Gowin. Add fix for Single Port BSRAM

Add description of BSRAM harness

In some cases, Gowin IDE adds a number of LUTs and DFFs to the BSRAM. Here we are trying to add similar elements.

More details with pictures: https://github.com/YosysHQ/apicula/blob/master/doc/bsram-fix.md

Signed-off-by: YRabbit <rabbit@yrabbit.cyou>

himbaechel/uarch/gowin/constids.inc

@@ -890,6 +890,8 @@ X(WRE)
 
 // BSRAM
 X(BSRAM_SUBTYPE)
+X(WRITE_MODE)
+X(READ_MODE)
 X(BIT_WIDTH)
 X(BIT_WIDTH_0)
 X(BIT_WIDTH_1)

himbaechel/uarch/gowin/gowin.cc

@@ -259,6 +259,7 @@ void GowinImpl::adjust_dsp_pin_mapping(void)
 void GowinImpl::prePlace() { assign_cell_info(); }
 void GowinImpl::postPlace()
 {
+    gwu.have_SP32();
     if (ctx->debug) {
         log_info("================== Final Placement ===================\n");
         for (auto &cell : ctx->cells) {

himbaechel/uarch/gowin/gowin.h

@@ -95,7 +95,8 @@ NPNR_PACKED_STRUCT(struct Extra_chip_data_POD {
     Bottom_io_POD bottom_io;
     RelSlice<IdString> diff_io_types;
     // chip flags
-    static constexpr int32_t HAS_SP32 = 0;
+    static constexpr int32_t HAS_SP32 = 1;
+    static constexpr int32_t NEED_SP_FIX = 2;
 });
 
 } // namespace

himbaechel/uarch/gowin/gowin_arch_gen.py

@@ -16,6 +16,7 @@ BEL_FLAG_SIMPLE_IO = 0x100
 
 # Chip flags
 CHIP_HAS_SP32    = 0x1
+CHIP_NEED_SP_FIX = 0x2
 
 # Z of the bels
 # sync with C++ part!
@@ -1008,8 +1009,15 @@ def main():
         db = pickle.load(f)
 
     chip_flags = 0;
+    # XXX compatibility
+    if not hasattr(db, "chip_flags"):
         if device not in {"GW1NS-4", "GW1N-9"}:
-        chip_flags &= CHIP_HAS_SP32;
+            chip_flags |= CHIP_HAS_SP32;
+    else:
+        if "HAS_SP32" in db.chip_flags:
+            chip_flags |= CHIP_HAS_SP32;
+        if "NEED_SP_FIX" in db.chip_flags:
+            chip_flags |= CHIP_NEED_SP_FIX;
 
     X = db.cols;
     Y = db.rows;

himbaechel/uarch/gowin/gowin_utils.cc

@@ -118,6 +118,12 @@ bool GowinUtils::have_SP32(void)
     return extra->chip_flags & Extra_chip_data_POD::HAS_SP32;
 }
 
+bool GowinUtils::need_SP_fix(void)
+{
+    const Extra_chip_data_POD *extra = reinterpret_cast<const Extra_chip_data_POD *>(ctx->chip_info->extra_data.get());
+    return extra->chip_flags & Extra_chip_data_POD::NEED_SP_FIX;
+}
+
 std::unique_ptr<CellInfo> GowinUtils::create_cell(IdString name, IdString type)
 {
     NPNR_ASSERT(!ctx->cells.count(name));

himbaechel/uarch/gowin/gowin_utils.h

@@ -34,6 +34,10 @@ struct GowinUtils
     Loc get_pair_iologic_bel(Loc loc);
     BelId get_io_bel_from_iologic(BelId bel);
 
+    // BSRAM
+    bool have_SP32(void);
+    bool need_SP_fix(void);
+
     // DSP
     inline int get_dsp_18_z(int z) const { return z & (~3); }
     inline int get_dsp_9_idx(int z) const { return z & 3; }
@@ -81,9 +85,6 @@ struct GowinUtils
         return is_global_wire(ctx->getPipSrcWire(pip)) || is_global_wire(ctx->getPipDstWire(pip));
     }
 
-    // chip dependent
-    bool have_SP32(void);
-
     // make cell but do not include it in the list of chip cells.
     std::unique_ptr<CellInfo> create_cell(IdString name, IdString type);
 };

himbaechel/uarch/gowin/pack.cc

@@ -508,13 +508,13 @@ struct GowinPacker
         make_iob_nets(*out_iob);
     }
 
-    IdString create_aux_name(IdString main_name, int idx = 0)
+    IdString create_aux_name(IdString main_name, int idx = 0, const char *str_suffix = "_aux$")
     {
         std::string sfx("");
         if (idx) {
             sfx = std::to_string(idx);
         }
-        return ctx->id(main_name.str(ctx) + std::string("_aux$") + sfx);
+        return ctx->id(main_name.str(ctx) + std::string(str_suffix) + sfx);
     }
 
     BelId get_aux_iologic_bel(const CellInfo &ci)
@@ -1331,22 +1331,103 @@ struct GowinPacker
         }
     }
 
-    // If the memory is controlled by the CE, then it is logical for the OCE to
+    // Analysis of the images generated by the IDE showed that some components
-    // also respond to this signal, unless the OCE is controlled separately.
+    // are being added at the input and output of the BSRAM.  Two LUTs are
-    void bsram_handle_sp_oce(CellInfo *ci, IdString ce_pin, IdString oce_pin)
+    // added on the WRE and CE inputs (strangely, OCE is not affected), a pair
+    // of LUT-DFFs on each DO output, and one or two flipflops of different
+    // types in the auxiliary network.
+    // The semantics of these additions are unclear, but we can replicate this behavior.
+    //  Fix BSRAM in single port mode.
+    void bsram_fix_sp(CellInfo *ci, std::vector<std::unique_ptr<CellInfo>> &new_cells)
     {
-        const NetInfo *net = ci->getPort(oce_pin);
+        int bit_width = ci->params.at(id_BIT_WIDTH).as_int64();
-        NPNR_ASSERT(ci->getPort(ce_pin) != nullptr);
+
-        if (net == nullptr || net->name == ctx->id("$PACKER_VCC") || net->name == ctx->id("$PACKER_GND")) {
+        // create WRE LUT
+        auto wre_lut_cell = gwu.create_cell(create_aux_name(ci->name, 0, "_wre_lut$"), id_LUT4);
+        CellInfo *wre_lut = wre_lut_cell.get();
+        wre_lut->setParam(id_INIT, 0x8888);
+        ci->movePortTo(id_CE, wre_lut, id_I0);
+        ci->movePortTo(id_WRE, wre_lut, id_I1);
+        wre_lut->addOutput(id_F);
+        ci->connectPorts(id_WRE, wre_lut, id_F);
+
+        // create CE LUT
+        auto ce_lut_cell = gwu.create_cell(create_aux_name(ci->name, 0, "_ce_lut$"), id_LUT4);
+        CellInfo *ce_lut = ce_lut_cell.get();
+        ce_lut->setParam(id_INIT, 0xeeee);
+        wre_lut->copyPortTo(id_I0, ce_lut, id_I0);
+        wre_lut->copyPortTo(id_I1, ce_lut, id_I1);
+        ce_lut->addOutput(id_F);
+        ci->connectPorts(id_CE, ce_lut, id_F);
+
+        // create ce reg
+        int write_mode = ci->params.at(id_WRITE_MODE).as_int64();
+        IdString dff_type = write_mode ? id_DFF : id_DFFR;
+        auto ce_pre_dff_cell = gwu.create_cell(create_aux_name(ci->name, 0, "_ce_pre_dff$"), dff_type);
+        CellInfo *ce_pre_dff = ce_pre_dff_cell.get();
+        ce_pre_dff->addInput(id_D);
+        ce_lut->copyPortTo(id_I0, ce_pre_dff, id_D);
+        ci->copyPortTo(id_CLK, ce_pre_dff, id_CLK);
+        if (dff_type == id_DFFR) {
+            wre_lut->copyPortTo(id_I1, ce_pre_dff, id_RESET);
+        }
+        ce_pre_dff->addOutput(id_Q);
+
+        // new ce src with Q pin (used by output pins, not by BSRAM itself)
+        CellInfo *new_ce_net_src = ce_pre_dff;
+
+        // add delay register in pipeline mode
+        int read_mode = ci->params.at(id_READ_MODE).as_int64();
+        if (read_mode) {
+            auto ce_pipe_dff_cell = gwu.create_cell(create_aux_name(ci->name, 0, "_ce_pipe_dff$"), id_DFF);
+            new_cells.push_back(std::move(ce_pipe_dff_cell));
+            CellInfo *ce_pipe_dff = new_cells.back().get();
+            ce_pipe_dff->addInput(id_D);
+            new_ce_net_src->connectPorts(id_Q, ce_pipe_dff, id_D);
+            ci->copyPortTo(id_CLK, ce_pipe_dff, id_CLK);
+            ce_pipe_dff->addOutput(id_Q);
+            new_ce_net_src = ce_pipe_dff;
+        }
+
+        // used outputs of the BSRAM convert to cached
+        for (int i = 0; i < bit_width; ++i) {
+            IdString do_name = ctx->idf("DO[%d]", i);
+            const NetInfo *net = ci->getPort(do_name);
             if (net != nullptr) {
-                ci->disconnectPort(oce_pin);
+                if (net->users.empty()) {
+                    ci->disconnectPort(do_name);
+                    continue;
                 }
-            ci->copyPortTo(ce_pin, ci, oce_pin);
+                // create cache lut
+                auto cache_lut_cell = gwu.create_cell(create_aux_name(ci->name, i, "_cache_lut$"), id_LUT4);
+                CellInfo *cache_lut = cache_lut_cell.get();
+                cache_lut->setParam(id_INIT, 0xcaca);
+                cache_lut->addInput(id_I0);
+                cache_lut->addInput(id_I1);
+                cache_lut->addInput(id_I2);
+                ci->movePortTo(do_name, cache_lut, id_F);
+                ci->connectPorts(do_name, cache_lut, id_I1);
+                new_ce_net_src->connectPorts(id_Q, cache_lut, id_I2);
+
+                // create cache DFF
+                auto cache_dff_cell = gwu.create_cell(create_aux_name(ci->name, i, "_cache_dff$"), id_DFFE);
+                CellInfo *cache_dff = cache_dff_cell.get();
+                cache_dff->addInput(id_CE);
+                cache_dff->addInput(id_D);
+                ci->copyPortTo(id_CLK, cache_dff, id_CLK);
+                new_ce_net_src->connectPorts(id_Q, cache_dff, id_CE);
+                cache_lut->copyPortTo(id_I1, cache_dff, id_D);
+                cache_dff->addOutput(id_Q);
+                cache_dff->connectPorts(id_Q, cache_lut, id_I0);
+
+                new_cells.push_back(std::move(cache_lut_cell));
+                new_cells.push_back(std::move(cache_dff_cell));
             }
-        if (ctx->verbose) {
-            log_info("%s: %s = %s = %s\n", ctx->nameOf(ci), ce_pin.c_str(ctx), oce_pin.c_str(ctx),
-                     ctx->nameOf(ci->getPort(oce_pin)));
         }
+
+        new_cells.push_back(std::move(wre_lut_cell));
+        new_cells.push_back(std::move(ce_lut_cell));
+        new_cells.push_back(std::move(ce_pre_dff_cell));
     }
 
     void pack_ROM(CellInfo *ci)
@@ -1432,7 +1513,6 @@ struct GowinPacker
             ci->renamePort(ctx->idf("BLKSELB[%d]", i), ctx->idf("BLKSELB%d", i));
         }
 
-        bsram_handle_sp_oce(ci, id_CEB, id_OCE);
         ci->copyPortTo(id_OCE, ci, id_OCEB);
 
         // Port A
@@ -1565,7 +1645,14 @@ struct GowinPacker
         }
 
         int bit_width = ci->params.at(id_BIT_WIDTH).as_int64();
-        bsram_handle_sp_oce(ci, id_CE, id_OCE);
+
+        // XXX strange WRE<->CE relations
+        // Gowin IDE adds two LUTs to the WRE and CE signals. The logic is
+        // unclear, but without them effects occur. Perhaps this is a
+        // correction of some BSRAM defects.
+        if (gwu.need_SP_fix()) {
+            bsram_fix_sp(ci, new_cells);
+        }
 
         // XXX UG285-1.3.6_E Gowin BSRAM & SSRAM User Guide:
         // For GW1N-9/GW1NR-9/GW1NS-4 series, 32/36-bit SP/SPX9 is divided into two
@@ -2726,9 +2813,6 @@ struct GowinPacker
         pack_gsr();
         ctx->check();
 
-        pack_inv();
-        ctx->check();
-
         pack_wideluts();
         ctx->check();
 
@@ -2751,6 +2835,9 @@ struct GowinPacker
         pack_dsp();
         ctx->check();
 
+        pack_inv();
+        ctx->check();
+
         pack_buffered_nets();
 
         ctx->fixupHierarchy();