Support 3D Custom Switch Blocks

vpr/src/base/SetupVPR.cpp

@@ -481,6 +481,7 @@ static void SetupRouterOpts(const t_options& Options, t_router_opts* RouterOpts)
     RouterOpts->generate_rr_node_overuse_report = Options.generate_rr_node_overuse_report;
     RouterOpts->flat_routing = Options.flat_routing;
     RouterOpts->has_choking_spot = Options.has_choking_spot;
+    RouterOpts->custom_3d_sb_fanin_fanout = Options.custom_3d_sb_fanin_fanout;
     RouterOpts->with_timing_analysis = Options.timing_analysis;
 }
 

vpr/src/base/read_options.cpp

@@ -2461,6 +2461,13 @@ argparse::ArgumentParser create_arg_parser(const std::string& prog_name, t_optio
         .default_value("false")
         .show_in(argparse::ShowIn::HELP_ONLY);
 
+    route_grp.add_argument(args.custom_3d_sb_fanin_fanout, "--custom_3d_sb_fanin_fanout")
+            .help(
+                    "Specifies the number of tracks that can drive a 3D switch block connection"
+                    "and the number of tracks that can be driven by a 3D switch block connection")
+            .default_value("1")
+            .show_in(argparse::ShowIn::HELP_ONLY);
+
     auto& route_timing_grp = parser.add_argument_group("timing-driven routing options");
 
     route_timing_grp.add_argument(args.astar_fac, "--astar_fac")

vpr/src/base/read_options.h

@@ -202,6 +202,7 @@ struct t_options {
     argparse::ArgValue<int> reorder_rr_graph_nodes_seed;
     argparse::ArgValue<bool> flat_routing;
     argparse::ArgValue<bool> has_choking_spot;
+    argparse::ArgValue<int> custom_3d_sb_fanin_fanout;
 
     /* Timing-driven router options only */
     argparse::ArgValue<float> astar_fac;

vpr/src/base/vpr_types.h

@@ -1464,6 +1464,8 @@ struct t_router_opts {
     bool flat_routing;
     bool has_choking_spot;
 
+    int custom_3d_sb_fanin_fanout = 1;
+
     bool with_timing_analysis;
 
     // Options related to rr_node reordering, for testing and possible cache optimization

vpr/src/route/build_switchblocks.h

@@ -92,20 +92,6 @@ struct t_switchblock_edge {
     short to_wire_layer;
 };
 
-/**
- * @brief contains the required information to create extra length-0 RR nodes to model 3D custom switch blocks connections within the RR graph
- *
- *  @from_tracks a vector containing source tracks ptc_num indices that are connected to the same destination track in above or below layer in multi-die FPGAs
- *  @offset_to_extra_chanx_node index (max_chan_width + "offset_to_extra_chanx_node") to the correct length-0 RR node that all tracks in "from_tracks" should be connected to in RR graph.
- *  @connected_to_dest this flag is used to avoid edge duplications while adding edges for all "from_tracks" RR nodes to the same node (length-0 RR node) in the destination layer
- *
- */
-struct t_inter_die_switchblock_edge {
-    std::vector<short> from_tracks;
-    short offset_to_extra_chanx_node = -1;
-    bool connected_to_dest = false;
-};
-
 /* Switchblock connections are made as [x][y][from_side][to_side][from_wire_ind].
  * The Switchblock_Lookup class specifies these dimensions.
  * Furthermore, a source_wire at a given 5-d coordinate may connect to multiple destination wires so the value

vpr/src/route/rr_graph.cpp

@@ -198,6 +198,7 @@ static std::function<void(t_chan_width*)> alloc_and_load_rr_graph(RRGraphBuilder
                                                                   const t_chan_width& chan_width,
                                                                   const int wire_to_ipin_switch,
                                                                   const int wire_to_pin_between_dice_switch,
+                                                                  const int custom_3d_sb_fanin_fanout,
                                                                   const int delayless_switch,
                                                                   const enum e_directionality directionality,
                                                                   bool* Fc_clipped,
@@ -482,6 +483,7 @@ static void build_rr_chan(RRGraphBuilder& rr_graph_builder,
                           t_rr_edge_info_set& des_3d_rr_edges_to_create,
                           const int wire_to_ipin_switch,
                           const int wire_to_pin_between_dice_switch,
+                          const int custom_3d_sb_fanin_fanout,
                           const int delayless_switch,
                           const enum e_directionality directionality);
 
@@ -654,6 +656,7 @@ static void build_rr_graph(const t_graph_type graph_type,
                            const int global_route_switch,
                            const int wire_to_arch_ipin_switch,
                            const int wire_to_pin_between_dice_switch,
+                           const int custom_3d_sb_fanin_fanout,
                            const int delayless_switch,
                            const float R_minW_nmos,
                            const float R_minW_pmos,
@@ -749,6 +752,7 @@ void create_rr_graph(const t_graph_type graph_type,
                            det_routing_arch->global_route_switch,
                            det_routing_arch->wire_to_arch_ipin_switch,
                            det_routing_arch->wire_to_arch_ipin_switch_between_dice,
+                           router_opts.custom_3d_sb_fanin_fanout,
                            det_routing_arch->delayless_switch,
                            det_routing_arch->R_minW_nmos,
                            det_routing_arch->R_minW_pmos,
@@ -970,6 +974,7 @@ static void build_rr_graph(const t_graph_type graph_type,
                            const int global_route_switch,
                            const int wire_to_arch_ipin_switch,
                            const int wire_to_pin_between_dice_switch,
+                           const int custom_3d_sb_fanin_fanout,
                            const int delayless_switch,
                            const float R_minW_nmos,
                            const float R_minW_pmos,
@@ -1271,7 +1276,7 @@ static void build_rr_graph(const t_graph_type graph_type,
      */
     if (grid.get_num_layers() > 1 && sb_type == CUSTOM) {
         //keep how many nodes each switchblock requires for each x,y location
-        auto extra_nodes_per_switchblock = get_number_track_to_track_inter_die_conn(sb_conn_map, device_ctx.rr_graph_builder);
+        auto extra_nodes_per_switchblock = get_number_track_to_track_inter_die_conn(sb_conn_map,custom_3d_sb_fanin_fanout, device_ctx.rr_graph_builder);
         //allocate new nodes in each switchblocks
         alloc_and_load_inter_die_rr_node_indices(device_ctx.rr_graph_builder, &nodes_per_chan, grid, extra_nodes_per_switchblock, &num_rr_nodes);
         device_ctx.rr_graph_builder.resize_nodes(num_rr_nodes);
@@ -1371,6 +1376,7 @@ static void build_rr_graph(const t_graph_type graph_type,
         nodes_per_chan,
         wire_to_arch_ipin_switch,
         wire_to_pin_between_dice_switch,
+        custom_3d_sb_fanin_fanout,
         delayless_switch,
         directionality,
         &Fc_clipped,
@@ -2034,6 +2040,7 @@ static std::function<void(t_chan_width*)> alloc_and_load_rr_graph(RRGraphBuilder
                                                                   const t_chan_width& chan_width,
                                                                   const int wire_to_ipin_switch,
                                                                   const int wire_to_pin_between_dice_switch,
+                                                                  const int custom_3d_sb_fanin_fanout,
                                                                   const int delayless_switch,
                                                                   const enum e_directionality directionality,
                                                                   bool* Fc_clipped,
@@ -2186,6 +2193,7 @@ static std::function<void(t_chan_width*)> alloc_and_load_rr_graph(RRGraphBuilder
                                   rr_edges_to_create, des_3d_rr_edges_to_create,
                                   wire_to_ipin_switch,
                                   wire_to_pin_between_dice_switch,
+                                  custom_3d_sb_fanin_fanout,
                                   delayless_switch,
                                   directionality);
 
@@ -2206,6 +2214,7 @@ static std::function<void(t_chan_width*)> alloc_and_load_rr_graph(RRGraphBuilder
                                   rr_edges_to_create, des_3d_rr_edges_to_create,
                                   wire_to_ipin_switch,
                                   wire_to_pin_between_dice_switch,
+                                  custom_3d_sb_fanin_fanout,
                                   delayless_switch,
                                   directionality);
 
@@ -3115,6 +3124,7 @@ static void build_rr_chan(RRGraphBuilder& rr_graph_builder,
                           t_rr_edge_info_set& des_3d_rr_edges_to_create,
                           const int wire_to_ipin_switch,
                           const int wire_to_pin_between_dice_switch,
+                          const int custom_3d_sb_fanin_fanout,
                           const int delayless_switch,
                           const enum e_directionality directionality) {
     /* this function builds both x and y-directed channel segments, so set up our
@@ -3198,7 +3208,7 @@ static void build_rr_chan(RRGraphBuilder& rr_graph_builder,
                                     opposite_chan_type,  seg_dimension, max_opposite_chan_width, grid,
                                     Fs_per_side, sblock_pattern, num_of_3d_conns_custom_SB, node, rr_edges_to_create,
                                     des_3d_rr_edges_to_create, from_seg_details, to_seg_details, opposite_chan_details,
-                                    directionality,delayless_switch,
+                                    directionality,custom_3d_sb_fanin_fanout,delayless_switch,
                                     switch_block_conn, sb_conn_map);
             }
         }
@@ -3218,7 +3228,7 @@ static void build_rr_chan(RRGraphBuilder& rr_graph_builder,
                                     opposite_chan_type,  seg_dimension, max_opposite_chan_width, grid,
                                     Fs_per_side, sblock_pattern, num_of_3d_conns_custom_SB, node, rr_edges_to_create,
                                     des_3d_rr_edges_to_create, from_seg_details, to_seg_details, opposite_chan_details,
-                                    directionality, delayless_switch, switch_block_conn, sb_conn_map);
+                                    directionality,custom_3d_sb_fanin_fanout, delayless_switch, switch_block_conn, sb_conn_map);
             }
         }
 
@@ -3250,7 +3260,7 @@ static void build_rr_chan(RRGraphBuilder& rr_graph_builder,
                                         chan_type,  seg_dimension, max_chan_width, grid,
                                         Fs_per_side, sblock_pattern, num_of_3d_conns_custom_SB, node, rr_edges_to_create,
                                         des_3d_rr_edges_to_create, from_seg_details, to_seg_details, from_chan_details,
-                                        directionality,delayless_switch,
+                                        directionality,custom_3d_sb_fanin_fanout, delayless_switch,
                                         switch_block_conn, sb_conn_map);
                 }
             }

vpr/src/route/rr_graph2.cpp

@@ -126,6 +126,7 @@ static void get_switchblocks_edges(RRGraphBuilder& rr_graph_builder,
                                    const int to_y,
                                    const t_rr_type to_chan_type,
                                    const int switch_override,
+                                   const int custom_3d_sb_fanin_fanout,
                                    const int delayless_switch,
                                    t_sb_connection_map* sb_conn_map,
                                    vtr::NdMatrix<int, 2>& num_of_3d_conns_custom_SB,
@@ -163,6 +164,7 @@ static int get_track_to_chan_seg(RRGraphBuilder& rr_graph_builder,
                                  const e_side from_side,
                                  const e_side to_side,
                                  const int swtich_override,
+                                 const int custom_3d_sb_fanin_fanout,
                                  const int delayless_switch,
                                  t_sb_connection_map* sb_conn_map,
                                  vtr::NdMatrix<int, 2>& num_of_3d_conns_custom_SB,
@@ -1245,6 +1247,7 @@ static bool check_3d_SB_RRnodes(RRGraphBuilder& rr_graph_builder, int x, int y,
 }
 
 vtr::NdMatrix<int, 2> get_number_track_to_track_inter_die_conn(t_sb_connection_map* sb_conn_map,
+                                                               const int custom_3d_sb_fanin_fanout,
                                                                RRGraphBuilder& rr_graph_builder) {
     auto& grid_ctx = g_vpr_ctx.device().grid;
     vtr::NdMatrix<int, 2> extra_nodes_per_switchblocks;
@@ -1283,7 +1286,7 @@ vtr::NdMatrix<int, 2> get_number_track_to_track_inter_die_conn(t_sb_connection_m
                         }
                     }
                 }
-                extra_nodes_per_switchblocks[x][y] += ((num_of_3d_conn + 39)/ 40);
+                extra_nodes_per_switchblocks[x][y] += ((num_of_3d_conn + custom_3d_sb_fanin_fanout - 1)/ custom_3d_sb_fanin_fanout);
             }
         }
     }
@@ -1924,6 +1927,7 @@ int get_track_to_tracks(RRGraphBuilder& rr_graph_builder,
                         const t_chan_seg_details* to_seg_details,
                         const t_chan_details& to_chan_details,
                         const enum e_directionality directionality,
+                        const int custom_3d_sb_fanin_fanout,
                         const int delayless_switch,
                         const vtr::NdMatrix<std::vector<int>, 3>& switch_block_conn,
                         t_sb_connection_map* sb_conn_map) {
@@ -2056,7 +2060,7 @@ int get_track_to_tracks(RRGraphBuilder& rr_graph_builder,
                 if (Direction::DEC == from_seg_details[from_track].direction() || BI_DIRECTIONAL == directionality) {
                     num_conn += get_track_to_chan_seg(rr_graph_builder, layer, max_chan_width, from_track, to_chan, to_seg,
                                                       to_type, from_side_a, to_side,
-                                                      switch_override, delayless_switch,
+                                                      switch_override, custom_3d_sb_fanin_fanout, delayless_switch,
                                                       sb_conn_map, num_of_3d_conns_custom_SB, from_rr_node, rr_edges_to_create, des_3d_rr_edges_to_create);
                 }
             } else {
@@ -2094,7 +2098,7 @@ int get_track_to_tracks(RRGraphBuilder& rr_graph_builder,
                 if (Direction::INC == from_seg_details[from_track].direction() || BI_DIRECTIONAL == directionality) {
                     num_conn += get_track_to_chan_seg(rr_graph_builder, layer, max_chan_width, from_track, to_chan, to_seg,
                                                       to_type, from_side_b, to_side,
-                                                      switch_override,delayless_switch,
+                                                      switch_override,custom_3d_sb_fanin_fanout, delayless_switch,
                                                       sb_conn_map, num_of_3d_conns_custom_SB, from_rr_node, rr_edges_to_create, des_3d_rr_edges_to_create);
                 }
             } else {
@@ -2240,6 +2244,7 @@ static void get_switchblocks_edges(RRGraphBuilder& rr_graph_builder,
                                    const int to_y,
                                    const t_rr_type to_chan_type,
                                    const int switch_override,
+                                   const int custom_3d_sb_fanin_fanout,
                                    const int delayless_switch,
                                    t_sb_connection_map* sb_conn_map,
                                    vtr::NdMatrix<int, 2>& num_of_3d_conns_custom_SB,
@@ -2309,7 +2314,7 @@ static void get_switchblocks_edges(RRGraphBuilder& rr_graph_builder,
                  *  +-------------+        +-------------+         +--------------+         +--------------+
                  *
                  * */
-                int offset = num_of_3d_conns_custom_SB[tile_x][tile_y] / 40;
+                int offset = num_of_3d_conns_custom_SB[tile_x][tile_y] / custom_3d_sb_fanin_fanout;
                 RRNodeId track_to_chanx_node = rr_graph_builder.node_lookup().find_node(layer, tile_x, tile_y, CHANX, max_chan_width + offset);
                 RRNodeId diff_layer_chanx_node = rr_graph_builder.node_lookup().find_node(to_layer, tile_x, tile_y, CHANX, max_chan_width + offset);
                 RRNodeId chanx_to_track_node = rr_graph_builder.node_lookup().find_node(to_layer, to_x, to_y, to_chan_type, to_wire);
@@ -2333,7 +2338,7 @@ static void get_switchblocks_edges(RRGraphBuilder& rr_graph_builder,
                 ++edge_count;
 
                 //we only add the following edge between intermediate nodes once for the first 3D connection for each pair of intermediate nodes
-                if (num_of_3d_conns_custom_SB[tile_x][tile_y] % 40 == 0) {
+                if (num_of_3d_conns_custom_SB[tile_x][tile_y] % custom_3d_sb_fanin_fanout == 0) {
                     rr_edges_to_create.emplace_back(track_to_chanx_node, diff_layer_chanx_node, delayless_switch, false);
                     ++edge_count;
                 }
@@ -2354,6 +2359,7 @@ static int get_track_to_chan_seg(RRGraphBuilder& rr_graph_builder,
                                  const e_side from_side,
                                  const e_side to_side,
                                  const int switch_override,
+                                 const int custom_3d_sb_fanin_fanout,
                                  const int delayless_switch,
                                  t_sb_connection_map* sb_conn_map,
                                  vtr::NdMatrix<int, 2>& num_of_3d_conns_custom_SB,
@@ -2393,6 +2399,7 @@ static int get_track_to_chan_seg(RRGraphBuilder& rr_graph_builder,
                            to_y,
                            to_chan_type,
                            switch_override,
+                           custom_3d_sb_fanin_fanout,
                            delayless_switch,
                            sb_conn_map,
                            num_of_3d_conns_custom_SB,
@@ -2415,6 +2422,7 @@ static int get_track_to_chan_seg(RRGraphBuilder& rr_graph_builder,
                                to_y,
                                to_chan_type,
                                switch_override,
+                               custom_3d_sb_fanin_fanout,
                                delayless_switch,
                                sb_conn_map,
                                num_of_3d_conns_custom_SB,

vpr/src/route/rr_graph2.h

@@ -87,6 +87,7 @@ int get_rr_node_index(const t_rr_node_indices& L_rr_node_indices,
  * @return number of die-crossing connection for each unique (x, y) location within the grid ([0..grid.width-1][0..grid.height-1])
  */
 vtr::NdMatrix<int, 2> get_number_track_to_track_inter_die_conn(t_sb_connection_map* sb_conn_map,
+                                                               const int custom_3d_sb_fanin_fanout,
                                                                RRGraphBuilder& rr_graph_builder);
 
 int find_average_rr_node_index(int device_width,
@@ -216,6 +217,7 @@ int get_track_to_tracks(RRGraphBuilder& rr_graph_builder,
                         const t_chan_seg_details* to_seg_details,
                         const t_chan_details& to_chan_details,
                         const enum e_directionality directionality,
+                        const int custom_3d_sb_fanin_fanout,
                         const int delayless_switch,
                         const vtr::NdMatrix<std::vector<int>, 3>& switch_block_conn,
                         t_sb_connection_map* sb_conn_map);