Actual source code: sfimpl.h

  1: #pragma once

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

  9: PETSC_EXTERN PetscLogEvent PETSCSF_SetGraph;
 10: PETSC_EXTERN PetscLogEvent PETSCSF_SetUp;
 11: PETSC_EXTERN PetscLogEvent PETSCSF_BcastBegin;
 12: PETSC_EXTERN PetscLogEvent PETSCSF_BcastEnd;
 13: PETSC_EXTERN PetscLogEvent PETSCSF_BcastBegin;
 14: PETSC_EXTERN PetscLogEvent PETSCSF_BcastEnd;
 15: PETSC_EXTERN PetscLogEvent PETSCSF_ReduceBegin;
 16: PETSC_EXTERN PetscLogEvent PETSCSF_ReduceEnd;
 17: PETSC_EXTERN PetscLogEvent PETSCSF_FetchAndOpBegin;
 18: PETSC_EXTERN PetscLogEvent PETSCSF_FetchAndOpEnd;
 19: PETSC_EXTERN PetscLogEvent PETSCSF_EmbedSF;
 20: PETSC_EXTERN PetscLogEvent PETSCSF_DistSect;
 21: PETSC_EXTERN PetscLogEvent PETSCSF_SectSF;
 22: PETSC_EXTERN PetscLogEvent PETSCSF_RemoteOff;
 23: PETSC_EXTERN PetscLogEvent PETSCSF_Pack;
 24: PETSC_EXTERN PetscLogEvent PETSCSF_Unpack;

 26: typedef enum {
 27:   PETSCSF_ROOT2LEAF = 0,
 28:   PETSCSF_LEAF2ROOT
 29: } PetscSFDirection;
 30: typedef enum {
 31:   PETSCSF_BCAST = 0,
 32:   PETSCSF_REDUCE,
 33:   PETSCSF_FETCH
 34: } PetscSFOperation;
 35: /* When doing device-aware MPI, a backend refers to the SF/device interface */
 36: typedef enum {
 37:   PETSCSF_BACKEND_INVALID = 0,
 38:   PETSCSF_BACKEND_CUDA,
 39:   PETSCSF_BACKEND_HIP,
 40:   PETSCSF_BACKEND_KOKKOS
 41: } PetscSFBackend;

 43: struct _PetscSFOps {
 44:   PetscErrorCode (*Reset)(PetscSF);
 45:   PetscErrorCode (*Destroy)(PetscSF);
 46:   PetscErrorCode (*SetUp)(PetscSF);
 47:   PetscErrorCode (*SetFromOptions)(PetscSF, PetscOptionItems *);
 48:   PetscErrorCode (*View)(PetscSF, PetscViewer);
 49:   PetscErrorCode (*Duplicate)(PetscSF, PetscSFDuplicateOption, PetscSF);
 50:   PetscErrorCode (*BcastBegin)(PetscSF, MPI_Datatype, PetscMemType, const void *, PetscMemType, void *, MPI_Op);
 51:   PetscErrorCode (*BcastEnd)(PetscSF, MPI_Datatype, const void *, void *, MPI_Op);
 52:   PetscErrorCode (*ReduceBegin)(PetscSF, MPI_Datatype, PetscMemType, const void *, PetscMemType, void *, MPI_Op);
 53:   PetscErrorCode (*ReduceEnd)(PetscSF, MPI_Datatype, const void *, void *, MPI_Op);
 54:   PetscErrorCode (*FetchAndOpBegin)(PetscSF, MPI_Datatype, PetscMemType, void *, PetscMemType, const void *, void *, MPI_Op);
 55:   PetscErrorCode (*FetchAndOpEnd)(PetscSF, MPI_Datatype, void *, const void *, void *, MPI_Op);
 56:   PetscErrorCode (*BcastToZero)(PetscSF, MPI_Datatype, PetscMemType, const void *, PetscMemType, void *); /* For internal use only */
 57:   PetscErrorCode (*GetRootRanks)(PetscSF, PetscInt *, const PetscMPIInt **, const PetscInt **, const PetscInt **, const PetscInt **);
 58:   PetscErrorCode (*GetLeafRanks)(PetscSF, PetscInt *, const PetscMPIInt **, const PetscInt **, const PetscInt **);
 59:   PetscErrorCode (*CreateLocalSF)(PetscSF, PetscSF *);
 60:   PetscErrorCode (*GetGraph)(PetscSF, PetscInt *, PetscInt *, const PetscInt **, const PetscSFNode **);
 61:   PetscErrorCode (*CreateEmbeddedRootSF)(PetscSF, PetscInt, const PetscInt *, PetscSF *);
 62:   PetscErrorCode (*CreateEmbeddedLeafSF)(PetscSF, PetscInt, const PetscInt *, PetscSF *);

 64:   PetscErrorCode (*Malloc)(PetscMemType, size_t, void **);
 65:   PetscErrorCode (*Free)(PetscMemType, void *);
 66: };

 68: typedef struct _n_PetscSFPackOpt *PetscSFPackOpt;

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

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

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

106:   PetscInt       nleafreqs;            /* Number of MPI requests for leaves */
107:   PetscInt      *rremote;              /* Concatenated array holding remote indices referenced for each remote rank */
108:   PetscBool      degreeknown;          /* The degree is currently known, do not have to recompute */
109:   PetscInt      *degree;               /* Degree of each of my root vertices */
110:   PetscInt      *degreetmp;            /* Temporary local array for computing degree */
111:   PetscBool      rankorder;            /* Sort ranks for gather and scatter operations */
112:   MPI_Group      ingroup;              /* Group of processes connected to my roots */
113:   MPI_Group      outgroup;             /* Group of processes connected to my leaves */
114:   PetscSF        multi;                /* Internal graph used to implement gather and scatter operations */
115:   PetscBool      graphset;             /* Flag indicating that the graph has been set, required before calling communication routines */
116:   PetscBool      setupcalled;          /* Type and communication structures have been set up */
117:   PetscSFPattern pattern;              /* Pattern of the graph */
118:   PetscBool      persistent;           /* Does this SF use MPI persistent requests for communication */
119:   PetscLayout    map;                  /* Layout of leaves over all processes when building a patterned graph */
120:   PetscBool      unknown_input_stream; /* If true, SF does not know which streams root/leafdata is on. Default is false, since we only use petsc default stream */
121:   PetscBool      use_gpu_aware_mpi;    /* If true, SF assumes it can pass GPU pointers to MPI */
122:   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 */
123:   PetscInt       maxResidentThreadsPerGPU;
124:   PetscBool      allow_multi_leaves;
125:   PetscSFBackend backend; /* The device backend (if any) SF will use */
126:   void          *data;    /* Pointer to implementation */

128: #if defined(PETSC_HAVE_NVSHMEM)
129:   PetscBool use_nvshmem;                 /* TRY to use nvshmem on cuda devices with this SF when possible */
130:   PetscBool use_nvshmem_get;             /* If true, use nvshmem_get based protocol, otherwise, use nvshmem_put based protocol */
131:   PetscBool checked_nvshmem_eligibility; /* Have we checked eligibility of using NVSHMEM on this sf? */
132:   PetscBool setup_nvshmem;               /* Have we already set up NVSHMEM related fields below? These fields are built on-demand */
133:   PetscInt  leafbuflen_rmax;             /* max leafbuflen[REMOTE] over comm */
134:   PetscInt  nRemoteRootRanks;            /* nranks - ndranks */
135:   PetscInt  nRemoteRootRanksMax;         /* max nranks-ndranks over comm */

137:   /* The following two fields look confusing but actually make sense: They are offsets of buffers at the remote side. We're doing one-sided communication! */
138:   PetscInt *rootsigdisp; /* [nRemoteRootRanks]. For my i-th remote root rank, I will access its rootsigdisp[i]-th root signal */
139:   PetscInt *rootbufdisp; /* [nRemoteRootRanks]. For my i-th remote root rank, I will access its root buf at offset rootbufdisp[i], in <unit> to be set */

141:   PetscInt    *rootbufdisp_d;
142:   PetscInt    *rootsigdisp_d; /* Copy of rootsigdisp[] on device */
143:   PetscMPIInt *ranks_d;       /* Copy of the remote part of (root) ranks[] on device */
144:   PetscInt    *roffset_d;     /* Copy of the remote part of roffset[] on device */
145: #endif
146: #if defined(PETSC_HAVE_MPIX_STREAM)
147:   MPIX_Stream mpi_stream;
148:   MPI_Comm    stream_comm; /* gpu stream aware MPI communicator */
149: #endif
150: };

152: PETSC_EXTERN PetscBool      PetscSFRegisterAllCalled;
153: PETSC_EXTERN PetscErrorCode PetscSFRegisterAll(void);

155: PETSC_INTERN PetscErrorCode PetscSFCreateLocalSF_Private(PetscSF, PetscSF *);
156: PETSC_INTERN PetscErrorCode PetscSFBcastToZero_Private(PetscSF, MPI_Datatype, const void *, void *) PETSC_ATTRIBUTE_MPI_POINTER_WITH_TYPE(3, 2) PETSC_ATTRIBUTE_MPI_POINTER_WITH_TYPE(4, 2);

158: PETSC_EXTERN PetscErrorCode MPIPetsc_Type_unwrap(MPI_Datatype, MPI_Datatype *, PetscBool *);
159: PETSC_EXTERN PetscErrorCode MPIPetsc_Type_compare(MPI_Datatype, MPI_Datatype, PetscBool *);
160: PETSC_EXTERN PetscErrorCode MPIPetsc_Type_compare_contig(MPI_Datatype, MPI_Datatype, PetscInt *);

162: #if defined(PETSC_HAVE_MPI_NONBLOCKING_COLLECTIVES)
163:   #define MPIU_Ibcast(a, b, c, d, e, req)                MPI_Ibcast(a, b, c, d, e, req)
164:   #define MPIU_Ireduce(a, b, c, d, e, f, g, req)         MPI_Ireduce(a, b, c, d, e, f, g, req)
165:   #define MPIU_Iscatter(a, b, c, d, e, f, g, h, req)     MPI_Iscatter(a, b, c, d, e, f, g, h, req)
166:   #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)
167:   #define MPIU_Igather(a, b, c, d, e, f, g, h, req)      MPI_Igather(a, b, c, d, e, f, g, h, req)
168:   #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)
169:   #define MPIU_Iallgather(a, b, c, d, e, f, g, req)      MPI_Iallgather(a, b, c, d, e, f, g, req)
170:   #define MPIU_Iallgatherv(a, b, c, d, e, f, g, h, req)  MPI_Iallgatherv(a, b, c, d, e, f, g, h, req)
171:   #define MPIU_Ialltoall(a, b, c, d, e, f, g, req)       MPI_Ialltoall(a, b, c, d, e, f, g, req)
172: #else
173:   /* Ignore req, the MPI_Request argument, and use MPI blocking collectives. One should initialize req
174:    to MPI_REQUEST_NULL so that one can do MPI_Wait(req,status) no matter the call is blocking or not.
175:  */
176:   #define MPIU_Ibcast(a, b, c, d, e, req)                MPI_Bcast(a, b, c, d, e)
177:   #define MPIU_Ireduce(a, b, c, d, e, f, g, req)         MPI_Reduce(a, b, c, d, e, f, g)
178:   #define MPIU_Iscatter(a, b, c, d, e, f, g, h, req)     MPI_Scatter(a, b, c, d, e, f, g, h)
179:   #define MPIU_Iscatterv(a, b, c, d, e, f, g, h, i, req) MPI_Scatterv(a, b, c, d, e, f, g, h, i)
180:   #define MPIU_Igather(a, b, c, d, e, f, g, h, req)      MPI_Gather(a, b, c, d, e, f, g, h)
181:   #define MPIU_Igatherv(a, b, c, d, e, f, g, h, i, req)  MPI_Gatherv(a, b, c, d, e, f, g, h, i)
182:   #define MPIU_Iallgather(a, b, c, d, e, f, g, req)      MPI_Allgather(a, b, c, d, e, f, g)
183:   #define MPIU_Iallgatherv(a, b, c, d, e, f, g, h, req)  MPI_Allgatherv(a, b, c, d, e, f, g, h)
184:   #define MPIU_Ialltoall(a, b, c, d, e, f, g, req)       MPI_Alltoall(a, b, c, d, e, f, g)
185: #endif

187: PETSC_EXTERN PetscErrorCode VecScatterGetRemoteCount_Private(VecScatter, PetscBool, PetscInt *, PetscInt *);
188: PETSC_EXTERN PetscErrorCode VecScatterGetRemote_Private(VecScatter, PetscBool, PetscInt *, const PetscInt **, const PetscInt **, const PetscMPIInt **, PetscInt *);
189: PETSC_EXTERN PetscErrorCode VecScatterGetRemoteOrdered_Private(VecScatter, PetscBool, PetscInt *, const PetscInt **, const PetscInt **, const PetscMPIInt **, PetscInt *);
190: PETSC_EXTERN PetscErrorCode VecScatterRestoreRemote_Private(VecScatter, PetscBool, PetscInt *, const PetscInt **, const PetscInt **, const PetscMPIInt **, PetscInt *);
191: PETSC_EXTERN PetscErrorCode VecScatterRestoreRemoteOrdered_Private(VecScatter, PetscBool, PetscInt *, const PetscInt **, const PetscInt **, const PetscMPIInt **, PetscInt *);

193: #if defined(PETSC_HAVE_CUDA)
194: PETSC_EXTERN PetscErrorCode PetscSFMalloc_CUDA(PetscMemType, size_t, void **);
195: PETSC_EXTERN PetscErrorCode PetscSFFree_CUDA(PetscMemType, void *);
196: #endif
197: #if defined(PETSC_HAVE_HIP)
198: PETSC_EXTERN PetscErrorCode PetscSFMalloc_HIP(PetscMemType, size_t, void **);
199: PETSC_EXTERN PetscErrorCode PetscSFFree_HIP(PetscMemType, void *);
200: #endif
201: #if defined(PETSC_HAVE_KOKKOS)
202: PETSC_EXTERN PetscErrorCode PetscSFMalloc_Kokkos(PetscMemType, size_t, void **);
203: PETSC_EXTERN PetscErrorCode PetscSFFree_Kokkos(PetscMemType, void *);
204: #endif

206: /* SF only supports CUDA and Kokkos devices. Even VIENNACL is a device, its device pointers are invisible to SF.
207:    Through VecGetArray(), we copy data of VECVIENNACL from device to host and pass host pointers to SF.
208:  */
209: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_KOKKOS) || defined(PETSC_HAVE_HIP)
210:   #define PetscSFMalloc(sf, mtype, sz, ptr) ((*(sf)->ops->Malloc)(mtype, sz, ptr))
211:   /* Free memory and set ptr to NULL when succeeded */
212:   #define PetscSFFree(sf, mtype, ptr) ((PetscErrorCode)((ptr) && ((*(sf)->ops->Free)(mtype, ptr) || ((ptr) = NULL, PETSC_SUCCESS))))
213: #else
214:   /* If pure host code, do with less indirection */
215:   #define PetscSFMalloc(sf, mtype, sz, ptr) PetscMalloc(sz, ptr)
216:   #define PetscSFFree(sf, mtype, ptr)       PetscFree(ptr)
217: #endif