ROCTx and NVTX Tracing

driver/ecrad_ifs_driver.F90

@@ -70,6 +70,10 @@ program ecrad_ifs_driver
 #ifdef HAVE_NVTX
   use nvtx
 #endif
+#ifdef HAVE_ROCTX
+  use roctx_profiling, only: roctxstartrange, roctxendrange
+  use iso_c_binding, only: c_null_char
+#endif
 
   implicit none
 
@@ -381,6 +385,9 @@ program ecrad_ifs_driver
 #ifdef HAVE_NVTX
      call nvtxStartRange("ecrad_offload")
 #endif
+#ifdef HAVE_ROCTX
+  call roctxStartRange("ecrad_offload"//c_null_char)
+#endif
 
 #if defined(_OPENACC) 
   !$ACC DATA COPYIN(yradiation, yradiation%rad_config, single_level, thermodynamics, &
@@ -420,6 +427,9 @@ program ecrad_ifs_driver
 #ifdef HAVE_NVTX
      call nvtxEndRange
 #endif
+#ifdef HAVE_ROCTX
+  call roctxEndRange
+#endif
 
   ! --------------------------------------------------------
   ! Section 4b: Call radiation_scheme per block
@@ -442,6 +452,9 @@ program ecrad_ifs_driver
 #ifdef HAVE_NVTX
     call nvtxStartRange("ecrad_it")
 #endif
+#ifdef HAVE_ROCTX
+      call roctxStartRange("ecrad_it"//c_null_char)
+#endif
 
       ! Compute number of blocks to process
       nblock = (driver_config%iendcol - driver_config%istartcol &
@@ -493,6 +506,12 @@ program ecrad_ifs_driver
              & )
       end do
       !$OMP END PARALLEL DO
+#ifdef HAVE_NVTX
+         call nvtxEndRange
+#endif
+#ifdef HAVE_ROCTX
+      call roctxEndRange
+#endif
 
   end do
 #ifdef DEBUG
@@ -537,6 +556,9 @@ program ecrad_ifs_driver
 #ifdef HAVE_NVTX
      call nvtxEndRange
 #endif
+#ifdef HAVE_ROCTX
+  call roctxEndRange
+#endif
 
 #ifndef NO_OPENMP
   if (driver_config%nrepeat > driver_config%nwarmup) then

driver/ecrad_ifs_driver_blocked.F90

@@ -72,6 +72,10 @@ program ecrad_ifs_driver
 #ifdef HAVE_NVTX
   use nvtx
 #endif
+#ifdef HAVE_ROCTX
+  use roctx_profiling, only: roctxstartrange, roctxendrange
+  use iso_c_binding, only: c_null_char
+#endif
 
   implicit none
 
@@ -400,6 +404,9 @@ program ecrad_ifs_driver
 #ifdef HAVE_NVTX
      call nvtxStartRange("ecrad_it")
 #endif
+#ifdef HAVE_ROCTX
+    call roctxStartRange("ecrad_it"//c_null_char)
+#endif
 
 #ifdef COPY_ASYNC
 #if defined(OMPGPU)
@@ -568,6 +575,9 @@ program ecrad_ifs_driver
 #ifdef HAVE_NVTX
      call nvtxEndRange
 #endif
+#ifdef HAVE_ROCTX
+     call roctxEndRange
+#endif
 
   end do
 

ifs/radiation_scheme.F90

@@ -93,6 +93,10 @@ SUBROUTINE RADIATION_SCHEME &
 #ifdef HAVE_NVTX
 USE NVTX
 #endif
+#ifdef HAVE_ROCTX
+USE ROCTX_PROFILING, ONLY: ROCTXSTARTRANGE, ROCTXENDRANGE
+USE ISO_C_BINDING, ONLY: C_NULL_CHAR
+#endif
 
 IMPLICIT NONE
 
@@ -321,6 +325,9 @@ SUBROUTINE RADIATION_SCHEME &
 #ifdef HAVE_NVTX
 call nvtxStartRange("allocate")
 #endif
+#ifdef HAVE_ROCTX
+call roctxStartRange("allocate"//c_null_char)
+#endif
 CALL SINGLE_LEVEL%ALLOCATE(KLON, YRERAD%NSW, YRERAD%NLWEMISS, &
      &                     USE_SW_ALBEDO_DIRECT=.TRUE.)
 CALL THERMODYNAMICS%ALLOCATE(KLON, KLEV, USE_H2O_SAT=.TRUE.)
@@ -338,6 +345,9 @@ SUBROUTINE RADIATION_SCHEME &
 #ifdef HAVE_NVTX
 call nvtxEndRange
 #endif
+#ifdef HAVE_ROCTX
+call roctxEndRange
+#endif
 
 #if defined(_OPENACC) || defined(OMPGPU)
 LLACC = .TRUE.
@@ -395,6 +405,9 @@ SUBROUTINE RADIATION_SCHEME &
 #ifdef HAVE_NVTX
 call nvtxStartRange("thermodynamics setup")
 #endif
+#ifdef HAVE_ROCTX
+call roctxStartRange("thermodynamics setup"//c_null_char)
+#endif
 ! Set thermodynamic profiles: simply copy over the half-level
 ! pressure and temperature
 !$ACC PARALLEL DEFAULT(NONE) ASYNC(1) IF(LLACC) &
@@ -460,6 +473,10 @@ SUBROUTINE RADIATION_SCHEME &
 call nvtxEndRange
 call nvtxStartRange("single level setup")
 #endif
+#ifdef HAVE_ROCTX
+call roctxEndRange
+call roctxStartRange("single level setup"//c_null_char)
+#endif
 
 ! Set single-level fileds
 SINGLE_LEVEL%SOLAR_IRRADIANCE              = PSOLAR_IRRADIANCE
@@ -580,6 +597,10 @@ SUBROUTINE RADIATION_SCHEME &
 call nvtxEndRange
 call nvtxStartRange("cloud setup")
 #endif
+#ifdef HAVE_ROCTX
+call roctxEndRange
+call roctxStartRange("cloud setup"//c_null_char)
+#endif
 
 ! Set cloud fields
 !$ACC PARALLEL DEFAULT(PRESENT) ASYNC(1) IF(LLACC)
@@ -719,6 +740,10 @@ SUBROUTINE RADIATION_SCHEME &
 call nvtxEndRange
 call nvtxStartRange("aerosol setup")
 #endif
+#ifdef HAVE_ROCTX
+call roctxEndRange
+call roctxStartRange("aerosol setup"//c_null_char)
+#endif
 
 ! Copy over aerosol mass mixing ratio
 IF (.NOT. LL_USE_TEGEN_AEROSOLS) THEN
@@ -804,6 +829,10 @@ SUBROUTINE RADIATION_SCHEME &
 call nvtxEndRange
 call nvtxStartRange("gas setup")
 #endif
+#ifdef HAVE_ROCTX
+call roctxEndRange
+call roctxStartRange("gas setup"//c_null_char)
+#endif
 
 ! Insert gas mixing ratios (mol mol-1 or kg kg-1)
 CALL GAS%PUT(GAS, IH2O,    IMASSMIXINGRATIO, PQ, LACC=LLACC)
@@ -824,18 +853,27 @@ SUBROUTINE RADIATION_SCHEME &
 #ifdef HAVE_NVTX
 call nvtxEndRange
 #endif
+#ifdef HAVE_ROCTX
+call roctxEndRange
+#endif
 
 ! Call radiation scheme
 #ifdef HAVE_NVTX
 call nvtxStartRange("radiation")
 #endif
+#ifdef HAVE_ROCTX
+call roctxStartRange("radiation"//c_null_char)
+#endif
 
 CALL RADIATION(KLON, KLEV, KIDIA, KFDIA, RAD_CONFIG,&
      &  SINGLE_LEVEL, THERMODYNAMICS, GAS, YLCLOUD, AEROSOL, FLUX)
 
 #ifdef HAVE_NVTX
 call nvtxEndRange
 #endif
+#ifdef HAVE_ROCTX
+call roctxEndRange
+#endif
 
 ! Check fluxes are within physical bounds
 IF (YRERAD%NDUMPBADINPUTS /= 0 &
@@ -919,6 +957,9 @@ SUBROUTINE RADIATION_SCHEME &
 #ifdef HAVE_NVTX
 call nvtxStartRange("compute fluxes")
 #endif
+#ifdef HAVE_ROCTX
+call roctxStartRange("compute fluxes"//c_null_char)
+#endif
 ! Compute required output fluxes
 ! First the net fluxes
 !$ACC PARALLEL DEFAULT(PRESENT) ASYNC(1) IF(LLACC)
@@ -1020,6 +1061,10 @@ SUBROUTINE RADIATION_SCHEME &
 call nvtxEndRange
 call nvtxStartRange("cleanup")
 #endif
+#ifdef HAVE_ROCTX
+call roctxEndRange
+call roctxStartRange("cleanup"//c_null_char)
+#endif
 
 !$ACC END DATA
 !$OMP TARGET EXIT DATA MAP(DELETE:ZRE_LIQUID_UM, ZRE_ICE_UM, ZDECORR_LEN_KM,ZLAYER_MASS)
@@ -1047,6 +1092,9 @@ SUBROUTINE RADIATION_SCHEME &
 #ifdef HAVE_NVTX
 call nvtxEndRange
 #endif
+#ifdef HAVE_ROCTX
+call roctxEndRange
+#endif
 
 END ASSOCIATE