diff options
-rw-r--r-- | common/nextpnr.h | 1 | ||||
-rw-r--r-- | common/placer1.cc | 11 | ||||
-rw-r--r-- | common/router1.cc | 3 | ||||
-rw-r--r-- | common/timing.cc | 108 | ||||
-rw-r--r-- | common/timing.h | 2 | ||||
-rw-r--r-- | ecp5/arch.cc | 5 | ||||
-rw-r--r-- | ecp5/arch.h | 1 | ||||
-rw-r--r-- | ecp5/main.cc | 7 | ||||
-rw-r--r-- | generic/arch.cc | 5 | ||||
-rw-r--r-- | generic/arch.h | 1 | ||||
-rw-r--r-- | ice40/arch.cc | 6 | ||||
-rw-r--r-- | ice40/arch.h | 1 | ||||
-rw-r--r-- | ice40/main.cc | 7 |
13 files changed, 141 insertions, 17 deletions
diff --git a/common/nextpnr.h b/common/nextpnr.h index 1cce21c0..4d9cf5f7 100644 --- a/common/nextpnr.h +++ b/common/nextpnr.h @@ -466,6 +466,7 @@ struct Context : Arch, DeterministicRNG bool force = false; bool timing_driven = true; float target_freq = 12e6; + bool user_freq = false; Context(ArchArgs args) : Arch(args) {} diff --git a/common/placer1.cc b/common/placer1.cc index 4659da11..30f0057c 100644 --- a/common/placer1.cc +++ b/common/placer1.cc @@ -155,6 +155,7 @@ class SAPlacer ctx->unlock(); int n_no_progress = 0; + wirelen_t min_metric = curr_metric; double avg_metric = curr_metric; temp = 10000; @@ -180,6 +181,11 @@ class SAPlacer } } + if (curr_metric < min_metric) { + min_metric = curr_metric; + improved = true; + } + // Heuristic to improve placement on the 8k if (improved) n_no_progress = 0; @@ -234,6 +240,9 @@ class SAPlacer ctx->shuffle(autoplaced); assign_budget(ctx); } + else { + update_budget(ctx); + } // Recalculate total metric entirely to avoid rounding errors // accumulating over time @@ -382,8 +391,6 @@ class SAPlacer // SA acceptance criterea if (delta < 0 || (temp > 1e-6 && (ctx->rng() / float(0x3fffffff)) <= std::exp(-delta / temp))) { n_accept++; - if (delta < 2) - improved = true; } else { if (other != IdString()) ctx->unbindBel(oldBel); diff --git a/common/router1.cc b/common/router1.cc index fbf3c467..dbf97af7 100644 --- a/common/router1.cc +++ b/common/router1.cc @@ -22,6 +22,7 @@ #include "log.h" #include "router1.h" +#include "timing.h" namespace { @@ -615,6 +616,8 @@ bool router1(Context *ctx) if (ctx->verbose || iterCnt == 1) log_info("routing queue contains %d jobs.\n", int(jobQueue.size())); + update_budget(ctx); + bool printNets = ctx->verbose && (jobQueue.size() < 10); while (!jobQueue.empty()) { diff --git a/common/timing.cc b/common/timing.cc index d91dea20..2c467961 100644 --- a/common/timing.cc +++ b/common/timing.cc @@ -22,19 +22,23 @@ #include <unordered_map> #include <utility> #include "log.h" +#include "util.h" NEXTPNR_NAMESPACE_BEGIN -static delay_t follow_net(Context *ctx, NetInfo *net, int path_length, delay_t slack); +typedef std::unordered_map<const PortInfo*, delay_t> UpdateMap; + +static delay_t follow_net(Context *ctx, NetInfo *net, int path_length, delay_t slack, UpdateMap &updates, delay_t &min_slack); // Follow a path, returning budget to annotate -static delay_t follow_user_port(Context *ctx, PortRef &user, int path_length, delay_t slack) +static delay_t follow_user_port(Context *ctx, PortRef &user, int path_length, delay_t slack, UpdateMap &updates, delay_t &min_slack) { delay_t value; if (ctx->getPortClock(user.cell, user.port) != IdString()) { // At the end of a timing path (arguably, should check setup time // here too) value = slack / path_length; + min_slack = std::min(min_slack, value); } else { // Default to the path ending here, if no further paths found value = slack / path_length; @@ -47,7 +51,7 @@ static delay_t follow_user_port(Context *ctx, PortRef &user, int path_length, de if (is_path) { NetInfo *net = port.second.net; if (net) { - delay_t path_budget = follow_net(ctx, net, path_length, slack - comb_delay); + delay_t path_budget = follow_net(ctx, net, path_length, slack - comb_delay, updates, min_slack); value = std::min(value, path_budget); } } @@ -55,23 +59,28 @@ static delay_t follow_user_port(Context *ctx, PortRef &user, int path_length, de } } - if (value < user.budget) { - user.budget = value; + auto ret = updates.emplace(&user.cell->ports.at(user.port), value); + if (!ret.second && value < ret.first->second) { + ret.first->second = value; } return value; } -static delay_t follow_net(Context *ctx, NetInfo *net, int path_length, delay_t slack) +static delay_t follow_net(Context *ctx, NetInfo *net, int path_length, delay_t slack, UpdateMap &updates, delay_t &min_slack) { delay_t net_budget = slack / (path_length + 1); - for (auto &usr : net->users) { - net_budget = std::min(net_budget, follow_user_port(ctx, usr, path_length + 1, slack)); + for (unsigned i = 0; i < net->users.size(); ++i) { + auto &usr = net->users[i]; + net_budget = std::min(net_budget, follow_user_port(ctx, usr, path_length + 1, slack - ctx->getNetinfoRouteDelay(net, i), updates, min_slack)); } return net_budget; } void assign_budget(Context *ctx) { + UpdateMap updates; + delay_t min_slack = delay_t(1.0e12 / ctx->target_freq); + log_break(); log_info("Annotating ports with timing budgets\n"); // Clear delays to a very high value first @@ -81,6 +90,7 @@ void assign_budget(Context *ctx) usr.budget = default_slack; } } + min_slack = default_slack; // Go through all clocked drivers and set up paths for (auto &cell : ctx->cells) { for (auto port : cell.second->ports) { @@ -91,17 +101,33 @@ void assign_budget(Context *ctx) delay_t clkToQ; if (ctx->getCellDelay(cell.second.get(), clock_domain, port.first, clkToQ)) slack -= clkToQ; - if (port.second.net) - follow_net(ctx, port.second.net, 0, slack); + if (port.second.net) { + log_break(); + follow_net(ctx, port.second.net, 0, slack, updates, min_slack); + } } } } } - // Post-allocation check + if (!ctx->user_freq) { + ctx->target_freq = 1e12 / (default_slack - min_slack); + if (ctx->verbose) + log_info("minimum slack for this assign = %d, target Fmax for next update = %f\n", min_slack, ctx->target_freq/1e6); + } + + // Update the budgets for (auto &net : ctx->nets) { - for (auto user : net.second->users) { - if (user.budget < 0) + for (size_t i = 0; i < net.second->users.size(); ++i) { + auto& user = net.second->users[i]; + auto pi = &user.cell->ports.at(user.port); + auto it = updates.find(pi); + if (it == updates.end()) continue; + auto budget = ctx->getNetinfoRouteDelay(net.second.get(), i) + it->second; + user.budget = ctx->getBudgetOverride(net.second->driver, budget); + + // Post-update check + if (ctx->user_freq && user.budget < 0) log_warning("port %s.%s, connected to net '%s', has negative " "timing budget of %fns\n", user.cell->name.c_str(ctx), user.port.c_str(ctx), net.first.c_str(ctx), @@ -117,4 +143,60 @@ void assign_budget(Context *ctx) log_info("Checksum: 0x%08x\n", ctx->checksum()); } +void update_budget(Context *ctx) +{ + UpdateMap updates; + delay_t default_slack = delay_t(1.0e12 / ctx->target_freq); + delay_t min_slack = delay_t(1.0e12 / ctx->target_freq); + + // Go through all clocked drivers and distribute the available path slack evenly into every budget + for (auto &cell : ctx->cells) { + for (auto& port : cell.second->ports) { + if (port.second.type == PORT_OUT) { + IdString clock_domain = ctx->getPortClock(cell.second.get(), port.first); + if (clock_domain != IdString()) { + delay_t slack = default_slack; // TODO: clock constraints + delay_t clkToQ; + if (ctx->getCellDelay(cell.second.get(), clock_domain, port.first, clkToQ)) + slack -= clkToQ; + if (port.second.net) + follow_net(ctx, port.second.net, 0, slack, updates, min_slack); + } + } + } + } + + if (!ctx->user_freq) { + ctx->target_freq = 1.05 * (1e12 / (default_slack - min_slack)); + if (ctx->verbose) + log_info("minimum slack for this update = %d, target Fmax for next update = %f\n", min_slack, ctx->target_freq/1e6); + } + + // Update the budgets + for (auto &net : ctx->nets) { + for (size_t i = 0; i < net.second->users.size(); ++i) { + auto& user = net.second->users[i]; + auto pi = &user.cell->ports.at(user.port); + auto it = updates.find(pi); + if (it == updates.end()) continue; + auto budget = ctx->getNetinfoRouteDelay(net.second.get(), i) + it->second; + user.budget = ctx->getBudgetOverride(net.second->driver, budget); + + // Post-update check + if (ctx->verbose) { + if (ctx->user_freq && user.budget < 0) + log_warning("port %s.%s, connected to net '%s', has negative " + "timing budget of %fns\n", + user.cell->name.c_str(ctx), user.port.c_str(ctx), net.first.c_str(ctx), + ctx->getDelayNS(user.budget)); + else + log_info("port %s.%s, connected to net '%s', has " + "timing budget of %fns\n", + user.cell->name.c_str(ctx), user.port.c_str(ctx), net.first.c_str(ctx), + ctx->getDelayNS(user.budget)); + } + } + } +} + NEXTPNR_NAMESPACE_END diff --git a/common/timing.h b/common/timing.h index 025e4a76..b5574392 100644 --- a/common/timing.h +++ b/common/timing.h @@ -27,6 +27,8 @@ NEXTPNR_NAMESPACE_BEGIN // Assign "budget" values for all user ports in the design void assign_budget(Context *ctx); +void update_budget(Context *ctx); + NEXTPNR_NAMESPACE_END #endif diff --git a/ecp5/arch.cc b/ecp5/arch.cc index 55fe5704..b518323e 100644 --- a/ecp5/arch.cc +++ b/ecp5/arch.cc @@ -413,6 +413,11 @@ delay_t Arch::estimateDelay(WireId src, WireId dst) const return 200 * (abs(src.location.x - dst.location.x) + abs(src.location.y - dst.location.y)); } +delay_t Arch::getBudgetOverride(const PortRef& pr, delay_t v) const +{ + return v; +} + // ----------------------------------------------------------------------- bool Arch::place() { return placer1(getCtx()); } diff --git a/ecp5/arch.h b/ecp5/arch.h index b5f3d817..22016fae 100644 --- a/ecp5/arch.h +++ b/ecp5/arch.h @@ -776,6 +776,7 @@ struct Arch : BaseCtx delay_t getRipupDelayPenalty() const { return 200; } float getDelayNS(delay_t v) const { return v * 0.001; } uint32_t getDelayChecksum(delay_t v) const { return v; } + delay_t getBudgetOverride(const PortRef& pr, delay_t v) const; // ------------------------------------------------- diff --git a/ecp5/main.cc b/ecp5/main.cc index f2db74d7..170f9c11 100644 --- a/ecp5/main.cc +++ b/ecp5/main.cc @@ -167,8 +167,13 @@ int main(int argc, char *argv[]) if (!ctx->pack() && !ctx->force) log_error("Packing design failed.\n"); - if (vm.count("freq")) + if (vm.count("freq")) { ctx->target_freq = vm["freq"].as<double>() * 1e6; + ctx->user_freq = true; + } + else { + log_warning("Target frequency not specified. Will optimise for max frequency.\n"); + } assign_budget(ctx.get()); ctx->check(); print_utilisation(ctx.get()); diff --git a/generic/arch.cc b/generic/arch.cc index d7401356..a12b7380 100644 --- a/generic/arch.cc +++ b/generic/arch.cc @@ -396,6 +396,11 @@ delay_t Arch::estimateDelay(WireId src, WireId dst) const return (dx + dy) * grid_distance_to_delay; } +delay_t Arch::getBudgetOverride(const PortRef& pr, delay_t v) const +{ + return v; +} + // --------------------------------------------------------------- bool Arch::place() { return placer1(getCtx()); } diff --git a/generic/arch.h b/generic/arch.h index 97ed1ac2..43f43842 100644 --- a/generic/arch.h +++ b/generic/arch.h @@ -196,6 +196,7 @@ struct Arch : BaseCtx delay_t getRipupDelayPenalty() const { return 1.0; } float getDelayNS(delay_t v) const { return v; } uint32_t getDelayChecksum(delay_t v) const { return 0; } + delay_t getBudgetOverride(const PortRef& pr, delay_t v) const; bool pack() { return true; } bool place(); diff --git a/ice40/arch.cc b/ice40/arch.cc index dedc59bc..daf0c78b 100644 --- a/ice40/arch.cc +++ b/ice40/arch.cc @@ -495,6 +495,12 @@ delay_t Arch::estimateDelay(WireId src, WireId dst) const return xscale * abs(xd) + yscale * abs(yd) + offset; } +delay_t Arch::getBudgetOverride(const PortRef& pr, delay_t v) const +{ + if (pr.port == id("COUT")) return 0; + return v; +} + // ----------------------------------------------------------------------- bool Arch::place() { return placer1(getCtx()); } diff --git a/ice40/arch.h b/ice40/arch.h index 123b408c..a9392f6d 100644 --- a/ice40/arch.h +++ b/ice40/arch.h @@ -684,6 +684,7 @@ struct Arch : BaseCtx delay_t getRipupDelayPenalty() const { return 200; } float getDelayNS(delay_t v) const { return v * 0.001; } uint32_t getDelayChecksum(delay_t v) const { return v; } + delay_t getBudgetOverride(const PortRef& pr, delay_t v) const; // ------------------------------------------------- diff --git a/ice40/main.cc b/ice40/main.cc index 6201831a..5b658288 100644 --- a/ice40/main.cc +++ b/ice40/main.cc @@ -363,8 +363,13 @@ int main(int argc, char *argv[]) } } - if (vm.count("freq")) + if (vm.count("freq")) { ctx->target_freq = vm["freq"].as<double>() * 1e6; + ctx->user_freq = true; + } + else { + log_warning("Target frequency not specified. Will optimise for max frequency.\n"); + } ctx->timing_driven = true; if (vm.count("no-tmdriv")) |