nexus: Promote CE/LSR globals

Signed-off-by: gatecat <[email protected]>

Author: gatecat

nexus/arch.cc

@@ -178,6 +178,7 @@ Arch::Arch(ArchArgs args) : args(args)
             // some tweaks to accomodate properly
             if (getBelType(bel) != id_DCC)
                 continue;
+            disabled_bels.insert(bel);
             WireId w = getBelPinWire(bel, id_CLKI);
             for (auto pip : getPipsUphill(w))
                 disabled_pips.insert(pip);

nexus/arch.h

@@ -785,9 +785,10 @@ enum CellPinStyle
 
     PINDEF_MASK = 0x30,
 
-    PINGLB_CLK = 0x100, // pin is a 'clock' for global purposes
+    PINGLB_CLK = 0x100,    // pin is a 'clock' for global purposes
+    PINGLB_CE_LSR = 0x200, // pin is CE/LSR for global purposes
 
-    PINGLB_MASK = 0x100,
+    PINGLB_MASK = 0x300,
 
     PINBIT_GATED = 0x1000,  // pin must be enabled in bitstream if used
     PINBIT_1 = 0x2000,      // pin has an explicit bit that must be set if tied to 1
@@ -796,8 +797,8 @@ enum CellPinStyle
     PINSTYLE_NONE = 0x0000,      // default
     PINSTYLE_CIB = 0x4012,       // 'CIB' signal, floats high but explicitly zeroed if not used
     PINSTYLE_CLK = 0x0107,       // CLK type signal, invertible and defaults to disconnected
-    PINSTYLE_CE = 0x0027,        // CE type signal, invertible and defaults to enabled
-    PINSTYLE_LSR = 0x0017,       // LSR type signal, invertible and defaults to not reset
+    PINSTYLE_CE = 0x0227,        // CE type signal, invertible and defaults to enabled
+    PINSTYLE_LSR = 0x0217,       // LSR type signal, invertible and defaults to not reset
     PINSTYLE_DEDI = 0x0000,      // dedicated signals, leave alone
     PINSTYLE_PU = 0x4022,        // signals that float high and default high
     PINSTYLE_PU_NONCIB = 0x0022, // signals that float high and default high
@@ -920,6 +921,7 @@ struct Arch : BaseArch<ArchRanges>
     dict<IdString, int> id_to_x, id_to_y;
 
     pool<PipId> disabled_pips;
+    pool<BelId> disabled_bels;
 
     // -------------------------------------------------
 

nexus/arch_place.cc

@@ -105,6 +105,10 @@ bool Arch::isBelLocationValid(BelId bel) const
         else
             return nexus_logic_tile_valid(*lts);
     } else {
+        if (getBelType(bel) == id_DCC) {
+            if (getBoundBelCell(bel) != nullptr && disabled_bels.count(bel))
+                return false;
+        }
         return true;
     }
 }

nexus/global.cc

@@ -107,6 +107,10 @@ struct NexusGlobalRouter
         while (!visit.empty() && (iter++ < iter_limit)) {
             WireId cursor = visit.front();
             visit.pop();
+            if (ctx->getBoundWireNet(cursor) == net) {
+                src = cursor;
+                break;
+            }
             // Search uphill pips
             for (PipId pip : ctx->getPipsUphill(cursor)) {
                 // Skip pip if unavailable, and not because it's already used for this net

nexus/pack.cc

@@ -852,7 +852,9 @@ struct NexusPacker
     void promote_globals()
     {
         std::vector<std::pair<int, IdString>> clk_fanout;
-        int available_globals = 16;
+        std::vector<std::pair<int, IdString>> ce_lsr_fanout;
+        // TODO: if we are more cunning about placement, the real limits are 16 per device half and 8 per row segment
+        int available_globals = 8;
         for (auto &net : ctx->nets) {
             NetInfo *ni = net.second.get();
             // Skip undriven nets; and nets that are already global
@@ -863,26 +865,41 @@ struct NexusPacker
                 continue;
             }
             // Count the number of clock ports
-            int clk_count = 0;
+            int clk_count = 0, ce_lsr_count = 0;
             for (const auto &usr : ni->users) {
                 auto port_style = ctx->get_cell_pin_style(usr.cell, usr.port);
                 if (port_style & PINGLB_CLK)
                     ++clk_count;
+                if (port_style & PINGLB_CE_LSR)
+                    ++ce_lsr_count;
             }
             if (clk_count > 0)
                 clk_fanout.emplace_back(clk_count, ni->name);
+            if (ce_lsr_count > 150)
+                ce_lsr_fanout.emplace_back(ce_lsr_count, ni->name);
         }
         if (available_globals <= 0)
             return;
         // Sort clocks by max fanout
         std::sort(clk_fanout.begin(), clk_fanout.end(), std::greater<std::pair<int, IdString>>());
+        std::sort(ce_lsr_fanout.begin(), ce_lsr_fanout.end(), std::greater<std::pair<int, IdString>>());
+
         log_info("Promoting globals...\n");
         // Promote the N highest fanout clocks
         for (size_t i = 0; i < std::min<size_t>(clk_fanout.size(), available_globals); i++) {
             NetInfo *net = ctx->nets.at(clk_fanout.at(i).second).get();
             log_info("     promoting clock net '%s'\n", ctx->nameOf(net));
             insert_buffer(net, id_DCC, "glb_clk", id_CLKI, id_CLKO, [](const PortRef &port) { return true; });
         }
+        // Limit to 1 LSR/CE to avoid routeability issues
+        int available_ce_lsr_globals = std::min<int>(available_globals, 1);
+        for (size_t i = 0; i < std::min<size_t>(ce_lsr_fanout.size(), available_ce_lsr_globals); i++) {
+            NetInfo *net = ctx->nets.at(ce_lsr_fanout.at(i).second).get();
+            log_info("     promoting CE/LSR net '%s'\n", ctx->nameOf(net));
+            insert_buffer(net, id_DCC, "glb_ce_lsr", id_CLKI, id_CLKO, [this](const PortRef &port) {
+                return ctx->get_cell_pin_style(port.cell, port.port) & PINGLB_CE_LSR;
+            });
+        }
     }
 
     // Place certain global cells