Question

我想用累积（MPI RMA）写入一个可以通过窗口访问的数组。为了便于检查，我写了每个单元格的数组索引。问题是，开始似乎总是正确的，但最终有时会出现错误的值：而不是其位置值，最后的条目被设置为最终索引。

我的代码：

Integer.class

我定义了一个新的MPI数据类型prec_mpi并使用了三个嵌套循环，因为我需要这个模式用于另一个（更大的）程序。使用整数来写入这个双精度数组也是有意的。

我输出了一个错误的输出，例如如下：

PROGRAM main         
     USE mpi
     IMPLICIT NONE

     REAL*8, PARAMETER                         :: prec = 1d-13
     REAL(kind(prec)),ALLOCATABLE,DIMENSION(:) :: speicher1D
     INTEGER                                   :: stelle, ierror
     INTEGER                                   :: ix,iy,iz,nz,i
     !--------------------------------------------------------------
     !---------------MPI-Variablen----------------------------------
     !--------------------------------------------------------------
     INTEGER                        :: myRank               = 0
     INTEGER                        :: numProcs             = 1
     INTEGER, PARAMETER             :: masterProc           = 0

     INTEGER                        :: s3Win1D,s3Win3D
     INTEGER                        :: startIndex, endIndex
     INTEGER(KIND=MPI_ADDRESS_KIND) :: sizeOfSchritt3D1
     INTEGER(KIND=MPI_ADDRESS_KIND) :: s3TargetDisp
     INTEGER                        :: prec_mpi, sizeOfprec_mpi
     !--------------------------------------------------------------
     !--------------MAIN-PROGRAM------------------------------------
     !--------------------------------------------------------------        


     !----MPI-Initialisierung----
     CALL MPI_INIT(ierror)
     CALL MPI_COMM_RANK(MPI_COMM_WORLD, myRank,ierror)
     CALL MPI_COMM_SIZE(MPI_COMM_WORLD,numProcs, ierror)

     !Create new MPI-datatype
     CALL MPI_TYPE_CREATE_F90_REAL(KIND(prec), MPI_UNDEFINED, prec_mpi,ierror)         
     CALL MPI_TYPE_SIZE(prec_mpi, sizeOfprec_mpi,ierror)

     nz                     = 3
     sizeOfSchritt3D1       = sizeOfprec_mpi*(nz*2+1)**3

     ALLOCATE(speicher1D((nz*2+1)**3))
     !------------MPI-RMA-Initialisierung--------------------------
     CALL MPI_WIN_CREATE(speicher1D, sizeOfSchritt3D1, sizeOfprec_mpi,MPI_INFO_NULL,MPI_COMM_WORLD,s3Win1D, ierror)
     !Determine each proc's area
     startIndex  = (nz*2+1)/numProcs*myRank-nz
     endIndex    = (nz*2+1)/numProcs*(myRank+1)-nz-1
     IF(myRank+1 .EQ. numProcs) THEN
        endIndex = nz
     END IF

     speicher1D          = 0.0d0  
     CALL MPI_BARRIER(MPI_COMM_WORLD,ierror)
     CALL MPI_WIN_FENCE(0,s3Win1D, ierror)
     DO iz=startIndex, endIndex
       DO iy=-nz,nz
         DO ix=-nz,nz
           stelle = 1 + (iz+nz)*(2*nz+1)**2 + (iy+nz)*2*nz+1)+ (ix+nz)
           s3targetDisp = stelle-1
           !WRITE(*,*) "Rank", myRank,"s3targetDisp: ",s3targetDisp, "index: ",stelle
           !1D
           CALL MPI_ACCUMULATE(REAL(stelle,kind(prec)), 1, prec_mpi, masterProc, s3targetDisp,1,prec_mpi, MPI_SUM,s3Win1D, ierror)
         END DO 
       END DO
     END DO     
     CALL MPI_WIN_FENCE(0,s3Win1D, ierror)
     CALL MPI_BARRIER(MPI_COMM_WORLD, ierror)
     DO i = 1,(nz*2+1)**3
       IF(myRank .EQ. masterProc) THEN
         WRITE(*,*) "Index: ",i,", entry: ",speicher1D(i)
       END IF
     END DO
     DEALLOCATE(speicher1D)

     CALL MPI_WIN_FREE(s3Win1D, ierror)
     CALL MPI_FINALIZE(ierror)
  END PROGRAM main

为什么条目会有这样的跳跃（在这种情况下是在索引275之后）？当我用嵌套循环中的[...] Index: 250 , entry: 250.000000000000 Index: 251 , entry: 251.000000000000 Index: 252 , entry: 252.000000000000 Index: 253 , entry: 253.000000000000 Index: 254 , entry: 254.000000000000 Index: 255 , entry: 255.000000000000 Index: 256 , entry: 256.000000000000 Index: 257 , entry: 257.000000000000 Index: 258 , entry: 258.000000000000 Index: 259 , entry: 259.000000000000 Index: 260 , entry: 260.000000000000 Index: 261 , entry: 261.000000000000 Index: 262 , entry: 262.000000000000 Index: 263 , entry: 263.000000000000 Index: 264 , entry: 264.000000000000 Index: 265 , entry: 265.000000000000 Index: 266 , entry: 266.000000000000 Index: 267 , entry: 267.000000000000 Index: 268 , entry: 268.000000000000 Index: 269 , entry: 269.000000000000 Index: 270 , entry: 270.000000000000 Index: 271 , entry: 271.000000000000 Index: 272 , entry: 272.000000000000 Index: 273 , entry: 273.000000000000 Index: 274 , entry: 274.000000000000 Index: 275 , entry: 275.000000000000 Index: 276 , entry: 302.000000000000 Index: 277 , entry: 303.000000000000 Index: 278 , entry: 304.000000000000 Index: 279 , entry: 305.000000000000 Index: 280 , entry: 306.000000000000 Index: 281 , entry: 308.000000000000 Index: 282 , entry: 310.000000000000 Index: 283 , entry: 313.000000000000 Index: 284 , entry: 316.000000000000 Index: 285 , entry: 319.000000000000 Index: 286 , entry: 322.000000000000 Index: 287 , entry: 325.000000000000 Index: 288 , entry: 329.000000000000 Index: 289 , entry: 332.000000000000 Index: 290 , entry: 335.000000000000 Index: 291 , entry: 338.000000000000 Index: 292 , entry: 341.000000000000 Index: 293 , entry: 342.000000000000 Index: 294 , entry: 343.000000000000 Index: 295 , entry: 343.000000000000 Index: 296 , entry: 343.000000000000 Index: 297 , entry: 343.000000000000 Index: 298 , entry: 343.000000000000 Index: 299 , entry: 343.000000000000 Index: 300 , entry: 343.000000000000 [...] Index: 341 , entry: 343.000000000000 Index: 342 , entry: 343.000000000000 Index: 343 , entry: 343.000000000000检查时，stelle和s3TargetDisp的值似乎没问题。也许这是同步的一些问题......但是在嵌套循环中只是累积（没有混合RMA /本地）并且有一个关闭栅栏。

使用MPI RMA写入数组

0 个答案: