Merge pull request #2465 from verilog-to-routing/par_router3_2

Improved parallel router: add DecompNetlistRouter

libs/EXTERNAL/libargparse/argparse_test.cpp

@@ -399,10 +399,11 @@ int main(
             .show_in(argparse::ShowIn::HELP_ONLY);
     route_grp.add_argument(args.router_algorithm, "--router_algorithm")
             .help("Specifies the router algorithm to use.\n"
-                  " * parallel: timing_driven with tricks to run on multiple cores (may be worse)\n"
-                  " * timing driven: focuses on routability and circuit speed\n")
+                  " * timing driven: focuses on routability and circuit speed [default]\n"
+                  " * parallel: timing_driven with nets in different regions of the chip routed in parallel\n"
+                  " * parallel_decomp: timing_driven with additional parallelism obtained by decomposing high-fanout nets, possibly reducing quality\n")
             .default_value("timing_driven")
-            .choices({"parallel", "timing_driven"})
+            .choices({"parallel", "parallel_decomp", "timing_driven"})
             .show_in(argparse::ShowIn::HELP_ONLY);
     route_grp.add_argument(args.min_incremental_reroute_fanout, "--min_incremental_reroute_fanout")
             .help("The net fanout thershold above which nets will be re-routed incrementally.")

libs/librrgraph/src/base/rr_graph_storage.h

@@ -667,11 +667,6 @@ class t_rr_graph_storage {
     static inline Direction get_node_direction(
         vtr::array_view_id<RRNodeId, const t_rr_node_data> node_storage,
         RRNodeId id) {
-        auto& node_data = node_storage[id];
-        if (node_data.type_ != CHANX && node_data.type_ != CHANY) {
-            VTR_LOG_ERROR("Attempted to access RR node 'direction' for non-channel type '%s'",
-                          rr_node_typename[node_data.type_]);
-        }
         return node_storage[id].dir_side_.direction;
     }
 

libs/libvtrutil/src/vtr_dynamic_bitset.h

@@ -20,6 +20,11 @@ class dynamic_bitset {
     static_assert(std::numeric_limits<Storage>::is_integer,
                   "dynamic_bitset storage must be integer!");
 
+    constexpr dynamic_bitset() = default;
+    constexpr dynamic_bitset(Index size) {
+        resize(size);
+    }
+
     ///@brief Reize to the determined size
     void resize(size_t size) {
         array_.resize((size + kWidth - 1) / kWidth);
@@ -63,6 +68,39 @@ class dynamic_bitset {
         return (array_[index_value / kWidth] & (1u << (index_value % kWidth))) != 0;
     }
 
+    ///@brief Return count of set bits.
+    constexpr size_t count(void) const {
+        size_t out = 0;
+        for (auto x : array_)
+            out += __builtin_popcount(x);
+        return out;
+    }
+
+    ///@brief Bitwise OR with rhs. Truncate the operation if one operand is smaller.
+    constexpr dynamic_bitset<Index, Storage>& operator|=(const dynamic_bitset<Index, Storage>& x) {
+        size_t n = std::min(array_.size(), x.array_.size());
+        for (size_t i = 0; i < n; i++)
+            array_[i] |= x.array_[i];
+        return *this;
+    }
+
+    ///@brief Bitwise AND with rhs. Truncate the operation if one operand is smaller.
+    constexpr dynamic_bitset<Index, Storage>& operator&=(const dynamic_bitset<Index, Storage>& x) {
+        size_t n = std::min(array_.size(), x.array_.size());
+        for (size_t i = 0; i < n; i++)
+            array_[i] &= x.array_[i];
+        return *this;
+    }
+
+    ///@brief Return inverted bitset.
+    inline dynamic_bitset<Index, Storage> operator~(void) const {
+        dynamic_bitset<Index, Storage> out(size());
+        size_t n = array_.size();
+        for (size_t i = 0; i < n; i++)
+            out.array_[i] = ~array_[i];
+        return out;
+    }
+
   private:
     std::vector<Storage> array_;
 };

vpr/src/base/ShowSetup.cpp

@@ -263,6 +263,9 @@ static void ShowRouterOpts(const t_router_opts& RouterOpts) {
             case PARALLEL:
                 VTR_LOG("PARALLEL\n");
                 break;
+            case PARALLEL_DECOMP:
+                VTR_LOG("PARALLEL_DECOMP\n");
+                break;
             case TIMING_DRIVEN:
                 VTR_LOG("TIMING_DRIVEN\n");
                 break;

vpr/src/base/read_options.cpp

@@ -174,6 +174,8 @@ struct ParseRouterAlgorithm {
         ConvertedValue<e_router_algorithm> conv_value;
         if (str == "parallel")
             conv_value.set_value(PARALLEL);
+        else if (str == "parallel_decomp")
+            conv_value.set_value(PARALLEL_DECOMP);
         else if (str == "timing_driven")
             conv_value.set_value(TIMING_DRIVEN);
         else {
@@ -2403,10 +2405,11 @@ argparse::ArgumentParser create_arg_parser(std::string prog_name, t_options& arg
     route_grp.add_argument<e_router_algorithm, ParseRouterAlgorithm>(args.RouterAlgorithm, "--router_algorithm")
         .help(
             "Specifies the router algorithm to use.\n"
-            " * parallel: [experimental] timing_driven but multithreaded\n"
-            " * timing_driven: focuses on routability and circuit speed\n")
+            " * timing driven: focuses on routability and circuit speed [default]\n"
+            " * parallel: timing_driven with nets in different regions of the chip routed in parallel\n"
+            " * parallel_decomp: timing_driven with additional parallelism obtained by decomposing high-fanout nets, possibly reducing quality\n")
         .default_value("timing_driven")
-        .choices({"parallel", "timing_driven"})
+        .choices({"parallel", "parallel_decomp", "timing_driven"})
         .show_in(argparse::ShowIn::HELP_ONLY);
 
     route_grp.add_argument(args.min_incremental_reroute_fanout, "--min_incremental_reroute_fanout")

vpr/src/base/vpr_types.h

@@ -1345,6 +1345,7 @@ struct t_placer_opts {
 
 enum e_router_algorithm {
     PARALLEL,
+    PARALLEL_DECOMP,
     TIMING_DRIVEN,
 };
 

vpr/src/route/DecompNetlistRouter.h

@@ -0,0 +1,126 @@
+#pragma once
+
+/** @file Parallel and net-decomposing case for NetlistRouter. Works like
+ * \see ParallelNetlistRouter, but tries to "decompose" nets and assign them to
+ * the next level of the partition tree where possible. */
+#include "netlist_routers.h"
+
+#include <tbb/task_group.h>
+
+/** Maximum number of iterations for net decomposition
+ * 5 is found experimentally: higher values get more speedup on initial iters but # of iters increases */
+const int MAX_DECOMP_ITER = 5;
+
+/** Maximum # of decomposition for a net: 2 means one net gets divided down to <4 virtual nets.
+ * Higher values are more aggressive: better thread utilization but worse congestion resolving */
+const int MAX_DECOMP_DEPTH = 2;
+
+/** Minimum # of fanouts of a net to consider decomp. */
+const int MIN_DECOMP_SINKS = 8;
+
+/** Minimum # of fanouts of a virtual net to consider decomp. */
+const int MIN_DECOMP_SINKS_VNET = 8;
+
+template<typename HeapType>
+class DecompNetlistRouter : public NetlistRouter {
+  public:
+    DecompNetlistRouter(
+        const Netlist<>& net_list,
+        const RouterLookahead* router_lookahead,
+        const t_router_opts& router_opts,
+        CBRR& connections_inf,
+        NetPinsMatrix<float>& net_delay,
+        const ClusteredPinAtomPinsLookup& netlist_pin_lookup,
+        std::shared_ptr<SetupHoldTimingInfo> timing_info,
+        NetPinTimingInvalidator* pin_timing_invalidator,
+        route_budgets& budgeting_inf,
+        const RoutingPredictor& routing_predictor,
+        const vtr::vector<ParentNetId, std::vector<std::unordered_map<RRNodeId, int>>>& choking_spots,
+        bool is_flat)
+        : _routers_th(_make_router(router_lookahead, is_flat))
+        , _net_list(net_list)
+        , _router_opts(router_opts)
+        , _connections_inf(connections_inf)
+        , _net_delay(net_delay)
+        , _netlist_pin_lookup(netlist_pin_lookup)
+        , _timing_info(timing_info)
+        , _pin_timing_invalidator(pin_timing_invalidator)
+        , _budgeting_inf(budgeting_inf)
+        , _routing_predictor(routing_predictor)
+        , _choking_spots(choking_spots)
+        , _is_flat(is_flat)
+        , _net_known_samples(net_list.nets().size())
+        , _is_decomp_disabled(net_list.nets().size()) {}
+    ~DecompNetlistRouter() {}
+
+    /** Run a single iteration of netlist routing for this->_net_list. This usually means calling
+     * \ref route_net for each net, which will handle other global updates.
+     * \return RouteIterResults for this iteration. */
+    RouteIterResults route_netlist(int itry, float pres_fac, float worst_neg_slack);
+    /** Set RCV enable flag for all routers managed by this netlist router.
+     * Net decomposition does not work with RCV, so calling this fn with x=true is a fatal error. */
+    void set_rcv_enabled(bool x);
+    void set_timing_info(std::shared_ptr<SetupHoldTimingInfo> timing_info);
+
+  private:
+    /** Should we decompose this net? */
+    bool should_decompose_net(ParentNetId net_id, const PartitionTreeNode& node);
+    /** Get a bitset with sinks to route before net decomposition */
+    vtr::dynamic_bitset<> get_decomposition_mask(ParentNetId net_id, const PartitionTreeNode& node);
+    /** Get a bitset with sinks to route before virtual net decomposition */
+    vtr::dynamic_bitset<> get_vnet_decomposition_mask(const VirtualNet& vnet, const PartitionTreeNode& node);
+    /** Decompose and route a regular net. Output the resulting vnets to \p left and \p right.
+     * \return Success status: true if routing is successful and left and right now contain valid virtual nets: false otherwise. */
+    bool decompose_and_route_net(ParentNetId net_id, const PartitionTreeNode& node, VirtualNet& left, VirtualNet& right);
+    /** Decompose and route a virtual net. Output the resulting vnets to \p left and \p right.
+     * \return Success status: true if routing is successful and left and right now contain valid virtual nets: false otherwise. */
+    bool decompose_and_route_vnet(VirtualNet& vnet, const PartitionTreeNode& node, VirtualNet& left, VirtualNet& right);
+    /** A single task to route nets inside a PartitionTree node and add tasks for its child nodes to task group \p g. */
+    void route_partition_tree_node(tbb::task_group& g, PartitionTreeNode& node);
+
+    ConnectionRouter<HeapType> _make_router(const RouterLookahead* router_lookahead, bool is_flat) {
+        auto& device_ctx = g_vpr_ctx.device();
+        auto& route_ctx = g_vpr_ctx.mutable_routing();
+
+        return ConnectionRouter<HeapType>(
+            device_ctx.grid,
+            *router_lookahead,
+            device_ctx.rr_graph.rr_nodes(),
+            &device_ctx.rr_graph,
+            device_ctx.rr_rc_data,
+            device_ctx.rr_graph.rr_switch(),
+            route_ctx.rr_node_route_inf,
+            is_flat);
+    }
+
+    /* Context fields. Most of them will be forwarded to route_net (see route_net.tpp) */
+    /** Per-thread storage for ConnectionRouters. */
+    tbb::enumerable_thread_specific<ConnectionRouter<HeapType>> _routers_th;
+    const Netlist<>& _net_list;
+    const t_router_opts& _router_opts;
+    CBRR& _connections_inf;
+    /** Per-thread storage for RouteIterResults. */
+    tbb::enumerable_thread_specific<RouteIterResults> _results_th;
+    NetPinsMatrix<float>& _net_delay;
+    const ClusteredPinAtomPinsLookup& _netlist_pin_lookup;
+    std::shared_ptr<SetupHoldTimingInfo> _timing_info;
+    NetPinTimingInvalidator* _pin_timing_invalidator;
+    route_budgets& _budgeting_inf;
+    const RoutingPredictor& _routing_predictor;
+    const vtr::vector<ParentNetId, std::vector<std::unordered_map<RRNodeId, int>>>& _choking_spots;
+    bool _is_flat;
+
+    /** Cached routing parameters for current iteration (inputs to \see route_netlist()) */
+    int _itry;
+    float _pres_fac;
+    float _worst_neg_slack;
+
+    /** Sinks to be always sampled for decomposition for each net: [0.._net_list.size()-1]
+     * (i.e. when routing fails after decomposition for a sink, sample it on next iteration) */
+    vtr::vector<ParentNetId, vtr::dynamic_bitset<>> _net_known_samples;
+
+    /** Is decomposition disabled for this net? [0.._net_list.size()-1] */
+    vtr::vector<ParentNetId, bool> _is_decomp_disabled;
+};
+
+#include "DecompNetlistRouter.tpp"