/**********************************************

   該函數的功能主要是進行16*16宏塊SAD
   運算

**********************************************/
.section L1_code;
.global __sad16;
.align 8;
__sad16:
       [--SP] = (R7:4,P5:3);
       I2 = R0;                //當前宏塊地址      cur
       P0 = R1;                //參考幀的開始地址  ref
       R6 = R2;                //stride            stride
       P5 = [SP+40];
       R7 = 16;
       R7 = R6-R7(S);          //stride-16
       M0 = R7;                //stride-16
       R7 = 12;
       R7 = R2-R7(S);
       M1 = R7;
       R7 = P0;                 
       P4 = R7;
       I0 = R0;
       I1 = R7;
       L2 = 0;
       L3 = 0;
       P2 = P4>>2;        
       P4 = P2<<2;             //最低兩位置零                
       P3 = 16;                //循環次數
       I3 = P4; 
       R0 = [I2++]||R2 = [I3++];   //開始進行SAD運算
       A1 = A0 = 0||R1 =[I2++]||R3 = [I3++];
        LSETUP(SAD_ST,SAD_END) LC1 = P3;
SAD_ST:
       SAA(R1:0,R3:2)||R0 = [I2++]||R2 = [I3++];
       SAA(R1:0,R3:2)(R)||R1 = [I2++M1]||R3 = [I3++];
       SAA(R1:0,R3:2)||R0 = [I2++]||R2 = [I3++M0];
       SAA(R1:0,R3:2)(R)||R1 = [I2++]||R2 = [I3++];
SAD_END:  R3 = [I3++];        
       R3 =A1.L+A1.H,R2 =A0.L+A0.H;
       R0 = R2+R3;
       [P5] = R0;
                                   //SAD運算結束
      (R7:4,P5:3) = [SP++];
       RTS;                       //函數返回
       NOP; 
__sad16.end: