[WIP] Parallel router: net decomposition

libs/EXTERNAL/libtatum/libtatum/tatum/analyzer_factory.hpp

@@ -7,6 +7,7 @@
 #include "tatum/TimingGraphFwd.hpp"
 #include "tatum/TimingConstraintsFwd.hpp"
 
+#include "tatum/analyzers/AdaptiveSetupHoldTimingAnalyzer.hpp"
 #include "tatum/graph_walkers.hpp"
 #include "tatum/timing_analyzers.hpp"
 #include "tatum/analyzers/full_timing_analyzers.hpp"
@@ -55,9 +56,9 @@ namespace tatum {
 ///Factor class to construct timing analyzers
 ///
 ///\tparam Visitor The analysis type visitor (e.g. SetupAnalysis)
-///\tparam GraphWalker The graph walker to use (defaults to serial traversals)
+///\tparam GraphWalker The graph walker to use
 template<class Visitor,
-         class GraphWalker>
+         class... GraphWalkers>
 struct AnalyzerFactory {
 
     //We use the dependent_false template to detect if the un-specialized AnalyzerFactor 
@@ -176,6 +177,21 @@ struct AnalyzerFactory<SetupHoldAnalysis,SerialIncrWalker> {
     }
 };
 
+template<>
+struct AnalyzerFactory<SetupHoldAnalysis,ParallelWalker,SerialIncrWalker> {
+
+    static std::unique_ptr<SetupHoldTimingAnalyzer> make(const TimingGraph& timing_graph,
+                                                         const TimingConstraints& timing_constraints,
+                                                         const DelayCalculator& delay_calc) {
+        return std::unique_ptr<SetupHoldTimingAnalyzer>(
+                new detail::AdaptiveSetupHoldTimingAnalyzer<ParallelWalker, SerialIncrWalker>(
+                    timing_graph, 
+                    timing_constraints, 
+                    delay_calc)
+                );
+    }
+};
+
 } //namepsace
 
 #endif

libs/EXTERNAL/libtatum/libtatum/tatum/analyzer_factory_fwd.hpp

@@ -10,7 +10,7 @@ namespace tatum {
 ///\tparam Visitor The analysis type visitor (e.g. SetupAnalysis)
 ///\tparam GraphWalker The graph walker to use (defaults to serial traversals)
 template<class Visitor,
-         class GraphWalker=SerialWalker>
+         class... GraphWalkers>
 struct AnalyzerFactory;
 
 } //namepsace

libs/EXTERNAL/libtatum/libtatum/tatum/analyzers/AdaptiveSetupHoldTimingAnalyzer.hpp

@@ -0,0 +1,202 @@
+#pragma once
+#include "tatum/TimingGraphFwd.hpp"
+#include "tatum/graph_walkers/SerialWalker.hpp"
+#include "tatum/graph_walkers/SerialIncrWalker.hpp"
+#include "tatum/SetupHoldAnalysis.hpp"
+#include "tatum/analyzers/SetupHoldTimingAnalyzer.hpp"
+#include "tatum/base/validate_timing_graph_constraints.hpp"
+#include "tatum/graph_walkers/TimingGraphWalker.hpp"
+
+namespace tatum { namespace detail {
+
+/** Threshold for AdaptiveSetupHoldTimingAnalyzer to use full updates. 
+* Expressed as fraction of all edges in timing graph. */
+constexpr float full_update_threshold = 0.1;
+
+/**
+ * A concrete implementation of a SetupHoldTimingAnalyzer.
+ *
+ * This is an adaptive analyzer: can do incremental updates if the number of invalidated
+ * nodes is small, falls back to a full update after a certain threshold to avoid the overhead.
+ */
+template<class FullWalker=SerialWalker, class IncrWalker=SerialIncrWalker>
+class AdaptiveSetupHoldTimingAnalyzer : public SetupHoldTimingAnalyzer {
+    public:
+        AdaptiveSetupHoldTimingAnalyzer(const TimingGraph& timing_graph, const TimingConstraints& timing_constraints, const DelayCalculator& delay_calculator)
+            : SetupHoldTimingAnalyzer()
+            , timing_graph_(timing_graph)
+            , timing_constraints_(timing_constraints)
+            , delay_calculator_(delay_calculator)
+            , setup_hold_visitor_(timing_graph_.nodes().size(), timing_graph_.edges().size()) {
+            validate_timing_graph_constraints(timing_graph_, timing_constraints_);
+
+            //Initialize profiling data. Use full walker to store data for both
+            full_walker_.set_profiling_data("total_analysis_sec", 0.);
+            full_walker_.set_profiling_data("analysis_sec", 0.);
+            full_walker_.set_profiling_data("num_full_updates", 0.);
+            full_walker_.set_profiling_data("num_incr_updates", 0.);
+
+            mode_ = Mode::INCR;
+            n_modified_edges_ = 0;
+            max_modified_edges_ = timing_graph_.edges().size() * full_update_threshold;
+        }
+
+    protected:
+        //Update both setup and hold simultaneously (this is more efficient than updating them sequentially)
+        virtual void update_timing_impl() override {
+            auto start_time = Clock::now();
+
+            if(mode_ == Mode::INCR)
+                update_timing_incr_(setup_hold_visitor_);
+            else
+                update_timing_full_(setup_hold_visitor_);
+
+            clear_timing_incr_();
+
+            double analysis_sec = std::chrono::duration_cast<dsec>(Clock::now() - start_time).count();
+
+            //Record profiling data (use full walker to store it) (arbitrary choice)
+            double total_analysis_sec = analysis_sec + full_walker_.get_profiling_data("total_analysis_sec");
+            full_walker_.set_profiling_data("total_analysis_sec", total_analysis_sec);
+            full_walker_.set_profiling_data("analysis_sec", analysis_sec);
+            if(mode_ == Mode::INCR)
+                full_walker_.set_profiling_data("num_incr_updates", full_walker_.get_profiling_data("num_incr_updates") + 1);
+            else
+                full_walker_.set_profiling_data("num_full_updates", full_walker_.get_profiling_data("num_full_updates") + 1);
+
+            mode_ = Mode::INCR; /* We did our update, try to use incr until too many edges are modified */
+        }
+
+        //Update only setup timing
+        virtual void update_setup_timing_impl() override {
+            auto& setup_visitor = setup_hold_visitor_.setup_visitor();
+
+            if(mode_ == Mode::INCR)
+                update_timing_incr_(setup_visitor);
+            else
+                update_timing_full_(setup_visitor);
+        }
+
+        //Update only hold timing
+        virtual void update_hold_timing_impl() override {
+            auto& hold_visitor = setup_hold_visitor_.hold_visitor();
+
+            if(mode_ == Mode::INCR)
+                update_timing_incr_(hold_visitor);
+            else
+                update_timing_full_(hold_visitor);
+        }
+
+        virtual void invalidate_edge_impl(const EdgeId edge) override {
+            if(mode_ == Mode::FULL)
+                return;
+            incr_walker_.invalidate_edge(edge);
+            n_modified_edges_++;
+            if(n_modified_edges_ > max_modified_edges_)
+                mode_ = Mode::FULL;
+        }
+
+        virtual node_range modified_nodes_impl() const override {
+            if(mode_ == Mode::FULL)
+                return full_walker_.modified_nodes();
+            else
+                return incr_walker_.modified_nodes();
+        }
+
+        double get_profiling_data_impl(std::string key) const override {
+            return full_walker_.get_profiling_data(key);
+        }
+
+        size_t num_unconstrained_startpoints_impl() const override {
+            if(mode_ == Mode::FULL)
+                return full_walker_.num_unconstrained_startpoints();
+            else
+                return incr_walker_.num_unconstrained_startpoints();
+        }
+
+        size_t num_unconstrained_endpoints_impl() const override {
+            if(mode_ == Mode::FULL)
+                return full_walker_.num_unconstrained_endpoints();
+            else
+                return incr_walker_.num_unconstrained_endpoints();
+        }
+
+        TimingTags::tag_range setup_tags_impl(NodeId node_id) const override { return setup_hold_visitor_.setup_tags(node_id); }
+        TimingTags::tag_range setup_tags_impl(NodeId node_id, TagType type) const override { return setup_hold_visitor_.setup_tags(node_id, type); }
+#ifdef TATUM_CALCULATE_EDGE_SLACKS
+        TimingTags::tag_range setup_edge_slacks_impl(EdgeId edge_id) const override { return setup_hold_visitor_.setup_edge_slacks(edge_id); }
+#endif
+        TimingTags::tag_range setup_node_slacks_impl(NodeId node_id) const override { return setup_hold_visitor_.setup_node_slacks(node_id); }
+
+        TimingTags::tag_range hold_tags_impl(NodeId node_id) const override { return setup_hold_visitor_.hold_tags(node_id); }
+        TimingTags::tag_range hold_tags_impl(NodeId node_id, TagType type) const override { return setup_hold_visitor_.hold_tags(node_id, type); }
+#ifdef TATUM_CALCULATE_EDGE_SLACKS
+        TimingTags::tag_range hold_edge_slacks_impl(EdgeId edge_id) const override { return setup_hold_visitor_.hold_edge_slacks(edge_id); }
+#endif
+        TimingTags::tag_range hold_node_slacks_impl(NodeId node_id) const override { return setup_hold_visitor_.hold_node_slacks(node_id); }
+
+    private:
+        /** Update using the full walker */
+        void update_timing_full_(GraphVisitor& visitor){
+            full_walker_.do_reset(timing_graph_, visitor);
+
+            full_walker_.do_arrival_pre_traversal(timing_graph_, timing_constraints_, visitor);            
+            full_walker_.do_arrival_traversal(timing_graph_, timing_constraints_, delay_calculator_, visitor);            
+
+            full_walker_.do_required_pre_traversal(timing_graph_, timing_constraints_, visitor);            
+            full_walker_.do_required_traversal(timing_graph_, timing_constraints_, delay_calculator_, visitor);            
+
+            full_walker_.do_update_slack(timing_graph_, delay_calculator_, visitor);
+        }
+
+        /** Update using the incremental walker */
+        void update_timing_incr_(GraphVisitor& visitor){
+            if (never_updated_incr_) {
+                //Invalidate all edges
+                for (EdgeId edge : timing_graph_.edges()) {
+                    incr_walker_.invalidate_edge(edge);
+                }
+
+                //Only need to pre-traverse the first update
+                incr_walker_.do_arrival_pre_traversal(timing_graph_, timing_constraints_, visitor);            
+            }
+
+            incr_walker_.do_arrival_traversal(timing_graph_, timing_constraints_, delay_calculator_, visitor);            
+
+            if (never_updated_incr_) {
+                //Only need to pre-traverse the first update
+                incr_walker_.do_required_pre_traversal(timing_graph_, timing_constraints_, visitor);            
+            }
+
+            incr_walker_.do_required_traversal(timing_graph_, timing_constraints_, delay_calculator_, visitor);            
+
+            incr_walker_.do_update_slack(timing_graph_, delay_calculator_, visitor);
+        }
+
+        /* Clear incremental timing info */
+        void clear_timing_incr_(){
+            incr_walker_.clear_invalidated_edges();
+
+            n_modified_edges_ = 0;
+            never_updated_incr_ = false;
+        }
+
+        const TimingGraph& timing_graph_;
+        const TimingConstraints& timing_constraints_;
+        const DelayCalculator& delay_calculator_;
+        SetupHoldAnalysis setup_hold_visitor_;
+
+        FullWalker full_walker_;
+        IncrWalker incr_walker_;
+        enum class Mode { FULL, INCR };
+        Mode mode_;
+
+        bool never_updated_incr_ = true;
+        size_t max_modified_edges_;
+        std::atomic_size_t n_modified_edges_ = 0;
+
+        typedef std::chrono::duration<double> dsec;
+        typedef std::chrono::high_resolution_clock Clock;
+};
+
+}} //namepsace

libs/EXTERNAL/libtatum/libtatum/tatum/analyzers/IncrHoldTimingAnalyzer.hpp

@@ -1,5 +1,5 @@
 #pragma once
-#include "tatum/graph_walkers/SerialWalker.hpp"
+#include "tatum/graph_walkers/SerialIncrWalker.hpp"
 #include "tatum/HoldAnalysis.hpp"
 #include "tatum/analyzers/HoldTimingAnalyzer.hpp"
 #include "tatum/base/validate_timing_graph_constraints.hpp"

libs/EXTERNAL/libtatum/libtatum/tatum/analyzers/IncrSetupHoldTimingAnalyzer.hpp

@@ -1,4 +1,5 @@
 #pragma once
+#include "tatum/graph_walkers/SerialIncrWalker.hpp"
 #include "tatum/graph_walkers/SerialWalker.hpp"
 #include "tatum/SetupHoldAnalysis.hpp"
 #include "tatum/analyzers/SetupHoldTimingAnalyzer.hpp"

libs/EXTERNAL/libtatum/libtatum/tatum/analyzers/IncrSetupTimingAnalyzer.hpp

@@ -1,5 +1,5 @@
 #pragma once
-#include "tatum/graph_walkers/SerialWalker.hpp"
+#include "tatum/graph_walkers/SerialIncrWalker.hpp"
 #include "tatum/SetupAnalysis.hpp"
 #include "tatum/analyzers/SetupTimingAnalyzer.hpp"
 #include "tatum/base/validate_timing_graph_constraints.hpp"

libs/EXTERNAL/libtatum/libtatum/tatum/graph_walkers/ParallelLevelizedWalker.hpp

@@ -11,7 +11,7 @@
 namespace tatum {
 
 /**
- * A parallel timing analyzer which traveres the timing graph in a levelized
+ * A parallel timing analyzer which traverses the timing graph in a levelized
  * manner.  However nodes within each level are processed in parallel using
  * Thread Building Blocks (TBB). If TBB is not available it operates serially and is 
  * equivalent to the SerialWalker.

libs/EXTERNAL/libtatum/libtatum/tatum/graph_walkers/SerialIncrWalker.hpp

@@ -17,14 +17,14 @@ namespace tatum {
  *
  * If TATUM_INCR_BLOCK_INVALIDATION is defined: 
  *      All of a nodes tags associated with an invalidated edge are invalidated.
- *      This is a robust but pessimisitc approach (it invalidates more tags than
+ *      This is a robust but pessimistic approach (it invalidates more tags than
  *      strictly required). As a result all nodes processed will report having been
  *      modified, meaning their decendents/predecessors will also be invalidated
  *      even if in reality the recalculated tags are identical to the previous ones
  *      (i.e. nothing has really changed).
  *
  * Ohterwise, the analyzer performs edge invalidation:
- *      Only node tags which are dominanted by an invalidated edge are invalidated.
+ *      Only node tags which are dominated by an invalidated edge are invalidated.
  *      This is a less pessimistic approach, and means when processed nodes which
  *      don't have any changed tags will report as being unmodified. This significantly
  *      prunes the amount of the timing graph which needs to be updated (as unmodified
@@ -37,7 +37,7 @@ namespace tatum {
  * manner. Unlike SerialWalker it attempts to incrementally (rather than
  * fully) update based on invalidated edges.
  *
- * To performan an incremental traversal, the st of invalidated edges
+ * To perform an incremental traversal, the set of invalidated edges
  * is processed to identify nodes which will need to be re-evaluated for
  * the arrival and/or required traversals.
  *

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_math.h

@@ -4,6 +4,7 @@
 #include <map>
 #include <vector>
 #include <cmath>
+#include <cstdint>
 
 #include "vtr_assert.h"
 
@@ -163,6 +164,17 @@ bool isclose(T a, T b) {
     return isclose<T>(a, b, DEFAULT_REL_TOL, DEFAULT_ABS_TOL);
 }
 
+/** Log2, round down.
+ * From https://stackoverflow.com/a/51351885 */
+static inline uint64_t log2_floor(uint64_t x) {
+    return 63U - __builtin_clzl(x);
+}
+
+/** Log2, round up */
+static inline uint64_t log2_ceil(uint64_t x) {
+    return log2_floor(x - 1) + 1;
+}
+
 } // namespace vtr
 
 #endif

utils/route_diag/src/main.cpp

@@ -117,13 +117,16 @@ static void do_one_route(const Netlist<>& net_list,
                                      -1,
                                      false,
                                      std::unordered_map<RRNodeId, int>());
-    std::tie(found_path, std::ignore, cheapest) = router.timing_driven_route_connection_from_route_tree(tree.root(),
-                                                                                                    sink_node,
-                                                                                                    cost_params,
-                                                                                                    bounding_box,
-                                                                                                    router_stats,
-                                                                                                    conn_params,
-                                                                                                    true);
+    std::tie(found_path, std::ignore, cheapest) = router.timing_driven_route_connection_from_route_tree(
+        tree.root(),
+        tree.root().inode,
+        sink_node,
+        cost_params,
+        bounding_box,
+        router_stats,
+        conn_params,
+        true
+    );
 
     if (found_path) {
         VTR_ASSERT(cheapest.index == sink_node);

vpr/src/base/SetupVPR.cpp

@@ -281,6 +281,9 @@ void SetupVPR(const t_options* Options,
     /* Set seed for pseudo-random placement, default seed to 1 */
     vtr::srandom(PlacerOpts->seed);
 
+    /* Make num_workers available to the router */
+    RouterOpts->num_workers = vpr_setup->num_workers;
+
     {
         vtr::ScopedStartFinishTimer t("Building complex block graph");
         alloc_and_load_all_pb_graphs(PowerOpts->do_power, RouterOpts->flat_routing);

vpr/src/base/read_netlist.cpp

@@ -1057,11 +1057,13 @@ static void load_external_nets_and_cb(ClusteredNetlist& clb_nlist) {
             int logical_pin = clb_nlist.pin_logical_index(pin_id);
             int physical_pin = get_physical_pin(tile_type, block_type, logical_pin);
 
+            /* XXX: Silence warning
             if (tile_type->is_ignored_pin[physical_pin] != is_ignored_net) {
                 VTR_LOG_WARN(
                     "Netlist connects net %s to both global and non-global pins.\n",
                     clb_nlist.net_name(net_id).c_str());
             }
+            */
         }
     }