Make baseline calculations more accurate for SANDGLASS_REALTICKS.

Perform the baseline calculation directly in the sandglass_bench() macro, so
that the baseline is calculated in the same place (and therefore optimization
level, etc.) as the benchmarked code.

2 files changed, 25 insertions, 18 deletions

diff --git a/src/sandglass.c b/src/sandglass.c
index b9316e5..6db28a3 100644
--- a/src/sandglass.c
+++ b/src/sandglass.c
@@ -189,16 +189,6 @@ sandglass_elapse(sandglass_t *sandglass)
     /* Magical correction for timespec-based grains */
     sandglass->grains += sandglass->adjustment;
 
-  if (sandglass->attributes.resolution == SANDGLASS_REALTICKS) {
-    baseline.attributes.incrementation = SANDGLASS_MONOTONIC;
-    baseline.attributes.resolution     = SANDGLASS_CPUTIME;
-    baseline.loops = sandglass->loops;
-
-    sandglass_bench(&baseline, { });
-    sandglass->grains -= baseline.grains;
-    sandglass->grains /= sandglass->loops;
-  }
-
   return 0;
 }
 
@@ -213,13 +203,6 @@ sandglass_real_gettime(sandglass_t *sandglass)
     case SANDGLASS_MONOTONIC:
       switch (sandglass->attributes.resolution) {
         case SANDGLASS_REALTICKS:
-#ifdef SANDGLASS_TSC
-          sandglass->grains = sandglass_get_tsc();
-          break;
-#else
-          return -1;
-#endif
-
         case SANDGLASS_CPUTIME:
 #ifdef SANDGLASS_TSC
           sandglass->grains = sandglass_get_tsc();
diff --git a/src/sandglass.h b/src/sandglass.h
index 7073fb9..0d5e5e5 100644
--- a/src/sandglass.h
+++ b/src/sandglass.h
@@ -102,8 +102,12 @@ struct sandglass_t
   /* Adjustment to be added for negative (i.e. overflowed) grains counts */
   long adjustment;
 
-  /* For SANDGLASS_TICKS looping support */
+  /* For SANDGLASS_REALTICKS looping support */
   unsigned int i, loops;
+
+  /* A field used by sandglass_bench() to store the overhead of
+     sandglass_begin()/_elapse(), and of looping */
+  long baseline;
 };
 typedef struct sandglass_t sandglass_t;
 
@@ -141,7 +145,23 @@ int sandglass_elapse(sandglass_t *sandglass);
  */
 #define sandglass_bench(sandglass, routine)                                    \
   do {                                                                         \
+    /* Warm up the cache for these functions */                                \
+    sandglass_begin(sandglass);                                                \
+    sandglass_elapse(sandglass);                                               \
+                                                                               \
+    /* Time an empty loop for our baseline */                                  \
+    sandglass_begin(sandglass);                                                \
+    for ((sandglass)->i = 0;                                                   \
+         (sandglass)->i < (sandglass)->loops;                                  \
+         ++(sandglass)->i) {                                                   \
+      SANDGLASS_NO_UNROLL();                                                   \
+    }                                                                          \
+    sandglass_elapse(sandglass);                                               \
+    (sandglass)->baseline = (sandglass)->grains;                               \
+                                                                               \
+    /* Warm up the cache for our routine */                                    \
     routine;                                                                   \
+    /* Time our routine in a loop */                                           \
     sandglass_begin(sandglass);                                                \
     for ((sandglass)->i = 0;                                                   \
          (sandglass)->i < (sandglass)->loops;                                  \
@@ -150,6 +170,10 @@ int sandglass_elapse(sandglass_t *sandglass);
       routine;                                                                 \
     }                                                                          \
     sandglass_elapse(sandglass);                                               \
+                                                                               \
+    /* Subtract the baseline and divide by the loop count */                   \
+    (sandglass)->grains -= (sandglass)->baseline;                              \
+    (sandglass)->grains /= (sandglass)->loops;                                 \
   } while (0)
 
 #ifdef __cplusplus