1*40e23c03SJunchao Zhang #if !defined(__SFPACK_H) 2*40e23c03SJunchao Zhang #define __SFPACK_H 3*40e23c03SJunchao Zhang 4*40e23c03SJunchao Zhang #include <petsc/private/sfimpl.h> /*I "petscsf.h" I*/ 5*40e23c03SJunchao Zhang 6*40e23c03SJunchao Zhang /* Optimization plans in packing(unpacking) for target processors. 7*40e23c03SJunchao Zhang 8*40e23c03SJunchao Zhang Indirect accesses in packing like p[i] = u[idx[i]] are expensive and are not vectorization friendly. We 9*40e23c03SJunchao Zhang try to optimize them if we found cenrtain patterns among indices in idx[]. As a result, a pack might be 10*40e23c03SJunchao Zhang optimized into 1) a small number of contiguous memory copies; OR 2) one strided memory copy. 11*40e23c03SJunchao Zhang 12*40e23c03SJunchao Zhang Each target has its own plan. n, the number of target processors, is nranks or niranks depending on the context. 13*40e23c03SJunchao Zhang */ 14*40e23c03SJunchao Zhang struct _n_PetscSFPackOpt { 15*40e23c03SJunchao Zhang PetscBool *optimized; /* [n] Is the packing to i-th target optimized? If yes, other fields give the opt plan */ 16*40e23c03SJunchao Zhang PetscInt *copy_offset; /* [n+1] We number all memory copies. Packing for target i is optimized into copies in [copy_offset[i],copy_offset[i+1]) */ 17*40e23c03SJunchao Zhang PetscInt *copy_start; /* [*] j-th copy starts at index copy_start[j] */ 18*40e23c03SJunchao Zhang PetscInt *copy_length; /* [*] with length copy_length[j] in unit of the <unit> used in for example, PetscSFReduceBegin(sf,unit,...) */ 19*40e23c03SJunchao Zhang PetscInt *stride_first; /* [n] If optimized[i] is TRUE but copy_offset[i] == copy_offset[i+1], then packing for remote i is strided. The first */ 20*40e23c03SJunchao Zhang PetscInt *stride_step; /* [n] index is stride_first[i], step is stride_step[i], */ 21*40e23c03SJunchao Zhang PetscInt *stride_n; /* [n] and total stride_n[i] steps */ 22*40e23c03SJunchao Zhang }; 23*40e23c03SJunchao Zhang 24*40e23c03SJunchao Zhang typedef struct _n_PetscSFPack* PetscSFPack; 25*40e23c03SJunchao Zhang 26*40e23c03SJunchao Zhang #define SFPACKHEADER \ 27*40e23c03SJunchao Zhang PetscErrorCode (*Pack) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,const void*,void*); \ 28*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndInsert)(PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 29*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndAdd) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 30*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndMin) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 31*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndMax) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 32*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndMinloc)(PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 33*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndMaxloc)(PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 34*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndMult) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 35*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndLAND) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 36*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndBAND) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 37*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndLOR) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 38*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndBOR) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 39*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndLXOR) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 40*40e23c03SJunchao Zhang PetscErrorCode (*UnpackAndBXOR) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*); \ 41*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndInsert) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 42*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndAdd) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 43*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndMin) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 44*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndMax) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 45*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndMinloc) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 46*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndMaxloc) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 47*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndMult) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 48*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndLAND) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 49*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndBAND) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 50*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndLOR) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 51*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndBOR) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 52*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndLXOR) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 53*40e23c03SJunchao Zhang PetscErrorCode (*FetchAndBXOR) (PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*); \ 54*40e23c03SJunchao Zhang PetscMPIInt tag; /* Each link has a tag so we can perform multiple SF ops at the same time */ \ 55*40e23c03SJunchao Zhang MPI_Datatype unit; \ 56*40e23c03SJunchao Zhang MPI_Datatype basicunit; /* unit is made of MPI builtin dataype basicunit */ \ 57*40e23c03SJunchao Zhang PetscBool isbuiltin; /* Is unit an MPI builtin datatype? If it is true, basicunit=unit, bs=1 */ \ 58*40e23c03SJunchao Zhang size_t unitbytes; /* Number of bytes in a unit */ \ 59*40e23c03SJunchao Zhang PetscInt bs; /* Number of basic units in a unit */ \ 60*40e23c03SJunchao Zhang const void *key; /* Array used as key for operation */ \ 61*40e23c03SJunchao Zhang PetscSFPack next 62*40e23c03SJunchao Zhang 63*40e23c03SJunchao Zhang /* An abstract class that defines a communication link, which includes how to 64*40e23c03SJunchao Zhang pack/unpack data. Subclasses may further contain fields for send/recv buffers, 65*40e23c03SJunchao Zhang MPI_Requests etc used in communication. 66*40e23c03SJunchao Zhang */ 67*40e23c03SJunchao Zhang struct _n_PetscSFPack { 68*40e23c03SJunchao Zhang SFPACKHEADER; 69*40e23c03SJunchao Zhang }; 70*40e23c03SJunchao Zhang 71*40e23c03SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFPackGetInUse(PetscSF,MPI_Datatype,const void*,PetscCopyMode,PetscSFPack*); 72*40e23c03SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFPackReclaim(PetscSF,PetscSFPack*); 73*40e23c03SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFPackSetupType(PetscSFPack,MPI_Datatype); 74*40e23c03SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFPackGetUnpackAndOp(PetscSF,PetscSFPack,MPI_Op,PetscErrorCode (**UnpackAndOp)(PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,const void*)); 75*40e23c03SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFPackGetFetchAndOp(PetscSF,PetscSFPack,MPI_Op,PetscErrorCode (**FetchAndOp)(PetscInt,PetscInt,const PetscInt*,PetscInt,PetscSFPackOpt,void*,void*)); 76*40e23c03SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFPackSetupOptimization(PetscInt,const PetscInt*,const PetscInt*,PetscSFPackOpt*); 77*40e23c03SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFPackDestoryOptimization(PetscSFPackOpt *out); 78*40e23c03SJunchao Zhang #endif 79