diff options
Diffstat (limited to 'common/place/timing_opt.cc')
| -rw-r--r-- | common/place/timing_opt.cc | 561 |
1 files changed, 561 insertions, 0 deletions
diff --git a/common/place/timing_opt.cc b/common/place/timing_opt.cc new file mode 100644 index 00000000..f9246292 --- /dev/null +++ b/common/place/timing_opt.cc @@ -0,0 +1,561 @@ +/* + * nextpnr -- Next Generation Place and Route + * + * Copyright (C) 2018 gatecat <gatecat@ds0.me> + * + * Permission to use, copy, modify, and/or distribute this software for any + * purpose with or without fee is hereby granted, provided that the above + * copyright notice and this permission notice appear in all copies. + * + * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES + * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF + * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR + * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES + * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN + * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF + * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. + * + */ + +/* + * Timing-optimised detailed placement algorithm using BFS of the neighbour graph created from cells + * on a critical path + * + * Based on "An Effective Timing-Driven Detailed Placement Algorithm for FPGAs" + * https://www.cerc.utexas.edu/utda/publications/C205.pdf + * + * Modifications made to deal with the smaller Bels that nextpnr uses instead of swapping whole tiles, + * and deal with the fact that not every cell on the crit path may be swappable. + */ + +#include "timing_opt.h" +#include <boost/range/adaptor/reversed.hpp> +#include <queue> +#include "nextpnr.h" +#include "timing.h" +#include "util.h" + +NEXTPNR_NAMESPACE_BEGIN + +class TimingOptimiser +{ + public: + TimingOptimiser(Context *ctx, TimingOptCfg cfg) : ctx(ctx), cfg(cfg), tmg(ctx){}; + bool optimise() + { + log_info("Running timing-driven placement optimisation...\n"); + ctx->lock(); + if (ctx->verbose) + timing_analysis(ctx, false, true, false, false); + tmg.setup(); + for (int i = 0; i < 30; i++) { + log_info(" Iteration %d...\n", i); + tmg.run(); + setup_delay_limits(); + auto crit_paths = find_crit_paths(0.98, 50000); + for (auto &path : crit_paths) + optimise_path(path); + if (ctx->verbose) + timing_analysis(ctx, false, true, false, false); + } + ctx->unlock(); + return true; + } + + private: + void setup_delay_limits() + { + max_net_delay.clear(); + for (auto &net : ctx->nets) { + NetInfo *ni = net.second.get(); + if (ni->driver.cell == nullptr) + continue; + for (auto usr : ni->users) { + max_net_delay[std::make_pair(usr.cell->name, usr.port)] = std::numeric_limits<delay_t>::max(); + } + for (auto usr : ni->users) { + delay_t net_delay = ctx->getNetinfoRouteDelay(ni, usr); + delay_t slack = tmg.get_setup_slack(CellPortKey(usr)); + delay_t domain_slack = tmg.get_domain_setup_slack(CellPortKey(usr)); + if (slack == std::numeric_limits<delay_t>::max()) + continue; + max_net_delay[std::make_pair(usr.cell->name, usr.port)] = net_delay + ((slack - domain_slack) / 10); + } + } + } + + bool check_cell_delay_limits(CellInfo *cell) + { + for (const auto &port : cell->ports) { + int nc; + if (ctx->getPortTimingClass(cell, port.first, nc) == TMG_IGNORE) + continue; + NetInfo *net = port.second.net; + if (net == nullptr) + continue; + if (port.second.type == PORT_IN) { + if (net->driver.cell == nullptr || net->driver.cell->bel == BelId()) + continue; + for (auto user : net->users) { + if (user.cell == cell && user.port == port.first) { + if (ctx->predictArcDelay(net, user) > + 1.1 * max_net_delay.at(std::make_pair(cell->name, port.first))) + return false; + } + } + + } else if (port.second.type == PORT_OUT) { + for (auto user : net->users) { + // This could get expensive for high-fanout nets?? + BelId dstBel = user.cell->bel; + if (dstBel == BelId()) + continue; + if (ctx->predictArcDelay(net, user) > + 1.1 * max_net_delay.at(std::make_pair(user.cell->name, user.port))) { + + return false; + } + } + } + } + return true; + } + + BelId cell_swap_bel(CellInfo *cell, BelId newBel) + { + BelId oldBel = cell->bel; + if (oldBel == newBel) + return oldBel; + CellInfo *other_cell = ctx->getBoundBelCell(newBel); + NPNR_ASSERT(other_cell == nullptr || other_cell->belStrength <= STRENGTH_WEAK); + ctx->unbindBel(oldBel); + if (other_cell != nullptr) { + ctx->unbindBel(newBel); + ctx->bindBel(oldBel, other_cell, STRENGTH_WEAK); + } + ctx->bindBel(newBel, cell, STRENGTH_WEAK); + return oldBel; + } + + // Check that a series of moves are both legal and remain within maximum delay bounds + // Moves are specified as a vector of pairs <cell, oldBel> + bool acceptable_move(std::vector<std::pair<CellInfo *, BelId>> &move, bool check_delays = true) + { + for (auto &entry : move) { + if (!ctx->isBelLocationValid(entry.first->bel)) + return false; + if (!ctx->isBelLocationValid(entry.second)) + return false; + if (!check_delays) + continue; + if (!check_cell_delay_limits(entry.first)) + return false; + // We might have swapped another cell onto the original bel. Check this for max delay violations + // too + CellInfo *swapped = ctx->getBoundBelCell(entry.second); + if (swapped != nullptr && !check_cell_delay_limits(swapped)) + return false; + } + return true; + } + + int find_neighbours(CellInfo *cell, IdString prev_cell, int d, bool allow_swap) + { + BelId curr = cell->bel; + Loc curr_loc = ctx->getBelLocation(curr); + int found_count = 0; + cell_neighbour_bels[cell->name] = pool<BelId>{}; + for (int dy = -d; dy <= d; dy++) { + for (int dx = -d; dx <= d; dx++) { + // Go through all the Bels at this location + // First, find all bels of the correct type that are either unbound or bound normally + // Strongly bound bels are ignored + // FIXME: This means that we cannot touch carry chains or similar relatively constrained macros + std::vector<BelId> free_bels_at_loc; + std::vector<BelId> bound_bels_at_loc; + for (auto bel : ctx->getBelsByTile(curr_loc.x + dx, curr_loc.y + dy)) { + if (!ctx->isValidBelForCellType(cell->type, bel)) + continue; + CellInfo *bound = ctx->getBoundBelCell(bel); + if (bound == nullptr) { + free_bels_at_loc.push_back(bel); + } else if (bound->belStrength <= STRENGTH_WEAK && bound->cluster == ClusterId()) { + bound_bels_at_loc.push_back(bel); + } + } + BelId candidate; + + while (!free_bels_at_loc.empty() || !bound_bels_at_loc.empty()) { + BelId try_bel; + if (!free_bels_at_loc.empty()) { + int try_idx = ctx->rng(int(free_bels_at_loc.size())); + try_bel = free_bels_at_loc.at(try_idx); + free_bels_at_loc.erase(free_bels_at_loc.begin() + try_idx); + } else { + int try_idx = ctx->rng(int(bound_bels_at_loc.size())); + try_bel = bound_bels_at_loc.at(try_idx); + bound_bels_at_loc.erase(bound_bels_at_loc.begin() + try_idx); + } + if (bel_candidate_cells.count(try_bel) && !allow_swap) { + // Overlap is only allowed if it is with the previous cell (this is handled by removing those + // edges in the graph), or if allow_swap is true to deal with cases where overlap means few + // neighbours are identified + if (bel_candidate_cells.at(try_bel).size() > 1 || + (bel_candidate_cells.at(try_bel).size() == 1 && + *(bel_candidate_cells.at(try_bel).begin()) != prev_cell)) + continue; + } + // TODO: what else to check here? + candidate = try_bel; + break; + } + + if (candidate != BelId()) { + cell_neighbour_bels[cell->name].insert(candidate); + bel_candidate_cells[candidate].insert(cell->name); + // Work out if we need to delete any overlap + std::vector<IdString> overlap; + for (auto other : bel_candidate_cells[candidate]) + if (other != cell->name && other != prev_cell) + overlap.push_back(other); + if (overlap.size() > 0) + NPNR_ASSERT(allow_swap); + for (auto ov : overlap) { + bel_candidate_cells[candidate].erase(ov); + cell_neighbour_bels[ov].erase(candidate); + } + } + } + } + return found_count; + } + + std::vector<std::vector<PortRef *>> find_crit_paths(float crit_thresh, size_t max_count) + { + std::vector<std::vector<PortRef *>> crit_paths; + std::vector<std::pair<NetInfo *, store_index<PortRef>>> crit_nets; + std::vector<IdString> netnames; + std::transform(ctx->nets.begin(), ctx->nets.end(), std::back_inserter(netnames), + [](const std::pair<IdString, std::unique_ptr<NetInfo>> &kv) { return kv.first; }); + ctx->sorted_shuffle(netnames); + for (auto net : netnames) { + if (crit_nets.size() >= max_count) + break; + float highest_crit = 0; + store_index<PortRef> crit_user_idx{}; + NetInfo *ni = ctx->nets.at(net).get(); + for (auto usr : ni->users.enumerate()) { + float crit = tmg.get_criticality(CellPortKey(usr.value)); + if (crit > highest_crit) { + highest_crit = crit; + crit_user_idx = usr.index; + } + } + if (highest_crit > crit_thresh) + crit_nets.emplace_back(ni, crit_user_idx); + } + + pool<PortRef *, hash_ptr_ops> used_ports; + + for (auto crit_net : crit_nets) { + + if (used_ports.count(&(crit_net.first->users.at(crit_net.second)))) + continue; + + std::deque<PortRef *> crit_path; + + // FIXME: This will fail badly on combinational loops + + // Iterate backwards following greatest criticality + NetInfo *back_cursor = crit_net.first; + while (back_cursor != nullptr) { + float max_crit = 0; + std::pair<NetInfo *, store_index<PortRef>> crit_sink{nullptr, {}}; + CellInfo *cell = back_cursor->driver.cell; + if (cell == nullptr) + break; + for (auto port : cell->ports) { + if (port.second.type != PORT_IN) + continue; + NetInfo *pn = port.second.net; + if (pn == nullptr) + continue; + int ccount; + DelayQuad combDelay; + TimingPortClass tpclass = ctx->getPortTimingClass(cell, port.first, ccount); + if (tpclass != TMG_COMB_INPUT) + continue; + bool is_path = ctx->getCellDelay(cell, port.first, back_cursor->driver.port, combDelay); + if (!is_path) + continue; + float usr_crit = tmg.get_criticality(CellPortKey(cell->name, port.first)); + if (used_ports.count(&(pn->users.at(port.second.user_idx)))) + continue; + if (usr_crit >= max_crit) { + max_crit = usr_crit; + crit_sink = std::make_pair(pn, port.second.user_idx); + } + } + + if (crit_sink.first != nullptr) { + crit_path.push_front(&(crit_sink.first->users.at(crit_sink.second))); + used_ports.insert(&(crit_sink.first->users.at(crit_sink.second))); + } + back_cursor = crit_sink.first; + } + // Iterate forwards following greatest criticiality + PortRef *fwd_cursor = &(crit_net.first->users.at(crit_net.second)); + while (fwd_cursor != nullptr) { + crit_path.push_back(fwd_cursor); + float max_crit = 0; + std::pair<NetInfo *, store_index<PortRef>> crit_sink{nullptr, {}}; + CellInfo *cell = fwd_cursor->cell; + for (auto port : cell->ports) { + if (port.second.type != PORT_OUT) + continue; + NetInfo *pn = port.second.net; + if (pn == nullptr) + continue; + + int ccount; + DelayQuad combDelay; + TimingPortClass tpclass = ctx->getPortTimingClass(cell, port.first, ccount); + if (tpclass != TMG_COMB_OUTPUT && tpclass != TMG_REGISTER_OUTPUT) + continue; + bool is_path = ctx->getCellDelay(cell, fwd_cursor->port, port.first, combDelay); + if (!is_path) + continue; + for (auto usr : pn->users.enumerate()) { + if (used_ports.count(&(pn->users.at(usr.index)))) + continue; + float crit = tmg.get_criticality(CellPortKey(usr.value)); + if (crit >= max_crit) { + max_crit = crit; + crit_sink = std::make_pair(pn, usr.index); + } + } + } + if (crit_sink.first != nullptr) { + fwd_cursor = &(crit_sink.first->users.at(crit_sink.second)); + used_ports.insert(&(crit_sink.first->users.at(crit_sink.second))); + } else { + fwd_cursor = nullptr; + } + } + + std::vector<PortRef *> crit_path_vec; + std::copy(crit_path.begin(), crit_path.end(), std::back_inserter(crit_path_vec)); + crit_paths.push_back(crit_path_vec); + } + + return crit_paths; + } + + void optimise_path(std::vector<PortRef *> &path) + { + path_cells.clear(); + cell_neighbour_bels.clear(); + bel_candidate_cells.clear(); + if (ctx->debug) + log_info("Optimising the following path: \n"); + + auto front_port = path.front(); + NetInfo *front_net = front_port->cell->ports.at(front_port->port).net; + if (front_net != nullptr && front_net->driver.cell != nullptr) { + auto front_cell = front_net->driver.cell; + if (front_cell->belStrength <= STRENGTH_WEAK && cfg.cellTypes.count(front_cell->type) && + front_cell->cluster == ClusterId()) { + path_cells.push_back(front_cell->name); + } + } + + for (auto port : path) { + if (ctx->debug) { + float crit = tmg.get_criticality(CellPortKey(*port)); + log_info(" %s.%s at %s crit %0.02f\n", port->cell->name.c_str(ctx), port->port.c_str(ctx), + ctx->nameOfBel(port->cell->bel), crit); + } + if (std::find(path_cells.begin(), path_cells.end(), port->cell->name) != path_cells.end()) + continue; + if (port->cell->belStrength > STRENGTH_WEAK || !cfg.cellTypes.count(port->cell->type) || + port->cell->cluster != ClusterId()) + continue; + if (ctx->debug) + log_info(" can move\n"); + path_cells.push_back(port->cell->name); + } + + if (path_cells.size() < 2) { + if (ctx->debug) { + log_info("Too few moveable cells; skipping path\n"); + log_break(); + } + + return; + } + + // Calculate original delay before touching anything + delay_t original_delay = 0; + + for (size_t i = 0; i < path.size(); i++) { + auto &port = path.at(i)->cell->ports.at(path.at(i)->port); + NetInfo *pn = port.net; + if (port.user_idx) + original_delay += ctx->predictArcDelay(pn, pn->users.at(port.user_idx)); + } + + IdString last_cell; + const int d = 2; // FIXME: how to best determine d + for (auto cell : path_cells) { + // FIXME: when should we allow swapping due to a lack of candidates + find_neighbours(ctx->cells[cell].get(), last_cell, d, false); + last_cell = cell; + } + + if (ctx->debug) { + for (auto cell : path_cells) { + log_info("Candidate neighbours for %s (%s):\n", cell.c_str(ctx), ctx->nameOfBel(ctx->cells[cell]->bel)); + for (auto neigh : cell_neighbour_bels.at(cell)) { + log_info(" %s\n", ctx->nameOfBel(neigh)); + } + } + } + + // Actual BFS path optimisation algorithm + dict<IdString, dict<BelId, delay_t>> cumul_costs; + dict<std::pair<IdString, BelId>, std::pair<IdString, BelId>> backtrace; + std::queue<std::pair<int, BelId>> visit; + pool<std::pair<int, BelId>> to_visit; + + for (auto startbel : cell_neighbour_bels[path_cells.front()]) { + // Swap for legality check + CellInfo *cell = ctx->cells.at(path_cells.front()).get(); + BelId origBel = cell_swap_bel(cell, startbel); + std::vector<std::pair<CellInfo *, BelId>> move{std::make_pair(cell, origBel)}; + if (acceptable_move(move)) { + auto entry = std::make_pair(0, startbel); + visit.push(entry); + cumul_costs[path_cells.front()][startbel] = 0; + } + // Swap back + cell_swap_bel(cell, origBel); + } + + while (!visit.empty()) { + auto entry = visit.front(); + visit.pop(); + auto cellname = path_cells.at(entry.first); + if (entry.first == int(path_cells.size()) - 1) + continue; + std::vector<std::pair<CellInfo *, BelId>> move; + // Apply the entire backtrace for accurate legality and delay checks + // This is probably pretty expensive (but also probably pales in comparison to the number of swaps + // SA will make...) + std::vector<std::pair<IdString, BelId>> route_to_entry; + auto cursor = std::make_pair(cellname, entry.second); + route_to_entry.push_back(cursor); + while (backtrace.count(cursor)) { + cursor = backtrace.at(cursor); + route_to_entry.push_back(cursor); + } + for (auto rt_entry : boost::adaptors::reverse(route_to_entry)) { + CellInfo *cell = ctx->cells.at(rt_entry.first).get(); + BelId origBel = cell_swap_bel(cell, rt_entry.second); + move.push_back(std::make_pair(cell, origBel)); + } + + // Have a look at where we can travel from here + for (auto neighbour : cell_neighbour_bels.at(path_cells.at(entry.first + 1))) { + // Edges between overlapping bels are deleted + if (neighbour == entry.second) + continue; + // Experimentally swap the next path cell onto the neighbour bel we are trying + IdString ncname = path_cells.at(entry.first + 1); + CellInfo *next_cell = ctx->cells.at(ncname).get(); + BelId origBel = cell_swap_bel(next_cell, neighbour); + move.push_back(std::make_pair(next_cell, origBel)); + + delay_t total_delay = 0; + + for (size_t i = 0; i < path.size(); i++) { + auto &port = path.at(i)->cell->ports.at(path.at(i)->port); + NetInfo *pn = port.net; + if (port.user_idx) + total_delay += ctx->predictArcDelay(pn, pn->users.at(port.user_idx)); + if (path.at(i)->cell == next_cell) + break; + } + + // First, check if the move is actually worthwhile from a delay point of view before the expensive + // legality check + if (!cumul_costs.count(ncname) || !cumul_costs.at(ncname).count(neighbour) || + cumul_costs.at(ncname).at(neighbour) > total_delay) { + // Now check that the swaps we have made to get here are legal and meet max delay requirements + if (acceptable_move(move)) { + cumul_costs[ncname][neighbour] = total_delay; + backtrace[std::make_pair(ncname, neighbour)] = std::make_pair(cellname, entry.second); + if (!to_visit.count(std::make_pair(entry.first + 1, neighbour))) + visit.push(std::make_pair(entry.first + 1, neighbour)); + } + } + // Revert the experimental swap + cell_swap_bel(move.back().first, move.back().second); + move.pop_back(); + } + + // Revert move by swapping cells back to their original order + // Execute swaps in reverse order to how we made them originally + for (auto move_entry : boost::adaptors::reverse(move)) { + cell_swap_bel(move_entry.first, move_entry.second); + } + } + + // Did we find a solution?? + if (cumul_costs.count(path_cells.back())) { + // Find the end position with the lowest total delay + auto &end_options = cumul_costs.at(path_cells.back()); + auto lowest = std::min_element(end_options.begin(), end_options.end(), + [](const std::pair<BelId, delay_t> &a, const std::pair<BelId, delay_t> &b) { + return a.second < b.second; + }); + NPNR_ASSERT(lowest != end_options.end()); + + std::vector<std::pair<IdString, BelId>> route_to_solution; + auto cursor = std::make_pair(path_cells.back(), lowest->first); + route_to_solution.push_back(cursor); + while (backtrace.count(cursor)) { + cursor = backtrace.at(cursor); + route_to_solution.push_back(cursor); + } + if (ctx->debug) + log_info("Found a solution with cost %.02f ns (existing path %.02f ns)\n", + ctx->getDelayNS(lowest->second), ctx->getDelayNS(original_delay)); + for (auto rt_entry : boost::adaptors::reverse(route_to_solution)) { + CellInfo *cell = ctx->cells.at(rt_entry.first).get(); + cell_swap_bel(cell, rt_entry.second); + if (ctx->debug) + log_info(" %s at %s\n", rt_entry.first.c_str(ctx), ctx->nameOfBel(rt_entry.second)); + } + + } else { + if (ctx->debug) + log_info("Solution was not found\n"); + } + if (ctx->debug) + log_break(); + } + + // Current candidate Bels for cells (linked in both direction> + std::vector<IdString> path_cells; + dict<IdString, pool<BelId>> cell_neighbour_bels; + dict<BelId, pool<IdString>> bel_candidate_cells; + // Map cell ports to net delay limit + dict<std::pair<IdString, IdString>, delay_t> max_net_delay; + Context *ctx; + TimingOptCfg cfg; + TimingAnalyser tmg; +}; + +bool timing_opt(Context *ctx, TimingOptCfg cfg) { return TimingOptimiser(ctx, cfg).optimise(); } + +NEXTPNR_NAMESPACE_END |