C $Header: /u/gcmpack/MITgcm/pkg/atm_ocn_coupler/CPP_EEOPTIONS.h,v 1.5 2009/09/14 18:24:51 jmc Exp $ C $Name: $ CBOP C !ROUTINE: CPP_EEOPTIONS.h C !INTERFACE: C include "CPP_EEOPTIONS.h" C C !DESCRIPTION: C *==========================================================* C | CPP\_EEOPTIONS.h | C *==========================================================* C | C preprocessor "execution environment" supporting | C | flags. Use this file to set flags controlling the | C | execution environment in which a model runs - as opposed | C | to the dynamical problem the model solves. | C | Note: Many options are implemented with both compile time| C | and run-time switches. This allows options to be | C | removed altogether, made optional at run-time or | C | to be permanently enabled. This convention helps | C | with the data-dependence analysis performed by the | C | adjoint model compiler. This data dependency | C | analysis can be upset by runtime switches that it | C | is unable to recoginise as being fixed for the | C | duration of an integration. | C | A reasonable way to use these flags is to | C | set all options as selectable at runtime but then | C | once an experimental configuration has been | C | identified, rebuild the code with the appropriate | C | options set at compile time. | C *==========================================================* CEOP #ifndef _CPP_EEOPTIONS_H_ #define _CPP_EEOPTIONS_H_ C In general the following convention applies: C ALLOW - indicates an feature will be included but it may C CAN have a run-time flag to allow it to be switched C on and off. C If ALLOW or CAN directives are "undef'd" this generally C means that the feature will not be available i.e. it C will not be included in the compiled code and so no C run-time option to use the feature will be available. C C ALWAYS - indicates the choice will be fixed at compile time C so no run-time option will be present C-- Flag used to indicate whether Fortran formatted write C and read are threadsafe. On SGI the routines can be thread C safe, on Sun it is not possible - if you are unsure then C undef this option. #undef FMTFTN_IO_THREAD_SAFE C-- Flag used to indicate whether Binary write to Local file (i.e., C a different file for each tile) and read are thread-safe. #undef LOCBIN_IO_THREAD_SAFE C-- Flag to turn off the writing of error message to ioUnit zero #undef DISABLE_WRITE_TO_UNIT_ZERO C-- Flag to turn on checking for errors from all threads and procs C (calling S/R STOP_IF_ERROR) before stopping. #define USE_ERROR_STOP C-- Flag turns off MPI_SEND ready_to_receive polling in the C gather_* subroutines to speed up integrations. #undef DISABLE_MPI_READY_TO_RECEIVE C-- Control MPI based parallel processing CXXX We no longer select the use of MPI via this file (CPP_EEOPTIONS.h) CXXX To use MPI, use an appropriate genmake2 options file or use CXXX genmake2 -mpi . CXXX #undef ALLOW_USE_MPI CXXX #undef ALWAYS_USE_MPI C-- Control use of communication that might overlap computation. C Under MPI selects/deselects "non-blocking" sends and receives. #define ALLOW_ASYNC_COMMUNICATION #undef ALLOW_ASYNC_COMMUNICATION #undef ALWAYS_USE_ASYNC_COMMUNICATION C-- Control use of communication that is atomic to computation. C Under MPI selects/deselects "blocking" sends and receives. #define ALLOW_SYNC_COMMUNICATION #undef ALWAYS_USE_SYNC_COMMUNICATION C-- Control storage of floating point operands C On many systems it improves performance only to use C 8-byte precision for time stepped variables. C Constant in time terms ( geometric factors etc.. ) C can use 4-byte precision, reducing memory utilisation and C boosting performance because of a smaller working C set size. However, on vector CRAY systems this degrades C performance. #define REAL4_IS_SLOW C-- Control use of "double" precision constants. C Use D0 where it means REAL*8 but not where it means REAL*16 #define D0 d0 C-- Control XY periodicity in processor to grid mappings C Note: Model code does not need to know whether a domain is C periodic because it has overlap regions for every box. C Model assume that these values have been C filled in some way. #undef ALWAYS_PREVENT_X_PERIODICITY #undef ALWAYS_PREVENT_Y_PERIODICITY #define CAN_PREVENT_X_PERIODICITY #define CAN_PREVENT_Y_PERIODICITY C-- Alternative formulation of BYTESWAP, faster than C compiler flag -byteswapio on the Altix. #undef FAST_BYTESWAP C-- Alternative way of doing global sum without MPI allreduce call C but instead, explicit MPI send & recv calls. #undef GLOBAL_SUM_SEND_RECV C-- Alternative way of doing global sum on a single CPU C to eliminate tiling-dependent roundoff errors. C Note: This is slow. #undef CG2D_SINGLECPU_SUM C--- Taken from file "CPP_EEMACROS.h": C Flag used to indicate which flavour of multi-threading C compiler directives to use. Only set one of these. C USE_SOLARIS_THREADING - Takes directives for SUN Workshop C compiler. C USE_KAP_THREADING - Takes directives for Kuck and C Associates multi-threading compiler C ( used on Digital platforms ). C USE_IRIX_THREADING - Takes directives for SGI MIPS C Pro Fortran compiler. C USE_EXEMPLAR_THREADING - Takes directives for HP SPP series C compiler. C USE_C90_THREADING - Takes directives for CRAY/SGI C90 C system F90 compiler. #ifdef TARGET_SUN #define USE_SOLARIS_THREADING #endif #ifdef TARGET_DEC #define USE_KAP_THREADING #endif #ifdef TARGET_SGI #define USE_IRIX_THREADING #endif #ifdef TARGET_HP #define USE_EXEMPLAR_THREADING #endif #ifdef TARGET_CRAY_VECTOR #define USE_C90_THREADING #endif C-- Define the mapping for the _BARRIER macro C On some systems low-level hardware support can be accessed through C compiler directives here. #define _BARRIER CALL BARRIER(myThid) C-- Define the mapping for the BEGIN_CRIT() and END_CRIT() macros. C On some systems we simply execute this section only using the C master thread i.e. its not really a critical section. We can C do this because we do not use critical sections in any critical C sections of our code! #define _BEGIN_CRIT(a) _BEGIN_MASTER(a) #define _END_CRIT(a) _END_MASTER(a) C-- Define the mapping for the BEGIN_MASTER_SECTION() and C END_MASTER_SECTION() macros. These are generally implemented by C simply choosing a particular thread to be "the master" and have C it alone execute the BEGIN_MASTER..., END_MASTER.. sections. #define _BEGIN_MASTER(a) IF ( a .EQ. 1 ) THEN #define _END_MASTER(a) ENDIF C- Note: global_sum/max macros were used to switch to JAM routines (obsolete); C in addition, since only the R4 & R8 S/R are coded, GLOBAL RS & RL macros C enable to call the corresponding R4 or R8 S/R. #ifdef REAL4_IS_SLOW #define _RS Real*8 #define RS_IS_REAL8 #define _GLOBAL_SUM_RS(a,b) CALL GLOBAL_SUM_R8( a, b ) #define _GLOBAL_MAX_RS(a,b) CALL GLOBAL_MAX_R8( a, b ) #define _MPI_TYPE_RS MPI_DOUBLE_PRECISION #else /* REAL4_IS_SLOW */ #define _RS Real*4 #define RS_IS_REAL4 #define _GLOBAL_SUM_RS(a,b) CALL GLOBAL_SUM_R4( a, b ) #define _GLOBAL_MAX_RS(a,b) CALL GLOBAL_MAX_R4( a, b ) #define _MPI_TYPE_RS MPI_REAL #endif /* REAL4_IS_SLOW */ #define _RL Real*8 #define RL_IS_REAL8 #define _GLOBAL_SUM_RL(a,b) CALL GLOBAL_SUM_R8( a, b ) #define _GLOBAL_MAX_RL(a,b) CALL GLOBAL_MAX_R8( a, b ) #define _MPI_TYPE_RL MPI_DOUBLE_PRECISION C- Note: a) exch macros were used to switch to JAM routines (obsolete) C b) exch R4 & R8 macros are not practically used ; if needed, C will directly call the corrresponding S/R. #define _EXCH_XY_RS(a,b) CALL EXCH_XY_RS ( a, b ) #define _EXCH_XY_RL(a,b) CALL EXCH_XY_RL ( a, b ) #define _EXCH_XYZ_RS(a,b) CALL EXCH_XYZ_RS ( a, b ) #define _EXCH_XYZ_RL(a,b) CALL EXCH_XYZ_RL ( a, b ) C-- Control use of "double" precision constants. C Use D0 where it means REAL*8 but not where it means REAL*16 #define D0 d0 #ifdef REAL_D0_IS_16BYTES #define D0 #endif C-- Substitue for 1.D variables C Sun compilers do not use 8-byte precision for literals C unless .Dnn is specified. CRAY vector machines use 16-byte C precision when they see .Dnn which runs very slowly! #ifdef REAL_D0_IS_16BYTES #define _d #define _F64( a ) a #endif #ifndef REAL_D0_IS_16BYTES #define _d D #define _F64( a ) DFLOAT( a ) #endif #endif /* _CPP_EEOPTIONS_H_ */