Google Git
Sign in
foss-fpga-tools / third_party / vtr-verilog-to-routing / 83fdb1529c902ff7390bcb0fb09d119aef837bbd / . / vpr / src / route / route_common.cpp
blob: 350a3406bb4690ec291ea070aadfb2cf94ab45d8 [file] [log] [blame]
#include <cstdio>
#include <ctime>
#include <cmath>
#include <algorithm>
#include <vector>
#include <iostream>
#include "vtr_assert.h"
#include "vtr_util.h"
#include "vtr_log.h"
#include "vtr_digest.h"
#include "vtr_memory.h"
#include "vpr_types.h"
#include "vpr_error.h"
#include "vpr_utils.h"
#include "stats.h"
#include "globals.h"
#include "route_export.h"
#include "route_common.h"
#include "route_tree_timing.h"
#include "route_timing.h"
#include "route_breadth_first.h"
#include "place_and_route.h"
#include "rr_graph.h"
#include "rr_graph2.h"
#include "read_xml_arch_file.h"
#include "draw.h"
#include "echo_files.h"
#include "route_profiling.h"
#include "timing_util.h"
#include "RoutingDelayCalculator.h"
#include "timing_info.h"
#include "tatum/echo_writer.hpp"
/**************** Types local to route_common.c ******************/
struct t_trace_branch {
t_trace* head;
t_trace* tail;
};
/**************** Static variables local to route_common.c ******************/
static t_heap** heap; /* Indexed from [1..heap_size] */
static int heap_size; /* Number of slots in the heap array */
static int heap_tail; /* Index of first unused slot in the heap array */
/* For managing my own list of currently free heap data structures. */
static t_heap* heap_free_head = nullptr;
/* For keeping track of the sudo malloc memory for the heap*/
static vtr::t_chunk heap_ch;
/* For managing my own list of currently free trace data structures. */
static t_trace* trace_free_head = nullptr;
/* For keeping track of the sudo malloc memory for the trace*/
static vtr::t_chunk trace_ch;
static int num_trace_allocated = 0; /* To watch for memory leaks. */
static int num_heap_allocated = 0;
static int num_linked_f_pointer_allocated = 0;
/* The numbering relation between the channels and clbs is: *
* *
* | IO | chan_ | CLB | chan_ | CLB | *
* |grid[0][2] | y[0][2] |grid[1][2] | y[1][2] | grid[2][2]| *
* +-----------+ +-----------+ +-----------+ *
* } capacity in *
* No channel chan_x[1][1] chan_x[2][1] } chan_width *
* } _x[1] *
* +-----------+ +-----------+ +-----------+ *
* | | chan_ | | chan_ | | *
* | IO | y[0][1] | CLB | y[1][1] | CLB | *
* |grid[0][1] | |grid[1][1] | |grid[2][1] | *
* | | | | | | *
* +-----------+ +-----------+ +-----------+ *
* } capacity in *
* chan_x[1][0] chan_x[2][0] } chan_width *
* } _x[0] *
* +-----------+ +-----------+ *
* No | | No | | *
* Channel | IO | Channel | IO | *
* |grid[1][0] | |grid[2][0] | *
* | | | | *
* +-----------+ +-----------+ *
* *
* {=======} {=======} *
* Capacity in Capacity in *
* chan_width_y[0] chan_width_y[1] *
* */
/******************** Subroutines local to route_common.c *******************/
static t_trace_branch traceback_branch(int node, std::unordered_set<int>& main_branch_visited);
static std::pair<t_trace*, t_trace*> add_trace_non_configurable(t_trace* head, t_trace* tail, int node, std::unordered_set<int>& visited);
static std::pair<t_trace*, t_trace*> add_trace_non_configurable_recurr(int node, std::unordered_set<int>& visited, int depth = 0);
static vtr::vector<ClusterNetId, std::vector<int>> load_net_rr_terminals(const t_rr_node_indices& L_rr_node_indices);
static vtr::vector<ClusterBlockId, std::vector<int>> load_rr_clb_sources(const t_rr_node_indices& L_rr_node_indices);
static t_clb_opins_used alloc_and_load_clb_opins_used_locally();
static void adjust_one_rr_occ_and_apcost(int inode, int add_or_sub, float pres_fac, float acc_fac);
bool validate_traceback_recurr(t_trace* trace, std::set<int>& seen_rr_nodes);
static bool validate_trace_nodes(t_trace* head, const std::unordered_set<int>& trace_nodes);
static float get_single_rr_cong_cost(int inode);
/************************** Subroutine definitions ***************************/
void save_routing(vtr::vector<ClusterNetId, t_trace*>& best_routing,
const t_clb_opins_used& clb_opins_used_locally,
t_clb_opins_used& saved_clb_opins_used_locally) {
/* This routing frees any routing currently held in best routing, *
* then copies over the current routing (held in route_ctx.trace), and *
* finally sets route_ctx.trace_head and route_ctx.trace_tail to all NULLs so that the *
* connection to the saved routing is broken. This is necessary so *
* that the next iteration of the router does not free the saved *
* routing elements. Also saves any data about locally used clb_opins, *
* since this is also part of the routing. */
t_trace *tptr, *tempptr;
auto& cluster_ctx = g_vpr_ctx.clustering();
auto& route_ctx = g_vpr_ctx.mutable_routing();
for (auto net_id : cluster_ctx.clb_nlist.nets()) {
/* Free any previously saved routing. It is no longer best. */
tptr = best_routing[net_id];
while (tptr != nullptr) {
tempptr = tptr->next;
free_trace_data(tptr);
tptr = tempptr;
}
/* Save a pointer to the current routing in best_routing. */
best_routing[net_id] = route_ctx.trace[net_id].head;
/* Set the current (working) routing to NULL so the current trace *
* elements won't be reused by the memory allocator. */
route_ctx.trace[net_id].head = nullptr;
route_ctx.trace[net_id].tail = nullptr;
route_ctx.trace_nodes[net_id].clear();
}
/* Save which OPINs are locally used. */
saved_clb_opins_used_locally = clb_opins_used_locally;
}
/* Deallocates any current routing in route_ctx.trace_head, and replaces it with *
* the routing in best_routing. Best_routing is set to NULL to show that *
* it no longer points to a valid routing. NOTE: route_ctx.trace_tail is not *
* restored -- it is set to all NULLs since it is only used in *
* update_traceback. If you need route_ctx.trace_tail restored, modify this *
* routine. Also restores the locally used opin data. */
void restore_routing(vtr::vector<ClusterNetId, t_trace*>& best_routing,
t_clb_opins_used& clb_opins_used_locally,
const t_clb_opins_used& saved_clb_opins_used_locally) {
auto& cluster_ctx = g_vpr_ctx.clustering();
auto& route_ctx = g_vpr_ctx.mutable_routing();
for (auto net_id : cluster_ctx.clb_nlist.nets()) {
/* Free any current routing. */
pathfinder_update_path_cost(route_ctx.trace[net_id].head, -1, 0.f);
free_traceback(net_id);
/* Set the current routing to the saved one. */
route_ctx.trace[net_id].head = best_routing[net_id];
pathfinder_update_path_cost(route_ctx.trace[net_id].head, 1, 0.f);
best_routing[net_id] = nullptr; /* No stored routing. */
}
/* Restore which OPINs are locally used. */
clb_opins_used_locally = saved_clb_opins_used_locally;
}
/* This routine finds a "magic cookie" for the routing and prints it. *
* Use this number as a routing serial number to ensure that programming *
* changes do not break the router. */
void get_serial_num() {
int serial_num, inode;
t_trace* tptr;
auto& cluster_ctx = g_vpr_ctx.clustering();
auto& route_ctx = g_vpr_ctx.routing();
auto& device_ctx = g_vpr_ctx.device();
serial_num = 0;
for (auto net_id : cluster_ctx.clb_nlist.nets()) {
/* Global nets will have null trace_heads (never routed) so they *
* are not included in the serial number calculation. */
tptr = route_ctx.trace[net_id].head;
while (tptr != nullptr) {
inode = tptr->index;
serial_num += (size_t(net_id) + 1)
* (device_ctx.rr_nodes[inode].xlow() * (device_ctx.grid.width()) - device_ctx.rr_nodes[inode].yhigh());
serial_num -= device_ctx.rr_nodes[inode].ptc_num() * (size_t(net_id) + 1) * 10;
serial_num -= device_ctx.rr_nodes[inode].type() * (size_t(net_id) + 1) * 100;
serial_num %= 2000000000; /* Prevent overflow */
tptr = tptr->next;
}
}
VTR_LOG("Serial number (magic cookie) for the routing is: %d\n", serial_num);
}
void try_graph(int width_fac, const t_router_opts& router_opts, t_det_routing_arch* det_routing_arch, std::vector<t_segment_inf>& segment_inf, t_chan_width_dist chan_width_dist, t_direct_inf* directs, int num_directs) {
auto& device_ctx = g_vpr_ctx.mutable_device();
t_graph_type graph_type;
if (router_opts.route_type == GLOBAL) {
graph_type = GRAPH_GLOBAL;
} else {
graph_type = (det_routing_arch->directionality == BI_DIRECTIONAL ? GRAPH_BIDIR : GRAPH_UNIDIR);
}
/* Set the channel widths */
t_chan_width chan_width = init_chan(width_fac, chan_width_dist);
/* Free any old routing graph, if one exists. */
free_rr_graph();
/* Set up the routing resource graph defined by this FPGA architecture. */
int warning_count;
create_rr_graph(graph_type,
device_ctx.physical_tile_types,
device_ctx.grid,
chan_width,
device_ctx.num_arch_switches,
det_routing_arch,
segment_inf,
router_opts.base_cost_type,
router_opts.trim_empty_channels,
router_opts.trim_obs_channels,
router_opts.clock_modeling,
directs, num_directs,
&warning_count);
}
bool try_route(int width_fac,
const t_router_opts& router_opts,
const t_analysis_opts& analysis_opts,
t_det_routing_arch* det_routing_arch,
std::vector<t_segment_inf>& segment_inf,
vtr::vector<ClusterNetId, float*>& net_delay,
std::shared_ptr<SetupHoldTimingInfo> timing_info,
std::shared_ptr<RoutingDelayCalculator> delay_calc,
t_chan_width_dist chan_width_dist,
t_direct_inf* directs,
int num_directs,
ScreenUpdatePriority first_iteration_priority) {
/* Attempts a routing via an iterated maze router algorithm. Width_fac *
* specifies the relative width of the channels, while the members of *
* router_opts determine the value of the costs assigned to routing *
* resource node, etc. det_routing_arch describes the detailed routing *
* architecture (connection and switch boxes) of the FPGA; it is used *
* only if a DETAILED routing has been selected. */
auto& device_ctx = g_vpr_ctx.mutable_device();
auto& cluster_ctx = g_vpr_ctx.clustering();
t_graph_type graph_type;
if (router_opts.route_type == GLOBAL) {
graph_type = GRAPH_GLOBAL;
} else {
graph_type = (det_routing_arch->directionality == BI_DIRECTIONAL ? GRAPH_BIDIR : GRAPH_UNIDIR);
}
/* Set the channel widths */
t_chan_width chan_width = init_chan(width_fac, chan_width_dist);
/* Set up the routing resource graph defined by this FPGA architecture. */
int warning_count;
create_rr_graph(graph_type,
device_ctx.physical_tile_types,
device_ctx.grid,
chan_width,
device_ctx.num_arch_switches,
det_routing_arch,
segment_inf,
router_opts.base_cost_type,
router_opts.trim_empty_channels,
router_opts.trim_obs_channels,
router_opts.clock_modeling,
directs, num_directs,
&warning_count);
//Initialize drawing, now that we have an RR graph
init_draw_coords(width_fac);
bool success = true;
/* Allocate and load additional rr_graph information needed only by the router. */
alloc_and_load_rr_node_route_structs();
init_route_structs(router_opts.bb_factor);
if (cluster_ctx.clb_nlist.nets().empty()) {
VTR_LOG_WARN("No nets to route\n");
}
if (router_opts.router_algorithm == BREADTH_FIRST) {
VTR_LOG("Confirming router algorithm: BREADTH_FIRST.\n");
success = try_breadth_first_route(router_opts);
} else { /* TIMING_DRIVEN route */
VTR_LOG("Confirming router algorithm: TIMING_DRIVEN.\n");
IntraLbPbPinLookup intra_lb_pb_pin_lookup(device_ctx.logical_block_types);
ClusteredPinAtomPinsLookup netlist_pin_lookup(cluster_ctx.clb_nlist, intra_lb_pb_pin_lookup);
success = try_timing_driven_route(router_opts,
analysis_opts,
segment_inf,
net_delay,
netlist_pin_lookup,
timing_info,
delay_calc,
first_iteration_priority);
profiling::time_on_fanout_analysis();
}
return (success);
}
bool feasible_routing() {
/* This routine checks to see if this is a resource-feasible routing. *
* That is, are all rr_node capacity limitations respected? It assumes *
* that the occupancy arrays are up to date when it is called. */
auto& device_ctx = g_vpr_ctx.device();
auto& route_ctx = g_vpr_ctx.routing();
for (size_t inode = 0; inode < device_ctx.rr_nodes.size(); inode++) {
if (route_ctx.rr_node_route_inf[inode].occ() > device_ctx.rr_nodes[inode].capacity()) {
return (false);
}
}
return (true);
}
//Returns all RR nodes in the current routing which are congested
std::vector<int> collect_congested_rr_nodes() {
auto& device_ctx = g_vpr_ctx.device();
auto& route_ctx = g_vpr_ctx.routing();
std::vector<int> congested_rr_nodes;
for (size_t inode = 0; inode < device_ctx.rr_nodes.size(); inode++) {
short occ = route_ctx.rr_node_route_inf[inode].occ();
short capacity = device_ctx.rr_nodes[inode].capacity();
if (occ > capacity) {
congested_rr_nodes.push_back(inode);
}
}
return congested_rr_nodes;
}
/* Returns a vector from [0..device_ctx.rr_nodes.size()-1] containing the set
* of nets using each RR node */
std::vector<std::set<ClusterNetId>> collect_rr_node_nets() {
auto& device_ctx = g_vpr_ctx.device();
auto& route_ctx = g_vpr_ctx.routing();
auto& cluster_ctx = g_vpr_ctx.clustering();
std::vector<std::set<ClusterNetId>> rr_node_nets(device_ctx.rr_nodes.size());
for (ClusterNetId inet : cluster_ctx.clb_nlist.nets()) {
t_trace* trace_elem = route_ctx.trace[inet].head;
while (trace_elem) {
int rr_node = trace_elem->index;
rr_node_nets[rr_node].insert(inet);
trace_elem = trace_elem->next;
}
}
return rr_node_nets;
}
void pathfinder_update_path_cost(t_trace* route_segment_start,
int add_or_sub,
float pres_fac) {
/* This routine updates the occupancy and pres_cost of the rr_nodes that are *
* affected by the portion of the routing of one net that starts at *
* route_segment_start. If route_segment_start is route_ctx.trace[net_id].head, the *
* cost of all the nodes in the routing of net net_id are updated. If *
* add_or_sub is -1 the net (or net portion) is ripped up, if it is 1 the *
* net is added to the routing. The size of pres_fac determines how severly *
* oversubscribed rr_nodes are penalized. */
t_trace* tptr;
tptr = route_segment_start;
if (tptr == nullptr) /* No routing yet. */
return;
for (;;) {
pathfinder_update_single_node_cost(tptr->index, add_or_sub, pres_fac);
if (tptr->iswitch == OPEN) { //End of branch
tptr = tptr->next; /* Skip next segment. */
if (tptr == nullptr)
break;
}
tptr = tptr->next;
} /* End while loop -- did an entire traceback. */
}
void pathfinder_update_single_node_cost(int inode, int add_or_sub, float pres_fac) {
/* Updates pathfinder's congestion cost by either adding or removing the
* usage of a resource node. pres_cost is Pn in the Pathfinder paper.
* pres_cost is set according to the overuse that would result from having
* ONE MORE net use this routing node. */
auto& route_ctx = g_vpr_ctx.mutable_routing();
auto& device_ctx = g_vpr_ctx.device();
int occ = route_ctx.rr_node_route_inf[inode].occ() + add_or_sub;
route_ctx.rr_node_route_inf[inode].set_occ(occ);
// can't have negative occupancy
VTR_ASSERT(occ >= 0);
int capacity = device_ctx.rr_nodes[inode].capacity();
if (occ < capacity) {
route_ctx.rr_node_route_inf[inode].pres_cost = 1.0;
} else {
route_ctx.rr_node_route_inf[inode].pres_cost = 1.0 + (occ + 1 - capacity) * pres_fac;
}
}
void pathfinder_update_cost(float pres_fac, float acc_fac) {
/* This routine recomputes the pres_cost and acc_cost of each routing *
* resource for the pathfinder algorithm after all nets have been routed. *
* It updates the accumulated cost to by adding in the number of extra *
* signals sharing a resource right now (i.e. after each complete iteration) *
* times acc_fac. It also updates pres_cost, since pres_fac may have *
* changed. THIS ROUTINE ASSUMES THE OCCUPANCY VALUES IN RR_NODE ARE UP TO *
* DATE. */
int occ, capacity;
auto& device_ctx = g_vpr_ctx.device();
auto& route_ctx = g_vpr_ctx.mutable_routing();
for (size_t inode = 0; inode < device_ctx.rr_nodes.size(); inode++) {
occ = route_ctx.rr_node_route_inf[inode].occ();
capacity = device_ctx.rr_nodes[inode].capacity();
if (occ > capacity) {
route_ctx.rr_node_route_inf[inode].acc_cost += (occ - capacity) * acc_fac;
route_ctx.rr_node_route_inf[inode].pres_cost = 1.0 + (occ + 1 - capacity) * pres_fac;
}
/* If occ == capacity, we don't need to increase acc_cost, but a change *
* in pres_fac could have made it necessary to recompute the cost anyway. */
else if (occ == capacity) {
route_ctx.rr_node_route_inf[inode].pres_cost = 1.0 + pres_fac;
}
}
}
void init_heap(const DeviceGrid& grid) {
if (heap != nullptr) {
vtr::free(heap + 1);
heap = nullptr;
}
heap_size = (grid.width() - 1) * (grid.height() - 1);
heap = (t_heap**)vtr::malloc(heap_size * sizeof(t_heap*));
heap--; /* heap stores from [1..heap_size] */
heap_tail = 1;
}
/* Call this before you route any nets. It frees any old traceback and *
* sets the list of rr_nodes touched to empty. */
void init_route_structs(int bb_factor) {
auto& cluster_ctx = g_vpr_ctx.clustering();
auto& device_ctx = g_vpr_ctx.device();
auto& route_ctx = g_vpr_ctx.mutable_routing();
//Free any old tracebacks
for (auto net_id : cluster_ctx.clb_nlist.nets())
free_traceback(net_id);
//Allocate new tracebacks
route_ctx.trace.resize(cluster_ctx.clb_nlist.nets().size());
route_ctx.trace_nodes.resize(cluster_ctx.clb_nlist.nets().size());
init_heap(device_ctx.grid);
//Various look-ups
route_ctx.net_rr_terminals = load_net_rr_terminals(device_ctx.rr_node_indices);
route_ctx.route_bb = load_route_bb(bb_factor);
route_ctx.rr_blk_source = load_rr_clb_sources(device_ctx.rr_node_indices);
route_ctx.clb_opins_used_locally = alloc_and_load_clb_opins_used_locally();
route_ctx.net_status.resize(cluster_ctx.clb_nlist.nets().size());
/* Check that things that should have been emptied after the last routing *
* really were. */
if (heap_tail != 1) {
VPR_FATAL_ERROR(VPR_ERROR_ROUTE,
"in init_route_structs. Heap is not empty.\n");
}
}
t_trace*
update_traceback(t_heap* hptr, ClusterNetId net_id) {
/* This routine adds the most recently finished wire segment to the *
* traceback linked list. The first connection starts with the net SOURCE *
* and begins at the structure pointed to by route_ctx.trace[net_id].head. Each *
* connection ends with a SINK. After each SINK, the next connection *
* begins (if the net has more than 2 pins). The first element after the *
* SINK gives the routing node on a previous piece of the routing, which is *
* the link from the existing net to this new piece of the net. *
* In each traceback I start at the end of a path and trace back through *
* its predecessors to the beginning. I have stored information on the *
* predecesser of each node to make traceback easy -- this sacrificies some *
* memory for easier code maintenance. This routine returns a pointer to *
* the first "new" node in the traceback (node not previously in trace). */
auto& route_ctx = g_vpr_ctx.mutable_routing();
auto& trace_nodes = route_ctx.trace_nodes[net_id];
VTR_ASSERT_SAFE(validate_trace_nodes(route_ctx.trace[net_id].head, trace_nodes));
t_trace_branch branch = traceback_branch(hptr->index, trace_nodes);
VTR_ASSERT_SAFE(validate_trace_nodes(branch.head, trace_nodes));
t_trace* ret_ptr = nullptr;
if (route_ctx.trace[net_id].tail != nullptr) {
route_ctx.trace[net_id].tail->next = branch.head; /* Traceback ends with tptr */
ret_ptr = branch.head->next; /* First new segment. */
} else { /* This was the first "chunk" of the net's routing */
route_ctx.trace[net_id].head = branch.head;
ret_ptr = branch.head; /* Whole traceback is new. */
}
route_ctx.trace[net_id].tail = branch.tail;
return (ret_ptr);
}
//Traces back a new routing branch starting from the specified 'node' and working backwards to any existing routing.
//Returns the new branch, and also updates trace_nodes for any new nodes which are included in the branches traceback.
static t_trace_branch traceback_branch(int node, std::unordered_set<int>& trace_nodes) {
auto& device_ctx = g_vpr_ctx.device();
auto& route_ctx = g_vpr_ctx.routing();
auto rr_type = device_ctx.rr_nodes[node].type();
if (rr_type != SINK) {
VPR_FATAL_ERROR(VPR_ERROR_ROUTE,
"in traceback_branch: Expected type = SINK (%d).\n");
}
//We construct the main traceback by walking from the given node back to the source,
//according to the previous edges/nodes recorded in rr_node_route_inf by the router.
t_trace* branch_head = alloc_trace_data();
t_trace* branch_tail = branch_head;
branch_head->index = node;
branch_head->iswitch = OPEN;
branch_head->next = nullptr;
trace_nodes.insert(node);
std::vector<int> new_nodes_added_to_traceback = {node};
auto iedge = route_ctx.rr_node_route_inf[node].prev_edge;
int inode = route_ctx.rr_node_route_inf[node].prev_node;
while (inode != NO_PREVIOUS) {
//Add the current node to the head of traceback
t_trace* prev_ptr = alloc_trace_data();
prev_ptr->index = inode;
prev_ptr->iswitch = device_ctx.rr_nodes[inode].edge_switch(iedge);
prev_ptr->next = branch_head;
branch_head = prev_ptr;
if (trace_nodes.count(inode)) {
break; //Connected to existing routing
}
trace_nodes.insert(inode); //Record this node as visited
new_nodes_added_to_traceback.push_back(inode);
iedge = route_ctx.rr_node_route_inf[inode].prev_edge;
inode = route_ctx.rr_node_route_inf[inode].prev_node;
}
//We next re-expand all the main-branch nodes to add any non-configurably connected side branches
// We are careful to do this *after* the main branch is constructed to ensure nodes which are both
// non-configurably connected *and* part of the main branch are only added to the traceback once.
for (int new_node : new_nodes_added_to_traceback) {
//Expand each main branch node
std::tie(branch_head, branch_tail) = add_trace_non_configurable(branch_head, branch_tail, new_node, trace_nodes);
}
return {branch_head, branch_tail};
}
//Traces any non-configurable subtrees from branch_head, returning the new branch_head and updating trace_nodes
//
//This effectively does a depth-first traversal
static std::pair<t_trace*, t_trace*> add_trace_non_configurable(t_trace* head, t_trace* tail, int node, std::unordered_set<int>& trace_nodes) {
//Trace any non-configurable subtrees
t_trace* subtree_head = nullptr;
t_trace* subtree_tail = nullptr;
std::tie(subtree_head, subtree_tail) = add_trace_non_configurable_recurr(node, trace_nodes);
//Add any non-empty subtree to tail of traceback
if (subtree_head && subtree_tail) {
if (!head) { //First subtree becomes head
head = subtree_head;
} else { //Later subtrees added to tail
VTR_ASSERT(tail);
tail->next = subtree_head;
}
tail = subtree_tail;
} else {
VTR_ASSERT(subtree_head == nullptr && subtree_tail == nullptr);
}
return {head, tail};
}
//Recursive helper function for add_trace_non_configurable()
static std::pair<t_trace*, t_trace*> add_trace_non_configurable_recurr(int node, std::unordered_set<int>& trace_nodes, int depth) {
t_trace* head = nullptr;
t_trace* tail = nullptr;
//Record the non-configurable out-going edges
std::vector<t_edge_size> unvisited_non_configurable_edges;
auto& device_ctx = g_vpr_ctx.device();
for (auto iedge : device_ctx.rr_nodes[node].non_configurable_edges()) {
VTR_ASSERT_SAFE(!device_ctx.rr_nodes[node].edge_is_configurable(iedge));
int to_node = device_ctx.rr_nodes[node].edge_sink_node(iedge);
if (!trace_nodes.count(to_node)) {
unvisited_non_configurable_edges.push_back(iedge);
}
}
if (unvisited_non_configurable_edges.size() == 0) {
//Base case: leaf node with no non-configurable edges
if (depth > 0) { //Arrived via non-configurable edge
VTR_ASSERT(!trace_nodes.count(node));
head = alloc_trace_data();
head->index = node;
head->iswitch = -1;
head->next = nullptr;
tail = head;
trace_nodes.insert(node);
}
} else {
//Recursive case: intermediate node with non-configurable edges
for (auto iedge : unvisited_non_configurable_edges) {
int to_node = device_ctx.rr_nodes[node].edge_sink_node(iedge);
int iswitch = device_ctx.rr_nodes[node].edge_switch(iedge);
VTR_ASSERT(!trace_nodes.count(to_node));
trace_nodes.insert(node);
//Recurse
t_trace* subtree_head = nullptr;
t_trace* subtree_tail = nullptr;
std::tie(subtree_head, subtree_tail) = add_trace_non_configurable_recurr(to_node, trace_nodes, depth + 1);
if (subtree_head && subtree_tail) {
//Add the non-empty sub-tree
//Duplicate the original head as the new tail (for the new branch)
t_trace* intermediate_head = alloc_trace_data();
intermediate_head->index = node;
intermediate_head->iswitch = iswitch;
intermediate_head->next = nullptr;
intermediate_head->next = subtree_head;
if (!head) { //First subtree becomes head
head = intermediate_head;
} else { //Later subtrees added to tail
VTR_ASSERT(tail);
tail->next = intermediate_head;
}
tail = subtree_tail;
} else {
VTR_ASSERT(subtree_head == nullptr && subtree_tail == nullptr);
}
}
}
return {head, tail};
}
/* The routine sets the path_cost to HUGE_POSITIVE_FLOAT for *
* all channel segments touched by previous routing phases. */
void reset_path_costs(const std::vector<int>& visited_rr_nodes) {
auto& route_ctx = g_vpr_ctx.mutable_routing();
for (auto node : visited_rr_nodes) {
route_ctx.rr_node_route_inf[node].path_cost = std::numeric_limits<float>::infinity();
route_ctx.rr_node_route_inf[node].backward_path_cost = std::numeric_limits<float>::infinity();
route_ctx.rr_node_route_inf[node].prev_node = NO_PREVIOUS;
route_ctx.rr_node_route_inf[node].prev_edge = NO_PREVIOUS;
}
}
/* Returns the *congestion* cost of using this rr_node. */
float get_rr_cong_cost(int inode) {
auto& device_ctx = g_vpr_ctx.device();
float cost = get_single_rr_cong_cost(inode);
auto itr = device_ctx.rr_node_to_non_config_node_set.find(inode);
if (itr != device_ctx.rr_node_to_non_config_node_set.end()) {
for (int node : device_ctx.rr_non_config_node_sets[itr->second]) {
if (node == inode) {
continue; //Already included above
}
cost += get_single_rr_cong_cost(node);
}
}
return (cost);
}
/* Returns the congestion cost of using this rr_node, *ignoring*
* non-configurable edges */
static float get_single_rr_cong_cost(int inode) {
auto& device_ctx = g_vpr_ctx.device();
auto& route_ctx = g_vpr_ctx.routing();
auto cost_index = device_ctx.rr_nodes[inode].cost_index();
float cost = device_ctx.rr_indexed_data[cost_index].base_cost
* route_ctx.rr_node_route_inf[inode].acc_cost
* route_ctx.rr_node_route_inf[inode].pres_cost;
return cost;
}
/* Mark all the SINKs of this net as targets by setting their target flags *
* to the number of times the net must connect to each SINK. Note that *
* this number can occasionally be greater than 1 -- think of connecting *
* the same net to two inputs of an and-gate (and-gate inputs are logically *
* equivalent, so both will connect to the same SINK). */
void mark_ends(ClusterNetId net_id) {
unsigned int ipin;
int inode;
auto& cluster_ctx = g_vpr_ctx.clustering();
auto& route_ctx = g_vpr_ctx.mutable_routing();
for (ipin = 1; ipin < cluster_ctx.clb_nlist.net_pins(net_id).size(); ipin++) {
inode = route_ctx.net_rr_terminals[net_id][ipin];
route_ctx.rr_node_route_inf[inode].target_flag++;
}
}
void mark_remaining_ends(const std::vector<int>& remaining_sinks) {
// like mark_ends, but only performs it for the remaining sinks of a net
auto& route_ctx = g_vpr_ctx.mutable_routing();
for (int sink_node : remaining_sinks)
++route_ctx.rr_node_route_inf[sink_node].target_flag;
}
void node_to_heap(int inode, float total_cost, int prev_node, int prev_edge, float backward_path_cost, float R_upstream) {
/* Puts an rr_node on the heap, if the new cost given is lower than the *
* current path_cost to this channel segment. The index of its predecessor *
* is stored to make traceback easy. The index of the edge used to get *
* from its predecessor to it is also stored to make timing analysis, etc. *
* easy. The backward_path_cost and R_upstream values are used only by the *
* timing-driven router -- the breadth-first router ignores them. */
auto& route_ctx = g_vpr_ctx.routing();
if (total_cost >= route_ctx.rr_node_route_inf[inode].path_cost)
return;
t_heap* hptr = alloc_heap_data();
hptr->index = inode;
hptr->cost = total_cost;
VTR_ASSERT_SAFE(hptr->u.prev.node == NO_PREVIOUS);
VTR_ASSERT_SAFE(hptr->u.prev.edge == NO_PREVIOUS);
hptr->u.prev.node = prev_node;
hptr->u.prev.edge = prev_edge;
hptr->backward_path_cost = backward_path_cost;
hptr->R_upstream = R_upstream;
add_to_heap(hptr);
}
void free_traceback(ClusterNetId net_id) {
/* Puts the entire traceback (old routing) for this net on the free list *
* and sets the route_ctx.trace_head pointers etc. for the net to NULL. */
auto& route_ctx = g_vpr_ctx.mutable_routing();
if (route_ctx.trace.empty()) {
return;
}
if (route_ctx.trace[net_id].head == nullptr) {
return;
}
free_traceback(route_ctx.trace[net_id].head);
route_ctx.trace[net_id].head = nullptr;
route_ctx.trace[net_id].tail = nullptr;
route_ctx.trace_nodes[net_id].clear();
}
void free_traceback(t_trace* tptr) {
while (tptr != nullptr) {
t_trace* tempptr = tptr->next;
free_trace_data(tptr);
tptr = tempptr;
}
}
/* Allocates data structures into which the key routing data can be saved, *
* allowing the routing to be recovered later (e.g. after a another routing *
* is attempted). */
vtr::vector<ClusterNetId, t_trace*> alloc_saved_routing() {
auto& cluster_ctx = g_vpr_ctx.clustering();
vtr::vector<ClusterNetId, t_trace*> best_routing(cluster_ctx.clb_nlist.nets().size());
return (best_routing);
}
//Calculates how many (and allocates space for) OPINs which must be reserved to
//respect 'instance' equivalence.
//
//TODO: At the moment this makes a significant simplifying assumption. It reserves
// all OPINs not connected to external nets. This ensures each signal leaving
// the logic block uses only a single OPIN. This is a safe, but pessmistic
// behaviour, which prevents any logic duplication (e.g. duplicating LUTs/BLEs)
// to drive multiple OPINs which could aid routability.
//
// Note that this only applies for 'intance' equivalence. 'full' equivalence
// does allow for multiple outputs to be used for a single signal.
static t_clb_opins_used alloc_and_load_clb_opins_used_locally() {
/* Allocates and loads the data needed to make the router reserve some CLB *
* output pins for connections made locally within a CLB (if the netlist *
* specifies that this is necessary). */
t_clb_opins_used clb_opins_used_locally;
int clb_pin, iclass, class_low, class_high;
auto& cluster_ctx = g_vpr_ctx.clustering();
clb_opins_used_locally.resize(cluster_ctx.clb_nlist.blocks().size());
for (auto blk_id : cluster_ctx.clb_nlist.blocks()) {
auto type = physical_tile_type(blk_id);
get_class_range_for_block(blk_id, &class_low, &class_high);
clb_opins_used_locally[blk_id].resize(type->num_class);
if (is_io_type(type)) continue;
int pin_low = 0;
int pin_high = 0;
get_pin_range_for_block(blk_id, &pin_low, &pin_high);
for (clb_pin = pin_low; clb_pin <= pin_high; clb_pin++) {
auto net = cluster_ctx.clb_nlist.block_net(blk_id, clb_pin);
if (!net || (net && cluster_ctx.clb_nlist.net_sinks(net).size() == 0)) {
//There is no external net connected to this pin
iclass = type->pin_class[clb_pin];
if (type->class_inf[iclass].equivalence == PortEquivalence::INSTANCE) {
//The pin is part of an instance equivalent class, hence we need to reserve a pin
VTR_ASSERT(type->class_inf[iclass].type == DRIVER);
/* Check to make sure class is in same range as that assigned to block */
VTR_ASSERT(iclass >= class_low && iclass <= class_high);
//We push back OPEN to reserve space to store the exact pin which
//will be reserved (determined later)
clb_opins_used_locally[blk_id][iclass].emplace_back(OPEN);
}
}
}
}
return (clb_opins_used_locally);
}
/*the trace lists are only freed after use by the timing-driven placer */
/*Do not free them after use by the router, since stats, and draw */
/*routines use the trace values */
void free_trace_structs() {
auto& cluster_ctx = g_vpr_ctx.clustering();
auto& route_ctx = g_vpr_ctx.mutable_routing();
if (route_ctx.trace.empty()) {
return;
}
for (auto net_id : cluster_ctx.clb_nlist.nets()) {
free_traceback(net_id);
if (route_ctx.trace[net_id].head) {
free(route_ctx.trace[net_id].head);
free(route_ctx.trace[net_id].tail);
}
route_ctx.trace[net_id].head = nullptr;
route_ctx.trace[net_id].tail = nullptr;
}
}
void free_route_structs() {
/* Frees the temporary storage needed only during the routing. The *
* final routing result is not freed. */
auto& route_ctx = g_vpr_ctx.mutable_routing();
if (heap != nullptr) {
//Free the individiaul heap elements (calls destructors)
for (int i = 1; i < num_heap_allocated; i++) {
VTR_LOG("Freeing %p\n", heap[i]);
vtr::chunk_delete(heap[i], &heap_ch);
}
// coverity[offset_free : Intentional]
free(heap + 1);
heap = nullptr; /* Defensive coding: crash hard if I use these. */
}
if (heap_free_head != nullptr) {
t_heap* curr = heap_free_head;
while (curr) {
t_heap* tmp = curr;
curr = curr->u.next;
vtr::chunk_delete(tmp, &heap_ch);
}
heap_free_head = nullptr;
}
if (route_ctx.route_bb.size() != 0) {
route_ctx.route_bb.clear();
}
/*free the memory chunks that were used by heap and linked f pointer */
free_chunk_memory(&heap_ch);
}
/* Frees the data structures needed to save a routing. */
void free_saved_routing(vtr::vector<ClusterNetId, t_trace*>& best_routing) {
auto& cluster_ctx = g_vpr_ctx.clustering();
for (auto net_id : cluster_ctx.clb_nlist.nets()) {
if (best_routing[net_id] != nullptr) {
free(best_routing[net_id]);
best_routing[net_id] = nullptr;
}
}
}
void alloc_and_load_rr_node_route_structs() {
/* Allocates some extra information about each rr_node that is used only *
* during routing. */
auto& route_ctx = g_vpr_ctx.mutable_routing();
auto& device_ctx = g_vpr_ctx.device();
route_ctx.rr_node_route_inf.resize(device_ctx.rr_nodes.size());
reset_rr_node_route_structs();
}
void reset_rr_node_route_structs() {
/* Resets some extra information about each rr_node that is used only *
* during routing. */
auto& route_ctx = g_vpr_ctx.mutable_routing();
auto& device_ctx = g_vpr_ctx.device();
VTR_ASSERT(route_ctx.rr_node_route_inf.size() == size_t(device_ctx.rr_nodes.size()));
for (size_t inode = 0; inode < device_ctx.rr_nodes.size(); inode++) {
route_ctx.rr_node_route_inf[inode].prev_node = NO_PREVIOUS;
route_ctx.rr_node_route_inf[inode].prev_edge = NO_PREVIOUS;
route_ctx.rr_node_route_inf[inode].pres_cost = 1.0;
route_ctx.rr_node_route_inf[inode].acc_cost = 1.0;
route_ctx.rr_node_route_inf[inode].path_cost = std::numeric_limits<float>::infinity();
route_ctx.rr_node_route_inf[inode].backward_path_cost = std::numeric_limits<float>::infinity();
route_ctx.rr_node_route_inf[inode].target_flag = 0;
route_ctx.rr_node_route_inf[inode].set_occ(0);
}
}
/* Allocates and loads the route_ctx.net_rr_terminals data structure. For each net it stores the rr_node *
* index of the SOURCE of the net and all the SINKs of the net [clb_nlist.nets()][clb_nlist.net_pins()]. *
* Entry [inet][pnum] stores the rr index corresponding to the SOURCE (opin) or SINK (ipin) of the pin. */
static vtr::vector<ClusterNetId, std::vector<int>> load_net_rr_terminals(const t_rr_node_indices& L_rr_node_indices) {
vtr::vector<ClusterNetId, std::vector<int>> net_rr_terminals;
int inode, i, j, node_block_pin, iclass;
auto& cluster_ctx = g_vpr_ctx.clustering();
auto& place_ctx = g_vpr_ctx.placement();
auto nets = cluster_ctx.clb_nlist.nets();
net_rr_terminals.resize(nets.size());
for (auto net_id : cluster_ctx.clb_nlist.nets()) {
auto net_pins = cluster_ctx.clb_nlist.net_pins(net_id);
net_rr_terminals[net_id].resize(net_pins.size());
int pin_count = 0;
for (auto pin_id : cluster_ctx.clb_nlist.net_pins(net_id)) {
auto block_id = cluster_ctx.clb_nlist.pin_block(pin_id);
i = place_ctx.block_locs[block_id].loc.x;
j = place_ctx.block_locs[block_id].loc.y;
auto type = physical_tile_type(block_id);
/* In the routing graph, each (x, y) location has unique pins on it
* so when there is capacity, blocks are packed and their pin numbers
* are offset to get their actual rr_node */
node_block_pin = cluster_ctx.clb_nlist.pin_physical_index(pin_id);
iclass = type->pin_class[node_block_pin];
inode = get_rr_node_index(L_rr_node_indices, i, j, (pin_count == 0 ? SOURCE : SINK), /* First pin is driver */
iclass);
net_rr_terminals[net_id][pin_count] = inode;
pin_count++;
}
}
return net_rr_terminals;
}
/* Saves the rr_node corresponding to each SOURCE and SINK in each CLB *
* in the FPGA. Currently only the SOURCE rr_node values are used, and *
* they are used only to reserve pins for locally used OPINs in the router. *
* [0..cluster_ctx.clb_nlist.blocks().size()-1][0..num_class-1]. *
* The values for blocks that are padsare NOT valid. */
static vtr::vector<ClusterBlockId, std::vector<int>> load_rr_clb_sources(const t_rr_node_indices& L_rr_node_indices) {
vtr::vector<ClusterBlockId, std::vector<int>> rr_blk_source;
int i, j, iclass, inode;
int class_low, class_high;
t_rr_type rr_type;
auto& cluster_ctx = g_vpr_ctx.clustering();
auto& place_ctx = g_vpr_ctx.placement();
rr_blk_source.resize(cluster_ctx.clb_nlist.blocks().size());
for (auto blk_id : cluster_ctx.clb_nlist.blocks()) {
auto type = physical_tile_type(blk_id);
get_class_range_for_block(blk_id, &class_low, &class_high);
rr_blk_source[blk_id].resize(type->num_class);
for (iclass = 0; iclass < type->num_class; iclass++) {
if (iclass >= class_low && iclass <= class_high) {
i = place_ctx.block_locs[blk_id].loc.x;
j = place_ctx.block_locs[blk_id].loc.y;
if (type->class_inf[iclass].type == DRIVER)
rr_type = SOURCE;
else
rr_type = SINK;
inode = get_rr_node_index(L_rr_node_indices, i, j, rr_type, iclass);
rr_blk_source[blk_id][iclass] = inode;
} else {
rr_blk_source[blk_id][iclass] = OPEN;
}
}
}
return rr_blk_source;
}
vtr::vector<ClusterNetId, t_bb> load_route_bb(int bb_factor) {
vtr::vector<ClusterNetId, t_bb> route_bb;
auto& cluster_ctx = g_vpr_ctx.clustering();
auto nets = cluster_ctx.clb_nlist.nets();
route_bb.resize(nets.size());
for (auto net_id : nets) {
route_bb[net_id] = load_net_route_bb(net_id, bb_factor);
}
return route_bb;
}
t_bb load_net_route_bb(ClusterNetId net_id, int bb_factor) {
/*
* This routine loads the bounding box used to limit the space
* searched by the maze router when routing a specific net. The search is
* limited to channels contained with the net bounding box expanded
* by bb_factor channels on each side. For example, if bb_factor is
* 0, the maze router must route each net within its bounding box.
* If bb_factor = max(device_ctx.grid.width()-1, device_cts.grid.height() - 1),
* the maze router will search every channel in
* the FPGA if necessary. The bounding box returned by this routine
* are different from the ones used by the placer in that they are
* clipped to lie within (0,0) and (device_ctx.grid.width()-1,device_ctx.grid.height()-1)
* rather than (1,1) and (device_ctx.grid.width()-1,device_ctx.grid.height()-1).
*/
auto& cluster_ctx = g_vpr_ctx.clustering();
auto& device_ctx = g_vpr_ctx.device();
auto& route_ctx = g_vpr_ctx.routing();
//Ensure bb_factor is bounded by the device size
//This ensures that if someone passes in an extremely large bb_factor
//(e.g. std::numeric_limits<int>::max()) the later addition/subtraction
//of bb_factor will not cause under/overflow
int max_dim = std::max<int>(device_ctx.grid.width() - 1, device_ctx.grid.height() - 1);
bb_factor = std::min(bb_factor, max_dim);
int driver_rr = route_ctx.net_rr_terminals[net_id][0];
const t_rr_node& source_node = device_ctx.rr_nodes[driver_rr];
VTR_ASSERT(source_node.type() == SOURCE);
VTR_ASSERT(source_node.xlow() <= source_node.xhigh());
VTR_ASSERT(source_node.ylow() <= source_node.yhigh());
int xmin = source_node.xlow();
int ymin = source_node.ylow();
int xmax = source_node.xhigh();
int ymax = source_node.yhigh();
auto net_sinks = cluster_ctx.clb_nlist.net_sinks(net_id);
for (size_t ipin = 1; ipin < net_sinks.size() + 1; ++ipin) { //Start at 1 since looping through sinks
int sink_rr = route_ctx.net_rr_terminals[net_id][ipin];
const t_rr_node& sink_node = device_ctx.rr_nodes[sink_rr];
VTR_ASSERT(sink_node.type() == SINK);
VTR_ASSERT(sink_node.xlow() <= sink_node.xhigh());
VTR_ASSERT(sink_node.ylow() <= sink_node.yhigh());
xmin = std::min<int>(xmin, sink_node.xlow());
xmax = std::max<int>(xmax, sink_node.xhigh());
ymin = std::min<int>(ymin, sink_node.ylow());
ymax = std::max<int>(ymax, sink_node.yhigh());
}
/* Want the channels on all 4 sides to be usuable, even if bb_factor = 0. */
xmin -= 1;
ymin -= 1;
/* Expand the net bounding box by bb_factor, then clip to the physical *
* chip area. */
t_bb bb;
bb.xmin = std::max<int>(xmin - bb_factor, 0);
bb.xmax = std::min<int>(xmax + bb_factor, device_ctx.grid.width() - 1);
bb.ymin = std::max<int>(ymin - bb_factor, 0);
bb.ymax = std::min<int>(ymax + bb_factor, device_ctx.grid.height() - 1);
return bb;
}
void add_to_mod_list(int inode, std::vector<int>& modified_rr_node_inf) {
auto& route_ctx = g_vpr_ctx.routing();
if (std::isinf(route_ctx.rr_node_route_inf[inode].path_cost)) {
modified_rr_node_inf.push_back(inode);
}
}
namespace heap_ {
size_t parent(size_t i);
size_t left(size_t i);
size_t right(size_t i);
size_t size();
void expand_heap_if_full();
size_t parent(size_t i) { return i >> 1; }
// child indices of a heap
size_t left(size_t i) { return i << 1; }
size_t right(size_t i) { return (i << 1) + 1; }
size_t size() { return static_cast<size_t>(heap_tail - 1); } // heap[0] is not valid element
// make a heap rooted at index i by **sifting down** in O(lgn) time
void sift_down(size_t hole) {
t_heap* head{heap[hole]};
size_t child{left(hole)};
while ((int)child < heap_tail) {
if ((int)child + 1 < heap_tail && heap[child + 1]->cost < heap[child]->cost)
++child;
if (heap[child]->cost < head->cost) {
heap[hole] = heap[child];
hole = child;
child = left(child);
} else
break;
}
heap[hole] = head;
}
// runs in O(n) time by sifting down; the least work is done on the most elements: 1 swap for bottom layer, 2 swap for 2nd, ... lgn swap for top
// 1*(n/2) + 2*(n/4) + 3*(n/8) + ... + lgn*1 = 2n (sum of i/2^i)
void build_heap() {
// second half of heap are leaves
for (size_t i = heap_tail >> 1; i != 0; --i)
sift_down(i);
}
// O(lgn) sifting up to maintain heap property after insertion (should sift down when building heap)
void sift_up(size_t leaf, t_heap* const node) {
while ((leaf > 1) && (node->cost < heap[parent(leaf)]->cost)) {
// sift hole up
heap[leaf] = heap[parent(leaf)];
leaf = parent(leaf);
}
heap[leaf] = node;
}
void expand_heap_if_full() {
if (heap_tail > heap_size) { /* Heap is full */
heap_size *= 2;
heap = (t_heap**)vtr::realloc((void*)(heap + 1),
heap_size * sizeof(t_heap*));
heap--; /* heap goes from [1..heap_size] */
}
}
// adds an element to the back of heap and expand if necessary, but does not maintain heap property
void push_back(t_heap* const hptr) {
expand_heap_if_full();
heap[heap_tail] = hptr;
++heap_tail;
}
void push_back_node(int inode, float total_cost, int prev_node, int prev_edge, float backward_path_cost, float R_upstream) {
/* Puts an rr_node on the heap with the same condition as node_to_heap,
* but do not fix heap property yet as that is more efficiently done from
* bottom up with build_heap */
auto& route_ctx = g_vpr_ctx.routing();
if (total_cost >= route_ctx.rr_node_route_inf[inode].path_cost)
return;
t_heap* hptr = alloc_heap_data();
hptr->index = inode;
hptr->cost = total_cost;
hptr->u.prev.node = prev_node;
hptr->u.prev.edge = prev_edge;
hptr->backward_path_cost = backward_path_cost;
hptr->R_upstream = R_upstream;
push_back(hptr);
}
bool is_valid() {
for (size_t i = 1; (int)i <= heap_tail >> 1; ++i) {
if ((int)left(i) < heap_tail && heap[left(i)]->cost < heap[i]->cost) return false;
if ((int)right(i) < heap_tail && heap[right(i)]->cost < heap[i]->cost) return false;
}
return true;
}
// extract every element and print it
void pop_heap() {
while (!is_empty_heap())
VTR_LOG("%e ", get_heap_head()->cost);
VTR_LOG("\n");
}
// print every element; not necessarily in order for minheap
void print_heap() {
for (int i = 1; i<heap_tail>> 1; ++i)
VTR_LOG("(%e %e %e) ", heap[i]->cost, heap[left(i)]->cost, heap[right(i)]->cost);
VTR_LOG("\n");
}
// verify correctness of extract top by making a copy, sorting it, and iterating it at the same time as extraction
void verify_extract_top() {
constexpr float float_epsilon = 1e-20;
std::cout << "copying heap\n";
std::vector<t_heap*> heap_copy{heap + 1, heap + heap_tail};
// sort based on cost with cheapest first
VTR_ASSERT(heap_copy.size() == size());
std::sort(begin(heap_copy), end(heap_copy),
[](const t_heap* a, const t_heap* b) {
return a->cost < b->cost;
});
std::cout << "starting to compare top elements\n";
size_t i = 0;
while (!is_empty_heap()) {
while (heap_copy[i]->index == OPEN)
++i; // skip the ones that won't be extracted
auto top = get_heap_head();
if (abs(top->cost - heap_copy[i]->cost) > float_epsilon)
std::cout << "mismatch with sorted " << top << '(' << top->cost << ") " << heap_copy[i] << '(' << heap_copy[i]->cost << ")\n";
++i;
}
if (i != heap_copy.size())
std::cout << "did not finish extracting: " << i << " vs " << heap_copy.size() << std::endl;
else
std::cout << "extract top working as intended\n";
}
} // namespace heap_
// adds to heap and maintains heap quality
void add_to_heap(t_heap* hptr) {
heap_::expand_heap_if_full();
// start with undefined hole
++heap_tail;
heap_::sift_up(heap_tail - 1, hptr);
}
/*WMF: peeking accessor :) */
bool is_empty_heap() {
return (bool)(heap_tail == 1);
}
t_heap*
get_heap_head() {
/* Returns a pointer to the smallest element on the heap, or NULL if the *
* heap is empty. Invalid (index == OPEN) entries on the heap are never *
* returned -- they are just skipped over. */
t_heap* cheapest;
size_t hole, child;
do {
if (heap_tail == 1) { /* Empty heap. */
VTR_LOG_WARN("Empty heap occurred in get_heap_head.\n");
return (nullptr);
}
cheapest = heap[1];
hole = 1;
child = 2;
--heap_tail;
while ((int)child < heap_tail) {
if (heap[child + 1]->cost < heap[child]->cost)
++child; // become right child
heap[hole] = heap[child];
hole = child;
child = heap_::left(child);
}
heap_::sift_up(hole, heap[heap_tail]);
} while (cheapest->index == OPEN); /* Get another one if invalid entry. */
return (cheapest);
}
void empty_heap() {
for (int i = 1; i < heap_tail; i++)
free_heap_data(heap[i]);
heap_tail = 1;
}
t_heap*
alloc_heap_data() {
if (heap_free_head == nullptr) { /* No elements on the free list */
heap_free_head = vtr::chunk_new<t_heap>(&heap_ch);
}
//Extract the head
t_heap* temp_ptr = heap_free_head;
heap_free_head = heap_free_head->u.next;
num_heap_allocated++;
//Reset
temp_ptr->u.next = nullptr;
temp_ptr->cost = 0.;
temp_ptr->backward_path_cost = 0.;
temp_ptr->R_upstream = 0.;
temp_ptr->index = OPEN;
temp_ptr->u.prev.node = NO_PREVIOUS;
temp_ptr->u.prev.edge = NO_PREVIOUS;
return (temp_ptr);
}
void free_heap_data(t_heap* hptr) {
hptr->u.next = heap_free_head;
heap_free_head = hptr;
num_heap_allocated--;
}
void invalidate_heap_entries(int sink_node, int ipin_node) {
/* Marks all the heap entries consisting of sink_node, where it was reached *
* via ipin_node, as invalid (OPEN). Used only by the breadth_first router *
* and even then only in rare circumstances. */
for (int i = 1; i < heap_tail; i++) {
if (heap[i]->index == sink_node) {
if (heap[i]->u.prev.node == ipin_node) {
heap[i]->index = OPEN; /* Invalid. */
break;
}
}
}
}
t_trace*
alloc_trace_data() {
t_trace* temp_ptr;
if (trace_free_head == nullptr) { /* No elements on the free list */
trace_free_head = (t_trace*)vtr::chunk_malloc(sizeof(t_trace), &trace_ch);
trace_free_head->next = nullptr;
}
temp_ptr = trace_free_head;
trace_free_head = trace_free_head->next;
num_trace_allocated++;
return (temp_ptr);
}
void free_trace_data(t_trace* tptr) {
/* Puts the traceback structure pointed to by tptr on the free list. */
tptr->next = trace_free_head;
trace_free_head = tptr;
num_trace_allocated--;
}
void print_route(FILE* fp, const vtr::vector<ClusterNetId, t_traceback>& tracebacks) {
if (tracebacks.empty()) return; //Only if routing exists
auto& place_ctx = g_vpr_ctx.placement();
auto& device_ctx = g_vpr_ctx.device();
auto& cluster_ctx = g_vpr_ctx.clustering();
auto& route_ctx = g_vpr_ctx.mutable_routing();
for (auto net_id : cluster_ctx.clb_nlist.nets()) {
if (!cluster_ctx.clb_nlist.net_is_ignored(net_id)) {
fprintf(fp, "\n\nNet %zu (%s)\n\n", size_t(net_id), cluster_ctx.clb_nlist.net_name(net_id).c_str());
if (cluster_ctx.clb_nlist.net_sinks(net_id).size() == false) {
fprintf(fp, "\n\nUsed in local cluster only, reserved one CLB pin\n\n");
} else {
t_trace* tptr = route_ctx.trace[net_id].head;
while (tptr != nullptr) {
int inode = tptr->index;
t_rr_type rr_type = device_ctx.rr_nodes[inode].type();
int ilow = device_ctx.rr_nodes[inode].xlow();
int jlow = device_ctx.rr_nodes[inode].ylow();
fprintf(fp, "Node:\t%d\t%6s (%d,%d) ", inode,
device_ctx.rr_nodes[inode].type_string(), ilow, jlow);
if ((ilow != device_ctx.rr_nodes[inode].xhigh())
|| (jlow != device_ctx.rr_nodes[inode].yhigh()))
fprintf(fp, "to (%d,%d) ", device_ctx.rr_nodes[inode].xhigh(),
device_ctx.rr_nodes[inode].yhigh());
switch (rr_type) {
case IPIN:
case OPIN:
if (is_io_type(device_ctx.grid[ilow][jlow].type)) {
fprintf(fp, " Pad: ");
} else { /* IO Pad. */
fprintf(fp, " Pin: ");
}
break;
case CHANX:
case CHANY:
fprintf(fp, " Track: ");
break;
case SOURCE:
case SINK:
if (is_io_type(device_ctx.grid[ilow][jlow].type)) {
fprintf(fp, " Pad: ");
} else { /* IO Pad. */
fprintf(fp, " Class: ");
}
break;
default:
VPR_FATAL_ERROR(VPR_ERROR_ROUTE,
"in print_route: Unexpected traceback element type: %d (%s).\n",
rr_type, device_ctx.rr_nodes[inode].type_string());
break;
}
fprintf(fp, "%d ", device_ctx.rr_nodes[inode].ptc_num());
if (!is_io_type(device_ctx.grid[ilow][jlow].type) && (rr_type == IPIN || rr_type == OPIN)) {
int pin_num = device_ctx.rr_nodes[inode].ptc_num();
int xoffset = device_ctx.grid[ilow][jlow].width_offset;
int yoffset = device_ctx.grid[ilow][jlow].height_offset;
ClusterBlockId iblock = place_ctx.grid_blocks[ilow - xoffset][jlow - yoffset].blocks[0];
VTR_ASSERT(iblock);
t_pb_graph_pin* pb_pin = get_pb_graph_node_pin_from_block_pin(iblock, pin_num);
t_pb_type* pb_type = pb_pin->parent_node->pb_type;
fprintf(fp, " %s.%s[%d] ", pb_type->name, pb_pin->port->name, pb_pin->pin_number);
}
/* Uncomment line below if you're debugging and want to see the switch types *
* used in the routing. */
fprintf(fp, "Switch: %d", tptr->iswitch);
fprintf(fp, "\n");
tptr = tptr->next;
}
}
} else { /* Global net. Never routed. */
fprintf(fp, "\n\nNet %zu (%s): global net connecting:\n\n", size_t(net_id),
cluster_ctx.clb_nlist.net_name(net_id).c_str());
for (auto pin_id : cluster_ctx.clb_nlist.net_pins(net_id)) {
ClusterBlockId block_id = cluster_ctx.clb_nlist.pin_block(pin_id);
int pin_index = cluster_ctx.clb_nlist.pin_physical_index(pin_id);
int iclass = physical_tile_type(block_id)->pin_class[pin_index];
fprintf(fp, "Block %s (#%zu) at (%d,%d), Pin class %d.\n",
cluster_ctx.clb_nlist.block_name(block_id).c_str(), size_t(block_id),
place_ctx.block_locs[block_id].loc.x,
place_ctx.block_locs[block_id].loc.y,
iclass);
}
}
}
}
/* Prints out the routing to file route_file. */
void print_route(const char* placement_file, const char* route_file) {
FILE* fp;
fp = fopen(route_file, "w");
auto& place_ctx = g_vpr_ctx.placement();
auto& device_ctx = g_vpr_ctx.device();
auto& route_ctx = g_vpr_ctx.mutable_routing();
fprintf(fp, "Placement_File: %s Placement_ID: %s\n", placement_file, place_ctx.placement_id.c_str());
fprintf(fp, "Array size: %zu x %zu logic blocks.\n", device_ctx.grid.width(), device_ctx.grid.height());
fprintf(fp, "\nRouting:");
print_route(fp, route_ctx.trace);
fclose(fp);
if (getEchoEnabled() && isEchoFileEnabled(E_ECHO_MEM)) {
fp = vtr::fopen(getEchoFileName(E_ECHO_MEM), "w");
fprintf(fp, "\nNum_heap_allocated: %d Num_trace_allocated: %d\n",
num_heap_allocated, num_trace_allocated);
fprintf(fp, "Num_linked_f_pointer_allocated: %d\n",
num_linked_f_pointer_allocated);
fclose(fp);
}
//Save the digest of the route file
route_ctx.routing_id = vtr::secure_digest_file(route_file);
}
//To ensure the router can only swaps pin which are actually logically equivalent some block output pins must be
//reserved in certain cases.
//
// In the RR graph, output pin equivalence is modelled by a single SRC node connected to (multiple) OPINs, modelling
// that each of the OPINs is logcially equivalent (i.e. it doesn't matter through which the router routes a signal,
// so long as it is from the appropriate SRC).
//
// This correctly models 'full' equivalence (e.g. if there is a full crossbar between the outputs), but is to
// optimistic for 'instance' equivalence (which typcially models the pin equivalence possible by swapping sub-block
// instances like BLEs). In particular, for the 'instance' equivalence case, some of the 'equivalent' block outputs
// may be used by internal signals which are routed entirely *within* the block (i.e. the signals which never leave
// the block). These signals effectively 'use-up' an output pin which should now be unavailable to the router.
//
// To model this we 'reserve' these locally used outputs, ensuring that the router will not use them (as if it did
// this would equate to duplicating a BLE into an already in-use BLE instance, which is clearly incorrect).
void reserve_locally_used_opins(float pres_fac, float acc_fac, bool rip_up_local_opins) {
int num_local_opin, inode, from_node, iconn, num_edges, to_node;
int iclass, ipin;
float cost;
t_heap* heap_head_ptr;
t_physical_tile_type_ptr type;
auto& cluster_ctx = g_vpr_ctx.clustering();
auto& route_ctx = g_vpr_ctx.mutable_routing();
auto& device_ctx = g_vpr_ctx.device();
if (rip_up_local_opins) {
for (auto blk_id : cluster_ctx.clb_nlist.blocks()) {
type = physical_tile_type(blk_id);
for (iclass = 0; iclass < type->num_class; iclass++) {
num_local_opin = route_ctx.clb_opins_used_locally[blk_id][iclass].size();
if (num_local_opin == 0) continue;
VTR_ASSERT(type->class_inf[iclass].equivalence == PortEquivalence::INSTANCE);
/* Always 0 for pads and for RECEIVER (IPIN) classes */
for (ipin = 0; ipin < num_local_opin; ipin++) {
inode = route_ctx.clb_opins_used_locally[blk_id][iclass][ipin];
adjust_one_rr_occ_and_apcost(inode, -1, pres_fac, acc_fac);
}
}
}
}
for (auto blk_id : cluster_ctx.clb_nlist.blocks()) {
type = physical_tile_type(blk_id);
for (iclass = 0; iclass < type->num_class; iclass++) {
num_local_opin = route_ctx.clb_opins_used_locally[blk_id][iclass].size();
if (num_local_opin == 0) continue;
VTR_ASSERT(type->class_inf[iclass].equivalence == PortEquivalence::INSTANCE);
//From the SRC node we walk through it's out going edges to collect the
//OPIN nodes. We then push them onto a heap so the the OPINs with lower
//congestion cost are popped-off/reserved first. (Intuitively, we want
//the reserved OPINs to move out of the way of congestion, by preferring
//to reserve OPINs with lower congestion costs).
from_node = route_ctx.rr_blk_source[blk_id][iclass];
num_edges = device_ctx.rr_nodes[from_node].num_edges();
for (iconn = 0; iconn < num_edges; iconn++) {
to_node = device_ctx.rr_nodes[from_node].edge_sink_node(iconn);
VTR_ASSERT(device_ctx.rr_nodes[to_node].type() == OPIN);
//Add the OPIN to the heap according to it's congestion cost
cost = get_rr_cong_cost(to_node);
node_to_heap(to_node, cost, OPEN, OPEN, 0., 0.);
}
for (ipin = 0; ipin < num_local_opin; ipin++) {
//Pop the nodes off the heap. We get them from the heap so we
//reserve those pins with lowest congestion cost first.
heap_head_ptr = get_heap_head();
inode = heap_head_ptr->index;
VTR_ASSERT(device_ctx.rr_nodes[inode].type() == OPIN);
adjust_one_rr_occ_and_apcost(inode, 1, pres_fac, acc_fac);
route_ctx.clb_opins_used_locally[blk_id][iclass][ipin] = inode;
free_heap_data(heap_head_ptr);
}
empty_heap();
}
}
}
static void adjust_one_rr_occ_and_apcost(int inode, int add_or_sub, float pres_fac, float acc_fac) {
/* Increments or decrements (depending on add_or_sub) the occupancy of *
* one rr_node, and adjusts the present cost of that node appropriately. */
auto& route_ctx = g_vpr_ctx.mutable_routing();
auto& device_ctx = g_vpr_ctx.device();
int new_occ = route_ctx.rr_node_route_inf[inode].occ() + add_or_sub;
int capacity = device_ctx.rr_nodes[inode].capacity();
route_ctx.rr_node_route_inf[inode].set_occ(new_occ);
if (new_occ < capacity) {
route_ctx.rr_node_route_inf[inode].pres_cost = 1.0;
} else {
route_ctx.rr_node_route_inf[inode].pres_cost = 1.0 + (new_occ + 1 - capacity) * pres_fac;
if (add_or_sub == 1) {
route_ctx.rr_node_route_inf[inode].acc_cost += (new_occ - capacity) * acc_fac;
}
}
}
void free_chunk_memory_trace() {
if (trace_ch.chunk_ptr_head != nullptr) {
free_chunk_memory(&trace_ch);
trace_ch.chunk_ptr_head = nullptr;
trace_free_head = nullptr;
}
}
// connection based overhaul (more specificity than nets)
// utility and debugging functions -----------------------
void print_traceback(ClusterNetId net_id) {
// linearly print linked list
auto& route_ctx = g_vpr_ctx.routing();
auto& device_ctx = g_vpr_ctx.device();
VTR_LOG("traceback %zu: ", size_t(net_id));
t_trace* head = route_ctx.trace[net_id].head;
while (head) {
int inode{head->index};
if (device_ctx.rr_nodes[inode].type() == SINK)
VTR_LOG("%d(sink)(%d)->", inode, route_ctx.rr_node_route_inf[inode].occ());
else
VTR_LOG("%d(%d)->", inode, route_ctx.rr_node_route_inf[inode].occ());
head = head->next;
}
VTR_LOG("\n");
}
void print_traceback(const t_trace* trace) {
auto& device_ctx = g_vpr_ctx.device();
auto& route_ctx = g_vpr_ctx.routing();
const t_trace* prev = nullptr;
while (trace) {
int inode = trace->index;
VTR_LOG("%d (%s)", inode, rr_node_typename[device_ctx.rr_nodes[inode].type()]);
if (trace->iswitch == OPEN) {
VTR_LOG(" !"); //End of branch
}
if (prev && prev->iswitch != OPEN && !device_ctx.rr_switch_inf[prev->iswitch].configurable()) {
VTR_LOG("*"); //Reached non-configurably
}
if (route_ctx.rr_node_route_inf[inode].occ() > device_ctx.rr_nodes[inode].capacity()) {
VTR_LOG(" x"); //Overused
}
VTR_LOG("\n");
prev = trace;
trace = trace->next;
}
VTR_LOG("\n");
}
bool validate_traceback(t_trace* trace) {
std::set<int> seen_rr_nodes;
return validate_traceback_recurr(trace, seen_rr_nodes);
}
bool validate_traceback_recurr(t_trace* trace, std::set<int>& seen_rr_nodes) {
if (!trace) {
return true;
}
seen_rr_nodes.insert(trace->index);
t_trace* next = trace->next;
if (next) {
if (trace->iswitch == OPEN) { //End of a branch
//Verify that the next element (branch point) has been already seen in the traceback so far
if (!seen_rr_nodes.count(next->index)) {
VPR_FATAL_ERROR(VPR_ERROR_ROUTE, "Traceback branch point %d not found", next->index);
} else {
//Recurse along the new branch
return validate_traceback_recurr(next, seen_rr_nodes);
}
} else { //Midway along branch
//Check there is an edge connecting trace and next
auto& device_ctx = g_vpr_ctx.device();
bool found = false;
for (t_edge_size iedge = 0; iedge < device_ctx.rr_nodes[trace->index].num_edges(); ++iedge) {
int to_node = device_ctx.rr_nodes[trace->index].edge_sink_node(iedge);
if (to_node == next->index) {
found = true;
//Verify that the switch matches
int rr_iswitch = device_ctx.rr_nodes[trace->index].edge_switch(iedge);
if (trace->iswitch != rr_iswitch) {
VPR_FATAL_ERROR(VPR_ERROR_ROUTE, "Traceback mismatched switch type: traceback %d rr_graph %d (RR nodes %d -> %d)\n",
trace->iswitch, rr_iswitch,
trace->index, to_node);
}
break;
}
}
if (!found) {
VPR_FATAL_ERROR(VPR_ERROR_ROUTE, "Traceback no RR edge between RR nodes %d -> %d\n", trace->index, next->index);
}
//Recurse
return validate_traceback_recurr(next, seen_rr_nodes);
}
}
VTR_ASSERT(!next);
return true; //End of traceback
}
//Print information about an invalid routing, caused by overused routing resources
void print_invalid_routing_info() {
auto& device_ctx = g_vpr_ctx.device();
auto& cluster_ctx = g_vpr_ctx.clustering();
auto& route_ctx = g_vpr_ctx.routing();
//Build a look-up of nets using each RR node
std::multimap<int, ClusterNetId> rr_node_nets;
for (auto net_id : cluster_ctx.clb_nlist.nets()) {
t_trace* tptr = route_ctx.trace[net_id].head;
while (tptr != nullptr) {
rr_node_nets.emplace(tptr->index, net_id);
tptr = tptr->next;
}
}
for (size_t inode = 0; inode < device_ctx.rr_nodes.size(); inode++) {
int occ = route_ctx.rr_node_route_inf[inode].occ();
int cap = device_ctx.rr_nodes[inode].capacity();
if (occ > cap) {
VTR_LOG(" %s is overused (occ=%d capacity=%d)\n", describe_rr_node(inode).c_str(), occ, cap);
auto range = rr_node_nets.equal_range(inode);
for (auto itr = range.first; itr != range.second; ++itr) {
auto net_id = itr->second;
VTR_LOG(" Used by net %s (%zu)\n", cluster_ctx.clb_nlist.net_name(net_id).c_str(), size_t(net_id));
}
}
}
}
void print_rr_node_route_inf() {
auto& route_ctx = g_vpr_ctx.routing();
auto& device_ctx = g_vpr_ctx.device();
for (size_t inode = 0; inode < route_ctx.rr_node_route_inf.size(); ++inode) {
if (!std::isinf(route_ctx.rr_node_route_inf[inode].path_cost)) {
int prev_node = route_ctx.rr_node_route_inf[inode].prev_node;
int prev_edge = route_ctx.rr_node_route_inf[inode].prev_edge;
VTR_LOG("rr_node: %d prev_node: %d prev_edge: %d",
inode, prev_node, prev_edge);
if (prev_node != OPEN && prev_edge != OPEN && !device_ctx.rr_nodes[prev_node].edge_is_configurable(prev_edge)) {
VTR_LOG("*");
}
VTR_LOG(" pcost: %g back_pcost: %g\n",
route_ctx.rr_node_route_inf[inode].path_cost, route_ctx.rr_node_route_inf[inode].backward_path_cost);
}
}
}
void print_rr_node_route_inf_dot() {
auto& route_ctx = g_vpr_ctx.routing();
auto& device_ctx = g_vpr_ctx.device();
VTR_LOG("digraph G {\n");
VTR_LOG("\tnode[shape=record]\n");
for (size_t inode = 0; inode < route_ctx.rr_node_route_inf.size(); ++inode) {
if (!std::isinf(route_ctx.rr_node_route_inf[inode].path_cost)) {
VTR_LOG("\tnode%zu[label=\"{%zu (%s)", inode, inode, device_ctx.rr_nodes[inode].type_string());
if (route_ctx.rr_node_route_inf[inode].occ() > device_ctx.rr_nodes[inode].capacity()) {
VTR_LOG(" x");
}
VTR_LOG("}\"]\n");
}
}
for (size_t inode = 0; inode < route_ctx.rr_node_route_inf.size(); ++inode) {
if (!std::isinf(route_ctx.rr_node_route_inf[inode].path_cost)) {
int prev_node = route_ctx.rr_node_route_inf[inode].prev_node;
int prev_edge = route_ctx.rr_node_route_inf[inode].prev_edge;
if (prev_node != OPEN && prev_edge != OPEN) {
VTR_LOG("\tnode%d -> node%zu [", prev_node, inode);
if (prev_node != OPEN && prev_edge != OPEN && !device_ctx.rr_nodes[prev_node].edge_is_configurable(prev_edge)) {
VTR_LOG("label=\"*\"");
}
VTR_LOG("];\n");
}
}
}
VTR_LOG("}\n");
}
static bool validate_trace_nodes(t_trace* head, const std::unordered_set<int>& trace_nodes) {
//Verifies that all nodes in the traceback 'head' are conatined in 'trace_nodes'
if (!head) {
return true;
}
std::vector<int> missing_from_trace_nodes;
for (t_trace* tptr = head; tptr != nullptr; tptr = tptr->next) {
if (!trace_nodes.count(tptr->index)) {
missing_from_trace_nodes.push_back(tptr->index);
}
}
if (!missing_from_trace_nodes.empty()) {
std::string msg = vtr::string_fmt(
"The following %zu nodes were found in traceback"
" but were missing from trace_nodes: %s\n",
missing_from_trace_nodes.size(),
vtr::join(missing_from_trace_nodes, ", ").c_str());
VPR_FATAL_ERROR(VPR_ERROR_ROUTE, msg.c_str());
return false;
}
return true;
}
// True if router will use a lookahead.
//
// This controls whether the router lookahead cache will be primed outside of
// the router ScopedStartFinishTimer.
bool router_needs_lookahead(enum e_router_algorithm router_algorithm) {
switch (router_algorithm) {
case BREADTH_FIRST:
return false;
case TIMING_DRIVEN:
return true;
default:
VPR_FATAL_ERROR(VPR_ERROR_ROUTE, "Unknown routing algorithm %d",
router_algorithm);
}
}