Fix post-processing bug that was destroying network stats (incl.

distributed stats) for outermost masked levels.

src/loop-analysis/tiling.cpp

@@ -179,6 +179,12 @@ void MaskTiles(std::vector<DataMovementInfo>& tile_nest, std::bitset<MaxTilingLe
       tile_nest[cur].content_accesses = 0;
       tile_nest[cur].parent_access_share = 0;
 
+      // There is some ugliness here. We have a follow-up call to
+      // ProcessOuterMaskedLevels() which will destroy even more stats from
+      // this level, so the above resets are perhaps not needed. Make sure
+      // there aren't any intervening calls between here and POML() and then
+      // delete most of the above lines *except* the trigger that POML needs.
+
       // Break out, we are done since there is no other unmasked level.
       break;
     }
@@ -284,8 +290,12 @@ void ProcessOuterMaskedLevels(std::vector<DataMovementInfo>& tile_nest, std::bit
     // Work on all outermost masked levels until we find an unmasked level.
     if (!mask[cur])
     {
-      // Blow up *all* stats (including network stats).
+      // Blow up all stats (including network stats), *except* distributed
+      // network stats (because distributed network traffic is transferred
+      // on behalf of the child level.
+      auto saved_distributed_stats = tile_nest[cur].distributed_access_stats;
       tile_nest[cur].Reset();
+      tile_nest[cur].distributed_access_stats = saved_distributed_stats;
     }
     else
     {
@@ -845,6 +855,33 @@ ComputeNest CollapseComputeNest(analysis::CompoundComputeNest& tiles, int num_ti
   return solution;
 }
 
+void PrintTiles(std::vector<DataMovementInfo>& tile_nest)
+{
+  int num_tiling_levels = tile_nest.size();
+
+  for (int cur = 0; cur < num_tiling_levels; cur++)
+  {
+    auto& tile = tile_nest[cur];
+
+    std::cout << "Level " << cur << std::endl;
+    std::cout << "  size = " << tile.size << std::endl;
+    std::cout << "  ca = " << tile.content_accesses << std::endl;
+    std::cout << "  access[" << tile.access_stats.stats.size() << "]:\n";
+    for (auto& x: tile.access_stats.stats)
+    {
+      if (x.second.accesses > 0)
+        std::cout << "    [" << x.first.first << "," << x.first.second << "]: accesses = "
+                  << x.second.accesses << " hops = " << x.second.hops << std::endl;
+    }
+    std::cout << "  dist access[" << tile.access_stats.stats.size() << "]:\n";
+    for (auto& x: tile.distributed_access_stats.stats)
+    {
+      if (x.second.accesses > 0)
+        std::cout << "    [" << x.first.first << "," << x.first.second << "]: accesses = "
+                  << x.second.accesses << " hops = " << x.second.hops << std::endl;
+    }
+  }
+}
 
 // Collapse tiles into a given number of levels.
 // Input and output are both arrays of tile nests,
@@ -945,6 +982,9 @@ CompoundDataMovementNest CollapseDataMovementNest(analysis::CompoundDataMovement
     // Calculate share of parent accesses (used to compute fills).
     ComputeParentAccessShare(solution[pv]);
 
+    // Perform distributed-multicast if supported.
+    DistributeTiles(solution[pv], distribution_supported[pv]);
+
     // Mask each solution according to the provided bit mask.
     MaskTiles(solution[pv], tile_mask[pv]);
 
@@ -954,9 +994,6 @@ CompoundDataMovementNest CollapseDataMovementNest(analysis::CompoundDataMovement
     // Set backing storage fill to zero place holder
     ResetBackingStorageFillsPlaceHolder(solution[pv]);
 
-    // Perform distributed-multicast if supported.
-    DistributeTiles(solution[pv], distribution_supported[pv]);
-
     // Calculate the extra accesses and fills due to link transfers
     ComputePeerAccesses(solution[pv]);
 
@@ -969,7 +1006,6 @@ CompoundDataMovementNest CollapseDataMovementNest(analysis::CompoundDataMovement
     // Find the parent and child levels for later compression/decompression logic
     SetParentLevel(solution[pv]);
     SetChildLevel(solution[pv]);
-
   }
 
   // flip the workload tensor set flag if necessary