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 PetscSFDataLink link; 6995fce210SBarry Smith PetscInt i,nranks; 7095fce210SBarry Smith const PetscInt *roffset,*rmine,*rremote; 7195fce210SBarry Smith const PetscMPIInt *ranks; 7295fce210SBarry Smith 7395fce210SBarry Smith PetscFunctionBegin; 7495fce210SBarry Smith /* Look for types in cache */ 7595fce210SBarry Smith for (link=w->link; link; link=link->next) { 7695fce210SBarry Smith PetscBool match; 775f80ce2aSJacob Faibussowitsch CHKERRQ(MPIPetsc_Type_compare(unit,link->unit,&match)); 7895fce210SBarry Smith if (match) { 7995fce210SBarry Smith *localtypes = link->mine; 8095fce210SBarry Smith *remotetypes = link->remote; 8195fce210SBarry Smith PetscFunctionReturn(0); 8295fce210SBarry Smith } 8395fce210SBarry Smith } 8495fce210SBarry Smith 8595fce210SBarry Smith /* Create new composite types for each send rank */ 865f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetRootRanks(sf,&nranks,&ranks,&roffset,&rmine,&rremote)); 875f80ce2aSJacob Faibussowitsch CHKERRQ(PetscNew(&link)); 885f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Type_dup(unit,&link->unit)); 895f80ce2aSJacob Faibussowitsch CHKERRQ(PetscMalloc2(nranks,&link->mine,nranks,&link->remote)); 9095fce210SBarry Smith for (i=0; i<nranks; i++) { 915b0d146aSStefano Zampini PetscInt rcount = roffset[i+1] - roffset[i]; 9295fce210SBarry Smith PetscMPIInt *rmine,*rremote; 9395fce210SBarry Smith #if !defined(PETSC_USE_64BIT_INDICES) 9495fce210SBarry Smith rmine = sf->rmine + sf->roffset[i]; 9595fce210SBarry Smith rremote = sf->rremote + sf->roffset[i]; 9695fce210SBarry Smith #else 9795fce210SBarry Smith PetscInt j; 985f80ce2aSJacob Faibussowitsch CHKERRQ(PetscMalloc2(rcount,&rmine,rcount,&rremote)); 9995fce210SBarry Smith for (j=0; j<rcount; j++) { 1005f80ce2aSJacob Faibussowitsch CHKERRQ(PetscMPIIntCast(sf->rmine[sf->roffset[i]+j],rmine+j)); 1015f80ce2aSJacob Faibussowitsch CHKERRQ(PetscMPIIntCast(sf->rremote[sf->roffset[i]+j],rremote+j)); 10295fce210SBarry Smith } 10395fce210SBarry Smith #endif 1045b0d146aSStefano Zampini 1055f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Type_create_indexed_block(rcount,1,rmine,link->unit,&link->mine[i])); 1065f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Type_create_indexed_block(rcount,1,rremote,link->unit,&link->remote[i])); 10795fce210SBarry Smith #if defined(PETSC_USE_64BIT_INDICES) 1085f80ce2aSJacob Faibussowitsch CHKERRQ(PetscFree2(rmine,rremote)); 10995fce210SBarry Smith #endif 1105f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Type_commit(&link->mine[i])); 1115f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Type_commit(&link->remote[i])); 11295fce210SBarry Smith } 11395fce210SBarry Smith link->next = w->link; 11495fce210SBarry Smith w->link = link; 11595fce210SBarry Smith 11695fce210SBarry Smith *localtypes = link->mine; 11795fce210SBarry Smith *remotetypes = link->remote; 11895fce210SBarry Smith PetscFunctionReturn(0); 11995fce210SBarry Smith } 12095fce210SBarry Smith 12195fce210SBarry Smith /*@C 1225b0d146aSStefano Zampini PetscSFWindowSetFlavorType - Set flavor type for MPI_Win creation 1235b0d146aSStefano Zampini 1245b0d146aSStefano Zampini Logically Collective 1255b0d146aSStefano Zampini 1264165533cSJose E. Roman Input Parameters: 1275b0d146aSStefano Zampini + sf - star forest for communication 1285b0d146aSStefano Zampini - flavor - flavor type 1295b0d146aSStefano Zampini 1305b0d146aSStefano Zampini Options Database Key: 1315b0d146aSStefano Zampini . -sf_window_flavor <flavor> - sets the flavor type CREATE, DYNAMIC, ALLOCATE or SHARED (see PetscSFWindowFlavorType) 1325b0d146aSStefano Zampini 1335b0d146aSStefano Zampini Level: advanced 1345b0d146aSStefano Zampini 1355b0d146aSStefano Zampini Notes: Windows reusage follow this rules: 1365b0d146aSStefano Zampini 1375b0d146aSStefano Zampini PETSCSF_WINDOW_FLAVOR_CREATE: creates a new window every time, uses MPI_Win_create 1385b0d146aSStefano Zampini 1395b0d146aSStefano 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. 1405b0d146aSStefano Zampini for i=1 to K 1415b0d146aSStefano Zampini PetscSFOperationBegin(rootdata1,leafdata_whatever); 1425b0d146aSStefano Zampini PetscSFOperationEnd(rootdata1,leafdata_whatever); 1435b0d146aSStefano Zampini ... 1445b0d146aSStefano Zampini PetscSFOperationBegin(rootdataN,leafdata_whatever); 1455b0d146aSStefano Zampini PetscSFOperationEnd(rootdataN,leafdata_whatever); 1465b0d146aSStefano Zampini endfor 1475b0d146aSStefano Zampini The following pattern will instead raise an error 1485b0d146aSStefano Zampini PetscSFOperationBegin(rootdata1,leafdata_whatever); 1495b0d146aSStefano Zampini PetscSFOperationEnd(rootdata1,leafdata_whatever); 1505b0d146aSStefano Zampini PetscSFOperationBegin(rank ? rootdata1 : rootdata2,leafdata_whatever); 1515b0d146aSStefano Zampini PetscSFOperationEnd(rank ? rootdata1 : rootdata2,leafdata_whatever); 1525b0d146aSStefano Zampini 1535b0d146aSStefano Zampini PETSCSF_WINDOW_FLAVOR_ALLOCATE: uses MPI_Win_allocate, reuses any pre-existing window which fits the data and it is not in use 1545b0d146aSStefano Zampini 1555b0d146aSStefano Zampini PETSCSF_WINDOW_FLAVOR_SHARED: uses MPI_Win_allocate_shared, reusage policy as for PETSCSF_WINDOW_FLAVOR_ALLOCATE 1565b0d146aSStefano Zampini 1575b0d146aSStefano Zampini .seealso: PetscSFSetFromOptions(), PetscSFWindowGetFlavorType() 1585b0d146aSStefano Zampini @*/ 1595b0d146aSStefano Zampini PetscErrorCode PetscSFWindowSetFlavorType(PetscSF sf,PetscSFWindowFlavorType flavor) 1605b0d146aSStefano Zampini { 1615b0d146aSStefano Zampini PetscFunctionBegin; 1625b0d146aSStefano Zampini PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1); 1635b0d146aSStefano Zampini PetscValidLogicalCollectiveEnum(sf,flavor,2); 1645f80ce2aSJacob Faibussowitsch CHKERRQ(PetscTryMethod(sf,"PetscSFWindowSetFlavorType_C",(PetscSF,PetscSFWindowFlavorType),(sf,flavor))); 1655b0d146aSStefano Zampini PetscFunctionReturn(0); 1665b0d146aSStefano Zampini } 1675b0d146aSStefano Zampini 1685b0d146aSStefano Zampini static PetscErrorCode PetscSFWindowSetFlavorType_Window(PetscSF sf,PetscSFWindowFlavorType flavor) 1695b0d146aSStefano Zampini { 1705b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 1715b0d146aSStefano Zampini 1725b0d146aSStefano Zampini PetscFunctionBegin; 1735b0d146aSStefano Zampini w->flavor = flavor; 1745b0d146aSStefano Zampini PetscFunctionReturn(0); 1755b0d146aSStefano Zampini } 1765b0d146aSStefano Zampini 1775b0d146aSStefano Zampini /*@C 1785b0d146aSStefano Zampini PetscSFWindowGetFlavorType - Get flavor type for PetscSF communication 1795b0d146aSStefano Zampini 1805b0d146aSStefano Zampini Logically Collective 1815b0d146aSStefano Zampini 1824165533cSJose E. Roman Input Parameter: 1835b0d146aSStefano Zampini . sf - star forest for communication 1845b0d146aSStefano Zampini 1854165533cSJose E. Roman Output Parameter: 1865b0d146aSStefano Zampini . flavor - flavor type 1875b0d146aSStefano Zampini 1885b0d146aSStefano Zampini Level: advanced 1895b0d146aSStefano Zampini 1905b0d146aSStefano Zampini .seealso: PetscSFSetFromOptions(), PetscSFWindowSetFlavorType() 1915b0d146aSStefano Zampini @*/ 1925b0d146aSStefano Zampini PetscErrorCode PetscSFWindowGetFlavorType(PetscSF sf,PetscSFWindowFlavorType *flavor) 1935b0d146aSStefano Zampini { 1945b0d146aSStefano Zampini PetscFunctionBegin; 1955b0d146aSStefano Zampini PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1); 1965b0d146aSStefano Zampini PetscValidPointer(flavor,2); 1975f80ce2aSJacob Faibussowitsch CHKERRQ(PetscUseMethod(sf,"PetscSFWindowGetFlavorType_C",(PetscSF,PetscSFWindowFlavorType*),(sf,flavor))); 1985b0d146aSStefano Zampini PetscFunctionReturn(0); 1995b0d146aSStefano Zampini } 2005b0d146aSStefano Zampini 2015b0d146aSStefano Zampini static PetscErrorCode PetscSFWindowGetFlavorType_Window(PetscSF sf,PetscSFWindowFlavorType *flavor) 2025b0d146aSStefano Zampini { 2035b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 2045b0d146aSStefano Zampini 2055b0d146aSStefano Zampini PetscFunctionBegin; 2065b0d146aSStefano Zampini *flavor = w->flavor; 2075b0d146aSStefano Zampini PetscFunctionReturn(0); 2085b0d146aSStefano Zampini } 2095b0d146aSStefano Zampini 2105b0d146aSStefano Zampini /*@C 2115b0d146aSStefano Zampini PetscSFWindowSetSyncType - Set synchronization type for PetscSF communication 21295fce210SBarry Smith 21395fce210SBarry Smith Logically Collective 21495fce210SBarry Smith 2154165533cSJose E. Roman Input Parameters: 21695fce210SBarry Smith + sf - star forest for communication 21795fce210SBarry Smith - sync - synchronization type 21895fce210SBarry Smith 21995fce210SBarry Smith Options Database Key: 22060263706SJed Brown . -sf_window_sync <sync> - sets the synchronization type FENCE, LOCK, or ACTIVE (see PetscSFWindowSyncType) 22195fce210SBarry Smith 22295fce210SBarry Smith Level: advanced 22395fce210SBarry Smith 22495fce210SBarry Smith .seealso: PetscSFSetFromOptions(), PetscSFWindowGetSyncType() 22595fce210SBarry Smith @*/ 22695fce210SBarry Smith PetscErrorCode PetscSFWindowSetSyncType(PetscSF sf,PetscSFWindowSyncType sync) 22795fce210SBarry Smith { 22895fce210SBarry Smith PetscFunctionBegin; 22995fce210SBarry Smith PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1); 23095fce210SBarry Smith PetscValidLogicalCollectiveEnum(sf,sync,2); 2315f80ce2aSJacob Faibussowitsch CHKERRQ(PetscTryMethod(sf,"PetscSFWindowSetSyncType_C",(PetscSF,PetscSFWindowSyncType),(sf,sync))); 23295fce210SBarry Smith PetscFunctionReturn(0); 23395fce210SBarry Smith } 23495fce210SBarry Smith 235f7a08781SBarry Smith static PetscErrorCode PetscSFWindowSetSyncType_Window(PetscSF sf,PetscSFWindowSyncType sync) 23695fce210SBarry Smith { 23795fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 23895fce210SBarry Smith 23995fce210SBarry Smith PetscFunctionBegin; 24095fce210SBarry Smith w->sync = sync; 24195fce210SBarry Smith PetscFunctionReturn(0); 24295fce210SBarry Smith } 24395fce210SBarry Smith 24495fce210SBarry Smith /*@C 2455b0d146aSStefano Zampini PetscSFWindowGetSyncType - Get synchronization type for PetscSF communication 24695fce210SBarry Smith 24795fce210SBarry Smith Logically Collective 24895fce210SBarry Smith 2494165533cSJose E. Roman Input Parameter: 25095fce210SBarry Smith . sf - star forest for communication 25195fce210SBarry Smith 2524165533cSJose E. Roman Output Parameter: 25395fce210SBarry Smith . sync - synchronization type 25495fce210SBarry Smith 25595fce210SBarry Smith Level: advanced 25695fce210SBarry Smith 2575b0d146aSStefano Zampini .seealso: PetscSFSetFromOptions(), PetscSFWindowSetSyncType() 25895fce210SBarry Smith @*/ 25995fce210SBarry Smith PetscErrorCode PetscSFWindowGetSyncType(PetscSF sf,PetscSFWindowSyncType *sync) 26095fce210SBarry Smith { 26195fce210SBarry Smith PetscFunctionBegin; 26295fce210SBarry Smith PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1); 26395fce210SBarry Smith PetscValidPointer(sync,2); 2645f80ce2aSJacob Faibussowitsch CHKERRQ(PetscUseMethod(sf,"PetscSFWindowGetSyncType_C",(PetscSF,PetscSFWindowSyncType*),(sf,sync))); 26595fce210SBarry Smith PetscFunctionReturn(0); 26695fce210SBarry Smith } 26795fce210SBarry Smith 268f7a08781SBarry Smith static PetscErrorCode PetscSFWindowGetSyncType_Window(PetscSF sf,PetscSFWindowSyncType *sync) 26995fce210SBarry Smith { 27095fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 27195fce210SBarry Smith 27295fce210SBarry Smith PetscFunctionBegin; 27395fce210SBarry Smith *sync = w->sync; 27495fce210SBarry Smith PetscFunctionReturn(0); 27595fce210SBarry Smith } 27695fce210SBarry Smith 27795fce210SBarry Smith /*@C 2785b0d146aSStefano Zampini PetscSFWindowSetInfo - Set the MPI_Info handle that will be used for subsequent windows allocation 2795b0d146aSStefano Zampini 2805b0d146aSStefano Zampini Logically Collective 2815b0d146aSStefano Zampini 2824165533cSJose E. Roman Input Parameters: 2835b0d146aSStefano Zampini + sf - star forest for communication 2845b0d146aSStefano Zampini - info - MPI_Info handle 2855b0d146aSStefano Zampini 2865b0d146aSStefano Zampini Level: advanced 2875b0d146aSStefano Zampini 2885b0d146aSStefano Zampini Notes: the info handle is duplicated with a call to MPI_Info_dup unless info = MPI_INFO_NULL. 2895b0d146aSStefano Zampini 2905b0d146aSStefano Zampini .seealso: PetscSFSetFromOptions(), PetscSFWindowGetInfo() 2915b0d146aSStefano Zampini @*/ 2925b0d146aSStefano Zampini PetscErrorCode PetscSFWindowSetInfo(PetscSF sf,MPI_Info info) 2935b0d146aSStefano Zampini { 2945b0d146aSStefano Zampini PetscFunctionBegin; 2955b0d146aSStefano Zampini PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1); 2965f80ce2aSJacob Faibussowitsch CHKERRQ(PetscTryMethod(sf,"PetscSFWindowSetInfo_C",(PetscSF,MPI_Info),(sf,info))); 2975b0d146aSStefano Zampini PetscFunctionReturn(0); 2985b0d146aSStefano Zampini } 2995b0d146aSStefano Zampini 3005b0d146aSStefano Zampini static PetscErrorCode PetscSFWindowSetInfo_Window(PetscSF sf,MPI_Info info) 3015b0d146aSStefano Zampini { 3025b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 3035b0d146aSStefano Zampini 3045b0d146aSStefano Zampini PetscFunctionBegin; 3055b0d146aSStefano Zampini if (w->info != MPI_INFO_NULL) { 3065f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Info_free(&w->info)); 3075b0d146aSStefano Zampini } 3085b0d146aSStefano Zampini if (info != MPI_INFO_NULL) { 3095f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Info_dup(info,&w->info)); 3105b0d146aSStefano Zampini } 3115b0d146aSStefano Zampini PetscFunctionReturn(0); 3125b0d146aSStefano Zampini } 3135b0d146aSStefano Zampini 3145b0d146aSStefano Zampini /*@C 3155b0d146aSStefano Zampini PetscSFWindowGetInfo - Get the MPI_Info handle used for windows allocation 3165b0d146aSStefano Zampini 3175b0d146aSStefano Zampini Logically Collective 3185b0d146aSStefano Zampini 3194165533cSJose E. Roman Input Parameter: 3205b0d146aSStefano Zampini . sf - star forest for communication 3215b0d146aSStefano Zampini 3224165533cSJose E. Roman Output Parameter: 3235b0d146aSStefano Zampini . info - MPI_Info handle 3245b0d146aSStefano Zampini 3255b0d146aSStefano Zampini Level: advanced 3265b0d146aSStefano Zampini 3275b0d146aSStefano Zampini Notes: if PetscSFWindowSetInfo() has not be called, this returns MPI_INFO_NULL 3285b0d146aSStefano Zampini 3295b0d146aSStefano Zampini .seealso: PetscSFSetFromOptions(), PetscSFWindowSetInfo() 3305b0d146aSStefano Zampini @*/ 3315b0d146aSStefano Zampini PetscErrorCode PetscSFWindowGetInfo(PetscSF sf,MPI_Info *info) 3325b0d146aSStefano Zampini { 3335b0d146aSStefano Zampini PetscFunctionBegin; 3345b0d146aSStefano Zampini PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1); 3355b0d146aSStefano Zampini PetscValidPointer(info,2); 3365f80ce2aSJacob Faibussowitsch CHKERRQ(PetscUseMethod(sf,"PetscSFWindowGetInfo_C",(PetscSF,MPI_Info*),(sf,info))); 3375b0d146aSStefano Zampini PetscFunctionReturn(0); 3385b0d146aSStefano Zampini } 3395b0d146aSStefano Zampini 3405b0d146aSStefano Zampini static PetscErrorCode PetscSFWindowGetInfo_Window(PetscSF sf,MPI_Info *info) 3415b0d146aSStefano Zampini { 3425b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 3435b0d146aSStefano Zampini 3445b0d146aSStefano Zampini PetscFunctionBegin; 3455b0d146aSStefano Zampini *info = w->info; 3465b0d146aSStefano Zampini PetscFunctionReturn(0); 3475b0d146aSStefano Zampini } 3485b0d146aSStefano Zampini 3495b0d146aSStefano Zampini /* 35095fce210SBarry Smith PetscSFGetWindow - Get a window for use with a given data type 35195fce210SBarry Smith 35295fce210SBarry Smith Collective on PetscSF 35395fce210SBarry Smith 3544165533cSJose E. Roman Input Parameters: 35595fce210SBarry Smith + sf - star forest 35695fce210SBarry Smith . unit - data type 35795fce210SBarry Smith . array - array to be sent 3585b0d146aSStefano Zampini . sync - type of synchronization PetscSFWindowSyncType 35995fce210SBarry Smith . epoch - PETSC_TRUE to acquire the window and start an epoch, PETSC_FALSE to just acquire the window 3605b0d146aSStefano Zampini . fenceassert - assert parameter for call to MPI_Win_fence(), if sync == PETSCSF_WINDOW_SYNC_FENCE 3615b0d146aSStefano Zampini . postassert - assert parameter for call to MPI_Win_post(), if sync == PETSCSF_WINDOW_SYNC_ACTIVE 362684a874aSStefano Zampini - startassert - assert parameter for call to MPI_Win_start(), if sync == PETSCSF_WINDOW_SYNC_ACTIVE 36395fce210SBarry Smith 3644165533cSJose E. Roman Output Parameters: 365684a874aSStefano Zampini + target_disp - target_disp argument for RMA calls (significative for PETSCSF_WINDOW_FLAVOR_DYNAMIC only) 366684a874aSStefano Zampini + reqs - array of requests (significative for sync == PETSCSF_WINDOW_SYNC_LOCK only) 367684a874aSStefano Zampini - win - window 36895fce210SBarry Smith 36995fce210SBarry Smith Level: developer 370dec1416fSJunchao Zhang .seealso: PetscSFGetRootRanks(), PetscSFWindowGetDataTypes() 3715b0d146aSStefano Zampini */ 372684a874aSStefano 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) 37395fce210SBarry Smith { 37495fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 37595fce210SBarry Smith MPI_Aint lb,lb_true,bytes,bytes_true; 37695fce210SBarry Smith PetscSFWinLink link; 377d547623eSJunchao Zhang #if defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW) 3785b0d146aSStefano Zampini MPI_Aint winaddr; 3795b0d146aSStefano Zampini PetscInt nranks; 380d547623eSJunchao Zhang #endif 3815b0d146aSStefano Zampini PetscBool reuse = PETSC_FALSE, update = PETSC_FALSE; 3825b0d146aSStefano Zampini PetscBool dummy[2]; 3835b0d146aSStefano Zampini MPI_Aint wsize; 38495fce210SBarry Smith 38595fce210SBarry Smith PetscFunctionBegin; 3865f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Type_get_extent(unit,&lb,&bytes)); 3875f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Type_get_true_extent(unit,&lb_true,&bytes_true)); 3882c71b3e2SJacob 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"); 3892c71b3e2SJacob 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"); 3905b0d146aSStefano Zampini if (w->flavor != PETSCSF_WINDOW_FLAVOR_CREATE) reuse = PETSC_TRUE; 3915b0d146aSStefano Zampini for (link=w->wins; reuse && link; link=link->next) { 3925b0d146aSStefano Zampini PetscBool winok = PETSC_FALSE; 3935b0d146aSStefano Zampini if (w->flavor != link->flavor) continue; 3945b0d146aSStefano Zampini switch (w->flavor) { 3955b0d146aSStefano 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) */ 3965b0d146aSStefano Zampini if (array == link->addr) { 39776bd3646SJed Brown if (PetscDefined(USE_DEBUG)) { 3985b0d146aSStefano Zampini dummy[0] = PETSC_TRUE; 3995b0d146aSStefano Zampini dummy[1] = PETSC_TRUE; 4005f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Allreduce(MPI_IN_PLACE,dummy,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)sf))); 4015f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Allreduce(MPI_IN_PLACE,dummy+1,1,MPIU_BOOL,MPI_LOR ,PetscObjectComm((PetscObject)sf))); 4022c71b3e2SJacob 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"); 40376bd3646SJed Brown } 404*28b400f6SJacob Faibussowitsch PetscCheck(!link->inuse,PetscObjectComm((PetscObject)sf),PETSC_ERR_PLIB,"Window in use"); 4052c71b3e2SJacob Faibussowitsch PetscCheckFalse(epoch && link->epoch,PetscObjectComm((PetscObject)sf),PETSC_ERR_PLIB,"Window epoch not finished"); 4065b0d146aSStefano Zampini winok = PETSC_TRUE; 4075b0d146aSStefano Zampini link->paddr = array; 40876bd3646SJed Brown } else if (PetscDefined(USE_DEBUG)) { 4095b0d146aSStefano Zampini dummy[0] = PETSC_FALSE; 4105b0d146aSStefano Zampini dummy[1] = PETSC_FALSE; 4115f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Allreduce(MPI_IN_PLACE,dummy ,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)sf))); 4125f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Allreduce(MPI_IN_PLACE,dummy+1,1,MPIU_BOOL,MPI_LOR ,PetscObjectComm((PetscObject)sf))); 4132c71b3e2SJacob 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"); 4145b0d146aSStefano Zampini } 4155b0d146aSStefano Zampini break; 4165b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_ALLOCATE: /* check available by matching size, allocate if in use */ 4175b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_SHARED: 4185b0d146aSStefano Zampini if (!link->inuse && bytes == (MPI_Aint)link->bytes) { 4195b0d146aSStefano Zampini update = PETSC_TRUE; 4205b0d146aSStefano Zampini link->paddr = array; 4215b0d146aSStefano Zampini winok = PETSC_TRUE; 4225b0d146aSStefano Zampini } 4235b0d146aSStefano Zampini break; 42498921bdaSJacob Faibussowitsch default: SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_SUP,"No support for flavor %s",PetscSFWindowFlavorTypes[w->flavor]); 4255b0d146aSStefano Zampini } 4265b0d146aSStefano Zampini if (winok) { 4275b0d146aSStefano Zampini *win = link->win; 4285f80ce2aSJacob Faibussowitsch CHKERRQ(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))); 4295b0d146aSStefano Zampini goto found; 4305b0d146aSStefano Zampini } 4315b0d146aSStefano Zampini } 4325b0d146aSStefano Zampini 4335b0d146aSStefano Zampini wsize = (MPI_Aint)bytes*sf->nroots; 4345f80ce2aSJacob Faibussowitsch CHKERRQ(PetscNew(&link)); 43595fce210SBarry Smith link->bytes = bytes; 43695fce210SBarry Smith link->next = w->wins; 4375b0d146aSStefano Zampini link->flavor = w->flavor; 4385b0d146aSStefano Zampini link->dyn_target_addr = NULL; 439684a874aSStefano Zampini link->reqs = NULL; 44095fce210SBarry Smith w->wins = link; 441684a874aSStefano Zampini if (sync == PETSCSF_WINDOW_SYNC_LOCK) { 442684a874aSStefano Zampini PetscInt i; 443684a874aSStefano Zampini 4445f80ce2aSJacob Faibussowitsch CHKERRQ(PetscMalloc1(sf->nranks,&link->reqs)); 445684a874aSStefano Zampini for (i = 0; i < sf->nranks; i++) link->reqs[i] = MPI_REQUEST_NULL; 446684a874aSStefano Zampini } 4475b0d146aSStefano Zampini switch (w->flavor) { 4485b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_CREATE: 4495f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_create(array,wsize,(PetscMPIInt)bytes,w->info,PetscObjectComm((PetscObject)sf),&link->win)); 4505b0d146aSStefano Zampini link->addr = array; 4515b0d146aSStefano Zampini link->paddr = array; 4525b0d146aSStefano Zampini break; 453d547623eSJunchao Zhang #if defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW) 4545b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_DYNAMIC: 4555f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_create_dynamic(w->info,PetscObjectComm((PetscObject)sf),&link->win)); 4565b0d146aSStefano Zampini #if defined(PETSC_HAVE_OMPI_MAJOR_VERSION) /* some OpenMPI versions do not support MPI_Win_attach(win,NULL,0); */ 4575f80ce2aSJacob Faibussowitsch int dummy = 0; 4585f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_attach(link->win,wsize ? array : (void*)&dummy,wsize)); 4595b0d146aSStefano Zampini #else 4605f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_attach(link->win,array,wsize)); 4615b0d146aSStefano Zampini #endif 4625b0d146aSStefano Zampini link->addr = array; 4635b0d146aSStefano Zampini link->paddr = array; 464*28b400f6SJacob Faibussowitsch PetscCheck(w->dynsf,PetscObjectComm((PetscObject)sf),PETSC_ERR_ORDER,"Must call PetscSFSetUp()"); 4655f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFSetUp(w->dynsf)); 4665f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetRootRanks(w->dynsf,&nranks,NULL,NULL,NULL,NULL)); 4675f80ce2aSJacob Faibussowitsch CHKERRQ(PetscMalloc1(nranks,&link->dyn_target_addr)); 4685f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Get_address(array,&winaddr)); 4695f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFBcastBegin(w->dynsf,MPI_AINT,&winaddr,link->dyn_target_addr,MPI_REPLACE)); 4705f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFBcastEnd(w->dynsf,MPI_AINT,&winaddr,link->dyn_target_addr,MPI_REPLACE)); 4715b0d146aSStefano Zampini break; 4725b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_ALLOCATE: 4735f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_allocate(wsize,(PetscMPIInt)bytes,w->info,PetscObjectComm((PetscObject)sf),&link->addr,&link->win)); 4745b0d146aSStefano Zampini update = PETSC_TRUE; 4755b0d146aSStefano Zampini link->paddr = array; 4765b0d146aSStefano Zampini break; 477d547623eSJunchao Zhang #endif 4785b0d146aSStefano Zampini #if defined(PETSC_HAVE_MPI_PROCESS_SHARED_MEMORY) 4795b0d146aSStefano Zampini case PETSCSF_WINDOW_FLAVOR_SHARED: 4805f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_allocate_shared(wsize,(PetscMPIInt)bytes,w->info,PetscObjectComm((PetscObject)sf),&link->addr,&link->win)); 4815b0d146aSStefano Zampini update = PETSC_TRUE; 4825b0d146aSStefano Zampini link->paddr = array; 4835b0d146aSStefano Zampini break; 4845b0d146aSStefano Zampini #endif 48598921bdaSJacob Faibussowitsch default: SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_SUP,"No support for flavor %s",PetscSFWindowFlavorTypes[w->flavor]); 4865b0d146aSStefano Zampini } 4875f80ce2aSJacob Faibussowitsch CHKERRQ(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))); 48895fce210SBarry Smith *win = link->win; 48995fce210SBarry Smith 4905b0d146aSStefano Zampini found: 4915b0d146aSStefano Zampini 492684a874aSStefano Zampini if (target_disp) *target_disp = link->dyn_target_addr; 493684a874aSStefano Zampini if (reqs) *reqs = link->reqs; 494684a874aSStefano Zampini if (update) { /* locks are needed for the "separate" memory model only, the fence guaranties memory-synchronization */ 495684a874aSStefano Zampini PetscMPIInt rank; 496684a874aSStefano Zampini 4975f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank)); 4985f80ce2aSJacob Faibussowitsch if (sync == PETSCSF_WINDOW_SYNC_LOCK) CHKERRMPI(MPI_Win_lock(MPI_LOCK_EXCLUSIVE,rank,MPI_MODE_NOCHECK,*win)); 4995f80ce2aSJacob Faibussowitsch CHKERRQ(PetscMemcpy(link->addr,array,sf->nroots*bytes)); 5005b0d146aSStefano Zampini if (sync == PETSCSF_WINDOW_SYNC_LOCK) { 5015f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_unlock(rank,*win)); 5025f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_fence(0,*win)); 5035b0d146aSStefano Zampini } 5045b0d146aSStefano Zampini } 5055b0d146aSStefano Zampini link->inuse = PETSC_TRUE; 5065b0d146aSStefano Zampini link->epoch = epoch; 50795fce210SBarry Smith if (epoch) { 5085b0d146aSStefano Zampini switch (sync) { 50995fce210SBarry Smith case PETSCSF_WINDOW_SYNC_FENCE: 5105f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_fence(fenceassert,*win)); 51195fce210SBarry Smith break; 51295fce210SBarry Smith case PETSCSF_WINDOW_SYNC_LOCK: /* Handled outside */ 51395fce210SBarry Smith break; 51495fce210SBarry Smith case PETSCSF_WINDOW_SYNC_ACTIVE: { 51595fce210SBarry Smith MPI_Group ingroup,outgroup; 5165b0d146aSStefano Zampini PetscMPIInt isize,osize; 5175b0d146aSStefano Zampini 5185b0d146aSStefano Zampini /* OpenMPI 4.0.2 with btl=vader does not like calling 5195b0d146aSStefano Zampini - MPI_Win_complete when ogroup is empty 5205b0d146aSStefano Zampini - MPI_Win_wait when igroup is empty 5215b0d146aSStefano Zampini So, we do not even issue the corresponding start and post calls 5225b0d146aSStefano Zampini The MPI standard (Sec. 11.5.2 of MPI 3.1) only requires that 5235b0d146aSStefano Zampini start(outgroup) has a matching post(ingroup) 5245b0d146aSStefano Zampini and this is guaranteed by PetscSF 5255b0d146aSStefano Zampini */ 5265f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetGroups(sf,&ingroup,&outgroup)); 5275f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Group_size(ingroup,&isize)); 5285f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Group_size(outgroup,&osize)); 5295f80ce2aSJacob Faibussowitsch if (isize) CHKERRMPI(MPI_Win_post(ingroup,postassert,*win)); 5305f80ce2aSJacob Faibussowitsch if (osize) CHKERRMPI(MPI_Win_start(outgroup,startassert,*win)); 53195fce210SBarry Smith } break; 53295fce210SBarry Smith default: SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_PLIB,"Unknown synchronization type"); 53395fce210SBarry Smith } 53495fce210SBarry Smith } 53595fce210SBarry Smith PetscFunctionReturn(0); 53695fce210SBarry Smith } 53795fce210SBarry Smith 5385b0d146aSStefano Zampini /* 53995fce210SBarry Smith PetscSFFindWindow - Finds a window that is already in use 54095fce210SBarry Smith 54195fce210SBarry Smith Not Collective 54295fce210SBarry Smith 5434165533cSJose E. Roman Input Parameters: 54495fce210SBarry Smith + sf - star forest 54595fce210SBarry Smith . unit - data type 54695fce210SBarry Smith - array - array with which the window is associated 54795fce210SBarry Smith 5484165533cSJose E. Roman Output Parameters: 549684a874aSStefano Zampini + win - window 550684a874aSStefano Zampini - reqs - outstanding requests associated to the window 55195fce210SBarry Smith 55295fce210SBarry Smith Level: developer 55395fce210SBarry Smith 55495fce210SBarry Smith .seealso: PetscSFGetWindow(), PetscSFRestoreWindow() 5555b0d146aSStefano Zampini */ 556684a874aSStefano Zampini static PetscErrorCode PetscSFFindWindow(PetscSF sf,MPI_Datatype unit,const void *array,MPI_Win *win,MPI_Request **reqs) 55795fce210SBarry Smith { 55895fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 55995fce210SBarry Smith PetscSFWinLink link; 56095fce210SBarry Smith 56195fce210SBarry Smith PetscFunctionBegin; 562c0cd0301SJed Brown *win = MPI_WIN_NULL; 56395fce210SBarry Smith for (link=w->wins; link; link=link->next) { 5645b0d146aSStefano Zampini if (array == link->paddr) { 5652abc8c78SJacob Faibussowitsch 5665f80ce2aSJacob Faibussowitsch CHKERRQ(PetscInfo(sf,"Window %" PETSC_MPI_WIN_FMT " of flavor %d for comm %" PETSC_MPI_COMM_FMT "\n",link->win,link->flavor,PetscObjectComm((PetscObject)sf))); 56795fce210SBarry Smith *win = link->win; 568684a874aSStefano Zampini *reqs = link->reqs; 56995fce210SBarry Smith PetscFunctionReturn(0); 57095fce210SBarry Smith } 57195fce210SBarry Smith } 57295fce210SBarry Smith SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Requested window not in use"); 57395fce210SBarry Smith } 57495fce210SBarry Smith 5755b0d146aSStefano Zampini /* 57695fce210SBarry Smith PetscSFRestoreWindow - Restores a window obtained with PetscSFGetWindow() 57795fce210SBarry Smith 57895fce210SBarry Smith Collective 57995fce210SBarry Smith 5804165533cSJose E. Roman Input Parameters: 58195fce210SBarry Smith + sf - star forest 58295fce210SBarry Smith . unit - data type 58395fce210SBarry Smith . array - array associated with window 5845b0d146aSStefano Zampini . sync - type of synchronization PetscSFWindowSyncType 58595fce210SBarry Smith . epoch - close an epoch, must match argument to PetscSFGetWindow() 5865b0d146aSStefano Zampini . update - if we have to update the local window array 58795fce210SBarry Smith - win - window 58895fce210SBarry Smith 58995fce210SBarry Smith Level: developer 59095fce210SBarry Smith 59195fce210SBarry Smith .seealso: PetscSFFindWindow() 5925b0d146aSStefano Zampini */ 5935b0d146aSStefano Zampini static PetscErrorCode PetscSFRestoreWindow(PetscSF sf,MPI_Datatype unit,void *array,PetscSFWindowSyncType sync,PetscBool epoch,PetscMPIInt fenceassert,PetscBool update,MPI_Win *win) 59495fce210SBarry Smith { 59595fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 59695fce210SBarry Smith PetscSFWinLink *p,link; 5975b0d146aSStefano Zampini PetscBool reuse = PETSC_FALSE; 5985b0d146aSStefano Zampini PetscSFWindowFlavorType flavor; 5995b0d146aSStefano Zampini void* laddr; 6005b0d146aSStefano Zampini size_t bytes; 60195fce210SBarry Smith 60295fce210SBarry Smith PetscFunctionBegin; 60395fce210SBarry Smith for (p=&w->wins; *p; p=&(*p)->next) { 60495fce210SBarry Smith link = *p; 60595fce210SBarry Smith if (*win == link->win) { 6062c71b3e2SJacob Faibussowitsch PetscCheckFalse(array != link->paddr,PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Matched window, but not array"); 60795fce210SBarry Smith if (epoch != link->epoch) { 608*28b400f6SJacob Faibussowitsch PetscCheck(!epoch,PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"No epoch to end"); 60995fce210SBarry Smith else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Restoring window without ending epoch"); 61095fce210SBarry Smith } 6115b0d146aSStefano Zampini laddr = link->addr; 6125b0d146aSStefano Zampini flavor = link->flavor; 6135b0d146aSStefano Zampini bytes = link->bytes; 6145b0d146aSStefano Zampini if (flavor != PETSCSF_WINDOW_FLAVOR_CREATE) reuse = PETSC_TRUE; 6155b0d146aSStefano Zampini else { *p = link->next; update = PETSC_FALSE; } /* remove from list */ 61695fce210SBarry Smith goto found; 61795fce210SBarry Smith } 61895fce210SBarry Smith } 61995fce210SBarry Smith SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Requested window not in use"); 62095fce210SBarry Smith 62195fce210SBarry Smith found: 6225f80ce2aSJacob Faibussowitsch CHKERRQ(PetscInfo(sf,"Window %" PETSC_MPI_WIN_FMT " of flavor %d for comm %" PETSC_MPI_COMM_FMT "\n",link->win,link->flavor,PetscObjectComm((PetscObject)sf))); 62395fce210SBarry Smith if (epoch) { 6245b0d146aSStefano Zampini switch (sync) { 62595fce210SBarry Smith case PETSCSF_WINDOW_SYNC_FENCE: 6265f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_fence(fenceassert,*win)); 62795fce210SBarry Smith break; 6285b0d146aSStefano Zampini case PETSCSF_WINDOW_SYNC_LOCK: /* Handled outside */ 6295b0d146aSStefano Zampini break; 63095fce210SBarry Smith case PETSCSF_WINDOW_SYNC_ACTIVE: { 6315b0d146aSStefano Zampini MPI_Group ingroup,outgroup; 6325b0d146aSStefano Zampini PetscMPIInt isize,osize; 6335b0d146aSStefano Zampini 6345b0d146aSStefano Zampini /* OpenMPI 4.0.2 with btl=wader does not like calling 6355b0d146aSStefano Zampini - MPI_Win_complete when ogroup is empty 6365b0d146aSStefano Zampini - MPI_Win_wait when igroup is empty 6375b0d146aSStefano Zampini The MPI standard (Sec. 11.5.2 of MPI 3.1) only requires that 6385b0d146aSStefano Zampini - each process who issues a call to MPI_Win_start issues a call to MPI_Win_Complete 6395b0d146aSStefano Zampini - each process who issues a call to MPI_Win_post issues a call to MPI_Win_Wait 6405b0d146aSStefano Zampini */ 6415f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetGroups(sf,&ingroup,&outgroup)); 6425f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Group_size(ingroup,&isize)); 6435f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Group_size(outgroup,&osize)); 6445f80ce2aSJacob Faibussowitsch if (osize) CHKERRMPI(MPI_Win_complete(*win)); 6455f80ce2aSJacob Faibussowitsch if (isize) CHKERRMPI(MPI_Win_wait(*win)); 64695fce210SBarry Smith } break; 64795fce210SBarry Smith default: SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_PLIB,"Unknown synchronization type"); 64895fce210SBarry Smith } 64995fce210SBarry Smith } 6505b0d146aSStefano Zampini if (update) { 6515b0d146aSStefano Zampini if (sync == PETSCSF_WINDOW_SYNC_LOCK) { 6525f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_fence(MPI_MODE_NOPUT|MPI_MODE_NOSUCCEED,*win)); 6535b0d146aSStefano Zampini } 6545f80ce2aSJacob Faibussowitsch CHKERRQ(PetscMemcpy(array,laddr,sf->nroots*bytes)); 6555b0d146aSStefano Zampini } 6565b0d146aSStefano Zampini link->epoch = PETSC_FALSE; 6575b0d146aSStefano Zampini link->inuse = PETSC_FALSE; 6585b0d146aSStefano Zampini link->paddr = NULL; 6595b0d146aSStefano Zampini if (!reuse) { 6605f80ce2aSJacob Faibussowitsch CHKERRQ(PetscFree(link->dyn_target_addr)); 6615f80ce2aSJacob Faibussowitsch CHKERRQ(PetscFree(link->reqs)); 6625f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_free(&link->win)); 6635f80ce2aSJacob Faibussowitsch CHKERRQ(PetscFree(link)); 66495fce210SBarry Smith *win = MPI_WIN_NULL; 6655b0d146aSStefano Zampini } 66695fce210SBarry Smith PetscFunctionReturn(0); 66795fce210SBarry Smith } 66895fce210SBarry Smith 66995fce210SBarry Smith static PetscErrorCode PetscSFSetUp_Window(PetscSF sf) 67095fce210SBarry Smith { 67195fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 67295fce210SBarry Smith MPI_Group ingroup,outgroup; 67395fce210SBarry Smith 67495fce210SBarry Smith PetscFunctionBegin; 6755f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFSetUpRanks(sf,MPI_GROUP_EMPTY)); 6765b0d146aSStefano Zampini if (!w->dynsf) { 6775b0d146aSStefano Zampini PetscInt i; 6785b0d146aSStefano Zampini PetscSFNode *remotes; 6795b0d146aSStefano Zampini 6805f80ce2aSJacob Faibussowitsch CHKERRQ(PetscMalloc1(sf->nranks,&remotes)); 6815b0d146aSStefano Zampini for (i=0;i<sf->nranks;i++) { 6825b0d146aSStefano Zampini remotes[i].rank = sf->ranks[i]; 6835b0d146aSStefano Zampini remotes[i].index = 0; 6845b0d146aSStefano Zampini } 6855f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFDuplicate(sf,PETSCSF_DUPLICATE_RANKS,&w->dynsf)); 6865f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFWindowSetFlavorType(w->dynsf,PETSCSF_WINDOW_FLAVOR_CREATE)); /* break recursion */ 6875f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFSetGraph(w->dynsf,1,sf->nranks,NULL,PETSC_OWN_POINTER,remotes,PETSC_OWN_POINTER)); 6885f80ce2aSJacob Faibussowitsch CHKERRQ(PetscLogObjectParent((PetscObject)sf,(PetscObject)w->dynsf)); 6895b0d146aSStefano Zampini } 69095fce210SBarry Smith switch (w->sync) { 69195fce210SBarry Smith case PETSCSF_WINDOW_SYNC_ACTIVE: 6925f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetGroups(sf,&ingroup,&outgroup)); 69395fce210SBarry Smith default: 69495fce210SBarry Smith break; 69595fce210SBarry Smith } 69695fce210SBarry Smith PetscFunctionReturn(0); 69795fce210SBarry Smith } 69895fce210SBarry Smith 6994416b707SBarry Smith static PetscErrorCode PetscSFSetFromOptions_Window(PetscOptionItems *PetscOptionsObject,PetscSF sf) 70095fce210SBarry Smith { 70195fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 7025b0d146aSStefano Zampini PetscSFWindowFlavorType flavor = w->flavor; 70395fce210SBarry Smith 70495fce210SBarry Smith PetscFunctionBegin; 7055f80ce2aSJacob Faibussowitsch CHKERRQ(PetscOptionsHead(PetscOptionsObject,"PetscSF Window options")); 7065f80ce2aSJacob Faibussowitsch CHKERRQ(PetscOptionsEnum("-sf_window_sync","synchronization type to use for PetscSF Window communication","PetscSFWindowSetSyncType",PetscSFWindowSyncTypes,(PetscEnum)w->sync,(PetscEnum*)&w->sync,NULL)); 7075f80ce2aSJacob Faibussowitsch CHKERRQ(PetscOptionsEnum("-sf_window_flavor","flavor to use for PetscSF Window creation","PetscSFWindowSetFlavorType",PetscSFWindowFlavorTypes,(PetscEnum)flavor,(PetscEnum*)&flavor,NULL)); 7085f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFWindowSetFlavorType(sf,flavor)); 7095f80ce2aSJacob Faibussowitsch CHKERRQ(PetscOptionsTail()); 71095fce210SBarry Smith PetscFunctionReturn(0); 71195fce210SBarry Smith } 71295fce210SBarry Smith 71395fce210SBarry Smith static PetscErrorCode PetscSFReset_Window(PetscSF sf) 71495fce210SBarry Smith { 71595fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 71695fce210SBarry Smith PetscSFDataLink link,next; 71795fce210SBarry Smith PetscSFWinLink wlink,wnext; 71895fce210SBarry Smith PetscInt i; 71995fce210SBarry Smith 72095fce210SBarry Smith PetscFunctionBegin; 72195fce210SBarry Smith for (link=w->link; link; link=next) { 72295fce210SBarry Smith next = link->next; 7235f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Type_free(&link->unit)); 72495fce210SBarry Smith for (i=0; i<sf->nranks; i++) { 7255f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Type_free(&link->mine[i])); 7265f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Type_free(&link->remote[i])); 72795fce210SBarry Smith } 7285f80ce2aSJacob Faibussowitsch CHKERRQ(PetscFree2(link->mine,link->remote)); 7295f80ce2aSJacob Faibussowitsch CHKERRQ(PetscFree(link)); 73095fce210SBarry Smith } 73195fce210SBarry Smith w->link = NULL; 73295fce210SBarry Smith for (wlink=w->wins; wlink; wlink=wnext) { 73395fce210SBarry Smith wnext = wlink->next; 734*28b400f6SJacob Faibussowitsch PetscCheck(!wlink->inuse,PetscObjectComm((PetscObject)sf),PETSC_ERR_ARG_WRONGSTATE,"Window still in use with address %p",(void*)wlink->addr); 7355f80ce2aSJacob Faibussowitsch CHKERRQ(PetscFree(wlink->dyn_target_addr)); 7365f80ce2aSJacob Faibussowitsch CHKERRQ(PetscFree(wlink->reqs)); 7375f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_free(&wlink->win)); 7385f80ce2aSJacob Faibussowitsch CHKERRQ(PetscFree(wlink)); 73995fce210SBarry Smith } 74095fce210SBarry Smith w->wins = NULL; 7415f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFDestroy(&w->dynsf)); 7425b0d146aSStefano Zampini if (w->info != MPI_INFO_NULL) { 7435f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Info_free(&w->info)); 7445b0d146aSStefano Zampini } 74595fce210SBarry Smith PetscFunctionReturn(0); 74695fce210SBarry Smith } 74795fce210SBarry Smith 74895fce210SBarry Smith static PetscErrorCode PetscSFDestroy_Window(PetscSF sf) 74995fce210SBarry Smith { 75095fce210SBarry Smith PetscFunctionBegin; 7515f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFReset_Window(sf)); 7525f80ce2aSJacob Faibussowitsch CHKERRQ(PetscFree(sf->data)); 7535f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowSetSyncType_C",NULL)); 7545f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowGetSyncType_C",NULL)); 7555f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowSetFlavorType_C",NULL)); 7565f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowGetFlavorType_C",NULL)); 7575f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowSetInfo_C",NULL)); 7585f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowGetInfo_C",NULL)); 75995fce210SBarry Smith PetscFunctionReturn(0); 76095fce210SBarry Smith } 76195fce210SBarry Smith 76295fce210SBarry Smith static PetscErrorCode PetscSFView_Window(PetscSF sf,PetscViewer viewer) 76395fce210SBarry Smith { 76495fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 76595fce210SBarry Smith PetscBool iascii; 7665b0d146aSStefano Zampini PetscViewerFormat format; 76795fce210SBarry Smith 76895fce210SBarry Smith PetscFunctionBegin; 7695f80ce2aSJacob Faibussowitsch CHKERRQ(PetscViewerGetFormat(viewer,&format)); 7705f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii)); 77195fce210SBarry Smith if (iascii) { 7725f80ce2aSJacob Faibussowitsch CHKERRQ(PetscViewerASCIIPrintf(viewer," current flavor=%s synchronization=%s MultiSF sort=%s\n",PetscSFWindowFlavorTypes[w->flavor],PetscSFWindowSyncTypes[w->sync],sf->rankorder ? "rank-order" : "unordered")); 7735b0d146aSStefano Zampini if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 7745b0d146aSStefano Zampini if (w->info != MPI_INFO_NULL) { 7755b0d146aSStefano Zampini PetscMPIInt k,nkeys; 7765b0d146aSStefano Zampini char key[MPI_MAX_INFO_KEY], value[MPI_MAX_INFO_VAL]; 7775b0d146aSStefano Zampini 7785f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Info_get_nkeys(w->info,&nkeys)); 7795f80ce2aSJacob Faibussowitsch CHKERRQ(PetscViewerASCIIPrintf(viewer," current info with %d keys. Ordered key-value pairs follow:\n",nkeys)); 7805b0d146aSStefano Zampini for (k = 0; k < nkeys; k++) { 7815b0d146aSStefano Zampini PetscMPIInt flag; 7825b0d146aSStefano Zampini 7835f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Info_get_nthkey(w->info,k,key)); 7845f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Info_get(w->info,key,MPI_MAX_INFO_VAL,value,&flag)); 785*28b400f6SJacob Faibussowitsch PetscCheck(flag,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Missing key %s",key); 7865f80ce2aSJacob Faibussowitsch CHKERRQ(PetscViewerASCIIPrintf(viewer," %s = %s\n",key,value)); 7875b0d146aSStefano Zampini } 7885b0d146aSStefano Zampini } else { 7895f80ce2aSJacob Faibussowitsch CHKERRQ(PetscViewerASCIIPrintf(viewer," current info=MPI_INFO_NULL\n")); 7905b0d146aSStefano Zampini } 7915b0d146aSStefano Zampini } 79295fce210SBarry Smith } 79395fce210SBarry Smith PetscFunctionReturn(0); 79495fce210SBarry Smith } 79595fce210SBarry Smith 79695fce210SBarry Smith static PetscErrorCode PetscSFDuplicate_Window(PetscSF sf,PetscSFDuplicateOption opt,PetscSF newsf) 79795fce210SBarry Smith { 79895fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 79995fce210SBarry Smith PetscSFWindowSyncType synctype; 80095fce210SBarry Smith 80195fce210SBarry Smith PetscFunctionBegin; 80295fce210SBarry Smith synctype = w->sync; 80395fce210SBarry Smith /* HACK: Must use FENCE or LOCK when called from PetscSFGetGroups() because ACTIVE here would cause recursion. */ 8045b0d146aSStefano Zampini if (!sf->setupcalled) synctype = PETSCSF_WINDOW_SYNC_LOCK; 8055f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFWindowSetSyncType(newsf,synctype)); 8065f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFWindowSetFlavorType(newsf,w->flavor)); 8075f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFWindowSetInfo(newsf,w->info)); 80895fce210SBarry Smith PetscFunctionReturn(0); 80995fce210SBarry Smith } 81095fce210SBarry Smith 811ad227feaSJunchao Zhang static PetscErrorCode PetscSFBcastBegin_Window(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,const void *rootdata,PetscMemType leafmtype,void *leafdata,MPI_Op op) 81295fce210SBarry Smith { 81395fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 81495fce210SBarry Smith PetscInt i,nranks; 81595fce210SBarry Smith const PetscMPIInt *ranks; 8165b0d146aSStefano Zampini const MPI_Aint *target_disp; 81795fce210SBarry Smith const MPI_Datatype *mine,*remote; 818684a874aSStefano Zampini MPI_Request *reqs; 81995fce210SBarry Smith MPI_Win win; 82095fce210SBarry Smith 82195fce210SBarry Smith PetscFunctionBegin; 8222c71b3e2SJacob Faibussowitsch PetscCheckFalse(op != MPI_REPLACE,PetscObjectComm((PetscObject)sf), PETSC_ERR_SUP, "PetscSFBcastBegin_Window with op!=MPI_REPLACE has not been implemented"); 8235f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetRootRanks(sf,&nranks,&ranks,NULL,NULL,NULL)); 8245f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFWindowGetDataTypes(sf,unit,&mine,&remote)); 8255f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetWindow(sf,unit,(void*)rootdata,w->sync,PETSC_TRUE,MPI_MODE_NOPUT|MPI_MODE_NOPRECEDE,MPI_MODE_NOPUT,0,&target_disp,&reqs,&win)); 82695fce210SBarry Smith for (i=0; i<nranks; i++) { 8275b0d146aSStefano Zampini MPI_Aint tdp = target_disp ? target_disp[i] : 0; 8285b0d146aSStefano Zampini 829684a874aSStefano Zampini if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) { 8305f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_lock(MPI_LOCK_SHARED,ranks[i],MPI_MODE_NOCHECK,win)); 831684a874aSStefano Zampini #if defined(PETSC_HAVE_MPI_RGET) 8325f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Rget(leafdata,1,mine[i],ranks[i],tdp,1,remote[i],win,&reqs[i])); 833684a874aSStefano Zampini #else 8345f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Get(leafdata,1,mine[i],ranks[i],tdp,1,remote[i],win)); 835684a874aSStefano Zampini #endif 836684a874aSStefano Zampini } else { 8375f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Get(leafdata,1,mine[i],ranks[i],tdp,1,remote[i],win)); 838684a874aSStefano Zampini } 83995fce210SBarry Smith } 84095fce210SBarry Smith PetscFunctionReturn(0); 84195fce210SBarry Smith } 84295fce210SBarry Smith 843ad227feaSJunchao Zhang PetscErrorCode PetscSFBcastEnd_Window(PetscSF sf,MPI_Datatype unit,const void *rootdata,void *leafdata,MPI_Op op) 84495fce210SBarry Smith { 8455b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 84695fce210SBarry Smith MPI_Win win; 8474b9acda6SJunchao Zhang MPI_Request *reqs = NULL; 84895fce210SBarry Smith 84995fce210SBarry Smith PetscFunctionBegin; 8505f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFFindWindow(sf,unit,rootdata,&win,&reqs)); 8515f80ce2aSJacob Faibussowitsch if (reqs) CHKERRMPI(MPI_Waitall(sf->nranks,reqs,MPI_STATUSES_IGNORE)); 852684a874aSStefano Zampini if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) { 853684a874aSStefano Zampini PetscInt i,nranks; 854684a874aSStefano Zampini const PetscMPIInt *ranks; 855684a874aSStefano Zampini 8565f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetRootRanks(sf,&nranks,&ranks,NULL,NULL,NULL)); 857684a874aSStefano Zampini for (i=0; i<nranks; i++) { 8585f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_unlock(ranks[i],win)); 859684a874aSStefano Zampini } 860684a874aSStefano Zampini } 8615f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFRestoreWindow(sf,unit,(void*)rootdata,w->sync,PETSC_TRUE,MPI_MODE_NOSTORE|MPI_MODE_NOSUCCEED,PETSC_FALSE,&win)); 86295fce210SBarry Smith PetscFunctionReturn(0); 86395fce210SBarry Smith } 86495fce210SBarry Smith 865eb02082bSJunchao Zhang PetscErrorCode PetscSFReduceBegin_Window(PetscSF sf,MPI_Datatype unit,PetscMemType leafmtype,const void *leafdata,PetscMemType rootmtype,void *rootdata,MPI_Op op) 86695fce210SBarry Smith { 86795fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 86895fce210SBarry Smith PetscInt i,nranks; 86995fce210SBarry Smith const PetscMPIInt *ranks; 8705b0d146aSStefano Zampini const MPI_Aint *target_disp; 87195fce210SBarry Smith const MPI_Datatype *mine,*remote; 87295fce210SBarry Smith MPI_Win win; 87395fce210SBarry Smith 87495fce210SBarry Smith PetscFunctionBegin; 8755f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetRootRanks(sf,&nranks,&ranks,NULL,NULL,NULL)); 8765f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFWindowGetDataTypes(sf,unit,&mine,&remote)); 8775f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFWindowOpTranslate(&op)); 8785f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetWindow(sf,unit,rootdata,w->sync,PETSC_TRUE,MPI_MODE_NOPRECEDE,0,0,&target_disp,NULL,&win)); 87995fce210SBarry Smith for (i=0; i<nranks; i++) { 8805b0d146aSStefano Zampini MPI_Aint tdp = target_disp ? target_disp[i] : 0; 8815b0d146aSStefano Zampini 8825f80ce2aSJacob Faibussowitsch if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) CHKERRMPI(MPI_Win_lock(MPI_LOCK_SHARED,ranks[i],MPI_MODE_NOCHECK,win)); 8835f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Accumulate((void*)leafdata,1,mine[i],ranks[i],tdp,1,remote[i],op,win)); 8845f80ce2aSJacob Faibussowitsch if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) CHKERRMPI(MPI_Win_unlock(ranks[i],win)); 88595fce210SBarry Smith } 88695fce210SBarry Smith PetscFunctionReturn(0); 88795fce210SBarry Smith } 88895fce210SBarry Smith 88900816365SJunchao Zhang static PetscErrorCode PetscSFReduceEnd_Window(PetscSF sf,MPI_Datatype unit,const void *leafdata,void *rootdata,MPI_Op op) 89095fce210SBarry Smith { 89195fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 89295fce210SBarry Smith MPI_Win win; 8934b9acda6SJunchao Zhang MPI_Request *reqs = NULL; 89495fce210SBarry Smith 89595fce210SBarry Smith PetscFunctionBegin; 8965f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFFindWindow(sf,unit,rootdata,&win,&reqs)); 8975f80ce2aSJacob Faibussowitsch if (reqs) CHKERRMPI(MPI_Waitall(sf->nranks,reqs,MPI_STATUSES_IGNORE)); 8985f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFRestoreWindow(sf,unit,rootdata,w->sync,PETSC_TRUE,MPI_MODE_NOSUCCEED,PETSC_TRUE,&win)); 89995fce210SBarry Smith PetscFunctionReturn(0); 90095fce210SBarry Smith } 9015b0d146aSStefano Zampini 902eb02082bSJunchao Zhang static PetscErrorCode PetscSFFetchAndOpBegin_Window(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,void *rootdata,PetscMemType leafmtype,const void *leafdata,void *leafupdate,MPI_Op op) 90395fce210SBarry Smith { 90495fce210SBarry Smith PetscInt i,nranks; 90595fce210SBarry Smith const PetscMPIInt *ranks; 90695fce210SBarry Smith const MPI_Datatype *mine,*remote; 9075b0d146aSStefano Zampini const MPI_Aint *target_disp; 90895fce210SBarry Smith MPI_Win win; 9095b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 9105b0d146aSStefano Zampini #if !defined(PETSC_HAVE_MPI_GET_ACCUMULATE) 9115b0d146aSStefano Zampini PetscSFWindowFlavorType oldf; 9125b0d146aSStefano Zampini #endif 91395fce210SBarry Smith 91495fce210SBarry Smith PetscFunctionBegin; 9155f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetRootRanks(sf,&nranks,&ranks,NULL,NULL,NULL)); 9165f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFWindowGetDataTypes(sf,unit,&mine,&remote)); 9175f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFWindowOpTranslate(&op)); 9185b0d146aSStefano Zampini #if !defined(PETSC_HAVE_MPI_GET_ACCUMULATE) 9195b0d146aSStefano Zampini /* FetchAndOp without MPI_Get_Accumulate requires locking. 9205b0d146aSStefano Zampini we create a new window every time to not interfere with user-defined MPI_Info which may have used "no_locks"="true" */ 9215b0d146aSStefano Zampini oldf = w->flavor; 9225b0d146aSStefano Zampini w->flavor = PETSCSF_WINDOW_FLAVOR_CREATE; 9235f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetWindow(sf,unit,rootdata,PETSCSF_WINDOW_SYNC_LOCK,PETSC_FALSE,0,0,0,&target_disp,NULL,&win)); 9245b0d146aSStefano Zampini #else 9255f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFGetWindow(sf,unit,rootdata,w->sync,PETSC_TRUE,MPI_MODE_NOPRECEDE,0,0,&target_disp,NULL,&win)); 9265b0d146aSStefano Zampini #endif 9275b0d146aSStefano Zampini for (i=0; i<nranks; i++) { 9285b0d146aSStefano Zampini MPI_Aint tdp = target_disp ? target_disp[i] : 0; 9295b0d146aSStefano Zampini 9305b0d146aSStefano Zampini #if !defined(PETSC_HAVE_MPI_GET_ACCUMULATE) 9315f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_lock(MPI_LOCK_EXCLUSIVE,ranks[i],0,win)); 9325f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Get(leafupdate,1,mine[i],ranks[i],tdp,1,remote[i],win)); 9335f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Accumulate((void*)leafdata,1,mine[i],ranks[i],tdp,1,remote[i],op,win)); 9345f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Win_unlock(ranks[i],win)); 9355b0d146aSStefano Zampini #else 9365f80ce2aSJacob Faibussowitsch if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) CHKERRMPI(MPI_Win_lock(MPI_LOCK_SHARED,ranks[i],0,win)); 9375f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Get_accumulate((void*)leafdata,1,mine[i],leafupdate,1,mine[i],ranks[i],tdp,1,remote[i],op,win)); 9385f80ce2aSJacob Faibussowitsch if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) CHKERRMPI(MPI_Win_unlock(ranks[i],win)); 9395b0d146aSStefano Zampini #endif 9405b0d146aSStefano Zampini } 9415b0d146aSStefano Zampini #if !defined(PETSC_HAVE_MPI_GET_ACCUMULATE) 9425b0d146aSStefano Zampini w->flavor = oldf; 9435b0d146aSStefano Zampini #endif 94495fce210SBarry Smith PetscFunctionReturn(0); 94595fce210SBarry Smith } 94695fce210SBarry Smith 94700816365SJunchao Zhang static PetscErrorCode PetscSFFetchAndOpEnd_Window(PetscSF sf,MPI_Datatype unit,void *rootdata,const void *leafdata,void *leafupdate,MPI_Op op) 94895fce210SBarry Smith { 94995fce210SBarry Smith MPI_Win win; 9505b0d146aSStefano Zampini #if defined(PETSC_HAVE_MPI_GET_ACCUMULATE) 9515b0d146aSStefano Zampini PetscSF_Window *w = (PetscSF_Window*)sf->data; 9525b0d146aSStefano Zampini #endif 9534b9acda6SJunchao Zhang MPI_Request *reqs = NULL; 95495fce210SBarry Smith 95595fce210SBarry Smith PetscFunctionBegin; 9565f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFFindWindow(sf,unit,rootdata,&win,&reqs)); 9575f80ce2aSJacob Faibussowitsch if (reqs) CHKERRMPI(MPI_Waitall(sf->nranks,reqs,MPI_STATUSES_IGNORE)); 9585b0d146aSStefano Zampini #if defined(PETSC_HAVE_MPI_GET_ACCUMULATE) 9595f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFRestoreWindow(sf,unit,rootdata,w->sync,PETSC_TRUE,MPI_MODE_NOSUCCEED,PETSC_TRUE,&win)); 9605b0d146aSStefano Zampini #else 9615f80ce2aSJacob Faibussowitsch CHKERRQ(PetscSFRestoreWindow(sf,unit,rootdata,PETSCSF_WINDOW_SYNC_LOCK,PETSC_FALSE,0,PETSC_TRUE,&win)); 9625b0d146aSStefano Zampini #endif 96395fce210SBarry Smith PetscFunctionReturn(0); 96495fce210SBarry Smith } 96595fce210SBarry Smith 966dec1416fSJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFCreate_Window(PetscSF sf) 96795fce210SBarry Smith { 96895fce210SBarry Smith PetscSF_Window *w = (PetscSF_Window*)sf->data; 96995fce210SBarry Smith 97095fce210SBarry Smith PetscFunctionBegin; 97195fce210SBarry Smith sf->ops->SetUp = PetscSFSetUp_Window; 97295fce210SBarry Smith sf->ops->SetFromOptions = PetscSFSetFromOptions_Window; 97395fce210SBarry Smith sf->ops->Reset = PetscSFReset_Window; 97495fce210SBarry Smith sf->ops->Destroy = PetscSFDestroy_Window; 97595fce210SBarry Smith sf->ops->View = PetscSFView_Window; 97695fce210SBarry Smith sf->ops->Duplicate = PetscSFDuplicate_Window; 977ad227feaSJunchao Zhang sf->ops->BcastBegin = PetscSFBcastBegin_Window; 978ad227feaSJunchao Zhang sf->ops->BcastEnd = PetscSFBcastEnd_Window; 97995fce210SBarry Smith sf->ops->ReduceBegin = PetscSFReduceBegin_Window; 98095fce210SBarry Smith sf->ops->ReduceEnd = PetscSFReduceEnd_Window; 98195fce210SBarry Smith sf->ops->FetchAndOpBegin = PetscSFFetchAndOpBegin_Window; 98295fce210SBarry Smith sf->ops->FetchAndOpEnd = PetscSFFetchAndOpEnd_Window; 98395fce210SBarry Smith 9845f80ce2aSJacob Faibussowitsch CHKERRQ(PetscNewLog(sf,&w)); 98595fce210SBarry Smith sf->data = (void*)w; 98695fce210SBarry Smith w->sync = PETSCSF_WINDOW_SYNC_FENCE; 9875b0d146aSStefano Zampini w->flavor = PETSCSF_WINDOW_FLAVOR_CREATE; 9885b0d146aSStefano Zampini w->info = MPI_INFO_NULL; 98995fce210SBarry Smith 9905f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowSetSyncType_C",PetscSFWindowSetSyncType_Window)); 9915f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowGetSyncType_C",PetscSFWindowGetSyncType_Window)); 9925f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowSetFlavorType_C",PetscSFWindowSetFlavorType_Window)); 9935f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowGetFlavorType_C",PetscSFWindowGetFlavorType_Window)); 9945f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowSetInfo_C",PetscSFWindowSetInfo_Window)); 9955f80ce2aSJacob Faibussowitsch CHKERRQ(PetscObjectComposeFunction((PetscObject)sf,"PetscSFWindowGetInfo_C",PetscSFWindowGetInfo_Window)); 99695fce210SBarry Smith 99795fce210SBarry Smith #if defined(OMPI_MAJOR_VERSION) && (OMPI_MAJOR_VERSION < 1 || (OMPI_MAJOR_VERSION == 1 && OMPI_MINOR_VERSION <= 6)) 99895fce210SBarry Smith { 99995fce210SBarry Smith PetscBool ackbug = PETSC_FALSE; 10005f80ce2aSJacob Faibussowitsch CHKERRQ(PetscOptionsGetBool(NULL,NULL,"-acknowledge_ompi_onesided_bug",&ackbug,NULL)); 100195fce210SBarry Smith if (ackbug) { 10025f80ce2aSJacob Faibussowitsch CHKERRQ(PetscInfo(sf,"Acknowledged Open MPI bug, proceeding anyway. Expect memory corruption.\n")); 100395fce210SBarry 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"); 100495fce210SBarry Smith } 100595fce210SBarry Smith #endif 100695fce210SBarry Smith PetscFunctionReturn(0); 100795fce210SBarry Smith } 1008