Merge pull request #89 from Xilinx-CNS/user/kieranm/ON-16400-rdtscp

ON-16400: switch zf*pingpong to use rdtscp for more accurate latency measurement

Author: kieranm-xilinx

src/tests/zf_apps/zf_timer.h

@@ -9,11 +9,25 @@
 #include <x86intrin.h>
 #include <stdint.h>
 
-inline uint64_t get_frc64_time(void)
+/* Return measure of free running counter, using rdtsc, so may be reordered with other code */
+inline uint64_t get_frc64_time_standard(void)
 {
   return __rdtsc();
 }
 
+/* Return measure of free running counter, using rdtscp, which gives higher accuracy when measuring 
+ * the duration of blocks of code as it can't be reordered with earlier instructions, but has 
+ * slightly higher overhead than get_frc64_time_standard() */
+inline uint64_t get_frc64_time_accurate(void)
+{
+  unsigned int aux;
+  return __rdtscp(&aux);
+}
+
+/* Default to using the more accurate measure as all current uses of this function are using it to
+ * measure ping pong deltas, and hence will benefit from the accuracy */
+#define get_frc64_time() get_frc64_time_accurate()
+
 void init_tsc_frequency(void);
 
 /* Given TSC ticks it will calculate and return number of nano seconds passed for it.