caiyu/nextpnr
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

router2: rework heuristics

Browse Source
This commit is contained in:
Lofty 2022-09-24 00:17:28 +01:00
parent f4e6bbd383
commit a1ad114069
3 changed files with 46 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
common/kernel/timing.h
View File

@ -99,6 +99,8 @@ struct TimingAnalyser
bool have_loops = false; bool have_loops = false;
bool updated_domains = false; bool updated_domains = false;
void print_fmax();
private: private:
void init_ports(); void init_ports();
void get_cell_delays(); void get_cell_delays();
@ -115,7 +117,6 @@ struct TimingAnalyser
void compute_slack(); void compute_slack();
void compute_criticality(); void compute_criticality();
void print_fmax();
// get the N most failing endpoints for a given domain pair // get the N most failing endpoints for a given domain pair
std::vector<CellPortKey> get_failing_eps(domain_id_t domain_pair, int count); std::vector<CellPortKey> get_failing_eps(domain_id_t domain_pair, int count);
// print the critical path for an endpoint and domain pair // print the critical path for an endpoint and domain pair

2
common/route/router1.cc
View File

@ -899,6 +899,8 @@ bool router1(Context *ctx, const Router1Cfg &cfg)
last_arcs_with_ripup = router.arcs_with_ripup; last_arcs_with_ripup = router.arcs_with_ripup;
last_arcs_without_ripup = router.arcs_without_ripup; last_arcs_without_ripup = router.arcs_without_ripup;
ctx->yield(); ctx->yield();
router.tmg.run();
router.tmg.print_fmax();
#ifndef NDEBUG #ifndef NDEBUG
router.check(); router.check();
#endif #endif

76
common/route/router2.cc
View File

@ -72,9 +72,11 @@ struct Router2
struct WireScore struct WireScore
{ {
float cost; float delay;
float togo_cost; float togo;
float total() const { return cost + togo_cost; } float congest;
float criticality;
float total() const { return (delay * criticality) + (congest * (1.0f - criticality)) + togo; }
}; };
struct PerWireData struct PerWireData
@ -232,7 +234,7 @@ struct Router2
{ {
explicit QueuedWire(int wire = -1, WireScore score = WireScore{}, int randtag = 0) explicit QueuedWire(int wire = -1, WireScore score = WireScore{}, int randtag = 0)
: wire(wire), score(score), randtag(randtag){}; : wire(wire), score(score), randtag(randtag) {};
int wire; int wire;
WireScore score; WireScore score;
@ -242,8 +244,8 @@ struct Router2
{ {
bool operator()(const QueuedWire &lhs, const QueuedWire &rhs) const noexcept bool operator()(const QueuedWire &lhs, const QueuedWire &rhs) const noexcept
{ {
float lhs_score = lhs.score.cost + lhs.score.togo_cost, float lhs_score = lhs.score.total(),
rhs_score = rhs.score.cost + rhs.score.togo_cost; rhs_score = rhs.score.total();
return lhs_score == rhs_score ? lhs.randtag > rhs.randtag : lhs_score > rhs_score; return lhs_score == rhs_score ? lhs.randtag > rhs.randtag : lhs_score > rhs_score;
} }
}; };
@ -350,27 +352,18 @@ struct Router2
ad.routed = false; ad.routed = false;
} }
float score_wire_for_arc(NetInfo *net, store_index<PortRef> user, size_t phys_pin, WireId wire, PipId pip, float score_wire_for_arc_congest(NetInfo *net, store_index<PortRef> user, size_t phys_pin, WireId wire, PipId pip,
float crit_weight) float crit_weight)
{ {
auto &wd = wire_data(wire); auto &wd = wire_data(wire);
auto &nd = nets.at(net->udata); float delay = cfg.get_base_cost(ctx, wire, pip, crit_weight);
float base_cost = cfg.get_base_cost(ctx, wire, pip, crit_weight); return (wd.hist_cong_cost + delay) * (1.0f + float(wd.curr_cong) * 5.0f);
int overuse = wd.curr_cong;
float hist_cost = 1.0f + crit_weight * (wd.hist_cong_cost - 1.0f);
float bias_cost = 0;
int source_uses = 0;
if (nd.wires.count(wire)) {
overuse -= 1;
source_uses = nd.wires.at(wire).second;
} }
float present_cost = 1.0f + overuse * curr_cong_weight * crit_weight;
if (pip != PipId()) { float score_wire_for_arc_delay(NetInfo *net, store_index<PortRef> user, size_t phys_pin, WireId wire, PipId pip,
Loc pl = ctx->getPipLocation(pip); float crit_weight)
bias_cost = cfg.bias_cost_factor * (base_cost / int(net->users.entries())) * {
((std::abs(pl.x - nd.cx) + std::abs(pl.y - nd.cy)) / float(nd.hpwl)); return cfg.get_base_cost(ctx, wire, pip, crit_weight);
}
return base_cost * hist_cost * present_cost / (1 + (source_uses * crit_weight)) + bias_cost;
} }
float get_togo_cost(NetInfo *net, store_index<PortRef> user, int wire, WireId src_sink, bool bwd, float crit_weight) float get_togo_cost(NetInfo *net, store_index<PortRef> user, int wire, WireId src_sink, bool bwd, float crit_weight)
@ -383,7 +376,7 @@ struct Router2
} }
// FIXME: timing/wirelength balance? // FIXME: timing/wirelength balance?
delay_t est_delay = ctx->estimateDelay(bwd ? src_sink : wd.w, bwd ? wd.w : src_sink); delay_t est_delay = ctx->estimateDelay(bwd ? src_sink : wd.w, bwd ? wd.w : src_sink);
return (ctx->getDelayNS(est_delay) / (1 + source_uses * crit_weight)) + cfg.ipin_cost_adder; return ctx->getDelayNS(est_delay) * cfg.estimate_weight;
} }
bool check_arc_routing(NetInfo *net, store_index<PortRef> usr, size_t phys_pin) bool check_arc_routing(NetInfo *net, store_index<PortRef> usr, size_t phys_pin)
@ -647,9 +640,11 @@ struct Router2
// Add 'forward' direction startpoints to queue // Add 'forward' direction startpoints to queue
auto seed_queue_fwd = [&](WireId wire, float wire_cost = 0) { auto seed_queue_fwd = [&](WireId wire, float wire_cost = 0) {
WireScore base_score; WireScore base_score;
base_score.cost = wire_cost; base_score.criticality = crit;
base_score.congest = 0.0;
base_score.delay = wire_cost;
int wire_idx = wire_to_idx.at(wire); int wire_idx = wire_to_idx.at(wire);
base_score.togo_cost = get_togo_cost(net, i, wire_idx, dst_wire, false, crit_weight); base_score.togo = get_togo_cost(net, i, wire_idx, dst_wire, false, crit);
t.fwd_queue.push(QueuedWire(wire_idx, base_score)); t.fwd_queue.push(QueuedWire(wire_idx, base_score));
set_visited_fwd(t, wire_idx, PipId()); set_visited_fwd(t, wire_idx, PipId());
}; };
@ -677,9 +672,11 @@ struct Router2
} }
auto seed_queue_bwd = [&](WireId wire) { auto seed_queue_bwd = [&](WireId wire) {
WireScore base_score; WireScore base_score;
base_score.cost = 0; base_score.criticality = crit;
base_score.congest = 0.0;
base_score.delay = 0;
int wire_idx = wire_to_idx.at(wire); int wire_idx = wire_to_idx.at(wire);
base_score.togo_cost = get_togo_cost(net, i, wire_idx, src_wire, true, crit_weight); base_score.togo = get_togo_cost(net, i, wire_idx, src_wire, true, crit);
t.bwd_queue.push(QueuedWire(wire_idx, base_score)); t.bwd_queue.push(QueuedWire(wire_idx, base_score));
set_visited_bwd(t, wire_idx, PipId()); set_visited_bwd(t, wire_idx, PipId());
}; };
@ -731,9 +728,10 @@ struct Router2
if (!thread_test_wire(t, nwd)) if (!thread_test_wire(t, nwd))
continue; // thread safety issue continue; // thread safety issue
WireScore next_score; WireScore next_score;
next_score.cost = curr.score.cost + score_wire_for_arc(net, i, phys_pin, next, dh, crit_weight); next_score.criticality = crit;
next_score.togo_cost = next_score.congest = curr.score.congest + score_wire_for_arc_congest(net, i, phys_pin, next, dh, crit_weight);
cfg.estimate_weight * get_togo_cost(net, i, next_idx, dst_wire, false, crit_weight); next_score.delay = curr.score.delay + score_wire_for_arc_delay(net, i, phys_pin, next, dh, crit_weight);
next_score.togo = get_togo_cost(net, i, next_idx, dst_wire, false, crit_weight);
set_visited_fwd(t, next_idx, dh); set_visited_fwd(t, next_idx, dh);
t.fwd_queue.push(QueuedWire(next_idx, next_score, t.rng.rng())); t.fwd_queue.push(QueuedWire(next_idx, next_score, t.rng.rng()));
} }
@ -777,9 +775,10 @@ struct Router2
if (!thread_test_wire(t, nwd)) if (!thread_test_wire(t, nwd))
continue; // thread safety issue continue; // thread safety issue
WireScore next_score; WireScore next_score;
next_score.cost = curr.score.cost + score_wire_for_arc(net, i, phys_pin, next, uh, crit_weight); next_score.criticality = crit;
next_score.togo_cost = next_score.congest = curr.score.congest + score_wire_for_arc_congest(net, i, phys_pin, next, uh, crit_weight);
cfg.estimate_weight * get_togo_cost(net, i, next_idx, src_wire, true, crit_weight); next_score.delay = curr.score.delay + score_wire_for_arc_delay(net, i, phys_pin, next, uh, crit_weight);
next_score.togo = get_togo_cost(net, i, next_idx, src_wire, true, crit_weight);
set_visited_bwd(t, next_idx, uh); set_visited_bwd(t, next_idx, uh);
t.bwd_queue.push(QueuedWire(next_idx, next_score, t.rng.rng())); t.bwd_queue.push(QueuedWire(next_idx, next_score, t.rng.rng()));
} }
@ -1422,7 +1421,7 @@ struct Router2
int tmgfail = 0; int tmgfail = 0;
if (timing_driven) if (timing_driven)
tmg.run(false); tmg.run(false);
if (timing_driven_ripup && iter < 500) { if (timing_driven_ripup && iter < 500 && overused_wires == 0) {
for (size_t i = 0; i < nets_by_udata.size(); i++) { for (size_t i = 0; i < nets_by_udata.size(); i++) {
NetInfo *ni = nets_by_udata.at(i); NetInfo *ni = nets_by_udata.at(i);
for (auto usr : ni->users.enumerate()) { for (auto usr : ni->users.enumerate()) {
@ -1433,6 +1432,12 @@ struct Router2
} }
} }
} }
if (overused_wires > 0) {
failed_nets.clear();
for (size_t i = 0; i < nets_by_udata.size(); i++) {
failed_nets.insert(i);
}
}
if (overused_wires == 0 && tmgfail == 0) { if (overused_wires == 0 && tmgfail == 0) {
// Try and actually bind nextpnr Arch API wires // Try and actually bind nextpnr Arch API wires
bind_and_check_all(); bind_and_check_all();
@ -1450,6 +1455,7 @@ struct Router2
++iter; ++iter;
if (curr_cong_weight < 1e9) if (curr_cong_weight < 1e9)
curr_cong_weight += cfg.curr_cong_mult; curr_cong_weight += cfg.curr_cong_mult;
tmg.print_fmax();
} while (!failed_nets.empty()); } while (!failed_nets.empty());
if (cfg.perf_profile) { if (cfg.perf_profile) {
std::vector<std::pair<int, IdString>> nets_by_runtime; std::vector<std::pair<int, IdString>> nets_by_runtime;