machinery for checking that our OUTER parallel iteration works - before we delete it

[moebius2.git] / parallel.h

diff --git a/parallel.h b/parallel.h
index 3d5082285824cf07bd0ab560fe032d0d0567b616..77025e512202f9a5323146e30abc3ce141bc6d55 100644 (file)
--- a/parallel.h
+++ b/parallel.h
@@ -14,13 +14,25 @@
 /* used anamorphically: section, nsections */
 
 #define OUTER_PERSECTION_BASE(zero,n, sect) \
-  ((zero) + sect * (((n)-(zero) + NSECTIONS-1) / NSECTIONS))
-#define OUTER(v,zero,n, precomp)                                         \
-  for ((v)= OUTER_PERSECTION_BASE((zero),(n), section);                          \
-       precomp,                                                                  \
-       (v) < OUTER_PERSECTION_BASE((zero),(n), section + 1) && (v) < (n); \
+  ((zero) + (sect) * (((n)-(zero) + NSECTIONS-1) / NSECTIONS))
+#define OUTER(v,zero,n, precomp)                                       \
+  for ((v)= OUTER_PERSECTION_BASE((zero),(n), (section));              \
+       precomp,                                                                \
+       (v) < OUTER_PERSECTION_BASE((zero),(n), (section) + 1) && (v) < (n); \
        (v)++)
 
+/*
+ * OUTER is a loop constructor like INNER (see mgraph.h).
+ *
+ * Constraints on its use:
+ *   - must be in exactly one loop of particular function
+ *   - function must not modify anything other than
+ *      its return value (for cost computation functions, COST()) or
+ *      its designated output (for precomputation functions, PRECOMP())
+ *       and in the latter case it may not read other parts of its output
+ *   - function must of course be reentrant
+ */
+
 #define nsections NSECTIONS
 
 typedef void Computation(const struct Vertices *vertices,
@@ -30,5 +42,16 @@ void inparallel(const struct Vertices *vertices,
                Computation *separately,
                Computation *combine,
                size_t secdatasz, void *gendata);
+  /* nsections copies of the computation `separately' are run in parallel
+   * with different values of `section' from 0 to nsections-1.
+   * Each copy is passed the same gendata as passed to inparallel.
+   * Each copy gets its own data block (struct aligned) of size
+   * secdatasz, passed as secdata, uninitialised on entry.
+   * After all the copies have finished, `combine' is invoked
+   * nsections times sequentially, with the same sets of arguments.
+   * inparallel is NOT itself reentrant.
+   */
+
+void inparallel_barrier(void);
 
 #endif /*PARALLEL_H*/