test/bench: add allreduce latency test

Rename bcast.def to coll_latency.def to add multiple collective latency
tests. Just add allreduce for now.

Author: hzhou

test/mpi/bench/bcast.def renamed to test/mpi/bench/coll_latency.def

@@ -12,10 +12,25 @@ page: bcast, bench_frame
     $(else)
         $call bench_coll, measure_bcast
 
-fncode: measure_bcast(int iter, int root, comm, size, pf_min, pf_max, pf_avg, pf_sigma)
+page: allreduce, bench_frame
+    SIZE_MIN: 4
+    data: buf, size/sizeof(int), MPI_INT
+
+    $call bench_coll, measure_allreduce
+
+#----------------------------------------
+subcode: measure_common
     &call measure_coll_latency, iter
         buf = (char *) gbuf + i * size % (MAX_BUFSIZE - size)
-        MPI_Bcast($(data), root, comm)
+        BLOCK
     $(for:min,max,avg,sigma)
         *pf_$1 = tf_$1
 
+fncode: measure_bcast(int iter, int root, comm, size, pf_min, pf_max, pf_avg, pf_sigma)
+    &call measure_common
+        MPI_Bcast($(data), root, comm)
+
+fncode: measure_allreduce(int iter, comm, size, pf_min, pf_max, pf_avg, pf_sigma)
+    &call measure_common
+        MPI_Allreduce(MPI_IN_PLACE, $(data), MPI_SUM, comm)
+

test/mpi/bench/macros/bench_coll.def

@@ -4,9 +4,14 @@ subcode: bench_coll(measure_func)
         $(set:params=iter, root, comm, size)
     $(else)
         $(set:params=iter, comm, size)
+    $(if:!SIZE_MIN)
+        $(set:SIZE_MIN=0)
+        $(set:SIZE_STEP=size=(size==0)?1:size*2)
+    $(else)
+        $(set:SIZE_STEP=size*=2)
     $if grank == 0
         $call header_coll_latency
-    &call foreach_size
+    $for int size = $(SIZE_MIN); size < $(MAX_BUFSIZE); $(SIZE_STEP)
         $my tf_min, tf_max, tf_avg, tf_sigma
         $(set:MIN_ITER=0.001/tf_max)
         &call coll_warmup