Question

当进程数P = 1,2,4,8和16时，运行以下3D复数FFT Fortran MPI程序。但是，当P = 12时它会失败，并且在调用函数fftw_mpi_plan_dft_3d的地方出现以下错误

“程序接收信号SIGSEGV：分段错误 - 无效的内存引用。”

为什么这个程序在P = 12时失败？

PROGRAM test
  USE, INTRINSIC :: ISO_C_BINDING

  IMPLICIT NONE
  INCLUDE 'mpif.h'
  INCLUDE 'fftw3-mpi.f03'

  INTEGER :: ier
  integer(C_INTPTR_T), parameter :: L = 256
  integer(C_INTPTR_T), parameter :: N = 256
  integer(C_INTPTR_T), parameter :: M = 48
  type(C_PTR) :: plan, cdata
  complex(C_DOUBLE_COMPLEX), pointer :: data(:,:,:)
  integer(C_INTPTR_T) :: i, j, alloc_local, local_M, local_j_offset

  CALL MPI_INIT(ier)

  CALL fftw_mpi_init

  alloc_local = fftw_mpi_local_size_3d(M, N, L, MPI_COMM_WORLD, local_M, local_j_offset)
  cdata = fftw_alloc_complex(alloc_local)
  CALL C_F_POINTER(cdata, data, [L, N, local_M])

  plan = fftw_mpi_plan_dft_3d(M, N, L, data, data, MPI_COMM_WORLD, FFTW_FORWARD, FFTW_ESTIMATE)

  CALL fftw_destroy_plan(plan)
  CALL fftw_free(cdata)

  CALL MPI_FINALIZE(ier)
  STOP

END PROGRAM test

Answer 1

我在使用cray-mpich 7.1.3的Cray XE6上看到了fftw_mpi_plan_dft_c2r的相同问题。如果使用gcc，Cray cc或Intel icc编译fftw并不重要。仅针对阵列维度和处理器数量的某些组合发生分段故障。导致segfault的一个组合是128个进程的数组大小1920,3840,3840。与原始帖子一样，我使用了FFTW_ESTIMATE，因此规划器不会进行任何转换。堆栈跟踪如下。

#0  0x00000000008f26f0 in ?? ()
#1  0x00000000004179bb in fftw_plan_destroy_internal (ego=0x8e2af0) at plan.c:49
#2  0x000000000040d755 in mkplan (ego_=0x8def60, p_=0x8f1440, plnr=0x8d6110)
at transpose-pairwise.c:467
#3  0x0000000000418a07 in invoke_solver (ego=0x8d6110, p=0x8f1440, s=0x8def60,
nflags=0x7fffffff4fc0) at planner.c:486
#4  0x0000000000418b9c in search0 (ego=0x8d6110, p=0x8f1440, slvndx=0x7fffffff4fcc,
flagsp=0x7fffffff4fc0) at planner.c:529
#5  0x0000000000418db8 in search (ego=0x8d6110, p=0x8f1440, slvndx=0x7fffffff4fcc,
flagsp=0x7fffffff4fc0) at planner.c:600
#6  0x00000000004191d8 in mkplan (ego=0x8d6110, p=0x8f1440) at planner.c:711
#7  0x0000000000419d35 in fftw_mkplan_d (ego=0x8d6110, p=0x8f1440) at planner.c:970
#8  0x0000000000411d47 in mkplan (ego_=0x8df140, p_=0x8df930, plnr=0x8d6110)
at dft-rank1-bigvec.c:165
#9  0x0000000000418a07 in invoke_solver (ego=0x8d6110, p=0x8df930, s=0x8df140,
nflags=0x7fffffff5290) at planner.c:486
#10 0x0000000000418b9c in search0 (ego=0x8d6110, p=0x8df930, slvndx=0x7fffffff529c,
flagsp=0x7fffffff5290) at planner.c:529
#11 0x0000000000418db8 in search (ego=0x8d6110, p=0x8df930, slvndx=0x7fffffff529c,
flagsp=0x7fffffff5290) at planner.c:600
#12 0x00000000004191d8 in mkplan (ego=0x8d6110, p=0x8df930) at planner.c:711
#13 0x0000000000419d35 in fftw_mkplan_d (ego=0x8d6110, p=0x8df930) at planner.c:970
#14 0x00000000004157b1 in mkplan (ego_=0x8df4a0, p_=0x8df650, plnr=0x8d6110)
at rdft2-rank-geq2.c:179
#15 0x0000000000418a07 in invoke_solver (ego=0x8d6110, p=0x8df650, s=0x8df4a0,
nflags=0x7fffffff5550) at planner.c:486
#16 0x0000000000418b9c in search0 (ego=0x8d6110, p=0x8df650, slvndx=0x7fffffff555c,
flagsp=0x7fffffff5550) at planner.c:529
#17 0x0000000000418db8 in search (ego=0x8d6110, p=0x8df650, slvndx=0x7fffffff555c,
flagsp=0x7fffffff5550) at planner.c:600
#18 0x00000000004191d8 in mkplan (ego=0x8d6110, p=0x8df650) at placner.c:711
#19 0x000000000041d1e2 in mkplan0 (plnr=0x8d6110, flags=64, prb=0x8df650, hash_info=0,
wisdom_state=WISDOM_NORMAL) at apiplan.c:34
#20 0x000000000041d229 in mkplan (plnr=0x8d6110, flags=64, prb=0x8df650, hash_info=0)
at apiplan.c:48
#21 0x000000000041d438 in fftw_mkapiplan (sign=0, flags=0, prb=0x8df650)
at apiplan.c:111
#22 0x000000000040a385 in plan_guru_rdft2 (rnk=3, dims0=0x8df540, howmany=1,
r=0x2aab74c21010, c=0x2aab73000010, comm=1140850688, kind=HC2R00, flags=64)
at api.c:784
#23 0x000000000040a542 in fftw_mpi_plan_many_dft_c2r (rnk=3, n=0x7fffffff5830,
howmany=1, iblock=0, oblock=0, in=0x2aab73000010, out=0x2aab74c21010,
comm=1140850688, flags=64) at api.c:831

fftw_mpi_plan_dft_3d因Segmentation故障而失败

1 个答案: