autocor.asm

davinci技术 源码 视频监控汇编源码

          .asg            A31,        A_prod8
          .asg            A30,        A_prodC
          .asg            B31,        B_prodA
          .asg            B30,        B_prodE
          .asg            A25,        A_prod3
          .asg            A24,        A_prod7
          .asg            B21,        B_prod0
          .asg            B20,        B_prod4
          .asg            A21,        A_prod2
          .asg            A20,        A_prod6
          .asg            B23,        B_prod9
          .asg            B22,        B_prodD
          .asg            A21,        A_prodB
          .asg            A20,        A_prodF
          .asg            B8,         B_s0
          .asg            A25,        A_s0

          .asg            B9,         B_t0
          .asg            A8,         A_t0

          .asg            B24,        B_u0
          .asg            A17,        A_u0

          .asg            B30,        B_v0
          .asg            A9,         A_v0

          .asg            A20,        A_temp0
          .asg            B21,        B_temp1
          .asg            A27,        A_temp2
          .asg            B23,        B_temp3

          .asg            A18,        A_sum_0
          .asg            B16,        B_sum_1
          .asg            A19,        A_sum_2
          .asg            B17,        B_sum_3




          SHRU     .S1X   B_nr,                2,              A_i
||        SHRU     .S2X   A_nx,                3,              B_ik_rst
||        ADDAH    .D2    B_x,                 B_nr,           B_yptr_rst

          MVC      .S2    B_ik_rst,            RILC
||        SUB      .D2    B_ik_rst,            1,              B_ik_rst
||        SUB      .L1X   B_yptr_rst,          16,             A_yptr_rst



   [A_i]  SPLOOPD         4
||        MVC      .S2    B_ik_rst,            ILC
||        MV       .L2X   A_yptr_rst,          B_xptr
||        MV       .L1    A_yptr_rst,          A_yptr


          LDDW     .D1T2  *++A_yptr[2],        B_y3y2:B_y1y0                       ;[ 1,1]

          SPMASK          S1,L2
||        LDDW     .D1T1  *-A_yptr[1],         A_z0z1:A_z2z3                       ;[ 2,1]
||        LDDW     .D2T2  *++B_xptr[2],        B_x3x2:B_x1x0                       ;[ 2,1]
||        SUB      .S1    A_yptr_rst,          8,              A_yptr_rst
||        MV       .L2X   A_yptr_rst,          B_xptr_rst


          SPMASK          S1,L2
||        LDDW     .D1T1  *A_yptr[1],          A_y7y6:A_y5y4                       ;[ 3,1]
||        SUB      .S1    A_i,                 1,              A_i
||        MV       .L2X   A_r,                 B_rptr

          LDDW     .D2T1  *B_xptr[1],          A_x7x6:A_x5x4                       ;[ 4,1]

          NOP             2

          DMV      .S2X   B_y1y0,              A_z0z1,         Bt_y1y0:Bt_z0z1     ;[ 7,1]
||        DDOTPH2  .M1X   A_z0z1:A_z2z3,       B_x1x0,         A_prod8:A_prodC     ;[ 7,1]
||        DDOTPH2  .M2    B_y3y2:B_y1y0,       B_x3x2,         B_prod1:B_prod5     ;[ 7,1]

          DMV      .S1X   A_y5y4,              B_y3y2,         At_y5y4:At_y3y2     ;[ 8,1]
||        DDOTPH2  .M2    Bt_y1y0:Bt_z0z1,     B_x1x0,         B_prod0:B_prod4     ;[ 8,1]

          DDOTPH2  .M1    At_y5y4:At_y3y2,     A_x7x6,         A_prodB:A_prodF     ;[ 9,1]
||        DDOTPH2  .M2X   B_y3y2:B_y1y0,       A_x5x4,         B_prodA:B_prodE     ;[ 9,1]

          DDOTPH2  .M2    Bt_y1y0:Bt_z0z1,     B_x3x2,         B_prod9:B_prodD     ;[10,1]
||        DDOTPH2  .M1    A_y7y6:A_y5y4,       A_x7x6,         A_prod3:A_prod7     ;[10,1]

          NOP             1

          DDOTPH2  .M1    At_y5y4:At_y3y2,     A_x5x4,         A_prod2:A_prod6     ;[12,1]

          ADD      .L1    A_prodC,             A_prodF,        A_v0                ;[13,1]
||        ADD      .D2    B_prod0,             B_prod1,        B_s0                ;[13,1]

          ADD      .L1    A_prod8,             A_prodB,        A_u0                ;[14,1]
||        ADD      .L2    B_prod4,             B_prod5,        B_t0                ;[14,1]


          SPMASK          L1,L2
||        ADD      .D2    B_prodE,             B_prodD,        B_v0                ;[15,1]
||        ZERO     .L1    A_sum_2:A_sum_0
||        ZERO     .L2    B_sum_3:B_sum_1


          ADD      .S2X   A_v0,                B_v0,           B_temp3             ;[16,1]
||        ADD      .L2    B_prodA,             B_prod9,        B_u0                ;[16,1]
||        ADD      .D1    A_prod6,             A_prod7,        A_t0                ;[16,1]
||        ADD      .L1    A_prod2,             A_prod3,        A_s0                ;[16,1]

          ADD      .L2    B_sum_3,             B_temp3,        B_sum_3             ;[17,1]
||        ADD      .S1X   A_s0,                B_s0,           A_temp0             ;[17,1]


          ADD      .S1X   A_u0,                B_u0,           A_temp2             ;[18,1]
||        ADD      .S2X   B_t0,                A_t0,           B_temp1             ;[18,1]

          SPKERNEL
||        ADD      .S1    A_sum_2,             A_temp2,        A_sum_2             ;[19,1]
||        ADD      .L2    B_sum_1,             B_temp1,        B_sum_1             ;[19,1]
||        ADD      .L1    A_sum_0,             A_temp0,        A_sum_0             ;[19,1]


POST_KERNEL:
;cycle 1
   [A_i]  MV       .D1    A_yptr_rst,          A_yptr
|| [A_i]  MV       .S2    B_xptr_rst,          B_xptr

;cycle 2
          NOP

;cycle 3
          SUB      .D1    A_yptr_rst,          8,              A_yptr_rst

;cycle 4
          SPMASKR

;Reload Point

;Reload Cycle  1  through 9
          NOP             9


;Reload Cycle  10
          SPMASK          S2
|| [!A_i] RET      .S2    B3
||        ADD      .L2X   B_t0,                A_t0,           B_temp1

;Reload Cycle  11
          NOP

;Reload Cycle  12
          RPACK2   .S2X   B_sum_3,             A_sum_2,        B_pack_0

;Reload Cycle  13
          NOP

;Reload Cycle  14
          RPACK2   .S2X   B_sum_1,             A_sum_0,        B_pack_1
|| [A_i]  B        .S1    POST_KERNEL

;Reload Cycle  15
          SPMASK          D2
||        STDW     .D2    B_pack_0:B_pack_1,   *B_rptr++[1]
||        ADD      .L2    B_prodE,             B_prodD,        B_v0                ;[15,1]
||[A_i]   SUB      .L1    A_i,                 1,              A_i

;Reload Cycle  16
          NOP

;Reload Cycle  17
          SPMASK          L2
||        ADD      .L2    0,                   B_temp3,        B_sum_3             ;[17,1]

;Reload Cycle  18
          NOP

;Reload Cycle  19
          SPMASK          S1,L1,L2
||        ADD      .S1    0,                   A_temp2,        A_sum_2             ;[19,1]
||        ADD      .L2    0,                   B_temp1,        B_sum_1             ;[19,1]
||        ADD      .L1    0,                   A_temp0,        A_sum_0             ;[19,1]

          .end

* ========================================================================= *
*  End of file:  autocor.asm                                                *
* ------------------------------------------------------------------------- *
*             Copyright (c) 2005 Texas Instruments, Incorporated.           *
*                            All Rights Reserved.                           *
* ========================================================================= *