Same changes for "triangle" multiplication phase as for the "square" one

(debugging output, simpler MAC clearing and index rotation logic).

1 files changed, 45 insertions, 20 deletions

diff --git a/modexpng_fpga_model.py b/modexpng_fpga_model.py
index 73a21d3..5632a6f 100644
--- a/modexpng_fpga_model.py
+++ b/modexpng_fpga_model.py
@@ -77,9 +77,9 @@ _VECTOR_CLASS = "Vector"
 DUMP_VECTORS = False
 DUMP_INDICES = False
 DUMP_MACS_CLEARING = False
-DUMP_MACS_ACCUMULATION = False
+DUMP_MACS_ACCUMULATION = True
 DUMP_MULT_PARTS = False
-DUMP_RCMB = True
+DUMP_RCMB = False
 
 
 #
@@ -395,6 +395,12 @@ class ModExpNG_WordMultiplier():
         if dump and DUMP_MULT_PARTS:
             print("t=%2d, col=%2d > parts[%2d]: mac[%d] = 0x%012x" %
                 (time, column, part_index, mac_index, parts[part_index]))
+
+    def _mult_store_part_aux(self, parts, time, column, part_index, mac_index, dump):
+        parts[part_index] = self._mac_aux[mac_index]
+        if dump and DUMP_MULT_PARTS:
+            print("t=%2d, col=%2d > parts[%2d]: mac_aux[%d] = 0x%012x" %
+                (time, column, part_index, mac_index, parts[part_index]))
                 
     def multiply_square(self, a_wide, b_narrow, ab_num_words, dump=False):
 
@@ -443,8 +449,6 @@ class ModExpNG_WordMultiplier():
                         part_index = t
                         self._mult_store_part(parts, t, col, part_index, x, dump)
 
-                            
-
                 if dump and DUMP_MACS_ACCUMULATION:
                     print("t=%2d, col=%2d > "% (t, col), end='')
                     for i in range(NUM_MULTS):
@@ -462,7 +466,9 @@ class ModExpNG_WordMultiplier():
 
         return parts
 
-    def multiply_triangle(self, a_wide, b_narrow, ab_num_words):
+    def multiply_triangle(self, a_wide, b_narrow, ab_num_words, dump=False):
+
+        if dump: print("multiply_triangle()")
 
         num_cols = ab_num_words // NUM_MULTS
 
@@ -474,15 +480,27 @@ class ModExpNG_WordMultiplier():
 
             last_col = col == (num_cols - 1)
 
-            self._clear_all_macs()
-            self._preset_indices(col)
-
-            if last_col:
-                self._clear_mac_aux()
-                self._preset_index_aux(num_cols)
-
             for t in range(ab_num_words + 1):
 
+                if t == 0:
+                    self._preset_indices(col)
+                    if last_col:
+                        self._preset_index_aux(num_cols)
+                else:
+                    self._rotate_indices(ab_num_words)
+                    if last_col:
+                        self._rotate_index_aux()
+            
+                if t == 0:
+                    self._clear_all_macs()
+                    if dump and DUMP_MACS_CLEARING:
+                        print("t= 0, col=%2d > clear > all" % (col))
+
+                    if last_col:
+                        self._clear_mac_aux()
+                        if dump and DUMP_MACS_CLEARING:
+                            print("t= 0, col=%2d > clear > aux" % (col))
+                        
                 # current b-word
                 bt = b_narrow.words[t]
 
@@ -492,7 +510,8 @@ class ModExpNG_WordMultiplier():
                     self._update_one_mac(x, ax, bt)
 
                     if t == (col * NUM_MULTS + x):
-                        parts[t] = self._macs[x]
+                        part_index = t
+                        self._mult_store_part(parts, t, col, part_index, x, dump)
 
                 # aux multiplier
                 if last_col:
@@ -500,17 +519,23 @@ class ModExpNG_WordMultiplier():
                     self._update_mac_aux(ax * bt)
 
                     if t == ab_num_words:
-                        parts[t] = self._mac_aux[0]
+                        part_index = t
+                        self._mult_store_part_aux(parts, t, col, part_index, 0, dump)
+
+                if dump and DUMP_MACS_ACCUMULATION:
+                    print("t=%2d, col=%2d > "% (t, col), end='')
+                    for i in range(NUM_MULTS):
+                        if i > 0: print(" | ", end='')
+                        print("mac[%d]: 0x%012x" % (i, self._macs[i]), end='')
+                    if last_col:
+                        print(" | mac_aux[ 0]: 0x%012x" % (self._mac_aux[0]), end='')
+                    print("")
 
+                        
                 # shortcut
                 if not last_col:
                     if t == (NUM_MULTS * (col + 1) - 1): break
 
-                # advance indices
-                self._rotate_indices(ab_num_words)
-                if last_col:
-                    self._rotate_index_aux()
-
         return parts
 
     def multiply_rectangle(self, a_wide, b_narrow, ab_num_words):
@@ -693,7 +718,7 @@ class ModExpNG_Worker():
             return ModExpNG_Operand(None, 2*ab_num_words, ab_words)
 
         # 2.
-        q_parts = self.multiplier.multiply_triangle(ab, n_coeff, ab_num_words)
+        q_parts = self.multiplier.multiply_triangle(ab, n_coeff, ab_num_words, dump)
         q_words = self.recombinator.recombine_triangle(q_parts, ab_num_words, dump)
         q = ModExpNG_Operand(None, ab_num_words + 1, q_words)