1af0996ceSBarry Smith #include <petsc/private/sfimpl.h> /*I "petscsf.h" I*/ 295fce210SBarry Smith 395fce210SBarry Smith typedef struct _n_PetscSFDataLink *PetscSFDataLink; 495fce210SBarry Smith typedef struct _n_PetscSFWinLink *PetscSFWinLink; 595fce210SBarry Smith 695fce210SBarry Smith typedef struct { 795fce210SBarry Smith PetscSFWindowSyncType sync; /* FENCE, LOCK, or ACTIVE synchronization */ 85b0d146aSStefano Zampini PetscSFDataLink link; /* List of MPI data types, lazily constructed for each data type */ 995fce210SBarry Smith PetscSFWinLink wins; /* List of active windows */ 105b0d146aSStefano Zampini PetscSFWindowFlavorType flavor; /* Current PETSCSF_WINDOW_FLAVOR_ */ 115b0d146aSStefano Zampini PetscSF dynsf; 125b0d146aSStefano Zampini MPI_Info info; 1395fce210SBarry Smith } PetscSF_Window; 1495fce210SBarry Smith 1595fce210SBarry Smith struct _n_PetscSFDataLink { 1695fce210SBarry Smith MPI_Datatype unit; 1795fce210SBarry Smith MPI_Datatype *mine; 1895fce210SBarry Smith MPI_Datatype *remote; 1995fce210SBarry Smith PetscSFDataLink next; 2095fce210SBarry Smith }; 2195fce210SBarry Smith 2295fce210SBarry Smith struct _n_PetscSFWinLink { 2395fce210SBarry Smith PetscBool inuse; 2495fce210SBarry Smith size_t bytes; 2595fce210SBarry Smith void *addr; 265b0d146aSStefano Zampini void *paddr; 2795fce210SBarry Smith MPI_Win win; 28684a874aSStefano Zampini MPI_Request *reqs; 295b0d146aSStefano Zampini PetscSFWindowFlavorType flavor; 305b0d146aSStefano Zampini MPI_Aint *dyn_target_addr; 3195fce210SBarry Smith PetscBool epoch; 3295fce210SBarry Smith PetscSFWinLink next; 3395fce210SBarry Smith }; 3495fce210SBarry Smith 354c8fdceaSLisandro Dalcin const char *const PetscSFWindowSyncTypes[] = {"FENCE","LOCK","ACTIVE","PetscSFWindowSyncType","PETSCSF_WINDOW_SYNC_",NULL}; 364c8fdceaSLisandro Dalcin const char *const PetscSFWindowFlavorTypes[] = {"CREATE","DYNAMIC","ALLOCATE","SHARED","PetscSFWindowFlavorType","PETSCSF_WINDOW_FLAVOR_",NULL}; 3795fce210SBarry Smith 38820f2d46SBarry Smith /* Built-in MPI_Ops act elementwise inside MPI_Accumulate, but cannot be used with composite types inside collectives (MPI_Allreduce) */ 3995fce210SBarry Smith static PetscErrorCode PetscSFWindowOpTranslate(MPI_Op *op) 4095fce210SBarry Smith { 4195fce210SBarry Smith PetscFunctionBegin; 4295fce210SBarry Smith if (*op == MPIU_SUM) *op = MPI_SUM; 4395fce210SBarry Smith else if (*op == MPIU_MAX) *op = MPI_MAX; 4495fce210SBarry Smith else if (*op == MPIU_MIN) *op = MPI_MIN; 4595fce210SBarry Smith PetscFunctionReturn(0); 4695fce210SBarry Smith } 4795fce210SBarry Smith 4895fce210SBarry Smith /*@C 4995fce210SBarry Smith PetscSFWindowGetDataTypes - gets composite local and remote data types for each rank 5095fce210SBarry Smith 5195fce210SBarry Smith Not Collective 5295fce210SBarry Smith 534165533cSJose E. Roman Input Parameters: 5495fce210SBarry Smith + sf - star forest 5595fce210SBarry Smith - unit - data type for each node 5695fce210SBarry Smith 574165533cSJose E. Roman Output Parameters: 5895fce210SBarry Smith + localtypes - types describing part of local leaf buffer referencing each remote rank 5995fce210SBarry Smith - remotetypes - types describing part of remote root buffer referenced for each remote rank 6095fce210SBarry Smith 6195fce210SBarry Smith Level: developer 6295fce210SBarry Smith 6395fce210SBarry Smith .seealso: PetscSFSetGraph(), PetscSFView() 6495fce210SBarry Smith @*/ 6595fce210SBarry Smith static PetscErrorCode PetscSFWindowGetDataTypes(PetscSF sf,MPI_Datatype unit,const MPI_Datatype **localtypes,const MPI_Datatype **remotetypes) 6695fce210SBarry Smith { 6795fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 6895fce210SBarry Smith PetscErrorCode ierr; 6995fce210SBarry Smith PetscSFDataLink link; 7095fce210SBarry Smith PetscInt i,nranks; 7195fce210SBarry Smith const PetscInt *roffset,*rmine,*rremote; 7295fce210SBarry Smith const PetscMPIInt *ranks; 7395fce210SBarry Smith 7495fce210SBarry Smith PetscFunctionBegin; 7595fce210SBarry Smith /* Look for types in cache */ 7695fce210SBarry Smith for (link=w->link; link; link=link->next) { 7795fce210SBarry Smith PetscBool match; 7895fce210SBarry Smith ierr = MPIPetsc_Type_compare(unit,link->unit,&match);CHKERRQ(ierr); 7995fce210SBarry Smith if (match) { 8095fce210SBarry Smith *localtypes = link->mine; 8195fce210SBarry Smith *remotetypes = link->remote; 8295fce210SBarry Smith PetscFunctionReturn(0); 8395fce210SBarry Smith } 8495fce210SBarry Smith } 8595fce210SBarry Smith 8695fce210SBarry Smith /* Create new composite types for each send rank */ 87dec1416fSJunchao Zhang ierr = PetscSFGetRootRanks(sf,&nranks,&ranks,&roffset,&rmine,&rremote);CHKERRQ(ierr); 8895dccacaSBarry Smith ierr = PetscNew(&link);CHKERRQ(ierr); 89ffc4695bSBarry Smith ierr = MPI_Type_dup(unit,&link->unit);CHKERRMPI(ierr); 90dcca6d9dSJed Brown ierr = PetscMalloc2(nranks,&link->mine,nranks,&link->remote);CHKERRQ(ierr); 9195fce210SBarry Smith for (i=0; i<nranks; i++) { 925b0d146aSStefano Zampini PetscInt rcount = roffset[i+1] - roffset[i]; 9395fce210SBarry Smith PetscMPIInt *rmine,*rremote; 9495fce210SBarry Smith #if !defined(PETSC_USE_64BIT_INDICES) 9595fce210SBarry Smith rmine = sf->rmine + sf->roffset[i]; 9695fce210SBarry Smith rremote = sf->rremote + sf->roffset[i]; 9795fce210SBarry Smith #else 9895fce210SBarry Smith PetscInt j; 99dcca6d9dSJed Brown ierr = PetscMalloc2(rcount,&rmine,rcount,&rremote);CHKERRQ(ierr); 10095fce210SBarry Smith for (j=0; j<rcount; j++) { 10195fce210SBarry Smith ierr = PetscMPIIntCast(sf->rmine[sf->roffset[i]+j],rmine+j);CHKERRQ(ierr); 10295fce210SBarry Smith ierr = PetscMPIIntCast(sf->rremote[sf->roffset[i]+j],rremote+j);CHKERRQ(ierr); 10395fce210SBarry Smith } 10495fce210SBarry Smith #endif 1055b0d146aSStefano Zampini 106ffc4695bSBarry Smith ierr = MPI_Type_create_indexed_block(rcount,1,rmine,link->unit,&link->mine[i]);CHKERRMPI(ierr); 107ffc4695bSBarry Smith ierr = MPI_Type_create_indexed_block(rcount,1,rremote,link->unit,&link->remote[i]);CHKERRMPI(ierr); 10895fce210SBarry Smith #if defined(PETSC_USE_64BIT_INDICES) 10995fce210SBarry Smith ierr = PetscFree2(rmine,rremote);CHKERRQ(ierr); 11095fce210SBarry Smith #endif 111ffc4695bSBarry Smith ierr = MPI_Type_commit(&link->mine[i]);CHKERRMPI(ierr); 112ffc4695bSBarry Smith ierr = MPI_Type_commit(&link->remote[i]);CHKERRMPI(ierr); 11395fce210SBarry Smith } 11495fce210SBarry Smith link->next = w->link; 11595fce210SBarry Smith w->link = link; 11695fce210SBarry Smith 11795fce210SBarry Smith *localtypes = link->mine; 11895fce210SBarry Smith *remotetypes = link->remote; 11995fce210SBarry Smith PetscFunctionReturn(0); 12095fce210SBarry Smith } 12195fce210SBarry Smith 12295fce210SBarry Smith /*@C 1235b0d146aSStefano Zampini PetscSFWindowSetFlavorType - Set flavor type for MPI_Win creation 1245b0d146aSStefano Zampini 1255b0d146aSStefano Zampini Logically Collective 1265b0d146aSStefano Zampini 1274165533cSJose E. Roman Input Parameters: 1285b0d146aSStefano Zampini + sf - star forest for communication 1295b0d146aSStefano Zampini - flavor - flavor type 1305b0d146aSStefano Zampini 1315b0d146aSStefano Zampini Options Database Key: 1325b0d146aSStefano Zampini . -sf_window_flavor <flavor> - sets the flavor type CREATE, DYNAMIC, ALLOCATE or SHARED (see PetscSFWindowFlavorType) 1335b0d146aSStefano Zampini 1345b0d146aSStefano Zampini Level: advanced 1355b0d146aSStefano Zampini 1365b0d146aSStefano Zampini Notes: Windows reusage follow this rules: 1375b0d146aSStefano Zampini 1385b0d146aSStefano Zampini PETSCSF_WINDOW_FLAVOR_CREATE: creates a new window every time, uses MPI_Win_create 1395b0d146aSStefano Zampini 1405b0d146aSStefano Zampini PETSCSF_WINDOW_FLAVOR_DYNAMIC: uses MPI_Win_create_dynamic/MPI_Win_attach and tries to reuse windows by comparing the root array. Intended to be used on repeated applications of the same SF, e.g. 1415b0d146aSStefano Zampini for i=1 to K 1425b0d146aSStefano Zampini PetscSFOperationBegin(rootdata1,leafdata_whatever); 1435b0d146aSStefano Zampini PetscSFOperationEnd(rootdata1,leafdata_whatever); 1445b0d146aSStefano Zampini ... 1455b0d146aSStefano Zampini PetscSFOperationBegin(rootdataN,leafdata_whatever); 1465b0d146aSStefano Zampini PetscSFOperationEnd(rootdataN,leafdata_whatever); 1475b0d146aSStefano Zampini endfor 1485b0d146aSStefano Zampini The following pattern will instead raise an error 1495b0d146aSStefano Zampini PetscSFOperationBegin(rootdata1,leafdata_whatever); 1505b0d146aSStefano Zampini PetscSFOperationEnd(rootdata1,leafdata_whatever); 1515b0d146aSStefano Zampini PetscSFOperationBegin(rank ? rootdata1 : rootdata2,leafdata_whatever); 1525b0d146aSStefano Zampini PetscSFOperationEnd(rank ? rootdata1 : rootdata2,leafdata_whatever); 1535b0d146aSStefano Zampini 1545b0d146aSStefano Zampini PETSCSF_WINDOW_FLAVOR_ALLOCATE: uses MPI_Win_allocate, reuses any pre-existing window which fits the data and it is not in use 1555b0d146aSStefano Zampini 1565b0d146aSStefano Zampini PETSCSF_WINDOW_FLAVOR_SHARED: uses MPI_Win_allocate_shared, reusage policy as for PETSCSF_WINDOW_FLAVOR_ALLOCATE 1575b0d146aSStefano Zampini 1585b0d146aSStefano Zampini .seealso: PetscSFSetFromOptions(), PetscSFWindowGetFlavorType() 1595b0d146aSStefano Zampini @*/ 1605b0d146aSStefano Zampini PetscErrorCode PetscSFWindowSetFlavorType(PetscSF sf,PetscSFWindowFlavorType flavor) 1615b0d146aSStefano Zampini { 1625b0d146aSStefano Zampini PetscErrorCode ierr; 1635b0d146aSStefano Zampini 1645b0d146aSStefano Zampini PetscFunctionBegin; 1655b0d146aSStefano Zampini PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1); 1665b0d146aSStefano Zampini PetscValidLogicalCollectiveEnum(sf,flavor,2); 1675b0d146aSStefano Zampini ierr = PetscTryMethod(sf,"PetscSFWindowSetFlavorType_C",(PetscSF,PetscSFWindowFlavorType),(sf,flavor));CHKERRQ(ierr); 1685b0d146aSStefano Zampini PetscFunctionReturn(0); 1695b0d146aSStefano Zampini } 1705b0d146aSStefano Zampini 1715b0d146aSStefano Zampini static PetscErrorCode PetscSFWindowSetFlavorType_Window(PetscSF sf,PetscSFWindowFlavorType flavor) 1725b0d146aSStefano Zampini { 1735b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 1745b0d146aSStefano Zampini 1755b0d146aSStefano Zampini PetscFunctionBegin; 1765b0d146aSStefano Zampini w->flavor = flavor; 1775b0d146aSStefano Zampini PetscFunctionReturn(0); 1785b0d146aSStefano Zampini } 1795b0d146aSStefano Zampini 1805b0d146aSStefano Zampini /*@C 1815b0d146aSStefano Zampini PetscSFWindowGetFlavorType - Get flavor type for PetscSF communication 1825b0d146aSStefano Zampini 1835b0d146aSStefano Zampini Logically Collective 1845b0d146aSStefano Zampini 1854165533cSJose E. Roman Input Parameter: 1865b0d146aSStefano Zampini . sf - star forest for communication 1875b0d146aSStefano Zampini 1884165533cSJose E. Roman Output Parameter: 1895b0d146aSStefano Zampini . flavor - flavor type 1905b0d146aSStefano Zampini 1915b0d146aSStefano Zampini Level: advanced 1925b0d146aSStefano Zampini 1935b0d146aSStefano Zampini .seealso: PetscSFSetFromOptions(), PetscSFWindowSetFlavorType() 1945b0d146aSStefano Zampini @*/ 1955b0d146aSStefano Zampini PetscErrorCode PetscSFWindowGetFlavorType(PetscSF sf,PetscSFWindowFlavorType *flavor) 1965b0d146aSStefano Zampini { 1975b0d146aSStefano Zampini PetscErrorCode ierr; 1985b0d146aSStefano Zampini 1995b0d146aSStefano Zampini PetscFunctionBegin; 2005b0d146aSStefano Zampini PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1); 2015b0d146aSStefano Zampini PetscValidPointer(flavor,2); 2025b0d146aSStefano Zampini ierr = PetscUseMethod(sf,"PetscSFWindowGetFlavorType_C",(PetscSF,PetscSFWindowFlavorType*),(sf,flavor));CHKERRQ(ierr); 2035b0d146aSStefano Zampini PetscFunctionReturn(0); 2045b0d146aSStefano Zampini } 2055b0d146aSStefano Zampini 2065b0d146aSStefano Zampini static PetscErrorCode PetscSFWindowGetFlavorType_Window(PetscSF sf,PetscSFWindowFlavorType *flavor) 2075b0d146aSStefano Zampini { 2085b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 2095b0d146aSStefano Zampini 2105b0d146aSStefano Zampini PetscFunctionBegin; 2115b0d146aSStefano Zampini *flavor = w->flavor; 2125b0d146aSStefano Zampini PetscFunctionReturn(0); 2135b0d146aSStefano Zampini } 2145b0d146aSStefano Zampini 2155b0d146aSStefano Zampini /*@C 2165b0d146aSStefano Zampini PetscSFWindowSetSyncType - Set synchronization type for PetscSF communication 21795fce210SBarry Smith 21895fce210SBarry Smith Logically Collective 21995fce210SBarry Smith 2204165533cSJose E. Roman Input Parameters: 22195fce210SBarry Smith + sf - star forest for communication 22295fce210SBarry Smith - sync - synchronization type 22395fce210SBarry Smith 22495fce210SBarry Smith Options Database Key: 22560263706SJed Brown . -sf_window_sync <sync> - sets the synchronization type FENCE, LOCK, or ACTIVE (see PetscSFWindowSyncType) 22695fce210SBarry Smith 22795fce210SBarry Smith Level: advanced 22895fce210SBarry Smith 22995fce210SBarry Smith .seealso: PetscSFSetFromOptions(), PetscSFWindowGetSyncType() 23095fce210SBarry Smith @*/ 23195fce210SBarry Smith PetscErrorCode PetscSFWindowSetSyncType(PetscSF sf,PetscSFWindowSyncType sync) 23295fce210SBarry Smith { 23395fce210SBarry Smith PetscErrorCode ierr; 23495fce210SBarry Smith 23595fce210SBarry Smith PetscFunctionBegin; 23695fce210SBarry Smith PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1); 23795fce210SBarry Smith PetscValidLogicalCollectiveEnum(sf,sync,2); 2385b0d146aSStefano Zampini ierr = PetscTryMethod(sf,"PetscSFWindowSetSyncType_C",(PetscSF,PetscSFWindowSyncType),(sf,sync));CHKERRQ(ierr); 23995fce210SBarry Smith PetscFunctionReturn(0); 24095fce210SBarry Smith } 24195fce210SBarry Smith 242f7a08781SBarry Smith static PetscErrorCode PetscSFWindowSetSyncType_Window(PetscSF sf,PetscSFWindowSyncType sync) 24395fce210SBarry Smith { 24495fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 24595fce210SBarry Smith 24695fce210SBarry Smith PetscFunctionBegin; 24795fce210SBarry Smith w->sync = sync; 24895fce210SBarry Smith PetscFunctionReturn(0); 24995fce210SBarry Smith } 25095fce210SBarry Smith 25195fce210SBarry Smith /*@C 2525b0d146aSStefano Zampini PetscSFWindowGetSyncType - Get synchronization type for PetscSF communication 25395fce210SBarry Smith 25495fce210SBarry Smith Logically Collective 25595fce210SBarry Smith 2564165533cSJose E. Roman Input Parameter: 25795fce210SBarry Smith . sf - star forest for communication 25895fce210SBarry Smith 2594165533cSJose E. Roman Output Parameter: 26095fce210SBarry Smith . sync - synchronization type 26195fce210SBarry Smith 26295fce210SBarry Smith Level: advanced 26395fce210SBarry Smith 2645b0d146aSStefano Zampini .seealso: PetscSFSetFromOptions(), PetscSFWindowSetSyncType() 26595fce210SBarry Smith @*/ 26695fce210SBarry Smith PetscErrorCode PetscSFWindowGetSyncType(PetscSF sf,PetscSFWindowSyncType *sync) 26795fce210SBarry Smith { 26895fce210SBarry Smith PetscErrorCode ierr; 26995fce210SBarry Smith 27095fce210SBarry Smith PetscFunctionBegin; 27195fce210SBarry Smith PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1); 27295fce210SBarry Smith PetscValidPointer(sync,2); 273163d334eSBarry Smith ierr = PetscUseMethod(sf,"PetscSFWindowGetSyncType_C",(PetscSF,PetscSFWindowSyncType*),(sf,sync));CHKERRQ(ierr); 27495fce210SBarry Smith PetscFunctionReturn(0); 27595fce210SBarry Smith } 27695fce210SBarry Smith 277f7a08781SBarry Smith static PetscErrorCode PetscSFWindowGetSyncType_Window(PetscSF sf,PetscSFWindowSyncType *sync) 27895fce210SBarry Smith { 27995fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 28095fce210SBarry Smith 28195fce210SBarry Smith PetscFunctionBegin; 28295fce210SBarry Smith *sync = w->sync; 28395fce210SBarry Smith PetscFunctionReturn(0); 28495fce210SBarry Smith } 28595fce210SBarry Smith 28695fce210SBarry Smith /*@C 2875b0d146aSStefano Zampini PetscSFWindowSetInfo - Set the MPI_Info handle that will be used for subsequent windows allocation 2885b0d146aSStefano Zampini 2895b0d146aSStefano Zampini Logically Collective 2905b0d146aSStefano Zampini 2914165533cSJose E. Roman Input Parameters: 2925b0d146aSStefano Zampini + sf - star forest for communication 2935b0d146aSStefano Zampini - info - MPI_Info handle 2945b0d146aSStefano Zampini 2955b0d146aSStefano Zampini Level: advanced 2965b0d146aSStefano Zampini 2975b0d146aSStefano Zampini Notes: the info handle is duplicated with a call to MPI_Info_dup unless info = MPI_INFO_NULL. 2985b0d146aSStefano Zampini 2995b0d146aSStefano Zampini .seealso: PetscSFSetFromOptions(), PetscSFWindowGetInfo() 3005b0d146aSStefano Zampini @*/ 3015b0d146aSStefano Zampini PetscErrorCode PetscSFWindowSetInfo(PetscSF sf,MPI_Info info) 3025b0d146aSStefano Zampini { 3035b0d146aSStefano Zampini PetscErrorCode ierr; 3045b0d146aSStefano Zampini 3055b0d146aSStefano Zampini PetscFunctionBegin; 3065b0d146aSStefano Zampini PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1); 3075b0d146aSStefano Zampini ierr = PetscTryMethod(sf,"PetscSFWindowSetInfo_C",(PetscSF,MPI_Info),(sf,info));CHKERRQ(ierr); 3085b0d146aSStefano Zampini PetscFunctionReturn(0); 3095b0d146aSStefano Zampini } 3105b0d146aSStefano Zampini 3115b0d146aSStefano Zampini static PetscErrorCode PetscSFWindowSetInfo_Window(PetscSF sf,MPI_Info info) 3125b0d146aSStefano Zampini { 3135b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 3145b0d146aSStefano Zampini PetscErrorCode ierr; 3155b0d146aSStefano Zampini 3165b0d146aSStefano Zampini PetscFunctionBegin; 3175b0d146aSStefano Zampini if (w->info != MPI_INFO_NULL) { 318ffc4695bSBarry Smith ierr = MPI_Info_free(&w->info);CHKERRMPI(ierr); 3195b0d146aSStefano Zampini } 3205b0d146aSStefano Zampini if (info != MPI_INFO_NULL) { 321ffc4695bSBarry Smith ierr = MPI_Info_dup(info,&w->info);CHKERRMPI(ierr); 3225b0d146aSStefano Zampini } 3235b0d146aSStefano Zampini PetscFunctionReturn(0); 3245b0d146aSStefano Zampini } 3255b0d146aSStefano Zampini 3265b0d146aSStefano Zampini /*@C 3275b0d146aSStefano Zampini PetscSFWindowGetInfo - Get the MPI_Info handle used for windows allocation 3285b0d146aSStefano Zampini 3295b0d146aSStefano Zampini Logically Collective 3305b0d146aSStefano Zampini 3314165533cSJose E. Roman Input Parameter: 3325b0d146aSStefano Zampini . sf - star forest for communication 3335b0d146aSStefano Zampini 3344165533cSJose E. Roman Output Parameter: 3355b0d146aSStefano Zampini . info - MPI_Info handle 3365b0d146aSStefano Zampini 3375b0d146aSStefano Zampini Level: advanced 3385b0d146aSStefano Zampini 3395b0d146aSStefano Zampini Notes: if PetscSFWindowSetInfo() has not be called, this returns MPI_INFO_NULL 3405b0d146aSStefano Zampini 3415b0d146aSStefano Zampini .seealso: PetscSFSetFromOptions(), PetscSFWindowSetInfo() 3425b0d146aSStefano Zampini @*/ 3435b0d146aSStefano Zampini PetscErrorCode PetscSFWindowGetInfo(PetscSF sf,MPI_Info *info) 3445b0d146aSStefano Zampini { 3455b0d146aSStefano Zampini PetscErrorCode ierr; 3465b0d146aSStefano Zampini 3475b0d146aSStefano Zampini PetscFunctionBegin; 3485b0d146aSStefano Zampini PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1); 3495b0d146aSStefano Zampini PetscValidPointer(info,2); 3505b0d146aSStefano Zampini ierr = PetscUseMethod(sf,"PetscSFWindowGetInfo_C",(PetscSF,MPI_Info*),(sf,info));CHKERRQ(ierr); 3515b0d146aSStefano Zampini PetscFunctionReturn(0); 3525b0d146aSStefano Zampini } 3535b0d146aSStefano Zampini 3545b0d146aSStefano Zampini static PetscErrorCode PetscSFWindowGetInfo_Window(PetscSF sf,MPI_Info *info) 3555b0d146aSStefano Zampini { 3565b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 3575b0d146aSStefano Zampini 3585b0d146aSStefano Zampini PetscFunctionBegin; 3595b0d146aSStefano Zampini *info = w->info; 3605b0d146aSStefano Zampini PetscFunctionReturn(0); 3615b0d146aSStefano Zampini } 3625b0d146aSStefano Zampini 3635b0d146aSStefano Zampini /* 36495fce210SBarry Smith PetscSFGetWindow - Get a window for use with a given data type 36595fce210SBarry Smith 36695fce210SBarry Smith Collective on PetscSF 36795fce210SBarry Smith 3684165533cSJose E. Roman Input Parameters: 36995fce210SBarry Smith + sf - star forest 37095fce210SBarry Smith . unit - data type 37195fce210SBarry Smith . array - array to be sent 3725b0d146aSStefano Zampini . sync - type of synchronization PetscSFWindowSyncType 37395fce210SBarry Smith . epoch - PETSC_TRUE to acquire the window and start an epoch, PETSC_FALSE to just acquire the window 3745b0d146aSStefano Zampini . fenceassert - assert parameter for call to MPI_Win_fence(), if sync == PETSCSF_WINDOW_SYNC_FENCE 3755b0d146aSStefano Zampini . postassert - assert parameter for call to MPI_Win_post(), if sync == PETSCSF_WINDOW_SYNC_ACTIVE 376684a874aSStefano Zampini - startassert - assert parameter for call to MPI_Win_start(), if sync == PETSCSF_WINDOW_SYNC_ACTIVE 37795fce210SBarry Smith 3784165533cSJose E. Roman Output Parameters: 379684a874aSStefano Zampini + target_disp - target_disp argument for RMA calls (significative for PETSCSF_WINDOW_FLAVOR_DYNAMIC only) 380684a874aSStefano Zampini + reqs - array of requests (significative for sync == PETSCSF_WINDOW_SYNC_LOCK only) 381684a874aSStefano Zampini - win - window 38295fce210SBarry Smith 38395fce210SBarry Smith Level: developer 384dec1416fSJunchao Zhang .seealso: PetscSFGetRootRanks(), PetscSFWindowGetDataTypes() 3855b0d146aSStefano Zampini */ 386684a874aSStefano Zampini static PetscErrorCode PetscSFGetWindow(PetscSF sf,MPI_Datatype unit,void *array,PetscSFWindowSyncType sync,PetscBool epoch,PetscMPIInt fenceassert,PetscMPIInt postassert,PetscMPIInt startassert,const MPI_Aint **target_disp, MPI_Request **reqs, MPI_Win *win) 38795fce210SBarry Smith { 38895fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 38995fce210SBarry Smith PetscErrorCode ierr; 39095fce210SBarry Smith MPI_Aint lb,lb_true,bytes,bytes_true; 39195fce210SBarry Smith PetscSFWinLink link; 392d547623eSJunchao Zhang #if defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW) 3935b0d146aSStefano Zampini MPI_Aint winaddr; 3945b0d146aSStefano Zampini PetscInt nranks; 395d547623eSJunchao Zhang #endif 3965b0d146aSStefano Zampini PetscBool reuse = PETSC_FALSE, update = PETSC_FALSE; 3975b0d146aSStefano Zampini PetscBool dummy[2]; 3985b0d146aSStefano Zampini MPI_Aint wsize; 39995fce210SBarry Smith 40095fce210SBarry Smith PetscFunctionBegin; 401ffc4695bSBarry Smith ierr = MPI_Type_get_extent(unit,&lb,&bytes);CHKERRMPI(ierr); 402ffc4695bSBarry Smith ierr = MPI_Type_get_true_extent(unit,&lb_true,&bytes_true);CHKERRMPI(ierr); 403*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(lb != 0 || lb_true != 0,PetscObjectComm((PetscObject)sf),PETSC_ERR_SUP,"No support for unit type with nonzero lower bound, write petsc-maint@mcs.anl.gov if you want this feature"); 404*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(bytes != bytes_true,PetscObjectComm((PetscObject)sf),PETSC_ERR_SUP,"No support for unit type with modified extent, write petsc-maint@mcs.anl.gov if you want this feature"); 4055b0d146aSStefano Zampini if (w->flavor != PETSCSF_WINDOW_FLAVOR_CREATE) reuse = PETSC_TRUE; 4065b0d146aSStefano Zampini for (link=w->wins; reuse && link; link=link->next) { 4075b0d146aSStefano Zampini PetscBool winok = PETSC_FALSE; 4085b0d146aSStefano Zampini if (w->flavor != link->flavor) continue; 4095b0d146aSStefano Zampini switch (w->flavor) { 4105b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_DYNAMIC: /* check available matching array, error if in use (we additionally check that the matching condition is the same across processes) */ 4115b0d146aSStefano Zampini if (array == link->addr) { 41276bd3646SJed Brown if (PetscDefined(USE_DEBUG)) { 4135b0d146aSStefano Zampini dummy[0] = PETSC_TRUE; 4145b0d146aSStefano Zampini dummy[1] = PETSC_TRUE; 415ffc4695bSBarry Smith ierr = MPI_Allreduce(MPI_IN_PLACE,dummy,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)sf));CHKERRMPI(ierr); 416ffc4695bSBarry Smith ierr = MPI_Allreduce(MPI_IN_PLACE,dummy+1,1,MPIU_BOOL,MPI_LOR ,PetscObjectComm((PetscObject)sf));CHKERRMPI(ierr); 417*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(dummy[0] != dummy[1],PetscObjectComm((PetscObject)sf),PETSC_ERR_SUP,"PETSCSF_WINDOW_FLAVOR_DYNAMIC requires root pointers to be consistently used across the comm. Use PETSCSF_WINDOW_FLAVOR_CREATE or PETSCSF_WINDOW_FLAVOR_ALLOCATE instead"); 41876bd3646SJed Brown } 419*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(link->inuse,PetscObjectComm((PetscObject)sf),PETSC_ERR_PLIB,"Window in use"); 420*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(epoch && link->epoch,PetscObjectComm((PetscObject)sf),PETSC_ERR_PLIB,"Window epoch not finished"); 4215b0d146aSStefano Zampini winok = PETSC_TRUE; 4225b0d146aSStefano Zampini link->paddr = array; 42376bd3646SJed Brown } else if (PetscDefined(USE_DEBUG)) { 4245b0d146aSStefano Zampini dummy[0] = PETSC_FALSE; 4255b0d146aSStefano Zampini dummy[1] = PETSC_FALSE; 426ffc4695bSBarry Smith ierr = MPI_Allreduce(MPI_IN_PLACE,dummy ,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)sf));CHKERRMPI(ierr); 427ffc4695bSBarry Smith ierr = MPI_Allreduce(MPI_IN_PLACE,dummy+1,1,MPIU_BOOL,MPI_LOR ,PetscObjectComm((PetscObject)sf));CHKERRMPI(ierr); 428*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(dummy[0] != dummy[1],PetscObjectComm((PetscObject)sf),PETSC_ERR_SUP,"PETSCSF_WINDOW_FLAVOR_DYNAMIC requires root pointers to be consistently used across the comm. Use PETSCSF_WINDOW_FLAVOR_CREATE or PETSCSF_WINDOW_FLAVOR_ALLOCATE instead"); 4295b0d146aSStefano Zampini } 4305b0d146aSStefano Zampini break; 4315b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_ALLOCATE: /* check available by matching size, allocate if in use */ 4325b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_SHARED: 4335b0d146aSStefano Zampini if (!link->inuse && bytes == (MPI_Aint)link->bytes) { 4345b0d146aSStefano Zampini update = PETSC_TRUE; 4355b0d146aSStefano Zampini link->paddr = array; 4365b0d146aSStefano Zampini winok = PETSC_TRUE; 4375b0d146aSStefano Zampini } 4385b0d146aSStefano Zampini break; 43998921bdaSJacob Faibussowitsch default: SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_SUP,"No support for flavor %s",PetscSFWindowFlavorTypes[w->flavor]); 4405b0d146aSStefano Zampini } 4415b0d146aSStefano Zampini if (winok) { 4425b0d146aSStefano Zampini *win = link->win; 4437d3de750SJacob Faibussowitsch ierr = PetscInfo(sf,"Reusing window %" PETSC_MPI_WIN_FMT " of flavor %d for comm %" PETSC_MPI_COMM_FMT "\n",link->win,link->flavor,PetscObjectComm((PetscObject)sf));CHKERRQ(ierr); 4445b0d146aSStefano Zampini goto found; 4455b0d146aSStefano Zampini } 4465b0d146aSStefano Zampini } 4475b0d146aSStefano Zampini 4485b0d146aSStefano Zampini wsize = (MPI_Aint)bytes*sf->nroots; 44995dccacaSBarry Smith ierr = PetscNew(&link);CHKERRQ(ierr); 45095fce210SBarry Smith link->bytes = bytes; 45195fce210SBarry Smith link->next = w->wins; 4525b0d146aSStefano Zampini link->flavor = w->flavor; 4535b0d146aSStefano Zampini link->dyn_target_addr = NULL; 454684a874aSStefano Zampini link->reqs = NULL; 45595fce210SBarry Smith w->wins = link; 456684a874aSStefano Zampini if (sync == PETSCSF_WINDOW_SYNC_LOCK) { 457684a874aSStefano Zampini PetscInt i; 458684a874aSStefano Zampini 459684a874aSStefano Zampini ierr = PetscMalloc1(sf->nranks,&link->reqs);CHKERRQ(ierr); 460684a874aSStefano Zampini for (i = 0; i < sf->nranks; i++) link->reqs[i] = MPI_REQUEST_NULL; 461684a874aSStefano Zampini } 4625b0d146aSStefano Zampini switch (w->flavor) { 4635b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_CREATE: 464ffc4695bSBarry Smith ierr = MPI_Win_create(array,wsize,(PetscMPIInt)bytes,w->info,PetscObjectComm((PetscObject)sf),&link->win);CHKERRMPI(ierr); 4655b0d146aSStefano Zampini link->addr = array; 4665b0d146aSStefano Zampini link->paddr = array; 4675b0d146aSStefano Zampini break; 468d547623eSJunchao Zhang #if defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW) 4695b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_DYNAMIC: 470ffc4695bSBarry Smith ierr = MPI_Win_create_dynamic(w->info,PetscObjectComm((PetscObject)sf),&link->win);CHKERRMPI(ierr); 4715b0d146aSStefano Zampini #if defined(PETSC_HAVE_OMPI_MAJOR_VERSION) /* some OpenMPI versions do not support MPI_Win_attach(win,NULL,0); */ 472ffc4695bSBarry Smith ierr = MPI_Win_attach(link->win,wsize ? array : &ierr,wsize);CHKERRMPI(ierr); 4735b0d146aSStefano Zampini #else 474ffc4695bSBarry Smith ierr = MPI_Win_attach(link->win,array,wsize);CHKERRMPI(ierr); 4755b0d146aSStefano Zampini #endif 4765b0d146aSStefano Zampini link->addr = array; 4775b0d146aSStefano Zampini link->paddr = array; 478*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(!w->dynsf,PetscObjectComm((PetscObject)sf),PETSC_ERR_ORDER,"Must call PetscSFSetUp()"); 4795b0d146aSStefano Zampini ierr = PetscSFSetUp(w->dynsf);CHKERRQ(ierr); 4805b0d146aSStefano Zampini ierr = PetscSFGetRootRanks(w->dynsf,&nranks,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 4815b0d146aSStefano Zampini ierr = PetscMalloc1(nranks,&link->dyn_target_addr);CHKERRQ(ierr); 482ffc4695bSBarry Smith ierr = MPI_Get_address(array,&winaddr);CHKERRMPI(ierr); 483ad227feaSJunchao Zhang ierr = PetscSFBcastBegin(w->dynsf,MPI_AINT,&winaddr,link->dyn_target_addr,MPI_REPLACE);CHKERRQ(ierr); 484ad227feaSJunchao Zhang ierr = PetscSFBcastEnd(w->dynsf,MPI_AINT,&winaddr,link->dyn_target_addr,MPI_REPLACE);CHKERRQ(ierr); 4855b0d146aSStefano Zampini break; 4865b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_ALLOCATE: 487ffc4695bSBarry Smith ierr = MPI_Win_allocate(wsize,(PetscMPIInt)bytes,w->info,PetscObjectComm((PetscObject)sf),&link->addr,&link->win);CHKERRMPI(ierr); 4885b0d146aSStefano Zampini update = PETSC_TRUE; 4895b0d146aSStefano Zampini link->paddr = array; 4905b0d146aSStefano Zampini break; 491d547623eSJunchao Zhang #endif 4925b0d146aSStefano Zampini #if defined(PETSC_HAVE_MPI_PROCESS_SHARED_MEMORY) 4935b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_SHARED: 494ffc4695bSBarry Smith ierr = MPI_Win_allocate_shared(wsize,(PetscMPIInt)bytes,w->info,PetscObjectComm((PetscObject)sf),&link->addr,&link->win);CHKERRMPI(ierr); 4955b0d146aSStefano Zampini update = PETSC_TRUE; 4965b0d146aSStefano Zampini link->paddr = array; 4975b0d146aSStefano Zampini break; 4985b0d146aSStefano Zampini #endif 49998921bdaSJacob Faibussowitsch default: SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_SUP,"No support for flavor %s",PetscSFWindowFlavorTypes[w->flavor]); 5005b0d146aSStefano Zampini } 5017d3de750SJacob Faibussowitsch ierr = PetscInfo(sf,"New window %" PETSC_MPI_WIN_FMT " of flavor %d for comm %" PETSC_MPI_COMM_FMT "\n",link->win,link->flavor,PetscObjectComm((PetscObject)sf));CHKERRQ(ierr); 50295fce210SBarry Smith *win = link->win; 50395fce210SBarry Smith 5045b0d146aSStefano Zampini found: 5055b0d146aSStefano Zampini 506684a874aSStefano Zampini if (target_disp) *target_disp = link->dyn_target_addr; 507684a874aSStefano Zampini if (reqs) *reqs = link->reqs; 508684a874aSStefano Zampini if (update) { /* locks are needed for the "separate" memory model only, the fence guaranties memory-synchronization */ 509684a874aSStefano Zampini PetscMPIInt rank; 510684a874aSStefano Zampini 511ffc4695bSBarry Smith ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);CHKERRMPI(ierr); 512ffc4695bSBarry Smith if (sync == PETSCSF_WINDOW_SYNC_LOCK) {ierr = MPI_Win_lock(MPI_LOCK_EXCLUSIVE,rank,MPI_MODE_NOCHECK,*win);CHKERRMPI(ierr);} 5135b0d146aSStefano Zampini ierr = PetscMemcpy(link->addr,array,sf->nroots*bytes);CHKERRQ(ierr); 5145b0d146aSStefano Zampini if (sync == PETSCSF_WINDOW_SYNC_LOCK) { 515ffc4695bSBarry Smith ierr = MPI_Win_unlock(rank,*win);CHKERRMPI(ierr); 516ffc4695bSBarry Smith ierr = MPI_Win_fence(0,*win);CHKERRMPI(ierr); 5175b0d146aSStefano Zampini } 5185b0d146aSStefano Zampini } 5195b0d146aSStefano Zampini link->inuse = PETSC_TRUE; 5205b0d146aSStefano Zampini link->epoch = epoch; 52195fce210SBarry Smith if (epoch) { 5225b0d146aSStefano Zampini switch (sync) { 52395fce210SBarry Smith case PETSCSF_WINDOW_SYNC_FENCE: 524ffc4695bSBarry Smith ierr = MPI_Win_fence(fenceassert,*win);CHKERRMPI(ierr); 52595fce210SBarry Smith break; 52695fce210SBarry Smith case PETSCSF_WINDOW_SYNC_LOCK: /* Handled outside */ 52795fce210SBarry Smith break; 52895fce210SBarry Smith case PETSCSF_WINDOW_SYNC_ACTIVE: { 52995fce210SBarry Smith MPI_Group ingroup,outgroup; 5305b0d146aSStefano Zampini PetscMPIInt isize,osize; 5315b0d146aSStefano Zampini 5325b0d146aSStefano Zampini /* OpenMPI 4.0.2 with btl=vader does not like calling 5335b0d146aSStefano Zampini - MPI_Win_complete when ogroup is empty 5345b0d146aSStefano Zampini - MPI_Win_wait when igroup is empty 5355b0d146aSStefano Zampini So, we do not even issue the corresponding start and post calls 5365b0d146aSStefano Zampini The MPI standard (Sec. 11.5.2 of MPI 3.1) only requires that 5375b0d146aSStefano Zampini start(outgroup) has a matching post(ingroup) 5385b0d146aSStefano Zampini and this is guaranteed by PetscSF 5395b0d146aSStefano Zampini */ 54095fce210SBarry Smith ierr = PetscSFGetGroups(sf,&ingroup,&outgroup);CHKERRQ(ierr); 541ffc4695bSBarry Smith ierr = MPI_Group_size(ingroup,&isize);CHKERRMPI(ierr); 542ffc4695bSBarry Smith ierr = MPI_Group_size(outgroup,&osize);CHKERRMPI(ierr); 543ffc4695bSBarry Smith if (isize) {ierr = MPI_Win_post(ingroup,postassert,*win);CHKERRMPI(ierr);} 544ffc4695bSBarry Smith if (osize) {ierr = MPI_Win_start(outgroup,startassert,*win);CHKERRMPI(ierr);} 54595fce210SBarry Smith } break; 54695fce210SBarry Smith default: SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_PLIB,"Unknown synchronization type"); 54795fce210SBarry Smith } 54895fce210SBarry Smith } 54995fce210SBarry Smith PetscFunctionReturn(0); 55095fce210SBarry Smith } 55195fce210SBarry Smith 5525b0d146aSStefano Zampini /* 55395fce210SBarry Smith PetscSFFindWindow - Finds a window that is already in use 55495fce210SBarry Smith 55595fce210SBarry Smith Not Collective 55695fce210SBarry Smith 5574165533cSJose E. Roman Input Parameters: 55895fce210SBarry Smith + sf - star forest 55995fce210SBarry Smith . unit - data type 56095fce210SBarry Smith - array - array with which the window is associated 56195fce210SBarry Smith 5624165533cSJose E. Roman Output Parameters: 563684a874aSStefano Zampini + win - window 564684a874aSStefano Zampini - reqs - outstanding requests associated to the window 56595fce210SBarry Smith 56695fce210SBarry Smith Level: developer 56795fce210SBarry Smith 56895fce210SBarry Smith .seealso: PetscSFGetWindow(), PetscSFRestoreWindow() 5695b0d146aSStefano Zampini */ 570684a874aSStefano Zampini static PetscErrorCode PetscSFFindWindow(PetscSF sf,MPI_Datatype unit,const void *array,MPI_Win *win,MPI_Request **reqs) 57195fce210SBarry Smith { 57295fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 57395fce210SBarry Smith PetscSFWinLink link; 5745b0d146aSStefano Zampini PetscErrorCode ierr; 57595fce210SBarry Smith 57695fce210SBarry Smith PetscFunctionBegin; 577c0cd0301SJed Brown *win = MPI_WIN_NULL; 57895fce210SBarry Smith for (link=w->wins; link; link=link->next) { 5795b0d146aSStefano Zampini if (array == link->paddr) { 5802abc8c78SJacob Faibussowitsch 5817d3de750SJacob Faibussowitsch ierr = PetscInfo(sf,"Window %" PETSC_MPI_WIN_FMT " of flavor %d for comm %" PETSC_MPI_COMM_FMT "\n",link->win,link->flavor,PetscObjectComm((PetscObject)sf));CHKERRQ(ierr); 58295fce210SBarry Smith *win = link->win; 583684a874aSStefano Zampini *reqs = link->reqs; 58495fce210SBarry Smith PetscFunctionReturn(0); 58595fce210SBarry Smith } 58695fce210SBarry Smith } 58795fce210SBarry Smith SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Requested window not in use"); 58895fce210SBarry Smith } 58995fce210SBarry Smith 5905b0d146aSStefano Zampini /* 59195fce210SBarry Smith PetscSFRestoreWindow - Restores a window obtained with PetscSFGetWindow() 59295fce210SBarry Smith 59395fce210SBarry Smith Collective 59495fce210SBarry Smith 5954165533cSJose E. Roman Input Parameters: 59695fce210SBarry Smith + sf - star forest 59795fce210SBarry Smith . unit - data type 59895fce210SBarry Smith . array - array associated with window 5995b0d146aSStefano Zampini . sync - type of synchronization PetscSFWindowSyncType 60095fce210SBarry Smith . epoch - close an epoch, must match argument to PetscSFGetWindow() 6015b0d146aSStefano Zampini . update - if we have to update the local window array 60295fce210SBarry Smith - win - window 60395fce210SBarry Smith 60495fce210SBarry Smith Level: developer 60595fce210SBarry Smith 60695fce210SBarry Smith .seealso: PetscSFFindWindow() 6075b0d146aSStefano Zampini */ 6085b0d146aSStefano Zampini static PetscErrorCode PetscSFRestoreWindow(PetscSF sf,MPI_Datatype unit,void *array,PetscSFWindowSyncType sync,PetscBool epoch,PetscMPIInt fenceassert,PetscBool update,MPI_Win *win) 60995fce210SBarry Smith { 61095fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 61195fce210SBarry Smith PetscErrorCode ierr; 61295fce210SBarry Smith PetscSFWinLink *p,link; 6135b0d146aSStefano Zampini PetscBool reuse = PETSC_FALSE; 6145b0d146aSStefano Zampini PetscSFWindowFlavorType flavor; 6155b0d146aSStefano Zampini void* laddr; 6165b0d146aSStefano Zampini size_t bytes; 61795fce210SBarry Smith 61895fce210SBarry Smith PetscFunctionBegin; 61995fce210SBarry Smith for (p=&w->wins; *p; p=&(*p)->next) { 62095fce210SBarry Smith link = *p; 62195fce210SBarry Smith if (*win == link->win) { 622*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(array != link->paddr,PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Matched window, but not array"); 62395fce210SBarry Smith if (epoch != link->epoch) { 624*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(epoch,PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"No epoch to end"); 62595fce210SBarry Smith else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Restoring window without ending epoch"); 62695fce210SBarry Smith } 6275b0d146aSStefano Zampini laddr = link->addr; 6285b0d146aSStefano Zampini flavor = link->flavor; 6295b0d146aSStefano Zampini bytes = link->bytes; 6305b0d146aSStefano Zampini if (flavor != PETSCSF_WINDOW_FLAVOR_CREATE) reuse = PETSC_TRUE; 6315b0d146aSStefano Zampini else { *p = link->next; update = PETSC_FALSE; } /* remove from list */ 63295fce210SBarry Smith goto found; 63395fce210SBarry Smith } 63495fce210SBarry Smith } 63595fce210SBarry Smith SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Requested window not in use"); 63695fce210SBarry Smith 63795fce210SBarry Smith found: 6387d3de750SJacob Faibussowitsch ierr = PetscInfo(sf,"Window %" PETSC_MPI_WIN_FMT " of flavor %d for comm %" PETSC_MPI_COMM_FMT "\n",link->win,link->flavor,PetscObjectComm((PetscObject)sf));CHKERRQ(ierr); 63995fce210SBarry Smith if (epoch) { 6405b0d146aSStefano Zampini switch (sync) { 64195fce210SBarry Smith case PETSCSF_WINDOW_SYNC_FENCE: 642ffc4695bSBarry Smith ierr = MPI_Win_fence(fenceassert,*win);CHKERRMPI(ierr); 64395fce210SBarry Smith break; 6445b0d146aSStefano Zampini case PETSCSF_WINDOW_SYNC_LOCK: /* Handled outside */ 6455b0d146aSStefano Zampini break; 64695fce210SBarry Smith case PETSCSF_WINDOW_SYNC_ACTIVE: { 6475b0d146aSStefano Zampini MPI_Group ingroup,outgroup; 6485b0d146aSStefano Zampini PetscMPIInt isize,osize; 6495b0d146aSStefano Zampini 6505b0d146aSStefano Zampini /* OpenMPI 4.0.2 with btl=wader does not like calling 6515b0d146aSStefano Zampini - MPI_Win_complete when ogroup is empty 6525b0d146aSStefano Zampini - MPI_Win_wait when igroup is empty 6535b0d146aSStefano Zampini The MPI standard (Sec. 11.5.2 of MPI 3.1) only requires that 6545b0d146aSStefano Zampini - each process who issues a call to MPI_Win_start issues a call to MPI_Win_Complete 6555b0d146aSStefano Zampini - each process who issues a call to MPI_Win_post issues a call to MPI_Win_Wait 6565b0d146aSStefano Zampini */ 6575b0d146aSStefano Zampini ierr = PetscSFGetGroups(sf,&ingroup,&outgroup);CHKERRQ(ierr); 658ffc4695bSBarry Smith ierr = MPI_Group_size(ingroup,&isize);CHKERRMPI(ierr); 659ffc4695bSBarry Smith ierr = MPI_Group_size(outgroup,&osize);CHKERRMPI(ierr); 660ffc4695bSBarry Smith if (osize) {ierr = MPI_Win_complete(*win);CHKERRMPI(ierr);} 661ffc4695bSBarry Smith if (isize) {ierr = MPI_Win_wait(*win);CHKERRMPI(ierr);} 66295fce210SBarry Smith } break; 66395fce210SBarry Smith default: SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_PLIB,"Unknown synchronization type"); 66495fce210SBarry Smith } 66595fce210SBarry Smith } 6665b0d146aSStefano Zampini if (update) { 6675b0d146aSStefano Zampini if (sync == PETSCSF_WINDOW_SYNC_LOCK) { 668ffc4695bSBarry Smith ierr = MPI_Win_fence(MPI_MODE_NOPUT|MPI_MODE_NOSUCCEED,*win);CHKERRMPI(ierr); 6695b0d146aSStefano Zampini } 6705b0d146aSStefano Zampini ierr = PetscMemcpy(array,laddr,sf->nroots*bytes);CHKERRQ(ierr); 6715b0d146aSStefano Zampini } 6725b0d146aSStefano Zampini link->epoch = PETSC_FALSE; 6735b0d146aSStefano Zampini link->inuse = PETSC_FALSE; 6745b0d146aSStefano Zampini link->paddr = NULL; 6755b0d146aSStefano Zampini if (!reuse) { 676684a874aSStefano Zampini ierr = PetscFree(link->dyn_target_addr);CHKERRQ(ierr); 677684a874aSStefano Zampini ierr = PetscFree(link->reqs);CHKERRQ(ierr); 678ffc4695bSBarry Smith ierr = MPI_Win_free(&link->win);CHKERRMPI(ierr); 67995fce210SBarry Smith ierr = PetscFree(link);CHKERRQ(ierr); 68095fce210SBarry Smith *win = MPI_WIN_NULL; 6815b0d146aSStefano Zampini } 68295fce210SBarry Smith PetscFunctionReturn(0); 68395fce210SBarry Smith } 68495fce210SBarry Smith 68595fce210SBarry Smith static PetscErrorCode PetscSFSetUp_Window(PetscSF sf) 68695fce210SBarry Smith { 68795fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 68895fce210SBarry Smith PetscErrorCode ierr; 68995fce210SBarry Smith MPI_Group ingroup,outgroup; 69095fce210SBarry Smith 69195fce210SBarry Smith PetscFunctionBegin; 692b5a8e515SJed Brown ierr = PetscSFSetUpRanks(sf,MPI_GROUP_EMPTY);CHKERRQ(ierr); 6935b0d146aSStefano Zampini if (!w->dynsf) { 6945b0d146aSStefano Zampini PetscInt i; 6955b0d146aSStefano Zampini PetscSFNode *remotes; 6965b0d146aSStefano Zampini 6975b0d146aSStefano Zampini ierr = PetscMalloc1(sf->nranks,&remotes);CHKERRQ(ierr); 6985b0d146aSStefano Zampini for (i=0;i<sf->nranks;i++) { 6995b0d146aSStefano Zampini remotes[i].rank = sf->ranks[i]; 7005b0d146aSStefano Zampini remotes[i].index = 0; 7015b0d146aSStefano Zampini } 7025b0d146aSStefano Zampini ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_RANKS,&w->dynsf);CHKERRQ(ierr); 7035b0d146aSStefano Zampini ierr = PetscSFWindowSetFlavorType(w->dynsf,PETSCSF_WINDOW_FLAVOR_CREATE);CHKERRQ(ierr); /* break recursion */ 7045b0d146aSStefano Zampini ierr = PetscSFSetGraph(w->dynsf,1,sf->nranks,NULL,PETSC_OWN_POINTER,remotes,PETSC_OWN_POINTER);CHKERRQ(ierr); 7055b0d146aSStefano Zampini ierr = PetscLogObjectParent((PetscObject)sf,(PetscObject)w->dynsf);CHKERRQ(ierr); 7065b0d146aSStefano Zampini } 70795fce210SBarry Smith switch (w->sync) { 70895fce210SBarry Smith case PETSCSF_WINDOW_SYNC_ACTIVE: 70995fce210SBarry Smith ierr = PetscSFGetGroups(sf,&ingroup,&outgroup);CHKERRQ(ierr); 71095fce210SBarry Smith default: 71195fce210SBarry Smith break; 71295fce210SBarry Smith } 71395fce210SBarry Smith PetscFunctionReturn(0); 71495fce210SBarry Smith } 71595fce210SBarry Smith 7164416b707SBarry Smith static PetscErrorCode PetscSFSetFromOptions_Window(PetscOptionItems *PetscOptionsObject,PetscSF sf) 71795fce210SBarry Smith { 71895fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 71995fce210SBarry Smith PetscErrorCode ierr; 7205b0d146aSStefano Zampini PetscSFWindowFlavorType flavor = w->flavor; 72195fce210SBarry Smith 72295fce210SBarry Smith PetscFunctionBegin; 723e55864a3SBarry Smith ierr = PetscOptionsHead(PetscOptionsObject,"PetscSF Window options");CHKERRQ(ierr); 72495fce210SBarry Smith ierr = PetscOptionsEnum("-sf_window_sync","synchronization type to use for PetscSF Window communication","PetscSFWindowSetSyncType",PetscSFWindowSyncTypes,(PetscEnum)w->sync,(PetscEnum*)&w->sync,NULL);CHKERRQ(ierr); 7255b0d146aSStefano Zampini ierr = PetscOptionsEnum("-sf_window_flavor","flavor to use for PetscSF Window creation","PetscSFWindowSetFlavorType",PetscSFWindowFlavorTypes,(PetscEnum)flavor,(PetscEnum*)&flavor,NULL);CHKERRQ(ierr); 7265b0d146aSStefano Zampini ierr = PetscSFWindowSetFlavorType(sf,flavor);CHKERRQ(ierr); 72795fce210SBarry Smith ierr = PetscOptionsTail();CHKERRQ(ierr); 72895fce210SBarry Smith PetscFunctionReturn(0); 72995fce210SBarry Smith } 73095fce210SBarry Smith 73195fce210SBarry Smith static PetscErrorCode PetscSFReset_Window(PetscSF sf) 73295fce210SBarry Smith { 73395fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 73495fce210SBarry Smith PetscErrorCode ierr; 73595fce210SBarry Smith PetscSFDataLink link,next; 73695fce210SBarry Smith PetscSFWinLink wlink,wnext; 73795fce210SBarry Smith PetscInt i; 73895fce210SBarry Smith 73995fce210SBarry Smith PetscFunctionBegin; 74095fce210SBarry Smith for (link=w->link; link; link=next) { 74195fce210SBarry Smith next = link->next; 742ffc4695bSBarry Smith ierr = MPI_Type_free(&link->unit);CHKERRMPI(ierr); 74395fce210SBarry Smith for (i=0; i<sf->nranks; i++) { 744ffc4695bSBarry Smith ierr = MPI_Type_free(&link->mine[i]);CHKERRMPI(ierr); 745ffc4695bSBarry Smith ierr = MPI_Type_free(&link->remote[i]);CHKERRMPI(ierr); 74695fce210SBarry Smith } 74795fce210SBarry Smith ierr = PetscFree2(link->mine,link->remote);CHKERRQ(ierr); 74895fce210SBarry Smith ierr = PetscFree(link);CHKERRQ(ierr); 74995fce210SBarry Smith } 75095fce210SBarry Smith w->link = NULL; 75195fce210SBarry Smith for (wlink=w->wins; wlink; wlink=wnext) { 75295fce210SBarry Smith wnext = wlink->next; 753*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(wlink->inuse,PetscObjectComm((PetscObject)sf),PETSC_ERR_ARG_WRONGSTATE,"Window still in use with address %p",(void*)wlink->addr); 7545b0d146aSStefano Zampini ierr = PetscFree(wlink->dyn_target_addr);CHKERRQ(ierr); 755684a874aSStefano Zampini ierr = PetscFree(wlink->reqs);CHKERRQ(ierr); 756ffc4695bSBarry Smith ierr = MPI_Win_free(&wlink->win);CHKERRMPI(ierr); 75795fce210SBarry Smith ierr = PetscFree(wlink);CHKERRQ(ierr); 75895fce210SBarry Smith } 75995fce210SBarry Smith w->wins = NULL; 7605b0d146aSStefano Zampini ierr = PetscSFDestroy(&w->dynsf);CHKERRQ(ierr); 7615b0d146aSStefano Zampini if (w->info != MPI_INFO_NULL) { 762ffc4695bSBarry Smith ierr = MPI_Info_free(&w->info);CHKERRMPI(ierr); 7635b0d146aSStefano Zampini } 76495fce210SBarry Smith PetscFunctionReturn(0); 76595fce210SBarry Smith } 76695fce210SBarry Smith 76795fce210SBarry Smith static PetscErrorCode PetscSFDestroy_Window(PetscSF sf) 76895fce210SBarry Smith { 76995fce210SBarry Smith PetscErrorCode ierr; 77095fce210SBarry Smith 77195fce210SBarry Smith PetscFunctionBegin; 77229046d53SLisandro Dalcin ierr = PetscSFReset_Window(sf);CHKERRQ(ierr); 77395fce210SBarry Smith ierr = PetscFree(sf->data);CHKERRQ(ierr); 774bdf89e91SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowSetSyncType_C",NULL);CHKERRQ(ierr); 775bdf89e91SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowGetSyncType_C",NULL);CHKERRQ(ierr); 7765b0d146aSStefano Zampini ierr = PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowSetFlavorType_C",NULL);CHKERRQ(ierr); 7775b0d146aSStefano Zampini ierr = PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowGetFlavorType_C",NULL);CHKERRQ(ierr); 7785b0d146aSStefano Zampini ierr = PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowSetInfo_C",NULL);CHKERRQ(ierr); 7795b0d146aSStefano Zampini ierr = PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowGetInfo_C",NULL);CHKERRQ(ierr); 78095fce210SBarry Smith PetscFunctionReturn(0); 78195fce210SBarry Smith } 78295fce210SBarry Smith 78395fce210SBarry Smith static PetscErrorCode PetscSFView_Window(PetscSF sf,PetscViewer viewer) 78495fce210SBarry Smith { 78595fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 78695fce210SBarry Smith PetscErrorCode ierr; 78795fce210SBarry Smith PetscBool iascii; 7885b0d146aSStefano Zampini PetscViewerFormat format; 78995fce210SBarry Smith 79095fce210SBarry Smith PetscFunctionBegin; 7915b0d146aSStefano Zampini ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 79295fce210SBarry Smith ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 79395fce210SBarry Smith if (iascii) { 79462152dedSBarry Smith ierr = PetscViewerASCIIPrintf(viewer," current flavor=%s synchronization=%s MultiSF sort=%s\n",PetscSFWindowFlavorTypes[w->flavor],PetscSFWindowSyncTypes[w->sync],sf->rankorder ? "rank-order" : "unordered");CHKERRQ(ierr); 7955b0d146aSStefano Zampini if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 7965b0d146aSStefano Zampini if (w->info != MPI_INFO_NULL) { 7975b0d146aSStefano Zampini PetscMPIInt k,nkeys; 7985b0d146aSStefano Zampini char key[MPI_MAX_INFO_KEY], value[MPI_MAX_INFO_VAL]; 7995b0d146aSStefano Zampini 800ffc4695bSBarry Smith ierr = MPI_Info_get_nkeys(w->info,&nkeys);CHKERRMPI(ierr); 8015b0d146aSStefano Zampini ierr = PetscViewerASCIIPrintf(viewer," current info with %d keys. Ordered key-value pairs follow:\n",nkeys);CHKERRQ(ierr); 8025b0d146aSStefano Zampini for (k = 0; k < nkeys; k++) { 8035b0d146aSStefano Zampini PetscMPIInt flag; 8045b0d146aSStefano Zampini 805ffc4695bSBarry Smith ierr = MPI_Info_get_nthkey(w->info,k,key);CHKERRMPI(ierr); 806ffc4695bSBarry Smith ierr = MPI_Info_get(w->info,key,MPI_MAX_INFO_VAL,value,&flag);CHKERRMPI(ierr); 807*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(!flag,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Missing key %s",key); 8085b0d146aSStefano Zampini ierr = PetscViewerASCIIPrintf(viewer," %s = %s\n",key,value);CHKERRQ(ierr); 8095b0d146aSStefano Zampini } 8105b0d146aSStefano Zampini } else { 8115b0d146aSStefano Zampini ierr = PetscViewerASCIIPrintf(viewer," current info=MPI_INFO_NULL\n");CHKERRQ(ierr); 8125b0d146aSStefano Zampini } 8135b0d146aSStefano Zampini } 81495fce210SBarry Smith } 81595fce210SBarry Smith PetscFunctionReturn(0); 81695fce210SBarry Smith } 81795fce210SBarry Smith 81895fce210SBarry Smith static PetscErrorCode PetscSFDuplicate_Window(PetscSF sf,PetscSFDuplicateOption opt,PetscSF newsf) 81995fce210SBarry Smith { 82095fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 82195fce210SBarry Smith PetscErrorCode ierr; 82295fce210SBarry Smith PetscSFWindowSyncType synctype; 82395fce210SBarry Smith 82495fce210SBarry Smith PetscFunctionBegin; 82595fce210SBarry Smith synctype = w->sync; 82695fce210SBarry Smith /* HACK: Must use FENCE or LOCK when called from PetscSFGetGroups() because ACTIVE here would cause recursion. */ 8275b0d146aSStefano Zampini if (!sf->setupcalled) synctype = PETSCSF_WINDOW_SYNC_LOCK; 82895fce210SBarry Smith ierr = PetscSFWindowSetSyncType(newsf,synctype);CHKERRQ(ierr); 8295b0d146aSStefano Zampini ierr = PetscSFWindowSetFlavorType(newsf,w->flavor);CHKERRQ(ierr); 8305b0d146aSStefano Zampini ierr = PetscSFWindowSetInfo(newsf,w->info);CHKERRQ(ierr); 83195fce210SBarry Smith PetscFunctionReturn(0); 83295fce210SBarry Smith } 83395fce210SBarry Smith 834ad227feaSJunchao Zhang static PetscErrorCode PetscSFBcastBegin_Window(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,const void *rootdata,PetscMemType leafmtype,void *leafdata,MPI_Op op) 83595fce210SBarry Smith { 83695fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 83795fce210SBarry Smith PetscErrorCode ierr; 83895fce210SBarry Smith PetscInt i,nranks; 83995fce210SBarry Smith const PetscMPIInt *ranks; 8405b0d146aSStefano Zampini const MPI_Aint *target_disp; 84195fce210SBarry Smith const MPI_Datatype *mine,*remote; 842684a874aSStefano Zampini MPI_Request *reqs; 84395fce210SBarry Smith MPI_Win win; 84495fce210SBarry Smith 84595fce210SBarry Smith PetscFunctionBegin; 846*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(op != MPI_REPLACE,PetscObjectComm((PetscObject)sf), PETSC_ERR_SUP, "PetscSFBcastBegin_Window with op!=MPI_REPLACE has not been implemented"); 847dec1416fSJunchao Zhang ierr = PetscSFGetRootRanks(sf,&nranks,&ranks,NULL,NULL,NULL);CHKERRQ(ierr); 84895fce210SBarry Smith ierr = PetscSFWindowGetDataTypes(sf,unit,&mine,&remote);CHKERRQ(ierr); 849684a874aSStefano Zampini ierr = PetscSFGetWindow(sf,unit,(void*)rootdata,w->sync,PETSC_TRUE,MPI_MODE_NOPUT|MPI_MODE_NOPRECEDE,MPI_MODE_NOPUT,0,&target_disp,&reqs,&win);CHKERRQ(ierr); 85095fce210SBarry Smith for (i=0; i<nranks; i++) { 8515b0d146aSStefano Zampini MPI_Aint tdp = target_disp ? target_disp[i] : 0; 8525b0d146aSStefano Zampini 853684a874aSStefano Zampini if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) { 854ffc4695bSBarry Smith ierr = MPI_Win_lock(MPI_LOCK_SHARED,ranks[i],MPI_MODE_NOCHECK,win);CHKERRMPI(ierr); 855684a874aSStefano Zampini #if defined(PETSC_HAVE_MPI_RGET) 856ffc4695bSBarry Smith ierr = MPI_Rget(leafdata,1,mine[i],ranks[i],tdp,1,remote[i],win,&reqs[i]);CHKERRMPI(ierr); 857684a874aSStefano Zampini #else 858ffc4695bSBarry Smith ierr = MPI_Get(leafdata,1,mine[i],ranks[i],tdp,1,remote[i],win);CHKERRMPI(ierr); 859684a874aSStefano Zampini #endif 860684a874aSStefano Zampini } else { 861ffc4695bSBarry Smith ierr = MPI_Get(leafdata,1,mine[i],ranks[i],tdp,1,remote[i],win);CHKERRMPI(ierr); 862684a874aSStefano Zampini } 86395fce210SBarry Smith } 86495fce210SBarry Smith PetscFunctionReturn(0); 86595fce210SBarry Smith } 86695fce210SBarry Smith 867ad227feaSJunchao Zhang PetscErrorCode PetscSFBcastEnd_Window(PetscSF sf,MPI_Datatype unit,const void *rootdata,void *leafdata,MPI_Op op) 86895fce210SBarry Smith { 8695b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 87095fce210SBarry Smith PetscErrorCode ierr; 87195fce210SBarry Smith MPI_Win win; 8724b9acda6SJunchao Zhang MPI_Request *reqs = NULL; 87395fce210SBarry Smith 87495fce210SBarry Smith PetscFunctionBegin; 875684a874aSStefano Zampini ierr = PetscSFFindWindow(sf,unit,rootdata,&win,&reqs);CHKERRQ(ierr); 876ffc4695bSBarry Smith if (reqs) {ierr = MPI_Waitall(sf->nranks,reqs,MPI_STATUSES_IGNORE);CHKERRMPI(ierr);} 877684a874aSStefano Zampini if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) { 878684a874aSStefano Zampini PetscInt i,nranks; 879684a874aSStefano Zampini const PetscMPIInt *ranks; 880684a874aSStefano Zampini 881684a874aSStefano Zampini ierr = PetscSFGetRootRanks(sf,&nranks,&ranks,NULL,NULL,NULL);CHKERRQ(ierr); 882684a874aSStefano Zampini for (i=0; i<nranks; i++) { 883ffc4695bSBarry Smith ierr = MPI_Win_unlock(ranks[i],win);CHKERRMPI(ierr); 884684a874aSStefano Zampini } 885684a874aSStefano Zampini } 8865b0d146aSStefano Zampini ierr = PetscSFRestoreWindow(sf,unit,(void*)rootdata,w->sync,PETSC_TRUE,MPI_MODE_NOSTORE|MPI_MODE_NOSUCCEED,PETSC_FALSE,&win);CHKERRQ(ierr); 88795fce210SBarry Smith PetscFunctionReturn(0); 88895fce210SBarry Smith } 88995fce210SBarry Smith 890eb02082bSJunchao Zhang PetscErrorCode PetscSFReduceBegin_Window(PetscSF sf,MPI_Datatype unit,PetscMemType leafmtype,const void *leafdata,PetscMemType rootmtype,void *rootdata,MPI_Op op) 89195fce210SBarry Smith { 89295fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 89395fce210SBarry Smith PetscErrorCode ierr; 89495fce210SBarry Smith PetscInt i,nranks; 89595fce210SBarry Smith const PetscMPIInt *ranks; 8965b0d146aSStefano Zampini const MPI_Aint *target_disp; 89795fce210SBarry Smith const MPI_Datatype *mine,*remote; 89895fce210SBarry Smith MPI_Win win; 89995fce210SBarry Smith 90095fce210SBarry Smith PetscFunctionBegin; 901dec1416fSJunchao Zhang ierr = PetscSFGetRootRanks(sf,&nranks,&ranks,NULL,NULL,NULL);CHKERRQ(ierr); 90295fce210SBarry Smith ierr = PetscSFWindowGetDataTypes(sf,unit,&mine,&remote);CHKERRQ(ierr); 90395fce210SBarry Smith ierr = PetscSFWindowOpTranslate(&op);CHKERRQ(ierr); 904684a874aSStefano Zampini ierr = PetscSFGetWindow(sf,unit,rootdata,w->sync,PETSC_TRUE,MPI_MODE_NOPRECEDE,0,0,&target_disp,NULL,&win);CHKERRQ(ierr); 90595fce210SBarry Smith for (i=0; i<nranks; i++) { 9065b0d146aSStefano Zampini MPI_Aint tdp = target_disp ? target_disp[i] : 0; 9075b0d146aSStefano Zampini 908ffc4695bSBarry Smith if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) {ierr = MPI_Win_lock(MPI_LOCK_SHARED,ranks[i],MPI_MODE_NOCHECK,win);CHKERRMPI(ierr);} 909c2c9d402SJose E. Roman ierr = MPI_Accumulate((void*)leafdata,1,mine[i],ranks[i],tdp,1,remote[i],op,win);CHKERRMPI(ierr); 910ffc4695bSBarry Smith if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) {ierr = MPI_Win_unlock(ranks[i],win);CHKERRMPI(ierr);} 91195fce210SBarry Smith } 91295fce210SBarry Smith PetscFunctionReturn(0); 91395fce210SBarry Smith } 91495fce210SBarry Smith 91500816365SJunchao Zhang static PetscErrorCode PetscSFReduceEnd_Window(PetscSF sf,MPI_Datatype unit,const void *leafdata,void *rootdata,MPI_Op op) 91695fce210SBarry Smith { 91795fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 91895fce210SBarry Smith PetscErrorCode ierr; 91995fce210SBarry Smith MPI_Win win; 9204b9acda6SJunchao Zhang MPI_Request *reqs = NULL; 92195fce210SBarry Smith 92295fce210SBarry Smith PetscFunctionBegin; 923684a874aSStefano Zampini ierr = PetscSFFindWindow(sf,unit,rootdata,&win,&reqs);CHKERRQ(ierr); 924ffc4695bSBarry Smith if (reqs) {ierr = MPI_Waitall(sf->nranks,reqs,MPI_STATUSES_IGNORE);CHKERRMPI(ierr);} 9255b0d146aSStefano Zampini ierr = PetscSFRestoreWindow(sf,unit,rootdata,w->sync,PETSC_TRUE,MPI_MODE_NOSUCCEED,PETSC_TRUE,&win);CHKERRQ(ierr); 92695fce210SBarry Smith PetscFunctionReturn(0); 92795fce210SBarry Smith } 9285b0d146aSStefano Zampini 929eb02082bSJunchao Zhang static PetscErrorCode PetscSFFetchAndOpBegin_Window(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,void *rootdata,PetscMemType leafmtype,const void *leafdata,void *leafupdate,MPI_Op op) 93095fce210SBarry Smith { 93195fce210SBarry Smith PetscErrorCode ierr; 93295fce210SBarry Smith PetscInt i,nranks; 93395fce210SBarry Smith const PetscMPIInt *ranks; 93495fce210SBarry Smith const MPI_Datatype *mine,*remote; 9355b0d146aSStefano Zampini const MPI_Aint *target_disp; 93695fce210SBarry Smith MPI_Win win; 9375b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 9385b0d146aSStefano Zampini #if !defined(PETSC_HAVE_MPI_GET_ACCUMULATE) 9395b0d146aSStefano Zampini PetscSFWindowFlavorType oldf; 9405b0d146aSStefano Zampini #endif 94195fce210SBarry Smith 94295fce210SBarry Smith PetscFunctionBegin; 943dec1416fSJunchao Zhang ierr = PetscSFGetRootRanks(sf,&nranks,&ranks,NULL,NULL,NULL);CHKERRQ(ierr); 94495fce210SBarry Smith ierr = PetscSFWindowGetDataTypes(sf,unit,&mine,&remote);CHKERRQ(ierr); 94595fce210SBarry Smith ierr = PetscSFWindowOpTranslate(&op);CHKERRQ(ierr); 9465b0d146aSStefano Zampini #if !defined(PETSC_HAVE_MPI_GET_ACCUMULATE) 9475b0d146aSStefano Zampini /* FetchAndOp without MPI_Get_Accumulate requires locking. 9485b0d146aSStefano Zampini we create a new window every time to not interfere with user-defined MPI_Info which may have used "no_locks"="true" */ 9495b0d146aSStefano Zampini oldf = w->flavor; 9505b0d146aSStefano Zampini w->flavor = PETSCSF_WINDOW_FLAVOR_CREATE; 951684a874aSStefano Zampini ierr = PetscSFGetWindow(sf,unit,rootdata,PETSCSF_WINDOW_SYNC_LOCK,PETSC_FALSE,0,0,0,&target_disp,NULL,&win);CHKERRQ(ierr); 9525b0d146aSStefano Zampini #else 953684a874aSStefano Zampini ierr = PetscSFGetWindow(sf,unit,rootdata,w->sync,PETSC_TRUE,MPI_MODE_NOPRECEDE,0,0,&target_disp,NULL,&win);CHKERRQ(ierr); 9545b0d146aSStefano Zampini #endif 9555b0d146aSStefano Zampini for (i=0; i<nranks; i++) { 9565b0d146aSStefano Zampini MPI_Aint tdp = target_disp ? target_disp[i] : 0; 9575b0d146aSStefano Zampini 9585b0d146aSStefano Zampini #if !defined(PETSC_HAVE_MPI_GET_ACCUMULATE) 959ffc4695bSBarry Smith ierr = MPI_Win_lock(MPI_LOCK_EXCLUSIVE,ranks[i],0,win);CHKERRMPI(ierr); 960ffc4695bSBarry Smith ierr = MPI_Get(leafupdate,1,mine[i],ranks[i],tdp,1,remote[i],win);CHKERRMPI(ierr); 961c2c9d402SJose E. Roman ierr = MPI_Accumulate((void*)leafdata,1,mine[i],ranks[i],tdp,1,remote[i],op,win);CHKERRMPI(ierr); 962ffc4695bSBarry Smith ierr = MPI_Win_unlock(ranks[i],win);CHKERRMPI(ierr); 9635b0d146aSStefano Zampini #else 964ffc4695bSBarry Smith if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) {ierr = MPI_Win_lock(MPI_LOCK_SHARED,ranks[i],0,win);CHKERRMPI(ierr);} 965c2c9d402SJose E. Roman ierr = MPI_Get_accumulate((void*)leafdata,1,mine[i],leafupdate,1,mine[i],ranks[i],tdp,1,remote[i],op,win);CHKERRMPI(ierr); 966ffc4695bSBarry Smith if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) {ierr = MPI_Win_unlock(ranks[i],win);CHKERRMPI(ierr);} 9675b0d146aSStefano Zampini #endif 9685b0d146aSStefano Zampini } 9695b0d146aSStefano Zampini #if !defined(PETSC_HAVE_MPI_GET_ACCUMULATE) 9705b0d146aSStefano Zampini w->flavor = oldf; 9715b0d146aSStefano Zampini #endif 97295fce210SBarry Smith PetscFunctionReturn(0); 97395fce210SBarry Smith } 97495fce210SBarry Smith 97500816365SJunchao Zhang static PetscErrorCode PetscSFFetchAndOpEnd_Window(PetscSF sf,MPI_Datatype unit,void *rootdata,const void *leafdata,void *leafupdate,MPI_Op op) 97695fce210SBarry Smith { 97795fce210SBarry Smith PetscErrorCode ierr; 97895fce210SBarry Smith MPI_Win win; 9795b0d146aSStefano Zampini #if defined(PETSC_HAVE_MPI_GET_ACCUMULATE) 9805b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 9815b0d146aSStefano Zampini #endif 9824b9acda6SJunchao Zhang MPI_Request *reqs = NULL; 98395fce210SBarry Smith 98495fce210SBarry Smith PetscFunctionBegin; 985684a874aSStefano Zampini ierr = PetscSFFindWindow(sf,unit,rootdata,&win,&reqs);CHKERRQ(ierr); 986ffc4695bSBarry Smith if (reqs) {ierr = MPI_Waitall(sf->nranks,reqs,MPI_STATUSES_IGNORE);CHKERRMPI(ierr);} 9875b0d146aSStefano Zampini #if defined(PETSC_HAVE_MPI_GET_ACCUMULATE) 9885b0d146aSStefano Zampini ierr = PetscSFRestoreWindow(sf,unit,rootdata,w->sync,PETSC_TRUE,MPI_MODE_NOSUCCEED,PETSC_TRUE,&win);CHKERRQ(ierr); 9895b0d146aSStefano Zampini #else 9905b0d146aSStefano Zampini ierr = PetscSFRestoreWindow(sf,unit,rootdata,PETSCSF_WINDOW_SYNC_LOCK,PETSC_FALSE,0,PETSC_TRUE,&win);CHKERRQ(ierr); 9915b0d146aSStefano Zampini #endif 99295fce210SBarry Smith PetscFunctionReturn(0); 99395fce210SBarry Smith } 99495fce210SBarry Smith 995dec1416fSJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFCreate_Window(PetscSF sf) 99695fce210SBarry Smith { 99795fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 99895fce210SBarry Smith PetscErrorCode ierr; 99995fce210SBarry Smith 100095fce210SBarry Smith PetscFunctionBegin; 100195fce210SBarry Smith sf->ops->SetUp = PetscSFSetUp_Window; 100295fce210SBarry Smith sf->ops->SetFromOptions = PetscSFSetFromOptions_Window; 100395fce210SBarry Smith sf->ops->Reset = PetscSFReset_Window; 100495fce210SBarry Smith sf->ops->Destroy = PetscSFDestroy_Window; 100595fce210SBarry Smith sf->ops->View = PetscSFView_Window; 100695fce210SBarry Smith sf->ops->Duplicate = PetscSFDuplicate_Window; 1007ad227feaSJunchao Zhang sf->ops->BcastBegin = PetscSFBcastBegin_Window; 1008ad227feaSJunchao Zhang sf->ops->BcastEnd = PetscSFBcastEnd_Window; 100995fce210SBarry Smith sf->ops->ReduceBegin = PetscSFReduceBegin_Window; 101095fce210SBarry Smith sf->ops->ReduceEnd = PetscSFReduceEnd_Window; 101195fce210SBarry Smith sf->ops->FetchAndOpBegin = PetscSFFetchAndOpBegin_Window; 101295fce210SBarry Smith sf->ops->FetchAndOpEnd = PetscSFFetchAndOpEnd_Window; 101395fce210SBarry Smith 1014b00a9115SJed Brown ierr = PetscNewLog(sf,&w);CHKERRQ(ierr); 101595fce210SBarry Smith sf->data = (void*)w; 101695fce210SBarry Smith w->sync = PETSCSF_WINDOW_SYNC_FENCE; 10175b0d146aSStefano Zampini w->flavor = PETSCSF_WINDOW_FLAVOR_CREATE; 10185b0d146aSStefano Zampini w->info = MPI_INFO_NULL; 101995fce210SBarry Smith 1020bdf89e91SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowSetSyncType_C",PetscSFWindowSetSyncType_Window);CHKERRQ(ierr); 1021bdf89e91SBarry Smith ierr = PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowGetSyncType_C",PetscSFWindowGetSyncType_Window);CHKERRQ(ierr); 10225b0d146aSStefano Zampini ierr = PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowSetFlavorType_C",PetscSFWindowSetFlavorType_Window);CHKERRQ(ierr); 10235b0d146aSStefano Zampini ierr = PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowGetFlavorType_C",PetscSFWindowGetFlavorType_Window);CHKERRQ(ierr); 10245b0d146aSStefano Zampini ierr = PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowSetInfo_C",PetscSFWindowSetInfo_Window);CHKERRQ(ierr); 10255b0d146aSStefano Zampini ierr = PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowGetInfo_C",PetscSFWindowGetInfo_Window);CHKERRQ(ierr); 102695fce210SBarry Smith 102795fce210SBarry Smith #if defined(OMPI_MAJOR_VERSION) && (OMPI_MAJOR_VERSION < 1 || (OMPI_MAJOR_VERSION == 1 && OMPI_MINOR_VERSION <= 6)) 102895fce210SBarry Smith { 102995fce210SBarry Smith PetscBool ackbug = PETSC_FALSE; 1030c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-acknowledge_ompi_onesided_bug",&ackbug,NULL);CHKERRQ(ierr); 103195fce210SBarry Smith if (ackbug) { 1032955c1f14SBarry Smith ierr = PetscInfo(sf,"Acknowledged Open MPI bug, proceeding anyway. Expect memory corruption.\n");CHKERRQ(ierr); 103395fce210SBarry Smith } else SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_LIB,"Open MPI is known to be buggy (https://svn.open-mpi.org/trac/ompi/ticket/1905 and 2656), use -acknowledge_ompi_onesided_bug to proceed"); 103495fce210SBarry Smith } 103595fce210SBarry Smith #endif 103695fce210SBarry Smith PetscFunctionReturn(0); 103795fce210SBarry Smith } 1038