|
| 1 | +! (C) Copyright 1988- ECMWF. |
| 2 | +! |
| 3 | +! This software is licensed under the terms of the Apache Licence Version 2.0 |
| 4 | +! which can be obtained at http://www.apache.org/licenses/LICENSE-2.0. |
| 5 | +! |
| 6 | +! In applying this licence, ECMWF does not waive the privileges and immunities |
| 7 | +! granted to it by virtue of its status as an intergovernmental organisation |
| 8 | +! nor does it submit to any jurisdiction. |
| 9 | + |
| 10 | +MODULE CLOUDSC_DRIVER_FIELD_LOKI_MOD |
| 11 | + USE PARKIND1, ONLY: JPIM, JPIB, JPRB, JPRD |
| 12 | + USE YOMPHYDER, ONLY: STATE_TYPE |
| 13 | + USE YOECLDP, ONLY : NCLV, YRECLDP |
| 14 | + USE CLOUDSC_MPI_MOD, ONLY: NUMPROC, IRANK |
| 15 | + USE TIMER_MOD, ONLY : PERFORMANCE_TIMER, GET_THREAD_NUM |
| 16 | + USE EC_PMON_MOD, ONLY: EC_PMON |
| 17 | + USE CLOUDSC_FIELD_STATE_MOD, ONLY: CLOUDSC_AUX_TYPE, CLOUDSC_FLUX_TYPE, CLOUDSC_STATE_TYPE |
| 18 | + |
| 19 | + USE CLOUDSC_MOD, ONLY : CLOUDSC |
| 20 | + |
| 21 | + IMPLICIT NONE |
| 22 | + |
| 23 | +CONTAINS |
| 24 | + |
| 25 | + SUBROUTINE CLOUDSC_DRIVER_FIELD( & |
| 26 | + & NUMOMP, NPROMA, NLEV, NGPTOT, NGPTOTG, KFLDX, PTSPHY, PAUX, FLUX, & |
| 27 | + & TENDENCY_TMP, TENDENCY_LOC) |
| 28 | + ! Driver routine that invokes the optimized CLAW-based CLOUDSC GPU kernel |
| 29 | + |
| 30 | + USE YOECLDP , ONLY : TECLDP |
| 31 | + USE YOMCST , ONLY : TOMCST |
| 32 | + USE YOETHF , ONLY : TOETHF |
| 33 | + |
| 34 | + INTEGER(KIND=JPIM) ,INTENT(IN) :: NUMOMP, NPROMA, NLEV, NGPTOT, NGPTOTG |
| 35 | + INTEGER(KIND=JPIM) ,INTENT(IN) :: KFLDX |
| 36 | + REAL(KIND=JPRB) ,INTENT(IN) :: PTSPHY ! PHYSICS TIMESTEP |
| 37 | + TYPE(CLOUDSC_AUX_TYPE) ,INTENT(IN) :: PAUX |
| 38 | + TYPE(CLOUDSC_FLUX_TYPE) ,INTENT(IN) :: FLUX |
| 39 | + TYPE(CLOUDSC_STATE_TYPE) ,INTENT(IN) :: TENDENCY_TMP |
| 40 | + TYPE(CLOUDSC_STATE_TYPE) ,INTENT(INOUT) :: TENDENCY_LOC |
| 41 | + |
| 42 | + INTEGER(KIND=JPIM) :: JKGLO,IBL,ICEND, NGPBLKS |
| 43 | + |
| 44 | + TYPE(TECLDP) :: LOCAL_YRECLDP |
| 45 | + |
| 46 | + TYPE(PERFORMANCE_TIMER) :: TIMER |
| 47 | + INTEGER(KIND=JPIM) :: TID ! thread id from 0 .. NUMOMP - 1 |
| 48 | + |
| 49 | + NGPBLKS = (NGPTOT / NPROMA) + MIN(MOD(NGPTOT,NPROMA), 1) |
| 50 | +1003 format(5x,'NUMPROC=',i0,', NUMOMP=',i0,', NGPTOTG=',i0,', NPROMA=',i0,', NGPBLKS=',i0) |
| 51 | + if (irank == 0) then |
| 52 | + write(0,1003) NUMPROC,NUMOMP,NGPTOTG,NPROMA,NGPBLKS |
| 53 | + end if |
| 54 | + |
| 55 | + ! Global timer for the parallel region |
| 56 | + CALL TIMER%START(NUMOMP) |
| 57 | + |
| 58 | + ! Workaround for PGI / OpenACC oddities: |
| 59 | + ! Create a local copy of the parameter struct to ensure they get |
| 60 | + ! moved to the device the in ``acc data`` clause below |
| 61 | + LOCAL_YRECLDP = YRECLDP |
| 62 | + |
| 63 | + !$loki data |
| 64 | + |
| 65 | + !$omp parallel default(shared) private(JKGLO,IBL,ICEND,TID) & |
| 66 | + !$omp& num_threads(NUMOMP) firstprivate(PAUX, FLUX, TENDENCY_TMP, TENDENCY_LOC) |
| 67 | + |
| 68 | + ! Local timer for each thread |
| 69 | + TID = GET_THREAD_NUM() |
| 70 | + CALL TIMER%THREAD_START(TID) |
| 71 | + |
| 72 | + !$omp do schedule(runtime) reduction(+:power_total,power_count) |
| 73 | + DO JKGLO=1,NGPTOT,NPROMA |
| 74 | + IBL=(JKGLO-1)/NPROMA+1 |
| 75 | + ICEND=MIN(NPROMA,NGPTOT-JKGLO+1) |
| 76 | + |
| 77 | + CALL PAUX%UPDATE_VIEW(IBL) |
| 78 | + CALL FLUX%UPDATE_VIEW(IBL) |
| 79 | + CALL TENDENCY_LOC%UPDATE_VIEW(IBL) |
| 80 | + CALL TENDENCY_TMP%UPDATE_VIEW(IBL) |
| 81 | + |
| 82 | + !-- These were uninitialized : meaningful only when we compare error differences |
| 83 | + PAUX%PCOVPTOT = 0.0_JPRB |
| 84 | + TENDENCY_LOC%CLD(:,:,NCLV) = 0.0_JPRB |
| 85 | + |
| 86 | + |
| 87 | + |
| 88 | + CALL CLOUDSC( 1, ICEND, NPROMA, NLEV, & ! These could also be accessed through FIELD_STATE |
| 89 | + & PTSPHY,& |
| 90 | + & PAUX%PT, PAUX%PQ, & |
| 91 | + & TENDENCY_TMP%T, TENDENCY_TMP%Q, TENDENCY_TMP%A, TENDENCY_TMP%CLD, & |
| 92 | + & TENDENCY_LOC%T, TENDENCY_LOC%Q, TENDENCY_LOC%A, TENDENCY_LOC%CLD, & |
| 93 | + & PAUX%PVFA, PAUX%PVFL, PAUX%PVFI, PAUX%PDYNA, PAUX%PDYNL, PAUX%PDYNI, & |
| 94 | + & PAUX%PHRSW, PAUX%PHRLW,& |
| 95 | + & PAUX%PVERVEL, PAUX%PAP, PAUX%PAPH,& |
| 96 | + & PAUX%PLSM, PAUX%LDCUM, PAUX%KTYPE, & |
| 97 | + & PAUX%PLU, PAUX%PLUDE, PAUX%PSNDE, PAUX%PMFU, PAUX%PMFD,& |
| 98 | + !---prognostic fields |
| 99 | + & PAUX%PA,& |
| 100 | + & PAUX%PCLV, & |
| 101 | + & PAUX%PSUPSAT,& |
| 102 | +! -- arrays for aerosol-cloud interactions |
| 103 | +! !! & PQAER, KAER, & |
| 104 | + & PAUX%PLCRIT_AER,PAUX%PICRIT_AER,& |
| 105 | + & PAUX%PRE_ICE,& |
| 106 | + & PAUX%PCCN, PAUX%PNICE,& |
| 107 | + !---diagnostic output |
| 108 | + & PAUX%PCOVPTOT, PAUX%PRAINFRAC_TOPRFZ,& |
| 109 | + !---resulting fluxes |
| 110 | + & FLUX%PFSQLF, FLUX%PFSQIF , FLUX%PFCQNNG, FLUX%PFCQLNG,& |
| 111 | + & FLUX%PFSQRF, FLUX%PFSQSF , FLUX%PFCQRNG, FLUX%PFCQSNG,& |
| 112 | + & FLUX%PFSQLTUR, FLUX%PFSQITUR , & |
| 113 | + & FLUX%PFPLSL, FLUX%PFPLSN, FLUX%PFHPSL, FLUX%PFHPSN, & |
| 114 | + & LOCAL_YRECLDP) |
| 115 | + |
| 116 | +#ifndef CLOUDSC_GPU_TIMING |
| 117 | + ! Log number of columns processed by this thread |
| 118 | + CALL TIMER%THREAD_LOG(TID, IGPC=ICEND) |
| 119 | +#endif |
| 120 | + ENDDO |
| 121 | + |
| 122 | + !-- The "nowait" is here to get correct local timings (tloc) per thread |
| 123 | + ! i.e. we should not wait for slowest thread to finish before measuring tloc |
| 124 | + !$omp end do nowait |
| 125 | + |
| 126 | + CALL TIMER%THREAD_END(TID) |
| 127 | + |
| 128 | + !$omp end parallel |
| 129 | + |
| 130 | + !$loki end data |
| 131 | + |
| 132 | + CALL TIMER%END() |
| 133 | + |
| 134 | +#ifdef CLOUDSC_GPU_TIMING |
| 135 | + ! On GPUs, adding block-level column totals is cumbersome and |
| 136 | + ! error prone, and of little value due to the large number of |
| 137 | + ! processing "thread teams". Instead we register the total here. |
| 138 | + CALL TIMER % THREAD_LOG(TID=TID, IGPC=NGPTOT) |
| 139 | +#endif |
| 140 | + CALL TIMER%PRINT_PERFORMANCE(NPROMA, NGPBLKS, NGPTOT) |
| 141 | + |
| 142 | + |
| 143 | + END SUBROUTINE CLOUDSC_DRIVER_FIELD |
| 144 | + |
| 145 | +END MODULE CLOUDSC_DRIVER_FIELD_LOKI_MOD |
| 146 | + |
0 commit comments