Question

我正在尝试将openmp添加到此代码中，因为它需要花费大量时间来执行。无论如何加速这个循环将非常感激。我已经尝试过一个解决方案，但它永远不会在高于0的线程数上执行。

!$OMP PARALLEL do private(pm25_tmp,pm10_tmp,V,aerombin,LVL,ROW,COL,isec)                                                                          
       write(*,*) 'num procs:',omp_get_num_procs()
       write(*,*) 'num of thread:',omp_get_num_threads()
       write(*,*) 'thread number:', OMP_GET_THREAD_NUM()
       do LVL = 1, NLAYS3D
         write(*,*) 'THIS IS THE LVL FOR PSO4', LVL
         DO ROW = 1, NROWS
           DO COL = 1, NCOLS
             pm25_tmp = 1.*PSO4(COL,ROW,LVL)
             pm10_tmp = 0.*PSO4(COL,ROW,LVL)
             if (PSO4(COL,ROW,LVL).gt.1.E-10) then
              call mode2sec(pm25_tmp,pm10_tmp,aerombin,pmtot,
 &        JDATE,JTIME,COL,ROW,LVL)
              DO isec=1,nsections
                V = ngasemis + isec+8*5
                EMIS1_SEC(COL,ROW,LVL,V) = aerombin(6,isec)
                aerombin(6,isec) = 0.0
              END DO
              end if
            end do
          end do
       end do
!$OMP END PARALLEL

这是我正在使用的makefile：

FC = ifort
M3LIB       =  /scratch3/NAGAPE/arl/Barry.Baker/cmaq502/CMAQv5.0.2/lib
icldir      = ${M3LIB}/ioapi_31/Linux2_x86_64ifort

# IOAPI LIBRARIES
IOAPI_INC  = ${M3LIB}/ioapi_31/fixed_src
IOAPI_LIB  = ${M3LIB}/ioapi_31/Linux2_x86_64ifort
NETCDF_LIB = ${M3LIB}/netCDF/Linux2_x86_64ifort
#LIB1   = /usr2/tmp/lib/ioapi/ioapi_22/Linux2_x86pg
#LIB2   = /usr2/tmp/lib/netCDF/Linux

#LIBS = -L$(LIB1) -lioapi -L$(LIB2) -lnetcdf
LIBS = -L$(IOAPI_LIB) -lioapi -L$(NETCDF_LIB) -lnetcdf
EXE  = senex_rwce3_8sec

# Set compiler version dependent flags and LINKTOOL

# Fortran flags

FLAGS2  = -openmp -I$(icldir)
#
#  Object files
#
OBJS  = senex_rwce3_eightsec.o \
       senex_mode2sec_eightsec.o \
       ../integrator/qk15.o   ../integrator/qk41.o   ../integrator/qpsrt.o \
       ../integrator/qag.o    ../integrator/qk21.o   ../integrator/qk51.o \
       ../integrator/r1mach.o ../integrator/qage.o   ../integrator/qk31.o \
       ../integrator/qk61.o   ../integrator/xerror.o

$EXE: $(OBJS)
        $(FC) $(FLAGS2) -o $(EXE) $(OBJS) $(LIBS)

Answer 1

您错过了一些重要信息，例如您使用的编译器以及casely评论过的，您如何编译应用程序。此外，您的OpenMP构造似乎是非法的，因为

  !$OMP PARALLEL DO

应该在DO循环之前出现。

话虽这么说，您可以尝试以下测试，看看您的系统是否可以成功生成OpenMP应用程序。您需要为Fortran编译器调整FC，为Makefile中的Fortran标志调整FFLAGS。请注意，在示例中我提供了添加-fopenmp，因为它告诉gfortran在代码中启用OpenMP（否则，将忽略OpenMP pragma！）。

以下代码的输出在我的旧版Lenovo T400中有以下代码：

# ./test
Serial: num procs:           2  num of threads =            1  thread num =            0
Parallel: num procs:           2  num of threads =            2  thread num =            0
Parallel: num procs:           2  num of threads =            2  thread num =        1

请注意，omp_get_num_procs（）给出了OpenMP检测到的处理器数量。例程omp_get_num_threads（）通知在并行区域内运行的线程数，并在任何并行区域外调用它时返回1。最后，omp_get_thread_num（）返回执行并行区域的线程标识符（从0到omp_get_num_threads（） - 1），如果在并行区域外调用它则返回0。

档案 test.f90

      PROGRAM TEST
      IMPLICIT NONE
      INTEGER I, &
        omp_get_num_procs, &
        omp_get_num_threads, &
        omp_get_thread_num

      write(*,*) 'Serial: num procs:',omp_get_num_procs(), &
        " num of threads = ", omp_get_num_threads(), &
        " thread num = ", omp_get_thread_num()
!$OMP PARALLEL DO 
      DO I = 1, omp_get_num_procs()
      write(*,*) 'Parallel: num procs:',omp_get_num_procs(), &
        " num of threads = ", omp_get_num_threads(), &
        " thread num = ", omp_get_thread_num()
      ENDDO

      END

文件 Makefile

FC = gfortran
FFLAGS = -fopenmp

test: test.f90
    $(FC) $(FFLAGS) $< -o $@

如果可能的话，将openmp应用于此循环

1 个答案: