ecp5: Add support for promotion of LSRs to global network

Signed-off-by: David Shah <dave@ds0.me>

2 files changed, 74 insertions, 5 deletions

diff --git a/ecp5/globals.cc b/ecp5/globals.cc
index da2ba8f0..4bfc2b16 100644
--- a/ecp5/globals.cc
+++ b/ecp5/globals.cc
@@ -62,7 +62,12 @@ class Ecp5GlobalRouter
             return true;
         return false;
     }
-
+    bool is_lsr_port(const PortRef &user)
+    {
+        if (user.cell->type == id_TRELLIS_SLICE && (user.port == id_LSR))
+            return true;
+        return false;
+    }
     std::vector<NetInfo *> get_clocks()
     {
         std::unordered_map<IdString, int> clockCount;
@@ -102,6 +107,32 @@ class Ecp5GlobalRouter
         return clocks;
     }
 
+    std::vector<NetInfo *> get_globalworthy_lsrs()
+    {
+        // Fanout threshold for LSRs is high (300), and we only promote up to two of them, in the nature of being
+        // conservative given the cost of a global
+        std::unordered_map<IdString, int> lsrCount;
+        for (auto &net : ctx->nets) {
+            NetInfo *ni = net.second.get();
+            lsrCount[ni->name] = 0;
+            for (const auto &user : ni->users)
+                if (is_lsr_port(user))
+                    lsrCount[ni->name]++;
+        }
+        // DCCAs must always drive globals
+        std::vector<NetInfo *> lsrs;
+        while (lsrs.size() < 2) {
+            auto max = std::max_element(lsrCount.begin(), lsrCount.end(),
+                                        [](const decltype(lsrCount)::value_type &a,
+                                           const decltype(lsrCount)::value_type &b) { return a.second < b.second; });
+            if (max == lsrCount.end() || max->second < 300)
+                break;
+            lsrs.push_back(ctx->nets.at(max->first).get());
+            lsrCount.erase(max->first);
+        }
+        return lsrs;
+    }
+
     PipId find_tap_pip(WireId tile_glb)
     {
         std::string wireName = ctx->getWireBasename(tile_glb).str(ctx);
@@ -140,6 +171,9 @@ class Ecp5GlobalRouter
         upstream.push(userWire);
         bool already_routed = false;
         WireId next;
+
+        IdString id_lsr1 = ctx->id("LSR1");
+
         // Search back from the pin until we reach the global network
         while (true) {
             next = upstream.front();
@@ -155,6 +189,11 @@ class Ecp5GlobalRouter
                 globalWire = next;
                 break;
             }
+
+            // Forbid to use LSR1 for non-WRE LSRs; as this is reserved for WRE
+            if (ctx->getWireBasename(next) == id_lsr1 && user.port != id_WRE)
+                continue;
+
             if (ctx->checkWireAvail(next)) {
                 for (auto pip : ctx->getPipsUphill(next)) {
                     WireId src = ctx->getPipSrcWire(pip);
@@ -464,16 +503,33 @@ class Ecp5GlobalRouter
             else
                 insert_dcc(clock);
         }
+        if (clocks.size() < 12 && !bool_or_default(ctx->settings, ctx->id("arch.nolsrglobal"), false)) {
+            // Promote up to two set/resets to global
+            auto lsrs = get_globalworthy_lsrs();
+            for (auto lsr : lsrs) {
+                bool is_noglobal = bool_or_default(lsr->attrs, ctx->id("noglobal"), false);
+                if (is_noglobal)
+                    continue;
+                log_info("    promoting LSR net %s to global network\n", lsr->name.c_str(ctx));
+                if (is_ooc) // Don't actually do anything in OOC mode, global routing will be done in the full design
+                    lsr->is_global = true;
+                else
+                    insert_dcc(lsr);
+            }
+        }
     }
 
     void route_globals()
     {
         log_info("Routing globals...\n");
-        std::set<int> all_globals, fab_globals;
+        std::set<int> all_globals, fab_globals, lsr_globals;
         for (int i = 0; i < 16; i++) {
             all_globals.insert(i);
-            if (i < 8)
+            if (i < 8) {
+                if (i >= 4)
+                    lsr_globals.insert(i);
                 fab_globals.insert(i);
+            }
         }
         std::vector<std::pair<PortRef *, int>> toroute;
         std::unordered_map<int, NetInfo *> clocks;
@@ -484,9 +540,18 @@ class Ecp5GlobalRouter
                 NPNR_ASSERT(clock != nullptr);
                 bool drives_fabric = std::any_of(clock->users.begin(), clock->users.end(),
                                                  [this](const PortRef &port) { return !is_clock_port(port); });
-
+                bool drives_lsr = std::any_of(clock->users.begin(), clock->users.end(),
+                                              [this](const PortRef &port) { return is_lsr_port(port); });
                 int glbid;
-                if (drives_fabric) {
+                if (drives_lsr) {
+                    if (lsr_globals.empty()) {
+                        // Route LSR through fabric (suboptimal)
+                        NPNR_ASSERT(!fab_globals.empty());
+                        glbid = *(fab_globals.begin());
+                    } else {
+                        glbid = *(lsr_globals.begin());
+                    }
+                } else if (drives_fabric) {
                     if (fab_globals.empty())
                         continue;
                     glbid = *(fab_globals.begin());
@@ -495,6 +560,7 @@ class Ecp5GlobalRouter
                 }
                 all_globals.erase(glbid);
                 fab_globals.erase(glbid);
+                lsr_globals.erase(glbid);
 
                 log_info("    routing clock net %s using global %d\n", clock->name.c_str(ctx), glbid);
                 bool routed = route_onto_global(clock, glbid);
diff --git a/ecp5/main.cc b/ecp5/main.cc
index 24a98df4..d1297a95 100644
--- a/ecp5/main.cc
+++ b/ecp5/main.cc
@@ -74,6 +74,7 @@ po::options_description ECP5CommandHandler::getArchOptions()
     specific.add_options()(
             "out-of-context",
             "disable IO buffer insertion and global promotion/routing, for building pre-routed blocks (experimental)");
+    specific.add_options()("no-lsr-globals", "don't promote local set/reset (LSR) signals to global networks");
 
     return specific;
 }
@@ -239,6 +240,8 @@ std::unique_ptr<Context> ECP5CommandHandler::createContext(std::unordered_map<st
         ctx->settings[ctx->id(val.first)] = val.second;
     ctx->settings[ctx->id("arch.package")] = ctx->archArgs().package;
     ctx->settings[ctx->id("arch.speed")] = speedString(ctx->archArgs().speed);
+    if (vm.count("no-lsr-globals"))
+        ctx->settings[ctx->id("arch.nolsrglobal")] = 1;
     if (vm.count("out-of-context"))
         ctx->settings[ctx->id("arch.ooc")] = 1;
     return ctx;