Merge pull request #73 from antoine-morvan/am/omp_num_thread_update

Add semantic for reading thread count from $OMP_NUM_THREADS

README.md

@@ -253,6 +253,9 @@ The different prototype variants of the dwarf create different binaries that
 all behave similarly. The basic three arguments define (in this order):
 
 - Number of OpenMP threads
+  - 1 : single thread mode, skip multithread MPI init; default value;
+  - 2 or higher : force OpenMP thread count, enables multithread MPI;
+  - 0 or negative : read OMP_NUM_THREADS variable if present or defaults to CPU count (`omp_get_max_threads()`);
 - Size of overall working set in columns
 - Block size (NPROMA) in columns
 

src/cloudsc_c/dwarf_cloudsc.c

@@ -33,6 +33,14 @@ int main( int argc, char *argv[] ) {
     omp_threads = atoi( argv[1] );
     ngptot      = atoi( argv[2] );
     nproma      = atoi( argv[3] );
+    if (omp_threads <= 0) {
+#ifdef _OPENMP
+      omp_threads = omp_get_max_threads();
+#else
+      // if arg is 0 or negative, and OpenMP disabled; defaults to 1
+      omp_threads = 1;
+#endif
+    }
     cloudsc_driver(omp_threads, ngptot, nproma);
   }
   else {

src/cloudsc_cuda/dwarf_cloudsc.cpp

@@ -33,6 +33,14 @@ int main( int argc, char *argv[] ) {
     omp_threads = atoi( argv[1] );
     ngptot      = atoi( argv[2] );
     nproma      = atoi( argv[3] );
+    if (omp_threads <= 0) {
+#ifdef _OPENMP
+      omp_threads = omp_get_max_threads();
+#else
+      // if arg is 0 or negative, and OpenMP disabled; defaults to 1
+      omp_threads = 1;
+#endif
+    }
     cloudsc_driver(omp_threads, ngptot, nproma);
   }
   else {

src/cloudsc_fortran/dwarf_cloudsc.F90

@@ -19,6 +19,10 @@ PROGRAM DWARF_CLOUDSC
 USE YOMCST   , ONLY : YRCST 
 USE YOETHF   , ONLY : YRTHF
 
+#ifdef _OPENMP
+USE OMP_LIB
+#endif
+
 IMPLICIT NONE
 
 CHARACTER(LEN=20) :: CLARG
@@ -45,8 +49,16 @@ PROGRAM DWARF_CLOUDSC
 
 ! Get the number of OpenMP threads to use for the benchmark
 if (IARGS >= 1) then
-   CALL GET_COMMAND_ARGUMENT(1, CLARG, LENARG)
-   READ(CLARG(1:LENARG),*) NUMOMP
+  CALL GET_COMMAND_ARGUMENT(1, CLARG, LENARG)
+  READ(CLARG(1:LENARG),*) NUMOMP
+  if (NUMOMP <= 0) then
+#ifdef _OPENMP
+    NUMOMP = OMP_GET_MAX_THREADS()
+#else
+    ! if arg is 0 or negative, and OpenMP disabled; defaults to 1
+    NUMOMP = 1
+#endif
+  end if
 end if
 
 ! Initialize MPI environment

src/cloudsc_gpu/dwarf_cloudsc_gpu.F90

@@ -46,6 +46,10 @@ PROGRAM DWARF_CLOUDSC
 USE CLOUDSC_DRIVER_GPU_SCC_FIELD_MOD, ONLY: CLOUDSC_DRIVER_GPU_SCC_FIELD
 #endif
 
+#ifdef _OPENMP
+USE OMP_LIB
+#endif
+
 IMPLICIT NONE
 
 CHARACTER(LEN=20) :: CLARG
@@ -71,8 +75,16 @@ PROGRAM DWARF_CLOUDSC
 
 ! Get the number of OpenMP threads to use for the benchmark
 if (IARGS >= 1) then
-   CALL GET_COMMAND_ARGUMENT(1, CLARG, LENARG)
-   READ(CLARG(1:LENARG),*) NUMOMP
+  CALL GET_COMMAND_ARGUMENT(1, CLARG, LENARG)
+  READ(CLARG(1:LENARG),*) NUMOMP
+  if (NUMOMP <= 0) then
+#ifdef _OPENMP
+    NUMOMP = OMP_GET_MAX_THREADS()
+#else
+     ! if arg is 0 or negative, and OpenMP disabled; defaults to 1
+    NUMOMP = 1
+#endif
+  end if
 end if
 
 ! Initialize MPI environment

src/cloudsc_loki/dwarf_cloudsc.F90

@@ -19,6 +19,10 @@ PROGRAM DWARF_CLOUDSC
 #endif
 USE EC_PMON_MOD, ONLY: EC_PMON
 
+#ifdef _OPENMP
+USE OMP_LIB
+#endif
+
 IMPLICIT NONE
 
 CHARACTER(LEN=20) :: CLARG
@@ -44,8 +48,16 @@ PROGRAM DWARF_CLOUDSC
 
 ! Get the number of OpenMP threads to use for the benchmark
 if (IARGS >= 1) then
-   CALL GET_COMMAND_ARGUMENT(1, CLARG, LENARG)
-   READ(CLARG(1:LENARG),*) NUMOMP
+  CALL GET_COMMAND_ARGUMENT(1, CLARG, LENARG)
+  READ(CLARG(1:LENARG),*) NUMOMP
+  if (NUMOMP <= 0) then
+#ifdef _OPENMP
+    NUMOMP = OMP_GET_MAX_THREADS()
+#else
+    ! if arg is 0 or negative, and OpenMP disabled; defaults to 1
+    NUMOMP = 1
+#endif
+  end if
 end if
 
 ! Initialize MPI environment