Merge pull request #8106 from The-OpenROAD-Project-staging/mpl-sa-fixes

Mpl sa fixes

Author: maliberty

src/mpl/src/SimulatedAnnealingCore.cpp

@@ -707,7 +707,6 @@ void SimulatedAnnealingCore<T>::fastSA()
 {
   float cost = calNormCost();
   float pre_cost = cost;
-  float delta_cost = 0.0;
   int step = 1;
   float temperature = init_temperature_;
   const float min_t = 1e-10;
@@ -718,31 +717,34 @@ void SimulatedAnnealingCore<T>::fastSA()
   // as it is too expensive
   notch_weight_ = 0.0;
 
-  if (isValid()) {
-    updateBestValidResult(cost);
-  }
+  updateBestResult(cost);
 
   while (step <= max_num_step_) {
     for (int i = 0; i < num_perturb_per_step_; i++) {
       saveState();
       perturb();
       cost = calNormCost();
 
-      const bool keep_result
-          = cost < pre_cost
-            || best_valid_result_.sequence_pair.pos_sequence.empty();
-      if (isValid() && keep_result) {
-        updateBestValidResult(cost);
+      const bool is_valid = isValid();
+      const bool improved = cost < pre_cost || best_result_.empty();
+      if ((!is_best_result_valid_ || is_valid) && improved) {
+        updateBestResult(cost);
+        is_best_result_valid_ = is_valid;
       }
 
-      delta_cost = cost - pre_cost;
-      const float num = distribution_(generator_);
-      const float prob
-          = (delta_cost > 0.0) ? std::exp((-1) * delta_cost / temperature) : 1;
-      if (num < prob) {
+      const float delta_cost = cost - pre_cost;
+      if (delta_cost <= 0) {
+        // always accept improvements
         pre_cost = cost;
       } else {
-        restoreState();
+        // probabilistically accept degradations for hill climbing
+        const float num = distribution_(generator_);
+        const float prob = std::exp(-delta_cost / temperature);
+        if (num < prob) {
+          pre_cost = cost;
+        } else {
+          restoreState();
+        }
       }
     }
 
@@ -765,40 +767,42 @@ void SimulatedAnnealingCore<T>::fastSA()
     graphics_->doNotSkip();
   }
   calPenalty();
+  cost = calNormCost();
 
-  if (!best_valid_result_.sequence_pair.pos_sequence.empty()
-      && (!isValid() || best_valid_result_.cost < calNormCost())) {
-    useBestValidResult();
+  const bool is_valid = isValid();
+  const bool improved = cost < best_result_.cost || best_result_.empty();
+  if ((is_best_result_valid_ && !is_valid) || !improved) {
+    useBestResult();
   }
 }
 
 template <class T>
-void SimulatedAnnealingCore<T>::updateBestValidResult(const float cost)
+void SimulatedAnnealingCore<T>::updateBestResult(const float cost)
 {
-  best_valid_result_.sequence_pair.pos_sequence = pos_seq_;
-  best_valid_result_.sequence_pair.neg_sequence = neg_seq_;
+  best_result_.sequence_pair.pos_sequence = pos_seq_;
+  best_result_.sequence_pair.neg_sequence = neg_seq_;
 
   if constexpr (std::is_same_v<T, SoftMacro>) {
     for (const int macro_id : pos_seq_) {
       SoftMacro& macro = macros_[macro_id];
-      best_valid_result_.macro_id_to_width[macro_id] = macro.getWidth();
+      best_result_.macro_id_to_width[macro_id] = macro.getWidth();
     }
   }
 
-  best_valid_result_.cost = cost;
+  best_result_.cost = cost;
 }
 
 template <class T>
-void SimulatedAnnealingCore<T>::useBestValidResult()
+void SimulatedAnnealingCore<T>::useBestResult()
 {
-  pos_seq_ = best_valid_result_.sequence_pair.pos_sequence;
-  neg_seq_ = best_valid_result_.sequence_pair.neg_sequence;
+  pos_seq_ = best_result_.sequence_pair.pos_sequence;
+  neg_seq_ = best_result_.sequence_pair.neg_sequence;
 
   if constexpr (std::is_same_v<T, SoftMacro>) {
     for (const int macro_id : pos_seq_) {
       SoftMacro& macro = macros_[macro_id];
       const float valid_result_width
-          = best_valid_result_.macro_id_to_width.at(macro_id);
+          = best_result_.macro_id_to_width.at(macro_id);
 
       if (macro.isMacroCluster()) {
         const float valid_result_height = macro.getArea() / valid_result_width;

src/mpl/src/SimulatedAnnealingCore.h

@@ -92,6 +92,8 @@ class SimulatedAnnealingCore
  protected:
   struct Result
   {
+    bool empty() const { return sequence_pair.pos_sequence.empty(); }
+
     float cost{std::numeric_limits<float>::max()};
     SequencePair sequence_pair;
     // [Only for SoftMacro] The same sequence pair can represent different
@@ -105,8 +107,8 @@ class SimulatedAnnealingCore
       const BoundaryRegionList& regions);
   void initSequencePair();
   void setDieArea(const Rect& die_area);
-  void updateBestValidResult(float cost);
-  void useBestValidResult();
+  void updateBestResult(float cost);
+  void useBestResult();
 
   virtual float calNormCost() const = 0;
   virtual void calPenalty() = 0;
@@ -210,7 +212,8 @@ class SimulatedAnnealingCore
   MplObserver* graphics_ = nullptr;
   odb::dbBlock* block_;
 
-  Result best_valid_result_;
+  Result best_result_;
+  bool is_best_result_valid_ = false;
 
   std::vector<float> cost_list_;  // store the cost in the list
   std::vector<float> T_list_;     // store the temperature

src/mpl/test/boundary_push1.defok

@@ -183,10 +183,10 @@ TRACKS Y 140 DO 282 STEP 3200 LAYER metal9 ;
 TRACKS X 190 DO 282 STEP 3200 LAYER metal10 ;
 TRACKS Y 140 DO 282 STEP 3200 LAYER metal10 ;
 COMPONENTS 4 ;
-    - MACRO_1 HM_100x100_1x1 + FIXED ( 600 610 ) FS ;
-    - MACRO_2 HM_100x100_1x1 + FIXED ( 600 241970 ) FS ;
-    - MACRO_3 HM_100x100_1x1 + FIXED ( 239440 241970 ) FS ;
-    - MACRO_4 HM_100x100_1x1 + FIXED ( 239440 610 ) FS ;
+    - MACRO_1 HM_100x100_1x1 + FIXED ( 239440 610 ) S ;
+    - MACRO_2 HM_100x100_1x1 + FIXED ( 600 241970 ) S ;
+    - MACRO_3 HM_100x100_1x1 + FIXED ( 600 610 ) S ;
+    - MACRO_4 HM_100x100_1x1 + FIXED ( 239440 241970 ) S ;
 END COMPONENTS
 PINS 4 ;
     - io_1 + NET io_1 + DIRECTION INPUT + USE SIGNAL ;

src/mpl/test/boundary_push2.defok

@@ -183,9 +183,9 @@ TRACKS Y 140 DO 282 STEP 3200 LAYER metal9 ;
 TRACKS X 190 DO 282 STEP 3200 LAYER metal10 ;
 TRACKS Y 140 DO 282 STEP 3200 LAYER metal10 ;
 COMPONENTS 54 ;
-    - MACRO_1 HM_100x100_1x1 + FIXED ( 600 20770 ) S ;
-    - MACRO_2 HM_100x100_1x1 + FIXED ( 239440 221810 ) FS ;
-    - MACRO_3 HM_100x100_1x1 + FIXED ( 600 221810 ) FS ;
+    - MACRO_1 HM_100x100_1x1 + FIXED ( 600 20770 ) FS ;
+    - MACRO_2 HM_100x100_1x1 + FIXED ( 600 221810 ) FS ;
+    - MACRO_3 HM_100x100_1x1 + FIXED ( 239440 221810 ) FS ;
     - MACRO_4 HM_100x100_1x1 + FIXED ( 239440 20770 ) FS ;
     - _001_ DFF_X1 + PLACED ( 216790 219801 ) N ;
     - _002_ DFF_X1 + PLACED ( 216790 219801 ) N ;

src/mpl/test/boundary_push3.defok

@@ -183,10 +183,10 @@ TRACKS Y 140 DO 282 STEP 3200 LAYER metal9 ;
 TRACKS X 190 DO 282 STEP 3200 LAYER metal10 ;
 TRACKS Y 140 DO 282 STEP 3200 LAYER metal10 ;
 COMPONENTS 54 ;
-    - MACRO_1 HM_100x100_1x1 + FIXED ( 220620 610 ) S ;
-    - MACRO_2 HM_100x100_1x1 + FIXED ( 19420 610 ) FS ;
-    - MACRO_3 HM_100x100_1x1 + FIXED ( 19420 241970 ) FS ;
-    - MACRO_4 HM_100x100_1x1 + FIXED ( 220620 241970 ) FS ;
+    - MACRO_1 HM_100x100_1x1 + FIXED ( 220620 241970 ) FS ;
+    - MACRO_2 HM_100x100_1x1 + FIXED ( 220620 610 ) FS ;
+    - MACRO_3 HM_100x100_1x1 + FIXED ( 19420 610 ) FS ;
+    - MACRO_4 HM_100x100_1x1 + FIXED ( 19420 241970 ) FS ;
     - _001_ DFF_X1 + PLACED ( 216790 219801 ) N ;
     - _002_ DFF_X1 + PLACED ( 216790 219801 ) N ;
     - _003_ DFF_X1 + PLACED ( 216790 219801 ) N ;

src/mpl/test/centralization1.defok

@@ -241,7 +241,7 @@ TRACKS Y 140 DO 282 STEP 3200 LAYER metal9 ;
 TRACKS X 190 DO 282 STEP 3200 LAYER metal10 ;
 TRACKS Y 140 DO 282 STEP 3200 LAYER metal10 ;
 COMPONENTS 1 ;
-    - MACRO_1 HM_100x100_1x1 + FIXED ( 202100 202490 ) FS ;
+    - MACRO_1 HM_100x100_1x1 + FIXED ( 202100 202490 ) S ;
 END COMPONENTS
 PINS 1 ;
     - io_1 + NET io_1 + DIRECTION INPUT + USE SIGNAL ;

src/mpl/test/clocked_macro.defok

@@ -339,7 +339,7 @@ TRACKS Y 140 DO 282 STEP 3200 LAYER metal9 ;
 TRACKS X 190 DO 282 STEP 3200 LAYER metal10 ;
 TRACKS Y 140 DO 282 STEP 3200 LAYER metal10 ;
 COMPONENTS 1 ;
-    - U1 CLOCKED_MACRO + FIXED ( 17880 16610 ) FS ;
+    - U1 CLOCKED_MACRO + FIXED ( 17880 16610 ) S ;
 END COMPONENTS
 NETS 3 ;
     - clk ( U1 CK ) + USE SIGNAL ;

src/mpl/test/guides2.defok

@@ -339,16 +339,16 @@ TRACKS Y 140 DO 282 STEP 3200 LAYER metal9 ;
 TRACKS X 190 DO 282 STEP 3200 LAYER metal10 ;
 TRACKS Y 140 DO 282 STEP 3200 LAYER metal10 ;
 COMPONENTS 10 ;
-    - U1 HM_100x400_4x4 + FIXED ( 665880 16610 ) S ;
-    - U10 HM_100x100_1x1 + FIXED ( 17880 232430 ) FN ;
-    - U2 HM_100x100_1x1 + FIXED ( 233880 448450 ) S ;
-    - U3 HM_100x100_1x1 + FIXED ( 233880 664610 ) S ;
-    - U4 HM_100x100_1x1 + FIXED ( 233880 16550 ) FN ;
-    - U5 HM_100x100_1x1 + FIXED ( 17880 448450 ) S ;
-    - U6 HM_100x400_4x4 + FIXED ( 449880 16590 ) N ;
-    - U7 HM_100x100_1x1 + FIXED ( 233880 232430 ) N ;
-    - U8 HM_100x100_1x1 + FIXED ( 17880 664610 ) S ;
-    - U9 HM_100x100_1x1 + FIXED ( 17880 16550 ) FN ;
+    - U1 HM_100x400_4x4 + FIXED ( 665880 16590 ) FN ;
+    - U10 HM_100x100_1x1 + FIXED ( 233880 664610 ) S ;
+    - U2 HM_100x100_1x1 + FIXED ( 449880 16550 ) FN ;
+    - U3 HM_100x100_1x1 + FIXED ( 449880 448450 ) S ;
+    - U4 HM_100x100_1x1 + FIXED ( 449880 232430 ) FN ;
+    - U5 HM_100x100_1x1 + FIXED ( 449880 664610 ) S ;
+    - U6 HM_100x400_4x4 + FIXED ( 17880 16590 ) FN ;
+    - U7 HM_100x100_1x1 + FIXED ( 233880 16550 ) FN ;
+    - U8 HM_100x100_1x1 + FIXED ( 233880 448450 ) S ;
+    - U9 HM_100x100_1x1 + FIXED ( 233880 232430 ) FN ;
 END COMPONENTS
 NETS 12 ;
     - w1 ( U2 I1 ) ( U1 O1 ) + USE SIGNAL ;