chebps11.asm

TI C54写的G729代码,视线8kbps高质量语音编码,汇编优化

;==========================================================================
;  File Name
;  ----------
;  LP_LSP.ASM
;
;  Brief Description of the Code:
;  ------------------------------
;  Evaluates the Chebichev polynomial series
;
;  Ref
;  ------
;  LPC.C
;==========================================================================

;--------------------------------------------------------------
; function  Word16 Chebps_11(Word16 x, Word16 f[], Word16 n)
;           ~~~~~~~~~~~~~~~~~~~~
;    Evaluates the Chebichev polynomial series
;--------------------------------------------------------------
;
;  The polynomial order is
;     n = M/2   (M is the prediction order)
;  The polynomial is given by
;    C(x) = T_n(x) + f(1)T_n-1(x) + ... +f(n-1)T_1(x) + f(n)/2
; Arguments:
;  x:     input value of evaluation; x = cos(frequency) in Q15
;  f[]:   coefficients of the pol.
;                         in Q11(Chebps_11), in Q10(Chebps_10)
;  n:     order of the pol.
;
; The value of C(x) is returned. (Saturated to +-1.99 in Q14)
;
;--------------------------------------------------------------
;  Input : DP  -> Addr_Chebps_Buf  ; Aligned on word boundary
;          BK  =  5
;          BRC =  n - 3
;
;          AR0 =  1     ; For dual memory operand addressing in
;                       ; circular buffer
;          AR4 -> X
;          AR3 -> f[1]  ; Use circular buffer, BK = 5
;
;  Output: A = Chebps_11(Word16 x, Word16 f[], Word16 n)
;          AR4 -> X
;          AR3 -> f[1]
;
;  Used and unchanged : AR3, AR4
;  Modified           : AR5
;
;--------------------------------------------------------------
;  AR3 -> f[i], AR4 -> X, AR5 -> b1_l ( BL )
;--------------------------------------------------------------



           .MMREGS

           .include ..\include\lpc_lsp.h

           .def     Chebps_11
           .def     Chebps_10

           .asg     "AR3", pCoef
           .asg     "AR4", pX
           .asg     "AR5", pB1_l


one_hi     .set     256                 ; 1.0 in Q24
one_lo     .set     0
one_hi1    .set     128                 ; 1.0 in Q24

MPY32_16   .macro
        LD       #0, A
        MACSU    *AR5+, *AR4, A
        LD       A, -16, A
        LD       A, -1, A
        LD       A,  1, A
        MAC      *AR5-, *AR4, A
           .endm

Chebps_11
        STM    BL, AR5

        LD     #one_hi, 16, A
        DST    A, B2                    ; b2 = 1.0 in Q24

        LD     *pX, 10, B
        ADD    *pCoef+0%, 13, B         ; B = b1 = 2*x + f[1]

        RPTB   BLK_END-1
        MPY32_16                        ; AR4 -> X, AR5 -> b1_l ( BL )
                                        ; A  = t0 = x*b1

        LD     A, 1, A                  ; t0 = 2.0*x*b1
        DSUB   B2, A                    ; t0 = 2.0*x*b1 - b2
        ADD    *pCoef+0%, 13, A         ; t0 = 2.0*x*b1 - b2 + f[i]
                                        ; Now A = b0, B = b1

        DST    B, B2                    ; b2 = b1
        LD     A, B                     ; b1 = b0
BLK_END
        MPY32_16                        ; t0 = x*b1
        DSUB   B2, A                    ; t0 = x*b1 - b2
        RETD
        ADD    *pCoef+0%, 12, A         ; t0 = x*b1 - b2 + f[i]/2
        LD     A, 6, A

Chebps_10
        STM    BL, AR5

        LD     #one_hi1, 16, A
        DST    A, B2                    ; b2 = 1.0 in Q23

        LD     *pX, 9, B
        ADD    *pCoef+0%, 13, B         ; B = b1 = 2*x + f[1]

        RPTB   BLK_END1-1
        MPY32_16                        ; AR4 -> X, AR5 -> b1_l ( BL )
                                        ; A  = t0 = x*b1

        LD     A, 1, A                  ; t0 = 2.0*x*b1
        DSUB   B2, A                    ; t0 = 2.0*x*b1 - b2
        ADD    *pCoef+0%, 13, A         ; t0 = 2.0*x*b1 - b2 + f[i]
                                        ; Now A = b0, B = b1

        DST    B, B2                    ; b2 = b1
        LD     A, B                     ; b1 = b0
BLK_END1
        MPY32_16                        ; t0 = x*b1
        DSUB   B2, A                    ; t0 = x*b1 - b2
        RETD
        ADD    *pCoef+0%, 12, A         ; t0 = x*b1 - b2 + f[i]/2
        LD     A, 7, A