normcorr1.asm

TI C54写的G729代码,视线8kbps高质量语音编码,汇编优化

;==========================================================================
;  File Name
;  ------------
;  NORMCORR.ASM
;
;  Brief Description of the Code:
;  ------------------------------
;  Find the normalized correlation between the target vector and the
;  filtered past excitation.
;
;  Ref
;  ------
;  PITCH.C
;==========================================================================
;  Norm_Corr
;--------------------------------------------------------------------------
;  Input arguments:
;     exc[]    : excitation buffer
;     xn[]     : target vector
;     h[]      : impulse response of synthesis and weighting filters (Q12)
;     L_subfr  : Length of subframe
;     t_min    : minimum value of pitch lag.
;     t_max    : maximum value of pitch lag.
;
;  Output arguments:
;     corr_norm[]:  normalized correlation (correlation between target and
;                   filtered excitation divided by the square root of
;                   energy of filtered excitation)
;--------------------------------------------------------------------------
;  static void Norm_Corr(Word16 exc[], Word16 xn[], Word16 h[], Word16 L_subfr,
;                 Word16 t_min, Word16 t_max, Word16 corr_norm[])
;--------------------------------------------------------------------------
;  Str_Norm_Corr      .struct
;  ptr_exc            .word   1
;  ptr_xn             .word   1
;  ptr_h              .word   1
;  t_min              .word   1
;  t_max              .word   1
;  ptr_corr           .word   1
;  Str_Norm_Corr_Len  .endstruct
;--------------------------------------------------------------------------
;  Constant : L_subfr = L_SUBFR
;--------------------------------------------------------------------------
;  Modified Registers : AR0, AR1, AR2, AR3, AR4, AR5, AR6
;--------------------------------------------------------------------------
		 .mmregs
         .include ..\include\const.h
         .include ..\include\struct.h
         .include ..\include\ld8amem.h
         .include ..\include\convolve.h

         .text
         .def     Norm_Corr
         .ref     Convolve
         .ref     Inv_sqrt

param_norm_corr .tag     Str_Norm_Corr

         .asg     "AR1", pCorr_norm
         .asg     "AR2", pExcf
         .asg     "AR3", pXn
         .asg     "AR4", pExcfTemp
         .asg     "AR5", pExc
         .asg     "AR6", pH
         .asg     "TRN", ASM_Cache            ; temp storage of ASM

         .asg     "Addr_norm_corr_buf", excf  ; circular buffer
         .asg     "excf+L_SUBFR", norm
         .asg     "norm+2", m_L_corr

Norm_Corr:
        LD      #param_norm_corr, DP

        LD      param_norm_corr.ptr_exc, A
        SUB     param_norm_corr.t_min, A
        STM     #excf, pConv_Y
        STLM    A, pConv_X

        STM     #L_SUBFR-1, BRC
        CALLD   Convolve
        MVDM    param_norm_corr.ptr_h, pConv_H

        STM     #1, AR0                         ; setup for circular buffer
        STM     #L_SUBFR, BK

        LD      #param_norm_corr, DP
        LD      param_norm_corr.ptr_exc, 16, A
        SUB     param_norm_corr.t_min, 16, A
        SUB     #1, 16, A
        MVDM    param_norm_corr.ptr_xn, pXn
        MVDM    param_norm_corr.ptr_h,  pH
        PSHM    AH
        MAR     *pH-%


        STM     #excf, pExcf
        RPTZ    A, #L_SUBFR-1
                SQURA   *pExcf+, A
        MAR     *+pExcf(-40)
        SUB     #400h, 16, A, B
        LD      #-2, ASM


        MVMM    pExcf, pExcfTemp
        LD      *pExcfTemp+, 16, A
        XC      2, BGT
        RPT     #L_SUBFR-1                      ; one word instruction
                ST      A, *pExcf+0%
            ||  LD      *pExcfTemp+, A


   ; ------------------------------------

        LD      param_norm_corr.t_max, A
        SUB     param_norm_corr.t_min, A
        STLM    A, AR4
        MVDM    param_norm_corr.ptr_corr, pCorr_norm
        MAR     *AR4-

        STM     #15-12,   ASM_Cache
        XC      2, BGT
        STM     #15-12-2, ASM_Cache

Outter_Loop_Start
                PSHM     AR4
                RPTZ     A, #L_SUBFR-1
                         SQURA   *pExcf+, A
                MAR     *+pExcf(-40)
                CALLD    Inv_sqrt               ; DP = 0

                STM      #norm, AR4
		        STL     A,-16,*AR4
          		LD      *AR4+,15,B
         		LD      A,-1,A
          		SUB     B,A
          		STL     A,*AR4

                RPTZ     B, #L_SUBFR-1
                         MAC     *pXn+0%, *pExcf+0%, B
                MAR      *pExcf-%



                STM      #m_L_corr, AR5
                 STL     B, -16, *AR5
                 LD      *AR5+,15,A
                 LD      B,-1,B
                 SUB     A,B
                 STL      B,*AR5-

                LD       #0, A
                LD       #0, B
                MACSU    *AR4-, *AR5+, A
                ld       A,-16,A
                LD       A,1,A
                MACSU    *AR5-, *AR4 ,B
                LD       B, -16, B
                LD       B,1,B
                ADD      B,A
                MAC      *AR5, *AR4, A


                STL      A, *pCorr_norm+
                POPM     AR4
                POPM     pExc
				NOP
				NOP
                LD       ASM_Cache, ASM

                STM      #L_SUBFR-2, BRC
                RPTBD    Inner_Loop_End-1
                LD       *pExc-, T
                PSHM     pExc

                         MAR     *pExcf-
                         MPY     *pH-, A
                         LD      A, ASM, A
                         ADD     *pExcf+, 16, A
                         STL     A, -16, *pExcf-
Inner_Loop_End
                NOP
				NOP
                LD       T, 16-3, A
Outter_Loop_End    
                MAR      *pH-%
                STL 	 A,ASM 
                BANZD    Outter_Loop_Start, *AR4-
                STL      A, -16, *pExcf
        FRAME   1
        RPTZ    A, #L_SUBFR-1
                SQURA   *pExcf+%, A
        CALLD   Inv_sqrt                        ; DP = 0

        STM     #norm, AR4
        STL     A,-16,*AR4
        LD      *AR4+,15,B
        LD      A,-1,A
        SUB     B,A
        STL      A,*AR4

        RPTZ    B, #L_SUBFR-1
                MAC     *pXn+0%, *pExcf+0%, B



        STM      #m_L_corr, AR5
                 STL     B,-16,*AR5
                 LD      *AR5+,15,A
                 LD      B,-1,B
                 SUB     A,B
                 STL     B,*AR5-

        LD       #0, A
        LD       #0, B
        MACSU    *AR4-, *AR5+, A
        LD       A,-16,A
        LD       A,1,A
        MACSU    *AR5-, *AR4 , B
        LD       B,-16,B
        LD       B,1,B
        ADD      B,A

        RETD
        MAC     *AR5, *AR4, A
        STL     A, *pCorr_norm+