Use a thread-specific cache variable (make sure to relink your app before using this).

diff --git a/include/smpi/smpif.h.in b/include/smpi/smpif.h.in
index 915d13e..bbcb661 100644 (file)
--- a/include/smpi/smpif.h.in
+++ b/include/smpi/smpif.h.in
@@ -17,7 +17,7 @@
 #include <smpi/smpi.h>
 #include <f2c.h>
 
 #include <smpi/smpi.h>
 #include <f2c.h>
 

-XBT_PUBLIC(int) smpi_current_rank;
+XBT_PUBLIC_DATA(__thread int) smpi_current_rank;

 
 XBT_PUBLIC(int) smpi_process_argc(void);
 XBT_PUBLIC(int) smpi_process_getarg(integer* index, char* dst, ftnlen len);
 
 XBT_PUBLIC(int) smpi_process_argc(void);
 XBT_PUBLIC(int) smpi_process_getarg(integer* index, char* dst, ftnlen len);

diff --git a/src/smpi/smpi_bench.c b/src/smpi/smpi_bench.c
index 6dcff87..2a38ef7 100644 (file)
--- a/src/smpi/smpi_bench.c
+++ b/src/smpi/smpi_bench.c
@@ -14,7 +14,7 @@ XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_bench, smpi,
 
 xbt_dict_t allocs = NULL;       /* Allocated on first use */
 xbt_dict_t samples = NULL;      /* Allocated on first use */
 
 xbt_dict_t allocs = NULL;       /* Allocated on first use */
 xbt_dict_t samples = NULL;      /* Allocated on first use */

-int smpi_current_rank = 0;      /* Updated after each MPI call */
+__thread int smpi_current_rank = 0;      /* Updated after each MPI call */

 
 typedef struct {
   int count;
 
 typedef struct {
   int count;