Zhang ierr = PetscCalloc2(nroots,&rootdata,maxlocal,&leafmem);CHKERRQ(ierr); 288cd620004SJunchao Zhang leafdata = leafmem - minleaf; 289f659e5c7SJunchao Zhang /* Tag selected roots */ 290f659e5c7SJunchao Zhang for (i=0; i<nselected; ++i) rootdata[selected[i]] = 1; 291f659e5c7SJunchao Zhang 292cd620004SJunchao Zhang ierr = PetscSFBcastBegin(sf,MPI_CHAR,rootdata,leafdata);CHKERRQ(ierr); 293cd620004SJunchao Zhang ierr = PetscSFBcastEnd(sf,MPI_CHAR,rootdata,leafdata);CHKERRQ(ierr); 294f659e5c7SJunchao Zhang ierr = PetscSFGetLeafInfo_Basic(sf,&nranks,&ndranks,&ranks,&roffset,&rmine,&rremote);CHKERRQ(ierr); /* Get send info */ 295cd620004SJunchao Zhang esf_nranks = esf_ndranks = esf_nleaves = 0; 296b23bfdefSJunchao Zhang for (i=0; i<nranks; i++) { 297cd620004SJunchao Zhang connected = PETSC_FALSE; /* Is this process still connected to this remote root rank? */ 298cd620004SJunchao Zhang for (j=roffset[i]; j<roffset[i+1]; j++) {if (leafdata[rmine[j]]) {esf_nleaves++; connected = PETSC_TRUE;}} 299f659e5c7SJunchao Zhang if (connected) {esf_nranks++; if (i < ndranks) esf_ndranks++;} 300f659e5c7SJunchao Zhang } 301f659e5c7SJunchao Zhang 302f659e5c7SJunchao Zhang /* Set graph of esf and also set up its outgoing communication (i.e., send info), which is usually done by PetscSFSetUpRanks */ 303cd620004SJunchao Zhang ierr = PetscMalloc1(esf_nleaves,&new_ilocal);CHKERRQ(ierr); 304cd620004SJunchao Zhang ierr = PetscMalloc1(esf_nleaves,&new_iremote);CHKERRQ(ierr); 305cd620004SJunchao Zhang ierr = PetscMalloc4(esf_nranks,&esf_ranks,esf_nranks+1,&esf_roffset,esf_nleaves,&esf_rmine,esf_nleaves,&esf_rremote);CHKERRQ(ierr); 306f659e5c7SJunchao Zhang p = 0; /* Counter for connected root ranks */ 307f659e5c7SJunchao Zhang q = 0; /* Counter for connected leaves */ 308f659e5c7SJunchao Zhang esf_roffset[0] = 0; 309f659e5c7SJunchao Zhang for (i=0; i<nranks; i++) { /* Scan leaf data again to fill esf arrays */ 310f659e5c7SJunchao Zhang connected = PETSC_FALSE; 311cd620004SJunchao Zhang for (j=roffset[i]; j<roffset[i+1]; j++) { 312cd620004SJunchao Zhang if (leafdata[rmine[j]]) { 313f659e5c7SJunchao Zhang esf_rmine[q] = new_ilocal[q] = rmine[j]; 314f659e5c7SJunchao Zhang esf_rremote[q] = rremote[j]; 315f659e5c7SJunchao Zhang new_iremote[q].index = rremote[j]; 316f659e5c7SJunchao Zhang new_iremote[q].rank = ranks[i]; 317f659e5c7SJunchao Zhang connected = PETSC_TRUE; 318f659e5c7SJunchao Zhang q++; 319f659e5c7SJunchao Zhang } 320f659e5c7SJunchao Zhang } 321f659e5c7SJunchao Zhang if (connected) { 322f659e5c7SJunchao Zhang esf_ranks[p] = ranks[i]; 323f659e5c7SJunchao Zhang esf_roffset[p+1] = q; 324f659e5c7SJunchao Zhang p++; 325f659e5c7SJunchao Zhang } 326f659e5c7SJunchao Zhang } 327f659e5c7SJunchao Zhang 328f659e5c7SJunchao Zhang /* SetGraph internally resets the SF, so we only set its fields after the call */ 329cd620004SJunchao Zhang ierr = PetscSFSetGraph(esf,nroots,esf_nleaves,new_ilocal,PETSC_OWN_POINTER,new_iremote,PETSC_OWN_POINTER);CHKERRQ(ierr); 330f659e5c7SJunchao Zhang esf->nranks = esf_nranks; 331f659e5c7SJunchao Zhang esf->ndranks = esf_ndranks; 332f659e5c7SJunchao Zhang esf->ranks = esf_ranks; 333f659e5c7SJunchao Zhang esf->roffset = esf_roffset; 334f659e5c7SJunchao Zhang esf->rmine = esf_rmine; 335f659e5c7SJunchao Zhang esf->rremote = esf_rremote; 336cd620004SJunchao Zhang esf->nleafreqs = esf_nranks - esf_ndranks; 337f659e5c7SJunchao Zhang 338f659e5c7SJunchao Zhang /* Set up the incoming communication (i.e., recv info) stored in esf->data, which is usually done by PetscSFSetUp_Basic */ 339f659e5c7SJunchao Zhang bas = (PetscSF_Basic*)esf->data; 340f659e5c7SJunchao Zhang ierr = PetscSFGetRootInfo_Basic(sf,&niranks,&ndiranks,&iranks,&ioffset,&irootloc);CHKERRQ(ierr); /* Get recv info */ 341f659e5c7SJunchao Zhang /* Embedded sf always has simpler communication than the original one. We might allocate longer arrays than needed here. But we 342cd620004SJunchao Zhang we do not care since these arrays are usually short. The benefit is we can fill these arrays by just parsing irootloc once. 343f659e5c7SJunchao Zhang */ 344f659e5c7SJunchao Zhang ierr = PetscMalloc2(niranks,&bas->iranks,niranks+1,&bas->ioffset);CHKERRQ(ierr); 345f659e5c7SJunchao Zhang ierr = PetscMalloc1(ioffset[niranks],&bas->irootloc);CHKERRQ(ierr); 346f659e5c7SJunchao Zhang bas->niranks = bas->ndiranks = bas->ioffset[0] = 0; 347f659e5c7SJunchao Zhang p = 0; /* Counter for connected leaf ranks */ 348f659e5c7SJunchao Zhang q = 0; /* Counter for connected roots */ 349f659e5c7SJunchao Zhang for (i=0; i<niranks; i++) { 350f659e5c7SJunchao Zhang connected = PETSC_FALSE; /* Is the current process still connected to this remote leaf rank? */ 351f659e5c7SJunchao Zhang for (j=ioffset[i]; j<ioffset[i+1]; j++) { 352cd620004SJunchao Zhang if (rootdata[irootloc[j]]) { 353f659e5c7SJunchao Zhang bas->irootloc[q++] = irootloc[j]; 354f659e5c7SJunchao Zhang connected = PETSC_TRUE; 355f659e5c7SJunchao Zhang } 356f659e5c7SJunchao Zhang } 357f659e5c7SJunchao Zhang if (connected) { 358f659e5c7SJunchao Zhang bas->niranks++; 359f659e5c7SJunchao Zhang if (i<ndiranks) bas->ndiranks++; /* Note that order of ranks (including distinguished ranks) is kept */ 360f659e5c7SJunchao Zhang bas->iranks[p] = iranks[i]; 361f659e5c7SJunchao Zhang bas->ioffset[p+1] = q; 362f659e5c7SJunchao Zhang p++; 363f659e5c7SJunchao Zhang } 364f659e5c7SJunchao Zhang } 365f659e5c7SJunchao Zhang bas->itotal = q; 366cd620004SJunchao Zhang bas->nrootreqs = bas->niranks - bas->ndiranks; 367cd620004SJunchao Zhang esf->persistent = PETSC_TRUE; 368cd620004SJunchao Zhang /* Setup packing related fields */ 369cd620004SJunchao Zhang ierr = PetscSFSetUpPackFields(esf);CHKERRQ(ierr); 370f659e5c7SJunchao Zhang 37120c24465SJunchao Zhang /* Copy from PetscSFSetUp(), since this method wants to skip PetscSFSetUp(). */ 37220c24465SJunchao Zhang #if defined(PETSC_HAVE_CUDA) 37320c24465SJunchao Zhang if (esf->backend == PETSCSF_BACKEND_CUDA) { 37420c24465SJunchao Zhang esf->ops->Malloc = PetscSFMalloc_Cuda; 37520c24465SJunchao Zhang esf->ops->Free = PetscSFFree_Cuda; 37620c24465SJunchao Zhang } 37720c24465SJunchao Zhang #endif 37820c24465SJunchao Zhang 37959af0bd3SScott Kruger #if defined(PETSC_HAVE_HIP) 38059af0bd3SScott Kruger /* TODO: Needs debugging */ 38159af0bd3SScott Kruger if (esf->backend == PETSCSF_BACKEND_HIP) { 38259af0bd3SScott Kruger esf->ops->Malloc = PetscSFMalloc_HIP; 38359af0bd3SScott Kruger esf->ops->Free = PetscSFFree_HIP; 38459af0bd3SScott Kruger } 38559af0bd3SScott Kruger #endif 38659af0bd3SScott Kruger 38720c24465SJunchao Zhang #if defined(PETSC_HAVE_KOKKOS) 38820c24465SJunchao Zhang if (esf->backend == PETSCSF_BACKEND_KOKKOS) { 38920c24465SJunchao Zhang esf->ops->Malloc = PetscSFMalloc_Kokkos; 39020c24465SJunchao Zhang esf->ops->Free = PetscSFFree_Kokkos; 39120c24465SJunchao Zhang } 39220c24465SJunchao Zhang #endif 393f659e5c7SJunchao Zhang esf->setupcalled = PETSC_TRUE; /* We have done setup ourselves! */ 394cd620004SJunchao Zhang ierr = PetscFree2(rootdata,leafmem);CHKERRQ(ierr); 395f659e5c7SJunchao Zhang *newsf = esf; 396f659e5c7SJunchao Zhang PetscFunctionReturn(0); 397f659e5c7SJunchao Zhang } 398f659e5c7SJunchao Zhang 3998cc058d9SJed Brown PETSC_EXTERN PetscErrorCode PetscSFCreate_Basic(PetscSF sf) 40095fce210SBarry Smith { 40140e23c03SJunchao Zhang PetscSF_Basic *dat; 40295fce210SBarry Smith PetscErrorCode ierr; 40395fce210SBarry Smith 40495fce210SBarry Smith PetscFunctionBegin; 40595fce210SBarry Smith sf->ops->SetUp = PetscSFSetUp_Basic; 40695fce210SBarry Smith sf->ops->Reset = PetscSFReset_Basic; 40795fce210SBarry Smith sf->ops->Destroy = PetscSFDestroy_Basic; 40895fce210SBarry Smith sf->ops->View = PetscSFView_Basic; 4093482bfa8SJunchao Zhang sf->ops->BcastAndOpBegin = PetscSFBcastAndOpBegin_Basic; 4103482bfa8SJunchao Zhang sf->ops->BcastAndOpEnd = PetscSFBcastAndOpEnd_Basic; 41195fce210SBarry Smith sf->ops->ReduceBegin = PetscSFReduceBegin_Basic; 41295fce210SBarry Smith sf->ops->ReduceEnd = PetscSFReduceEnd_Basic; 41395fce210SBarry Smith sf->ops->FetchAndOpBegin = PetscSFFetchAndOpBegin_Basic; 41495fce210SBarry Smith sf->ops->FetchAndOpEnd = PetscSFFetchAndOpEnd_Basic; 4158750ddebSJunchao Zhang sf->ops->GetLeafRanks = PetscSFGetLeafRanks_Basic; 416f659e5c7SJunchao Zhang sf->ops->CreateEmbeddedSF = PetscSFCreateEmbeddedSF_Basic; 41795fce210SBarry Smith 41840e23c03SJunchao Zhang ierr = PetscNewLog(sf,&dat);CHKERRQ(ierr); 41940e23c03SJunchao Zhang sf->data = (void*)dat; 42095fce210SBarry Smith PetscFunctionReturn(0); 42195fce210SBarry Smith } 422