some more optimizations

buildz80com.py

@@ -34,11 +34,24 @@ def pack_2bit_weights(weights: np.ndarray) -> bytes:
         chunk = mapped[i:i+4]
         if len(chunk) < 4:
             chunk = np.pad(chunk, (0, 4 - len(chunk)), constant_values=2)
+        # 0,1,2,3 are as -2,-1,0,1 but in muladd for speed will be treated as -2,0,1,-1                        
+        for c in range(4):
+            match int(chunk[c]):
+                case 0:
+                    chunk[c] = 0 # -2 in position 0
+                case 1:
+                    chunk[c] = 3 # -1 in position 3
+                case 2:
+                    chunk[c] = 1 # 0 in position 1
+                case 3:
+                    chunk[c] = 2 # 1 in position 2
+                case _:
+                    raise Exception("weight value not valid")
         byte = \
-            (int(chunk[2]) << 6) | \
-            (int(chunk[1]) << 4) | \
-            (int(chunk[0]) << 2) | \
-            (int(chunk[3])) # chunk 3 as last since in the evaluation there will be first a rotation
+            (chunk[2] << 6) | \
+            (chunk[1] << 4) | \
+            (chunk[0] << 2) | \
+            (chunk[3]) # chunk 3 as last since in the evaluation there will be first a rotation
         packed.append(byte)
 
     return bytes(packed)
@@ -453,8 +466,9 @@ def build_autoreg(model_path: str = 'command_model_autoreg.pt'):
     b.inc_hl()                  #   inc hl            
     b.ld_d_hl()                 #   ld d, (hl)             
     b.inc_hl()                  #   inc hl            
-    b.ld_mem_label_hl('CURIN')  #   ld (CURIN), hl                            
-    b.call('MULADD')            #   call MULADD                  
+    b.ld_mem_label_hl('CURIN')  #   ld (CURIN), hl               
+    b.dec_a()                   #   dec a ; to verify if is 1 (that means weitgh "0 equivalent")
+    b.call_nz('MULADD')         #   call nz, MULADD ; "0 equivalent" is the most common state, so call will be performed more rarely                  
     b.inc_c()                   #   inc c           
     b.djnz('LWT')               #   djnz LWT               
                                 # 
@@ -467,11 +481,11 @@ def build_autoreg(model_path: str = 'command_model_autoreg.pt'):
     b.ld_hl_mem_label('ACC')    #   ld hl, (ACC)                          
     b.add_hl_de()               #   add hl, de               
     b.ld_mem_label_hl('ACC')    #   ld (ACC), hl
-                                #   ; After each layer, arithmetic right-shift by 2 to prevent overflow                          
-    b.sra_h()                   #   sra h          ; Shift right arithmetic (preserves sign)    
+                                #   ; after each layer, arithmetic right-shift by 2 to prevent overflow                          
+    b.sra_h()                   #   sra h ; shift right arithmetic (preserves sign)    
     b.rr_l()                    #   rr l          
     b.sra_h()                   #   sra h           
-    b.rr_l()                    #   rr l           ; ACC = ACC / 4
+    b.rr_l()                    #   rr l ; ACC = ACC / 4
     b.ld_iyd_l(0)               #   ld (iy+0), l                
     b.ld_iyd_h(1)               #   ld (iy+1), h              
     b.inc_iy()                  #   inc iy            
@@ -480,21 +494,16 @@ def build_autoreg(model_path: str = 'command_model_autoreg.pt'):
     b.djnz('LNEUR')             #   dec b : jp nz, LNEUR                  
     b.ret()                     #   ret         
                                 # 
-    # === MULADD ===            #                     
-    # a is 0,1,2,3 that are as -2,-1,0,+1
+    # === MULADD ===            #   ; since last dec here a is $FF,1,2 that are as -2,1,-1
     b.label('MULADD')           # MULADD:                     
-    # TODO: ACC should be put in a register
     b.ld_hl_mem_label('ACC')    #   ld hl, (ACC)                          
     b.dec_a()                   #   dec a          
-    b.jr_z('MA_M1')             #   jr z, MA_M1 ; jump if a is -1 equivalent
+    b.jr_z('MA_P1')             #   jr z, MA_P1 ; jump if a is +1 equivalent (33% jump probability)
+    b.sbc_hl_de()               #   sbc hl, de ; a is -1 or -2 equivalent             
     b.dec_a()                   #   dec a          
-    b.ret_z()                   #   ret z ; a is zero equivalent
-    b.dec_a()                   #   dec a          
-    b.jr_z('MA_P1')             #   jr z, MA_P1 ; jump if a is +1 equivalent
-    b.label('MA_M2')            # MA_M2: ; a is -2 equivalent                    
-    b.sbc_hl_de()               #   sbc hl, de              
-    b.label('MA_M1')            # MA_M1: ; -1                   
-    b.sbc_hl_de()               #   sbc hl, de               
+    b.jr_z('MA_MRET')           #   jr z, MA_MRET ; skip next sbc if a is just -1 equivalent (50% jump probability)
+    b.sbc_hl_de()               #   sbc hl, de ; second time since a is -2 equivalent
+    b.label('MA_MRET')          # MA_MRET:               
     b.ld_mem_label_hl('ACC')    #   ld (ACC), hl                          
     b.ret()                     #   ret         
     b.label('MA_P1')            # MA_P1: ; a is +1                     

libz80.py

@@ -83,6 +83,10 @@ def call(self, label: str):
         self.emit(0xCD)
         self.fixup_word(label)
 
+    def call_nz(self, label: str):
+        self.emit(0xC4)
+        self.fixup_word(label)
+
     def call_addr(self, addr: int):
         self.emit(0xCD, addr & 0xFF, (addr >> 8) & 0xFF)