C $Header: /u/gcmpack/MITgcm/eesupp/src/main.F,v 1.27 2013/11/26 00:52:09 jmc Exp $ C $Name: $ CBOI C C !TITLE: WRAPPER CODE SYNOPSIS C !AUTHORS: mitgcm developers ( support@mitgcm.org ) C !AFFILIATION: Massachussetts Institute of Technology C !DATE: C !INTRODUCTION: C Wrapper synopsis and code Routines in the subdirectories under C eesupp/ ( src/ and inc/ ) provide the core framework within which C numerical and ancilliary software of MITgcm operates. The eesupp/ C directories provide a collection of software we call {\bf WRAPPER} C ( ({\bf W}rappable {\bf A}pplication {\bf P}aralell {\bf C P}rogramming {\bf E}nvironment {\bf R}esource). The {bf WRAPPER} C provides a generic bootstrapping capability to start applications C in a manner that allows them to exploit single and C multi-processing environments on all present day hardware C platforms (spanning vector SMP systems to distributed memory and C processing cluster systems). Numerical applications must be coded C to fit within the {\bf WRAPPER}. This entails applications C adopting a particular style for declaring data structures C representing grids and values on grids. The {\bf WRAPPER} C currently provides support for grid point models using a single C global indexing system. This is sufficient for latitude-logitude, C cylindrical, and cartesian coordinate configurations. There is C also limited support for composing grids in which no single, C sructured global index can be defined. At present, this support is C limited to specific configurations of projections of a cube onto C the sphere. C C The main functions supported by the current {\bf WRAPPER} code are C \begin{itemize} C \item program startup and termination including C creation/management of multiple threads and/or processes C \item communication and synchronisatioin operations between C multiple processes and/or threads C \item multi-process input and output operations to disk and to C other applications C \end{itemize} C C Multi-process execution assumes the existence of MPI for process C startup and termination. However, MPI does not have to be used for C performance critical operations. Instead, {\bf WRAPPER} C performance critical parallel primitives are implemented to allow C them to bind to different low-level system software C layers. Bindings exist for using {\bf WRAPPER} with portable C systems such as MPI and UNIX System V IPC memory mapping, as well C bindings for high-performance propreitary systems such as Myrinet C GM software and Compaq IMC memory channel technology. C CEOI C-- Get C preprocessor options #include "PACKAGES_CONFIG.h" #include "CPP_OPTIONS.h" C---+----1----+----2----+----3----+----4----+----5----+----6----+----7-|--+----| CBOP C !ROUTINE: MAIN C !INTERFACE: PROGRAM MAIN C !DESCRIPTION: C *==========================================================* C | PROGRAM MAIN C | o MAIN wrapper for MITgcm UV implementation. C *==========================================================* C | MAIN controls the "execution environment". C | Its main functions are C | 1. call procedure EEBOOT to perform execution environment C | initialisation. C | 2. call procedure THE\_MODEL\_MAIN once for each concurrent C | thread. THE\_MODEL\_MAIN is the user supplied top-level C | routine. C | 3. call procedure EEDIE to perform execution environment C | shutdown. C *==========================================================* C !CALLING SEQUENCE: C C main() C | C |--eeboot() :: WRAPPER initilization C | C |--check_threads() :: Validate multiple thread start up. C | C |--the_model_main() :: Numerical code top-level driver routine C | C |--eedie() :: WRAPPER termination C !USES: IMPLICIT NONE C == Global variables == C Include all the "shared" data here. That means all common C blocks used in the model. On many implementations this is not C necessary but doing this is the safest method. #include "SIZE.h" #include "EEPARAMS.h" #include "EESUPPORT.h" #include "PARAMS.h" #include "GRID.h" #include "DYNVARS.h" #include "FFIELDS.h" #include "SURFACE.h" #ifdef HAVE_SIGREG #include "SIGREG.h" #endif C !LOCAL VARIABLES: C msgBuf :: I/O message buffer C I :: loop counter C myThid :: thread Id number CHARACTER*(MAX_LEN_MBUF) msgBuf INTEGER myThid INTEGER I #ifdef USE_OMP_THREADING INTEGER OMP_GET_THREAD_NUM EXTERNAL OMP_GET_THREAD_NUM #endif CEOP #ifdef USE_GSL_IEEE CALL FGSL_IEEE_ENV_SETUP () #endif C-- Set up the execution environment C EEBOOT loads a execution environment parameter file C ( called "eedata" by default ) and sets variables C accordingly. CALL EEBOOT C-- Trap errors IF ( eeBootError ) THEN fatalError = .TRUE. GOTO 999 ENDIF #ifdef HAVE_SETRLSTK IF (useSETRLSTK) THEN CALL setrlstk ENDIF #endif #ifdef HAVE_SIGREG IF (useSIGREG) THEN i_got_signal = 0 CALL sigreg( i_got_signal ) ENDIF #endif #ifdef HAVE_PTHREADS c IF (usePTHREADS) THEN CALL PTINIT(nThreads) c ELSE #else C-- Start nThreads concurrent threads. C Note: We do a fiddly check here. The check is performed C by CHECK_THREADS. CHECK_THREADS does a count C of all the threads. If after ten seconds it has not C found nThreads threads are running it flags an C error. This traps the case in which the input C parameter nThreads is different from the actual C number of concurrent threads the OS gives us. This C case causes a deadlock if we do not trap it here. #include "MAIN_PDIRECTIVES1.h" DO I=1,nThreads #ifdef USE_OMP_THREADING IF ( OMP_GET_THREAD_NUM() .EQ. I-1 ) THEN #endif myThid = I C-- Do check to see if there are nThreads threads running IF ( .NOT. eeBootError ) THEN CALL CHECK_THREADS( myThid ) ENDIF C-- Invoke nThreads instances of the numerical model IF ( .NOT. eeBootError ) THEN #if (defined (ALLOW_ADMTLM)) CALL ADMTLM_DSVD(myThid) #elif (defined (ALLOW_HESSIAN_CODE)) CALL HESSIAN_MAIN(myThid) #else CALL THE_MODEL_MAIN(myThid) #endif ENDIF C-- Each threads sets flag indicating it is done threadIsComplete(myThid) = .TRUE. IF ( .NOT. eeBootError ) THEN _BARRIER ENDIF #ifdef USE_OMP_THREADING ENDIF #endif ENDDO #include "MAIN_PDIRECTIVES2.h" #endif /* HAVE_PTHREADS */ 999 CONTINUE C-- Shut down execution environment CALL EEDIE C-- Write closedown status IF ( fatalError ) THEN WRITE( msgBuf,'(A)') 'PROGRAM MAIN: ends with fatal Error' CALL PRINT_ERROR( msgBuf, 1 ) WRITE(standardMessageUnit,'(A)') & 'PROGRAM MAIN: ends with fatal Error' STOP 'ABNORMAL END: PROGRAM MAIN' ELSE WRITE(standardMessageUnit,'(A)') & 'PROGRAM MAIN: Execution ended Normally' STOP 'NORMAL END' ENDIF END C---+----1----+----2----+----3----+----4----+----5----+----6----+----7-|--+----| CBOP 0 C !ROUTINE: PTREENTRY C !INTERFACE: SUBROUTINE PTREENTRY( I myThid ) C !DESCRIPTION: C Re-entry point for a pthreads-based threading mechanism. The C intent is to produce a threading hack that will work with gcc/g77. C !USES: IMPLICIT NONE #include "SIZE.h" #include "EEPARAMS.h" #include "EESUPPORT.h" #include "PARAMS.h" #include "GRID.h" #include "DYNVARS.h" #include "FFIELDS.h" #include "SURFACE.h" C !INPUT PARAMETERS: C myThid :: my thread Id number INTEGER myThid CEOP WRITE(*,*) 'myThid = ', myThid CALL CHECK_THREADS( myThid ) c CALL THE_MODEL_MAIN(myThid) threadIsComplete(myThid) = .TRUE. RETURN END C---+----1----+----2----+----3----+----4----+----5----+----6----+----7-|--+----|