gatemate: support multiple clock distribution strategies (#1574)

* gatemate: support multiple clock distribution strategies

* error out on non supported cases

* Implement full use strategy

* Address review comments

common/place/placer_heap.cc

@@ -921,8 +921,10 @@ class HeAPPlacer
                     y0 = std::max(y0, r.y0);
                     x1 = std::min(x1, r.x1);
                     y1 = std::min(y1, r.y1);
-                    if (x0 > x1) std::swap(x0, x1);
+                    if (x0 > x1)
-                    if (y0 > y1) std::swap(y0, y1);
+                        std::swap(x0, x1);
+                    if (y0 > y1)
+                        std::swap(y0, y1);
                 }
 
                 // Pick a random X and Y location within our search radius / search box

himbaechel/uarch/gatemate/cells.cc

@@ -125,6 +125,11 @@ CellInfo *GateMatePacker::create_cell_ptr(IdString type, IdString name)
     } else if (type.in(id_CPE_IBUF)) {
         add_port(id_Y, PORT_OUT);
         add_port(id_I, PORT_IN);
+    } else if (type.in(id_CPE_IOBUF)) {
+        add_port(id_Y, PORT_OUT);
+        add_port(id_T, PORT_IN);
+        add_port(id_A, PORT_IN);
+        add_port(id_IO, PORT_INOUT);
     } else if (type.in(id_PLL)) {
         add_port(id_CLK_REF, PORT_IN);
         add_port(id_USR_CLK_REF, PORT_IN);

himbaechel/uarch/gatemate/gatemate.cc

@@ -162,6 +162,7 @@ void GateMateImpl::init(Context *ctx)
     int index = 0;
     for (auto &die : extra->dies) {
         IdString name(die.name);
+        index_to_die[index] = name;
         die_to_index[name] = index++;
         ctx->createRectangularRegion(name, die.x1, die.y1, die.x2, die.y2);
         if (die_name == name.c_str(ctx)) {

himbaechel/uarch/gatemate/gatemate.h

@@ -30,6 +30,13 @@
 
 NEXTPNR_NAMESPACE_BEGIN
 
+enum MultiDieStrategy
+{
+    CLOCK_MIRROR,
+    REUSE_CLK1,
+    FULL_USE,
+};
+
 struct GateMateImpl : HimbaechelAPI
 {
     ~GateMateImpl();
@@ -91,9 +98,14 @@ struct GateMateImpl : HimbaechelAPI
     int fpga_mode;
     int timing_mode;
     std::map<const NetInfo *, int> global_signals;
+    dict<std::pair<IdString, int>, NetInfo *> global_mapping;
+    dict<std::pair<IdString, int>, IdString> global_clk_mapping;
     std::vector<CellInfo *> clkin;
     std::vector<CellInfo *> glbout;
     std::vector<CellInfo *> pll;
+    pool<IdString> ignore;
+    MultiDieStrategy strategy;
+    dict<int, IdString> index_to_die;
 
   private:
     bool getChildPlacement(const BaseClusterInfo *cluster, Loc root_loc,

himbaechel/uarch/gatemate/pack.cc

@@ -397,6 +397,37 @@ void GateMatePacker::remove_double_constrained()
     }
 }
 
+void GateMatePacker::recursiveAddToRegion(CellInfo *root, IdString die)
+{
+    if (root->region && root->region->name != die)
+        log_error("Trying to assign cell '%s' to multiple regions.\n", root->name.c_str(ctx));
+    ctx->constrainCellToRegion(root->name, die);
+    for (auto cell : root->constr_children) {
+        if (cell->region && cell->region->name != die)
+            log_error("Trying to assign cell '%s' to multiple regions.\n", cell->name.c_str(ctx));
+        ctx->constrainCellToRegion(cell->name, die);
+        recursiveAddToRegion(cell, die);
+    }
+}
+
+void GateMatePacker::assign_clocks()
+{
+    log_info("Assign cells based on clock..\n");
+    for (auto &glob : uarch->global_signals) {
+        const NetInfo *net = glob.first;
+        for (auto &user : net->users) {
+            IdString die = uarch->index_to_die[uarch->tile_extra_data(net->driver.cell->bel.tile)->die];
+            if (user.cell->region && user.cell->region->name != die)
+                log_error("Trying to assign cell '%s' to multiple regions.\n", user.cell->name.c_str(ctx));
+            ctx->constrainCellToRegion(user.cell->name, die);
+            if (user.cell->cluster != ClusterId()) {
+                CellInfo *root = ctx->getClusterRootCell(user.cell->cluster);
+                recursiveAddToRegion(root, die);
+            }
+        }
+    }
+}
+
 void GateMateImpl::pack()
 {
     const ArchArgs &args = ctx->args;
@@ -404,8 +435,34 @@ void GateMateImpl::pack()
         parse_ccf(args.options.at("ccf"));
     }
 
-    if (forced_die != IdString())
+    if (args.options.count("multi")) {
+        std::string val = args.options.at("multi");
+        if (val == "mirror") {
+            strategy = MultiDieStrategy::CLOCK_MIRROR;
+            log_info("Multidie mode: CLOCK MIRROR\n");
+        } else if (val == "clk1") {
+            strategy = MultiDieStrategy::REUSE_CLK1;
+            log_info("Multidie mode: REUSE CLK1\n");
+        } else if (val == "full") {
+            strategy = MultiDieStrategy::FULL_USE;
+            log_info("Multidie mode: FULL USE\n");
+        } else {
+            log_error("Unknown value for 'multi' option. Allowed values are 'mirror', 'full' and 'clk1'.\n");
+        }
+    } else {
+        strategy = MultiDieStrategy::CLOCK_MIRROR;
+        if (dies != 1)
+            log_warning("Multi die clock placement strategy set to 'mirror'.\n");
+    }
+
+    if (forced_die != IdString()) {
         preferred_die = die_to_index[forced_die];
+        if (strategy == MultiDieStrategy::FULL_USE)
+            log_error("Not allowed to use forced die in FULL USE mode.\n");
+    }
+
+    if (strategy == MultiDieStrategy::REUSE_CLK1 || strategy == MultiDieStrategy::FULL_USE)
+        preferred_die = 0;
 
     GateMatePacker packer(ctx, this);
     packer.pack_constants();
@@ -431,6 +488,9 @@ void GateMateImpl::pack()
                 ctx->constrainCellToRegion(cell.second->name, forced_die);
         }
     }
+
+    if (strategy == MultiDieStrategy::FULL_USE)
+        packer.assign_clocks();
 }
 
 void GateMateImpl::repack()
@@ -438,7 +498,8 @@ void GateMateImpl::repack()
     GateMatePacker packer(ctx, this);
     packer.repack_ram();
     packer.repack_cpe();
-    packer.reassign_clocks();
+    if (strategy != MultiDieStrategy::FULL_USE)
+        packer.reassign_clocks();
     packer.remove_clocking();
 }
 

himbaechel/uarch/gatemate/pack.h

@@ -53,7 +53,7 @@ struct GateMatePacker
     void pack_cpe();
     void pack_addf();
     void pack_bufg();
-    void sort_bufg();
+    void sort_bufg(unsigned max_num);
     void insert_clocking();
     void pack_pll();
     void pack_misc();
@@ -71,6 +71,7 @@ struct GateMatePacker
     void repack_ram();
     void reassign_clocks();
     void copy_clocks();
+    void assign_clocks();
 
   private:
     void rename_param(CellInfo *cell, IdString name, IdString new_name, int width);
@@ -88,6 +89,12 @@ struct GateMatePacker
     void move_connections(NetInfo *from_net, NetInfo *to_net);
     void remap_ram_half(CellInfo *half, CellInfo *cell, int num);
 
+    void strategy_mirror();
+    void strategy_clk1();
+    void strategy_full();
+
+    void recursiveAddToRegion(CellInfo *root, IdString die);
+
     PllCfgRecord get_pll_settings(double f_ref, double f_core, int mode, int low_jitter, bool pdiv0_mux, bool feedback);
 
     std::pair<CellInfo *, CellInfo *> move_ram_i(CellInfo *cell, IdString origPort, bool place = true,

himbaechel/uarch/gatemate/pack_clocking.cc

@@ -30,7 +30,7 @@ NEXTPNR_NAMESPACE_BEGIN
 
 inline bool is_bufg(const BaseCtx *ctx, const CellInfo *cell) { return cell->type.in(id_CC_BUFG); }
 
-void GateMatePacker::sort_bufg()
+void GateMatePacker::sort_bufg(unsigned max_num)
 {
     struct ItemBufG
     {
@@ -61,8 +61,8 @@ void GateMatePacker::sort_bufg()
         bufg.push_back(ItemBufG(&ci, o_net->users.entries()));
     }
 
-    if (bufg.size() > 4) {
+    if (bufg.size() > max_num) {
-        log_warning("More than 4 BUFG used. Those with highest fan-out will be used.\n");
+        log_warning("More than %d BUFG used. Those with highest fan-out will be used.\n", max_num);
         std::sort(bufg.begin(), bufg.end(), [](const ItemBufG &a, const ItemBufG &b) { return a.fan_out > b.fan_out; });
         for (size_t i = 4; i < bufg.size(); i++) {
             log_warning("Removing BUFG cell %s.\n", bufg.at(i).cell->name.c_str(ctx));
@@ -92,7 +92,7 @@ static int glb_mux_mapping[] = {
 
 void GateMatePacker::pack_bufg()
 {
-    sort_bufg();
+    sort_bufg(uarch->strategy == MultiDieStrategy::FULL_USE ? 4 * uarch->dies : 4);
 
     log_info("Packing BUFGs..\n");
     auto update_bufg_port = [&](std::vector<CellInfo *> &bufg, CellInfo *cell, int port_num, int pll_num) {
@@ -113,7 +113,7 @@ void GateMatePacker::pack_bufg()
         return true;
     };
 
-    unsigned max_plls = 4; // * uarch->dies;
+    unsigned max_plls = 4 * uarch->dies;
     // Index vector for permutation
     std::vector<unsigned> indexes(max_plls);
     for (unsigned i = 0; i < max_plls; ++i)
@@ -682,7 +682,117 @@ void GateMatePacker::copy_clocks()
 {
     if (uarch->dies == 1)
         return;
-    log_info("Copy clocks..\n");
+    switch (uarch->strategy) {
+    case MultiDieStrategy::REUSE_CLK1:
+        if (uarch->global_signals.size() > 1 || uarch->pll.size() > 1)
+            log_error("Unable to use REUSE CLK1 strategy when there is more than one clock/PLL.\n");
+        strategy_clk1();
+        break;
+    case MultiDieStrategy::CLOCK_MIRROR:
+        if (uarch->global_signals.size() > 4 || uarch->pll.size() > 4)
+            log_error("Unable to use MIRROR CLOCK strategy when there is more than 4 clocks/PLLs.\n");
+        strategy_mirror();
+        break;
+    case MultiDieStrategy::FULL_USE:
+        strategy_full();
+        break;
+    }
+}
+
+void GateMatePacker::strategy_full()
+{
+    log_info("All resources for clock distribution..\n");
+
+    for (int new_die = 1; new_die < uarch->dies; new_die++) {
+        // Create appropriate GPIO and IOSEL cells
+        for (int i = 0; i < 4; i++) {
+            NetInfo *in_net = uarch->clkin[new_die]->getPort(ctx->idf("CLK%d", i));
+            if (in_net && in_net->driver.cell) {
+                CellInfo *iosel = in_net->driver.cell;
+                auto pad_info = uarch->bel_to_pad[iosel->bel];
+                Loc l = uarch->locations[std::make_pair(IdString(pad_info->package_pin), new_die)];
+                CellInfo *iosel_new = ctx->getBoundBelCell(ctx->getBelByLocation(l));
+                if (!iosel_new) {
+                    iosel_new = create_cell_ptr(iosel->type, ctx->idf("%s$die%d", iosel->name.c_str(ctx), new_die));
+                    iosel_new->params = iosel->params;
+                    ctx->bindBel(ctx->getBelByLocation(l), iosel_new, PlaceStrength::STRENGTH_FIXED);
+
+                    CellInfo *gpio = iosel->getPort(id_GPIO_IN)->driver.cell;
+                    CellInfo *gpio_new =
+                            create_cell_ptr(gpio->type, ctx->idf("%s$die%d", gpio->name.c_str(ctx), new_die));
+                    gpio_new->params = gpio->params;
+                    ctx->bindBel(ctx->getBelByLocation({l.x, l.y, 0}), gpio_new, PlaceStrength::STRENGTH_FIXED);
+
+                    // Duplicate input connection
+                    gpio_new->connectPort(id_I, gpio->getPort(id_I));
+                    // Connect IOSEL and CPE_IBUF
+                    gpio_new->connectPorts(id_Y, iosel_new, id_GPIO_IN);
+                }
+                uarch->clkin[new_die]->disconnectPort(ctx->idf("CLK%d", i));
+                if (iosel_new->getPort(id_IN1))
+                    uarch->clkin[new_die]->connectPort(ctx->idf("CLK%d", i), iosel_new->getPort(id_IN1));
+                else
+                    iosel_new->connectPorts(id_IN1, uarch->clkin[new_die], ctx->idf("CLK%d", i));
+            }
+        }
+    }
+}
+
+void GateMatePacker::strategy_clk1()
+{
+    log_info("Reuse CLK1 for clock distribution..\n");
+    NetInfo *net = uarch->glbout[0]->getPort(id_GLB0);
+    NetInfo *new_clk1 = ctx->createNet(ctx->id("$clk1$pin"));
+    for (int new_die = 0; new_die < uarch->dies; new_die++) {
+        CellInfo *iosel = create_cell_ptr(id_IOSEL, ctx->idf("$iosel_clk1$die%d", new_die));
+        iosel->setParam(id_DELAY_IBF, Property(1, 16));
+        iosel->setParam(id_INPUT_ENABLE, Property(1, 1));
+        if (new_die == 0) {
+            // On die 0 it should be output as well
+            iosel->setParam(id_DELAY_OBF, Property(1, 16));
+            iosel->setParam(id_OE_ENABLE, Property(1, 1));
+            iosel->setParam(id_OUT_SIGNAL, Property(1, 1));
+            iosel->setParam(id_SLEW, Property(1, 1));
+        }
+
+        BelId bel = ctx->getBelByLocation(uarch->locations[std::make_pair(ctx->id("IO_SB_A7"), new_die)]);
+        ctx->bindBel(bel, iosel, PlaceStrength::STRENGTH_FIXED);
+
+        CellInfo *gpio = create_cell_ptr((new_die ? id_CPE_IBUF : id_CPE_IOBUF), ctx->idf("$clk1$die%d", new_die));
+        Loc loc = ctx->getBelLocation(bel);
+        ctx->bindBel(ctx->getBelByLocation({loc.x, loc.y, 0}), gpio, PlaceStrength::STRENGTH_FIXED);
+
+        uarch->clkin[new_die]->connectPort(id_CLK1, new_clk1);
+        uarch->clkin[new_die]->params[ctx->id("REF1")] = Property(1, 3);
+        uarch->glbout[new_die]->params[ctx->id("GLB1_EN")] = Property(Property::State::S1);
+        uarch->glbout[new_die]->params[ctx->id("GLB1_CFG")] = Property(0, 3);
+        uarch->clkin[new_die]->connectPorts(ctx->id("CLK_REF1"), uarch->glbout[new_die], ctx->id("CLK_REF_OUT1"));
+
+        gpio->connectPorts(id_Y, iosel, id_GPIO_IN);
+
+        if (new_die == 0) {
+            iosel->connectPort(id_OUT1, ctx->getNetByAlias(uarch->global_signals.begin()->first->name));
+            CellInfo *cpe = move_ram_o_fixed(iosel, id_OUT1, loc).first;
+            uarch->ignore.emplace(cpe->name);
+
+            iosel->connectPorts(id_GPIO_OUT, gpio, id_A);
+            iosel->connectPorts(id_GPIO_EN, gpio, id_T);
+            gpio->connectPort(id_IO, new_clk1);
+        } else {
+            gpio->connectPort(id_I, new_clk1);
+        }
+
+        NetInfo *new_signal = ctx->createNet(ctx->idf("%s$die%d", net->name.c_str(ctx), new_die));
+        uarch->glbout[new_die]->connectPort(ctx->id("GLB1"), new_signal);
+        copy_constraint(net, new_signal);
+        uarch->global_mapping.emplace(std::make_pair(net->name, new_die), new_signal);
+        uarch->global_clk_mapping.emplace(std::make_pair(id_CLOCK1, new_die), id_CLOCK2);
+    }
+}
+
+void GateMatePacker::strategy_mirror()
+{
+    log_info("Mirror clocks..\n");
 
     // Save first CLKIN inputs
     std::vector<CellInfo *> clk_iosel(4, nullptr);
@@ -793,8 +903,52 @@ void GateMatePacker::copy_clocks()
                                                         ctx->idf("CLK_REF_OUT%d", i));
             }
         }
+        for (int i = 0; i < 4; i++) {
+            NetInfo *net = uarch->glbout[0]->getPort(ctx->idf("GLB%d", i));
+            if (net) {
+                if (new_die != 0) {
+                    NetInfo *new_signal = ctx->createNet(ctx->idf("%s$die%d", net->name.c_str(ctx), new_die));
+                    uarch->glbout[new_die]->connectPort(ctx->idf("GLB%d", i), new_signal);
+                    copy_constraint(net, new_signal);
+                    uarch->global_mapping.emplace(std::make_pair(net->name, new_die), new_signal);
+                } else {
+                    uarch->global_mapping.emplace(std::make_pair(net->name, new_die), net);
+                }
+            }
+        }
     }
 }
+
+static int clk_config_val(IdString name)
+{
+    switch (name.index) {
+    case id_CLOCK1.index:
+        return 0b00100011;
+    case id_CLOCK2.index:
+        return 0b00110011;
+    case id_CLOCK3.index:
+        return 0b00000011;
+    case id_CLOCK4.index:
+        return 0b00010011;
+    }
+    return 0;
+}
+
+static int ioclk_config_val(IdString name)
+{
+    switch (name.index) {
+    case id_CLOCK1.index:
+        return 0;
+    case id_CLOCK2.index:
+        return 1;
+    case id_CLOCK3.index:
+        return 2;
+    case id_CLOCK4.index:
+        return 3;
+    }
+    return 0;
+}
+
 void GateMatePacker::reassign_clocks()
 {
     if (uarch->dies == 1)
@@ -805,22 +959,55 @@ void GateMatePacker::reassign_clocks()
 
     for (auto &glob : uarch->global_signals) {
         const NetInfo *net = glob.first;
-        int index = glob.second;
-        int drv_die = uarch->tile_extra_data(net->driver.cell->bel.tile)->die;
         auto users = net->users; // make a copy
         int count = 0;
         for (auto &user : users) {
             int cell_die = uarch->tile_extra_data(user.cell->bel.tile)->die;
-            if (cell_die != drv_die) {
+            if (uarch->global_mapping.count(std::make_pair(net->name, cell_die))) {
-                if (!new_bufg[cell_die][index]) {
+                NetInfo *new_net = uarch->global_mapping.at(std::make_pair(net->name, cell_die));
-                    NetInfo *new_signal = ctx->createNet(ctx->idf("%s$die%d", net->name.c_str(ctx), cell_die));
+                if (uarch->ignore.count(user.cell->name))
-                    new_bufg[cell_die][index] = new_signal;
+                    continue;
-                    uarch->glbout[cell_die]->connectPort(ctx->idf("GLB%d", index), new_signal);
+
-                    copy_constraint(net, new_signal);
+                if (new_net == net)
-                }
+                    continue;
+
                 user.cell->disconnectPort(user.port);
-                user.cell->connectPort(user.port, new_bufg[cell_die][index]);
+
+                if (user.port.in(id_CLOCK1, id_CLOCK2, id_CLOCK3, id_CLOCK4) &&
+                    uarch->global_clk_mapping.count(std::make_pair(user.port, cell_die))) {
+                    IdString newPort = uarch->global_clk_mapping.at(std::make_pair(user.port, cell_die));
+                    if (!user.cell->ports.count(newPort))
+                        user.cell->addInput(newPort);
+                    user.cell->connectPort(newPort, new_net);
+
+                    if (user.cell->type == id_RAM) {
+                        int a0_clk = int_or_default(user.cell->params, id_RAM_cfg_forward_a0_clk, 0);
+                        int a1_clk = int_or_default(user.cell->params, id_RAM_cfg_forward_a1_clk, 0);
+                        int b0_clk = int_or_default(user.cell->params, id_RAM_cfg_forward_b0_clk, 0);
+                        int b1_clk = int_or_default(user.cell->params, id_RAM_cfg_forward_b1_clk, 0);
+
+                        if (a0_clk == clk_config_val(user.port))
+                            user.cell->params[id_RAM_cfg_forward_a0_clk] = Property(clk_config_val(newPort), 8);
+                        if (a1_clk == clk_config_val(user.port))
+                            user.cell->params[id_RAM_cfg_forward_a1_clk] = Property(clk_config_val(newPort), 8);
+                        if (b0_clk == clk_config_val(user.port))
+                            user.cell->params[id_RAM_cfg_forward_b0_clk] = Property(clk_config_val(newPort), 8);
+                        if (b1_clk == clk_config_val(user.port))
+                            user.cell->params[id_RAM_cfg_forward_b1_clk] = Property(clk_config_val(newPort), 8);
+                    }
+                    if (user.cell->type == id_IOSEL) {
+                        int in_clk = int_or_default(user.cell->params, id_IN_CLOCK, 0);
+                        int out_clk = int_or_default(user.cell->params, id_OUT_CLOCK, 0);
+                        if (in_clk == ioclk_config_val(user.port))
+                            user.cell->params[id_IN_CLOCK] = Property(ioclk_config_val(newPort), 2);
+                        if (out_clk == ioclk_config_val(user.port))
+                            user.cell->params[id_OUT_CLOCK] = Property(ioclk_config_val(newPort), 2);
+                    }
+                } else
+                    user.cell->connectPort(user.port, new_net);
                 count++;
+            } else {
+                log_error("Global signal '%s' is not available in die %d.\n", net->name.c_str(ctx), cell_die);
             }
         }
         if (count)

himbaechel/uarch/gatemate/pack_io.cc

@@ -376,7 +376,14 @@ void GateMatePacker::pack_io_sel()
                     return true;
                 } else {
                     int index = uarch->global_signals[clk_net];
-                    cell->movePortTo(id_CLK, target, ctx->idf("CLOCK%d", index + 1));
+                    NetInfo *net = target->getPort(ctx->idf("CLOCK%d", index + 1));
+                    if (net) {
+                        if (net != clk_net)
+                            log_error("Not able to connected different CLK signal to cell '%s'.\n",
+                                      cell->name.c_str(ctx));
+                    } else {
+                        cell->movePortTo(id_CLK, target, ctx->idf("CLOCK%d", index + 1));
+                    }
                     target->params[id_OUT_CLOCK] = Property(index, 2);
                 }
             }
@@ -396,7 +403,14 @@ void GateMatePacker::pack_io_sel()
                     target->params[id_SEL_IN_CLOCK] = Property(Property::State::S1);
                 } else {
                     int index = uarch->global_signals[clk_net];
-                    cell->movePortTo(id_CLK, target, ctx->idf("CLOCK%d", index + 1));
+                    NetInfo *net = target->getPort(ctx->idf("CLOCK%d", index + 1));
+                    if (net) {
+                        if (net != clk_net)
+                            log_error("Not able to connected different CLK signal to cell '%s'.\n",
+                                      cell->name.c_str(ctx));
+                    } else {
+                        cell->movePortTo(id_CLK, target, ctx->idf("CLOCK%d", index + 1));
+                    }
                     target->params[id_IN_CLOCK] = Property(index, 2);
                 }
             }