Merge pull request #285 from bosilca/master

Reenable high accuracy timers

ompi/mpi/c/wtick.c

@@ -40,7 +40,9 @@ double MPI_Wtick(void)
 {
     OPAL_CR_NOOP_PROGRESS();
 
-#if OPAL_TIMER_USEC_NATIVE
+#if OPAL_TIMER_CYCLE_NATIVE
+    return opal_timer_base_get_freq();
+#elif OPAL_TIMER_USEC_NATIVE
     return 0.000001;
 #else
     /* Otherwise, we already return usec precision. */

ompi/mpi/c/wtime.c

@@ -40,7 +40,9 @@ double MPI_Wtime(void)
 {
     double wtime;
 
-#if OPAL_TIMER_USEC_NATIVE
+#if OPAL_TIMER_CYCLE_NATIVE
+    wtime = ((double) opal_timer_base_get_cycles()) / opal_timer_base_get_freq();
+#elif OPAL_TIMER_USEC_NATIVE
     wtime = ((double) opal_timer_base_get_usec()) / 1000000.0;
 #else
     /* Fall back to gettimeofday() if we have nothing else */

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

@@ -2,7 +2,7 @@
  * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
  *                         University Research and Technology
  *                         Corporation.  All rights reserved.
- * Copyright (c) 2004-2005 The University of Tennessee and The University
+ * Copyright (c) 2004-2014 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
  * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
@@ -25,29 +25,34 @@ typedef uint64_t opal_timer_t;
 
 #if OPAL_GCC_INLINE_ASSEMBLY
 
-#if 0
+/**
+ * http://www.intel.com/content/www/us/en/intelligent-systems/embedded-systems-training/ia-32-ia-64-benchmark-code-execution-paper.html
+ */
 static inline opal_timer_t
 opal_sys_timer_get_cycles(void)
 {
-    opal_timer_t ret;
-
-    __asm__ __volatile__("rdtsc" : "=A"(ret));
-
-    return ret;
-}
-
+     unsigned a, d;
+#if 0
+     __asm__ __volatile__ ("cpuid\n\t"
+                           "rdtsc\n\t"
+                           : "=a" (a), "=d" (d)
+                           :: "rbx", "rcx");
 #else
-
-static inline opal_timer_t 
-opal_sys_timer_get_cycles(void)
-{
-     unsigned a, d; 
-     __asm__ __volatile__ ("rdtsc" : "=a" (a), "=d" (d));
+     /* 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" (a), "=r" (d)
+                           :: "rax", "rbx", "rcx", "rdx");
+#endif
      return ((opal_timer_t)a) | (((opal_timer_t)d) << 32);
 }
 
-#endif
-
 #define OPAL_HAVE_SYS_TIMER_GET_CYCLES 1
 
 #else

opal/include/opal/sys/ia32/timer.h

@@ -2,7 +2,7 @@
  * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
  *                         University Research and Technology
  *                         Corporation.  All rights reserved.
- * Copyright (c) 2004-2005 The University of Tennessee and The University
+ * Copyright (c) 2004-2014 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
  * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
@@ -30,7 +30,10 @@ opal_sys_timer_get_cycles(void)
 {
     opal_timer_t ret;
 
-    __asm__ __volatile__("rdtsc" : "=A"(ret));
+    __asm__ __volatile__("cpuid\n"
+                         "rdtsc\n"
+                         : "=A"(ret)
+                         :: "ebx", "ecx", "edx");
 
     return ret;
 }

opal/mca/timer/aix/timer_aix.h

@@ -2,7 +2,7 @@
  * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
  *                         University Research and Technology
  *                         Corporation.  All rights reserved.
- * Copyright (c) 2004-2006 The University of Tennessee and The University
+ * Copyright (c) 2004-2014 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
  * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
@@ -39,7 +39,7 @@ opal_timer_base_get_usec()
     retval = (t.tb_high * 1000000) + t.tb_low / 1000;
 
     return retval;
-}    
+}
 
 static inline opal_timer_t
 opal_timer_base_get_cycles()

opal/mca/timer/altix/timer_altix.h

@@ -2,7 +2,7 @@
  * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
  *                         University Research and Technology
  *                         Corporation.  All rights reserved.
- * Copyright (c) 2004-2005 The University of Tennessee and The University
+ * Copyright (c) 2004-2014 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
  * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
@@ -36,7 +36,7 @@ static inline opal_timer_t
 opal_timer_base_get_usec(void)
 {
     return opal_timer_base_get_cycles() / opal_timer_altix_usec_conv;
-}    
+}
 
 
 static inline opal_timer_t

opal/mca/timer/darwin/timer_darwin.h

@@ -2,7 +2,7 @@
  * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
  *                         University Research and Technology
  *                         Corporation.  All rights reserved.
- * Copyright (c) 2004-2005 The University of Tennessee and The University
+ * Copyright (c) 2004-2014 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
  * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
@@ -26,24 +26,39 @@ typedef uint64_t opal_timer_t;
 
 /* frequency in mhz */
 OPAL_DECLSPEC extern opal_timer_t opal_timer_darwin_freq;
+OPAL_DECLSPEC extern mach_timebase_info_data_t opal_timer_darwin_info;
+OPAL_DECLSPEC extern opal_timer_t opal_timer_darwin_bias;
 
-
+/**
+ * Use the pragmatic solution proposed at
+ * http://stackoverflow.com/questions/23378063/how-can-i-use-mach-absolute-time-without-overflowing/23378064#23378064
+ */
 static inline opal_timer_t
 opal_timer_base_get_cycles(void)
 {
-    /* this is basically a wrapper around the "right" assembly to get
-       the tick counter off the PowerPC Time Base.  I believe it's
-       something similar on x86 */
-    return mach_absolute_time();
+    uint64_t now = mach_absolute_time();
+
+    if( opal_timer_darwin_info.denom == 0 ) {
+        (void)mach_timebase_info(&opal_timer_darwin_info);
+        if( opal_timer_darwin_info.denom > 1024 ) {
+            double frac = (double)opal_timer_darwin_info.numer/opal_timer_darwin_info.denom;
+            opal_timer_darwin_info.denom = 1024;
+            opal_timer_darwin_info.numer = opal_timer_darwin_info.denom * frac + 0.5;
+        }
+        opal_timer_darwin_bias = now;
+    }
+    /* this is basically a wrapper around the "right" assembly to convert
+       the tick counter off the PowerPC Time Base into nanos. */
+    return (now - opal_timer_darwin_bias) * opal_timer_darwin_info.numer / opal_timer_darwin_info.denom;
 }
 
 
 static inline opal_timer_t
 opal_timer_base_get_usec(void)
 {
     /* freq is in Hz, so this gives usec */
-    return mach_absolute_time() * 1000000  / opal_timer_darwin_freq;
-}    
+    return opal_timer_base_get_cycles() / 1000;
+}
 
 
 static inline opal_timer_t
@@ -53,9 +68,9 @@ opal_timer_base_get_freq(void)
 }
 
 
-#define OPAL_TIMER_CYCLE_NATIVE 1
+#define OPAL_TIMER_CYCLE_NATIVE 0
 #define OPAL_TIMER_CYCLE_SUPPORTED 1
-#define OPAL_TIMER_USEC_NATIVE 0
+#define OPAL_TIMER_USEC_NATIVE 1
 #define OPAL_TIMER_USEC_SUPPORTED 1
 
 #endif

opal/mca/timer/darwin/timer_darwin_component.c

@@ -2,7 +2,7 @@
  * Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
  *                         University Research and Technology
  *                         Corporation.  All rights reserved.
- * Copyright (c) 2004-2005 The University of Tennessee and The University
+ * Copyright (c) 2004-2014 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
  * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
@@ -20,17 +20,16 @@
 
 #include "opal_config.h"
 
-#include <mach/mach_time.h>
-
 #include "opal/mca/timer/timer.h"
 #include "opal/mca/timer/darwin/timer_darwin.h"
 #include "opal/constants.h"
 
 opal_timer_t opal_timer_darwin_freq;
+mach_timebase_info_data_t opal_timer_darwin_info = {.denom = 0};
+opal_timer_t opal_timer_darwin_bias;
 
 static int opal_timer_darwin_open(void);
 
-
 const opal_timer_base_component_2_0_0_t mca_timer_darwin_component = {
     /* First, the mca_component_t struct containing meta information
        about the component itself */
@@ -53,55 +52,48 @@ const opal_timer_base_component_2_0_0_t mca_timer_darwin_component = {
     },
 };
 
-
+/* mach_timebase_info() returns a fraction that can be multiplied
+   by the difference between two calls to mach_absolute_time() to
+   get the number of nanoseconds that passed between the two
+   calls.  
+
+   On PPC, mach_timebase_info returns numer = 1000000000 and denom
+   = 33333335 (or possibly 25000000, depending on the machine).
+   mach_absolute_time() returns a cycle count from the global
+   clock, which runs at 25 - 33MHz, so dividing the cycle count by
+   the frequency gives you seconds between the interval, then
+   multiplying by 1000000000 gives you nanoseconds.  Of course,
+   you should do the multiply first, then the divide to reduce
+   arithmetic errors due to integer math.  But since we want the
+   least amount of math in the critical path as possible and
+   mach_absolute_time is already a cycle counter, we claim we have
+   native cycle count support and set the frequencey to be the
+   frequencey of the global clock, which is sTBI.denom *
+   (1000000000 / sTBI.numer), which is sTBI.denom * (1 / 1), or
+   sTBI.denom.
+
+   On Intel, mach_timebase_info returns numer = 1 nd denom = 1,
+   meaning that mach_absolute_time() returns some global clock
+   time in nanoseconds.  Because PPC returns a frequency and
+   returning a time in microseconds would still require math in
+   the critical path (a divide, at that), we pretend that the
+   nanosecond timer is instead a cycle counter for a 1GHz clock
+   and that we're returning a cycle count natively.  so sTBI.denom
+   * (1000000000 / sTBI.numer) gives us 1 * (1000000000 / 1), or
+   1000000000, meaning we have a 1GHz clock.
+
+   More generally, since mach_timebase_info() gives the "keys" to
+   transition the return from mach_absolute_time() into
+   nanoseconds, taking the reverse of that and multipling by
+   1000000000 will give you a frequency in cycles / second if you
+   think of mach_absolute_time() always returning a cycle count.
+*/
 int opal_timer_darwin_open(void)
 {
-    mach_timebase_info_data_t sTBI;
-
-    mach_timebase_info(&sTBI);
-
-    /* mach_timebase_info() returns a fraction that can be multiplied
-       by the difference between two calls to mach_absolute_time() to
-       get the number of nanoseconds that passed between the two
-       calls.  
-
-       On PPC, mach_timebase_info returns numer = 1000000000 and denom
-       = 33333335 (or possibly 25000000, depending on the machine).
-       mach_absolute_time() returns a cycle count from the global
-       clock, which runs at 25 - 33MHz, so dividing the cycle count by
-       the frequency gives you seconds between the interval, then
-       multiplying by 1000000000 gives you nanoseconds.  Of course,
-       you should do the multiply first, then the divide to reduce
-       arithmetic errors due to integer math.  But since we want the
-       least amount of math in the critical path as possible and
-       mach_absolute_time is already a cycle counter, we claim we have
-       native cycle count support and set the frequencey to be the
-       frequencey of the global clock, which is sTBI.denom *
-       (1000000000 / sTBI.numer), which is sTBI.denom * (1 / 1), or
-       sTBI.denom.
-
-       On Intel, mach_timebase_info returns numer = 1 nd denom = 1,
-       meaning that mach_absolute_time() returns some global clock
-       time in nanoseconds.  Because PPC returns a frequency and
-       returning a time in microseconds would still require math in
-       the critical path (a divide, at that), we pretend that the
-       nanosecond timer is instead a cycle counter for a 1GHz clock
-       and that we're returning a cycle count natively.  so sTBI.denom
-       * (1000000000 / sTBI.numer) gives us 1 * (1000000000 / 1), or
-       1000000000, meaning we have a 1GHz clock.
-
-       More generally, since mach_timebase_info() gives the "keys" to
-       transition the return from mach_absolute_time() into
-       nanoseconds, taking the reverse of that and multipling by
-       1000000000 will give you a frequency in cycles / second if you
-       think of mach_absolute_time() always returning a cycle count.
+    /* Call the opal_timer_base_get_cycles once to start the enging */
+    (void)opal_timer_base_get_cycles();
 
-       By the way, it's interesting to note that because these are
-       library functions and because of how rosetta works, a PPC
-       binary running under rosetta on an Intel Mac will behave
-       exactly like an Intel binary running on an Intel Mac.
-    */
-    opal_timer_darwin_freq = sTBI.denom * (1000000000 / sTBI.numer);
+    opal_timer_darwin_freq = opal_timer_darwin_info.denom * (1000000000 / opal_timer_darwin_info.numer);
 
     return OPAL_SUCCESS;
 }

opal/mca/timer/linux/timer_linux.h

@@ -2,7 +2,7 @@
  * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
  *                         University Research and Technology
  *                         Corporation.  All rights reserved.
- * Copyright (c) 2004-2005 The University of Tennessee and The University
+ * Copyright (c) 2004-2014 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
  * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
@@ -44,7 +44,7 @@ opal_timer_base_get_usec(void)
 #else
     return 0;
 #endif
-}    
+}
 
 
 static inline opal_timer_t
@@ -56,7 +56,7 @@ opal_timer_base_get_freq(void)
 
 #define OPAL_TIMER_CYCLE_NATIVE OPAL_HAVE_SYS_TIMER_GET_CYCLES
 #define OPAL_TIMER_CYCLE_SUPPORTED OPAL_HAVE_SYS_TIMER_GET_CYCLES
-#define OPAL_TIMER_USEC_NATIVE 0
+#define OPAL_TIMER_USEC_NATIVE OPAL_HAVE_SYS_TIMER_GET_CYCLES
 #define OPAL_TIMER_USEC_SUPPORTED OPAL_HAVE_SYS_TIMER_GET_CYCLES
 
 #endif

opal/mca/timer/solaris/timer_solaris.h

@@ -2,7 +2,7 @@
  * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
  *                         University Research and Technology
  *                         Corporation.  All rights reserved.
- * Copyright (c) 2004-2005 The University of Tennessee and The University
+ * Copyright (c) 2004-2014 The University of Tennessee and The University
  *                         of Tennessee Research Foundation.  All rights
  *                         reserved.
  * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
@@ -35,7 +35,7 @@ opal_timer_base_get_usec(void)
 {
     /* gethrtime returns nanoseconds */
     return gethrtime() / 1000;
-}    
+}
 
 static inline opal_timer_t
 opal_timer_base_get_freq(void)