diff --git a/c/runtime.h b/c/runtime.h
index a36b1f7..9fc2120 100644
--- a/c/runtime.h
+++ b/c/runtime.h
@@ -53,12 +53,12 @@ static inline void process_y(int val, int i_py, Bundle_t *p_bundle, int ip, int
     return;
   }
 
-  // ------ ADD BIAS ------ 
-  if (p_bundle->is_bias)
-    val = (val << p_bundle->b_val_shift) + (mem.b[it_bias] << p_bundle->b_bias_shift);
+  // // ------ ADD BIAS ------ 
+  // if (p_bundle->is_bias)
+  //   val = (val << p_bundle->b_val_shift) + (mem.b[it_bias] << p_bundle->b_bias_shift);
   
-  // ------ CORE ACT ------
-  val = quant_lrelu(val, p_bundle->ca_nzero, p_bundle->ca_shift, p_bundle->ca_pl_scale);
+  // // ------ CORE ACT ------
+  // val = quant_lrelu(val, p_bundle->ca_nzero, p_bundle->ca_shift, p_bundle->ca_pl_scale);
 
 
   // ------ MAX/AVG POOL ------
@@ -120,10 +120,10 @@ extern EXT_C void load_y (unsigned char *p_done, unsigned char *pt_done_proc,  c
   //  inverted to increment once per call
   ++iw; if (iw >= p_bundle->w_kw2) { iw = 0;
     //after_each(in) = after_all(iw):
-    ++in; if (in >= p_bundle->n) { in = 0;
-      //after_each(il) = after_all(in):
-      ++il; if (il >= p_bundle->l) { il = 0;
-        //after_each(it) = after_all(il):
+    ++il; if (il >= p_bundle->l) { il = 0;
+      //after_each(in) = after_all(il):
+      ++in; if (in >= p_bundle->n) { in = 0;
+        //after_each(it) = after_all(in):
         ++it; if (it >= p_bundle->t) { it = 0;
           //after_each(ip) = after_all(it):
           printf("done p!! iw:%d in:%d il:%d it:%d ip:%d ib:%d\n", iw, in, il, it, ip, ib);
@@ -146,8 +146,8 @@ extern EXT_C void load_y (unsigned char *p_done, unsigned char *pt_done_proc,  c
           i_py = 0;
         }//new(it):
         it_bias = p_bundle->b_offset + p_bundle->coe*it;
-      }//new(il):
-    }//new(in):
+      }//new(in):
+    }//new(il):
   }//new(iw):
   *pt_done_proc = !(*pt_done_proc);
 }
diff --git a/test/py/bundle.py b/test/py/bundle.py
index 5c118d3..8e1fda7 100644
--- a/test/py/bundle.py
+++ b/test/py/bundle.py
@@ -263,6 +263,7 @@ def add (p, p_frac, p_bits, q, q_frac, q_bits):
         clog2_add = int(np.ceil(np.log2(np.prod(self.w['int'].shape[:-1]))))
         self.proc['bits'] = self.inp['bits'] + self.w['bits'] + clog2_add
         self.proc['frac'] = self.inp['frac'] + self.w['frac']
+        # self.o_exp = self.proc['int']
 
         if self.b is not None:
             (self.proc['int'], self.proc['frac'], self.proc['bits']), (self.bias_val_shift, self.bias_b_shift) = add(