1 #if !defined(__SFPACK_H) 2 #define __SFPACK_H 3 4 #include <../src/vec/is/sf/impls/basic/sfbasic.h> 5 #if defined(PETSC_HAVE_CUDA) 6 #include <petscdevice.h> 7 typedef cudaStream_t cupmStream_t; 8 typedef cudaEvent_t cupmEvent_t; 9 #endif 10 11 #if defined(PETSC_HAVE_HIP) 12 #include <petscdevice.h> 13 typedef hipStream_t cupmStream_t; 14 typedef hipEvent_t cupmEvent_t; 15 #endif 16 17 /* In terms of function overloading, long long int is a different type than int64_t, which PetscInt might be defined to. 18 We perfer long long int over PetscInt (int64_t), since CUDA atomics are built around (unsigned) long long int. 19 */ 20 typedef long long int llint; 21 typedef unsigned long long int ullint; 22 23 /* We separate SF communications for SFBasic and SFNeighbor in two parts: local (self,intra-rank) and remote (inter-rank) */ 24 typedef enum {PETSCSF_LOCAL=0, PETSCSF_REMOTE} PetscSFScope; 25 26 /* Optimizations in packing & unpacking for destination ranks. 27 28 Suppose there are m indices stored in idx[], and two addresses u, p. We want to do packing: 29 p[i] = u[idx[i]], for i in [0,m) 30 31 Indices are associated with n ranks and each rank's indices are stored consecutively in idx[]. 32 We go through indices for each rank and see if they are indices of a 3D submatrix of size [dx,dy,dz] in 33 a parent matrix of size [X,Y,Z], with the submatrix's first index being <start>. 34 35 E.g., for indices 1,2,3, 6,7,8, 11,12,13, the submatrix size is [3,3,1] with start=1, and the parent matrix's size 36 is [5,3,1]. For simplicity, if any destination rank does not have this pattern, we give up the optimization. 37 38 Note before using this per-rank optimization, one should check leafcontig[], rootcontig[], which say 39 indices in whole are contiguous, and therefore much more useful than this one when true. 40 */ 41 struct _n_PetscSFPackOpt { 42 PetscInt *array; /* [7*n+2] Memory pool for other fields in this struct. Used to easily copy this struct to GPU */ 43 PetscInt n; /* Number of destination ranks */ 44 PetscInt *offset; /* [n+1] Offsets of indices for each rank. offset[0]=0, offset[i+1]=offset[i]+dx[i]*dy[i]*dz[i] */ 45 PetscInt *start; /* [n] First index */ 46 PetscInt *dx,*dy,*dz; /* [n] Lengths of the submatrix in X, Y, Z dimension. */ 47 PetscInt *X,*Y; /* [n] Lengths of the outer matrix in X, Y. We do not care Z. */ 48 }; 49 50 /* An abstract class that defines a communication link, which includes how to pack/unpack data and send/recv buffers 51 */ 52 struct _n_PetscSFLink { 53 PetscErrorCode (*Memcpy) (PetscSFLink,PetscMemType,void*,PetscMemType,const void*,size_t); /* Async device memcopy might use stream in the link */ 54 PetscErrorCode (*PrePack) (PetscSF,PetscSFLink,PetscSFDirection); 55 PetscErrorCode (*PostUnpack) (PetscSF,PetscSFLink,PetscSFDirection); 56 PetscErrorCode (*StartCommunication) (PetscSF,PetscSFLink,PetscSFDirection); 57 PetscErrorCode (*FinishCommunication) (PetscSF,PetscSFLink,PetscSFDirection); 58 PetscErrorCode (*SyncDevice) (PetscSFLink); 59 PetscErrorCode (*SyncStream) (PetscSFLink); 60 PetscErrorCode (*Destroy) (PetscSF,PetscSFLink); 61 62 PetscErrorCode (*BuildDependenceBegin)(PetscSF,PetscSFLink,PetscSFDirection); 63 PetscErrorCode (*BuildDependenceEnd) (PetscSF,PetscSFLink,PetscSFDirection); 64 65 PetscErrorCode (*h_Pack) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,void*); 66 PetscErrorCode (*h_UnpackAndInsert) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 67 PetscErrorCode (*h_UnpackAndAdd) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 68 PetscErrorCode (*h_UnpackAndMin) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 69 PetscErrorCode (*h_UnpackAndMax) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 70 PetscErrorCode (*h_UnpackAndMinloc) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 71 PetscErrorCode (*h_UnpackAndMaxloc) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 72 PetscErrorCode (*h_UnpackAndMult) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 73 PetscErrorCode (*h_UnpackAndLAND) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 74 PetscErrorCode (*h_UnpackAndBAND) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 75 PetscErrorCode (*h_UnpackAndLOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 76 PetscErrorCode (*h_UnpackAndBOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 77 PetscErrorCode (*h_UnpackAndLXOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 78 PetscErrorCode (*h_UnpackAndBXOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 79 PetscErrorCode (*h_FetchAndAdd) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*, void*); 80 81 PetscErrorCode (*h_ScatterAndInsert)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 82 PetscErrorCode (*h_ScatterAndAdd) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 83 PetscErrorCode (*h_ScatterAndMin) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 84 PetscErrorCode (*h_ScatterAndMax) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 85 PetscErrorCode (*h_ScatterAndMinloc)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 86 PetscErrorCode (*h_ScatterAndMaxloc)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 87 PetscErrorCode (*h_ScatterAndMult) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 88 PetscErrorCode (*h_ScatterAndLAND) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 89 PetscErrorCode (*h_ScatterAndBAND) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 90 PetscErrorCode (*h_ScatterAndLOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 91 PetscErrorCode (*h_ScatterAndBOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 92 PetscErrorCode (*h_ScatterAndLXOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 93 PetscErrorCode (*h_ScatterAndBXOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 94 95 PetscErrorCode (*h_FetchAndAddLocal)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,void*); 96 97 PetscBool deviceinited; /* Are device related fields initialized? */ 98 #if defined(PETSC_HAVE_DEVICE) 99 /* These fields are lazily initialized in a sense that only when device pointers are passed to an SF, the SF 100 will set them, otherwise it just leaves them alone. Packing routines using regular ops when there are no data race chances. 101 */ 102 PetscErrorCode (*d_Pack) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,void*); 103 PetscErrorCode (*d_UnpackAndInsert) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 104 PetscErrorCode (*d_UnpackAndAdd) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 105 PetscErrorCode (*d_UnpackAndMin) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 106 PetscErrorCode (*d_UnpackAndMax) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 107 PetscErrorCode (*d_UnpackAndMinloc) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 108 PetscErrorCode (*d_UnpackAndMaxloc) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 109 PetscErrorCode (*d_UnpackAndMult) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 110 PetscErrorCode (*d_UnpackAndLAND) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 111 PetscErrorCode (*d_UnpackAndBAND) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 112 PetscErrorCode (*d_UnpackAndLOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 113 PetscErrorCode (*d_UnpackAndBOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 114 PetscErrorCode (*d_UnpackAndLXOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 115 PetscErrorCode (*d_UnpackAndBXOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 116 PetscErrorCode (*d_FetchAndAdd) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*, void*); 117 118 PetscErrorCode (*d_ScatterAndInsert)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 119 PetscErrorCode (*d_ScatterAndAdd) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 120 PetscErrorCode (*d_ScatterAndMin) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 121 PetscErrorCode (*d_ScatterAndMax) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 122 PetscErrorCode (*d_ScatterAndMinloc)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 123 PetscErrorCode (*d_ScatterAndMaxloc)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 124 PetscErrorCode (*d_ScatterAndMult) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 125 PetscErrorCode (*d_ScatterAndLAND) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 126 PetscErrorCode (*d_ScatterAndBAND) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 127 PetscErrorCode (*d_ScatterAndLOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 128 PetscErrorCode (*d_ScatterAndBOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 129 PetscErrorCode (*d_ScatterAndLXOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 130 PetscErrorCode (*d_ScatterAndBXOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 131 PetscErrorCode (*d_FetchAndAddLocal)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,void*); 132 133 /* Packing routines using atomics when there are data race chances */ 134 PetscErrorCode (*da_UnpackAndInsert)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 135 PetscErrorCode (*da_UnpackAndAdd) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 136 PetscErrorCode (*da_UnpackAndMin) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 137 PetscErrorCode (*da_UnpackAndMax) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 138 PetscErrorCode (*da_UnpackAndMinloc)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 139 PetscErrorCode (*da_UnpackAndMaxloc)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 140 PetscErrorCode (*da_UnpackAndMult) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 141 PetscErrorCode (*da_UnpackAndLAND) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 142 PetscErrorCode (*da_UnpackAndBAND) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 143 PetscErrorCode (*da_UnpackAndLOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 144 PetscErrorCode (*da_UnpackAndBOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 145 PetscErrorCode (*da_UnpackAndLXOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 146 PetscErrorCode (*da_UnpackAndBXOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*); 147 PetscErrorCode (*da_FetchAndAdd) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*, void*); 148 149 PetscErrorCode (*da_ScatterAndInsert)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 150 PetscErrorCode (*da_ScatterAndAdd) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 151 PetscErrorCode (*da_ScatterAndMin) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 152 PetscErrorCode (*da_ScatterAndMax) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 153 PetscErrorCode (*da_ScatterAndMinloc)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 154 PetscErrorCode (*da_ScatterAndMaxloc)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 155 PetscErrorCode (*da_ScatterAndMult) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 156 PetscErrorCode (*da_ScatterAndLAND) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 157 PetscErrorCode (*da_ScatterAndBAND) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 158 PetscErrorCode (*da_ScatterAndLOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 159 PetscErrorCode (*da_ScatterAndBOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 160 PetscErrorCode (*da_ScatterAndLXOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 161 PetscErrorCode (*da_ScatterAndBXOR) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*); 162 PetscErrorCode (*da_FetchAndAddLocal)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,void*); 163 #if defined (PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP) 164 PetscInt maxResidentThreadsPerGPU; /* It is a copy from SF for convenience */ 165 cupmStream_t stream; /* stream on which input/output root/leafdata is computed on (default is PetscDefaultCudaStream) */ 166 #endif 167 #endif 168 PetscMPIInt tag; /* Each link has a tag so we can perform multiple SF ops at the same time */ 169 MPI_Datatype unit; /* The MPI datatype this PetscSFLink is built for */ 170 MPI_Datatype basicunit; /* unit is made of MPI builtin dataype basicunit */ 171 PetscBool isbuiltin; /* Is unit an MPI/PETSc builtin datatype? If it is true, then bs=1 and basicunit is equivalent to unit */ 172 size_t unitbytes; /* Number of bytes in a unit */ 173 PetscInt bs; /* Number of basic units in a unit */ 174 const void *rootdata,*leafdata; /* rootdata and leafdata the link is working on. They are used as keys for pending links. */ 175 PetscMemType rootmtype,leafmtype; /* root/leafdata's memory type */ 176 177 /* For local and remote communication */ 178 PetscMemType rootmtype_mpi,leafmtype_mpi; /* Mtypes of buffers passed to MPI. If use_gpu_aware_mpi, they are same as root/leafmtype. Otherwise they are PETSC_MEMTYPE_HOST */ 179 PetscBool rootdirect[2],leafdirect[2]; /* Can root/leafdata be directly passed to SF (i.e., without buffering). In layout of [PETSCSF_LOCAL/REMOTE]. See more in PetscSFLinkCreate() */ 180 PetscInt rootdirect_mpi,leafdirect_mpi;/* Can root/leafdata for remote be directly passed to MPI? 1: yes, 0: no. See more in PetscSFLinkCreate() */ 181 const void *rootdatadirect[2][2]; /* The root/leafdata used to init root/leaf requests, in layout of [PETSCSF_DIRECTION][PETSC_MEMTYPE]. */ 182 const void *leafdatadirect[2][2]; /* ... We need them to look up links when root/leafdirect_mpi are true */ 183 char *rootbuf[2][2]; /* Buffers for packed roots, in layout of [PETSCSF_LOCAL/REMOTE][PETSC_MEMTYPE]. PETSCSF_LOCAL does not need MPI, .. */ 184 /* .. but in case rootmtype is different from leafmtype, we still need to pack local roots and then copy them to memory of leafmtype */ 185 char *rootbuf_alloc[2][2]; /* Log memory allocated by petsc. We need it since rootbuf[][] may point to rootdata given by user */ 186 char *leafbuf[2][2]; /* Buffers for packed leaves, in layout of [PETSCSF_LOCAL/REMOTE][PETSC_MEMTYPE] */ 187 char *leafbuf_alloc[2][2]; 188 MPI_Request *rootreqs[2][2][2]; /* Root requests in layout of [PETSCSF_DIRECTION][PETSC_MEMTYPE][rootdirect_mpi] */ 189 MPI_Request *leafreqs[2][2][2]; /* Leaf requests in layout of [PETSCSF_DIRECTION][PETSC_MEMTYPE][leafdirect_mpi] */ 190 PetscBool rootreqsinited[2][2][2]; /* Are root requests initialized? Also in layout of [PETSCSF_DIRECTION][PETSC_MEMTYPE][rootdirect_mpi]*/ 191 PetscBool leafreqsinited[2][2][2]; /* Are leaf requests initialized? Also in layout of [PETSCSF_DIRECTION][PETSC_MEMTYPE][leafdirect_mpi]*/ 192 MPI_Request *reqs; /* An array of length (nrootreqs+nleafreqs)*8. Pointers in rootreqs[][][] and leafreqs[][][] point here */ 193 PetscSFLink next; 194 195 PetscBool use_nvshmem; /* Does this link use nvshem (vs. MPI) for communication? */ 196 #if defined(PETSC_HAVE_NVSHMEM) 197 cupmEvent_t dataReady; /* Events to mark readiness of root/leafdata */ 198 cupmEvent_t endRemoteComm; /* Events to mark end of local/remote communication */ 199 cupmStream_t remoteCommStream; /* Streams for remote (i.e., inter-rank) communication */ 200 201 /* The buffers are allocated in device symmetric heap. Their length is the maximal length over all ranks in the comm, and therefore is the same. */ 202 uint64_t *rootSendSig,*rootRecvSig; /* [max{niranks-ndiranks}], signals used when rootbuf works as send/recv buf */ 203 uint64_t *leafSendSig,*leafRecvSig; /* [max{nranks-ndranks}], signals used when leafbuf works as send/recv buf */ 204 #endif 205 }; 206 207 PETSC_INTERN PetscErrorCode PetscSFSetErrorOnUnsupportedOverlap(PetscSF,MPI_Datatype,const void*,const void*); 208 209 /* Create/setup/retrieve/destroy a link */ 210 PETSC_INTERN PetscErrorCode PetscSFLinkCreate(PetscSF,MPI_Datatype,PetscMemType,const void*,PetscMemType,const void*,MPI_Op,PetscSFOperation,PetscSFLink*); 211 PETSC_INTERN PetscErrorCode PetscSFLinkSetUp_Host(PetscSF,PetscSFLink,MPI_Datatype); 212 PETSC_INTERN PetscErrorCode PetscSFLinkGetInUse(PetscSF,MPI_Datatype,const void*,const void*,PetscCopyMode,PetscSFLink*); 213 PETSC_INTERN PetscErrorCode PetscSFLinkReclaim(PetscSF,PetscSFLink*); 214 PETSC_INTERN PetscErrorCode PetscSFLinkDestroy(PetscSF,PetscSFLink); 215 216 /* Get pack/unpack function pointers from a link */ 217 static inline PetscErrorCode PetscSFLinkGetPack(PetscSFLink link,PetscMemType mtype,PetscErrorCode (**Pack)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,void*)) 218 { 219 PetscFunctionBegin; 220 if (PetscMemTypeHost(mtype)) *Pack = link->h_Pack; 221 #if defined(PETSC_HAVE_DEVICE) 222 else *Pack = link->d_Pack; 223 #endif 224 PetscFunctionReturn(0); 225 } 226 227 PETSC_INTERN PetscErrorCode PetscSFLinkGetUnpackAndOp(PetscSFLink,PetscMemType,MPI_Op,PetscBool,PetscErrorCode (**UnpackAndOp)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,const void*)); 228 PETSC_INTERN PetscErrorCode PetscSFLinkGetFetchAndOp (PetscSFLink,PetscMemType,MPI_Op,PetscBool,PetscErrorCode (**FetchAndOp) (PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,void*)); 229 PETSC_INTERN PetscErrorCode PetscSFLinkGetScatterAndOp(PetscSFLink,PetscMemType,MPI_Op,PetscBool,PetscErrorCode (**ScatterAndOp)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,PetscInt,PetscSFPackOpt,const PetscInt*,void*)); 230 PETSC_INTERN PetscErrorCode PetscSFLinkGetFetchAndOpLocal(PetscSFLink,PetscMemType,MPI_Op,PetscBool,PetscErrorCode (**FetchAndOpLocal)(PetscSFLink,PetscInt,PetscInt,PetscSFPackOpt,const PetscInt*,void*,PetscInt,PetscSFPackOpt,const PetscInt*,const void*,void*)); 231 PETSC_INTERN PetscErrorCode PetscSFLinkGetMPIBuffersAndRequests(PetscSF,PetscSFLink,PetscSFDirection,void**,void**,MPI_Request**,MPI_Request**); 232 233 /* Do Pack/Unpack/Fetch/Scatter with the link */ 234 PETSC_INTERN PetscErrorCode PetscSFLinkPackRootData (PetscSF,PetscSFLink,PetscSFScope,const void*); 235 PETSC_INTERN PetscErrorCode PetscSFLinkPackLeafData (PetscSF,PetscSFLink,PetscSFScope,const void*); 236 PETSC_INTERN PetscErrorCode PetscSFLinkUnpackRootData(PetscSF,PetscSFLink,PetscSFScope,void*,MPI_Op); 237 PETSC_INTERN PetscErrorCode PetscSFLinkUnpackLeafData(PetscSF,PetscSFLink,PetscSFScope,void*,MPI_Op); 238 PETSC_INTERN PetscErrorCode PetscSFLinkFetchAndOpRemote (PetscSF,PetscSFLink,void*,MPI_Op); 239 240 PETSC_INTERN PetscErrorCode PetscSFLinkScatterLocal(PetscSF,PetscSFLink,PetscSFDirection,void*,void*,MPI_Op); 241 PETSC_INTERN PetscErrorCode PetscSFLinkFetchAndOpLocal(PetscSF,PetscSFLink,void*,const void*,void*,MPI_Op); 242 243 PETSC_INTERN PetscErrorCode PetscSFSetUpPackFields(PetscSF); 244 PETSC_INTERN PetscErrorCode PetscSFResetPackFields(PetscSF); 245 PETSC_INTERN PetscErrorCode PetscSFLinkCreate_MPI(PetscSF,MPI_Datatype,PetscMemType,const void*,PetscMemType,const void*,MPI_Op,PetscSFOperation,PetscSFLink*); 246 247 #if defined(PETSC_HAVE_CUDA) 248 PETSC_INTERN PetscErrorCode PetscSFLinkSetUp_CUDA(PetscSF,PetscSFLink,MPI_Datatype); 249 #endif 250 251 #if defined(PETSC_HAVE_HIP) 252 PETSC_INTERN PetscErrorCode PetscSFLinkSetUp_HIP(PetscSF,PetscSFLink,MPI_Datatype); 253 #endif 254 255 #if defined(PETSC_HAVE_KOKKOS) 256 PETSC_INTERN PetscErrorCode PetscSFLinkSetUp_Kokkos(PetscSF,PetscSFLink,MPI_Datatype); 257 #endif 258 259 #if defined(PETSC_HAVE_NVSHMEM) 260 PETSC_INTERN PetscErrorCode PetscSFLinkCreate_NVSHMEM(PetscSF,MPI_Datatype,PetscMemType,const void*,PetscMemType,const void*,MPI_Op,PetscSFOperation,PetscSFLink*); 261 PETSC_INTERN PetscErrorCode PetscSFLinkNvshmemCheck(PetscSF,PetscMemType,const void*,PetscMemType,const void*,PetscBool*); 262 #endif 263 264 static inline PetscErrorCode PetscSFLinkStartCommunication(PetscSF sf,PetscSFLink link,PetscSFDirection direction) 265 { 266 PetscFunctionBegin; 267 if (link->StartCommunication) PetscCall((*link->StartCommunication)(sf,link,direction)); 268 PetscFunctionReturn(0); 269 } 270 271 static inline PetscErrorCode PetscSFLinkFinishCommunication(PetscSF sf,PetscSFLink link,PetscSFDirection direction) 272 { 273 PetscFunctionBegin; 274 if (link->FinishCommunication) PetscCall((*link->FinishCommunication)(sf,link,direction)); 275 PetscFunctionReturn(0); 276 } 277 278 /* A set of helper routines for Pack/Unpack/Scatter on GPUs */ 279 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP) || defined(PETSC_HAVE_SYCL) 280 /* PetscSFLinkCopyXxxxBufferInCaseNotUseGpuAwareMPI routines are simple: if not use_gpu_aware_mpi, we need 281 to copy the buffer from GPU to CPU before MPI calls, and from CPU to GPU after MPI calls. 282 */ 283 static inline PetscErrorCode PetscSFLinkCopyRootBufferInCaseNotUseGpuAwareMPI(PetscSF sf,PetscSFLink link,PetscBool device2host) 284 { 285 PetscSF_Basic *bas = (PetscSF_Basic*)sf->data; 286 287 PetscFunctionBegin; 288 /* rootdata is on device but we use regular MPI for communication */ 289 if (PetscMemTypeDevice(link->rootmtype) && PetscMemTypeHost(link->rootmtype_mpi) && bas->rootbuflen[PETSCSF_REMOTE]) { 290 void *h_buf = link->rootbuf[PETSCSF_REMOTE][PETSC_MEMTYPE_HOST]; 291 void *d_buf = link->rootbuf[PETSCSF_REMOTE][PETSC_MEMTYPE_DEVICE]; 292 size_t count = bas->rootbuflen[PETSCSF_REMOTE]*link->unitbytes; 293 if (device2host) { 294 PetscCall((*link->Memcpy)(link,PETSC_MEMTYPE_HOST,h_buf,PETSC_MEMTYPE_DEVICE,d_buf,count)); 295 PetscCall(PetscLogGpuToCpu(count)); 296 } else { 297 PetscCall((*link->Memcpy)(link,PETSC_MEMTYPE_DEVICE,d_buf,PETSC_MEMTYPE_HOST,h_buf,count)); 298 PetscCall(PetscLogCpuToGpu(count)); 299 } 300 } 301 PetscFunctionReturn(0); 302 } 303 304 static inline PetscErrorCode PetscSFLinkCopyLeafBufferInCaseNotUseGpuAwareMPI(PetscSF sf,PetscSFLink link,PetscBool device2host) 305 { 306 PetscFunctionBegin; 307 if (PetscMemTypeDevice(link->leafmtype) && PetscMemTypeHost(link->leafmtype_mpi) && sf->leafbuflen[PETSCSF_REMOTE]) { 308 void *h_buf = link->leafbuf[PETSCSF_REMOTE][PETSC_MEMTYPE_HOST]; 309 void *d_buf = link->leafbuf[PETSCSF_REMOTE][PETSC_MEMTYPE_DEVICE]; 310 size_t count = sf->leafbuflen[PETSCSF_REMOTE]*link->unitbytes; 311 if (device2host) { 312 PetscCall((*link->Memcpy)(link,PETSC_MEMTYPE_HOST,h_buf,PETSC_MEMTYPE_DEVICE,d_buf,count)); 313 PetscCall(PetscLogGpuToCpu(count)); 314 } else { 315 PetscCall((*link->Memcpy)(link,PETSC_MEMTYPE_DEVICE,d_buf,PETSC_MEMTYPE_HOST,h_buf,count)); 316 PetscCall(PetscLogCpuToGpu(count)); 317 } 318 } 319 PetscFunctionReturn(0); 320 } 321 322 /* Make sure root/leafbuf for the remote is ready for MPI */ 323 static inline PetscErrorCode PetscSFLinkSyncStreamBeforeCallMPI(PetscSF sf,PetscSFLink link,PetscSFDirection direction) 324 { 325 PetscSF_Basic *bas; 326 PetscInt buflen; 327 PetscMemType mtype; 328 329 PetscFunctionBegin; 330 if (direction == PETSCSF_ROOT2LEAF) { 331 bas = (PetscSF_Basic*)sf->data; 332 mtype = link->rootmtype; 333 buflen = bas->rootbuflen[PETSCSF_REMOTE]; 334 } else { 335 mtype = link->leafmtype; 336 buflen = sf->leafbuflen[PETSCSF_REMOTE]; 337 } 338 339 if (PetscMemTypeDevice(mtype) && buflen) { 340 PetscCall((*link->SyncStream)(link)); 341 } 342 PetscFunctionReturn(0); 343 } 344 #else /* Host only */ 345 #define PetscSFLinkCopyRootBufferInCaseNotUseGpuAwareMPI(a,b,c) 0 346 #define PetscSFLinkCopyLeafBufferInCaseNotUseGpuAwareMPI(a,b,c) 0 347 #define PetscSFLinkSyncStreamBeforeCallMPI(a,b,c) 0 348 #endif 349 350 /* Get root indices used for pack/unpack 351 352 Input arguments: 353 +sf - StarForest 354 .link - The link, which provides the stream for the async memcpy (In SF, we make all GPU operations asynchronous to avoid unexpected pipeline stalls) 355 .mtype - In what type of memory? (PETSC_MEMTYPE_DEVICE or PETSC_MEMTYPE_HOST) 356 -scope - Which part of the indices? (PETSCSF_LOCAL or PETSCSF_REMOTE) 357 358 Output arguments: 359 +count - Count of indices 360 .start - The first index (only useful when indices is NULL) 361 .opt - Packing optimizations 362 -indices - Indices of roots for pack/unpack. NULL means indices are contiguous 363 */ 364 static inline PetscErrorCode PetscSFLinkGetRootPackOptAndIndices(PetscSF sf,PetscSFLink link,PetscMemType mtype,PetscSFScope scope,PetscInt *count,PetscInt *start,PetscSFPackOpt *opt,const PetscInt **indices) 365 { 366 PetscSF_Basic *bas = (PetscSF_Basic*)sf->data; 367 PetscInt offset; 368 369 PetscFunctionBegin; 370 *count = bas->rootbuflen[scope]; 371 *start = bas->rootstart[scope]; 372 *opt = NULL; 373 *indices = NULL; 374 375 /* We have these rules: 376 1) opt == NULL && indices == NULL ==> indices are contiguous. 377 2) opt != NULL ==> indices are in 3D but not contiguous. On host, indices != NULL since indices are already available and we do not 378 want to enforce all operations to use opt; but on device, indices = NULL since we do not want to copy indices to device. 379 */ 380 if (!bas->rootcontig[scope]) { 381 offset = (scope == PETSCSF_LOCAL)? 0 : bas->ioffset[bas->ndiranks]; 382 if (PetscMemTypeHost(mtype)) {*opt = bas->rootpackopt[scope]; *indices = bas->irootloc + offset;} 383 else { 384 size_t size; 385 if (bas->rootpackopt[scope]) { 386 if (!bas->rootpackopt_d[scope]) { 387 PetscCall(PetscMalloc1(1,&bas->rootpackopt_d[scope])); 388 PetscCall(PetscArraycpy(bas->rootpackopt_d[scope],bas->rootpackopt[scope],1)); /* Make pointers in bas->rootpackopt_d[] still work on host */ 389 size = (bas->rootpackopt[scope]->n*7+2)*sizeof(PetscInt); /* See comments at struct _n_PetscSFPackOpt*/ 390 PetscCall(PetscSFMalloc(sf,PETSC_MEMTYPE_DEVICE,size,(void **)&bas->rootpackopt_d[scope]->array)); 391 PetscCall((*link->Memcpy)(link,PETSC_MEMTYPE_DEVICE,bas->rootpackopt_d[scope]->array,PETSC_MEMTYPE_HOST,bas->rootpackopt[scope]->array,size)); 392 } 393 *opt = bas->rootpackopt_d[scope]; 394 } else { /* On device, we only provide indices when there is no optimization. We're reluctant to copy indices to device. */ 395 if (!bas->irootloc_d[scope]) { 396 size = bas->rootbuflen[scope]*sizeof(PetscInt); 397 PetscCall(PetscSFMalloc(sf,PETSC_MEMTYPE_DEVICE,size,(void **)&bas->irootloc_d[scope])); 398 PetscCall((*link->Memcpy)(link,PETSC_MEMTYPE_DEVICE,bas->irootloc_d[scope],PETSC_MEMTYPE_HOST,bas->irootloc+offset,size)); 399 } 400 *indices = bas->irootloc_d[scope]; 401 } 402 } 403 } 404 PetscFunctionReturn(0); 405 } 406 407 /* Get leaf indices used for pack/unpack 408 409 See also PetscSFLinkGetRootPackOptAndIndices() 410 */ 411 static inline PetscErrorCode PetscSFLinkGetLeafPackOptAndIndices(PetscSF sf,PetscSFLink link,PetscMemType mtype,PetscSFScope scope,PetscInt *count,PetscInt *start,PetscSFPackOpt *opt,const PetscInt **indices) 412 { 413 PetscInt offset; 414 415 PetscFunctionBegin; 416 *count = sf->leafbuflen[scope]; 417 *start = sf->leafstart[scope]; 418 *opt = NULL; 419 *indices = NULL; 420 if (!sf->leafcontig[scope]) { 421 offset = (scope == PETSCSF_LOCAL)? 0 : sf->roffset[sf->ndranks]; 422 if (PetscMemTypeHost(mtype)) {*opt = sf->leafpackopt[scope]; *indices = sf->rmine + offset;} 423 else { 424 size_t size; 425 if (sf->leafpackopt[scope]) { 426 if (!sf->leafpackopt_d[scope]) { 427 PetscCall(PetscMalloc1(1,&sf->leafpackopt_d[scope])); 428 PetscCall(PetscArraycpy(sf->leafpackopt_d[scope],sf->leafpackopt[scope],1)); 429 size = (sf->leafpackopt[scope]->n*7+2)*sizeof(PetscInt); /* See comments at struct _n_PetscSFPackOpt*/ 430 PetscCall(PetscSFMalloc(sf,PETSC_MEMTYPE_DEVICE,size,(void **)&sf->leafpackopt_d[scope]->array)); /* Change ->array to a device pointer */ 431 PetscCall((*link->Memcpy)(link,PETSC_MEMTYPE_DEVICE,sf->leafpackopt_d[scope]->array,PETSC_MEMTYPE_HOST,sf->leafpackopt[scope]->array,size)); 432 } 433 *opt = sf->leafpackopt_d[scope]; 434 } else { 435 if (!sf->rmine_d[scope]) { 436 size = sf->leafbuflen[scope]*sizeof(PetscInt); 437 PetscCall(PetscSFMalloc(sf,PETSC_MEMTYPE_DEVICE,size,(void **)&sf->rmine_d[scope])); 438 PetscCall((*link->Memcpy)(link,PETSC_MEMTYPE_DEVICE,sf->rmine_d[scope],PETSC_MEMTYPE_HOST,sf->rmine+offset,size)); 439 } 440 *indices = sf->rmine_d[scope]; 441 } 442 } 443 } 444 PetscFunctionReturn(0); 445 } 446 #endif 447