Merge pull request #720 from acomodi/interchange-clusters

interchange: enable clusters support

11 files changed, 711 insertions, 29 deletions

diff --git a/fpga_interchange/arch.cc b/fpga_interchange/arch.cc
index 8e7fe2a3..1bbe9a02 100644
--- a/fpga_interchange/arch.cc
+++ b/fpga_interchange/arch.cc
@@ -758,6 +758,7 @@ bool Arch::pack()
     pack_ports();
     pack_default_conns();
     expand_macros();
+    pack_cluster();
     return true;
 }
 
diff --git a/fpga_interchange/arch.h b/fpga_interchange/arch.h
index c8a61430..77a7ae59 100644
--- a/fpga_interchange/arch.h
+++ b/fpga_interchange/arch.h
@@ -94,6 +94,15 @@ struct TileStatus
     PseudoPipModel pseudo_pip_model;
 };
 
+struct Cluster
+{
+    uint32_t index;
+    CellInfo *root;
+    std::vector<CellInfo *> cluster_nodes;
+    dict<IdString, IdString> cell_cluster_node_map;
+    dict<IdString, std::vector<std::pair<IdString, CellInfo *>>> cluster_node_cells;
+};
+
 struct Arch : ArchAPI<ArchRanges>
 {
     boost::iostreams::mapped_file_source blob_file;
@@ -258,6 +267,20 @@ struct Arch : ArchAPI<ArchRanges>
                 map_cell_pins(cell, mapping, /*bind_constants=*/false);
             }
             constraints.bindBel(tile_status.tags.data(), get_cell_constraints(bel, cell->type));
+
+            // Clean previous cell placement in tile
+            if (cell->bel != BelId()) {
+                TileStatus &prev_tile_status = get_tile_status(cell->bel.tile);
+                NPNR_ASSERT(prev_tile_status.boundcells[cell->bel.index] != nullptr);
+
+                const auto &prev_bel_data = bel_info(chip_info, cell->bel);
+                NPNR_ASSERT(prev_bel_data.category == BEL_CATEGORY_LOGIC);
+
+                get_site_status(prev_tile_status, prev_bel_data).unbindBel(cell);
+                prev_tile_status.boundcells[cell->bel.index] = nullptr;
+
+                constraints.unbindBel(prev_tile_status.tags.data(), get_cell_constraints(cell->bel, cell->type));
+            }
         } else {
             map_port_pins(bel, cell);
             // FIXME: Probably need to actually constraint io port cell/bel,
@@ -687,7 +710,14 @@ struct Arch : ArchAPI<ArchRanges>
 
     void place_iobufs(WireId pad_wire, NetInfo *net, const pool<CellInfo *, hash_ptr_ops> &tightly_attached_bels,
                       pool<CellInfo *, hash_ptr_ops> *placed_cells);
+
     void pack_ports();
+
+    // Clusters
+    void pack_cluster();
+    void prepare_cluster(const ClusterPOD *cluster, uint32_t index);
+    dict<ClusterId, Cluster> clusters;
+
     void decode_lut_cells();
 
     const GlobalCellPOD *global_cell_info(IdString cell_type) const;
@@ -821,10 +851,10 @@ struct Arch : ArchAPI<ArchRanges>
         }
         const TileStatus &tile_status = iter->second;
         const CellInfo *cell = tile_status.boundcells[bel.index];
+
         if (cell != nullptr) {
-            if (!dedicated_interconnect.isBelLocationValid(bel, cell)) {
+            if (cell->cluster == ClusterId() && !dedicated_interconnect.isBelLocationValid(bel, cell))
                 return false;
-            }
 
             if (io_port_types.count(cell->type)) {
                 // FIXME: Probably need to actually constraint io port cell/bel,
@@ -837,24 +867,21 @@ struct Arch : ArchAPI<ArchRanges>
                 return false;
             }
         }
+
         // Still check site status if cell is nullptr; as other bels in the site could be illegal (for example when
         // dedicated paths can no longer be used after ripping up a cell)
         auto &bel_data = bel_info(chip_info, bel);
-        return get_site_status(tile_status, bel_data).checkSiteRouting(getCtx(), tile_status);
-    }
+        bool site_status = get_site_status(tile_status, bel_data).checkSiteRouting(getCtx(), tile_status);
 
-    // -------------------------------------------------
+        return site_status;
+    }
 
-    // TODO
-    CellInfo *getClusterRootCell(ClusterId cluster) const override { NPNR_ASSERT_FALSE("unimplemented"); }
-    ArcBounds getClusterBounds(ClusterId cluster) const override { NPNR_ASSERT_FALSE("unimplemented"); }
-    Loc getClusterOffset(const CellInfo *cell) const override { NPNR_ASSERT_FALSE("unimplemented"); }
-    bool isClusterStrict(const CellInfo *cell) const override { NPNR_ASSERT_FALSE("unimplemented"); }
+    CellInfo *getClusterRootCell(ClusterId cluster) const override;
+    ArcBounds getClusterBounds(ClusterId cluster) const override;
+    Loc getClusterOffset(const CellInfo *cell) const override;
+    bool isClusterStrict(const CellInfo *cell) const override;
     bool getClusterPlacement(ClusterId cluster, BelId root_bel,
-                             std::vector<std::pair<CellInfo *, BelId>> &placement) const override
-    {
-        NPNR_ASSERT_FALSE("unimplemented");
-    }
+                             std::vector<std::pair<CellInfo *, BelId>> &placement) const override;
 
     IdString get_bel_tiletype(BelId bel) const { return IdString(loc_info(chip_info, bel).name); }
 
diff --git a/fpga_interchange/arch_pack_clusters.cc b/fpga_interchange/arch_pack_clusters.cc
new file mode 100644
index 00000000..18162aa4
--- /dev/null
+++ b/fpga_interchange/arch_pack_clusters.cc
@@ -0,0 +1,572 @@
+/*
+ *  nextpnr -- Next Generation Place and Route
+ *
+ *  Copyright (C) 2021  Symbiflow Authors
+ *
+ *
+ *  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.
+ *
+ */
+
+#include "arch.h"
+#include "design_utils.h"
+#include "log.h"
+#include "nextpnr.h"
+#include "util.h"
+
+#include <boost/algorithm/string.hpp>
+#include <queue>
+
+NEXTPNR_NAMESPACE_BEGIN
+
+enum ClusterWireNodeState
+{
+    IN_SINK_SITE = 0,
+    IN_ROUTING = 1,
+    IN_SOURCE_SITE = 2,
+    ONLY_IN_SOURCE_SITE = 3
+};
+
+enum ExpansionDirection
+{
+    CLUSTER_UPHILL_DIR = 0,
+    CLUSTER_DOWNHILL_DIR = 1
+};
+
+struct ClusterWireNode
+{
+    WireId wire;
+    ClusterWireNodeState state;
+    int depth;
+};
+
+static void handle_expansion_node(const Context *ctx, WireId prev_wire, PipId pip, ClusterWireNode curr_node,
+                                  std::vector<ClusterWireNode> &nodes_to_expand, pool<BelId> &bels,
+                                  ExpansionDirection direction)
+{
+    WireId wire;
+
+    if (direction == CLUSTER_UPHILL_DIR)
+        wire = ctx->getPipSrcWire(pip);
+    else
+        wire = ctx->getPipDstWire(pip);
+
+    if (wire == WireId())
+        return;
+
+    ClusterWireNode next_node;
+    next_node.wire = wire;
+    next_node.depth = curr_node.depth;
+
+    if (next_node.depth >= 2)
+        return;
+
+    auto const &wire_data = ctx->wire_info(wire);
+
+    bool expand_node = true;
+    if (ctx->is_site_port(pip)) {
+        switch (curr_node.state) {
+        case ONLY_IN_SOURCE_SITE:
+            expand_node = false;
+            break;
+        case IN_SOURCE_SITE:
+            NPNR_ASSERT(wire_data.site == -1);
+            next_node.state = IN_ROUTING;
+            break;
+        case IN_ROUTING:
+            NPNR_ASSERT(wire_data.site != -1);
+            next_node.state = IN_SINK_SITE;
+            break;
+        case IN_SINK_SITE:
+            expand_node = false;
+            break;
+        default:
+            // Unreachable!!!
+            NPNR_ASSERT(false);
+        }
+    } else {
+        if (next_node.state == IN_ROUTING)
+            next_node.depth++;
+        next_node.state = curr_node.state;
+    }
+
+    if (expand_node)
+        nodes_to_expand.push_back(next_node);
+    else
+        return;
+
+    if (next_node.state == IN_SINK_SITE || next_node.state == ONLY_IN_SOURCE_SITE) {
+        for (BelPin bel_pin : ctx->getWireBelPins(wire)) {
+            BelId bel = bel_pin.bel;
+            auto const &bel_data = bel_info(ctx->chip_info, bel);
+
+            if (bels.count(bel))
+                continue;
+
+            if (bel_data.category != BEL_CATEGORY_LOGIC)
+                return;
+
+            if (bel_data.synthetic)
+                return;
+
+            if (direction == CLUSTER_UPHILL_DIR) {
+                // Check that the BEL is indeed the one reached by backward exploration,
+                // by checking the previous visited wire.
+                for (IdString check_pin : ctx->getBelPins(bel)) {
+                    if (prev_wire == ctx->getBelPinWire(bel, check_pin)) {
+                        bels.insert(bel);
+                        break;
+                    }
+                }
+            } else {
+                bels.insert(bel);
+            }
+        }
+    }
+
+    return;
+}
+
+static pool<BelId> find_cluster_bels(const Context *ctx, WireId wire, ExpansionDirection direction,
+                                     bool out_of_site_expansion = false)
+{
+    std::vector<ClusterWireNode> nodes_to_expand;
+    pool<BelId> bels;
+
+    const auto &wire_data = ctx->wire_info(wire);
+    NPNR_ASSERT(wire_data.site != -1);
+
+    ClusterWireNode wire_node;
+    wire_node.wire = wire;
+    wire_node.state = IN_SOURCE_SITE;
+    if (!out_of_site_expansion)
+        wire_node.state = ONLY_IN_SOURCE_SITE;
+    wire_node.depth = 0;
+
+    nodes_to_expand.push_back(wire_node);
+
+    while (!nodes_to_expand.empty()) {
+        ClusterWireNode node_to_expand = nodes_to_expand.back();
+        WireId prev_wire = node_to_expand.wire;
+        nodes_to_expand.pop_back();
+
+        if (direction == CLUSTER_DOWNHILL_DIR) {
+            for (PipId pip : ctx->getPipsDownhill(node_to_expand.wire)) {
+                if (ctx->is_pip_synthetic(pip))
+                    continue;
+
+                handle_expansion_node(ctx, prev_wire, pip, node_to_expand, nodes_to_expand, bels, direction);
+            }
+        } else {
+            NPNR_ASSERT(direction == CLUSTER_UPHILL_DIR);
+            for (PipId pip : ctx->getPipsUphill(node_to_expand.wire)) {
+                if (ctx->is_pip_synthetic(pip))
+                    continue;
+
+                handle_expansion_node(ctx, prev_wire, pip, node_to_expand, nodes_to_expand, bels, direction);
+            }
+        }
+    }
+
+    return bels;
+}
+
+CellInfo *Arch::getClusterRootCell(ClusterId cluster) const
+{
+    NPNR_ASSERT(cluster != ClusterId());
+    return clusters.at(cluster).root;
+}
+
+bool Arch::getClusterPlacement(ClusterId cluster, BelId root_bel,
+                               std::vector<std::pair<CellInfo *, BelId>> &placement) const
+{
+    const Context *ctx = getCtx();
+    const Cluster &packed_cluster = clusters.at(cluster);
+
+    CellInfo *root_cell = getClusterRootCell(cluster);
+    if (!ctx->isValidBelForCellType(root_cell->type, root_bel))
+        return false;
+
+    BelId next_bel;
+
+    // Place cluster
+    for (CellInfo *cluster_node : packed_cluster.cluster_nodes) {
+        if (cluster_node == root_cell) {
+            next_bel = root_bel;
+        } else {
+            // Find next chained cluster node
+            auto &cluster_data = cluster_info(chip_info, packed_cluster.index);
+
+            IdString next_bel_pin(cluster_data.chainable_ports[0].bel_source);
+            WireId next_bel_pin_wire = ctx->getBelPinWire(next_bel, next_bel_pin);
+            next_bel = BelId();
+            for (BelId bel :
+                 find_cluster_bels(ctx, next_bel_pin_wire, CLUSTER_DOWNHILL_DIR, /*out_of_site_expansion=*/true)) {
+                if (ctx->isValidBelForCellType(cluster_node->type, bel)) {
+                    next_bel = bel;
+                    break;
+                }
+            }
+
+            if (next_bel == BelId())
+                return false;
+        }
+
+        // Build a cell to bell mapping required to find BELs connected to the cluster ports.
+        dict<IdString, std::vector<IdString>> cell_bel_pins;
+
+        int32_t mapping = bel_info(chip_info, next_bel).pin_map[get_cell_type_index(cluster_node->type)];
+        NPNR_ASSERT(mapping >= 0);
+
+        const CellBelMapPOD &cell_pin_map = chip_info->cell_map->cell_bel_map[mapping];
+        for (const auto &pin_map : cell_pin_map.common_pins) {
+            IdString cell_pin(pin_map.cell_pin);
+            IdString bel_pin(pin_map.bel_pin);
+
+            cell_bel_pins[cell_pin].push_back(bel_pin);
+        }
+
+        placement.emplace_back(cluster_node, next_bel);
+
+        // Place cluster node cells at the same site
+        for (auto port_cell : packed_cluster.cluster_node_cells.at(cluster_node->name)) {
+            bool placed_cell = false;
+
+            IdString port = port_cell.first;
+            CellInfo *cell = port_cell.second;
+
+            NPNR_ASSERT(cell_bel_pins.count(port));
+
+            PortType port_type = cluster_node->ports.at(port).type;
+
+            if (port_type == PORT_INOUT)
+                continue;
+
+            for (auto &bel_pin : cell_bel_pins.at(port)) {
+                WireId bel_pin_wire = ctx->getBelPinWire(next_bel, bel_pin);
+
+                ExpansionDirection direction = port_type == PORT_IN ? CLUSTER_UPHILL_DIR : CLUSTER_DOWNHILL_DIR;
+                pool<BelId> cluster_bels = find_cluster_bels(ctx, bel_pin_wire, direction);
+
+                if (cluster_bels.size() == 0)
+                    continue;
+
+                for (BelId bel : cluster_bels) {
+                    if (ctx->isValidBelForCellType(cell->type, bel)) {
+                        placement.emplace_back(cell, bel);
+                        placed_cell = true;
+                        break;
+                    }
+                }
+
+                if (placed_cell)
+                    break;
+            }
+
+            if (!placed_cell)
+                return false;
+        }
+    }
+
+    return true;
+}
+
+ArcBounds Arch::getClusterBounds(ClusterId cluster) const
+{
+    // TODO: Implement this
+    ArcBounds bounds(0, 0, 0, 0);
+    return bounds;
+}
+
+Loc Arch::getClusterOffset(const CellInfo *cell) const
+{
+    Loc offset;
+    CellInfo *root = getClusterRootCell(cell->cluster);
+
+    if (cell->bel != BelId() && root->bel != BelId()) {
+        Loc root_loc = getBelLocation(root->bel);
+        Loc cell_loc = getBelLocation(cell->bel);
+        offset.x = cell_loc.x - root_loc.x;
+        offset.y = cell_loc.y - root_loc.y;
+        offset.z = cell_loc.z - root_loc.z;
+    } else {
+        Cluster cluster = clusters.at(cell->cluster);
+        auto &cluster_data = cluster_info(chip_info, cluster.index);
+
+        if (cluster_data.chainable_ports.size() == 0)
+            return offset;
+
+        auto &chainable_port = cluster_data.chainable_ports[0];
+
+        IdString cluster_node = cluster.cell_cluster_node_map.at(cell->name);
+        CellInfo *cluster_node_cell = cells.at(cluster_node).get();
+
+        auto res = std::find(cluster.cluster_nodes.begin(), cluster.cluster_nodes.end(), cluster_node_cell);
+        NPNR_ASSERT(res != cluster.cluster_nodes.end());
+
+        auto distance = std::distance(cluster.cluster_nodes.begin(), res);
+
+        offset.x = chainable_port.avg_x_offset * distance;
+        offset.y = chainable_port.avg_y_offset * distance;
+    }
+
+    return offset;
+}
+
+bool Arch::isClusterStrict(const CellInfo *cell) const { return true; }
+
+static void dump_clusters(const ChipInfoPOD *chip_info, Context *ctx)
+{
+    for (size_t i = 0; i < chip_info->clusters.size(); ++i) {
+        const auto &cluster = chip_info->clusters[i];
+        IdString cluster_name(cluster.name);
+        log_info("Cluster '%s' loaded! Parameters:\n", cluster_name.c_str(ctx));
+
+        log_info("  - root cell types:\n");
+        for (auto cell : cluster.root_cell_types)
+            log_info("      - %s\n", IdString(cell).c_str(ctx));
+
+        for (auto chain_ports : cluster.chainable_ports)
+            log_info("  - chainable pair: source %s - sink %s\n", IdString(chain_ports.cell_source).c_str(ctx),
+                     IdString(chain_ports.cell_sink).c_str(ctx));
+
+        if (cluster.cluster_cells_map.size() != 0)
+            log_info("  - cell port maps:\n");
+        for (auto cluster_cell : cluster.cluster_cells_map) {
+            log_info("    - cell: %s - port: %s\n", IdString(cluster_cell.cell).c_str(ctx),
+                     IdString(cluster_cell.port).c_str(ctx));
+        }
+    }
+}
+
+static bool check_cluster_cells_compatibility(CellInfo *old_cell, CellInfo *new_cell, pool<IdString> &exclude_nets)
+{
+    NPNR_ASSERT(new_cell->type == old_cell->type);
+    for (auto &new_port_pair : new_cell->ports) {
+        PortInfo new_port_info = new_port_pair.second;
+        PortInfo old_port_info = old_cell->ports.at(new_port_pair.first);
+
+        if (exclude_nets.count(new_port_info.net->name))
+            continue;
+
+        if (new_port_info.type != PORT_IN)
+            continue;
+
+        if (new_port_info.net != old_port_info.net)
+            return false;
+    }
+
+    return true;
+}
+
+void Arch::prepare_cluster(const ClusterPOD *cluster, uint32_t index)
+{
+    Context *ctx = getCtx();
+    IdString cluster_name(cluster->name);
+
+    pool<IdString> cluster_cell_types;
+    for (auto cell_type : cluster->root_cell_types)
+        cluster_cell_types.insert(IdString(cell_type));
+
+    // Find cluster roots
+    std::vector<CellInfo *> roots;
+    for (auto &cell : cells) {
+        CellInfo *ci = cell.second.get();
+
+        if (ci->cluster != ClusterId())
+            continue;
+
+        if (!cluster_cell_types.count(ci->type))
+            continue;
+
+        if (cluster->chainable_ports.size() == 0) {
+            ci->cluster.set(ctx, ci->name.str(ctx));
+            roots.push_back(ci);
+            continue;
+        }
+
+        // Only one type of dedicated interconnect is allowed.
+        auto chain_ports = cluster->chainable_ports[0];
+        IdString source_port(chain_ports.cell_source);
+        IdString sink_port(chain_ports.cell_sink);
+
+        PortRef driver = ci->ports[sink_port].net->driver;
+
+        if (driver.cell == nullptr || driver.port != source_port) {
+            // We hit a root cell
+            ci->cluster.set(ctx, ci->name.c_str(ctx));
+            roots.push_back(ci);
+
+            // Chained cells use dedicated connections, usually not exposed to the
+            // general interconnect resources. The port disconnection is required for
+            // sink ports which are connected to GND or VCC by default, which are not
+            // reachable due to the fixed dedicated interconnect.
+            // E.g.: The CI input of carry chains in 7series corresponds to the CIN bel port,
+            //       which can only be connected to the COUT output of the tile below.
+            disconnect_port(ctx, ci, sink_port);
+        }
+    }
+
+    dict<IdString, pool<IdString>> port_cell_maps;
+    for (auto cell_port_map : cluster->cluster_cells_map) {
+        IdString cell(cell_port_map.cell);
+        IdString port(cell_port_map.port);
+
+        pool<IdString> cells_pool({cell});
+
+        port_cell_maps.emplace(port, cells_pool).first->second.insert(cell);
+    }
+
+    // Generate unique clusters starting from each root
+    for (auto root : roots) {
+        Cluster cluster_info;
+        cluster_info.root = root;
+        cluster_info.index = index;
+
+        CellInfo *next_cluster_node = root;
+        if (ctx->verbose)
+            log_info("  - forming cluster starting from root cell: %s\n", next_cluster_node->name.c_str(ctx));
+
+        // counter to determine whether this cluster needs to exist
+        uint32_t count_cluster_cells = 0;
+        do {
+            std::vector<std::pair<IdString, CellInfo *>> cluster_cells;
+
+            // type -> cells map to verify compatibility of cells in the same cluster
+            dict<IdString, CellInfo *> cell_type_dict;
+            pool<IdString> exclude_nets;
+
+            count_cluster_cells++;
+
+            for (auto port : next_cluster_node->ports) {
+                if (!port_cell_maps.count(port.first))
+                    continue;
+
+                PortInfo port_info = port.second;
+
+                if (port_info.type == PORT_OUT) {
+                    exclude_nets.insert(port_info.net->name);
+                    auto &users = port_info.net->users;
+                    if (users.size() != 1)
+                        continue;
+
+                    CellInfo *user_cell = users[0].cell;
+                    if (user_cell == nullptr)
+                        continue;
+
+                    if (!port_cell_maps.at(port.first).count(user_cell->type))
+                        continue;
+
+                    auto res = cell_type_dict.emplace(user_cell->type, user_cell);
+                    bool compatible = true;
+                    if (!res.second)
+                        // Check whether a cell of the same type has all the required nets compatible with
+                        // all other nets for the same type. If not, discard the cell.
+                        // An example is multiple FFs belonging to the same cluster, where one of them has a different
+                        // Set/Reset or CE net w.r.t. the others, making the cluster unplaceable.
+                        compatible = check_cluster_cells_compatibility(res.first->second, user_cell, exclude_nets);
+
+                    if (!compatible)
+                        continue;
+
+                    user_cell->cluster = root->cluster;
+                    cluster_cells.push_back(std::make_pair(port.first, user_cell));
+                    cluster_info.cell_cluster_node_map.emplace(user_cell->name, next_cluster_node->name);
+                    count_cluster_cells++;
+
+                    if (ctx->verbose)
+                        log_info("      - adding user cell: %s\n", user_cell->name.c_str(ctx));
+
+                } else if (port_info.type == PORT_IN) {
+                    auto &driver = port_info.net->driver;
+                    auto &users = port_info.net->users;
+                    if (users.size() != 1)
+                        continue;
+
+                    CellInfo *driver_cell = driver.cell;
+                    if (driver_cell == nullptr)
+                        continue;
+
+                    if (!port_cell_maps.at(port.first).count(driver_cell->type))
+                        continue;
+
+                    driver_cell->cluster = root->cluster;
+                    cluster_cells.push_back(std::make_pair(port.first, driver_cell));
+                    cluster_info.cell_cluster_node_map.emplace(driver_cell->name, next_cluster_node->name);
+                    count_cluster_cells++;
+
+                    if (ctx->verbose)
+                        log_info("      - adding driver cell: %s\n", driver_cell->name.c_str(ctx));
+                }
+            }
+
+            cluster_info.cell_cluster_node_map.emplace(next_cluster_node->name, next_cluster_node->name);
+            cluster_info.cluster_nodes.push_back(next_cluster_node);
+            cluster_info.cluster_node_cells.emplace(next_cluster_node->name, cluster_cells);
+
+            if (cluster->chainable_ports.size() == 0)
+                break;
+
+            // Only one type of dedicated interconnect is allowed.
+            auto chain_ports = cluster->chainable_ports[0];
+            IdString source_port(chain_ports.cell_source);
+            IdString sink_port(chain_ports.cell_sink);
+
+            NetInfo *next_net = next_cluster_node->ports.at(source_port).net;
+
+            if (next_net == nullptr)
+                continue;
+
+            next_cluster_node = nullptr;
+            for (auto &user : next_net->users) {
+                CellInfo *user_cell = user.cell;
+
+                if (user_cell == nullptr)
+                    continue;
+
+                if (cluster_cell_types.count(user_cell->type)) {
+                    user_cell->cluster = root->cluster;
+                    next_cluster_node = user_cell;
+                    break;
+                }
+            }
+
+            if (next_cluster_node == nullptr)
+                break;
+
+        } while (true);
+
+        if (count_cluster_cells == 1 && cluster->chainable_ports.size() == 0) {
+            root->cluster = ClusterId();
+            continue;
+        }
+
+        clusters.emplace(root->cluster, cluster_info);
+    }
+}
+
+void Arch::pack_cluster()
+{
+    Context *ctx = getCtx();
+
+    if (ctx->verbose)
+        dump_clusters(chip_info, ctx);
+
+    for (uint32_t i = 0; i < chip_info->clusters.size(); ++i) {
+        const auto &cluster = chip_info->clusters[i];
+
+        prepare_cluster(&cluster, i);
+    }
+}
+
+NEXTPNR_NAMESPACE_END
diff --git a/fpga_interchange/chipdb.h b/fpga_interchange/chipdb.h
index 155c2bb2..fde35e7f 100644
--- a/fpga_interchange/chipdb.h
+++ b/fpga_interchange/chipdb.h
@@ -34,7 +34,7 @@ NEXTPNR_NAMESPACE_BEGIN
  * kExpectedChipInfoVersion
  */
 
-static constexpr int32_t kExpectedChipInfoVersion = 10;
+static constexpr int32_t kExpectedChipInfoVersion = 11;
 
 // Flattened site indexing.
 //
@@ -402,6 +402,27 @@ NPNR_PACKED_STRUCT(struct MacroExpansionPOD {
     RelSlice<MacroParamMapRulePOD> param_rules;
 });
 
+NPNR_PACKED_STRUCT(struct ClusterCellPortPOD {
+    uint32_t cell;
+    uint32_t port;
+});
+
+NPNR_PACKED_STRUCT(struct ChainablePortPOD {
+    uint32_t cell_source;
+    uint32_t cell_sink;
+    uint32_t bel_source;
+    uint32_t bel_sink;
+    int16_t avg_x_offset;
+    int16_t avg_y_offset;
+});
+
+NPNR_PACKED_STRUCT(struct ClusterPOD {
+    uint32_t name;
+    RelSlice<uint32_t> root_cell_types;
+    RelSlice<ChainablePortPOD> chainable_ports;
+    RelSlice<ClusterCellPortPOD> cluster_cells_map;
+});
+
 NPNR_PACKED_STRUCT(struct ChipInfoPOD {
     RelPtr<char> name;
     RelPtr<char> generator;
@@ -421,6 +442,8 @@ NPNR_PACKED_STRUCT(struct ChipInfoPOD {
     RelSlice<MacroPOD> macros;
     RelSlice<MacroExpansionPOD> macro_rules;
 
+    RelSlice<ClusterPOD> clusters;
+
     // BEL bucket constids.
     RelSlice<int32_t> bel_buckets;
 
@@ -460,6 +483,11 @@ inline const SiteInstInfoPOD &site_inst_info(const ChipInfoPOD *chip_info, int32
     return chip_info->sites[chip_info->tiles[tile].sites[site]];
 }
 
+inline const ClusterPOD &cluster_info(const ChipInfoPOD *chip_info, int32_t cluster)
+{
+    return chip_info->clusters[cluster];
+}
+
 enum SyntheticType
 {
     NOT_SYNTH = 0,
diff --git a/fpga_interchange/examples/tests.cmake b/fpga_interchange/examples/tests.cmake
index 3c97fe26..48b1cee3 100644
--- a/fpga_interchange/examples/tests.cmake
+++ b/fpga_interchange/examples/tests.cmake
@@ -77,13 +77,14 @@ function(add_interchange_test)
 
     # Synthesis
     set(synth_json ${CMAKE_CURRENT_BINARY_DIR}/${name}.json)
+    set(synth_log ${CMAKE_CURRENT_BINARY_DIR}/${name}.json.log)
     add_custom_command(
         OUTPUT ${synth_json}
         COMMAND ${CMAKE_COMMAND} -E env
             SOURCES="${sources}"
             OUT_JSON=${synth_json}
             TECHMAP=${techmap}
-            yosys -c ${tcl}
+            yosys -c ${tcl} -l ${synth_log}
         DEPENDS ${sources} ${techmap} ${tcl}
     )
 
@@ -134,6 +135,7 @@ function(add_interchange_test)
     get_property(chipdb_bin_loc TARGET device-${device} PROPERTY CHIPDB_BIN_LOC)
 
     set(phys ${CMAKE_CURRENT_BINARY_DIR}/${name}.phys)
+    set(phys_log ${CMAKE_CURRENT_BINARY_DIR}/${name}.phys.log)
     add_custom_command(
         OUTPUT ${phys}
         COMMAND
@@ -143,6 +145,7 @@ function(add_interchange_test)
                 --netlist ${netlist}
                 --phys ${phys}
                 --package ${package}
+                --log ${phys_log}
         DEPENDS
             nextpnr-fpga_interchange
             ${netlist}
@@ -151,6 +154,7 @@ function(add_interchange_test)
             ${chipdb_bin_loc}
     )
 
+    set(phys_verbose_log ${CMAKE_CURRENT_BINARY_DIR}/${name}.phys.verbose.log)
     add_custom_target(
         test-${family}-${name}-phys-verbose
         COMMAND
@@ -161,6 +165,7 @@ function(add_interchange_test)
                 --phys ${phys}
                 --package ${package}
                 --verbose
+                --log ${phys_verbose_log}
         DEPENDS
             ${netlist}
             ${xdc}
diff --git a/fpga_interchange/examples/tests/counter/counter.v b/fpga_interchange/examples/tests/counter/counter.v
index 00f52a20..4b3f343b 100644
--- a/fpga_interchange/examples/tests/counter/counter.v
+++ b/fpga_interchange/examples/tests/counter/counter.v
@@ -1,13 +1,15 @@
 module top(input clk, input rst, output [7:4] io_led);
 
-reg [31:0] counter = 32'b0;
+localparam SIZE = 32;
 
-assign io_led = counter >> 22;
+reg [SIZE-1:0] counter = SIZE'b0;
+
+assign io_led = {counter[SIZE-1], counter[25:23]};
 
 always @(posedge clk)
 begin
     if(rst)
-        counter <= 32'b0;
+        counter <= SIZE'b0;
     else
         counter <= counter + 1;
 end
diff --git a/fpga_interchange/examples/tests/counter/run_xilinx.tcl b/fpga_interchange/examples/tests/counter/run_xilinx.tcl
index ffea3b2e..c02cf933 100644
--- a/fpga_interchange/examples/tests/counter/run_xilinx.tcl
+++ b/fpga_interchange/examples/tests/counter/run_xilinx.tcl
@@ -2,7 +2,7 @@ yosys -import
 
 read_verilog $::env(SOURCES)
 
-synth_xilinx -nolutram -nowidelut -nosrl -nocarry -nodsp
+synth_xilinx -nolutram -nowidelut -nosrl -nodsp
 techmap -map $::env(TECHMAP)
 
 # opt_expr -undriven makes sure all nets are driven, if only by the $undef
diff --git a/fpga_interchange/fpga_interchange.cpp b/fpga_interchange/fpga_interchange.cpp
index 1d08b128..ac2a7e96 100644
--- a/fpga_interchange/fpga_interchange.cpp
+++ b/fpga_interchange/fpga_interchange.cpp
@@ -539,10 +539,14 @@ void FpgaInterchange::write_physical_netlist(const Context * ctx, const std::str
     auto net_iter = nets.begin();
     for(auto & net_pair : ctx->nets) {
         auto &net = *net_pair.second;
-        auto net_out = *net_iter++;
 
         const CellInfo *driver_cell = net.driver.cell;
 
+        if (driver_cell == nullptr)
+            continue;
+
+        auto net_out = *net_iter++;
+
         // Handle GND and VCC nets.
         if(driver_cell->bel == ctx->get_gnd_bel()) {
             IdString gnd_net_name(ctx->chip_info->constants->gnd_net_name);
diff --git a/fpga_interchange/site_router.cc b/fpga_interchange/site_router.cc
index 090b9342..12dce9f5 100644
--- a/fpga_interchange/site_router.cc
+++ b/fpga_interchange/site_router.cc
@@ -60,7 +60,9 @@ bool check_initial_wires(const Context *ctx, SiteInformation *site_info)
             if (!cell->ports.count(pin_pair.first))
                 continue;
             const PortInfo &port = cell->ports.at(pin_pair.first);
-            NPNR_ASSERT(port.net != nullptr);
+
+            if (port.net == nullptr)
+                continue;
 
             for (IdString bel_pin_name : pin_pair.second) {
                 BelPin bel_pin;
@@ -297,7 +299,11 @@ struct SiteExpansionLoop
         // already unroutable!
         solution.clear();
         solution.store_solution(ctx, node_storage, net->driver, completed_routes);
-        solution.verify(ctx, *net);
+        bool verify = solution.verify(ctx, *net);
+
+        if (!verify)
+            return false;
+
         for (size_t route : completed_routes) {
             SiteWire wire = node_storage->get_node(route)->wire;
             targets.erase(wire);
@@ -1086,6 +1092,43 @@ static void block_lut_outputs(SiteArch *site_arch, const pool<std::pair<IdString
     }
 }
 
+// Block wires corresponding to dedicated interconnections that are not
+// exposed to the general interconnect.
+// These wires cannot be chosen for the first element in a BEL chain, as they
+// would result in an unroutability error.
+static void block_cluster_wires(SiteArch *site_arch)
+{
+    const Context *ctx = site_arch->site_info->ctx;
+    auto &cells_in_site = site_arch->site_info->cells_in_site;
+
+    for (auto &cell : cells_in_site) {
+        if (cell->cluster == ClusterId())
+            continue;
+
+        if (ctx->getClusterRootCell(cell->cluster) != cell)
+            continue;
+
+        Cluster cluster = ctx->clusters.at(cell->cluster);
+
+        uint32_t cluster_id = cluster.index;
+        auto &cluster_data = cluster_info(ctx->chip_info, cluster_id);
+
+        if (cluster_data.chainable_ports.size() == 0)
+            continue;
+
+        IdString cluster_chain_input(cluster_data.chainable_ports[0].cell_sink);
+
+        if (cluster_chain_input == IdString())
+            continue;
+
+        auto &cell_bel_pins = cell->cell_bel_pins.at(cluster_chain_input);
+        for (auto &bel_pin : cell_bel_pins) {
+            SiteWire bel_pin_wire = site_arch->getBelPinWire(cell->bel, bel_pin);
+            site_arch->bindWire(bel_pin_wire, &site_arch->blocking_site_net);
+        }
+    }
+}
+
 // Recursively visit downhill PIPs until a SITE_PORT_SINK is reached.
 // Marks all PIPs for all valid paths.
 static bool visit_downhill_pips(const SiteArch *site_arch, const SiteWire &site_wire, std::vector<PipId> &valid_pips)
@@ -1205,6 +1248,7 @@ bool SiteRouter::checkSiteRouting(const Context *ctx, const TileStatus &tile_sta
     // site_arch.archcheck();
 
     block_lut_outputs(&site_arch, blocked_wires);
+    block_cluster_wires(&site_arch);
 
     // Do a detailed routing check to see if the site has at least 1 valid
     // routing solution.
@@ -1264,6 +1308,7 @@ void SiteRouter::bindSiteRouting(Context *ctx)
 
     SiteArch site_arch(&site_info);
     block_lut_outputs(&site_arch, blocked_wires);
+    block_cluster_wires(&site_arch);
     NPNR_ASSERT(route_site(&site_arch, &ctx->site_routing_cache, &ctx->node_storage, /*explain=*/false));
 
     check_routing(site_arch);
diff --git a/fpga_interchange/site_routing_cache.cc b/fpga_interchange/site_routing_cache.cc
index 512ca2ac..cd16cfff 100644
--- a/fpga_interchange/site_routing_cache.cc
+++ b/fpga_interchange/site_routing_cache.cc
@@ -68,7 +68,7 @@ void SiteRoutingSolution::store_solution(const SiteArch *ctx, const RouteNodeSto
     solution_offsets.push_back(solution_storage.size());
 }
 
-void SiteRoutingSolution::verify(const SiteArch *ctx, const SiteNetInfo &net)
+bool SiteRoutingSolution::verify(const SiteArch *ctx, const SiteNetInfo &net)
 {
     pool<SiteWire> seen_users;
     for (size_t i = 0; i < num_solutions(); ++i) {
@@ -88,7 +88,7 @@ void SiteRoutingSolution::verify(const SiteArch *ctx, const SiteNetInfo &net)
         NPNR_ASSERT(net.driver == cursor);
     }
 
-    NPNR_ASSERT(seen_users.size() == net.users.size());
+    return seen_users.size() == net.users.size();
 }
 
 SiteRoutingKey SiteRoutingKey::make(const SiteArch *ctx, const SiteNetInfo &site_net)
@@ -194,9 +194,7 @@ bool SiteRoutingCache::get_solution(const SiteArch *ctx, const SiteNetInfo &net,
         }
     }
 
-    solution->verify(ctx, net);
-
-    return true;
+    return solution->verify(ctx, net);
 }
 
 void SiteRoutingCache::add_solutions(const SiteArch *ctx, const SiteNetInfo &net, const SiteRoutingSolution &solution)
diff --git a/fpga_interchange/site_routing_cache.h b/fpga_interchange/site_routing_cache.h
index b4baf65a..a430d206 100644
--- a/fpga_interchange/site_routing_cache.h
+++ b/fpga_interchange/site_routing_cache.h
@@ -32,7 +32,7 @@ struct SiteRoutingSolution
 {
     void store_solution(const SiteArch *ctx, const RouteNodeStorage *node_storage, const SiteWire &driver,
                         std::vector<size_t> solutions);
-    void verify(const SiteArch *ctx, const SiteNetInfo &net);
+    bool verify(const SiteArch *ctx, const SiteNetInfo &net);
 
     void clear()
     {