DENDRO: /net/www/csela/dendro/releases/Dendro/include/par/parUtils.h Source File

00001 
00011 #ifndef __PAR_UTILS_H_
00012 #define __PAR_UTILS_H_
00013 
00014 #define KEEP_HIGH 100
00015 #define KEEP_LOW  101
00016 
00017 #ifdef __DEBUG__
00018 #ifndef __DEBUG_PAR__
00019 #define __DEBUG_PAR__
00020 #endif
00021 #endif
00022 
00023 #include "mpi.h"
00024 #include <vector>
00025 #include "dendro.h"
00026 
00027 #ifdef PETSC_USE_LOG
00028 
00029 #include "petscsys.h"
00030 
00031 namespace par {
00032   extern int sortEvent;
00033   extern int concatEvent;
00034   extern int remdupEvent;
00035   extern int partwEvent;
00036   extern int searchEvent;
00037   extern int parScatterEvent;
00038   extern int a2avWaitEvent;
00039   extern int all2AllvSparseEvent;
00040   extern int all2AllvDenseEvent;
00041   extern int allGatherEvent;
00042   extern int reduceEvent;
00043   extern int sendRecvEvent;
00044   extern int allReduceEvent;
00045   extern int all2AllEvent;
00046   extern int allGathervEvent;
00047   extern int gatherEvent;
00048   extern int scanEvent;
00049   extern int bcastEvent;
00050   extern int splitComm2wayEvent;
00051   extern int splitCommEvent;
00052 }
00053 
00054 #define PROF_A2AV_WAIT_BEGIN PetscLogEventBegin(a2avWaitEvent,0,0,0,0);
00055 #define PROF_A2AV_WAIT_END  PetscLogEventEnd(a2avWaitEvent,0,0,0,0); 
00056 
00057 #define PROF_SPLIT_COMM_2WAY_BEGIN \
00058   PetscFunctionBegin; \
00059 PetscLogEventBegin(splitComm2wayEvent,0,0,0,0);
00060 #define PROF_SPLIT_COMM_2WAY_END         \
00061   PetscLogEventEnd(splitComm2wayEvent,0,0,0,0); \
00062 PetscFunctionReturn(0);
00063 
00064 #define PROF_SPLIT_COMM_BEGIN \
00065   PetscFunctionBegin; \
00066 PetscLogEventBegin(splitCommEvent,0,0,0,0);
00067 #define PROF_SPLIT_COMM_END         \
00068   PetscLogEventEnd(splitCommEvent,0,0,0,0); \
00069 PetscFunctionReturn(0);
00070 
00071 #define PROF_SEARCH_BEGIN \
00072   PetscFunctionBegin; \
00073 PetscLogEventBegin(searchEvent,0,0,0,0);
00074 #define PROF_SEARCH_END         \
00075   PetscLogEventEnd(searchEvent,0,0,0,0); \
00076 PetscFunctionReturn(0);
00077 
00078 #define PROF_PAR_CONCAT_BEGIN \
00079   PetscFunctionBegin; \
00080 PetscLogEventBegin(concatEvent,0,0,0,0);
00081 #define PROF_PAR_CONCAT_END         \
00082   PetscLogEventEnd(concatEvent,0,0,0,0); \
00083 PetscFunctionReturn(0);
00084 
00085 #define PROF_PAR_SCATTER_BEGIN \
00086   PetscFunctionBegin; \
00087 PetscLogEventBegin(parScatterEvent,0,0,0,0);
00088 #define PROF_PAR_SCATTER_END         \
00089   PetscLogEventEnd(parScatterEvent,0,0,0,0); \
00090 PetscFunctionReturn(0);
00091 
00092 #define PROF_PAR_SENDRECV_BEGIN \
00093   PetscFunctionBegin; \
00094 PetscLogEventBegin(sendRecvEvent,0,0,0,0);
00095 #define PROF_PAR_SENDRECV_END \
00096   PetscLogEventEnd(sendRecvEvent,0,0,0,0); \
00097 PetscFunctionReturn(0);
00098 
00099 #define PROF_PAR_BCAST_BEGIN \
00100   PetscFunctionBegin; \
00101 PetscLogEventBegin(bcastEvent,0,0,0,0);
00102 #define PROF_PAR_BCAST_END \
00103   PetscLogEventEnd(bcastEvent,0,0,0,0); \
00104 PetscFunctionReturn(0);
00105 
00106 #define PROF_PAR_SCAN_BEGIN \
00107   PetscFunctionBegin; \
00108 PetscLogEventBegin(scanEvent,0,0,0,0);
00109 #define PROF_PAR_SCAN_END \
00110   PetscLogEventEnd(scanEvent,0,0,0,0); \
00111 PetscFunctionReturn(0);
00112 
00113 #define PROF_PAR_GATHER_BEGIN \
00114   PetscFunctionBegin; \
00115 PetscLogEventBegin(gatherEvent,0,0,0,0);
00116 #define PROF_PAR_GATHER_END \
00117   PetscLogEventEnd(gatherEvent,0,0,0,0); \
00118 PetscFunctionReturn(0);
00119 
00120 #define PROF_PAR_REDUCE_BEGIN \
00121   PetscFunctionBegin; \
00122 PetscLogEventBegin(reduceEvent,0,0,0,0);
00123 #define PROF_PAR_REDUCE_END \
00124   PetscLogEventEnd(reduceEvent,0,0,0,0); \
00125 PetscFunctionReturn(0);
00126 
00127 #define PROF_PAR_ALLREDUCE_BEGIN \
00128   PetscFunctionBegin; \
00129 PetscLogEventBegin(allReduceEvent,0,0,0,0);
00130 #define PROF_PAR_ALLREDUCE_END \
00131   PetscLogEventEnd(allReduceEvent,0,0,0,0); \
00132 PetscFunctionReturn(0);
00133 
00134 #define PROF_PAR_ALL2ALL_BEGIN \
00135   PetscFunctionBegin; \
00136 PetscLogEventBegin(all2AllEvent,0,0,0,0);
00137 #define PROF_PAR_ALL2ALL_END \
00138   PetscLogEventEnd(all2AllEvent,0,0,0,0); \
00139 PetscFunctionReturn(0);
00140 
00141 #define PROF_PAR_ALLGATHER_BEGIN \
00142   PetscFunctionBegin; \
00143 PetscLogEventBegin(allGatherEvent,0,0,0,0);
00144 #define PROF_PAR_ALLGATHER_END \
00145   PetscLogEventEnd(allGatherEvent,0,0,0,0); \
00146 PetscFunctionReturn(0);
00147 
00148 #define PROF_PAR_ALLGATHERV_BEGIN \
00149   PetscFunctionBegin; \
00150 PetscLogEventBegin(allGathervEvent,0,0,0,0);
00151 #define PROF_PAR_ALLGATHERV_END \
00152   PetscLogEventEnd(allGathervEvent,0,0,0,0); \
00153 PetscFunctionReturn(0);
00154 
00155 #define PROF_PAR_ALL2ALLV_SPARSE_BEGIN \
00156   PetscFunctionBegin; \
00157 PetscLogEventBegin(all2AllvSparseEvent,0,0,0,0);
00158 #define PROF_PAR_ALL2ALLV_SPARSE_END         \
00159   PetscLogEventEnd(all2AllvSparseEvent,0,0,0,0); \
00160 PetscFunctionReturn(0);
00161 
00162 #define PROF_PAR_ALL2ALLV_DENSE_BEGIN \
00163   PetscFunctionBegin; \
00164 PetscLogEventBegin(all2AllvDenseEvent,0,0,0,0);
00165 #define PROF_PAR_ALL2ALLV_DENSE_END \
00166   PetscLogEventEnd(all2AllvDenseEvent,0,0,0,0); \
00167 PetscFunctionReturn(0);
00168 
00169 #define PROF_SORT_BEGIN         \
00170   PetscFunctionBegin; \
00171 PetscLogEventBegin(sortEvent,0,0,0,0);
00172 #define PROF_SORT_END         \
00173   PetscLogEventEnd(sortEvent,0,0,0,0); \
00174 PetscFunctionReturn(0);
00175 
00176 #define PROF_REMDUP_BEGIN       \
00177   PetscFunctionBegin; \
00178 PetscLogEventBegin(remdupEvent,0,0,0,0);
00179 #define PROF_REMDUP_END         \
00180   PetscLogEventEnd(remdupEvent,0,0,0,0); \
00181 PetscFunctionReturn(0);
00182 
00183 #define PROF_PARTW_BEGIN        \
00184   PetscFunctionBegin; \
00185 PetscLogEventBegin(partwEvent,0,0,0,0);
00186 #define PROF_PARTW_END         \
00187   PetscLogEventEnd(partwEvent,0,0,0,0); \
00188 PetscFunctionReturn(0);
00189 
00190 #else
00191 
00192 #define PROF_A2AV_WAIT_BEGIN 
00193 #define PROF_SPLIT_COMM_2WAY_BEGIN
00194 #define PROF_SPLIT_COMM_BEGIN
00195 #define PROF_SEARCH_BEGIN
00196 #define PROF_PAR_SCATTER_BEGIN
00197 #define PROF_PAR_SENDRECV_BEGIN 
00198 #define PROF_PAR_BCAST_BEGIN 
00199 #define PROF_PAR_GATHER_BEGIN 
00200 #define PROF_PAR_SCAN_BEGIN 
00201 #define PROF_PAR_REDUCE_BEGIN
00202 #define PROF_PAR_ALLREDUCE_BEGIN
00203 #define PROF_PAR_ALL2ALL_BEGIN
00204 #define PROF_PAR_ALLGATHERV_BEGIN
00205 #define PROF_PAR_ALLGATHER_BEGIN
00206 #define PROF_PAR_ALL2ALLV_SPARSE_BEGIN
00207 #define PROF_PAR_ALL2ALLV_DENSE_BEGIN
00208 #define PROF_PAR_CONCAT_BEGIN
00209 #define PROF_SORT_BEGIN
00210 #define PROF_REMDUP_BEGIN
00211 #define PROF_PARTW_BEGIN
00212 
00213 #define PROF_A2AV_WAIT_END  
00214 #define PROF_SPLIT_COMM_2WAY_END return 1; 
00215 #define PROF_SPLIT_COMM_END return 1; 
00216 #define PROF_SEARCH_END return 1; 
00217 #define PROF_PAR_SCATTER_END return 1; 
00218 #define PROF_PAR_SENDRECV_END return 1; 
00219 #define PROF_PAR_BCAST_END return 1; 
00220 #define PROF_PAR_GATHER_END return 1; 
00221 #define PROF_PAR_SCAN_END return 1; 
00222 #define PROF_PAR_REDUCE_END return 1; 
00223 #define PROF_PAR_ALLREDUCE_END return 1; 
00224 #define PROF_PAR_ALL2ALL_END return 1; 
00225 #define PROF_PAR_ALLGATHERV_END return 1; 
00226 #define PROF_PAR_ALLGATHER_END return 1; 
00227 #define PROF_PAR_ALL2ALLV_SPARSE_END return 1; 
00228 #define PROF_PAR_ALL2ALLV_DENSE_END return 1; 
00229 #define PROF_PAR_CONCAT_END return 1; 
00230 #define PROF_SORT_END return 1; 
00231 #define PROF_REMDUP_END return 1;
00232 #define PROF_PARTW_END return 1; 
00233 
00234 #endif
00235 
00242 namespace par {
00243 
00244   template <typename T>
00245     int Mpi_Issend(T* buf, int count, int dest, int tag, MPI_Comm comm, MPI_Request* request);
00246 
00247   template <typename T>
00248     int Mpi_Recv(T* buf, int count, int source, int tag, MPI_Comm comm, MPI_Status* status);
00249 
00250   template <typename T>
00251     int Mpi_Irecv(T* buf, int count, int source, int tag, MPI_Comm comm, MPI_Request* request);
00252 
00256   template <typename T>
00257     int Mpi_Gather( T* sendBuffer, T* recvBuffer, int count, int root, MPI_Comm comm);
00258 
00262   template <typename T, typename S>
00263     int Mpi_Sendrecv( T* sendBuf, int sendCount, int dest, int sendTag,
00264         S* recvBuf, int recvCount, int source, int recvTag,
00265         MPI_Comm comm, MPI_Status* status);
00266 
00270   template <typename T>
00271     int Mpi_Bcast( T* buffer, int count, int root, MPI_Comm comm);
00272 
00276   template <typename T>
00277     int Mpi_Scan( T* sendbuf, T* recvbuf, int count, MPI_Op op, MPI_Comm comm);
00278 
00282   template <typename T>
00283     int Mpi_Reduce( T* sendbuf, T* recvbuf, int count, MPI_Op op, int root, MPI_Comm comm);
00284 
00288   template <typename T> 
00289     int Mpi_Allreduce( T* sendbuf, T* recvbuf, int count, MPI_Op op, MPI_Comm comm);
00290 
00294   template <typename T>
00295     int Mpi_Alltoall(T* sendbuf, T* recvbuf, int count, MPI_Comm comm); 
00296 
00300   template <typename T>
00301     int Mpi_Allgatherv(T* sendbuf, int sendcount, T* recvbuf,
00302         int* recvcounts, int* displs, MPI_Comm comm);
00303 
00307   template <typename T>
00308     int Mpi_Allgather(T* sendbuf, T* recvbuf, int count, MPI_Comm comm);
00309 
00313   template <typename T>
00314     int Mpi_Alltoallv_sparse(T* sendbuf, int* sendcnts, int* sdispls, 
00315         T* recvbuf, int* recvcnts, int* rdispls, MPI_Comm comm);
00316 
00320   template <typename T>
00321     int Mpi_Alltoallv_dense(T* sendbuf, int* sendcnts, int* sdispls, 
00322         T* recvbuf, int* recvcnts, int* rdispls, MPI_Comm comm);
00323 
00335   template <typename T> 
00336     int scatterValues(std::vector<T> & in, std::vector<T> & out, 
00337         DendroIntL outSz, MPI_Comm comm );
00338 
00352   template <typename T>
00353     int maxLowerBound(const std::vector<T> & keys, const std::vector<T> & searchList,
00354         std::vector<T> & results, MPI_Comm comm);
00355 
00356   template<typename T>
00357     unsigned int defaultWeight(const T *a);
00358 
00370   template<typename T>
00371     int partitionW(std::vector<T>& vec,
00372         unsigned int (*getWeight)(const T *), MPI_Comm comm);
00373 
00386   template<typename T>
00387     int concatenate(std::vector<T> & listA,
00388         std::vector<T> & listB, MPI_Comm comm);
00389 
00405   template<typename T>
00406     int sampleSort(std::vector<T>& in, std::vector<T> & out, MPI_Comm comm); 
00407 
00416   template<typename T>
00417     int removeDuplicates(std::vector<T>& nodes, bool isSorted, MPI_Comm comm);
00418 
00428   int splitComm2way(bool iAmEmpty, MPI_Comm* new_comm, MPI_Comm orig_comm);
00429 
00439   int splitComm2way(const bool* isEmptyList, MPI_Comm* new_comm, MPI_Comm orig_comm);
00440 
00441   /*
00442      @author Rahul Sampath
00443      @brief Splits a communication group into two, processors with 
00444      ranks less than splittingRank form one group and the other 
00445      processors form the second group. Both the groups are sorted in 
00446      the ascending order of their ranks in the old comm.
00447      @param splittingRank The rank used for splitting the communicator
00448      @param orig_comm    The comm group that needs to be split.
00449      @param new_comm     The new comm group.
00450      */
00451   int splitCommUsingSplittingRank(int splittingRank, MPI_Comm* new_comm, MPI_Comm orig_comm);
00452 
00462   unsigned int splitCommBinary( MPI_Comm orig_comm, MPI_Comm* new_comm);
00463 
00464 
00473   unsigned int splitCommBinaryNoFlip( MPI_Comm orig_comm, MPI_Comm* new_comm);
00474 
00500   template <typename T>
00501     void MergeLists( std::vector<T> &listA, std::vector<T> &listB, int KEEP_WHAT) ;
00502 
00511   template <typename T>
00512     void MergeSplit( std::vector<T> &local_list, int which_keys, int partner, MPI_Comm  comm);
00513 
00517   template <typename T>
00518     void Par_bitonic_sort_incr( std::vector<T> &local_list, int proc_set_size, MPI_Comm  comm );
00519 
00523   template <typename T>
00524     void Par_bitonic_sort_decr( std::vector<T> &local_list, int proc_set_size, MPI_Comm  comm);
00525 
00529   template <typename T>
00530     void Par_bitonic_merge_incr( std::vector<T> &local_list, int proc_set_size, MPI_Comm  comm );
00531 
00541   template <typename T>
00542     void bitonicSort_binary(std::vector<T> & in, MPI_Comm comm) ;
00543 
00556   template <typename T>
00557     void bitonicSort(std::vector<T> & in, MPI_Comm comm) ;
00558 
00559 }//end namespace
00560 
00561 #include "parUtils.txx"
00562 
00563 #endif
00564 
00565 
00566 
00567 
00568 
00569 
00570 
00571 
00572 
00573 
00574 
00575 
00576 
00577 
00578