Initial XLUT support

himbaechel/uarch/ng-ultra/extra_data.h

@@ -42,6 +42,10 @@ enum ClusterPlacement
     PLACE_CY_FE2,
     PLACE_CY_FE3,
     PLACE_CY_FE4,
+    PLACE_XLUT_FE1,
+    PLACE_XLUT_FE2,
+    PLACE_XLUT_FE3,
+    PLACE_XLUT_FE4,
     PLACE_XRF_I1,
     PLACE_XRF_I2,
     PLACE_XRF_I3,

himbaechel/uarch/ng-ultra/gen/arch_gen.py

@@ -376,6 +376,10 @@ def create_tile_types(ch: Chip, bels, bel_pins, crossbars, interconnects, muxes,
                     # LUT bypass
                     by = tt.create_pip(f"{name}.I1",f"{name}.LO","BYPASS")
                     by.extra_data = PipExtraData(ch.strs.id(name),PIP_EXTRA_BYPASS,1,0)
+                elif (tile_type.startswith("TILE") and bel=="XLUT"):
+                    for out in ["G1","G2","G3","G4"]:
+                        vi = tt.create_pip(f"{name}.J",f"{name}.{out}","Virtual")
+                        vi.extra_data = PipExtraData(ch.strs.id(name),PIP_EXTRA_VIRTUAL,0,0)
                 elif (tile_type.startswith("TILE") and bel=="CY"):
                     # matrix of each input to each output combination
                     # crossbar but use mux placeholder for convenience
@@ -384,16 +388,16 @@ def create_tile_types(ch: Chip, bels, bel_pins, crossbars, interconnects, muxes,
                             pd = tt.create_pip(f"{name}."+inp,f"{name}."+out,"MATRIX_PIP")
                             pd.extra_data = PipExtraData(ch.strs.id(f"{name}."+inp),PIP_EXTRA_MUX,int(inp[1:])-1,int(out[1:])-1)
 
-                #elif (tile_type.startswith("CKG") and bel=="WFG"):
+                elif (tile_type.startswith("CKG") and bel=="WFG"):
-                #    by = tt.create_pip(f"{name}.ZI",f"{name}.ZO","BYPASS")
+                    by = tt.create_pip(f"{name}.ZI",f"{name}.ZO","BYPASS")
-                #    by.extra_data = PipExtraData(ch.strs.id(name),PIP_EXTRA_BYPASS,0,0)
+                    by.extra_data = PipExtraData(ch.strs.id(name),PIP_EXTRA_BYPASS,0,0)
-                #elif (tile_type.startswith("TUBE") and bel=="GCK"):
+                elif (tile_type.startswith("TUBE") and bel=="GCK"):
-                #    # 20 clock signals comming to 20 GCK, SI1 is bypass
+                    # 20 clock signals comming to 20 GCK, SI1 is bypass
-                #    by = tt.create_pip(f"{name}.SI1",f"{name}.SO","BYPASS")
+                    by = tt.create_pip(f"{name}.SI1",f"{name}.SO","BYPASS")
-                #    by.extra_data = PipExtraData(ch.strs.id(name),PIP_EXTRA_BYPASS,0,0)
+                    by.extra_data = PipExtraData(ch.strs.id(name),PIP_EXTRA_BYPASS,0,0)
-                #    # there are CMD signals that can be bypassed as well
+                    # there are CMD signals that can be bypassed as well
-                #    by = tt.create_pip(f"{name}.CMD",f"{name}.SO","BYPASS")
+                    by = tt.create_pip(f"{name}.CMD",f"{name}.SO","BYPASS")
-                #    by.extra_data = PipExtraData(ch.strs.id(name),PIP_EXTRA_BYPASS,1,0)
+                    by.extra_data = PipExtraData(ch.strs.id(name),PIP_EXTRA_BYPASS,1,0)
                 
 
             # Add LUT permutation

himbaechel/uarch/ng-ultra/ng_ultra.cc

@@ -512,6 +512,15 @@ Loc getCYFE(Loc root, int pos)
     return result;
 }
 
+Loc getXLUTFE(Loc root, int pos)
+{
+    Loc result;
+    result.x = root.x;
+    result.y = root.y;
+    result.z = root.z - BEL_XLUT_Z + 8 * pos;
+    return result;
+}
+
 Loc getXRFFE(Loc root, int pos)
 {
     static const std::vector<Loc> map = 
@@ -591,7 +600,6 @@ Loc getXRFFE(Loc root, int pos)
     return result;
 }
 
-
 Loc getCDCFE(Loc root, int pos)
 {
     static const std::vector<Loc> cdc1 = 
@@ -963,10 +971,8 @@ bool NgUltraImpl::getChildPlacement(const BaseClusterInfo *cluster, Loc root_loc
         Loc child_loc = if_using_basecluster<Loc>(child, [&](const BaseClusterInfo *child) {
             switch(child->constr_z) {
                 case PLACE_CY_CHAIN : { Loc l = getNextLocInCYChain(prev); prev = l; return l; }
-                case PLACE_CY_FE1: return getCYFE(root_loc,0);
+                case PLACE_CY_FE1 ... PLACE_CY_FE4: return getCYFE(root_loc, child->constr_z - PLACE_CY_FE1 );
-                case PLACE_CY_FE2: return getCYFE(root_loc,1);
+                case PLACE_XLUT_FE1 ... PLACE_XLUT_FE4: return getXLUTFE(root_loc, child->constr_z - PLACE_XLUT_FE1 );
-                case PLACE_CY_FE3: return getCYFE(root_loc,2);
-                case PLACE_CY_FE4: return getCYFE(root_loc,3);
                 case PLACE_XRF_I1 ... PLACE_XRF_WEA:
                                     return getXRFFE(root_loc, child->constr_z - PLACE_XRF_I1 );
                 case PLACE_CDC_AI1 ... PLACE_CDC_DDRSTI:

himbaechel/uarch/ng-ultra/pack.cc

@@ -294,6 +294,63 @@ void NgUltraPacker::bind_attr_loc(CellInfo *cell, dict<IdString, Property> *attr
     }
 }
 
+void NgUltraPacker::pack_xluts(void)
+{
+    log_info("Pack XLUTs...\n");
+    int lut_only = 0;//, lut_and_ff = 0;
+    for (auto &cell : ctx->cells) {
+        CellInfo &ci = *cell.second;
+        if (!ci.type.in(id_NX_LUT))
+            continue;
+        if (!ci.params.count(id_lut_table))
+            log_error("Cell '%s' missing lut_table\n", ci.name.c_str(ctx));
+
+        if (ci.cluster!=ClusterId())
+            continue;
+        CellInfo *lut[4];
+        if (!ci.getPort(id_I1))
+            continue;
+        if (!ci.getPort(id_I2))
+            continue;
+        if (!ci.getPort(id_I3))
+            continue;
+        if (!ci.getPort(id_I4))
+            continue;
+        lut[0] = net_driven_by(ctx, ci.getPort(id_I1), is_lut, id_O);
+        if (!lut[0] || ci.getPort(id_I1)->users.entries()!=1)
+            continue;
+        lut[1] = net_driven_by(ctx, ci.getPort(id_I2), is_lut, id_O);
+        if (!lut[1] || ci.getPort(id_I2)->users.entries()!=1)
+            continue;
+        lut[2] = net_driven_by(ctx, ci.getPort(id_I3), is_lut, id_O);
+        if (!lut[2] || ci.getPort(id_I3)->users.entries()!=1)
+            continue;
+        lut[3] = net_driven_by(ctx, ci.getPort(id_I4), is_lut, id_O);
+        if (!lut[3] || ci.getPort(id_I4)->users.entries()!=1)
+            continue;
+        
+        ci.type = id_XLUT;
+        bind_attr_loc(&ci, &ci.attrs);
+        ci.cluster = ci.name;
+        lut_only++;
+        for (int i=0;i<4;i++) {
+            ci.constr_children.push_back(lut[i]);
+            lut[i]->cluster = ci.cluster;
+            lut[i]->type = id_BEYOND_FE;
+            lut[i]->constr_z = PLACE_XLUT_FE1 + i;
+            lut[i]->renamePort(id_O, id_LO);
+            lut[i]->params[id_lut_used] = Property(1,1);
+            lut[i]->timing_index = ctx->get_cell_timing_idx(ctx->id("BEYOND_FE_LUT"));
+            lut_only++;
+        }
+    }
+    if (lut_only)
+        log_info("    %6d FEs used as LUT only\n", lut_only);
+    //if (lut_and_ff)
+    //    log_info("    %6d FEs used as LUT and DFF\n", lut_and_ff);
+    flush_cells();
+}
+
 void NgUltraPacker::pack_lut_dffs(void)
 {
     log_info("Pack LUT-DFFs...\n");
@@ -1774,6 +1831,7 @@ void NgUltraImpl::pack()
     packer.pack_cdcs();
     packer.pack_fifos();
     packer.pack_cys();
+    packer.pack_xluts();
     packer.pack_lut_dffs();
     packer.pack_dffs();
 

himbaechel/uarch/ng-ultra/pack.h

@@ -46,6 +46,7 @@ struct NgUltraPacker
     // LUTs & FFs
     void update_lut_init();
     void update_dffs();
+    void pack_xluts();
     void pack_lut_dffs();
     void pack_dffs();
     void pack_cys();