Merge pull request #2926 from Zzzoom/amd64_timer_perf

Improve x86-64 timer performance

Author: hjelmn

opal/include/opal/sys/amd64/timer.h

@@ -31,31 +31,14 @@ typedef uint64_t opal_timer_t;
 
 #if OPAL_GCC_INLINE_ASSEMBLY
 
-/**
- * http://www.intel.com/content/www/us/en/intelligent-systems/embedded-systems-training/ia-32-ia-64-benchmark-code-execution-paper.html
- */
+/* TODO: add AMD mfence version and dispatch at init */
 static inline opal_timer_t
 opal_sys_timer_get_cycles(void)
 {
-     unsigned l, h;
-#if !OPAL_ASSEMBLY_SUPPORTS_RDTSCP
-     __asm__ __volatile__ ("cpuid\n\t"
+     uint32_t l, h;
+     __asm__ __volatile__ ("lfence\n\t"
                            "rdtsc\n\t"
-                           : "=a" (l), "=d" (h)
-                           :: "rbx", "rcx");
-#else
-     /* If we need higher accuracy we should implement the algorithm proposed
-      * on the Intel document referenced above. However, in the context of MPI
-      * this function will be used as the backend for MPI_Wtime and as such
-      * can afford a small inaccuracy.
-      */
-     __asm__ __volatile__ ("rdtscp\n\t"
-                           "mov %%edx, %0\n\t"
-                           "mov %%eax, %1\n\t"
-                           "cpuid\n\t"
-                           : "=r" (h), "=r" (l)
-                           :: "rax", "rbx", "rcx", "rdx");
-#endif
+                           : "=a" (l), "=d" (h));
      return ((opal_timer_t)l) | (((opal_timer_t)h) << 32);
 }
 

opal/runtime/opal_progress.c

@@ -37,6 +37,7 @@
 #include "opal/runtime/opal_params.h"
 
 #define OPAL_PROGRESS_USE_TIMERS (OPAL_TIMER_CYCLE_SUPPORTED || OPAL_TIMER_USEC_SUPPORTED)
+#define OPAL_PROGRESS_ONLY_USEC_NATIVE (OPAL_TIMER_USEC_NATIVE && !OPAL_TIMER_CYCLE_NATIVE)
 
 #if OPAL_ENABLE_DEBUG
 bool opal_progress_debug = false;
@@ -189,11 +190,11 @@ opal_progress(void)
     if( opal_progress_event_flag != 0 ) {
 #if OPAL_HAVE_WORKING_EVENTOPS
 #if OPAL_PROGRESS_USE_TIMERS
-#if OPAL_TIMER_USEC_NATIVE
+#if OPAL_PROGRESS_ONLY_USEC_NATIVE
         opal_timer_t now = opal_timer_base_get_usec();
 #else
         opal_timer_t now = opal_timer_base_get_cycles();
-#endif  /* OPAL_TIMER_USEC_NATIVE */
+#endif  /* OPAL_PROGRESS_ONLY_USEC_NATIVE */
     /* trip the event library if we've reached our tick rate and we are
        enabled */
         if (now - event_progress_last_time > event_progress_delta ) {
@@ -316,7 +317,7 @@ opal_progress_set_event_poll_rate(int polltime)
 
 #if OPAL_PROGRESS_USE_TIMERS
     event_progress_delta = 0;
-#  if OPAL_TIMER_USEC_NATIVE
+#  if OPAL_PROGRESS_ONLY_USEC_NATIVE
     event_progress_last_time = opal_timer_base_get_usec();
 #  else
     event_progress_last_time = opal_timer_base_get_cycles();
@@ -343,7 +344,7 @@ opal_progress_set_event_poll_rate(int polltime)
 #endif
     }
 
-#if OPAL_PROGRESS_USE_TIMERS && !OPAL_TIMER_USEC_NATIVE
+#if OPAL_PROGRESS_USE_TIMERS && !OPAL_PROGRESS_ONLY_USEC_NATIVE
     /*  going to use cycles for counter.  Adjust specified usec into cycles */
     event_progress_delta = event_progress_delta * opal_timer_base_get_freq() / 1000000;
 #endif