Ensure bias values for flatten layers are allocated first (#152)

* Ensure bias values for flatten layers are allocated first
* Do not include sampleoutput.h in RTL sims

Author: Robert Muchsel

izer/backend/max7800x.py

@@ -875,6 +875,7 @@ def create_net(self) -> str:  # pylint: disable=too-many-locals,too-many-branche
                     output_processor_map,
                     out_expand,
                     list(set().union(groups_used)),
+                    flatten,
                 )
 
             apb.function_header(function='init')
@@ -1069,6 +1070,7 @@ def create_net(self) -> str:  # pylint: disable=too-many-locals,too-many-branche
                     output_processor_map,
                     out_expand,
                     list(set().union(groups_used)),
+                    flatten,
                 )
 
             if verbose:

izer/kbias.py

@@ -32,6 +32,7 @@ def load(
         output_processor_map,
         out_expand,
         groups_used,
+        flatten,
 ):
     """
     Write `bias` values for the network to C code.
@@ -78,7 +79,7 @@ def load(
                 bias_len[ll] += 1  # Work around a problem on AI85
 
             bias_map += [(ll, groups_used if bias_group_map[ll] is None else bias_group_map[ll],
-                          bias_len[ll])]
+                          bias_len[ll], flatten[ll])]
             continue
 
         # For depth-wise convolutions, and qupac, each group needs to have 'its own' bias values
@@ -175,7 +176,10 @@ def rearrange_processor(p, bc_mode):
     # Reverse-sort the data structure for a Fit-First Descending (FFD) algorithm
     def bias_sort(e):
         """Order bias tuples"""
-        # Highest priority: If only one group is allowed
+        # Highest priority: flatten
+        if tc.dev.REQUIRE_FLATTEN_BIAS_FIRST and e[3]:  # flatten
+            return -1
+        # Next highest priority: If only one group is allowed
         if len(e[1]) == 1:
             return 0
         # For all other cases, reverse sort by size; sorted() will keep the original order
@@ -187,7 +191,7 @@ def bias_sort(e):
 
     bias_map = sorted(bias_map, key=bias_sort)
 
-    for _, (ll, gmap, blen) in enumerate(bias_map):
+    for _, (ll, gmap, blen, _) in enumerate(bias_map):
         if not calcx4[ll]:
             group = gmap[argmin(group_bias_max[t] for t in gmap)]
         else:

izer/toplevel.py

@@ -118,7 +118,8 @@ def header(
             memfile.write(f'#include "{state.weight_filename}"\n')
         if not lib and (embedded_code or state.compact_data):
             memfile.write(f'#include "{state.sample_filename}"\n')
-            if not cmsis_nn and state.generate_kat and state.result_filename.lower() != 'none':
+            if not cmsis_nn and embedded_code and state.generate_kat \
+               and state.result_filename.lower() != 'none':
                 memfile.write(f'#include "{state.result_filename}"\n')
     memfile.write('\n')
 

izer/tornadocnn.py

@@ -44,6 +44,7 @@ class Dev:
     REQUIRE_ONESHOT_CLEAR = True
     REQUIRE_NEW_STREAMING = False
     REQUIRE_FIFO_CPL = True
+    REQUIRE_FLATTEN_BIAS_FIRST = False
     EMULATE_ELTWISE_MP = False
     EMULATE_1X1_STREAMING = True
     USE_PROCESSORS = True
@@ -262,6 +263,7 @@ class DevAI85(Dev):
     FAST_FIFO_DR = 4  # Data register
     FAST_FIFO_DMA = 5  # DMA register (reserved function, not yet supported)
 
+    REQUIRE_FLATTEN_BIAS_FIRST = True
     EMULATE_ELTWISE_MP = True
     SUPPORTED_X2D_PADS = [0, 1, 2]
     SUPPORTED_X2D_OUTPUT_PADS = [1]
@@ -483,7 +485,7 @@ class DevAI87(Dev):
     C_PAD = 2
 
     # PLL Speed in MHz
-    PLL_SPEED = 240
+    PLL_SPEED = 200
 
     def __str__(self):
         return self.__class__.__name__