Merge pull request #2457 from verilog-to-routing/compressed_router_lookahead

Updating router lookahead

Author: vaughnbetz

vpr/src/base/ShowSetup.cpp

@@ -373,6 +373,9 @@ static void ShowRouterOpts(const t_router_opts& RouterOpts) {
                 case e_router_lookahead::MAP:
                     VTR_LOG("MAP\n");
                     break;
+                case e_router_lookahead::COMPRESSED_MAP:
+                    VTR_LOG("COMPRESSED_MAP\n");
+                    break;
                 case e_router_lookahead::EXTENDED_MAP:
                     VTR_LOG("EXTENDED_MAP\n");
                     break;
@@ -512,6 +515,9 @@ static void ShowRouterOpts(const t_router_opts& RouterOpts) {
                 case e_router_lookahead::MAP:
                     VTR_LOG("MAP\n");
                     break;
+                case e_router_lookahead::COMPRESSED_MAP:
+                    VTR_LOG("COMPRESSED_MAP\n");
+                    break;
                 case e_router_lookahead::EXTENDED_MAP:
                     VTR_LOG("EXTENDED_MAP\n");
                     break;
@@ -628,8 +634,8 @@ static void ShowPlacerOpts(const t_placer_opts& PlacerOpts,
                 VPR_FATAL_ERROR(VPR_ERROR_UNKNOWN, "Unknown delay_model_reducer\n");
             VTR_LOG("PlacerOpts.delay_model_reducer: %s\n", e_reducer_strings[(size_t)PlacerOpts.delay_model_reducer].c_str());
 
-            std::string place_delay_model_strings[2] = {"DELTA", "DELTA_OVERRIDE"};
-            if ((size_t)PlacerOpts.delay_model_type > 1)
+            std::string place_delay_model_strings[3] = {"SIMPLE", "DELTA", "DELTA_OVERRIDE"};
+            if ((size_t)PlacerOpts.delay_model_type > 2)
                 VPR_FATAL_ERROR(VPR_ERROR_UNKNOWN, "Unknown delay_model_type\n");
             VTR_LOG("PlacerOpts.delay_model_type: %s\n", place_delay_model_strings[(size_t)PlacerOpts.delay_model_type].c_str());
         }

vpr/src/base/read_options.cpp

@@ -902,11 +902,14 @@ struct ParseRouteBBUpdate {
 
 struct ParseRouterLookahead {
     ConvertedValue<e_router_lookahead> from_str(std::string str) {
+        std::transform(str.begin(), str.end(), str.begin(), ::tolower);
         ConvertedValue<e_router_lookahead> conv_value;
         if (str == "classic")
             conv_value.set_value(e_router_lookahead::CLASSIC);
         else if (str == "map")
             conv_value.set_value(e_router_lookahead::MAP);
+        else if (str == "compressed_map")
+            conv_value.set_value(e_router_lookahead::COMPRESSED_MAP);
         else if (str == "extended_map")
             conv_value.set_value(e_router_lookahead::EXTENDED_MAP);
         else {
@@ -926,6 +929,8 @@ struct ParseRouterLookahead {
             conv_value.set_value("classic");
         else if (val == e_router_lookahead::MAP) {
             conv_value.set_value("map");
+        } else if (val == e_router_lookahead::COMPRESSED_MAP) {
+            conv_value.set_value("compressed_map");
         } else {
             VTR_ASSERT(val == e_router_lookahead::EXTENDED_MAP);
             conv_value.set_value("extended_map");
@@ -934,14 +939,16 @@ struct ParseRouterLookahead {
     }
 
     std::vector<std::string> default_choices() {
-        return {"classic", "map", "extended_map"};
+        return {"classic", "map", "compressed_map", "extended_map"};
     }
 };
 
 struct ParsePlaceDelayModel {
     ConvertedValue<PlaceDelayModelType> from_str(std::string str) {
         ConvertedValue<PlaceDelayModelType> conv_value;
-        if (str == "delta")
+        if (str == "simple") {
+            conv_value.set_value(PlaceDelayModelType::SIMPLE);
+        } else if (str == "delta")
             conv_value.set_value(PlaceDelayModelType::DELTA);
         else if (str == "delta_override")
             conv_value.set_value(PlaceDelayModelType::DELTA_OVERRIDE);
@@ -955,7 +962,9 @@ struct ParsePlaceDelayModel {
 
     ConvertedValue<std::string> to_str(PlaceDelayModelType val) {
         ConvertedValue<std::string> conv_value;
-        if (val == PlaceDelayModelType::DELTA)
+        if (val == PlaceDelayModelType::SIMPLE)
+            conv_value.set_value("simple");
+        else if (val == PlaceDelayModelType::DELTA)
             conv_value.set_value("delta");
         else if (val == PlaceDelayModelType::DELTA_OVERRIDE)
             conv_value.set_value("delta_override");
@@ -968,7 +977,7 @@ struct ParsePlaceDelayModel {
     }
 
     std::vector<std::string> default_choices() {
-        return {"delta", "delta_override"};
+        return {"simple", "delta", "delta_override"};
     }
 };
 
@@ -2245,6 +2254,7 @@ argparse::ArgumentParser create_arg_parser(std::string prog_name, t_options& arg
             "This option controls what information is considered and how"
             " the placement delay model is constructed.\n"
             "Valid options:\n"
+            " * 'simple' uses map router lookahead\n"
             " * 'delta' uses differences in position only\n"
             " * 'delta_override' uses differences in position with overrides for direct connects\n")
         .default_value("delta")
@@ -2566,6 +2576,8 @@ argparse::ArgumentParser create_arg_parser(std::string prog_name, t_options& arg
             " * classic: The classic VPR lookahead (may perform better on un-buffered routing\n"
             "            architectures)\n"
             " * map: An advanced lookahead which accounts for diverse wire type\n"
+            " * compressed_map: The algorithm is similar to map lookahead with the exception of saprse sampling of the chip"
+            " to reduce the run-time to build the router lookahead and also its memory footprint\n"
             " * extended_map: A more advanced and extended lookahead which accounts for a more\n"
             "                 exhaustive node sampling method\n"
             "\n"

vpr/src/base/vpr_types.h

@@ -119,10 +119,11 @@ constexpr auto INVALID_BLOCK_ID = ClusterBlockId(-2);
 #endif
 
 enum class e_router_lookahead {
-    CLASSIC,      ///<VPR's classic lookahead (assumes uniform wire types)
-    MAP,          ///<Lookahead considering different wire types (see Oleg Petelin's MASc Thesis)
-    EXTENDED_MAP, ///<Lookahead with a more extensive node sampling method
-    NO_OP         ///<A no-operation lookahead which always returns zero
+    CLASSIC,        ///<VPR's classic lookahead (assumes uniform wire types)
+    MAP,            ///<Lookahead considering different wire types (see Oleg Petelin's MASc Thesis)
+    COMPRESSED_MAP, /// Similar to MAP, but use a sparse sampling of the chip
+    EXTENDED_MAP,   ///<Lookahead with a more extensive node sampling method
+    NO_OP           ///<A no-operation lookahead which always returns zero
 };
 
 enum class e_route_bb_update {
@@ -1129,6 +1130,7 @@ enum e_place_effort_scaling {
 };
 
 enum class PlaceDelayModelType {
+    SIMPLE,
     DELTA,          ///<Delta x/y based delay model
     DELTA_OVERRIDE, ///<Delta x/y based delay model with special case delay overrides
 };

vpr/src/place/place.cpp

@@ -619,7 +619,6 @@ void try_place(const Netlist<>& net_list,
     if (placer_opts.place_algorithm.is_timing_driven()) {
         /*do this before the initial placement to avoid messing up the initial placement */
         place_delay_model = alloc_lookups_and_delay_model(net_list,
-                                                          device_ctx.arch_switch_inf,
                                                           chan_width_dist,
                                                           placer_opts,
                                                           router_opts,

vpr/src/place/place_delay_model.cpp

@@ -160,6 +160,14 @@ void OverrideDelayModel::set_base_delay_model(std::unique_ptr<DeltaDelayModel> b
     base_delay_model_ = std::move(base_delay_model_obj);
 }
 
+float SimpleDelayModel::delay(const t_physical_tile_loc& from_loc, int /*from_pin*/, const t_physical_tile_loc& to_loc, int /*to_pin*/) const {
+    int delta_x = std::abs(from_loc.x - to_loc.x);
+    int delta_y = std::abs(from_loc.y - to_loc.y);
+
+    int from_tile_idx = g_vpr_ctx.device().grid.get_physical_type(from_loc)->index;
+    return delays_[from_tile_idx][from_loc.layer_num][to_loc.layer_num][delta_x][delta_y];
+}
+
 /**
  * When writing capnp targetted serialization, always allow compilation when
  * VTR_ENABLE_CAPNPROTO=OFF. Generally this means throwing an exception instead.
@@ -319,7 +327,6 @@ void OverrideDelayModel::write(const std::string& file) const {
 
 ///@brief Initialize the placer delay model.
 std::unique_ptr<PlaceDelayModel> alloc_lookups_and_delay_model(const Netlist<>& net_list,
-                                                               const std::vector<t_arch_switch_inf>& arch_switch_inf,
                                                                t_chan_width_dist chan_width_dist,
                                                                const t_placer_opts& placer_opts,
                                                                const t_router_opts& router_opts,
@@ -331,7 +338,6 @@ std::unique_ptr<PlaceDelayModel> alloc_lookups_and_delay_model(const Netlist<>&
     return compute_place_delay_model(placer_opts,
                                      router_opts,
                                      net_list,
-                                     arch_switch_inf,
                                      det_routing_arch,
                                      segment_inf,
                                      chan_width_dist,

vpr/src/place/place_delay_model.h

@@ -29,7 +29,6 @@ class PlaceDelayModel;
 
 ///@brief Initialize the placer delay model.
 std::unique_ptr<PlaceDelayModel> alloc_lookups_and_delay_model(const Netlist<>& net_list,
-                                                               const std::vector<t_arch_switch_inf>& arch_switch_inf,
                                                                t_chan_width_dist chan_width_dist,
                                                                const t_placer_opts& place_opts,
                                                                const t_router_opts& router_opts,
@@ -113,7 +112,13 @@ class DeltaDelayModel : public PlaceDelayModel {
 
   private:
     vtr::NdMatrix<float, 3> delays_; // [0..num_layers-1][0..max_dx][0..max_dy]
+    /**
+     * @brief The minimum delay of inter-layer connections
+     */
     float cross_layer_delay_;
+    /**
+     * @brief Indicates whether the router is a two-stage or run-flat
+     */
     bool is_flat_;
 };
 
@@ -144,8 +149,13 @@ class OverrideDelayModel : public PlaceDelayModel {
 
   private:
     std::unique_ptr<DeltaDelayModel> base_delay_model_;
-    /* Minimum delay of cross-layer connections */
+    /**
+     * @brief Minimum delay of cross-layer connections
+     */
     float cross_layer_delay_;
+    /**
+     * @brief Indicates whether the router is a two-stage or run-flat
+     */
     bool is_flat_;
 
     void compute_override_delay_model(RouterDelayProfiler& router,
@@ -217,3 +227,33 @@ class OverrideDelayModel : public PlaceDelayModel {
     static_assert(sizeof(t_override::delta_x) == sizeof(short), "Expect all t_override data members to be shorts");
     static_assert(sizeof(t_override::delta_y) == sizeof(short), "Expect all t_override data members to be shorts");
 };
+
+///@brief A simple delay model based on the information stored in router lookahead
+///  This is in contrast to other placement delay models that get the cost of getting from one location to another by running the router
+class SimpleDelayModel : public PlaceDelayModel {
+  public:
+    SimpleDelayModel() {}
+
+    void compute(
+        RouterDelayProfiler& router,
+        const t_placer_opts& placer_opts,
+        const t_router_opts& router_opts,
+        int longest_length) override;
+    float delay(const t_physical_tile_loc& from_loc, int /*from_pin*/, const t_physical_tile_loc& to_loc, int /*to_pin*/) const override;
+    void dump_echo(std::string /*filepath*/) const override {}
+
+    void read(const std::string& /*file*/) override {}
+    void write(const std::string& /*file*/) const override {}
+
+  private:
+    /**
+     * @brief The matrix to store the minimum delay between different points on different layers.
+     *
+     *The matrix used to store delay information is a 5D matrix. This data structure stores the minimum delay for each tile type on each layer to other layers
+     *for each dx and dy. We decided to separate the delay for each physical type on each die to accommodate cases where the connectivity of a physical type differs
+     *on each layer. Additionally, instead of using d_layer, we distinguish between the destination layer to handle scenarios where connectivity between layers
+     *is not uniform. For example, if the number of inter-layer connections between layer 1 and 2 differs from the number of connections between layer 0 and 1.
+     *One might argue that this variability could also occur for dx and dy. However, we are operating under the assumption that the FPGA fabric architecture is regular.
+     */
+    vtr::NdMatrix<float, 5> delays_; // [0..num_physical_type-1][0..num_layers-1][0..num_layers-1][0..max_dx][0..max_dy]
+};

vpr/src/place/timing_place_lookup.cpp

@@ -143,6 +143,13 @@ static vtr::NdMatrix<float, 3> compute_delta_delay_model(
     int longest_length,
     bool is_flat);
 
+/**
+ * @brief Use the information in the router lookahead to fill the delay matrix instead of running the router
+ * @param route_profiler
+ * @return The delay matrix that contain the minimum cost between two locations
+ */
+static vtr::NdMatrix<float, 5> compute_simple_delay_model(RouterDelayProfiler& route_profiler);
+
 static bool find_direct_connect_sample_locations(const t_direct_inf* direct,
                                                  t_physical_tile_type_ptr from_type,
                                                  int from_pin,
@@ -167,7 +174,6 @@ static float find_neightboring_average(vtr::NdMatrix<float, 3>& matrix, t_physic
 std::unique_ptr<PlaceDelayModel> compute_place_delay_model(const t_placer_opts& placer_opts,
                                                            const t_router_opts& router_opts,
                                                            const Netlist<>& net_list,
-                                                           const std::vector<t_arch_switch_inf>& arch_switch_inf,
                                                            t_det_routing_arch* det_routing_arch,
                                                            std::vector<t_segment_inf>& segment_inf,
                                                            t_chan_width_dist chan_width_dist,
@@ -195,10 +201,11 @@ std::unique_ptr<PlaceDelayModel> compute_place_delay_model(const t_placer_opts&
 
     /*now setup and compute the actual arrays */
     std::unique_ptr<PlaceDelayModel> place_delay_model;
-    float min_cross_layer_delay = get_min_cross_layer_delay(arch_switch_inf,
-                                                            segment_inf,
-                                                            det_routing_arch->wire_to_arch_ipin_switch_between_dice);
-    if (placer_opts.delay_model_type == PlaceDelayModelType::DELTA) {
+    float min_cross_layer_delay = get_min_cross_layer_delay();
+
+    if (placer_opts.delay_model_type == PlaceDelayModelType::SIMPLE) {
+        place_delay_model = std::make_unique<SimpleDelayModel>();
+    } else if (placer_opts.delay_model_type == PlaceDelayModelType::DELTA) {
         place_delay_model = std::make_unique<DeltaDelayModel>(min_cross_layer_delay, is_flat);
     } else if (placer_opts.delay_model_type == PlaceDelayModelType::DELTA_OVERRIDE) {
         place_delay_model = std::make_unique<OverrideDelayModel>(min_cross_layer_delay, is_flat);
@@ -250,6 +257,14 @@ void OverrideDelayModel::compute(
     compute_override_delay_model(route_profiler, router_opts);
 }
 
+void SimpleDelayModel::compute(
+    RouterDelayProfiler& router,
+    const t_placer_opts& /*placer_opts*/,
+    const t_router_opts& /*router_opts*/,
+    int /*longest_length*/) {
+    delays_ = compute_simple_delay_model(router);
+}
+
 /******* File Accessible Functions **********/
 
 std::vector<int> get_best_classes(enum e_pin_type pintype, t_physical_tile_type_ptr type) {
@@ -1004,6 +1019,37 @@ static vtr::NdMatrix<float, 3> compute_delta_delay_model(
     return delta_delays;
 }
 
+static vtr::NdMatrix<float, 5> compute_simple_delay_model(RouterDelayProfiler& route_profiler) {
+    const auto& grid = g_vpr_ctx.device().grid;
+    int num_physical_tile_types = static_cast<int>(g_vpr_ctx.device().physical_tile_types.size());
+    // Initializing the delay matrix to [num_physical_types][num_layers][num_layers][width][height]
+    // The second index related to the layer that the source location is on and the third index is for the sink layer
+    vtr::NdMatrix<float, 5> delta_delays({static_cast<unsigned long>(num_physical_tile_types),
+                                          static_cast<unsigned long>(grid.get_num_layers()),
+                                          static_cast<unsigned long>(grid.get_num_layers()),
+                                          grid.width(),
+                                          grid.height()});
+
+    for (int physical_tile_type_idx = 0; physical_tile_type_idx < num_physical_tile_types; ++physical_tile_type_idx) {
+        for (int from_layer = 0; from_layer < grid.get_num_layers(); ++from_layer) {
+            for (int to_layer = 0; to_layer < grid.get_num_layers(); ++to_layer) {
+                for (int dx = 0; dx < static_cast<int>(grid.width()); ++dx) {
+                    for (int dy = 0; dy < static_cast<int>(grid.height()); ++dy) {
+                        float min_delay = route_profiler.get_min_delay(physical_tile_type_idx,
+                                                                       from_layer,
+                                                                       to_layer,
+                                                                       dx,
+                                                                       dy);
+                        delta_delays[physical_tile_type_idx][from_layer][to_layer][dx][dy] = min_delay;
+                    }
+                }
+            }
+        }
+    }
+
+    return delta_delays;
+}
+
 //Finds a src_rr and sink_rr appropriate for measuring the delay of the current direct specification
 static bool find_direct_connect_sample_locations(const t_direct_inf* direct,
                                                  t_physical_tile_type_ptr from_type,

vpr/src/place/timing_place_lookup.h

@@ -5,7 +5,6 @@
 std::unique_ptr<PlaceDelayModel> compute_place_delay_model(const t_placer_opts& placer_opts,
                                                            const t_router_opts& router_opts,
                                                            const Netlist<>& net_list,
-                                                           const std::vector<t_arch_switch_inf>& arch_switch_inf,
                                                            t_det_routing_arch* det_routing_arch,
                                                            std::vector<t_segment_inf>& segment_inf,
                                                            t_chan_width_dist chan_width_dist,