ecp5: LUT permutation support

Signed-off-by: gatecat <gatecat@ds0.me>

7 files changed, 117 insertions, 7 deletions

diff --git a/ecp5/arch.cc b/ecp5/arch.cc
index ec64fb82..95a27682 100644
--- a/ecp5/arch.cc
+++ b/ecp5/arch.cc
@@ -144,6 +144,8 @@ Arch::Arch(ArchArgs args) : args(args)
         n_pips++;
     }
     pip2net.resize(n_pips, nullptr);
+
+    lutperm_allowed.resize(chip_info->width * chip_info->height * 4);
 }
 
 // -----------------------------------------------------------------------
@@ -625,6 +627,8 @@ bool Arch::route()
 {
     std::string router = str_or_default(settings, id("router"), defaultRouter);
 
+    disable_router_lutperm = getCtx()->setting<bool>("arch.disable_router_lutperm", false);
+
     setup_wire_locations();
     route_ecp5_globals(getCtx());
     assignArchInfo();
diff --git a/ecp5/arch.h b/ecp5/arch.h
index e7bf64fe..51a919bb 100644
--- a/ecp5/arch.h
+++ b/ecp5/arch.h
@@ -58,8 +58,15 @@ NPNR_PACKED_STRUCT(struct PipInfoPOD {
     int16_t timing_class;
     int8_t tile_type;
     int8_t pip_type;
+    int16_t lutperm_flags;
+    int16_t padding;
 });
 
+inline bool is_lutperm_pip(int16_t flags) { return flags & 0x4000; }
+inline uint8_t lutperm_lut(int16_t flags) { return (flags >> 4) & 0x7; }
+inline uint8_t lutperm_out(int16_t flags) { return (flags >> 2) & 0x3; }
+inline uint8_t lutperm_in(int16_t flags) { return flags & 0x3; }
+
 NPNR_PACKED_STRUCT(struct PipLocatorPOD {
     LocationPOD rel_loc;
     int32_t index;
@@ -446,6 +453,14 @@ struct Arch : BaseArch<ArchRanges>
     mutable dict<IdStringList, PipId> pip_by_name;
 
     std::vector<CellInfo *> bel_to_cell;
+    enum class LutPermRule
+    {
+        NONE,
+        CARRY,
+        ALL,
+    };
+    std::vector<LutPermRule> lutperm_allowed;
+    bool disable_router_lutperm = false;
 
     // faster replacements for base_pip2net, base_wire2net
     // indexed by get_pip_vecidx()
@@ -509,6 +524,12 @@ struct Arch : BaseArch<ArchRanges>
         return (bel.location.y * chip_info->width + bel.location.x) * max_loc_bels + bel.index;
     }
 
+    int get_slice_index(int x, int y, int slice) const
+    {
+        NPNR_ASSERT(slice >= 0 && slice < 4);
+        return (y * chip_info->width + x) * 4 + slice;
+    }
+
     void bindBel(BelId bel, CellInfo *cell, PlaceStrength strength) override
     {
         NPNR_ASSERT(bel != BelId());
@@ -517,6 +538,11 @@ struct Arch : BaseArch<ArchRanges>
         bel_to_cell[idx] = cell;
         cell->bel = bel;
         cell->belStrength = strength;
+        if (getBelType(bel) == id_TRELLIS_SLICE) {
+            lutperm_allowed.at(get_slice_index(bel.location.x, bel.location.y, getBelLocation(bel).z)) =
+                    (cell->sliceInfo.is_memory ? LutPermRule::NONE
+                                               : (cell->sliceInfo.is_carry ? LutPermRule::CARRY : LutPermRule::ALL));
+        }
         refreshUiBel(bel);
     }
 
@@ -744,11 +770,31 @@ struct Arch : BaseArch<ArchRanges>
         p2n_entry->wires.erase(dst);
         p2n_entry = nullptr;
     }
-    bool checkPipAvail(PipId pip) const override { return getBoundPipNet(pip) == nullptr; }
+    bool is_pip_blocked(PipId pip) const
+    {
+        auto &pip_data = loc_info(pip)->pip_data[pip.index];
+        int lp = pip_data.lutperm_flags;
+        if (is_lutperm_pip(lp)) {
+            if (disable_router_lutperm)
+                return true;
+            auto rule = lutperm_allowed.at(get_slice_index(pip.location.x, pip.location.y, lutperm_lut(lp) / 2));
+            if (rule == LutPermRule::NONE) {
+                // Permutation not allowed
+                return true;
+            } else if (rule == LutPermRule::CARRY) {
+                // Can swap A/B and C/D only
+                int i = lutperm_out(lp), j = lutperm_in(lp);
+                if ((i / 2) != (j / 2))
+                    return true;
+            }
+        }
+        return false;
+    }
+    bool checkPipAvail(PipId pip) const override { return (getBoundPipNet(pip) == nullptr) && !is_pip_blocked(pip); }
     bool checkPipAvailForNet(PipId pip, NetInfo *net) const override
     {
         NetInfo *bound_net = getBoundPipNet(pip);
-        return bound_net == nullptr || bound_net == net;
+        return (bound_net == nullptr || bound_net == net) && !is_pip_blocked(pip);
     }
     NetInfo *getBoundPipNet(PipId pip) const override { return pip2net.at(get_pip_vecidx(pip)); }
 
diff --git a/ecp5/archdefs.h b/ecp5/archdefs.h
index 80e7810c..dd260a3e 100644
--- a/ecp5/archdefs.h
+++ b/ecp5/archdefs.h
@@ -165,6 +165,7 @@ struct ArchCellInfo : BaseClusterInfo
         bool using_dff;
         bool has_l6mux;
         bool is_carry;
+        bool is_memory;
         IdString clk_sig, lsr_sig, clkmux, lsrmux, srmode;
         int sd0, sd1;
     } sliceInfo;
diff --git a/ecp5/bitstream.cc b/ecp5/bitstream.cc
index 338c4f20..de6c711e 100644
--- a/ecp5/bitstream.cc
+++ b/ecp5/bitstream.cc
@@ -522,6 +522,55 @@ static void set_pip(Context *ctx, ChipConfig &cc, PipId pip)
     cc.tiles[tile].add_arc(sink, source);
 }
 
+static unsigned permute_lut(Context *ctx, CellInfo *cell, pool<IdString> &used_phys_pins, int k, unsigned orig_init)
+{
+    std::array<std::vector<unsigned>, 4> phys_to_log;
+    const std::array<IdString, 4> ports{k ? id_A1 : id_A0, k ? id_B1 : id_B0, k ? id_C1 : id_C0, k ? id_D1 : id_D0};
+    for (unsigned i = 0; i < 4; i++) {
+        WireId pin_wire = ctx->getBelPinWire(cell->bel, ports[i]);
+        for (PipId pip : ctx->getPipsUphill(pin_wire)) {
+            if (!ctx->getBoundPipNet(pip))
+                continue;
+            unsigned lp = ctx->loc_info(pip)->pip_data[pip.index].lutperm_flags;
+            if (!is_lutperm_pip(lp)) { // non-permuting
+                phys_to_log[i].push_back(i);
+            } else { // permuting
+                unsigned from_pin = lutperm_in(lp);
+                unsigned to_pin = lutperm_out(lp);
+                NPNR_ASSERT(to_pin == i);
+                phys_to_log[from_pin].push_back(i);
+            }
+        }
+    }
+    for (unsigned i = 0; i < 4; i++)
+        if (!phys_to_log.at(i).empty())
+            used_phys_pins.insert(ports.at(i));
+    if (cell->sliceInfo.is_carry) {
+        // Insert dummy entries to ensure we keep the split between the two halves of a CCU2
+        for (unsigned i = 0; i < 4; i++) {
+            if (!phys_to_log.at(i).empty())
+                continue;
+            for (unsigned j = 2 * (i / 2); j < 2 * ((i / 2) + 1); j++) {
+                if (!ctx->getBoundWireNet(ctx->getBelPinWire(cell->bel, ports[j])))
+                    phys_to_log.at(i).push_back(j);
+            }
+        }
+    }
+    unsigned permuted_init = 0;
+    for (unsigned i = 0; i < 16; i++) {
+        unsigned log_idx = 0;
+        for (unsigned j = 0; j < 4; j++) {
+            if ((i >> j) & 0x1) {
+                for (auto log_pin : phys_to_log[j])
+                    log_idx |= (1 << log_pin);
+            }
+        }
+        if ((orig_init >> log_idx) & 0x1)
+            permuted_init |= (1 << i);
+    }
+    return permuted_init;
+}
+
 static std::vector<bool> parse_config_str(const Property &p, int length)
 {
     std::vector<bool> word;
@@ -787,12 +836,15 @@ void write_bitstream(Context *ctx, std::string base_config_file, std::string tex
         }
         BelId bel = ci->bel;
         if (ci->type == ctx->id("TRELLIS_SLICE")) {
+            pool<IdString> used_phys_pins;
             std::string tname = ctx->get_tile_by_type_loc(bel.location.y, bel.location.x, "PLC2");
             std::string slice = ctx->loc_info(bel)->bel_data[bel.index].name.get();
             int lut0_init = int_or_default(ci->params, ctx->id("LUT0_INITVAL"));
             int lut1_init = int_or_default(ci->params, ctx->id("LUT1_INITVAL"));
-            cc.tiles[tname].add_word(slice + ".K0.INIT", int_to_bitvector(lut0_init, 16));
-            cc.tiles[tname].add_word(slice + ".K1.INIT", int_to_bitvector(lut1_init, 16));
+            cc.tiles[tname].add_word(slice + ".K0.INIT",
+                                     int_to_bitvector(permute_lut(ctx, ci, used_phys_pins, 0, lut0_init), 16));
+            cc.tiles[tname].add_word(slice + ".K1.INIT",
+                                     int_to_bitvector(permute_lut(ctx, ci, used_phys_pins, 1, lut1_init), 16));
             cc.tiles[tname].add_enum(slice + ".MODE", str_or_default(ci->params, ctx->id("MODE"), "LOGIC"));
             cc.tiles[tname].add_enum(slice + ".GSR", str_or_default(ci->params, ctx->id("GSR"), "ENABLED"));
             cc.tiles[tname].add_enum(slice + ".REG0.SD", intstr_or_default(ci->params, ctx->id("REG0_SD"), "0"));
@@ -854,7 +906,7 @@ void write_bitstream(Context *ctx, std::string base_config_file, std::string tex
 
             // Tie unused inputs high
             for (auto input : {id_A0, id_B0, id_C0, id_D0, id_A1, id_B1, id_C1, id_D1}) {
-                if (ci->ports.find(input) == ci->ports.end() || ci->ports.at(input).net == nullptr) {
+                if (!used_phys_pins.count(input)) {
                     cc.tiles[tname].add_enum(slice + "." + input.str(ctx) + "MUX", "1");
                 }
             }
diff --git a/ecp5/main.cc b/ecp5/main.cc
index f3861149..1864548b 100644
--- a/ecp5/main.cc
+++ b/ecp5/main.cc
@@ -85,6 +85,7 @@ 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()("disable-router-lutperm", "don't allow the router to permute LUT inputs");
 
     return specific;
 }
@@ -255,6 +256,8 @@ std::unique_ptr<Context> ECP5CommandHandler::createContext(dict<std::string, Pro
     ctx->settings[ctx->id("arch.speed")] = speedString(ctx->archArgs().speed);
     if (vm.count("out-of-context"))
         ctx->settings[ctx->id("arch.ooc")] = 1;
+    if (vm.count("disable-router-lutperm"))
+        ctx->settings[ctx->id("arch.disable_router_lutperm")] = 1;
     return ctx;
 }
 
diff --git a/ecp5/pack.cc b/ecp5/pack.cc
index cbf882a8..85d92336 100644
--- a/ecp5/pack.cc
+++ b/ecp5/pack.cc
@@ -3253,7 +3253,9 @@ void Arch::assignArchInfo()
             ci->sliceInfo.clkmux = id(str_or_default(ci->params, id_CLKMUX, "CLK"));
             ci->sliceInfo.lsrmux = id(str_or_default(ci->params, id_LSRMUX, "LSR"));
             ci->sliceInfo.srmode = id(str_or_default(ci->params, id_SRMODE, "LSR_OVER_CE"));
-            ci->sliceInfo.is_carry = str_or_default(ci->params, id("MODE"), "LOGIC") == "CCU2";
+            std::string mode = str_or_default(ci->params, id("MODE"), "LOGIC");
+            ci->sliceInfo.is_carry = (mode == "CCU2");
+            ci->sliceInfo.is_memory = (mode == "DPRAM" || mode == "RAMW");
             ci->sliceInfo.sd0 = std::stoi(str_or_default(ci->params, id("REG0_SD"), "0"));
             ci->sliceInfo.sd1 = std::stoi(str_or_default(ci->params, id("REG1_SD"), "0"));
             ci->sliceInfo.has_l6mux = false;
diff --git a/ecp5/trellis_import.py b/ecp5/trellis_import.py
index 2e76fb74..a586db7b 100755
--- a/ecp5/trellis_import.py
+++ b/ecp5/trellis_import.py
@@ -426,6 +426,8 @@ def write_database(dev_name, chip, ddrg, endianness):
                 if cls == 1 and "PCS" in snk_name or "DCU" in snk_name or "DCU" in src_name:
                    cls = 2
                 bba.u8(cls, "pip_type")
+                bba.u16(arc.lutperm_flags, "lutperm_flags")
+                bba.u16(0, "padding")
         if len(loctype.wires) > 0:
             for wire_idx in range(len(loctype.wires)):
                 wire = loctype.wires[wire_idx]
@@ -623,7 +625,7 @@ def main():
     # print("Initialising chip...")
     chip = pytrellis.Chip(dev_names[args.device])
     # print("Building routing graph...")
-    ddrg = pytrellis.make_dedup_chipdb(chip)
+    ddrg = pytrellis.make_dedup_chipdb(chip, include_lutperm_pips=True)
     max_row = chip.get_max_row()
     max_col = chip.get_max_col()
     process_timing_data()