Adjust the preRA schedule heuristic

The heuristic are experience based, but using the same threshold for
scheduling the beneifit comparision should be a always choice.

Author: bcheng0127

visa/LocalScheduler/G4_Sched.cpp

@@ -1694,14 +1694,19 @@ bool BB_Scheduler::scheduleBlockForLatency(unsigned &MaxPressure,
     unsigned Thresholds[] = {144, 128, 112, 104, 96};
     unsigned Iterations = 5;
     float Ratio = (std::max(NumGrfs, 128u) - 48u) / 80.0f;
-    // limit the iterative approach to certain platforms for now
-    if (config.DoNotIterate) {
-      Thresholds[0] = getLatencyHidingThreshold(kernel, NumGrfs);
-      Iterations = 1;
-      Ratio = 1.0f; // already adjusted inside getLatencyHidingThreshold
-    }
-    for (unsigned i = 0; i < Iterations; ++i) {
-      auto GroupingThreshold = Thresholds[i];
+
+    // the last iteration uses the threshold from getLatencyHidingThreshold
+    for (unsigned i = 0; i < Iterations + 1; ++i) {
+      unsigned GroupingThreshold = 0;
+      if (i == Iterations) {
+        GroupingThreshold = getLatencyHidingThreshold(kernel, NumGrfs);
+        if (GroupingThreshold * 1.0 > Thresholds[i - 1] * Ratio) {
+          break;
+        }
+        Ratio = 1.0f; // already adjusted inside getLatencyHidingThreshold
+      } else {
+        GroupingThreshold = Thresholds[i];
+      }
       ddd.reset(ReassignID);
       ReassignID = false; // only reassign inst-local-id at most once
       LatencyScheduling(unsigned(GroupingThreshold * Ratio));