1 #if !defined(__SFPACK_H) 2 #define __SFPACK_H 3 4 #include <petsc/private/sfimpl.h> /*I "petscsf.h" I*/ 5 6 /* Optimization plans in packing & unpacking for destination ranks. 7 8 Suppose there are count indices stored in idx[], and two addresses u, p. We want to do packing: 9 p[i] = u[idx[i]], for i in [0,count) 10 11 Often, the indices are associated with n ranks. Each rank's indices are stored consecutively in idx[]. 12 We analyze indices for each rank and see if they are patterns that can be used to optimize the packing. 13 The result is stored in PetscSFPackOpt. Packing for a rank might be not optimizable, or optimized in 14 to a small number of contiguous memory copies or one strided memory copy. 15 */ 16 typedef enum {PETSCSF_PACKOPT_NONE=0, PETSCSF_PACKOPT_MULTICOPY, PETSCSF_PACKOPT_STRIDE} PetscSFPackOptType; 17 18 struct _n_PetscSFPackOpt { 19 PetscInt n; /* Number of destination ranks */ 20 PetscSFPackOptType *type; /* [n] Optimization types for the n ranks */ 21 PetscInt *offset; /* [n+1] Indices for i-th rank are in [offset[i],offset[i+1]) of idx[] */ 22 PetscInt *copy_offset; /* [n+1] If type[i] = PETSCSF_PACKOPT_MULTICOPY, packing for i-th rank is optimized into copies numbered between [copy_offset[i],copy_offset[i+1]) */ 23 PetscInt *copy_start; /* [*] j-th copy starts at copy_start[j] in idx[]. In other words, there are copy_length[j] contiguous indices */ 24 PetscInt *copy_length; /* [*] starting from idx[copy_start[j]] */ 25 PetscInt *stride_step; /* [n] If type[i] = PETSCSF_PACKOPT_STRIDE, then packing for i-th rank is strided, with first index being idx[offset[i]] and step stride_step[i], */ 26 PetscInt *stride_n; /* [n] and total stride_n[i] steps */ 27 }; 28 29 typedef struct _n_PetscSFPack* PetscSFPack; 30 31 #define SFPACKHEADER \ 32 PetscErrorCode (*Pack) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,const void*,void*); \ 33 PetscErrorCode (*UnpackAndInsert)(PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 34 PetscErrorCode (*UnpackAndAdd) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 35 PetscErrorCode (*UnpackAndMin) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 36 PetscErrorCode (*UnpackAndMax) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 37 PetscErrorCode (*UnpackAndMinloc)(PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 38 PetscErrorCode (*UnpackAndMaxloc)(PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 39 PetscErrorCode (*UnpackAndMult) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 40 PetscErrorCode (*UnpackAndLAND) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 41 PetscErrorCode (*UnpackAndBAND) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 42 PetscErrorCode (*UnpackAndLOR) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 43 PetscErrorCode (*UnpackAndBOR) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 44 PetscErrorCode (*UnpackAndLXOR) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 45 PetscErrorCode (*UnpackAndBXOR) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 46 PetscErrorCode (*FetchAndInsert) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 47 PetscErrorCode (*FetchAndAdd) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 48 PetscErrorCode (*FetchAndMin) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 49 PetscErrorCode (*FetchAndMax) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 50 PetscErrorCode (*FetchAndMinloc) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 51 PetscErrorCode (*FetchAndMaxloc) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 52 PetscErrorCode (*FetchAndMult) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 53 PetscErrorCode (*FetchAndLAND) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 54 PetscErrorCode (*FetchAndBAND) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 55 PetscErrorCode (*FetchAndLOR) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 56 PetscErrorCode (*FetchAndBOR) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 57 PetscErrorCode (*FetchAndLXOR) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 58 PetscErrorCode (*FetchAndBXOR) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 59 PetscMPIInt tag; /* Each link has a tag so we can perform multiple SF ops at the same time */\ 60 MPI_Datatype unit; \ 61 MPI_Datatype basicunit; /* unit is made of MPI builtin dataype basicunit */ \ 62 PetscBool isbuiltin; /* Is unit an MPI builtin datatype? If it is true, basicunit=unit, bs=1 */ \ 63 size_t unitbytes; /* Number of bytes in a unit */ \ 64 PetscInt bs; /* Number of basic units in a unit */ \ 65 const void *rkey,*lkey; /* rootdata and leafdata used as keys for operation */ \ 66 char *rootbuf; /* Buffer for packed roots in send/recv */ \ 67 char *leafbuf; /* Buffer for packed leaves in send/recv */ \ 68 char *selfbuf; /* If self communication does not use MPI, this is the shared buffer for packed roots or leaves */ \ 69 PetscSFPack next 70 71 /* An abstract class that defines a communication link, which includes how to 72 pack/unpack data. Subclasses may further contain fields for send/recv buffers, 73 MPI_Requests etc used in communication. 74 */ 75 struct _n_PetscSFPack { 76 SFPACKHEADER; 77 }; 78 79 PETSC_INTERN PetscErrorCode PetscSFPackGetInUse(PetscSF,MPI_Datatype,const void*,const void*,PetscCopyMode,PetscSFPack*); 80 PETSC_INTERN PetscErrorCode PetscSFPackReclaim(PetscSF,PetscSFPack*); 81 PETSC_INTERN PetscErrorCode PetscSFPackSetupType(PetscSFPack,MPI_Datatype); 82 PETSC_INTERN PetscErrorCode PetscSFPackGetUnpackAndOp(PetscSF,PetscSFPack,MPI_Op,PetscErrorCode (**UnpackAndOp)(PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*)); 83 PETSC_INTERN PetscErrorCode PetscSFPackGetFetchAndOp (PetscSF,PetscSFPack,MPI_Op,PetscErrorCode (**FetchAndOp) (PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*, void*)); 84 PETSC_INTERN PetscErrorCode PetscSFPackSetupOptimization(PetscInt,const PetscInt*,const PetscInt*,PetscSFPackOpt*); 85 PETSC_INTERN PetscErrorCode PetscSFPackDestoryOptimization(PetscSFPackOpt *out); 86 PETSC_INTERN PetscErrorCode PetscSFPackSetErrorOnUnsupportedOverlap(PetscSF,MPI_Datatype,const void*,const void*); 87 88 #endif 89