Actual source code: sfimpl.h

petsc-master 2020-11-24
Report Typos and Errors
  1: #if !defined(PETSCSFIMPL_H)
  2: #define PETSCSFIMPL_H

  4: #include <petscvec.h>
  5: #include <petscsf.h>
  6: #include <petsc/private/petscimpl.h>
  7: #include <petscviewer.h>

  9: #if defined(PETSC_HAVE_CUDA)
 10:   #include <cuda_runtime.h>
 11: #endif

 13: #if defined(PETSC_HAVE_HIP)
 14:   #include <hip/hip_runtime.h>
 15: #endif

 17: PETSC_EXTERN PetscLogEvent PETSCSF_SetGraph;
 18: PETSC_EXTERN PetscLogEvent PETSCSF_SetUp;
 19: PETSC_EXTERN PetscLogEvent PETSCSF_BcastBegin;
 20: PETSC_EXTERN PetscLogEvent PETSCSF_BcastEnd;
 21: PETSC_EXTERN PetscLogEvent PETSCSF_BcastAndOpBegin;
 22: PETSC_EXTERN PetscLogEvent PETSCSF_BcastAndOpEnd;
 23: PETSC_EXTERN PetscLogEvent PETSCSF_ReduceBegin;
 24: PETSC_EXTERN PetscLogEvent PETSCSF_ReduceEnd;
 25: PETSC_EXTERN PetscLogEvent PETSCSF_FetchAndOpBegin;
 26: PETSC_EXTERN PetscLogEvent PETSCSF_FetchAndOpEnd;
 27: PETSC_EXTERN PetscLogEvent PETSCSF_EmbedSF;
 28: PETSC_EXTERN PetscLogEvent PETSCSF_DistSect;
 29: PETSC_EXTERN PetscLogEvent PETSCSF_SectSF;
 30: PETSC_EXTERN PetscLogEvent PETSCSF_RemoteOff;
 31: PETSC_EXTERN PetscLogEvent PETSCSF_Pack;
 32: PETSC_EXTERN PetscLogEvent PETSCSF_Unpack;

 34: typedef enum {PETSCSF_../../..2LEAF=0, PETSCSF_LEAF2../../..} PetscSFDirection;
 35: typedef enum {PETSCSF_BCAST=0, PETSCSF_REDUCE, PETSCSF_FETCH} PetscSFOperation;
 36: /* When doing device-aware MPI, a backend refers to the SF/device interface */
 37: typedef enum {PETSCSF_BACKEND_INVALID=0,PETSCSF_BACKEND_CUDA,PETSCSF_BACKEND_KOKKOS} PetscSFBackend;

 39: struct _PetscSFOps {
 40:   PetscErrorCode (*Reset)(PetscSF);
 41:   PetscErrorCode (*Destroy)(PetscSF);
 42:   PetscErrorCode (*SetUp)(PetscSF);
 43:   PetscErrorCode (*SetFromOptions)(PetscOptionItems*,PetscSF);
 44:   PetscErrorCode (*View)(PetscSF,PetscViewer);
 45:   PetscErrorCode (*Duplicate)(PetscSF,PetscSFDuplicateOption,PetscSF);
 46:   PetscErrorCode (*BcastAndOpBegin)(PetscSF,MPI_Datatype,PetscMemType,const void*,PetscMemType,void*,MPI_Op);
 47:   PetscErrorCode (*BcastAndOpEnd)  (PetscSF,MPI_Datatype,const void*,void*,MPI_Op);
 48:   PetscErrorCode (*ReduceBegin)    (PetscSF,MPI_Datatype,PetscMemType,const void*,PetscMemType,void*,MPI_Op);
 49:   PetscErrorCode (*ReduceEnd)      (PetscSF,MPI_Datatype,const void*,void*,MPI_Op);
 50:   PetscErrorCode (*FetchAndOpBegin)(PetscSF,MPI_Datatype,PetscMemType,void*,PetscMemType,const void*,void*,MPI_Op);
 51:   PetscErrorCode (*FetchAndOpEnd)  (PetscSF,MPI_Datatype,void*,const void*,void*,MPI_Op);
 52:   PetscErrorCode (*BcastToZero)    (PetscSF,MPI_Datatype,PetscMemType,const void*,PetscMemType,      void*); /* For interal use only */
 53:   PetscErrorCode (*GetRootRanks)(PetscSF,PetscInt*,const PetscMPIInt**,const PetscInt**,const PetscInt**,const PetscInt**);
 54:   PetscErrorCode (*GetLeafRanks)(PetscSF,PetscInt*,const PetscMPIInt**,const PetscInt**,const PetscInt**);
 55:   PetscErrorCode (*CreateLocalSF)(PetscSF,PetscSF*);
 56:   PetscErrorCode (*GetGraph)(PetscSF,PetscInt*,PetscInt*,const PetscInt**,const PetscSFNode**);
 57:   PetscErrorCode (*CreateEmbeddedSF)(PetscSF,PetscInt,const PetscInt*,PetscSF*);
 58:   PetscErrorCode (*CreateEmbeddedLeafSF)(PetscSF,PetscInt,const PetscInt*,PetscSF*);

 60:   PetscErrorCode (*Malloc)(PetscMemType,size_t,void**);
 61:   PetscErrorCode (*Free)(PetscMemType,void*);
 62: };

 64: typedef struct _n_PetscSFPackOpt *PetscSFPackOpt;

 66: struct _p_PetscSF {
 67:   PETSCHEADER(struct _PetscSFOps);
 68:   struct { /* Fields needed to implement VecScatter behavior */
 69:     PetscInt          from_n,to_n;   /* Recorded local sizes of the input from/to vectors in VecScatterCreate(). Used subsequently for error checking. */
 70:     PetscBool         beginandendtogether;  /* Indicates that the scatter begin and end  function are called together, VecScatterEnd() is then treated as a nop */
 71:     PetscBool         packongpu;     /* For GPU vectors, pack needed entries on GPU instead of pulling the whole vector down to CPU and then packing on CPU */
 72:     const PetscScalar *xdata;        /* Vector data to read from */
 73:     PetscScalar       *ydata;        /* Vector data to write to. The two pointers are recorded in VecScatterBegin. Memory is not managed by SF. */
 74:     PetscSF           lsf;           /* The local part of the scatter, used in SCATTER_LOCAL. Built on demand. */
 75:     PetscInt          bs;            /* Block size, determined by IS passed to VecScatterCreate */
 76:     MPI_Datatype      unit;          /* one unit = bs PetscScalars */
 77:     PetscBool         logging;       /* Indicate if vscat log events are happening. If yes, avoid duplicated SF logging to have clear -log_view */
 78:   } vscat;

 80:   /* Fields for generic PetscSF functionality */
 81:   PetscInt        nroots;          /* Number of root vertices on current process (candidates for incoming edges) */
 82:   PetscInt        nleaves;         /* Number of leaf vertices on current process (this process specifies a root for each leaf) */
 83:   PetscInt        *mine;           /* Location of leaves in leafdata arrays provided to the communication routines */
 84:   PetscInt        *mine_alloc;
 85:   PetscInt        minleaf,maxleaf;
 86:   PetscSFNode     *remote;         /* Remote references to roots for each local leaf */
 87:   PetscSFNode     *remote_alloc;
 88:   PetscInt        nranks;          /* Number of ranks owning roots connected to my leaves */
 89:   PetscInt        ndranks;         /* Number of ranks in distinguished group holding roots connected to my leaves */
 90:   PetscMPIInt     *ranks;          /* List of ranks referenced by "remote" */
 91:   PetscInt        *roffset;        /* Array of length nranks+1, offset in rmine/rremote for each rank */
 92:   PetscInt        *rmine;          /* Concatenated array holding local indices referencing each remote rank */
 93:   PetscInt        *rmine_d[2];     /* A copy of rmine[local/remote] in device memory if needed */

 95:   /* Some results useful in packing by analyzing rmine[] */
 96:   PetscInt        leafbuflen[2];   /* Length (in unit) of leaf buffers, in layout of [PETSCSF_LOCAL/REMOTE] */
 97:   PetscBool       leafcontig[2];   /* True means indices in rmine[self part] or rmine[remote part] are contiguous, and they start from ... */
 98:   PetscInt        leafstart[2];    /* ... leafstart[0] and leafstart[1] respectively */
 99:   PetscSFPackOpt  leafpackopt[2];  /* Optimization plans to (un)pack leaves connected to remote roots, based on index patterns in rmine[]. NULL for no optimization */
100:   PetscSFPackOpt  leafpackopt_d[2];/* Copy of leafpackopt_d[] on device if needed */
101:   PetscBool       leafdups[2];     /* Indices in rmine[] for self(0)/remote(1) communication have dups? TRUE implies theads working on them in parallel may have data race. */

103:   PetscInt        nleafreqs;       /* Number of MPI reqests for leaves */
104:   PetscInt        *rremote;        /* Concatenated array holding remote indices referenced for each remote rank */
105:   PetscBool       degreeknown;     /* The degree is currently known, do not have to recompute */
106:   PetscInt        *degree;         /* Degree of each of my root vertices */
107:   PetscInt        *degreetmp;      /* Temporary local array for computing degree */
108:   PetscBool       rankorder;       /* Sort ranks for gather and scatter operations */
109:   MPI_Group       ingroup;         /* Group of processes connected to my roots */
110:   MPI_Group       outgroup;        /* Group of processes connected to my leaves */
111:   PetscSF         multi;           /* Internal graph used to implement gather and scatter operations */
112:   PetscBool       graphset;        /* Flag indicating that the graph has been set, required before calling communication routines */
113:   PetscBool       setupcalled;     /* Type and communication structures have been set up */
114:   PetscSFPattern  pattern;         /* Pattern of the graph */
115:   PetscBool       persistent;      /* Does this SF use MPI persistent requests for communication */
116:   PetscLayout     map;             /* Layout of leaves over all processes when building a patterned graph */
117:   PetscBool       use_default_stream;  /* If true, SF assumes root/leafdata is on the default stream upon input and will also leave them there upon output */
118:   PetscBool       use_gpu_aware_mpi;   /* If true, SF assumes it can pass GPU pointers to MPI */
119:   PetscBool       use_stream_aware_mpi;/* If true, SF assumes the underlying MPI is cuda-stream aware and we won't sync streams for send/recv buffers passed to MPI */
120: #if defined(PETSC_HAVE_CUDA)
121:   PetscInt        maxResidentThreadsPerGPU;
122: #endif
123:   PetscSFBackend  backend;         /* The device backend (if any) SF will use */
124:   void *data;                      /* Pointer to implementation */
125: };

127: PETSC_EXTERN PetscBool PetscSFRegisterAllCalled;
128: PETSC_EXTERN PetscErrorCode PetscSFRegisterAll(void);

130: PETSC_INTERN PetscErrorCode PetscSFCreateLocalSF_Private(PetscSF,PetscSF*);
131: PETSC_INTERN PetscErrorCode PetscSFBcastToZero_Private(PetscSF,MPI_Datatype,const void*,void*);

133: PETSC_EXTERN PetscErrorCode MPIPetsc_Type_unwrap(MPI_Datatype,MPI_Datatype*,PetscBool*);
134: PETSC_EXTERN PetscErrorCode MPIPetsc_Type_compare(MPI_Datatype,MPI_Datatype,PetscBool*);
135: PETSC_EXTERN PetscErrorCode MPIPetsc_Type_compare_contig(MPI_Datatype,MPI_Datatype,PetscInt*);

137: #if defined(PETSC_HAVE_MPI_NONBLOCKING_COLLECTIVES)
138: #define MPIU_Iscatter(a,b,c,d,e,f,g,h,req)     MPI_Iscatter(a,b,c,d,e,f,g,h,req)
139: #define MPIU_Iscatterv(a,b,c,d,e,f,g,h,i,req)  MPI_Iscatterv(a,b,c,d,e,f,g,h,i,req)
140: #define MPIU_Igather(a,b,c,d,e,f,g,h,req)      MPI_Igather(a,b,c,d,e,f,g,h,req)
141: #define MPIU_Igatherv(a,b,c,d,e,f,g,h,i,req)   MPI_Igatherv(a,b,c,d,e,f,g,h,i,req)
142: #define MPIU_Iallgather(a,b,c,d,e,f,g,req)     MPI_Iallgather(a,b,c,d,e,f,g,req)
143: #define MPIU_Iallgatherv(a,b,c,d,e,f,g,h,req)  MPI_Iallgatherv(a,b,c,d,e,f,g,h,req)
144: #define MPIU_Ialltoall(a,b,c,d,e,f,g,req)      MPI_Ialltoall(a,b,c,d,e,f,g,req)
145: #else
146: /* Ignore req, the MPI_Request argument, and use MPI blocking collectives. One should initialize req
147:    to MPI_REQUEST_NULL so that one can do MPI_Wait(req,status) no matter the call is blocking or not.
148:  */
149: #define MPIU_Iscatter(a,b,c,d,e,f,g,h,req)     MPI_Scatter(a,b,c,d,e,f,g,h)
150: #define MPIU_Iscatterv(a,b,c,d,e,f,g,h,i,req)  MPI_Scatterv(a,b,c,d,e,f,g,h,i)
151: #define MPIU_Igather(a,b,c,d,e,f,g,h,req)      MPI_Gather(a,b,c,d,e,f,g,h)
152: #define MPIU_Igatherv(a,b,c,d,e,f,g,h,i,req)   MPI_Gatherv(a,b,c,d,e,f,g,h,i)
153: #define MPIU_Iallgather(a,b,c,d,e,f,g,req)     MPI_Allgather(a,b,c,d,e,f,g)
154: #define MPIU_Iallgatherv(a,b,c,d,e,f,g,h,req)  MPI_Allgatherv(a,b,c,d,e,f,g,h)
155: #define MPIU_Ialltoall(a,b,c,d,e,f,g,req)      MPI_Alltoall(a,b,c,d,e,f,g)
156: #endif

158: PETSC_EXTERN PetscErrorCode VecScatterGetRemoteCount_Private(VecScatter,PetscBool,PetscInt*,PetscInt*);
159: PETSC_EXTERN PetscErrorCode VecScatterGetRemote_Private(VecScatter,PetscBool,PetscInt*,const PetscInt**,const PetscInt**,const PetscMPIInt**,PetscInt*);
160: PETSC_EXTERN PetscErrorCode VecScatterGetRemoteOrdered_Private(VecScatter,PetscBool,PetscInt*,const PetscInt**,const PetscInt**,const PetscMPIInt**,PetscInt*);
161: PETSC_EXTERN PetscErrorCode VecScatterRestoreRemote_Private(VecScatter,PetscBool,PetscInt*,const PetscInt**,const PetscInt**,const PetscMPIInt**,PetscInt*);
162: PETSC_EXTERN PetscErrorCode VecScatterRestoreRemoteOrdered_Private(VecScatter,PetscBool,PetscInt*,const PetscInt**,const PetscInt**,const PetscMPIInt**,PetscInt*);

164: #if defined(PETSC_HAVE_CUDA)
165: PETSC_EXTERN PetscErrorCode PetscSFMalloc_Cuda(PetscMemType,size_t,void**);
166: PETSC_EXTERN PetscErrorCode PetscSFFree_Cuda(PetscMemType,void*);
167: #endif

169: #if defined(PETSC_HAVE_KOKKOS)
170: PETSC_EXTERN PetscErrorCode PetscSFMalloc_Kokkos(PetscMemType,size_t,void**);
171: PETSC_EXTERN PetscErrorCode PetscSFFree_Kokkos(PetscMemType,void*);
172: #endif

174: /* SF only supports CUDA and Kokkos devices. Even VIENNACL is a device, its device pointers are invisible to SF.
175:    Through VecGetArray(), we copy data of VECVIENNACL from device to host and pass host pointers to SF.
176:  */
177: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_KOKKOS)
178:   #define PetscSFMalloc(sf,mtype,sz,ptr)  ((*(sf)->ops->Malloc)(mtype,sz,ptr))
179:   /* Free memory and set ptr to NULL when succeeded */
180:   #define PetscSFFree(sf,mtype,ptr)       ((ptr) && ((*(sf)->ops->Free)(mtype,ptr) || ((ptr)=NULL,0)))
181: #else
182:   /* If pure host code, do with less indirection */
183:   #define PetscSFMalloc(sf,mtype,sz,ptr)  PetscMalloc(sz,ptr)
184:   #define PetscSFFree(sf,mtype,ptr)       PetscFree(ptr)
185: #endif

187: #endif