caiyu/nextpnr
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
f3d9b45387
nextpnr/common/placer_heap.cc
David Shah f3d9b45387 HeAP: Add SA-based iterative refinement after AP 
Signed-off-by: David Shah <dave@ds0.me>
2019-03-22 10:31:54 +00:00

1230 lines
52 KiB
C++
Raw Blame History

/*
* nextpnr -- Next Generation Place and Route
*
* Copyright (C) 2019 David Shah <david@symbioticeda.com>
*
* 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.
*
* [[cite]] HeAP
* Analytical Placement for Heterogeneous FPGAs, Marcel Gort and Jason H. Anderson
* https://janders.eecg.utoronto.ca/pdfs/marcelfpl12.pdf
*
* [[cite]] SimPL
* SimPL: An Effective Placement Algorithm, Myung-Chul Kim, Dong-Jin Lee and Igor L. Markov
* http://www.ece.umich.edu/cse/awards/pdfs/iccad10-simpl.pdf
*/
#include <deque>
#include <numeric>
#include <queue>
#include <unordered_map>
#include <boost/optional.hpp>
#include <fstream>
#include "log.h"
#include "nextpnr.h"
#include "place_common.h"
#include "placer_math.h"
#include "placer1.h"
#include "util.h"
NEXTPNR_NAMESPACE_BEGIN
namespace {
// A simple internal representation for a sparse system of equations Ax = rhs
// This is designed to decouple the functions that build the matrix to the engine that
// solves it, and the representation that requires
template <typename T> struct EquationSystem
{
EquationSystem(size_t rows, size_t cols)
{
A.resize(cols);
rhs.resize(rows);
}
// Simple sparse format, easy to convert to CCS for solver
std::vector<std::vector<std::pair<int, T>>> A; // col -> (row, x[row, col]) sorted by row
std::vector<T> rhs; // RHS vector
void reset()
{
for (auto &col : A)
col.clear();
std::fill(rhs.begin(), rhs.end(), T());
}
void add_coeff(int row, int col, T val)
{
auto &Ac = A.at(col);
// Binary search
int b = 0, e = int(Ac.size()) - 1;
while (b <= e) {
int i = (b + e) / 2;
if (Ac.at(i).first == row) {
Ac.at(i).second += val;
return;
}
if (Ac.at(i).first > row)
e = i - 1;
else
b = i + 1;
}
Ac.insert(Ac.begin() + b, std::make_pair(row, val));
}
void add_rhs(int row, T val) { rhs[row] += val; }
void solve(std::vector<double> &x)
{
NPNR_ASSERT(x.size() == A.size());
int nnz = std::accumulate(A.begin(), A.end(), 0,
[](int a, const std::vector<std::pair<int, T>> &vec) { return a + int(vec.size()); });
taucif_system *sys = taucif_create_system(int(rhs.size()), int(A.size()), nnz);
for (int col = 0; col < int(A.size()); col++) {
auto &Ac = A[col];
for (auto &el : Ac) {
if (col <= el.first) {
// log_info("%d %d %f\n", el.first, col, el.second);
taucif_add_matrix_value(sys, el.first, col, el.second);
}
// FIXME: in debug mode, assert really is symmetric
}
}
taucif_finalise_matrix(sys);
int result = taucif_solve_system(sys, x.data(), rhs.data());
NPNR_ASSERT(result == 0);
taucif_free_system(sys);
// for (int i = 0; i < int(x.size()); i++)
// log_info("x[%d] = %f\n", i, x.at(i));
}
};
} // namespace
class HeAPPlacer
{
public:
HeAPPlacer(Context *ctx) : ctx(ctx) {}
bool place()
{
ctx->lock();
taucif_init_solver();
place_constraints();
build_fast_bels();
seed_placement();
update_all_chains();
wirelen_t hpwl = total_hpwl();
log_info("Initial placer starting hpwl = %d\n", int(hpwl));
for (int i = 0; i < 20; i++) {
setup_solve_cells();
EquationSystem<double> esx(solve_cells.size(), solve_cells.size());
build_equations(esx, false);
// log_info("x-axis\n");
solve_equations(esx, false);
EquationSystem<double> esy(solve_cells.size(), solve_cells.size());
build_equations(esy, true);
// log_info("y-axis\n");
solve_equations(esy, true);
update_all_chains();
hpwl = total_hpwl();
log_info("Initial placer iter %d, hpwl = %d\n", i, int(hpwl));
}
// legalise_with_cuts(true);
// CutLegaliser(this, ctx->id("ICESTORM_LC")).run();
//NPNR_ASSERT(false);
bool valid = true;
wirelen_t solved_hpwl = 0, legal_hpwl = 0, best_hpwl = std::numeric_limits<wirelen_t>::max();
int iter = 0, stalled = 0;
while (!valid || (stalled < 5 && (solved_hpwl <= legal_hpwl * 0.8))) {
if (!valid && ((solved_hpwl > legal_hpwl * 0.8) || (stalled > 5))) {
stalled = 0;
best_hpwl = std::numeric_limits<wirelen_t>::max();
valid = true;
}
setup_solve_cells();
EquationSystem<double> esx(solve_cells.size(), solve_cells.size());
build_equations(esx, false, iter);
// log_info("x-axis\n");
solve_equations(esx, false);
EquationSystem<double> esy(solve_cells.size(), solve_cells.size());
build_equations(esy, true, iter);
// log_info("y-axis\n");
solve_equations(esy, true);
solved_hpwl = total_hpwl();
log_info("Solved HPWL = %d\n", int(solved_hpwl));
update_all_chains();
CutLegaliser(this, ctx->id("ICESTORM_LC")).run();
update_all_chains();
legal_hpwl = total_hpwl();
log_info("Spread HPWL = %d\n", int(legal_hpwl));
legalise_placement_simple(valid);
update_all_chains();
legal_hpwl = total_hpwl();
log_info("Legalised HPWL = %d (%s)\n", int(legal_hpwl), valid ? "valid" : "invalid");
if (legal_hpwl < best_hpwl) {
best_hpwl = legal_hpwl;
stalled = 0;
} else {
++stalled;
}
ctx->yield();
++iter;
}
ctx->unlock();
placer1_refine(ctx, Placer1Cfg(ctx));
return true;
}
private:
Context *ctx;
int max_x = 0, max_y = 0;
std::vector<std::vector<std::vector<std::vector<BelId>>>> fast_bels;
std::unordered_map<IdString, std::tuple<int, int>> bel_types;
// For fast handling of heterogeneosity during initial placement without full legalisation,
// for each Bel type this goes from x or y to the nearest x or y where a Bel of a given type exists
// This is particularly important for the iCE40 architecture, where multipliers and BRAM only exist at the
// edges and corners respectively
std::vector<std::vector<int>> nearest_row_with_bel;
std::vector<std::vector<int>> nearest_col_with_bel;
// In some cases, we can't use bindBel because we allow overlap in the earlier stages. So we use this custom
// structure instead
struct CellLocation
{
int x, y;
double rawx, rawy;
bool locked, global;
};
std::unordered_map<IdString, CellLocation> cell_locs;
// The set of cells that we will actually place. This excludes locked cells and children cells of macros/chains
// (only the root of each macro is placed.)
std::vector<CellInfo *> place_cells;
// The cells in the current equation being solved (a subset of place_cells in some cases, where we only place
// cells of a certain type)
std::vector<CellInfo *> solve_cells;
// For cells in a chain, this is the ultimate root cell of the chain (sometimes this is not constr_parent
// where chains are within chains
std::unordered_map<IdString, CellInfo *> chain_root;
std::unordered_map<IdString, int> chain_size;
// The offset from chain_root to a cell in the chain
std::unordered_map<IdString, std::pair<int, int>> cell_offsets;
// Place cells with the BEL attribute set to constrain them
void place_constraints()
{
size_t placed_cells = 0;
// Initial constraints placer
for (auto &cell_entry : ctx->cells) {
CellInfo *cell = cell_entry.second.get();
auto loc = cell->attrs.find(ctx->id("BEL"));
if (loc != cell->attrs.end()) {
std::string loc_name = loc->second;
BelId bel = ctx->getBelByName(ctx->id(loc_name));
if (bel == BelId()) {
log_error("No Bel named \'%s\' located for "
"this chip (processing BEL attribute on \'%s\')\n",
loc_name.c_str(), cell->name.c_str(ctx));
}
IdString bel_type = ctx->getBelType(bel);
if (bel_type != cell->type) {
log_error("Bel \'%s\' of type \'%s\' does not match cell "
"\'%s\' of type \'%s\'\n",
loc_name.c_str(), bel_type.c_str(ctx), cell->name.c_str(ctx), cell->type.c_str(ctx));
}
if (!ctx->isValidBelForCell(cell, bel)) {
log_error("Bel \'%s\' of type \'%s\' is not valid for cell "
"\'%s\' of type \'%s\'\n",
loc_name.c_str(), bel_type.c_str(ctx), cell->name.c_str(ctx), cell->type.c_str(ctx));
}
auto bound_cell = ctx->getBoundBelCell(bel);
if (bound_cell) {
log_error("Cell \'%s\' cannot be bound to bel \'%s\' since it is already bound to cell \'%s\'\n",
cell->name.c_str(ctx), loc_name.c_str(), bound_cell->name.c_str(ctx));
}
ctx->bindBel(bel, cell, STRENGTH_USER);
placed_cells++;
}
}
int constr_placed_cells = placed_cells;
log_info("Placed %d cells based on constraints.\n", int(placed_cells));
ctx->yield();
}
// Construct the fast_bels, nearest_row_with_bel and nearest_col_with_bel
void build_fast_bels()
{
int num_bel_types = 0;
for (auto bel : ctx->getBels()) {
IdString type = ctx->getBelType(bel);
if (bel_types.find(type) == bel_types.end()) {
bel_types[type] = std::tuple<int, int>(num_bel_types++, 1);
} else {
std::get<1>(bel_types.at(type))++;
}
}
for (auto bel : ctx->getBels()) {
if (!ctx->checkBelAvail(bel))
continue;
Loc loc = ctx->getBelLocation(bel);
IdString type = ctx->getBelType(bel);
int type_idx = std::get<0>(bel_types.at(type));
if (int(fast_bels.size()) < type_idx + 1)
fast_bels.resize(type_idx + 1);
if (int(fast_bels.at(type_idx).size()) < (loc.x + 1))
fast_bels.at(type_idx).resize(loc.x + 1);
if (int(fast_bels.at(type_idx).at(loc.x).size()) < (loc.y + 1))
fast_bels.at(type_idx).at(loc.x).resize(loc.y + 1);
max_x = std::max(max_x, loc.x);
max_y = std::max(max_y, loc.y);
fast_bels.at(type_idx).at(loc.x).at(loc.y).push_back(bel);
}
nearest_row_with_bel.resize(num_bel_types, std::vector<int>(max_y + 1, -1));
nearest_col_with_bel.resize(num_bel_types, std::vector<int>(max_x + 1, -1));
for (auto bel : ctx->getBels()) {
if (!ctx->checkBelAvail(bel))
continue;
Loc loc = ctx->getBelLocation(bel);
int type_idx = std::get<0>(bel_types.at(ctx->getBelType(bel)));
auto &nr = nearest_row_with_bel.at(type_idx), &nc = nearest_col_with_bel.at(type_idx);
// Traverse outwards through nearest_row_with_bel and nearest_col_with_bel, stopping once
// another row/col is already recorded as being nearer
for (int x = loc.x; x <= max_x; x++) {
if (nc.at(x) != -1 && std::abs(loc.x - nc.at(x)) <= (x - loc.x))
break;
nc.at(x) = loc.x;
}
for (int x = loc.x - 1; x >= 0; x--) {
if (nc.at(x) != -1 && std::abs(loc.x - nc.at(x)) <= (loc.x - x))
break;
nc.at(x) = loc.x;
}
for (int y = loc.y; y <= max_y; y++) {
if (nr.at(y) != -1 && std::abs(loc.y - nr.at(y)) <= (y - loc.y))
break;
nr.at(y) = loc.y;
}
for (int y = loc.y - 1; y >= 0; y--) {
if (nr.at(y) != -1 && std::abs(loc.y - nr.at(y)) <= (loc.y - y))
break;
nr.at(y) = loc.y;
}
}
}
// Check if a cell has any meaningful connectivity
bool has_connectivity(CellInfo *cell)
{
for (auto port : cell->ports) {
if (port.second.net != nullptr && port.second.net->driver.cell != nullptr &&
!port.second.net->users.empty())
return true;
}
return false;
}
// Build up a random initial placement, without regard to legality
// FIXME: Are there better approaches to the initial placement (e.g. greedy?)
void seed_placement()
{
std::unordered_map<IdString, std::deque<BelId>> available_bels;
for (auto bel : ctx->getBels()) {
if (!ctx->checkBelAvail(bel))
continue;
available_bels[ctx->getBelType(bel)].push_back(bel);
}
for (auto &t : available_bels) {
std::random_shuffle(t.second.begin(), t.second.end(), [&](size_t n){
return ctx->rng(int(n));
});
}
for (auto cell : sorted(ctx->cells)) {
CellInfo *ci = cell.second;
if (ci->bel != BelId()) {
Loc loc = ctx->getBelLocation(ci->bel);
cell_locs[cell.first].x = loc.x;
cell_locs[cell.first].y = loc.y;
cell_locs[cell.first].locked = true;
cell_locs[cell.first].global = ctx->getBelGlobalBuf(ci->bel);
} else if (ci->constr_parent == nullptr) {
bool placed = false;
while (!placed) {
if (!available_bels.count(ci->type) || available_bels.at(ci->type).empty())
log_error("Unable to place cell '%s', no Bels remaining of type '%s'\n", ci->name.c_str(ctx),
ci->type.c_str(ctx));
BelId bel = available_bels.at(ci->type).back();
available_bels.at(ci->type).pop_back();
Loc loc = ctx->getBelLocation(bel);
cell_locs[cell.first].x = loc.x;
cell_locs[cell.first].y = loc.y;
cell_locs[cell.first].locked = false;
cell_locs[cell.first].global = ctx->getBelGlobalBuf(bel);
// FIXME
if (has_connectivity(cell.second) && cell.second->type != ctx->id("SB_IO")) {
place_cells.push_back(ci);
placed = true;
} else {
if (ctx->isValidBelForCell(ci, bel)) {
ctx->bindBel(bel, ci, STRENGTH_STRONG);
cell_locs[cell.first].locked = true;
placed = true;
} else {
available_bels.at(ci->type).push_front(bel);
}
}
}
}
}
}
// Setup the cells to be solved, returns the number of rows
int setup_solve_cells(std::unordered_set<IdString> *celltypes = nullptr)
{
int row = 0;
solve_cells.clear();
// First clear the udata of all cells
for (auto cell : sorted(ctx->cells))
cell.second->udata = dont_solve;
// Then update cells to be placed, which excludes cell children
for (auto cell : place_cells) {
if (celltypes && !celltypes->count(cell->type))
continue;
cell->udata = row++;
solve_cells.push_back(cell);
}
// Finally, update the udata of children
for (auto chained : chain_root)
ctx->cells.at(chained.first)->udata = chained.second->udata;
return row;
}
// Update the location of all children of a chain
void update_chain(CellInfo *cell, CellInfo *root)
{
const auto &base = cell_locs[cell->name];
for (auto child : cell->constr_children) {
chain_size[root->name]++;
if (child->constr_x != child->UNCONSTR)
cell_locs[child->name].x = std::min(max_x, base.x + child->constr_x);
else
cell_locs[child->name].x = base.x; // better handling of UNCONSTR?
if (child->constr_y != child->UNCONSTR)
cell_locs[child->name].y = std::min(max_y, base.y + child->constr_y);
else
cell_locs[child->name].y = base.y; // better handling of UNCONSTR?
chain_root[cell->name] = root;
if (!child->constr_children.empty())
update_chain(child, root);
}
}
// Update all chains
void update_all_chains()
{
for (auto cell : place_cells) {
chain_size[cell->name] = 1;
if (!cell->constr_children.empty())
update_chain(cell, cell);
}
}
// Run a function on all ports of a net - including the driver and all users
template <typename Tf> void foreach_port(NetInfo *net, Tf func)
{
if (net->driver.cell != nullptr)
func(net->driver);
for (auto &user : net->users)
func(user);
}
// Build the system of equations for either X or Y
void build_equations(EquationSystem<double> &es, bool yaxis, int iter = -1)
{
// Return the x or y position of a cell, depending on ydir
auto cell_pos = [&](CellInfo *cell) { return yaxis ? cell_locs.at(cell->name).y : cell_locs.at(cell->name).x; };
es.reset();
for (auto net : sorted(ctx->nets)) {
NetInfo *ni = net.second;
if (ni->driver.cell == nullptr)
continue;
if (ni->users.empty())
continue;
if (cell_locs.at(ni->driver.cell->name).global)
continue;
// Find the bounds of the net in this axis, and the ports that correspond to these bounds
PortRef *lbport = nullptr, *ubport = nullptr;
int lbpos = std::numeric_limits<int>::max(), ubpos = std::numeric_limits<int>::min();
foreach_port(ni, [&](PortRef &port) {
int pos = cell_pos(port.cell);
if (pos < lbpos) {
lbpos = pos;
lbport = &port;
}
if (pos > ubpos) {
ubpos = pos;
ubport = &port;
}
});
NPNR_ASSERT(lbport != nullptr);
NPNR_ASSERT(ubport != nullptr);
auto stamp_equation = [&](PortRef &var, PortRef &eqn, double weight) {
if (eqn.cell->udata == dont_solve)
return;
int row = eqn.cell->udata;
int v_pos = cell_pos(var.cell);
if (var.cell->udata != dont_solve) {
es.add_coeff(row, var.cell->udata, weight);
} else {
es.add_rhs(row, -v_pos * weight);
}
if (cell_offsets.count(var.cell->name)) {
es.add_rhs(row, -(yaxis ? cell_offsets.at(var.cell->name).second
: cell_offsets.at(var.cell->name).first) *
weight);
}
};
// Add all relevant connections to the matrix
foreach_port(ni, [&](PortRef &port) {
int this_pos = cell_pos(port.cell);
auto process_arc = [&](PortRef *other) {
if (other == &port)
return;
int o_pos = cell_pos(other->cell);
// if (o_pos == this_pos)
// return; // FIXME: or clamp to 1?
double weight = 1.0 / (ni->users.size() * std::max(1, std::abs(o_pos - this_pos)));
// FIXME: add criticality to weighting
// If cell 0 is not fixed, it will stamp +w on its equation and -w on the other end's equation,
// if the other end isn't fixed
stamp_equation(port, port, weight);
stamp_equation(port, *other, -weight);
stamp_equation(*other, *other, weight);
stamp_equation(*other, port, -weight);
};
process_arc(lbport);
process_arc(ubport);
});
}
if (iter != -1) {
const float alpha = 0.1;
float weight = alpha * iter;
for (size_t row = 0; row < solve_cells.size(); row++) {
// Add an arc from legalised to current position
es.add_coeff(row, row, weight);
es.add_rhs(row, weight * cell_pos(solve_cells.at(row)));
}
}
}
// Build the system of equations for either X or Y
void solve_equations(EquationSystem<double> &es, bool yaxis)
{
// Return the x or y position of a cell, depending on ydir
auto cell_pos = [&](CellInfo *cell) { return yaxis ? cell_locs.at(cell->name).y : cell_locs.at(cell->name).x; };
std::vector<double> vals;
std::transform(solve_cells.begin(), solve_cells.end(), std::back_inserter(vals), cell_pos);
es.solve(vals);
for (size_t i = 0; i < vals.size(); i++)
if (yaxis) {
cell_locs.at(solve_cells.at(i)->name).rawy = vals.at(i);
cell_locs.at(solve_cells.at(i)->name).y = std::min(max_y, std::max(0, int(vals.at(i))));
} else {
cell_locs.at(solve_cells.at(i)->name).rawx = vals.at(i);
cell_locs.at(solve_cells.at(i)->name).x = std::min(max_x, std::max(0, int(vals.at(i))));
}
}
// Compute HPWL
wirelen_t total_hpwl()
{
wirelen_t hpwl = 0;
for (auto net : sorted(ctx->nets)) {
NetInfo *ni = net.second;
if (ni->driver.cell == nullptr)
continue;
CellLocation &drvloc = cell_locs.at(ni->driver.cell->name);
if (drvloc.global)
continue;
int xmin = drvloc.x, xmax = drvloc.x, ymin = drvloc.y, ymax = drvloc.y;
for (auto &user : ni->users) {
CellLocation &usrloc = cell_locs.at(user.cell->name);
xmin = std::min(xmin, usrloc.x);
xmax = std::max(xmax, usrloc.x);
ymin = std::min(ymin, usrloc.y);
ymax = std::max(ymax, usrloc.y);
}
hpwl += (xmax - xmin) + (ymax - ymin);
}
return hpwl;
}
// Swap the Bel of a cell with another, return the original location
BelId swap_cell_bels(CellInfo *cell, BelId newBel)
{
BelId oldBel = cell->bel;
CellInfo *bound = ctx->getBoundBelCell(newBel);
if (bound != nullptr)
ctx->unbindBel(newBel);
ctx->unbindBel(oldBel);
ctx->bindBel(newBel, cell, STRENGTH_WEAK);
if (bound != nullptr)
ctx->bindBel(oldBel, bound, STRENGTH_WEAK);
return oldBel;
}
// Placement legalisation
// Note that there are *two meanings* of legalisation in nextpnr placement
// The first kind, as in HeAP, simply ensures that there is no overlap (each Bel maps only to one cell)
// The second kind also ensures that validity rules (isValidBelForCell) are met, because there is no guarantee
// in nextpnr that Bels are freely swappable (indeed many a architectures Bel is a logic cell with complex
// validity rules for control sets, etc, rather than a CLB/tile as in a more conventional pack&place flow)
void legalise_placement_simple(bool require_validity = false)
{
// Unbind all cells placed in this solution
for (auto cell : sorted(ctx->cells)) {
CellInfo *ci = cell.second;
if (ci->udata != dont_solve && ci->bel != BelId())
ctx->unbindBel(ci->bel);
}
// At the moment we don't follow the full HeAP algorithm using cuts for legalisation, instead using
// the simple greedy largest-macro-first approach.
std::priority_queue<std::pair<int, IdString>> remaining;
for (auto cell : solve_cells) {
remaining.emplace(chain_size[cell->name], cell->name);
}
while (!remaining.empty()) {
auto top = remaining.top();
remaining.pop();
CellInfo *ci = ctx->cells.at(top.second).get();
// Was now placed, ignore
if (ci->bel != BelId())
continue;
// log_info(" Legalising %s\n", top.second.c_str(ctx));
int bt = std::get<0>(bel_types.at(ci->type));
auto &fb = fast_bels.at(bt);
int radius = 0;
int iter = 0;
bool placed = false;
while (!placed) {
int nx = ctx->rng(2 * radius + 1) + std::max(cell_locs.at(ci->name).x - radius, 0);
int ny = ctx->rng(2 * radius + 1) + std::max(cell_locs.at(ci->name).y - radius, 0);
iter++;
if ((iter % (20 * (radius + 1))) == 0)
radius = std::min(std::max(max_x, max_y), radius + 1);
if (nx < 0 || nx > max_x)
continue;
if (ny < 0 || ny > max_x)
continue;
// ny = nearest_row_with_bel.at(bt).at(ny);
// nx = nearest_col_with_bel.at(bt).at(nx);
if (nx >= int(fb.size()))
continue;
if (ny >= int(fb.at(nx).size()))
continue;
if (fb.at(nx).at(ny).empty())
continue;
if (ci->constr_children.empty()) {
for (auto sz : fb.at(nx).at(ny)) {
if (ctx->checkBelAvail(sz) || radius > (max_x / 4)) {
CellInfo *bound = ctx->getBoundBelCell(sz);
if (bound != nullptr) {
if (bound->constr_parent != nullptr || !bound->constr_children.empty())
continue;
ctx->unbindBel(bound->bel);
remaining.emplace(chain_size[bound->name], bound->name);
}
ctx->bindBel(sz, ci, STRENGTH_WEAK);
if (require_validity && !ctx->isBelLocationValid(sz)) {
ctx->unbindBel(sz);
if (bound != nullptr)
ctx->bindBel(sz, bound, STRENGTH_WEAK);
} else {
Loc loc = ctx->getBelLocation(sz);
cell_locs[ci->name].x = loc.x;
cell_locs[ci->name].y = loc.y;
placed = true;
break;
}
}
}
} else {
// FIXME
NPNR_ASSERT(false);
}
}
}
}
static constexpr float beta = 0.9;
struct ChainExtent
{
int x0, y0, x1, y1;
};
struct LegaliserRegion
{
int id;
int x0, y0, x1, y1;
int cells, bels;
std::unordered_set<IdString> included_chains;
bool overused() const
{
if (bels < 4)
return cells > bels;
else
return cells > beta * bels;
}
};
class CutLegaliser
{
public:
CutLegaliser(HeAPPlacer *p, IdString beltype)
: p(p), ctx(p->ctx), beltype(beltype), fb(p->fast_bels.at(std::get<0>(p->bel_types.at(beltype))))
{
}
static int seq;
void run()
{
init();
find_overused_regions();
expand_regions();
std::queue<std::pair<int, bool>> workqueue;
std::vector<std::pair<double, double>> orig;
if (ctx->debug)
for (auto c : p->solve_cells)
orig.emplace_back(p->cell_locs[c->name].rawx, p->cell_locs[c->name].rawy);
for (auto &r : regions) {
if (merged_regions.count(r.id))
continue;
log_info("%s (%d, %d) |_> (%d, %d) %d/%d\n", beltype.c_str(ctx), r.x0, r.y0, r.x1, r.y1, r.cells,
r.bels);
workqueue.emplace(r.id, false);
//cut_region(r, false);
}
while (!workqueue.empty()) {
auto front = workqueue.front();
workqueue.pop();
auto &r = regions.at(front.first);
//log_info("%s (%d, %d) |_> (%d, %d) %d/%d\n", beltype.c_str(ctx), r.x0, r.y0, r.x1, r.y1, r.cells,
//r.bels);
auto res = cut_region(r, front.second);
if (res) {
workqueue.emplace(res->first, !front.second);
workqueue.emplace(res->second, !front.second);
}
}
if (ctx->debug) {
std::ofstream sp("spread" + std::to_string(seq) + ".csv");
for (size_t i = 0; i < p->solve_cells.size(); i++) {
auto &c = p->solve_cells.at(i);
sp << orig.at(i).first << "," << orig.at(i).second << "," << p->cell_locs[c->name].rawx << "," << p->cell_locs[c->name].rawy << std::endl;
}
++seq;
}
}
private:
HeAPPlacer *p;
Context *ctx;
IdString beltype;
std::vector<std::vector<int>> occupancy;
std::vector<std::vector<int>> groups;
std::vector<std::vector<ChainExtent>> chaines;
std::map<IdString, ChainExtent> cell_extents;
std::vector<std::vector<std::vector<BelId>>> &fb;
std::vector<LegaliserRegion> regions;
std::unordered_set<int> merged_regions;
// Cells at a location, sorted by real (not integer) x and y
std::vector<std::vector<std::vector<CellInfo *>>> cells_at_location;
int occ_at(int x, int y) { return occupancy.at(x).at(y); }
int bels_at(int x, int y)
{
if (x >= int(fb.size()) || y >= int(fb.at(x).size()))
return 0;
return int(fb.at(x).at(y).size());
}
void init()
{
occupancy.resize(p->max_x + 1, std::vector<int>(p->max_y + 1, 0));
groups.resize(p->max_x + 1, std::vector<int>(p->max_y + 1, -1));
chaines.resize(p->max_x + 1, std::vector<ChainExtent>(p->max_y + 1));
cells_at_location.resize(p->max_x + 1, std::vector<std::vector<CellInfo *>>(p->max_y + 1));
for (int x = 0; x <= p->max_x; x++)
for (int y = 0; y <= p->max_y; y++) {
occupancy.at(x).at(y) = 0;
groups.at(x).at(y) = -1;
chaines.at(x).at(y) = {x, y, x, y};
}
auto set_chain_ext = [&](IdString cell, int x, int y) {
if (!cell_extents.count(cell))
cell_extents[cell] = {x, y, x, y};
else {
cell_extents[cell].x0 = std::min(cell_extents[cell].x0, x);
cell_extents[cell].y0 = std::min(cell_extents[cell].y0, y);
cell_extents[cell].x1 = std::max(cell_extents[cell].x1, x);
cell_extents[cell].y1 = std::max(cell_extents[cell].y1, y);
}
};
for (auto &cell : p->cell_locs) {
if (ctx->cells.at(cell.first)->type == beltype)
occupancy.at(cell.second.x).at(cell.second.y)++;
// Compute ultimate extent of each chain root
if (p->chain_root.count(cell.first)) {
set_chain_ext(p->chain_root.at(cell.first)->name, cell.second.x, cell.second.y);
} else if (!ctx->cells.at(cell.first)->constr_children.empty()) {
set_chain_ext(cell.first, cell.second.x, cell.second.y);
}
}
for (auto &cell : p->cell_locs) {
// Transfer chain extents to the actual chaines structure
ChainExtent *ce = nullptr;
if (p->chain_root.count(cell.first))
ce = &(cell_extents.at(p->chain_root.at(cell.first)->name));
else if (!ctx->cells.at(cell.first)->constr_children.empty())
ce = &(cell_extents.at(cell.first));
if (ce) {
auto &lce = chaines.at(cell.second.x).at(cell.second.y);
lce.x0 = std::min(lce.x0, ce->x0);
lce.y0 = std::min(lce.y0, ce->y0);
lce.x1 = std::max(lce.x1, ce->x1);
lce.y1 = std::max(lce.y1, ce->y1);
}
}
for (auto cell : p->solve_cells) {
cells_at_location.at(p->cell_locs.at(cell->name).x)
.at(p->cell_locs.at(cell->name).y)
.push_back(cell);
}
}
void merge_regions(LegaliserRegion &merged, LegaliserRegion &mergee)
{
// Prevent grow_region from recursing while doing this
for (int x = mergee.x0; x <= mergee.x1; x++)
for (int y = mergee.y0; y <= mergee.y1; y++) {
// log_info("%d %d\n", groups.at(x).at(y), mergee.id);
NPNR_ASSERT(groups.at(x).at(y) == mergee.id);
groups.at(x).at(y) = merged.id;
merged.cells += occ_at(x, y);
merged.bels += bels_at(x, y);
}
merged_regions.insert(mergee.id);
grow_region(merged, mergee.x0, mergee.y0, mergee.x1, mergee.y1);
}
void grow_region(LegaliserRegion &r, int x0, int y0, int x1, int y1, bool init = false)
{
// log_info("growing to (%d, %d) |_> (%d, %d)\n", x0, y0, x1, y1);
if ((x0 >= r.x0 && y0 >= r.y0 && x1 <= r.x1 && y1 <= r.y1) || init)
return;
int old_x0 = r.x0 + (init ? 1 : 0), old_y0 = r.y0, old_x1 = r.x1, old_y1 = r.y1;
r.x0 = std::min(r.x0, x0);
r.y0 = std::min(r.y0, y0);
r.x1 = std::max(r.x1, x1);
r.y1 = std::max(r.y1, y1);
auto process_location = [&](int x, int y) {
// Merge with any overlapping regions
if (groups.at(x).at(y) == -1) {
r.bels += bels_at(x, y);
r.cells += occ_at(x, y);
}
if (groups.at(x).at(y) != -1 && groups.at(x).at(y) != r.id)
merge_regions(r, regions.at(groups.at(x).at(y)));
groups.at(x).at(y) = r.id;
// Grow to cover any chains
auto &chaine = chaines.at(x).at(y);
grow_region(r, chaine.x0, chaine.y0, chaine.x1, chaine.y1);
};
for (int x = r.x0; x < old_x0; x++)
for (int y = r.y0; y <= r.y1; y++)
process_location(x, y);
for (int x = old_x1 + 1; x <= x1; x++)
for (int y = r.y0; y <= r.y1; y++)
process_location(x, y);
for (int y = r.y0; y < old_y0; y++)
for (int x = r.x0; x <= r.x1; x++)
process_location(x, y);
for (int y = old_y1 + 1; y <= r.y1; y++)
for (int x = r.x0; x <= r.x1; x++)
process_location(x, y);
}
void find_overused_regions()
{
for (int x = 0; x <= p->max_x; x++)
for (int y = 0; y <= p->max_y; y++) {
// Either already in a group, or not overutilised. Ignore
if (groups.at(x).at(y) != -1 || (occ_at(x, y) <= bels_at(x, y)))
continue;
// log_info("%d %d %d\n", x, y, occ_at(x, y));
int id = int(regions.size());
groups.at(x).at(y) = id;
LegaliserRegion reg;
reg.id = id;
reg.x0 = reg.x1 = x;
reg.y0 = reg.y1 = y;
reg.bels = bels_at(x, y);
reg.cells = occ_at(x, y);
// Make sure we cover carries, etc
grow_region(reg, reg.x0, reg.y0, reg.x1, reg.y1, true);
bool expanded = true;
while (expanded) {
expanded = false;
// Keep trying expansion in x and y, until we find no over-occupancy cells
// or hit grouped cells
// First try expanding in x
if (reg.x1 < p->max_x) {
bool over_occ_x = false;
for (int y1 = reg.y0; y1 <= reg.y1; y1++) {
if (occ_at(reg.x1 + 1, y1) > bels_at(reg.x1 + 1, y1)) {
// log_info("(%d, %d) occ %d bels %d\n", reg.x1+ 1, y1, occ_at(reg.x1 + 1, y1),
// bels_at(reg.x1 + 1, y1));
over_occ_x = true;
break;
}
}
if (over_occ_x) {
expanded = true;
grow_region(reg, reg.x0, reg.y0, reg.x1 + 1, reg.y1);
}
}
if (reg.y1 < p->max_y) {
bool over_occ_y = false;
for (int x1 = reg.x0; x1 <= reg.x1; x1++) {
if (occ_at(x1, reg.y1 + 1) > bels_at(x1, reg.y1 + 1)) {
// log_info("(%d, %d) occ %d bels %d\n", x1, reg.y1 + 1, occ_at(x1, reg.y1 + 1),
// bels_at(x1, reg.y1 + 1));
over_occ_y = true;
break;
}
}
if (over_occ_y) {
expanded = true;
grow_region(reg, reg.x0, reg.y0, reg.x1, reg.y1 + 1);
}
}
}
regions.push_back(reg);
}
}
void expand_regions()
{
std::queue<int> overu_regions;
for (auto &r : regions) {
if (!merged_regions.count(r.id) && r.overused())
overu_regions.push(r.id);
}
while (!overu_regions.empty()) {
int rid = overu_regions.front();
overu_regions.pop();
if (merged_regions.count(rid))
continue;
auto &reg = regions.at(rid);
while (reg.overused()) {
bool changed = false;
if (reg.x0 > 0) {
grow_region(reg, reg.x0 - 1, reg.y0, reg.x1, reg.y1);
changed = true;
if (!reg.overused())
break;
}
if (reg.x1 < p->max_x) {
grow_region(reg, reg.x0, reg.y0, reg.x1 + 1, reg.y1);
changed = true;
if (!reg.overused())
break;
}
if (reg.y0 > 0) {
grow_region(reg, reg.x0, reg.y0 - 1, reg.x1, reg.y1);
changed = true;
if (!reg.overused())
break;
}
if (reg.y1 < p->max_y) {
grow_region(reg, reg.x0, reg.y0, reg.x1, reg.y1 + 1);
changed = true;
if (!reg.overused())
break;
}
if (!changed)
log_error("Failed to expand region (%d, %d) |_> (%d, %d) of %d %ss\n", reg.x0, reg.y0, reg.x1,
reg.y1, reg.cells, beltype.c_str(ctx));
}
}
}
// Implementation of the recursive cut-based spreading as described in the HeAP paper
// Note we use "left" to mean "-x/-y" depending on dir and "right" to mean "+x/+y" depending on dir
std::vector<CellInfo *> cut_cells;
boost::optional<std::pair<int, int>> cut_region(LegaliserRegion &r, bool dir)
{
cut_cells.clear();
auto &cal = cells_at_location;
for (int x = r.x0; x <= r.x1; x++) {
for (int y = r.y0; y <= r.y1; y++) {
std::copy(cal.at(x).at(y).begin(), cal.at(x).at(y).end(), std::back_inserter(cut_cells));
}
}
std::sort(cut_cells.begin(), cut_cells.end(), [&](const CellInfo *a, const CellInfo *b) {
return dir ? (p->cell_locs.at(a->name).rawy < p->cell_locs.at(b->name).rawy) : (p->cell_locs.at(a->name).rawx < p->cell_locs.at(b->name).rawx);
});
if (cut_cells.empty())
return {};
// Find the cells midpoint, counting chains in terms of their total size - making the initial source cut
int pivot_cells = 0;
int pivot = 0;
for (auto &cell : cut_cells) {
pivot_cells += p->chain_size.count(cell->name) ? p->chain_size.at(cell->name) : 1;
if (pivot_cells >= r.cells / 2)
break;
pivot++;
}
if (pivot == int(cut_cells.size()))
pivot = int(cut_cells.size()) - 1;
//log_info("orig pivot %d lc %d rc %d\n", pivot, pivot_cells, r.cells - pivot_cells);
// Find the clearance required either side of the pivot
int clearance_l = 0, clearance_r = 0;
for (size_t i = 0; i < cut_cells.size(); i++) {
int size;
if (cell_extents.count(cut_cells.at(i)->name)) {
auto &ce = cell_extents.at(cut_cells.at(i)->name);
size = dir ? (ce.y1 - ce.y0 + 1) : (ce.x1 - ce.x0 + 1);
} else {
size = 1;
}
if (int(i) < pivot)
clearance_l = std::max(clearance_l, size);
else
clearance_r = std::max(clearance_r, size);
}
// Find the target cut that minimises difference in utilisation, whilst trying to ensure that all chains
// still fit
// First trim the boundaries of the region in the axis-of-interest, skipping any rows/cols without any
// bels of the appropriate type
int trimmed_l = dir ? r.y0 : r.x0, trimmed_r = dir ? r.y1 : r.x1;
while (trimmed_l < (dir ? r.y1 : r.x1)) {
bool have_bels = false;
for (int i = dir ? r.x0 : r.y0; i <= (dir ? r.x1 : r.y1); i++)
if (bels_at(dir ? i : trimmed_l, dir ? trimmed_l : i) > 0) {
have_bels = true;
break;
}
if (have_bels)
break;
trimmed_l++;
}
while (trimmed_r > (dir ? r.y0 : r.x0)) {
bool have_bels = false;
for (int i = dir ? r.x0 : r.y0; i <= (dir ? r.x1 : r.y1); i++)
if (bels_at(dir ? i : trimmed_r, dir ? trimmed_r : i) > 0) {
have_bels = true;
break;
}
if (have_bels)
break;
trimmed_r--;
}
//log_info("tl %d tr %d cl %d cr %d\n", trimmed_l, trimmed_r, clearance_l, clearance_r);
if ((trimmed_r - trimmed_l + 1) <= std::max(clearance_l, clearance_r))
return {};
// Now find the initial target cut that minimises utilisation imbalance, whilst
// meeting the clearance requirements for any large macros
int left_cells = pivot_cells, right_cells = r.cells - pivot_cells;
int left_bels = 0, right_bels = r.bels;
int best_tgt_cut = -1;
double best_deltaU = std::numeric_limits<double>::max();
std::pair<int, int> target_cut_bels;
for (int i = trimmed_l; i <= trimmed_r; i++) {
int slither_bels = 0;
for (int j = dir ? r.x0 : r.y0; j <= (dir ? r.x1 : r.y1); j++) {
slither_bels += dir ? bels_at(j, i) : bels_at(i, j);
}
left_bels += slither_bels;
right_bels -= slither_bels;
if (((i - trimmed_l) + 1) >= clearance_l && ((trimmed_r - i) + 1) >= clearance_r) {
// Solution is potentially valid
double aU =
std::abs(double(left_cells) / double(left_bels) - double(right_cells) / double(right_bels));
if (aU < best_deltaU) {
best_deltaU = aU;
best_tgt_cut = i;
target_cut_bels = std::make_pair(left_bels, right_bels);
}
}
}
NPNR_ASSERT(best_tgt_cut != -1);
left_bels = target_cut_bels.first;
right_bels = target_cut_bels.second;
//log_info("pivot %d target cut %d lc %d lb %d rc %d rb %d\n", pivot, best_tgt_cut, left_cells, left_bels, right_cells, right_bels);
// Peturb the source cut to eliminate overutilisation
while (pivot > 0 && (double(left_cells) / double(left_bels) > double(right_cells) / double(right_bels))) {
auto &move_cell = cut_cells.at(pivot);
int size = p->chain_size.count(move_cell->name) ? p->chain_size.at(move_cell->name) : 1;
left_cells -= size;
right_cells += size;
pivot--;
}
while (pivot < int(cut_cells.size()) - 1 && (double(left_cells) / double(left_bels) < double(right_cells) / double(right_bels))) {
auto &move_cell = cut_cells.at(pivot + 1);
int size = p->chain_size.count(move_cell->name) ? p->chain_size.at(move_cell->name) : 1;
left_cells += size;
right_cells -= size;
pivot++;
}
//log_info("peturbed pivot %d lc %d lb %d rc %d rb %d\n", pivot, left_cells, left_bels, right_cells, right_bels);
// Split regions into bins, and then spread cells by linear interpolation within those bins
auto spread_binlerp = [&](int cells_start, int cells_end, double area_l, double area_r) {
int N = cells_end - cells_start;
if (N <= 2) {
for (int i = cells_start; i < cells_end; i++) {
auto &pos = dir ? p->cell_locs.at(cut_cells.at(i)->name).rawy
: p->cell_locs.at(cut_cells.at(i)->name).rawx;
pos = area_l + i * ((area_r - area_l) / N);
}
return;
}
// Split region into up to 10 (K) bins
int K = std::min<int>(N, 10);
std::vector<std::pair<int, double>> bin_bounds; // [start, end]
bin_bounds.emplace_back(cells_start, area_l);
for (int i = 1; i < K; i++)
bin_bounds.emplace_back(cells_start + (N * i) / K,
area_l + ((area_r - area_l + 0.9) * i) / K);
bin_bounds.emplace_back(cells_end, area_r + 0.9);
//log("bins ");
//for (auto b : bin_bounds) log("%d, %.01f; ", b.first, b.second);
//log("\n");
for (int i = 0; i < K; i++) {
auto &bl = bin_bounds.at(i), br = bin_bounds.at(i + 1);
double orig_left = dir ? p->cell_locs.at(cut_cells.at(bl.first)->name).rawy
: p->cell_locs.at(cut_cells.at(bl.first)->name).rawx;
double orig_right = dir ? p->cell_locs.at(cut_cells.at(br.first - 1)->name).rawy
: p->cell_locs.at(cut_cells.at(br.first - 1)->name).rawx;
double m = (br.second - bl.second) / (1 + orig_right - orig_left);
for (int j = bl.first; j < br.first; j++) {
auto &pos = dir ? p->cell_locs.at(cut_cells.at(j)->name).rawy
: p->cell_locs.at(cut_cells.at(j)->name).rawx;
NPNR_ASSERT(pos >= orig_left && pos <= orig_right);
pos = bl.second + m * (pos - orig_left);
}
}
};
spread_binlerp(0, pivot + 1, trimmed_l, best_tgt_cut);
spread_binlerp(pivot + 1, int(cut_cells.size()), best_tgt_cut + 1, trimmed_r);
// Update various data structures
for (int x = r.x0; x <= r.x1; x++)
for (int y = r.y0; y <= r.y1; y++) {
cells_at_location.at(x).at(y).clear();
}
for (auto cell : cut_cells) {
auto &cl = p->cell_locs.at(cell->name);
cl.x = std::min(r.x1, std::max(r.x0, int(cl.rawx)));
cl.y = std::min(r.y1, std::max(r.y0, int(cl.rawy)));
cells_at_location.at(cl.x).at(cl.y).push_back(cell);
//log_info("spread pos %d %d\n", cl.x, cl.y);
}
LegaliserRegion rl, rr;
rl.id = int(regions.size());
rl.x0 = r.x0;
rl.y0 = r.y0;
rl.x1 = dir ? r.x1 : best_tgt_cut;
rl.y1 = dir ? best_tgt_cut : r.y1;
rl.cells = left_cells;
rl.bels = left_bels;
rr.id = int(regions.size()) + 1;
rr.x0 = dir ? r.x0 : (best_tgt_cut + 1);
rr.y0 = dir ? (best_tgt_cut + 1) : r.y0;
rr.x1 = r.x1;
rr.y1 = r.y1;
rr.cells = right_cells;
rr.bels = right_bels;
regions.push_back(rl);
regions.push_back(rr);
for (int x = rl.x0; x <= rl.x1; x++)
for (int y = rl.y0; y <= rl.y1; y++)
groups.at(x).at(y) = rl.id;
for (int x = rr.x0; x <= rr.x1; x++)
for (int y = rr.y0; y <= rr.y1; y++)
groups.at(x).at(y) = rr.id;
return std::make_pair(rl.id, rr.id);
};
};
typedef decltype(CellInfo::udata) cell_udata_t;
cell_udata_t dont_solve = std::numeric_limits<cell_udata_t>::max();
};
int HeAPPlacer::CutLegaliser::seq = 0;
bool placer_heap(Context *ctx) { return HeAPPlacer(ctx).place(); }
NEXTPNR_NAMESPACE_END
Reference in New Issue View Git Blame Copy Permalink