diff --git a/common/place/placer_static.cc b/common/place/placer_static.cc index d47429c2..5a64065a 100644 --- a/common/place/placer_static.cc +++ b/common/place/placer_static.cc @@ -56,6 +56,7 @@ namespace { struct PlacerGroup { + bool enabled = true; int total_bels = 0; double concrete_area = 0; double dark_area = 0; @@ -70,6 +71,9 @@ struct PlacerGroup FFTArray density_fft; FFTArray electro_phi; FFTArray electro_fx, electro_fy; + + double init_potential = 0; + double curr_potential = 0; }; // Could be an actual concrete netlist cell; or just a spacer @@ -301,7 +305,7 @@ class StaticPlacer dict beltype2group; for (int i = 0; i < int(groups.size()); i++) { groups.at(i).loc_area.reset(width, height); - for (const auto &bel_type : cfg.cell_groups.at(i).cell_area) + for (const auto &bel_type : cfg.cell_groups.at(i).bel_area) beltype2group[bel_type.first] = i; } for (auto bel : ctx->getBels()) { @@ -409,14 +413,16 @@ class StaticPlacer StaticRect rect; // Mismatched group case if (!lookup_group(ci->type, cell_group, rect)) { - for (auto bel : ctx->getBels()) { - if (ctx->isValidBelForCellType(ci->type, bel) && ctx->checkBelAvail(bel)) { - ctx->bindBel(bel, ci, STRENGTH_STRONG); - if (!ctx->isBelLocationValid(bel)) { - ctx->unbindBel(bel); - } else { - log_info(" placed potpourri cell '%s' at bel '%s'\n", ctx->nameOf(ci), ctx->nameOfBel(bel)); - break; + if (ci->bel == BelId()) { + for (auto bel : ctx->getBels()) { + if (ctx->isValidBelForCellType(ci->type, bel) && ctx->checkBelAvail(bel)) { + ctx->bindBel(bel, ci, STRENGTH_STRONG); + if (!ctx->isBelLocationValid(bel)) { + ctx->unbindBel(bel); + } else { + log_info(" placed potpourri cell '%s' at bel '%s'\n", ctx->nameOf(ci), ctx->nameOfBel(bel)); + break; + } } } } @@ -506,26 +512,31 @@ class StaticPlacer } } - const double target_util = 0.8; + const double target_util = 1.0; void insert_dark() { log_info("⌁ inserting dark nodes...\n"); for (int group = 0; group < int(groups.size()); group++) { + const auto &cg = cfg.cell_groups.at(group); auto &g = groups.at(group); + int dark_count = 0; for (auto tile : g.loc_area) { if (tile.value > 0.5f) continue; StaticRect dark_area(1.0f, 1.0f - tile.value); - int cell_idx = add_cell(dark_area, group, RealPair(tile.x + 0.5f, tile.y + 0.5f), nullptr /*spacer*/); + int cell_idx = add_cell(dark_area, group, RealPair(tile.x, tile.y), nullptr /*spacer*/); mcells.at(cell_idx).is_dark = true; + ++dark_count; } + log_info("⌁ group %s inserted a total of %d dark nodes\n", ctx->nameOf(cg.name), dark_count); } } void insert_spacer() { log_info("⌁ inserting spacers...\n"); + int inserted_spacers = 0; for (int group = 0; group < int(groups.size()); group++) { const auto &cg = cfg.cell_groups.at(group); @@ -537,8 +548,13 @@ class StaticPlacer int spacer_count = (g.total_area * target_util - g.concrete_area) / cg.spacer_rect.area(); if (spacer_count <= 0) continue; - for (int i = 0; i < spacer_count; i++) { - add_cell(cg.spacer_rect, group, RealPair(ctx->rngf(width), ctx->rngf(height)), nullptr /*spacer*/); + while (inserted_spacers < spacer_count) { + int x = ctx->rng(width); + int y = ctx->rng(height); + // avoid placing spacers at locations with dark nodes + if (ctx->rngf(1.0) > g.loc_area.at(x, y)) + continue; + add_cell(cg.spacer_rect, group, RealPair(x + ctx->rngf(1.0), y + ctx->rngf(1.0)), nullptr /*spacer*/); ++inserted_spacers; } } @@ -556,7 +572,7 @@ class StaticPlacer { // TODO: a m x m grid follows the paper and makes the DCTs easier, but is it actually ideal for non-square // FPGAs? - m = 1 << int(std::ceil(std::log2(std::sqrt(mcells.size() / groups.size())))); + m = 1 << int(std::ceil(std::log2(std::max(width, height)))); bin_w = double(width) / m; bin_h = double(height) / m; @@ -626,11 +642,10 @@ class StaticPlacer } } - void compute_overlap() + void compute_conc_density() { - // populate for concrete cells only for (auto &g : groups) - g.conc_density.reset(m, m, 0); + g.conc_density.reset(width, height, 0); for (int idx = 0; idx < int(ccells.size()); idx++) { auto &mc = mcells.at(idx); auto &g = groups.at(mc.group); @@ -641,10 +656,18 @@ class StaticPlacer for (int dx = 0; dx <= int(size.w); dx++) { float h = (dy == int(size.h)) ? (size.h - int(size.h)) : 1; float w = (dx == int(size.w)) ? (size.w - int(size.w)) : 1; - g.conc_density.at(loc.x + dx, loc.y + dy) += w * h; + if ((loc.x + dx) >= 0 && (loc.x + dx) < width && (loc.y + dy) >= 0 && (loc.y + dy) < height) + g.conc_density.at(loc.x + dx, loc.y + dy) += w * h; } } } + } + + void compute_overlap() + { + // populate for concrete cells only + compute_conc_density(); + std::string overlap_str = ""; for (int idx = 0; idx < int(groups.size()); idx++) { auto &g = groups.at(idx); @@ -868,28 +891,17 @@ class StaticPlacer } if (init_penalty) { // set initial density penalty - dict wirelen_sum; - dict force_sum; - for (auto &cell : ctx->cells) { - CellInfo *ci = cell.second.get(); - if (ci->udata == -1) - continue; - auto &mc = mcells.at(ci->udata); - auto res1 = wirelen_sum.insert({mc.group, std::abs(mc.ref_wl_grad.x) + std::abs(mc.ref_wl_grad.y)}); - if (!res1.second) - res1.first->second += std::abs(mc.ref_wl_grad.x) + std::abs(mc.ref_wl_grad.y); - auto res2 = force_sum.insert({mc.group, std::abs(mc.ref_dens_grad.x) + std::abs(mc.ref_dens_grad.y)}); - if (!res2.second) - res2.first->second += std::abs(mc.ref_dens_grad.x) + std::abs(mc.ref_dens_grad.y); - } - dens_penalty = std::vector(wirelen_sum.size(), 0.0); - for (auto &item : wirelen_sum) { - auto group = item.first; - auto wirelen = item.second; - dens_penalty[group] = wirelen / force_sum.at(group); - log_info(" initial density penalty for %s: %f\n", cfg.cell_groups.at(group).name.c_str(ctx), - dens_penalty[group]); + double wirelen_sum = 0, force_sum = 0; + for (int i = 0; i < int(ccells.size()); i++) { + auto mc = mcells.at(i); + wirelen_sum += std::abs(mc.ref_wl_grad.x) + std::abs(mc.ref_wl_grad.y); + force_sum += std::abs(mc.ref_dens_grad.x) + std::abs(mc.ref_dens_grad.y); } + const float eta = 1e-1; + float init_dens_penalty = eta * (wirelen_sum / force_sum); + log_info("initial density penalty: %f\n", init_dens_penalty); + dens_penalty.resize(groups.size(), init_dens_penalty); + update_potentials(true); // set initial potential } // Third loop: compute total gradient, and precondition // TODO: ALM as well as simple penalty @@ -953,17 +965,59 @@ class StaticPlacer return hpwl; } - float system_potential() + void update_potentials(bool init = false) { - float pot = 0; + for (auto &group : groups) + group.curr_potential = 0; for (auto &cell : mcells) { auto &g = groups.at(cell.group); iter_slithers(cell.ref_pos, cell.rect, - [&](int x, int y, float area) { pot += g.electro_phi.at(x, y) * area; }); + [&](int x, int y, float area) { g.curr_potential += g.electro_phi.at(x, y) * area; }); } + if (init) { + for (auto &group : groups) + group.init_potential = group.curr_potential; + } + } + + float system_potential() + { + float pot = 0; + for (auto &group : groups) + pot += group.curr_potential; return pot; } + float penalty_beta = 2.0e3f; + float alpha_l = 1.05f, alpha_h = 1.06f; + double penalty_incr = alpha_h - 1; + void update_penalties() { + float pot_norm = 0; + // compute L2-norm of relative system potential + std::vector rel_pot; + for (int g = 0; g < int(groups.size()); g++) { + auto &group = groups.at(g); + if (!group.enabled) + continue; + float phi_hat = group.curr_potential / group.init_potential; + rel_pot.push_back(phi_hat); + pot_norm += phi_hat * phi_hat; + } + pot_norm = sqrt(pot_norm); + log_info("pot_norm: %f\n", pot_norm); + // update penalty multiplier (ELFPlace equation 22) + double log_term = std::log(penalty_beta * pot_norm + 1); + penalty_incr = penalty_incr * ((log_term / (log_term + 1)) * (alpha_h - alpha_l) + alpha_l); + // update density penalties (ELFPlace equation 21) + for (int g = 0; g < int(groups.size()); g++) { + if (!groups.at(g).enabled) + continue; + float next_penalty = dens_penalty.at(g) + (penalty_incr * (rel_pot.at(g) / pot_norm)); + dens_penalty.at(g) = next_penalty; + } + } + + void initialise() { float initial_steplength = 0.01f; @@ -1016,6 +1070,7 @@ class StaticPlacer // Move the post-solve position of a chain towards be the weighted average of its constituents // The strength increases with iterations float alpha = std::min(std::pow(1.002f, iter) - 1, 1.0f); + float dist = 0; for (int i = 0; i < int(macros.size()); i++) { auto ¯o = macros.at(i); float total_area = 0; @@ -1033,10 +1088,13 @@ class StaticPlacer for (int c : macro.conc_cells) { auto &cc = ccells.at(c); auto &mc = mcells.at(c); + auto last_pos = mc.pos; mc.pos = mc.pos * (1 - alpha) + (pos + RealPair(cc.chunk_dx, cc.chunk_dy)) * alpha; mc.ref_pos = mc.ref_pos * (1 - alpha) + (ref_pos + RealPair(cc.chunk_dx, cc.chunk_dy)) * alpha; + dist += std::sqrt(std::pow(last_pos.x - mc.pos.x, 2) + std::pow(last_pos.y - mc.pos.y, 2)); } } + log_info(" update_chains distance %.2f\n", dist); } void step() @@ -1064,9 +1122,10 @@ class StaticPlacer nesterov_a = a_next; update_chains(); update_gradients(true); + update_potentials(); log_info(" system potential: %f hpwl: %f\n", system_potential(), system_hpwl()); compute_overlap(); - if ((iter % 5) == 0) + if ((iter % 10) == 0) update_timing(); } @@ -1410,8 +1469,9 @@ class StaticPlacer bool legalised_ip = false; while (true) { step(); - for (auto &penalty : dens_penalty) - penalty *= 1.025; + for (auto &p : dens_penalty) + p *= 1.025; + // update_penalties(); if (!legalised_ip) { float ip_overlap = 0; for (int i = cfg.logic_groups; i < int(groups.size()); i++) @@ -1419,6 +1479,8 @@ class StaticPlacer if (ip_overlap < 0.15) { legalise_step(true); legalised_ip = true; + for (int i = cfg.logic_groups; i < int(groups.size()); i++) + groups.at(i).enabled = false; } } else { float logic_overlap = 0;