Added Ortree maybe

dataflow_neuro/treegates.act

@@ -34,6 +34,9 @@ namespace tmpl {
 
                 
 export deftype power (bool?! vdd, vss) { }
+
+
+
 export template<pint N; pbool invert>
 defproc ortree (bool? in[N]; bool out)
 {
@@ -43,14 +46,48 @@ defproc ortree (bool? in[N]; bool out)
 
   pint i, end, j;
   pbool isinv;
-  isinv = false;
+  isinv = invert;
   i = 0;
   end = N-1;
 
-  /* dynamic array that holds all the nodes in the completion tree */
-  bool tmp[N];
+  pint lenTree2Count, lenTree3Count;
+  /* Pre"calculate" the number of C cells required, look below if confused */
+  *[ i != end ->
+     j = 0;
+     *[ i < end ->
+        j = j + 1;
+        [ i+1 >= end ->
+          i = end;
+          lenTree2Count = lenTree2Count +1;
+        [] i+2 >= end ->
+          i = end;
+          lenTree3Count = lenTree3Count +1;
+        [] else ->
+          i = i + 2;
+          lenTree2Count = lenTree2Count +1;
+        ]
+      ]
+      /*-- update range that has to be combined --*/
+      i = end+1;
+      end = end+j;
+      j = 0;
+  ]
+
+  /* array that holds ALL the nodes in the completion tree */
+  bool tmp[end];
   (k:N:tmp[k] = in[k];)
 
+  /* array to hold the actual C-elments, either A2C or A3C */
+  OR2_X1 C2Els[lenTree2Count];
+  OR3_X1 C3Els[lenTree3Count];
+
+  /* Reset the variables we just stole lol */
+  i = 0;
+  end = N-1;
+  j = 0;
+  pint tree2Index = 0;
+  pint tree3Index = 0;
+
   /* Invariant: i <= end */
     
   *[ i != end ->
@@ -64,51 +101,36 @@ defproc ortree (bool? in[N]; bool out)
         /*-- there are still signals that need to be combined --*/
         j = j + 1;
         bool tmp[end+j..end+j];
-        [ i+2 >= end ->
-          /*-- last piece: use either a 2 or 3 input NAND/NOR gate --*/
-          [isinv ->
-            prs { (&k:i..end:tmp[k]) => tmp[end+j]- }
-          [] else ->
-            prs { (|k:i..end:tmp[k]) => tmp[end+j]- }
-          ]
+        [ i+1 >= end ->
+          /*-- last piece: use either a 2 input C-element --*/
+          C2Els[tree2Index].c1 = tmp[i];
+          C2Els[tree2Index].c2 = tmp[i+1];
+          C2Els[tree2Index].y = tmp[end+j];
+          tree2Index = tree2Index +1;
+          i = end;
+        [] i+2 >= end ->
+          /*-- last piece: use either a 3 input C-element --*/
+          C3Els[tree3Index].c1 = tmp[i];
+          C3Els[tree3Index].c2 = tmp[i+1];
+          C3Els[tree3Index].c3 = tmp[i+2];
+          C3Els[tree3Index].y = tmp[end+j];
+
+          tree3Index = tree3Index +1;
           i = end;
         [] else ->
           /*-- more to come; so use a two input C-element --*/
-          [isinv ->
-            prs { (&k:i..i+1:tmp[k]) => tmp[end+j]- }
-          [] else ->
-            prs { (|k:i..i+1:tmp[k]) => tmp[end+j]- }
-          ]
+          C2Els[tree2Index].c1 = tmp[i];
+          C2Els[tree2Index].c2 = tmp[i+1];
+          C2Els[tree2Index].y = tmp[end+j];
+          tree2Index = tree2Index +1;
           i = i + 2;
         ]
-        sizing {
-          leak_adjust <- 1; 
-          p_n_mode <- 1;
-          tmp[end+j]{-1}
-        }
       ]
-      /*-- we just added an inverting layer --*/
-      isinv = ~isinv;
-
       /*-- update range that has to be combined --*/
       i = end+1;
       end = end+j;
       j = 0;
   ]
-
-  isinv = invert ? ~isinv : isinv;
-
-  /*-- invert the signal if needed --*/
-  [isinv -> prs { tmp[end] => out- } 
-  [] else -> tmp[end] = out;
-  ]
-  [isinv ->
-  sizing {
-    leak_adjust <- 1; 
-    p_n_mode <- 1;
-    out{-1}
-  }
-  ]
 }
   
 /*