vpr: Fix slow look-up of clb to atom pins mapping during placement

Previously this mapping was needlessly recalculated once per temperature
in-order to update clustered net criticalities.

We now calculate this mapping once at the beginning of placement, and
re-use it.  This yeilds ~22% placement speed-up on the LU8 benchmark.

Also update the timing driven router to use the faster look-up.

Author: kmurray

libs/libvtrutil/src/vtr_vector.h

@@ -1,6 +1,7 @@
 #ifndef VTR_VECTOR
 #define VTR_VECTOR
 #include <vector>
+#include <cstddef>
 
 namespace vtr {
 

vpr/src/base/clustered_netlist_utils.cpp

@@ -0,0 +1,24 @@
+#include "clustered_netlist_utils.h"
+ClusteredPinAtomPinsLookup::ClusteredPinAtomPinsLookup(const ClusteredNetlist& clustered_netlist, const IntraLbPbPinLookup& pb_gpin_lookup) {
+    init_lookup(clustered_netlist, pb_gpin_lookup);
+}
+
+ClusteredPinAtomPinsLookup::atom_pin_range ClusteredPinAtomPinsLookup::connected_atom_pins(ClusterPinId clustered_pin) const {
+    VTR_ASSERT(clustered_pin);
+    //return vtr::make_range(clustered_pin_connected_atom_pins_[clustered_pin]);
+    return vtr::make_range(clustered_pin_connected_atom_pins_[clustered_pin].begin(),
+                           clustered_pin_connected_atom_pins_[clustered_pin].end());
+    //return atom_pin_range(clustered_pin_connected_atom_pins_[clustered_pin].begin(), clustered_pin_connected_atom_pins_[clustered_pin].end());
+}
+
+void ClusteredPinAtomPinsLookup::init_lookup(const ClusteredNetlist& clustered_netlist, const IntraLbPbPinLookup& pb_gpin_lookup) {
+    auto clustered_pins = clustered_netlist.pins();
+    clustered_pin_connected_atom_pins_.clear();
+    clustered_pin_connected_atom_pins_.resize(clustered_pins.size());
+    for (ClusterPinId clustered_pin : clustered_pins) {
+
+        auto clustered_block = clustered_netlist.pin_block(clustered_pin);
+        int phys_pin_index = clustered_netlist.pin_physical_index(clustered_pin);
+        clustered_pin_connected_atom_pins_[clustered_pin] = find_clb_pin_connected_atom_pins(clustered_block, phys_pin_index, pb_gpin_lookup);
+    }
+}

vpr/src/base/clustered_netlist_utils.h

@@ -0,0 +1,27 @@
+#ifndef CLUSTERED_NETLIST_UTILS_H
+#define CLUSTERED_NETLIST_UTILS_H
+
+#include "vtr_vector.h"
+#include "vtr_range.h"
+
+#include "vpr_utils.h"
+#include "atom_netlist_fwd.h"
+#include "clustered_netlist_fwd.h"
+
+class ClusteredPinAtomPinsLookup {
+    public:
+        typedef std::vector<AtomPinId>::const_iterator      atom_pin_iterator;
+        typedef typename vtr::Range<atom_pin_iterator>   atom_pin_range;
+    public:
+        ClusteredPinAtomPinsLookup(const ClusteredNetlist& clustered_netlist, const IntraLbPbPinLookup& pb_gpin_lookup);
+
+        atom_pin_range connected_atom_pins(ClusterPinId clustered_pin) const;
+
+    private:
+        void init_lookup(const ClusteredNetlist& clustered_netlist, const IntraLbPbPinLookup& pb_gpin_lookup);
+
+    private:
+        vtr::vector<ClusterPinId,std::vector<AtomPinId>> clustered_pin_connected_atom_pins_;
+};
+
+#endif

vpr/src/base/netlist.h

@@ -561,6 +561,7 @@ class Netlist {
         //Returns the net associated with the specified pin
         NetId       pin_net(const PinId pin_id) const;
 
+        //Returns the index of the specified pin within it's connected net
         int         pin_net_index(const PinId pin_id) const;
 
         //Returns the port associated with the specified pin

vpr/src/place/place.cpp

@@ -297,7 +297,7 @@ static void outer_loop_recompute_criticalities(t_placer_opts placer_opts,
 	float * place_delay_value, float * timing_cost, float * delay_cost,
 	int * outer_crit_iter_count, float * inverse_prev_timing_cost,
 	float * inverse_prev_bb_cost,
-    const IntraLbPbPinLookup& pb_gpin_lookup,
+    const ClusteredPinAtomPinsLookup& netlist_pin_lookup,
 #ifdef ENABLE_CLASSIC_VPR_STA
     t_slack* slacks,
     t_timing_inf timing_inf,
@@ -313,7 +313,7 @@ static void placement_inner_loop(float t, float rlim, t_placer_opts placer_opts,
     t_slack* slacks,
     t_timing_inf timing_inf,
 #endif
-    const IntraLbPbPinLookup& pb_gpin_lookup,
+    const ClusteredPinAtomPinsLookup& netlist_pin_lookup,
     SetupTimingInfo& timing_info);
 
 /*****************************************************************************/
@@ -385,6 +385,8 @@ void try_place(t_placer_opts placer_opts,
 	initial_placement(placer_opts.pad_loc_type, placer_opts.pad_loc_file.c_str());
 	init_draw_coords((float) width_fac);
 
+    //Enables fast look-up of atom pins connect to CLB pins
+    ClusteredPinAtomPinsLookup netlist_pin_lookup(cluster_ctx.clb_nlist, pb_gpin_lookup);
 
 	/* Gets initial cost and loads bounding boxes. */
 
@@ -414,7 +416,7 @@ void try_place(t_placer_opts placer_opts,
         timing_info->set_warn_unconstrained(false); //Don't warn again about unconstrained nodes again during placement
 
         //Initial slack estimates
-        load_criticalities(*timing_info, crit_exponent, pb_gpin_lookup);
+        load_criticalities(*timing_info, crit_exponent, netlist_pin_lookup);
 
         critical_path = timing_info->least_slack_critical_path();
 
@@ -572,7 +574,7 @@ void try_place(t_placer_opts placer_opts,
 		outer_loop_recompute_criticalities(placer_opts, num_connections,
 			crit_exponent, bb_cost, &place_delay_value, &timing_cost, &delay_cost,
 			&outer_crit_iter_count, &inverse_prev_timing_cost, &inverse_prev_bb_cost,
-            pb_gpin_lookup,
+            netlist_pin_lookup,
 #ifdef ENABLE_CLASSIC_VPR_STA
             slacks,
             timing_inf,
@@ -586,7 +588,7 @@ void try_place(t_placer_opts placer_opts,
             slacks,
             timing_inf,
 #endif
-            pb_gpin_lookup,
+            netlist_pin_lookup,
             *timing_info);
 
 		/* Lines below prevent too much round-off error from accumulating *
@@ -694,7 +696,7 @@ void try_place(t_placer_opts placer_opts,
 	outer_loop_recompute_criticalities(placer_opts, num_connections,
 			crit_exponent, bb_cost, &place_delay_value, &timing_cost, &delay_cost,
 			&outer_crit_iter_count, &inverse_prev_timing_cost, &inverse_prev_bb_cost,
-            pb_gpin_lookup,
+            netlist_pin_lookup,
 #ifdef ENABLE_CLASSIC_VPR_STA
             slacks,
             timing_inf,
@@ -712,7 +714,7 @@ void try_place(t_placer_opts placer_opts,
             slacks,
             timing_inf,
 #endif
-            pb_gpin_lookup,
+            netlist_pin_lookup,
             *timing_info);
 
 	tot_iter += move_lim;
@@ -868,7 +870,7 @@ static void outer_loop_recompute_criticalities(t_placer_opts placer_opts,
 	float * place_delay_value, float * timing_cost, float * delay_cost,
 	int * outer_crit_iter_count, float * inverse_prev_timing_cost,
 	float * inverse_prev_bb_cost, 
-    const IntraLbPbPinLookup& pb_gpin_lookup,
+    const ClusteredPinAtomPinsLookup& netlist_pin_lookup,
 #ifdef ENABLE_CLASSIC_VPR_STA
     t_slack* slacks,
     t_timing_inf timing_inf,
@@ -892,7 +894,7 @@ static void outer_loop_recompute_criticalities(t_placer_opts placer_opts,
 
         //Per-temperature timing update
         timing_info.update();
-		load_criticalities(timing_info, crit_exponent, pb_gpin_lookup);
+		load_criticalities(timing_info, crit_exponent, netlist_pin_lookup);
 
 #ifdef ENABLE_CLASSIC_VPR_STA
         load_timing_graph_net_delays(point_to_point_delay_cost);
@@ -922,7 +924,7 @@ static void placement_inner_loop(float t, float rlim, t_placer_opts placer_opts,
 	t_slack* slacks,
 	t_timing_inf timing_inf,
 #endif
-	const IntraLbPbPinLookup& pb_gpin_lookup,
+	const ClusteredPinAtomPinsLookup& netlist_pin_lookup,
 	SetupTimingInfo& timing_info) {
 
 	int inner_crit_iter_count, inner_iter;
@@ -978,7 +980,7 @@ static void placement_inner_loop(float t, float rlim, t_placer_opts placer_opts,
 				 */
 				 //Inner loop timing update
 				timing_info.update();
-				load_criticalities(timing_info, crit_exponent, pb_gpin_lookup);
+				load_criticalities(timing_info, crit_exponent, netlist_pin_lookup);
 
 #ifdef ENABLE_CLASSIC_VPR_STA
 				load_timing_graph_net_delays(point_to_point_delay_cost);

vpr/src/place/timing_place.cpp

@@ -49,7 +49,7 @@ static void free_crit(vtr::t_chunk *chunk_list_ptr){
 }
 
 /**************************************/
-void load_criticalities(SetupTimingInfo& timing_info, float crit_exponent, const IntraLbPbPinLookup& pb_gpin_lookup) {
+void load_criticalities(SetupTimingInfo& timing_info, float crit_exponent, const ClusteredPinAtomPinsLookup& pin_lookup) {
 	/* Performs a 1-to-1 mapping from criticality to f_timing_place_crit.  
 	  For every pin on every net (or, equivalently, for every tedge ending 
 	  in that pin), f_timing_place_crit = criticality^(criticality exponent) */
@@ -58,14 +58,17 @@ void load_criticalities(SetupTimingInfo& timing_info, float crit_exponent, const
 	for (auto net_id : cluster_ctx.clb_nlist.nets()) {
 		if (cluster_ctx.clb_nlist.net_is_global(net_id))
 			continue;
-		for (size_t ipin = 1; ipin < cluster_ctx.clb_nlist.net_pins(net_id).size(); ipin++) {
-            float clb_pin_crit = calculate_clb_net_pin_criticality(timing_info, pb_gpin_lookup, net_id, ipin);
+
+        for (auto clb_pin : cluster_ctx.clb_nlist.net_sinks(net_id)) {
+            int ipin = cluster_ctx.clb_nlist.pin_net_index(clb_pin);
+
+            float clb_pin_crit = calculate_clb_net_pin_criticality(timing_info, pin_lookup, clb_pin);
 
             /* The placer likes a great deal of contrast between criticalities. 
             Since path criticality varies much more than timing, we "sharpen" timing 
             criticality by taking it to some power, crit_exponent (between 1 and 8 by default). */
             f_timing_place_crit[net_id][ipin] = pow(clb_pin_crit, crit_exponent);
-		}
+        }
 	}
 }
 

vpr/src/place/timing_place.h

@@ -2,6 +2,7 @@
 #define TIMING_PLACE
 
 #include "timing_info_fwd.h"
+#include "clustered_netlist_utils.h"
 
 void alloc_lookups_and_criticalities(t_chan_width_dist chan_width_dist,
 		t_router_opts router_opts,
@@ -11,7 +12,7 @@ void alloc_lookups_and_criticalities(t_chan_width_dist chan_width_dist,
 
 void free_lookups_and_criticalities();
 
-void load_criticalities(SetupTimingInfo& timing_info, float crit_exponent, const IntraLbPbPinLookup& pb_gpin_lookup);
+void load_criticalities(SetupTimingInfo& timing_info, float crit_exponent, const ClusteredPinAtomPinsLookup& pin_lookup);
 
 float get_timing_place_crit(ClusterNetId net_id, int ipin);
 void set_timing_place_crit(ClusterNetId net_id, int ipin, float val);

vpr/src/route/connection_based_routing.h

@@ -120,7 +120,7 @@ class Connection_based_routing_resources {
 	// and if so, mark them to be rerouted
 	bool forcibly_reroute_connections(float max_criticality, 
             std::shared_ptr<const SetupTimingInfo> timing_info,
-            const IntraLbPbPinLookup& pb_gpin_lookup,
+            const ClusteredPinAtomPinsLookup& netlist_pin_lookup,
             vtr::vector_map<ClusterNetId, float *> &net_delay);
 
 };