A call of PAPI_thread_init() seems necessary

Also, reliable results can only be obtained for threads that are bound
to a specific core (socket at least?), otherwise the FLOP and CPU
counters from different CPU's get mixed.

Since PAPI works now on my IvyBridge system with the newest version I
can do more detailed tests there.

ftimings/papi.c

@@ -20,6 +20,7 @@
 #include <stdlib.h>
 #include <string.h>
 #include <errno.h>
+#include <pthread.h>
 
 #ifdef HAVE_CONFIG_H
 #include "config.h"
@@ -35,6 +36,13 @@ static int cycle_available = 0;
 #ifdef HAVE_LIBPAPI
 #include <papi.h>
 
+#define papi_try(x, on_error) if ((ret = x) < 0) { \
+		fprintf(stderr, "ftimings: %s:%d: PAPI_library_init(%d): %s\n", \
+				__FILE__, __LINE__, PAPI_VER_CURRENT, PAPI_strerror(ret)); \
+		on_error; \
+	}
+
+
 static void papi_uninit(void) {
 	PAPI_shutdown();
 }
@@ -46,7 +54,9 @@ static int papi_init(void) {
 		return papi_available;
 	}
 
+#ifdef _OPENMP
 #pragma omp critical
+#endif
 	do {
 		/* Think about it :) */
 		if (tried_papi_init) {
@@ -63,6 +73,7 @@ static int papi_init(void) {
 			papi_available = 0;
 			break;
 		}
+		papi_try(PAPI_thread_init(pthread_self), papi_available = 0; break)
 
 		if ((ret = PAPI_create_eventset(&event_set)) < 0) {
 			fprintf(stderr, "ftimings: %s:%d PAPI_create_eventset(): %s\n",