Replace rr_node_indices in block_rr_indices with RRSpatialLookup; Deposit invalid node ids when allocating RRSpatialLookup; No long need complex sanity checks. The vectors in RRSpatialLookup can be well aligned and allocated just in need; Remove the dead loop between rr_graph2.h and rr_graph_storage.h

tangxifan · tangxifan · commit 69f83bf9dd66 · 2021-05-28T16:44:50.000-06:00
diff --git a/vpr/src/device/rr_spatial_lookup.cpp b/vpr/src/device/rr_spatial_lookup.cpp
@@ -80,8 +80,33 @@ void RRSpatialLookup::add_node(RRNodeId node,
                                t_rr_type type,
                                int ptc,
                                e_side side) {
+    VTR_ASSERT(node); /* Must have a valid node id to be added */
     VTR_ASSERT_SAFE(3 == rr_node_indices_[type].ndims());
 
+    resize_nodes(x, y, type, side);
+
+    if (size_t(ptc) >= rr_node_indices_[type][x][y][side].size()) {
+        /* Deposit invalid ids to newly allocated elements while original elements are untouched */
+        rr_node_indices_[type][x][y][side].resize(ptc + 1, int(size_t(RRNodeId::INVALID())));
+    }
+
+    /* Resize on demand finished; Register the node */
+    rr_node_indices_[type][x][y][side][ptc] = int(size_t(node));
+}
+
+void RRSpatialLookup::mirror_nodes(const vtr::Point<int>& src_coord,
+                                   const vtr::Point<int>& des_coord,
+                                   t_rr_type type,
+                                   e_side side) {
+    VTR_ASSERT(SOURCE == type || SINK == type);
+    resize_nodes(des_coord.x(), des_coord.y(), type, side);
+    rr_node_indices_[type][des_coord.x()][des_coord.y()][side] = rr_node_indices_[type][src_coord.x()][src_coord.y()][side];
+}
+
+void RRSpatialLookup::resize_nodes(int x,
+                                   int y,
+                                   t_rr_type type,
+                                   e_side side) {
     /* Expand the fast look-up if the new node is out-of-range
      * This may seldom happen because the rr_graph building function
      * should ensure the fast look-up well organized  
@@ -97,11 +122,4 @@ void RRSpatialLookup::add_node(RRNodeId node,
                                        std::max(rr_node_indices_[type].dim_size(1), size_t(y) + 1),
                                        std::max(rr_node_indices_[type].dim_size(2), size_t(side) + 1)});
     }
-
-    if (size_t(ptc) >= rr_node_indices_[type][x][y][side].size()) {
-        rr_node_indices_[type][x][y][side].resize(ptc + 1);
-    }
-
-    /* Resize on demand finished; Register the node */
-    rr_node_indices_[type][x][y][side][ptc] = int(size_t(node));
 }
diff --git a/vpr/src/device/rr_spatial_lookup.h b/vpr/src/device/rr_spatial_lookup.h
@@ -1,6 +1,7 @@
 #ifndef RR_SPATIAL_LOOKUP_H
 #define RR_SPATIAL_LOOKUP_H
 
+#include "vtr_geometry.h"
 #include "vpr_types.h"
 
 /******************************************************************** 
@@ -89,7 +90,32 @@ class RRSpatialLookup {
                   int ptc,
                   e_side side);
 
-    /* TODO: Add an API remove_node() to unregister a node from the look-up */
+     /* Mirror the last dimension of a look-up, i.e., a list of nodes, from a source coordinate to 
+     * a destination coordinate.
+     * This function is mostly need by SOURCE and SINK nodes which are indexable in multiple locations.
+     * Considering a bounding box (x, y)->(x + width, y + height) of a multi-height and multi-width grid, 
+     * SOURCE and SINK nodes are indexable in any location inside the boundry.
+     *
+     * Note: currently this function only accept SOURCE/SINK nodes. May unlock for other depending on needs
+     */
+    void mirror_nodes(const vtr::Point<int>& src_coord,
+                      const vtr::Point<int>& des_coord,
+                      t_rr_type type,
+                      e_side side);
+
+    /* Resize three dimensions of the lookup under a given type of node to be memory efficient
+     * This function is called to expand the matrix when x, y or side 
+     * when one or more of them beyond current capacity 
+     *
+     * Strongly recommend to use when the sizes of dimensions are deterministic
+     *
+     * TODO: should have a reserve function but vtd::ndmatrix does not have such API
+     *       as a result, resize can be an internal one while reserve function is a public mutator
+     */
+    void resize_nodes(int x,
+                      int y,
+                      t_rr_type type,
+                      e_side side);
 
     /* -- Internal data storage -- */
   private:
diff --git a/vpr/src/route/rr_edge.h b/vpr/src/route/rr_edge.h
@@ -0,0 +1,25 @@
+#ifndef RR_EDGE_H
+#define RR_EDGE_H
+
+struct t_rr_edge_info {
+    t_rr_edge_info(int from, int to, short type) noexcept
+        : from_node(from)
+        , to_node(to)
+        , switch_type(type) {}
+
+    int from_node = OPEN;
+    int to_node = OPEN;
+    short switch_type = OPEN;
+
+    friend bool operator<(const t_rr_edge_info& lhs, const t_rr_edge_info& rhs) {
+        return std::tie(lhs.from_node, lhs.to_node, lhs.switch_type) < std::tie(rhs.from_node, rhs.to_node, rhs.switch_type);
+    }
+
+    friend bool operator==(const t_rr_edge_info& lhs, const t_rr_edge_info& rhs) {
+        return std::tie(lhs.from_node, lhs.to_node, lhs.switch_type) == std::tie(rhs.from_node, rhs.to_node, rhs.switch_type);
+    }
+};
+
+typedef std::vector<t_rr_edge_info> t_rr_edge_info_set;
+
+#endif /* RR_EDGE */
diff --git a/vpr/src/route/rr_graph.cpp b/vpr/src/route/rr_graph.cpp
@@ -573,7 +573,7 @@ static void build_rr_graph(const t_graph_type graph_type,
     /* Alloc node lookups, count nodes, alloc rr nodes */
     int num_rr_nodes = 0;
 
-    alloc_and_load_rr_node_indices(device_ctx.rr_node_indices,
+    alloc_and_load_rr_node_indices(device_ctx.rr_graph_builder,
                                    max_chan_width, grid,
                                    &num_rr_nodes, chan_details_x, chan_details_y);
 
diff --git a/vpr/src/route/rr_graph2.cpp b/vpr/src/route/rr_graph2.cpp
@@ -32,11 +32,10 @@ static void load_chan_rr_indices(const int max_chan_width,
                                  const int num_chans,
                                  const t_rr_type type,
                                  const t_chan_details& chan_details,
-                                 t_rr_node_indices& indices,
+                                 RRGraphBuilder& rr_graph_builder,
                                  int* index);
 
 static void load_block_rr_indices(RRGraphBuilder& rr_graph_builder,
-                                  t_rr_node_indices& indices,
                                   const DeviceGrid& grid,
                                   int* index);
 
@@ -928,44 +927,40 @@ static void load_chan_rr_indices(const int max_chan_width,
                                  const int num_chans,
                                  const t_rr_type type,
                                  const t_chan_details& chan_details,
-                                 t_rr_node_indices& indices,
+                                 RRGraphBuilder& rr_graph_builder,
                                  int* index) {
-    VTR_ASSERT(indices[type].dim_size(0) == size_t(num_chans));
-    VTR_ASSERT(indices[type].dim_size(1) == size_t(chan_len));
-    VTR_ASSERT(indices[type].dim_size(2) == NUM_SIDES);
-    for (int chan = 0; chan < num_chans - 1; ++chan) {
-        for (int seg = 1; seg < chan_len - 1; ++seg) {
-            /* Alloc the track inode lookup list */
-            //Since channels have no side, we just use the first side
-            indices[type][chan][seg][0].resize(max_chan_width, OPEN);
-        }
-    }
-
     for (int chan = 0; chan < num_chans - 1; ++chan) {
         for (int seg = 1; seg < chan_len - 1; ++seg) {
             /* Assign an inode to the starts of tracks */
             int x = (type == CHANX ? seg : chan);
             int y = (type == CHANX ? chan : seg);
             const t_chan_seg_details* seg_details = chan_details[x][y].data();
 
-            for (unsigned track = 0; track < indices[type][chan][seg][0].size(); ++track) {
+            for (int track = 0; track < max_chan_width; ++track) {
                 if (seg_details[track].length() <= 0)
                     continue;
 
                 int start = get_seg_start(seg_details, track, chan, seg);
 
+                /* TODO: Now we still use the (y, x) convention here for CHANX. Should rework later */
+                int node_x = chan;
+                int node_y = start;
+                if (CHANX == type) {
+                   std::swap(node_x, node_y);
+                }
+
                 /* If the start of the wire doesn't have a inode,
                  * assign one to it. */
-                int inode = indices[type][chan][start][0][track];
-                if (OPEN == inode) {
-                    inode = *index;
+                RRNodeId inode = rr_graph_builder.node_lookup().find_node(node_x, node_y, type, track, SIDES[0]);
+                if (!inode) {
+                    inode = RRNodeId(*index);
                     ++(*index);
 
-                    indices[type][chan][start][0][track] = inode;
+                    rr_graph_builder.node_lookup().add_node(inode, chan, start, type, track, SIDES[0]);
                 }
 
                 /* Assign inode of start of wire to current position */
-                indices[type][chan][seg][0][track] = inode;
+                rr_graph_builder.node_lookup().add_node(inode, chan, seg, type, track, SIDES[0]);
             }
         }
     }
@@ -975,7 +970,6 @@ static void load_chan_rr_indices(const int max_chan_width,
  * TODO: these building functions should only talk to a RRGraphBuilder object
  */
 static void load_block_rr_indices(RRGraphBuilder& rr_graph_builder,
-                                  t_rr_node_indices& indices,
                                   const DeviceGrid& grid,
                                   int* index) {
     //Walk through the grid assigning indices to SOURCE/SINK IPIN/OPIN
@@ -991,16 +985,12 @@ static void load_block_rr_indices(RRGraphBuilder& rr_graph_builder,
                     auto class_type = type->class_inf[iclass].type;
                     if (class_type == DRIVER) {
                         rr_graph_builder.node_lookup().add_node(RRNodeId(*index), x, y, SOURCE, iclass, SIDES[0]);
-                        rr_graph_builder.node_lookup().add_node(RRNodeId::INVALID(), x, y, SINK, iclass, SIDES[0]);
                     } else {
                         VTR_ASSERT(class_type == RECEIVER);
                         rr_graph_builder.node_lookup().add_node(RRNodeId(*index), x, y, SINK, iclass, SIDES[0]);
-                        rr_graph_builder.node_lookup().add_node(RRNodeId::INVALID(), x, y, SOURCE, iclass, SIDES[0]);
                     }
                     ++(*index);
                 }
-                VTR_ASSERT(indices[SOURCE][x][y][0].size() == type->class_inf.size());
-                VTR_ASSERT(indices[SINK][x][y][0].size() == type->class_inf.size());
 
                 /* Limited sides for grids
                  *   The wanted side depends on the location of the grid.
@@ -1068,18 +1058,13 @@ static void load_block_rr_indices(RRGraphBuilder& rr_graph_builder,
                                     auto class_type = type->class_inf[iclass].type;
 
                                     if (class_type == DRIVER) {
-                                        indices[OPIN][x_tile][y_tile][side].push_back(*index);
-                                        indices[IPIN][x_tile][y_tile][side].push_back(OPEN);
+                                        rr_graph_builder.node_lookup().add_node(RRNodeId(*index), x_tile, y_tile, OPIN, ipin, side);
                                         assigned_to_rr_node = true;
                                     } else {
                                         VTR_ASSERT(class_type == RECEIVER);
-                                        indices[OPIN][x_tile][y_tile][side].push_back(OPEN);
-                                        indices[IPIN][x_tile][y_tile][side].push_back(*index);
+                                        rr_graph_builder.node_lookup().add_node(RRNodeId(*index), x_tile, y_tile, IPIN, ipin, side);
                                         assigned_to_rr_node = true;
                                     }
-                                } else {
-                                    indices[IPIN][x_tile][y_tile][side].push_back(OPEN);
-                                    indices[OPIN][x_tile][y_tile][side].push_back(OPEN);
                                 }
                             }
                         }
@@ -1099,25 +1084,6 @@ static void load_block_rr_indices(RRGraphBuilder& rr_graph_builder,
                         ++(*index);
                     }
                 }
-
-                //Sanity check
-                for (int width_offset = 0; width_offset < type->width; ++width_offset) {
-                    int x_tile = x + width_offset;
-                    for (int height_offset = 0; height_offset < type->height; ++height_offset) {
-                        int y_tile = y + height_offset;
-                        for (e_side side : SIDES) {
-                            //Note that the fast look-up stores all the indices for the pins on each side
-                            //It has a fixed size (either 0 or the number of pins)
-                            //Case 0 pins: the side is skipped as no pins are located on it
-                            //Case number of pins: there are pins on this side
-                            //and data query can be applied any pin id on this side
-                            VTR_ASSERT((indices[IPIN][x_tile][y_tile][side].size() == size_t(type->num_pins))
-                                       || (0 == indices[IPIN][x_tile][y_tile][side].size()));
-                            VTR_ASSERT((indices[OPIN][x_tile][y_tile][side].size() == size_t(type->num_pins))
-                                       || (0 == indices[OPIN][x_tile][y_tile][side].size()));
-                        }
-                    }
-                }
             }
         }
     }
@@ -1132,8 +1098,14 @@ static void load_block_rr_indices(RRGraphBuilder& rr_graph_builder,
                 int root_x = x - width_offset;
                 int root_y = y - height_offset;
 
-                indices[SOURCE][x][y][0] = indices[SOURCE][root_x][root_y][0];
-                indices[SINK][x][y][0] = indices[SINK][root_x][root_y][0];
+                rr_graph_builder.node_lookup().mirror_nodes(vtr::Point<int>(root_x, root_y),
+                                                            vtr::Point<int>(x, y),
+                                                            SOURCE,
+                                                            SIDES[0]);
+                rr_graph_builder.node_lookup().mirror_nodes(vtr::Point<int>(root_x, root_y),
+                                                            vtr::Point<int>(x, y),
+                                                            SINK,
+                                                            SIDES[0]);
             }
         }
     }
@@ -1149,7 +1121,7 @@ static void load_block_rr_indices(RRGraphBuilder& rr_graph_builder,
  *         This will block us when putting the RRGraphBuilder object as an input arguement 
  *         of this function
  */
-void alloc_and_load_rr_node_indices(t_rr_node_indices& indices,
+void alloc_and_load_rr_node_indices(RRGraphBuilder& rr_graph_builder,
                                     const int max_chan_width,
                                     const DeviceGrid& grid,
                                     int* index,
@@ -1162,20 +1134,20 @@ void alloc_and_load_rr_node_indices(t_rr_node_indices& indices,
     /* Alloc the lookup table */
     for (t_rr_type rr_type : RR_TYPES) {
         if (rr_type == CHANX) {
-            indices[rr_type].resize({grid.height(), grid.width(), NUM_SIDES});
+            rr_graph_builder.node_lookup().resize_nodes(grid.height(), grid.width(), rr_type, NUM_SIDES);
         } else {
-            indices[rr_type].resize({grid.width(), grid.height(), NUM_SIDES});
+            rr_graph_builder.node_lookup().resize_nodes(grid.width(), grid.height(), rr_type, NUM_SIDES);
         }
     }
 
     /* Assign indices for block nodes */
-    load_block_rr_indices(g_vpr_ctx.mutable_device().rr_graph_builder, indices, grid, index);
+    load_block_rr_indices(rr_graph_builder, grid, index);
 
     /* Load the data for x and y channels */
     load_chan_rr_indices(max_chan_width, grid.width(), grid.height(),
-                         CHANX, chan_details_x, indices, index);
+                         CHANX, chan_details_x, rr_graph_builder, index);
     load_chan_rr_indices(max_chan_width, grid.height(), grid.width(),
-                         CHANY, chan_details_y, indices, index);
+                         CHANY, chan_details_y, rr_graph_builder, index);
 }
 
 bool verify_rr_node_indices(const DeviceGrid& grid, const t_rr_node_indices& rr_node_indices, const t_rr_graph_storage& rr_nodes) {
diff --git a/vpr/src/route/rr_graph2.h b/vpr/src/route/rr_graph2.h
@@ -5,6 +5,7 @@
 #include "build_switchblocks.h"
 #include "rr_graph_fwd.h"
 #include "rr_graph_util.h"
+#include "rr_graph_builder.h"
 #include "rr_types.h"
 #include "device_grid.h"
 
@@ -15,27 +16,6 @@ enum e_seg_details_type {
     SEG_DETAILS_Y
 };
 
-struct t_rr_edge_info {
-    t_rr_edge_info(int from, int to, short type) noexcept
-        : from_node(from)
-        , to_node(to)
-        , switch_type(type) {}
-
-    int from_node = OPEN;
-    int to_node = OPEN;
-    short switch_type = OPEN;
-
-    friend bool operator<(const t_rr_edge_info& lhs, const t_rr_edge_info& rhs) {
-        return std::tie(lhs.from_node, lhs.to_node, lhs.switch_type) < std::tie(rhs.from_node, rhs.to_node, rhs.switch_type);
-    }
-
-    friend bool operator==(const t_rr_edge_info& lhs, const t_rr_edge_info& rhs) {
-        return std::tie(lhs.from_node, lhs.to_node, lhs.switch_type) == std::tie(rhs.from_node, rhs.to_node, rhs.switch_type);
-    }
-};
-
-typedef std::vector<t_rr_edge_info> t_rr_edge_info_set;
-
 typedef vtr::NdMatrix<short, 6> t_sblock_pattern;
 
 struct t_opin_connections_scratchpad {
@@ -44,7 +24,7 @@ struct t_opin_connections_scratchpad {
 
 /******************* Subroutines exported by rr_graph2.c *********************/
 
-void alloc_and_load_rr_node_indices(t_rr_node_indices& indices,
+void alloc_and_load_rr_node_indices(RRGraphBuilder& rr_graph_builder,
                                     const int max_chan_width,
                                     const DeviceGrid& grid,
                                     int* index,
diff --git a/vpr/src/route/rr_graph_storage.h b/vpr/src/route/rr_graph_storage.h
@@ -6,7 +6,7 @@
 
 #include "rr_graph_fwd.h"
 #include "rr_node_fwd.h"
-#include "rr_graph2.h"
+#include "rr_edge.h"
 #include "vtr_log.h"
 #include "vtr_memory.h"
 #include "vpr_utils.h"
