void Update_Err(Word16 Olp, Word16 AcLg, Word16 AcGn,ENC_HANDLE *handle)
{
Word16 *ptr_tab;
Word16 i, iz, temp1, temp2;
Word16 Lag;
Word32 Worst1, Worst0, L_temp;
Word16 beta;
Lag = Olp - (Word16)Pstep + AcLg;
/* Select Quantization tables */
i = 0 ;
ptr_tab = tabgain85;
if ( handle->mode == G723_63 ) {
if ( Olp >= (Word16) (SubFrLen-2) ) ptr_tab = tabgain170;
}
else {
ptr_tab = tabgain170;
}
beta = ptr_tab[(int)AcGn]; /* beta = gain * 8192 */
if(Lag <= (SubFrLen/2)) {
Worst0 = L_mls(handle->CodStat.Err[0], beta);
Worst0 = L_shl(Worst0, 2);
Worst0 = L_add(Err0, Worst0);
Worst1 = Worst0;
}
else {
iz = mult(Lag, 1092); /* Lag / 30 */
temp1 = add(iz, 1);
temp2 = sub(shl(temp1, 5), shl(temp1, 1)); /* 30 (iz+1) */
if(temp2 != Lag) {
if(iz == 1) {
Worst0 = L_mls(handle->CodStat.Err[0], beta);
Worst0 = L_shl(Worst0, 2);
Worst0 = L_add(Err0, Worst0);
Worst1 = L_mls(handle->CodStat.Err[1], beta);
Worst1 = L_shl(Worst1, 2);
Worst1 = L_add(Err0, Worst1);
if(Worst0 > Worst1) Worst1 = Worst0;
else Worst0 = Worst1;
}
else {
Worst0 = L_mls(handle->CodStat.Err[iz-2], beta);
Worst0 = L_shl(Worst0, 2);
Worst0 = L_add(Err0, Worst0);
L_temp = L_mls(handle->CodStat.Err[iz-1], beta);
L_temp = L_shl(L_temp, 2);
L_temp = L_add(Err0, L_temp);
if(L_temp > Worst0) Worst0 = L_temp;
Worst1 = L_mls(handle->CodStat.Err[iz], beta);
Worst1 = L_shl(Worst1, 2);
Worst1 = L_add(Err0, Worst1);
if(L_temp > Worst1) Worst1 = L_temp;
}
}
else { /* Lag % SubFrLen = 0 */
Worst0 = L_mls(handle->CodStat.Err[iz-1], beta);
Worst0 = L_shl(Worst0, 2);
Worst0 = L_add(Err0, Worst0);
Worst1 = L_mls(handle->CodStat.Err[iz], beta);
Worst1 = L_shl(Worst1, 2);
Worst1 = L_add(Err0, Worst1);
}
}
for(i=4; i>=2; i--) {
handle->CodStat.Err[i] = handle->CodStat.Err[i-2];
}
handle->CodStat.Err[0] = Worst0;
handle->CodStat.Err[1] = Worst1;
return;
}
void Coder_ld8g(
Word16 ana[], /* (o) : analysis parameters */
Word16 frame, /* input : frame counter */
Word16 dtx_enable, /* input : DTX enable flag */
Word16 rate /* input : rate selector/frame =0 6.4kbps , =1 8kbps,= 2 11.8 kbps*/
)
{
/* LPC analysis */
Word16 r_l_fwd[NP+1], r_h_fwd[NP+1]; /* Autocorrelations low and hi (forward) */
Word32 r_bwd[M_BWDP1]; /* Autocorrelations (backward) */
Word16 r_l_bwd[M_BWDP1]; /* Autocorrelations low (backward) */
Word16 r_h_bwd[M_BWDP1]; /* Autocorrelations high (backward) */
Word16 rc_fwd[M]; /* Reflection coefficients : forward analysis */
Word16 rc_bwd[M_BWD]; /* Reflection coefficients : backward analysis */
Word16 A_t_fwd[MP1*2]; /* A(z) forward unquantized for the 2 subframes */
Word16 A_t_fwd_q[MP1*2]; /* A(z) forward quantized for the 2 subframes */
Word16 A_t_bwd[2*M_BWDP1]; /* A(z) backward for the 2 subframes */
Word16 *Aq; /* A(z) "quantized" for the 2 subframes */
Word16 *Ap; /* A(z) "unquantized" for the 2 subframes */
Word16 *pAp, *pAq;
Word16 Ap1[M_BWDP1]; /* A(z) with spectral expansion */
Word16 Ap2[M_BWDP1]; /* A(z) with spectral expansion */
Word16 lsp_new[M], lsp_new_q[M]; /* LSPs at 2th subframe */
Word16 lsf_int[M]; /* Interpolated LSF 1st subframe. */
Word16 lsf_new[M];
Word16 lp_mode; /* Backward / Forward Indication mode */
Word16 m_ap, m_aq, i_gamma;
Word16 code_lsp[2];
/* Other vectors */
Word16 h1[L_SUBFR]; /* Impulse response h1[] */
Word16 xn[L_SUBFR]; /* Target vector for pitch search */
Word16 xn2[L_SUBFR]; /* Target vector for codebook search */
Word16 code[L_SUBFR]; /* Fixed codebook excitation */
Word16 y1[L_SUBFR]; /* Filtered adaptive excitation */
Word16 y2[L_SUBFR]; /* Filtered fixed codebook excitation */
Word16 g_coeff[4]; /* Correlations between xn & y1 */
Word16 res2[L_SUBFR]; /* residual after long term prediction*/
Word16 g_coeff_cs[5];
Word16 exp_g_coeff_cs[5]; /* Correlations between xn, y1, & y2
<y1,y1>, -2<xn,y1>,
<y2,y2>, -2<xn,y2>, 2<y1,y2> */
/* Scalars */
Word16 i, j, k, i_subfr;
Word16 T_op, T0, T0_min, T0_max, T0_frac;
Word16 gain_pit, gain_code, index;
Word16 taming, pit_sharp;
Word16 sat_filter;
Word32 L_temp;
Word16 freq_cur[M];
/* For G.729B */
Word16 rh_nbe[MP1];
Word16 lsfq_mem[MA_NP][M];
Word16 exp_R0, Vad;
Word16 tmp1, tmp2,avg_lag;
Word16 temp, Energy_db;
/*------------------------------------------------------------------------*
* - Perform LPC analysis: *
* * autocorrelation + lag windowing *
* * Levinson-durbin algorithm to find a[] *
* * convert a[] to lsp[] *
* * quantize and code the LSPs *
* * find the interpolated LSPs and convert to a[] for the 2 *
* subframes (both quantized and unquantized) *
*------------------------------------------------------------------------*/
/* ------------------- */
/* LP Forward analysis */
/* ------------------- */
Autocorrg(p_window, NP, r_h_fwd, r_l_fwd, &exp_R0); /* Autocorrelations */
Copy(r_h_fwd, rh_nbe, MP1);
Lag_window(NP, r_h_fwd, r_l_fwd); /* Lag windowing */
Levinsong(M, r_h_fwd, r_l_fwd, &A_t_fwd[MP1], rc_fwd, old_A_fwd, old_rc_fwd,&temp ); /* Levinson Durbin */
Az_lsp(&A_t_fwd[MP1], lsp_new, lsp_old); /* From A(z) to lsp */
/* For G.729B */
/* ------ VAD ------- */
if (dtx_enable == 1) {
Lsp_lsf(lsp_new, lsf_new, M);
vadg(rc_fwd[1], lsf_new, r_h_fwd, r_l_fwd, exp_R0, p_window, frame,
pastVad, ppastVad, &Vad, &Energy_db);
musdetect( rate, r_h_fwd[0], r_l_fwd[0], exp_R0,rc_fwd ,lag_buf , pgain_buf,
prev_lp_mode, frame,pastVad, &Vad, Energy_db);
Update_cng(rh_nbe, exp_R0, Vad);
}
else Vad = 1;
/* -------------------- */
/* LP Backward analysis */
/* -------------------- */
/* -------------------- */
/* LP Backward analysis */
/* -------------------- */
if ( (rate-(1-Vad))== G729E) {
/* LPC recursive Window as in G728 */
autocorr_hyb_window(synth, r_bwd, rexp); /* Autocorrelations */
Lag_window_bwd(r_bwd, r_h_bwd, r_l_bwd); /* Lag windowing */
/* Fixed Point Levinson (as in G729) */
Levinsong(M_BWD, r_h_bwd, r_l_bwd, &A_t_bwd[M_BWDP1], rc_bwd,
old_A_bwd, old_rc_bwd, &temp );
/* Tests saturation of A_t_bwd */
sat_filter = 0;
for (i=M_BWDP1; i<2*M_BWDP1; i++) if (A_t_bwd[i] >= 32767) sat_filter = 1;
if (sat_filter == 1) Copy(A_t_bwd_mem, &A_t_bwd[M_BWDP1], M_BWDP1);
else Copy(&A_t_bwd[M_BWDP1], A_t_bwd_mem, M_BWDP1);
/* Additional bandwidth expansion on backward filter */
Weight_Az(&A_t_bwd[M_BWDP1], GAMMA_BWD, M_BWD, &A_t_bwd[M_BWDP1]);
}
/*--------------------------------------------------*
* Update synthesis signal for next frame. *
*--------------------------------------------------*/
Copy(&synth[L_FRAME], &synth[0], MEM_SYN_BWD);
/*--------------------------------------------------------------------*
* Find interpolated LPC parameters in all subframes (unquantized). *
* The interpolated parameters are in array A_t[] of size (M+1)*4 *
*--------------------------------------------------------------------*/
if( prev_lp_mode == 0) {
Int_lpc(lsp_old, lsp_new, lsf_int, lsf_new, A_t_fwd);
}
else {
/* no interpolation */
/* unquantized */
Lsp_Az(lsp_new, A_t_fwd); /* Subframe 1 */
Lsp_lsf(lsp_new, lsf_new, M); /* transformation from LSP to LSF (freq.domain) */
Copy(lsf_new, lsf_int, M); /* Subframe 1 */
}
if(Vad == 1) {
/* ---------------- */
/* LSP quantization */
/* ---------------- */
Qua_lspe(lsp_new, lsp_new_q, code_lsp, freq_prev, freq_cur);
/*--------------------------------------------------------------------*
* Find interpolated LPC parameters in all subframes (quantized) *
* the quantized interpolated parameters are in array Aq_t[] *
*--------------------------------------------------------------------*/
if( prev_lp_mode == 0) {
Int_qlpc(lsp_old_q, lsp_new_q, A_t_fwd_q);
}
else {
/* no interpolation */
Lsp_Az(lsp_new_q, &A_t_fwd_q[MP1]); /* Subframe 2 */
Copy(&A_t_fwd_q[MP1], A_t_fwd_q, MP1); /* Subframe 1 */
}
/*---------------------------------------------------------------------*
* - Decision for the switch Forward / Backward *
*---------------------------------------------------------------------*/
if(rate == G729E) {
set_lpc_modeg(speech, A_t_fwd_q, A_t_bwd, &lp_mode,
lsp_new, lsp_old, &bwd_dominant, prev_lp_mode, prev_filter,
&C_int, &glob_stat, &stat_bwd, &val_stat_bwd);
}
else {
update_bwd( &lp_mode, &bwd_dominant, &C_int, &glob_stat);
}
}
else update_bwd( &lp_mode, &bwd_dominant, &C_int, &glob_stat);
/* ---------------------------------- */
/* update the LSPs for the next frame */
/* ---------------------------------- */
Copy(lsp_new, lsp_old, M);
/*----------------------------------------------------------------------*
* - Find the weighted input speech w_sp[] for the whole speech frame *
*----------------------------------------------------------------------*/
if(lp_mode == 0) {
m_ap = M;
if (bwd_dominant == 0) Ap = A_t_fwd;
else Ap = A_t_fwd_q;
perc_var(gamma1, gamma2, lsf_int, lsf_new, rc_fwd);
}
else {
if (bwd_dominant == 0) {
m_ap = M;
Ap = A_t_fwd;
}
else {
m_ap = M_BWD;
Ap = A_t_bwd;
}
perc_vare(gamma1, gamma2, bwd_dominant);
}
pAp = Ap;
for (i=0; i<2; i++) {
Weight_Az(pAp, gamma1[i], m_ap, Ap1);
Weight_Az(pAp, gamma2[i], m_ap, Ap2);
Residue(m_ap, Ap1, &speech[i*L_SUBFR], &wsp[i*L_SUBFR], L_SUBFR);
Syn_filte(m_ap, Ap2, &wsp[i*L_SUBFR], &wsp[i*L_SUBFR], L_SUBFR,
&mem_w[M_BWD-m_ap], 0);
for(j=0; j<M_BWD; j++) mem_w[j] = wsp[i*L_SUBFR+L_SUBFR-M_BWD+j];
pAp += m_ap+1;
}
/* ---------------------- */
/* Case of Inactive frame */
/* ---------------------- */
if (Vad == 0){
for (i=0; i<MA_NP; i++) Copy(&freq_prev[i][0], &lsfq_mem[i][0], M);
Cod_cngg(exc, pastVad, lsp_old_q, old_A_fwd, old_rc_fwd, A_t_fwd_q, ana, lsfq_mem, &seed);
for (i=0; i<MA_NP; i++) Copy(&lsfq_mem[i][0], &freq_prev[i][0], M);
ppastVad = pastVad;
pastVad = Vad;
/* UPDATE wsp, mem_w, mem_syn, mem_err, and mem_w0 */
pAp = A_t_fwd; /* pointer to interpolated LPC parameters */
pAq = A_t_fwd_q; /* pointer to interpolated quantized LPC parameters */
i_gamma = 0;
for(i_subfr=0; i_subfr < L_FRAME; i_subfr += L_SUBFR) {
Weight_Az(pAp, gamma1[i_gamma], M, Ap1);
Weight_Az(pAp, gamma2[i_gamma], M, Ap2);
i_gamma = add(i_gamma,1);
/* update mem_syn */
Syn_filte(M, pAq, &exc[i_subfr], &synth_ptr[i_subfr], L_SUBFR, &mem_syn[M_BWD-M], 0);
for(j=0; j<M_BWD; j++) mem_syn[j] = synth_ptr[i_subfr+L_SUBFR-M_BWD+j];
/* update mem_w0 */
for (i=0; i<L_SUBFR; i++)
error[i] = speech[i_subfr+i] - synth_ptr[i_subfr+i];
Residue(M, Ap1, error, xn, L_SUBFR);
Syn_filte(M, Ap2, xn, xn, L_SUBFR, &mem_w0[M_BWD-M], 0);
for(j=0; j<M_BWD; j++) mem_w0[j] = xn[L_SUBFR-M_BWD+j];
/* update mem_err */
for (i = L_SUBFR-M_BWD, j = 0; i < L_SUBFR; i++, j++)
mem_err[j] = error[i];
for (i= 0; i< 4; i++)
pgain_buf[i] = pgain_buf[i+1];
pgain_buf[4] = 8192;
pAp += MP1;
pAq += MP1;
}
/* update previous filter for next frame */
Copy(&A_t_fwd_q[MP1], prev_filter, MP1);
for(i=MP1; i <M_BWDP1; i++) prev_filter[i] = 0;
prev_lp_mode = lp_mode;
sharp = SHARPMIN;
/* Update memories for next frames */
Copy(&old_speech[L_FRAME], &old_speech[0], L_TOTAL-L_FRAME);
Copy(&old_wsp[L_FRAME], &old_wsp[0], PIT_MAX);
Copy(&old_exc[L_FRAME], &old_exc[0], PIT_MAX+L_INTERPOL);
return;
} /* End of inactive frame case */
/* -------------------- */
/* Case of Active frame */
/* -------------------- */
*ana++ = rate+ (Word16)2; /* bit rate mode */
if(lp_mode == 0) {
m_aq = M;
Aq = A_t_fwd_q;
/* update previous filter for next frame */
Copy(&Aq[MP1], prev_filter, MP1);
for(i=MP1; i <M_BWDP1; i++) prev_filter[i] = 0;
for(j=MP1; j<M_BWDP1; j++) ai_zero[j] = 0;
}
else {
m_aq = M_BWD;
Aq = A_t_bwd;
if (bwd_dominant == 0) {
for(j=MP1; j<M_BWDP1; j++) ai_zero[j] = 0;
}
/* update previous filter for next frame */
Copy(&Aq[M_BWDP1], prev_filter, M_BWDP1);
}
if(dtx_enable == 1) {
seed = INIT_SEED;
ppastVad = pastVad;
pastVad = Vad;
}
if (rate == G729E) *ana++ = lp_mode;
/*----------------------------------------------------------------------*
* - Find the weighted input speech w_sp[] for the whole speech frame *
* - Find the open-loop pitch delay *
*----------------------------------------------------------------------*/
if( lp_mode == 0) {
Copy(lsp_new_q, lsp_old_q, M);
Lsp_prev_update(freq_cur, freq_prev);
*ana++ = code_lsp[0];
*ana++ = code_lsp[1];
}
/* Find open loop pitch lag */
T_op = Pitch_ol(wsp, PIT_MIN, PIT_MAX, L_FRAME);
for (i= 0; i< 4; i++)
lag_buf[i] = lag_buf[i+1];
avg_lag = add(lag_buf[0],lag_buf[1]);
avg_lag = add(avg_lag,lag_buf[2]);
avg_lag = add(avg_lag,lag_buf[3]);
avg_lag = mult_r(avg_lag,8192);
tmp1 = sub(T_op,shl(avg_lag,1));
tmp2 = sub(T_op,add(shl(avg_lag,1),avg_lag));
if( sub(abs_s(tmp1), 4)<0){
lag_buf[4] = shr(T_op,1);
}
else if( sub(abs_s(tmp2),6)<0){
lag_buf[4] = mult(T_op,10923);
}
else{
lag_buf[4] = T_op;
}
/* Range for closed loop pitch search in 1st subframe */
T0_min = sub(T_op, 3);
if (sub(T0_min,PIT_MIN)<0) {
T0_min = PIT_MIN;
}
T0_max = add(T0_min, 6);
if (sub(T0_max ,PIT_MAX)>0)
{
T0_max = PIT_MAX;
T0_min = sub(T0_max, 6);
}
/*------------------------------------------------------------------------*
* Loop for every subframe in the analysis frame *
*------------------------------------------------------------------------*
* To find the pitch and innovation parameters. The subframe size is *
* L_SUBFR and the loop is repeated 2 times. *
* - find the weighted LPC coefficients *
* - find the LPC residual signal res[] *
* - compute the target signal for pitch search *
* - compute impulse response of weighted synthesis filter (h1[]) *
* - find the closed-loop pitch parameters *
* - encode the pitch delay *
* - update the impulse response h1[] by including fixed-gain pitch *
* - find target vector for codebook search *
* - codebook search *
* - encode codebook address *
* - VQ of pitch and codebook gains *
* - find synthesis speech *
* - update states of weighting filter *
*------------------------------------------------------------------------*/
pAp = Ap; /* pointer to interpolated "unquantized"LPC parameters */
pAq = Aq; /* pointer to interpolated "quantized" LPC parameters */
i_gamma = 0;
for (i_subfr = 0; i_subfr < L_FRAME; i_subfr += L_SUBFR) {
/*---------------------------------------------------------------*
* Find the weighted LPC coefficients for the weighting filter. *
*---------------------------------------------------------------*/
Weight_Az(pAp, gamma1[i_gamma], m_ap, Ap1);
Weight_Az(pAp, gamma2[i_gamma], m_ap, Ap2);
/*---------------------------------------------------------------*
* Compute impulse response, h1[], of weighted synthesis filter *
*---------------------------------------------------------------*/
for (i = 0; i <=m_ap; i++) ai_zero[i] = Ap1[i];
Syn_filte(m_aq, pAq, ai_zero, h1, L_SUBFR, zero, 0);
Syn_filte(m_ap, Ap2, h1, h1, L_SUBFR, zero, 0);
/*------------------------------------------------------------------------*
* *
* Find the target vector for pitch search: *
* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ *
* *
* |------| res[n] *
* speech[n]---| A(z) |-------- *
* |------| | |--------| error[n] |------| *
* zero -- (-)--| 1/A(z) |-----------| W(z) |-- target *
* exc |--------| |------| *
* *
* Instead of subtracting the zero-input response of filters from *
* the weighted input speech, the above configuration is used to *
* compute the target vector. This configuration gives better performance *
* with fixed-point implementation. The memory of 1/A(z) is updated by *
* filtering (res[n]-exc[n]) through 1/A(z), or simply by subtracting *
* the synthesis speech from the input speech: *
* error[n] = speech[n] - syn[n]. *
* The memory of W(z) is updated by filtering error[n] through W(z), *
* or more simply by subtracting the filtered adaptive and fixed *
* codebook excitations from the target: *
* target[n] - gain_pit*y1[n] - gain_code*y2[n] *
* as these signals are already available. *
* *
*------------------------------------------------------------------------*/
Residue(m_aq, pAq, &speech[i_subfr], &exc[i_subfr], L_SUBFR); /* LPC residual */
for (i=0; i<L_SUBFR; i++) res2[i] = exc[i_subfr+i];
Syn_filte(m_aq, pAq, &exc[i_subfr], error, L_SUBFR,
&mem_err[M_BWD-m_aq], 0);
Residue(m_ap, Ap1, error, xn, L_SUBFR);
Syn_filte(m_ap, Ap2, xn, xn, L_SUBFR, &mem_w0[M_BWD-m_ap], 0); /* target signal xn[]*/
/*----------------------------------------------------------------------*
* Closed-loop fractional pitch search *
*----------------------------------------------------------------------*/
T0 = Pitch_fr3(&exc[i_subfr], xn, h1, L_SUBFR, T0_min, T0_max,
i_subfr, &T0_frac);
index = Enc_lag3(T0, T0_frac, &T0_min, &T0_max,PIT_MIN,PIT_MAX,
i_subfr);
*ana++ = index;
if ( (i_subfr == 0) ) {
*ana = Parity_Pitch(index);
if( rate == G729E) {
*ana ^= (shr(index, 1) & 0x0001);
}
ana++;
}
/*-----------------------------------------------------------------*
* - find unity gain pitch excitation (adaptive codebook entry) *
* with fractional interpolation. *
* - find filtered pitch exc. y1[]=exc[] convolve with h1[]) *
* - compute pitch gain and limit between 0 and 1.2 *
* - update target vector for codebook search *
* - find LTP residual. *
*-----------------------------------------------------------------*/
Pred_lt_3(&exc[i_subfr], T0, T0_frac, L_SUBFR);
Convolve(&exc[i_subfr], h1, y1, L_SUBFR);
gain_pit = G_pitch(xn, y1, g_coeff, L_SUBFR);
/* clip pitch gain if taming is necessary */
taming = test_err(T0, T0_frac);
if( taming == 1){
if (sub(gain_pit, GPCLIP) > 0) {
gain_pit = GPCLIP;
}
}
/* xn2[i] = xn[i] - y1[i] * gain_pit */
for (i = 0; i < L_SUBFR; i++) {
L_temp = L_mult(y1[i], gain_pit);
L_temp = L_shl(L_temp, 1); /* gain_pit in Q14 */
xn2[i] = sub(xn[i], extract_h(L_temp));
}
/*-----------------------------------------------------*
* - Innovative codebook search. *
*-----------------------------------------------------*/
switch (rate) {
case G729: /* 8 kbit/s */
{
/* case 8 kbit/s */
index = ACELP_Codebook(xn2, h1, T0, sharp, i_subfr, code, y2, &i);
*ana++ = index; /* Positions index */
*ana++ = i; /* Signs index */
break;
}
case G729E: /* 11.8 kbit/s */
{
/*-----------------------------------------------------------------*
* Include fixed-gain pitch contribution into impulse resp. h[] *
*-----------------------------------------------------------------*/
pit_sharp = shl(sharp, 1); /* From Q14 to Q15 */
if(T0 < L_SUBFR) {
for (i = T0; i < L_SUBFR; i++){ /* h[i] += pitch_sharp*h[i-T0] */
h1[i] = add(h1[i], mult(h1[i-T0], pit_sharp));
}
}
/* calculate residual after long term prediction */
/* res2[i] -= exc[i+i_subfr] * gain_pit */
for (i = 0; i < L_SUBFR; i++) {
L_temp = L_mult(exc[i+i_subfr], gain_pit);
L_temp = L_shl(L_temp, 1); /* gain_pit in Q14 */
res2[i] = sub(res2[i], extract_h(L_temp));
}
if (lp_mode == 0) ACELP_10i40_35bits(xn2, res2, h1, code, y2, ana); /* Forward */
else ACELP_12i40_44bits(xn2, res2, h1, code, y2, ana); /* Backward */
ana += 5;
/*-----------------------------------------------------------------*
* Include fixed-gain pitch contribution into code[]. *
*-----------------------------------------------------------------*/
if(T0 < L_SUBFR) {
for (i = T0; i < L_SUBFR; i++) { /* code[i] += pitch_sharp*code[i-T0] */
code[i] = add(code[i], mult(code[i-T0], pit_sharp));
}
}
break;
}
default : {
printf("Unrecognized bit rate\n");
exit(-1);
}
} /* end of switch */
/*-----------------------------------------------------*
* - Quantization of gains. *
*-----------------------------------------------------*/
g_coeff_cs[0] = g_coeff[0]; /* <y1,y1> */
exp_g_coeff_cs[0] = negate(g_coeff[1]); /* Q-Format:XXX -> JPN */
g_coeff_cs[1] = negate(g_coeff[2]); /* (xn,y1) -> -2<xn,y1> */
exp_g_coeff_cs[1] = negate(add(g_coeff[3], 1)); /* Q-Format:XXX -> JPN */
Corr_xy2( xn, y1, y2, g_coeff_cs, exp_g_coeff_cs ); /* Q0 Q0 Q12 ^Qx ^Q0 */
/* g_coeff_cs[3]:exp_g_coeff_cs[3] = <y2,y2> */
/* g_coeff_cs[4]:exp_g_coeff_cs[4] = -2<xn,y2> */
/* g_coeff_cs[5]:exp_g_coeff_cs[5] = 2<y1,y2> */
index = Qua_gain(code, g_coeff_cs, exp_g_coeff_cs, L_SUBFR,
&gain_pit, &gain_code, taming);
*ana++ = index;
/*------------------------------------------------------------*
* - Update pitch sharpening "sharp" with quantized gain_pit *
*------------------------------------------------------------*/
for (i= 0; i< 4; i++)
pgain_buf[i] = pgain_buf[i+1];
pgain_buf[4] = gain_pit;
sharp = gain_pit;
if (sub(sharp, SHARPMAX) > 0) sharp = SHARPMAX;
else {
if (sub(sharp, SHARPMIN) < 0) sharp = SHARPMIN;
}
/*------------------------------------------------------*
* - Find the total excitation *
* - find synthesis speech corresponding to exc[] *
* - update filters memories for finding the target *
* vector in the next subframe *
* (update error[-m..-1] and mem_w_err[]) *
* update error function for taming process *
*------------------------------------------------------*/
for (i = 0; i < L_SUBFR; i++) {
/* exc[i] = gain_pit*exc[i] + gain_code*code[i]; */
/* exc[i] in Q0 gain_pit in Q14 */
/* code[i] in Q13 gain_cod in Q1 */
L_temp = L_mult(exc[i+i_subfr], gain_pit);
L_temp = L_mac(L_temp, code[i], gain_code);
L_temp = L_shl(L_temp, 1);
exc[i+i_subfr] = round(L_temp);
}
update_exc_err(gain_pit, T0);
Syn_filte(m_aq, pAq, &exc[i_subfr], &synth_ptr[i_subfr], L_SUBFR,
&mem_syn[M_BWD-m_aq], 0);
for(j=0; j<M_BWD; j++) mem_syn[j] = synth_ptr[i_subfr+L_SUBFR-M_BWD+j];
for (i = L_SUBFR-M_BWD, j = 0; i < L_SUBFR; i++, j++) {
mem_err[j] = sub(speech[i_subfr+i], synth_ptr[i_subfr+i]);
temp = extract_h(L_shl( L_mult(y1[i], gain_pit), 1) );
k = extract_h(L_shl( L_mult(y2[i], gain_code), 2) );
mem_w0[j] = sub(xn[i], add(temp, k));
}
pAp += m_ap+1;
pAq += m_aq+1;
i_gamma = add(i_gamma,1);
}
/*--------------------------------------------------*
* Update signal for next frame. *
* -> shift to the left by L_FRAME: *
* speech[], wsp[] and exc[] *
*--------------------------------------------------*/
Copy(&old_speech[L_FRAME], &old_speech[0], L_TOTAL-L_FRAME);
Copy(&old_wsp[L_FRAME], &old_wsp[0], PIT_MAX);
Copy(&old_exc[L_FRAME], &old_exc[0], PIT_MAX+L_INTERPOL);
prev_lp_mode = lp_mode;
return;
}
/*----------------------------------------------------------------------------
; FUNCTION CODE
----------------------------------------------------------------------------*/
Word16 hp_max(
Word32 corr[], /* i : correlation vector. */
Word16 scal_sig[], /* i : scaled signal. */
Word16 L_frame, /* i : length of frame to compute pitch */
Word16 lag_max, /* i : maximum lag */
Word16 lag_min, /* i : minimum lag */
Word16 *cor_hp_max, /* o : max high-pass filtered norm. correlation */
Flag *pOverflow /* i/o : overflow Flag */
)
{
Word16 i;
Word16 *p, *p1;
Word32 max, t0, t1;
Word16 max16, t016, cor_max;
Word16 shift, shift1, shift2;
Word32 L_temp;
max = MIN_32;
t0 = 0L;
for (i = lag_max - 1; i > lag_min; i--)
{
/* high-pass filtering */
t0 = L_shl(corr[-i], 1, pOverflow);
L_temp = L_sub(t0, corr[-i-1], pOverflow);
t0 = L_sub(L_temp, corr[-i+1], pOverflow);
t0 = L_abs(t0);
if (t0 >= max)
{
max = t0;
}
}
/* compute energy */
p = scal_sig;
p1 = &scal_sig[0];
t0 = 0L;
for (i = 0; i < L_frame; i++, p++, p1++)
{
t0 = L_mac(t0, *p, *p1, pOverflow);
}
p = scal_sig;
p1 = &scal_sig[-1];
t1 = 0L;
for (i = 0; i < L_frame; i++, p++, p1++)
{
t1 = L_mac(t1, *p, *p1, pOverflow);
}
/* high-pass filtering */
L_temp = L_shl(t0, 1, pOverflow);
t1 = L_shl(t1, 1, pOverflow);
t0 = L_sub(L_temp, t1, pOverflow);
t0 = L_abs(t0);
/* max/t0 */
/* shift1 = sub(norm_l(max), 1);
max16 = extract_h(L_shl(max, shift1));
shift2 = norm_l(t0);
t016 = extract_h(L_shl(t0, shift2)); */
t016 = norm_l(max);
shift1 = sub(t016, 1, pOverflow);
L_temp = L_shl(max, shift1, pOverflow);
max16 = (Word16)(L_temp >> 16);
shift2 = norm_l(t0);
L_temp = L_shl(t0, shift2, pOverflow);
t016 = (Word16)(L_temp >> 16);
if (t016 != 0)
{
cor_max = div_s(max16, t016);
}
else
{
cor_max = 0;
}
shift = sub(shift1, shift2, pOverflow);
if (shift >= 0)
{
*cor_hp_max = shr(cor_max, shift, pOverflow); /* Q15 */
}
else
{
*cor_hp_max = shl(cor_max, negate(shift), pOverflow); /* Q15 */
}
return 0;
}
Word16 Pitch_ol( /* output: open loop pitch lag */
Word16 signal[], /* input : signal used to compute the open loop pitch */
/* signal[-pit_max] to signal[-1] should be known */
Word16 pit_min, /* input : minimum pitch lag */
Word16 pit_max, /* input : maximum pitch lag */
Word16 L_frame /* input : length of frame to compute pitch */
)
{
Word16 i, j;
Word16 max1, max2, max3;
Word16 p_max1, p_max2, p_max3;
Word32 t0, L_temp;
/* Scaled signal */
Word16 scaled_signal[L_FRAME+PIT_MAX];
Word16 *scal_sig;
scal_sig = &scaled_signal[pit_max];
/*--------------------------------------------------------*
* Verification for risk of overflow. *
*--------------------------------------------------------*/
Overflow = 0;
t0 = 0;
for(i= -pit_max; i< L_frame; i++)
t0 = L_mac(t0, signal[i], signal[i]);
/*--------------------------------------------------------*
* Scaling of input signal. *
* *
* if Overflow -> scal_sig[i] = signal[i]>>3 *
* else if t0 < 1^20 -> scal_sig[i] = signal[i]<<3 *
* else -> scal_sig[i] = signal[i] *
*--------------------------------------------------------*/
if(Overflow == 1)
{
for(i=-pit_max; i<L_frame; i++)
scal_sig[i] = shr(signal[i], 3);
}
else {
L_temp = L_sub(t0, (Word32)1048576L);
if ( L_temp < (Word32)0 ) /* if (t0 < 2^20) */
{
for(i=-pit_max; i<L_frame; i++)
scal_sig[i] = shl(signal[i], 3);
}
else
{
for(i=-pit_max; i<L_frame; i++)
scal_sig[i] = signal[i];
}
}
/*--------------------------------------------------------------------*
* The pitch lag search is divided in three sections. *
* Each section cannot have a pitch multiple. *
* We find a maximum for each section. *
* We compare the maximum of each section by favoring small lag. *
* *
* First section: lag delay = pit_max downto 4*pit_min *
* Second section: lag delay = 4*pit_min-1 downto 2*pit_min *
* Third section: lag delay = 2*pit_min-1 downto pit_min *
*--------------------------------------------------------------------*/
j = shl(pit_min, 2);
p_max1 = Lag_max(scal_sig, L_frame, pit_max, j, &max1);
i = sub(j, 1); j = shl(pit_min, 1);
p_max2 = Lag_max(scal_sig, L_frame, i, j, &max2);
i = sub(j, 1);
p_max3 = Lag_max(scal_sig, L_frame, i, pit_min , &max3);
/*--------------------------------------------------------------------*
* Compare the 3 sections maximum, and favor small lag. *
*--------------------------------------------------------------------*/
if( sub(mult(max1, THRESHPIT), max2) < 0)
{
max1 = max2;
p_max1 = p_max2;
}
if( sub(mult(max1, THRESHPIT), max3) < 0)
{
p_max1 = p_max3;
}
return (p_max1);
}
/*************************************************************************
*
* FUNCTION: Dec_gain()
*
* PURPOSE: Decode the pitch and codebook gains
*
************************************************************************/
void Dec_gain(
gc_predState *pred_state, /* i/o: MA predictor state */
enum Mode mode, /* i : AMR mode */
Word16 index, /* i : index of quantization. */
Word16 code[], /* i : Innovative vector. */
Word16 evenSubfr, /* i : Flag for even subframes */
Word16 * gain_pit, /* o : Pitch gain. */
Word16 * gain_cod /* o : Code gain. */
)
{
const Word16 *p;
Word16 frac, gcode0, exp, qua_ener, qua_ener_MR122;
Word16 g_code;
Word32 L_tmp;
/* Read the quantized gains (table depends on mode) */
index = shl (index, 2);
test(); test(); test();
if ( sub (mode, MR102) == 0
|| sub (mode, MR74) == 0
|| sub (mode, MR67) == 0)
{
p = &table_gain_highrates[index]; move16 ();
*gain_pit = *p++; move16 ();
g_code = *p++; move16 ();
qua_ener_MR122 = *p++; move16 ();
qua_ener = *p; move16 ();
}
else
{
test();
if (sub (mode, MR475) == 0)
{
index = add (index, shl(sub(1, evenSubfr), 1));
p = &table_gain_MR475[index]; move16 ();
*gain_pit = *p++; move16 ();
g_code = *p++; move16 ();
/*---------------------------------------------------------*
* calculate predictor update values (not stored in 4.75 *
* quantizer table to save space): *
* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ *
* *
* qua_ener = log2(g) *
* qua_ener_MR122 = 20*log10(g) *
*---------------------------------------------------------*/
/* Log2(x Q12) = log2(x) + 12 */
Log2 (L_deposit_l (g_code), &exp, &frac);
exp = sub(exp, 12);
qua_ener_MR122 = add (shr_r (frac, 5), shl (exp, 10));
/* 24660 Q12 ~= 6.0206 = 20*log10(2) */
L_tmp = Mpy_32_16(exp, frac, 24660);
qua_ener = round (L_shl (L_tmp, 13)); /* Q12 * Q0 = Q13 -> Q10 */
}
else
{
p = &table_gain_lowrates[index]; move16 ();
*gain_pit = *p++; move16 ();
g_code = *p++; move16 ();
qua_ener_MR122 = *p++; move16 ();
qua_ener = *p; move16 ();
}
}
/*-------------------------------------------------------------------*
* predict codebook gain *
* ~~~~~~~~~~~~~~~~~~~~~ *
* gc0 = Pow2(int(d)+frac(d)) *
* = 2^exp + 2^frac *
* *
* gcode0 (Q14) = 2^14*2^frac = gc0 * 2^(14-exp) *
*-------------------------------------------------------------------*/
gc_pred(pred_state, mode, code, &exp, &frac, NULL, NULL);
gcode0 = extract_l(Pow2(14, frac));
/*------------------------------------------------------------------*
* read quantized gains, update table of past quantized energies *
* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ *
* st->past_qua_en(Q10) = 20 * Log10(g_fac) / constant *
* = Log2(g_fac) *
* = qua_ener *
* constant = 20*Log10(2) *
*------------------------------------------------------------------*/
L_tmp = L_mult(g_code, gcode0);
L_tmp = L_shr(L_tmp, sub(10, exp));
*gain_cod = extract_h(L_tmp);
/* update table of past quantized energies */
gc_pred_update(pred_state, qua_ener_MR122, qua_ener);
return;
}
void musdetect(
Word16 rate,
Word16 r_h,
Word16 r_l,
Word16 exp_R0,
Word16 *rc,
Word16 *lags,
Word16 *pgains,
Word16 stat_flg,
Word16 frm_count,
Word16 prev_vad,
Word16 *Vad,
Word16 LLenergy)
{
Word16 i,j;
static Word16 count_music=0;
static Word16 Mcount_music=0;
static Word16 count_consc=0;
Word16 std;
static Word16 MeanPgain =8192;
Word16 PFLAG1, PFLAG2, PFLAG;
static Word16 count_pflag=0;
static Word16 Mcount_pflag=0;
static Word16 count_consc_pflag=0;
static Word16 count_consc_rflag=0;
static Word16 mrc[10]={0,0, 0, 0, 0, 0, 0, 0,0,0};
static Word16 MeanSE =0;
Word16 pderr, Lenergy , SD;
Word16 Thres, Coeff, C_Coeff;
Word32 acc0;
Word16 exp, frac, lagsum;
pderr =32767;
for (i=0; i<4; i++){
j = mult(rc[i], rc[i]);
j = sub(32767, j);
pderr = mult(pderr, j);
}
/* compute the frame energy */
acc0 = Mpy_32_16(r_h, r_l, pderr);
Log2(acc0, &exp, &frac);
acc0 = Mpy_32_16(exp, frac, 9864);
i = sub(exp_R0, 1);
i = sub(i, 1);
acc0 = L_mac(acc0, 9864, i);
acc0 = L_shl(acc0, 11);
Lenergy = extract_h(acc0);
Lenergy = sub(Lenergy, 4875);
acc0 = 0L;
for (i=0; i<10; i++){
j = sub(mrc[i], rc[i]);
acc0 = L_mac(acc0, j, j);
}
SD = extract_h(acc0);
if( *Vad == NOISE ){
for (i=0; i<10; i++){
acc0 = L_mult(29491, mrc[i]);
acc0 = L_mac(acc0, 3277, rc[i]);
mrc[i] = extract_h(acc0);
}
acc0 = L_mult(29491, MeanSE);
acc0 = L_mac(acc0, 3277, Lenergy);
MeanSE = extract_h(acc0);
}
/* determine the PFLAG */
acc0 = 0L;
lagsum = 0;
for (i=0; i<5; i++){
acc0 = L_mac(acc0, pgains[i], 6554); /* 1/5 in Q15 */
lagsum = add(lagsum, lags[i]);
}
acc0 = L_mult(extract_h(acc0), 6554);
acc0 = L_mac(acc0, MeanPgain, 26214);
MeanPgain = extract_h(acc0); /* compute the mean pitch gain */
acc0 = 0L;
for (i=0; i<5; i++){
/* NOTE: the variance of the lag is scaled up by 25 */
j = shl(lags[i], 2);
j = add(j,lags[i]); /* j equals 5*lags[i] */
j = sub(j, lagsum);
acc0 = L_mac(acc0, j, j);
}
acc0 = L_shl(acc0, 22);
/* NOTE: the final variance of the lag is scaled up by 25x128 */
std = extract_h(acc0);
if ( rate == G729D)
Thres = 11960; /* 0.73 in Q14 */
else
Thres = 10322; /* 0.63 in Q14 */
if ( sub(MeanPgain,Thres) > 0)
PFLAG2 =1;
else
PFLAG2 =0;
/* 21632 = 1.3*1.3*25*4*128*/
if ( (sub(std, 21632) < 0) && (sub(MeanPgain, 7373) > 0))
PFLAG1 =1;
else
PFLAG1 =0;
PFLAG= (Word16)( ((Word16)prev_vad & (Word16)(PFLAG1 | PFLAG2))| (Word16)(PFLAG2));
if( (sub(rc[1], 14746) <= 0) && (rc[1] >= 0) && (sub(MeanPgain,8192) < 0))
count_consc_rflag = add(count_consc_rflag,1);
else
count_consc_rflag =0;
if( (stat_flg == 1) && (*Vad == VOICE))
count_music = add(count_music,256); /* Q8 */
if( (frm_count & 0x003f) == 0){
if( frm_count == 64)
Mcount_music = count_music;
else{
acc0 = L_mult(29491, Mcount_music);
acc0 = L_mac(acc0, 3277, count_music);
Mcount_music = extract_h(acc0);
}
}
if( count_music == 0)
count_consc = add(count_consc,1);
else
count_consc = 0;
if( ((sub(count_consc, 500)>0) || (sub(count_consc_rflag , 150)>0))) Mcount_music = 0;
if( (frm_count & 0x003f) == 0){
count_music = 0;
}
if( PFLAG== 1 )
count_pflag = add(count_pflag,256); /* Q8 */
if( (frm_count & 0x003f) == 0){
if( frm_count == 64)
Mcount_pflag = count_pflag;
else{
if( sub(count_pflag , 6400)> 0){
Coeff = 32113;
C_Coeff = 655;
}
else if( sub(count_pflag , 5120)> 0){
Coeff = 31130;
C_Coeff = 1638;
}
else {
Coeff = 29491;
C_Coeff = 3277;
}
acc0 = L_mult(Coeff, Mcount_pflag);
acc0 = L_mac(acc0, C_Coeff, count_pflag);
Mcount_pflag = extract_h(acc0);
}
}
if( count_pflag == 0)
count_consc_pflag = add(count_consc_pflag,1);
else
count_consc_pflag = 0;
if( ((sub(count_consc_pflag, 100)>0) || (sub(count_consc_rflag , 150)>0))) Mcount_pflag = 0;
if( (frm_count & 0x003f) == 0)
count_pflag = 0;
if (rate == G729E){
if( (sub(SD,4915) > 0) && (sub(Lenergy ,MeanSE)> 819) && (sub(LLenergy,10240) >0 ) )
*Vad =VOICE;
else if( ((sub(SD,12452) > 0) || (sub(Lenergy ,MeanSE)> 819)) &&
(sub(LLenergy,10240) >0 ) )
*Vad =VOICE;
else if( ( (sub(Mcount_pflag ,2560) >0) || (sub(Mcount_music ,280)>0) || (sub(frm_count,64) < 0))
&& (sub(LLenergy,1433) >0))
*Vad =VOICE;
}
return;
}
int SetAlphaIntoColor(int col, int alpha)
{
int newCol = and(col,16777215) + shl(alpha,24);
return newCol;
}
void GSDrawScanlineCodeGenerator::Init(int params)
{
const int _top = params + 4;
const int _v = params + 8;
// int skip = left & 3;
mov(ebx, edx);
and(edx, 3);
// left -= skip;
sub(ebx, edx);
// int steps = right - left - 4;
sub(ecx, ebx);
sub(ecx, 4);
// GSVector4i test = m_test[skip] | m_test[7 + (steps & (steps >> 31))];
shl(edx, 4);
movdqa(xmm7, xmmword[edx + (size_t)&m_test[0]]);
mov(eax, ecx);
sar(eax, 31);
and(eax, ecx);
shl(eax, 4);
por(xmm7, xmmword[eax + (size_t)&m_test[7]]);
// GSVector2i* fza_base = &m_env.fzbr[top];
mov(esi, dword[esp + _top]);
lea(esi, ptr[esi * 8]);
add(esi, dword[&m_env.fzbr]);
// GSVector2i* fza_offset = &m_env.fzbc[left >> 2];
lea(edi, ptr[ebx * 2]);
add(edi, dword[&m_env.fzbc]);
if(!m_sel.sprite && (m_sel.fwrite && m_sel.fge || m_sel.zb) || m_sel.fb && (m_sel.edge || m_sel.tfx != TFX_NONE || m_sel.iip))
{
// edx = &m_env.d[skip]
shl(edx, 4);
lea(edx, ptr[edx + (size_t)m_env.d]);
// ebx = &v
mov(ebx, dword[esp + _v]);
}
if(!m_sel.sprite)
{
if(m_sel.fwrite && m_sel.fge || m_sel.zb)
{
movaps(xmm0, xmmword[ebx + 16]); // v.p
if(m_sel.fwrite && m_sel.fge)
{
// f = GSVector4i(vp).zzzzh().zzzz().add16(m_env.d[skip].f);
cvttps2dq(xmm1, xmm0);
pshufhw(xmm1, xmm1, _MM_SHUFFLE(2, 2, 2, 2));
pshufd(xmm1, xmm1, _MM_SHUFFLE(2, 2, 2, 2));
paddw(xmm1, xmmword[edx + 16 * 6]);
movdqa(xmmword[&m_env.temp.f], xmm1);
}
if(m_sel.zb)
{
// z = vp.zzzz() + m_env.d[skip].z;
shufps(xmm0, xmm0, _MM_SHUFFLE(2, 2, 2, 2));
addps(xmm0, xmmword[edx]);
movaps(xmmword[&m_env.temp.z], xmm0);
}
}
}
else
{
if(m_sel.ztest)
{
movdqa(xmm0, xmmword[&m_env.p.z]);
}
}
if(m_sel.fb)
{
if(m_sel.edge || m_sel.tfx != TFX_NONE)
{
movaps(xmm4, xmmword[ebx + 32]); // v.t
}
if(m_sel.edge)
{
pshufhw(xmm3, xmm4, _MM_SHUFFLE(2, 2, 2, 2));
pshufd(xmm3, xmm3, _MM_SHUFFLE(3, 3, 3, 3));
psrlw(xmm3, 9);
movdqa(xmmword[&m_env.temp.cov], xmm3);
}
if(m_sel.tfx != TFX_NONE)
{
if(m_sel.fst)
{
// GSVector4i vti(vt);
cvttps2dq(xmm4, xmm4);
// si = vti.xxxx() + m_env.d[skip].si;
// ti = vti.yyyy(); if(!sprite) ti += m_env.d[skip].ti;
pshufd(xmm2, xmm4, _MM_SHUFFLE(0, 0, 0, 0));
pshufd(xmm3, xmm4, _MM_SHUFFLE(1, 1, 1, 1));
paddd(xmm2, xmmword[edx + 16 * 7]);
if(!m_sel.sprite)
{
paddd(xmm3, xmmword[edx + 16 * 8]);
}
else
{
if(m_sel.ltf)
{
movdqa(xmm4, xmm3);
pshuflw(xmm4, xmm4, _MM_SHUFFLE(2, 2, 0, 0));
pshufhw(xmm4, xmm4, _MM_SHUFFLE(2, 2, 0, 0));
psrlw(xmm4, 1);
movdqa(xmmword[&m_env.temp.vf], xmm4);
}
}
movdqa(xmmword[&m_env.temp.s], xmm2);
movdqa(xmmword[&m_env.temp.t], xmm3);
}
else
{
// s = vt.xxxx() + m_env.d[skip].s;
// t = vt.yyyy() + m_env.d[skip].t;
// q = vt.zzzz() + m_env.d[skip].q;
movaps(xmm2, xmm4);
movaps(xmm3, xmm4);
shufps(xmm2, xmm2, _MM_SHUFFLE(0, 0, 0, 0));
shufps(xmm3, xmm3, _MM_SHUFFLE(1, 1, 1, 1));
shufps(xmm4, xmm4, _MM_SHUFFLE(2, 2, 2, 2));
addps(xmm2, xmmword[edx + 16 * 1]);
addps(xmm3, xmmword[edx + 16 * 2]);
addps(xmm4, xmmword[edx + 16 * 3]);
movaps(xmmword[&m_env.temp.s], xmm2);
movaps(xmmword[&m_env.temp.t], xmm3);
movaps(xmmword[&m_env.temp.q], xmm4);
rcpps(xmm4, xmm4);
mulps(xmm2, xmm4);
mulps(xmm3, xmm4);
}
}
if(!(m_sel.tfx == TFX_DECAL && m_sel.tcc))
{
if(m_sel.iip)
{
// GSVector4i vc = GSVector4i(v.c);
cvttps2dq(xmm6, xmmword[ebx]); // v.c
// vc = vc.upl16(vc.zwxy());
pshufd(xmm5, xmm6, _MM_SHUFFLE(1, 0, 3, 2));
punpcklwd(xmm6, xmm5);
// rb = vc.xxxx().add16(m_env.d[skip].rb);
// ga = vc.zzzz().add16(m_env.d[skip].ga);
pshufd(xmm5, xmm6, _MM_SHUFFLE(0, 0, 0, 0));
pshufd(xmm6, xmm6, _MM_SHUFFLE(2, 2, 2, 2));
paddw(xmm5, xmmword[edx + 16 * 4]);
paddw(xmm6, xmmword[edx + 16 * 5]);
movdqa(xmmword[&m_env.temp.rb], xmm5);
movdqa(xmmword[&m_env.temp.ga], xmm6);
}
else
{
if(m_sel.tfx == TFX_NONE)
{
movdqa(xmm5, xmmword[&m_env.c.rb]);
movdqa(xmm6, xmmword[&m_env.c.ga]);
}
}
}
}
}
void GSDrawScanlineCodeGenerator::Step()
{
// steps -= 4;
sub(ecx, 4);
// fza_offset++;
add(edi, 8);
if(!m_sel.sprite)
{
// z += m_env.d4.z;
if(m_sel.zb)
{
movaps(xmm0, xmmword[&m_env.temp.z]);
addps(xmm0, xmmword[&m_env.d4.z]);
movaps(xmmword[&m_env.temp.z], xmm0);
}
// f = f.add16(m_env.d4.f);
if(m_sel.fwrite && m_sel.fge)
{
movdqa(xmm1, xmmword[&m_env.temp.f]);
paddw(xmm1, xmmword[&m_env.d4.f]);
movdqa(xmmword[&m_env.temp.f], xmm1);
}
}
else
{
if(m_sel.ztest)
{
movdqa(xmm0, xmmword[&m_env.p.z]);
}
}
if(m_sel.fb)
{
if(m_sel.tfx != TFX_NONE)
{
if(m_sel.fst)
{
// GSVector4i st = m_env.d4.st;
// si += st.xxxx();
// if(!sprite) ti += st.yyyy();
movdqa(xmm4, xmmword[&m_env.d4.st]);
pshufd(xmm2, xmm4, _MM_SHUFFLE(0, 0, 0, 0));
paddd(xmm2, xmmword[&m_env.temp.s]);
movdqa(xmmword[&m_env.temp.s], xmm2);
if(!m_sel.sprite)
{
pshufd(xmm3, xmm4, _MM_SHUFFLE(1, 1, 1, 1));
paddd(xmm3, xmmword[&m_env.temp.t]);
movdqa(xmmword[&m_env.temp.t], xmm3);
}
else
{
movdqa(xmm3, xmmword[&m_env.temp.t]);
}
}
else
{
// GSVector4 stq = m_env.d4.stq;
// s += stq.xxxx();
// t += stq.yyyy();
// q += stq.zzzz();
movaps(xmm2, xmmword[&m_env.d4.stq]);
movaps(xmm3, xmm2);
movaps(xmm4, xmm2);
shufps(xmm2, xmm2, _MM_SHUFFLE(0, 0, 0, 0));
shufps(xmm3, xmm3, _MM_SHUFFLE(1, 1, 1, 1));
shufps(xmm4, xmm4, _MM_SHUFFLE(2, 2, 2, 2));
addps(xmm2, xmmword[&m_env.temp.s]);
addps(xmm3, xmmword[&m_env.temp.t]);
addps(xmm4, xmmword[&m_env.temp.q]);
movaps(xmmword[&m_env.temp.s], xmm2);
movaps(xmmword[&m_env.temp.t], xmm3);
movaps(xmmword[&m_env.temp.q], xmm4);
rcpps(xmm4, xmm4);
mulps(xmm2, xmm4);
mulps(xmm3, xmm4);
}
}
if(!(m_sel.tfx == TFX_DECAL && m_sel.tcc))
{
if(m_sel.iip)
{
// GSVector4i c = m_env.d4.c;
// rb = rb.add16(c.xxxx());
// ga = ga.add16(c.yyyy());
movdqa(xmm7, xmmword[&m_env.d4.c]);
pshufd(xmm5, xmm7, _MM_SHUFFLE(0, 0, 0, 0));
pshufd(xmm6, xmm7, _MM_SHUFFLE(1, 1, 1, 1));
paddw(xmm5, xmmword[&m_env.temp.rb]);
paddw(xmm6, xmmword[&m_env.temp.ga]);
movdqa(xmmword[&m_env.temp.rb], xmm5);
movdqa(xmmword[&m_env.temp.ga], xmm6);
}
else
{
if(m_sel.tfx == TFX_NONE)
{
movdqa(xmm5, xmmword[&m_env.c.rb]);
movdqa(xmm6, xmmword[&m_env.c.ga]);
}
}
}
}
// test = m_test[7 + (steps & (steps >> 31))];
mov(edx, ecx);
sar(edx, 31);
and(edx, ecx);
shl(edx, 4);
movdqa(xmm7, xmmword[edx + (size_t)&m_test[7]]);
}
/* Standard Long-Term Postfilter */
void postfilter(
Word16 *s, /* input : quantized speech signal */
Word16 pp, /* input : pitch period */
Word16 *ma_a,
Word16 *b_prv,
Word16 *pp_prv,
Word16 *e) /* output: enhanced speech signal */
{
int n;
Word16 len, t0, t1, t2, t3, shift, aa, R0norm, R0_exp;
Word32 a0, a1, R0, R1, R01, R01max, Rx;
Word16 *fp1;
Word16 ppt, pptmin, pptmax, ppnew;
Word16 bb[2];
Word16 R1max_exp, R1max, R01Sqmax_exp, R01Sqmax, R01Sq_exp, R01Sq, R1_exp, R1n;
Word16 gainn, Rx_exp;
Word16 buf[MAXPP+FRSZ];
Word16 *ps, ww1, ww2;
Word32 step, delta;
Word16 bi0, bi1c, bi1p;
ps = s+XQOFF;
/********************************************************************/
/* pitch search around decoded pitch */
/********************************************************************/
pptmin = sub(pp, DPPQNS);
pptmax = add(pp, DPPQNS);
if (pptmin<MINPP)
{
pptmin = MINPP;
pptmax = add(pptmin, 2*DPPQNS);
}
else if (pptmax>MAXPP)
{
pptmax = MAXPP;
pptmin = sub(pptmax, 2*DPPQNS);
}
fp1 = &s[XQOFF-pptmax];
len = add(FRSZ, pptmax);
a0 = 0;
for (n=0;n<len;n++)
{
t1 = shr(*fp1++, 3);
a0 = L_mac0(a0,t1,t1);
}
shift = norm_l(a0);
if (a0==0) shift=31;
shift = sub(6, shift);
if (shift > 0)
{
ps = buf+pptmax;
fp1 = &s[XQOFF-pptmax];
shift = shr(add(shift, 1), 1);
for (n=0;n<len;n++)
{
buf[n] = shr(fp1[n], shift);
}
}
else shift=0;
R0 = 0;
R1 = 0;
R01 = 0;
for(n=0; n<FRSZ; n++)
{
R0 = L_mac0(R0, ps[n], ps[n]);
R1 = L_mac0(R1, ps[n-pptmin], ps[n-pptmin]);
R01 = L_mac0(R01,ps[n], ps[n-pptmin]);
}
R0_exp = norm_l(R0);
R0norm = extract_h(L_shl(R0, R0_exp));
R0_exp = R0_exp-16;
ppnew = pptmin;
R1max_exp = norm_l(R1);
R1max = extract_h(L_shl(R1, R1max_exp));
R01Sqmax_exp = norm_l(R01);
t1 = extract_h(L_shl(R01, R01Sqmax_exp));
R01Sqmax_exp = shl(R01Sqmax_exp, 1);
R01Sqmax = extract_h(L_mult(t1, t1));
R01max = R01;
for(ppt=pptmin+1; ppt<=pptmax; ppt++)
{
R1 = L_msu0(R1,ps[FRSZ-ppt], ps[FRSZ-ppt]);
R1 = L_mac0(R1,ps[-ppt], ps[-ppt]);
R01= 0;
for(n=0; n<FRSZ; n++)
{
R01 = L_mac0(R01, ps[n], ps[n-ppt]);
}
R01Sq_exp = norm_l(R01);
t1 = extract_h(L_shl(R01, R01Sq_exp));
R01Sq_exp = shl(R01Sq_exp, 1);
R01Sq = extract_h(L_mult(t1, t1));
R1_exp = norm_l(R1);
R1n = extract_h(L_shl(R1, R1_exp));
a0 = L_mult(R01Sq, R1max);
a1 = L_mult(R01Sqmax, R1n);
t1 = add(R01Sq_exp, R1max_exp);
t2 = add(R01Sqmax_exp, R1_exp);
t2 = sub(t1, t2);
if (t2>=0) a0 = L_shr(a0, t2);
if (t2<0) a1 = L_shl(a1, t2);
if (L_sub(a0, a1)>0)
{
R01Sqmax = R01Sq;
R01Sqmax_exp = R01Sq_exp;
R1max = R1n; R1max_exp = R1_exp;
ppnew = ppt;
R01max = R01;
}
}
/******************************************************************/
/* calculate all-zero pitch postfilter */
/******************************************************************/
if (R1max==0 || R0==0 || R01max <= 0)
{
aa = 0;
}
else
{
a0 = R1max_exp-16;
t1 = mult(R1max, R0norm);
a0 = a0+R0_exp-15;
sqrt_i(t1, (Word16)a0, &t1, &t2);
t0 = norm_l(R01max);
t3 = extract_h(L_shl(R01max, t0));
t0 = t0-16;
aa = mult(t3, t1);
t0 = t0+t2-15;
t0 = t0-15;
if (t0<0) aa = shl(aa, sub(0,t0));
else aa = shr(aa, t0);
}
a0 = L_mult(8192, aa);
a0 = L_mac(a0, 24576, *ma_a);
*ma_a = intround(a0);
if((*ma_a < ATHLD1) && (aa < (ATHLD2)))
aa = 0;
bb[1] = mult(ScLTPF, aa);
/******************************************************************/
/* calculate normalization energies */
/******************************************************************/
Rx = 0;
R0 = 0;
for(n=0; n<FRSZ; n++)
{
a0 = L_shl(s[XQOFF+n], 15);
a0 = L_add(a0, L_mult0(bb[1], s[XQOFF+n-ppnew]));
e[n] = intround(a0);
t1 = shr(e[n], shift);
t2 = shr(s[XQOFF+n], shift);
Rx = L_mac0(Rx, t1, t1);
R0 = L_mac0(R0, t2, t2);
}
R0 = L_shr(R0, 2);
if(R0 == 0 || Rx == 0)
gainn = 32767;
else
{
Rx_exp = norm_l(Rx);
t1 = extract_h(L_shl(Rx, Rx_exp));
t2 = extract_h(L_shl(R0, Rx_exp));
if (t2>= t1)
gainn = 32767;
else
{
t1 = div_s(t2, t1);
gainn = sqrts(t1);
}
}
/******************************************************************/
/* interpolate from the previous postfilter to the current */
/******************************************************************/
bb[0] = gainn;
bb[1] = mult(gainn, bb[1]);
step = (Word32)((1.0/(NINT+1))*(2147483648.0));
delta = 0;
for(n=0; n<NINT; n++)
{
delta = L_add(delta, step);
ww1 = intround(delta);
ww2 = add(sub(32767, ww1), 1);
/* interpolate between two filters */
bi0 = intround(L_mac(L_mult(ww1, bb[0]), ww2, b_prv[0]));
bi1c= mult(ww1, bb[1]);
bi1p= mult(ww2, b_prv[1]);
e[n] = intround(L_mac(L_mac(L_mult(bi1c, s[XQOFF+n-ppnew]), bi1p, s[XQOFF+n-(*pp_prv)]), bi0, s[XQOFF+n]));
}
for(n=NINT; n<FRSZ; n++)
{
e[n] = intround(L_shl(L_mult(gainn, e[n]),1));
}
/******************************************************************/
/* save state memory */
/******************************************************************/
*pp_prv = ppnew;
b_prv[0] = bb[0];
b_prv[1] = bb[1];
return;
}
void GSDrawScanlineCodeGenerator::WriteFrame(int params)
{
const int _top = params + 4;
if(!m_sel.fwrite)
{
return;
}
if(m_sel.colclamp == 0)
{
// c[0] &= 0x000000ff;
// c[1] &= 0x000000ff;
pcmpeqd(xmm7, xmm7);
psrlw(xmm7, 8);
pand(xmm5, xmm7);
pand(xmm6, xmm7);
}
if(m_sel.fpsm == 2 && m_sel.dthe)
{
mov(eax, dword[esp + _top]);
and(eax, 3);
shl(eax, 5);
paddw(xmm5, xmmword[eax + (size_t)&m_env.dimx[0]]);
paddw(xmm6, xmmword[eax + (size_t)&m_env.dimx[1]]);
}
// GSVector4i fs = c[0].upl16(c[1]).pu16(c[0].uph16(c[1]));
movdqa(xmm7, xmm5);
punpcklwd(xmm5, xmm6);
punpckhwd(xmm7, xmm6);
packuswb(xmm5, xmm7);
if(m_sel.fba && m_sel.fpsm != 1)
{
// fs |= 0x80000000;
pcmpeqd(xmm7, xmm7);
pslld(xmm7, 31);
por(xmm5, xmm7);
}
if(m_sel.fpsm == 2)
{
// GSVector4i rb = fs & 0x00f800f8;
// GSVector4i ga = fs & 0x8000f800;
mov(eax, 0x00f800f8);
movd(xmm6, eax);
pshufd(xmm6, xmm6, _MM_SHUFFLE(0, 0, 0, 0));
mov(eax, 0x8000f800);
movd(xmm7, eax);
pshufd(xmm7, xmm7, _MM_SHUFFLE(0, 0, 0, 0));
movdqa(xmm4, xmm5);
pand(xmm4, xmm6);
pand(xmm5, xmm7);
// fs = (ga >> 16) | (rb >> 9) | (ga >> 6) | (rb >> 3);
movdqa(xmm6, xmm4);
movdqa(xmm7, xmm5);
psrld(xmm4, 3);
psrld(xmm6, 9);
psrld(xmm5, 6);
psrld(xmm7, 16);
por(xmm5, xmm4);
por(xmm7, xmm6);
por(xmm5, xmm7);
}
if(m_sel.rfb)
{
// fs = fs.blend(fd, fm);
blend(xmm5, xmm2, xmm3); // TODO: could be skipped in certain cases, depending on fpsm and fm
}
bool fast = m_sel.rfb && m_sel.fpsm < 2;
WritePixel(xmm5, xmm0, ebx, dl, fast, m_sel.fpsm);
}
void perc_var (
Word16 *gamma1, /* Bandwidth expansion parameter */
Word16 *gamma2, /* Bandwidth expansion parameter */
Word16 *LsfInt, /* Interpolated LSP vector : 1st subframe */
Word16 *LsfNew, /* New LSP vector : 2nd subframe */
Word16 *r_c /* Reflection coefficients */
)
{
Word32 L_temp;
Word16 cur_rc; /* Q11 */
Word16 Lar[4]; /* Q11 */
Word16 *LarNew; /* Q11 */
Word16 *Lsf; /* Q15 */
Word16 CritLar0, CritLar1; /* Q11 */
Word16 temp;
Word16 d_min; /* Q10 */
Word16 i, k;
for (k=0; k<M; k++) {
LsfInt[k] = shl(LsfInt[k], 1);
LsfNew[k] = shl(LsfNew[k], 1);
}
LarNew = &Lar[2];
/* ---------------------------------------- */
/* Reflection coefficients ---> Lar */
/* Lar(i) = log10( (1+rc) / (1-rc) ) */
/* Approximated by */
/* x <= SEG1 y = x */
/* SEG1 < x <= SEG2 y = A1 x - B1_L */
/* SEG2 < x <= SEG3 y = A2 x - B2_L */
/* x > SEG3 y = A3 x - B3_L */
/* ---------------------------------------- */
for (i=0; i<2; i++) {
cur_rc = abs_s(r_c[i]);
cur_rc = shr(cur_rc, 4);
if (sub(cur_rc ,SEG1)<= 0) {
LarNew[i] = cur_rc;
}
else {
if (sub(cur_rc,SEG2)<= 0) {
cur_rc = shr(cur_rc, 1);
L_temp = L_mult(cur_rc, A1);
L_temp = L_sub(L_temp, L_B1);
L_temp = L_shr(L_temp, 11);
LarNew[i] = extract_l(L_temp);
}
else {
if (sub(cur_rc ,SEG3)<= 0) {
cur_rc = shr(cur_rc, 1);
L_temp = L_mult(cur_rc, A2);
L_temp = L_sub(L_temp, L_B2);
L_temp = L_shr(L_temp, 11);
LarNew[i] = extract_l(L_temp);
}
else {
cur_rc = shr(cur_rc, 1);
L_temp = L_mult(cur_rc, A3);
L_temp = L_sub(L_temp, L_B3);
L_temp = L_shr(L_temp, 11);
LarNew[i] = extract_l(L_temp);
}
}
}
if (r_c[i] < 0) {
LarNew[i] = sub(0, LarNew[i]);
}
}
/* Interpolation of Lar for the 1st subframe */
temp = add(LarNew[0], LarOld[0]);
Lar[0] = shr(temp, 1);
LarOld[0] = LarNew[0];
temp = add(LarNew[1], LarOld[1]);
Lar[1] = shr(temp, 1);
LarOld[1] = LarNew[1];
for (k=0; k<2; k++) { /* LOOP : gamma2 for 1st to 2nd subframes */
/* ---------------------------------------------------------- */
/* First criterion based on the first two Lars */
/* smooth == 1 ==> gamma2 can vary from 0.4 to 0.7 */
/* smooth == 0 ==> gamma2 is set to 0.6 */
/* */
/* Double threshold + hysteresis : */
/* if smooth = 1 */
/* if (CritLar0 < THRESH_L1) and (CritLar1 > THRESH_H1) */
/* smooth = 0 */
/* if smooth = 0 */
/* if (CritLar0 > THRESH_L2) or (CritLar1 < THRESH_H2) */
/* smooth = 1 */
/* ---------------------------------------------------------- */
CritLar0 = Lar[2*k];
CritLar1 = Lar[2*k+1];
if (smooth != 0) {
if ((sub(CritLar0,THRESH_L1)<0)&&( sub(CritLar1,THRESH_H1)>0)) {
smooth = 0;
}
}
else {
if ( (sub(CritLar0 ,THRESH_L2)>0) || (sub(CritLar1,THRESH_H2) <0) ) {
smooth = 1;
}
}
if (smooth == 0) {
/* ------------------------------------------------------ */
/* Second criterion based on the minimum distance between */
/* two successives LSPs */
/* */
/* gamma2[k] = -6.0 * pi * d_min + 1.0 */
/* */
/* with Lsfs normalized range 0.0 <= val <= 1.0 */
/* ------------------------------------------------------ */
gamma1[k] = GAMMA1_0;
if (k == 0) {
Lsf = LsfInt;
}
else {
Lsf = LsfNew;
}
d_min = sub(Lsf[1], Lsf[0]);
for (i=1; i<M-1; i++) {
temp = sub(Lsf[i+1],Lsf[i]);
if (sub(temp,d_min)<0) {
d_min = temp;
}
}
temp = mult(ALPHA, d_min);
temp = sub(BETA, temp);
temp = shl(temp, 5);
gamma2[k] = temp;
if (sub(gamma2[k] , GAMMA2_0_H)>0) {
gamma2[k] = GAMMA2_0_H;
}
if (sub(gamma2[k] ,GAMMA2_0_L)<0) {
gamma2[k] = GAMMA2_0_L;
}
}
else {
gamma1[k] = GAMMA1_1;
gamma2[k] = GAMMA2_1;
}
}
return;
}
Word16 vad2 (Word16 * farray_ptr, vadState2 * st)
{
/*
* The channel table is defined below. In this table, the
* lower and higher frequency coefficients for each of the 16
* channels are specified. The table excludes the coefficients
* with numbers 0 (DC), 1, and 64 (Foldover frequency).
*/
const static Word16 ch_tbl[NUM_CHAN][2] =
{
{2, 3},
{4, 5},
{6, 7},
{8, 9},
{10, 11},
{12, 13},
{14, 16},
{17, 19},
{20, 22},
{23, 26},
{27, 30},
{31, 35},
{36, 41},
{42, 48},
{49, 55},
{56, 63}
};
/* channel energy scaling table - allows efficient division by number
* of DFT bins in the channel: 1/2, 1/3, 1/4, etc.
*/
const static Word16 ch_tbl_sh[NUM_CHAN] =
{
16384, 16384, 16384, 16384, 16384, 16384, 10923, 10923,
10923, 8192, 8192, 6554, 5461, 4681, 4681, 4096
};
/*
* The voice metric table is defined below. It is a non-
* linear table with a deadband near zero. It maps the SNR
* index (quantized SNR value) to a number that is a measure
* of voice quality.
*/
const static Word16 vm_tbl[90] =
{
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
3, 3, 3, 3, 3, 4, 4, 4, 5, 5, 5, 6, 6, 7, 7, 7,
8, 8, 9, 9, 10, 10, 11, 12, 12, 13, 13, 14, 15,
15, 16, 17, 17, 18, 19, 20, 20, 21, 22, 23, 24,
24, 25, 26, 27, 28, 28, 29, 30, 31, 32, 33, 34,
35, 36, 37, 37, 38, 39, 40, 41, 42, 43, 44, 45,
46, 47, 48, 49, 50, 50, 50, 50, 50, 50, 50, 50,
50, 50
};
/* hangover as a function of peak SNR (3 dB steps) */
const static Word16 hangover_table[20] =
{
30, 30, 30, 30, 30, 30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 8, 8, 8
};
/* burst sensitivity as a function of peak SNR (3 dB steps) */
const static Word16 burstcount_table[20] =
{
8, 8, 8, 8, 8, 8, 8, 8, 7, 6, 5, 4, 4, 4, 4, 4, 4, 4, 4, 4
};
/* voice metric sensitivity as a function of peak SNR (3 dB steps) */
const static Word16 vm_threshold_table[20] =
{
34, 34, 34, 34, 34, 34, 34, 34, 34, 34, 34, 40, 51, 71, 100, 139, 191, 257, 337, 432
};
/* State tables that use 22,9 or 27,4 scaling for ch_enrg[] */
const static Word16 noise_floor_chan[2] = {NOISE_FLOOR_CHAN_0, NOISE_FLOOR_CHAN_1};
const static Word16 min_chan_enrg[2] = {MIN_CHAN_ENRG_0, MIN_CHAN_ENRG_1};
const static Word16 ine_noise[2] = {INE_NOISE_0, INE_NOISE_1};
const static Word16 fbits[2] = {FRACTIONAL_BITS_0, FRACTIONAL_BITS_1};
const static Word16 state_change_shift_r[2] = {STATE_1_TO_0_SHIFT_R, STATE_0_TO_1_SHIFT_R};
/* Energy scale table given 30,1 input scaling (also account for -6 dB shift on input) */
const static Word16 enrg_norm_shift[2] = {(FRACTIONAL_BITS_0-1+2), (FRACTIONAL_BITS_1-1+2)};
/* Automatic variables */
Word32 Lenrg; /* scaled as 30,1 */
Word32 Ltne; /* scaled as 22,9 */
Word32 Ltce; /* scaled as 22,9 or 27,4 */
Word16 tne_db; /* scaled as 7,8 */
Word16 tce_db; /* scaled as 7,8 */
Word16 input_buffer[FRM_LEN]; /* used for block normalising input data */
Word16 data_buffer[FFT_LEN]; /* used for in-place FFT */
Word16 ch_snr[NUM_CHAN]; /* scaled as 7,8 */
Word16 ch_snrq; /* scaled as 15,0 (in 0.375 dB steps) */
Word16 vm_sum; /* scaled as 15,0 */
Word16 ch_enrg_dev; /* scaled as 7,8 */
Word32 Lpeak; /* maximum channel energy */
Word16 p2a_flag; /* flag to indicate spectral peak-to-average ratio > 10 dB */
Word16 ch_enrg_db[NUM_CHAN]; /* scaled as 7,8 */
Word16 ch_noise_db; /* scaled as 7,8 */
Word16 alpha; /* scaled as 0,15 */
Word16 one_m_alpha; /* scaled as 0,15 */
Word16 update_flag; /* set to indicate a background noise estimate update */
Word16 i, j, j1, j2; /* Scratch variables */
Word16 hi1, lo1;
Word32 Ltmp, Ltmp1, Ltmp2;
Word16 tmp;
Word16 normb_shift; /* block norm shift count */
Word16 ivad; /* intermediate VAD decision (return value) */
Word16 tsnrq; /* total signal-to-noise ratio (quantized 3 dB steps) scaled as 15,0 */
Word16 xt; /* instantaneous frame SNR in dB, scaled as 7,8 */
Word16 state_change;
/* Increment frame counter */
st->Lframe_cnt = L_add(st->Lframe_cnt, 1);
/* Block normalize the input */
normb_shift = block_norm(farray_ptr, input_buffer, FRM_LEN, FFT_HEADROOM);
/* Pre-emphasize the input data and store in the data buffer with the appropriate offset */
for (i = 0; i < DELAY; i++)
{
data_buffer[i] = 0; move16();
}
st->pre_emp_mem = shr_r(st->pre_emp_mem, sub(st->last_normb_shift, normb_shift));
st->last_normb_shift = normb_shift; move16();
data_buffer[DELAY] = add(input_buffer[0], mult(PRE_EMP_FAC, st->pre_emp_mem)); move16();
for (i = DELAY + 1, j = 1; i < DELAY + FRM_LEN; i++, j++)
{
data_buffer[i] = add(input_buffer[j], mult(PRE_EMP_FAC, input_buffer[j-1])); move16();
}
st->pre_emp_mem = input_buffer[FRM_LEN-1]; move16();
for (i = DELAY + FRM_LEN; i < FFT_LEN; i++)
{
data_buffer[i] = 0; move16();
}
/* Perform FFT on the data buffer */
r_fft(data_buffer);
/* Use normb_shift factor to determine the scaling of the energy estimates */
state_change = 0; move16();
test();
if (st->shift_state == 0)
{ test();
if (sub(normb_shift, -FFT_HEADROOM+2) <= 0)
{
state_change = 1; move16();
st->shift_state = 1; move16();
}
}
else
{ test();
if (sub(normb_shift, -FFT_HEADROOM+5) >= 0)
{
state_change = 1; move16();
st->shift_state = 0; move16();
}
}
/* Scale channel energy estimate */ test();
if (state_change)
{
for (i = LO_CHAN; i <= HI_CHAN; i++)
{
st->Lch_enrg[i] = L_shr(st->Lch_enrg[i], state_change_shift_r[st->shift_state]); move32();
}
}
/* Estimate the energy in each channel */
test();
if (L_sub(st->Lframe_cnt, 1) == 0)
{
alpha = 32767; move16();
one_m_alpha = 0; move16();
}
else
{
alpha = CEE_SM_FAC; move16();
one_m_alpha = ONE_MINUS_CEE_SM_FAC; move16();
}
for (i = LO_CHAN; i <= HI_CHAN; i++)
{
Lenrg = 0; move16();
j1 = ch_tbl[i][0]; move16();
j2 = ch_tbl[i][1]; move16();
for (j = j1; j <= j2; j++)
{
Lenrg = L_mac(Lenrg, data_buffer[2 * j], data_buffer[2 * j]);
Lenrg = L_mac(Lenrg, data_buffer[2 * j + 1], data_buffer[2 * j + 1]);
}
/* Denorm energy & scale 30,1 according to the state */
Lenrg = L_shr_r(Lenrg, sub(shl(normb_shift, 1), enrg_norm_shift[st->shift_state]));
/* integrate over time: e[i] = (1-alpha)*e[i] + alpha*enrg/num_bins_in_chan */
tmp = mult(alpha, ch_tbl_sh[i]);
L_Extract (Lenrg, &hi1, &lo1);
Ltmp = Mpy_32_16(hi1, lo1, tmp);
L_Extract (st->Lch_enrg[i], &hi1, &lo1);
st->Lch_enrg[i] = L_add(Ltmp, Mpy_32_16(hi1, lo1, one_m_alpha)); move32();
test();
if (L_sub(st->Lch_enrg[i], min_chan_enrg[st->shift_state]) < 0)
{
st->Lch_enrg[i] = min_chan_enrg[st->shift_state]; move32();
}
}
/* Compute the total channel energy estimate (Ltce) */
Ltce = 0; move16();
for (i = LO_CHAN; i <= HI_CHAN; i++)
{
Ltce = L_add(Ltce, st->Lch_enrg[i]);
}
/* Calculate spectral peak-to-average ratio, set flag if p2a > 10 dB */
Lpeak = 0; move32();
for (i = LO_CHAN+2; i <= HI_CHAN; i++) /* Sine waves not valid for low frequencies */
{ test();
if (L_sub(st->Lch_enrg [i], Lpeak) > 0)
{
Lpeak = st->Lch_enrg [i]; move32();
}
}
/* Set p2a_flag if peak (dB) > average channel energy (dB) + 10 dB */
/* Lpeak > Ltce/num_channels * 10^(10/10) */
/* Lpeak > (10/16)*Ltce */
L_Extract (Ltce, &hi1, &lo1);
Ltmp = Mpy_32_16(hi1, lo1, 20480);
test();
if (L_sub(Lpeak, Ltmp) > 0)
{
p2a_flag = TRUE; move16();
}
else
{
p2a_flag = FALSE; move16();
}
/* Initialize channel noise estimate to either the channel energy or fixed level */
/* Scale the energy appropriately to yield state 0 (22,9) scaling for noise */
test();
if (L_sub(st->Lframe_cnt, 4) <= 0)
{ test();
if (p2a_flag == TRUE)
{
for (i = LO_CHAN; i <= HI_CHAN; i++)
{
st->Lch_noise[i] = INE_NOISE_0; move32();
}
}
else
{
for (i = LO_CHAN; i <= HI_CHAN; i++)
{ test();
if (L_sub(st->Lch_enrg[i], ine_noise[st->shift_state]) < 0)
{
st->Lch_noise[i] = INE_NOISE_0; move32();
}
else
{ test();
if (st->shift_state == 1)
{
st->Lch_noise[i] = L_shr(st->Lch_enrg[i], state_change_shift_r[0]);
move32();
}
else
{
st->Lch_noise[i] = st->Lch_enrg[i]; move32();
}
}
}
}
}
/* Compute the channel energy (in dB), the channel SNRs, and the sum of voice metrics */
vm_sum = 0; move16();
for (i = LO_CHAN; i <= HI_CHAN; i++)
{
ch_enrg_db[i] = fn10Log10(st->Lch_enrg[i], fbits[st->shift_state]); move16();
ch_noise_db = fn10Log10(st->Lch_noise[i], FRACTIONAL_BITS_0);
ch_snr[i] = sub(ch_enrg_db[i], ch_noise_db); move16();
/* quantize channel SNR in 3/8 dB steps (scaled 7,8 => 15,0) */
/* ch_snr = round((snr/(3/8))>>8) */
/* = round(((0.6667*snr)<<2)>>8) */
/* = round((0.6667*snr)>>6) */
ch_snrq = shr_r(mult(21845, ch_snr[i]), 6);
/* Accumulate the sum of voice metrics */ test();
if (sub(ch_snrq, 89) < 0)
{ test();
if (ch_snrq > 0)
{
j = ch_snrq; move16();
}
else
{
j = 0; move16();
}
}
else
{
j = 89; move16();
}
vm_sum = add(vm_sum, vm_tbl[j]);
}
/* Initialize NOMINAL peak voice energy and average noise energy, calculate instantaneous SNR */
test(),test(),logic16();
if (L_sub(st->Lframe_cnt, 4) <= 0 || st->fupdate_flag == TRUE)
{
/* tce_db = (96 - 22 - 10*log10(64) (due to FFT)) scaled as 7,8 */
tce_db = 14320; move16();
st->negSNRvar = 0; move16();
st->negSNRbias = 0; move16();
/* Compute the total noise estimate (Ltne) */
Ltne = 0; move32();
for (i = LO_CHAN; i <= HI_CHAN; i++)
{
Ltne = L_add(Ltne, st->Lch_noise[i]);
}
/* Get total noise in dB */
tne_db = fn10Log10(Ltne, FRACTIONAL_BITS_0);
/* Initialise instantaneous and long-term peak signal-to-noise ratios */
xt = sub(tce_db, tne_db);
st->tsnr = xt; move16();
}
else
{
/* Calculate instantaneous frame signal-to-noise ratio */
/* xt = 10*log10( sum(2.^(ch_snr*0.1*log2(10)))/length(ch_snr) ) */
Ltmp1 = 0; move32();
for (i=LO_CHAN; i<=HI_CHAN; i++) {
/* Ltmp2 = ch_snr[i] * 0.1 * log2(10); (ch_snr scaled as 7,8) */
Ltmp2 = L_shr(L_mult(ch_snr[i], 10885), 8);
L_Extract(Ltmp2, &hi1, &lo1);
hi1 = add(hi1, 3); /* 2^3 to compensate for negative SNR */
Ltmp1 = L_add(Ltmp1, Pow2(hi1, lo1));
}
xt = fn10Log10(Ltmp1, 4+3); /* average by 16, inverse compensation 2^3 */
/* Estimate long-term "peak" SNR */ test(),test();
if (sub(xt, st->tsnr) > 0)
{
/* tsnr = 0.9*tsnr + 0.1*xt; */
st->tsnr = round(L_add(L_mult(29491, st->tsnr), L_mult(3277, xt)));
}
/* else if (xt > 0.625*tsnr) */
else if (sub(xt, mult(20480, st->tsnr)) > 0)
{
/* tsnr = 0.998*tsnr + 0.002*xt; */
st->tsnr = round(L_add(L_mult(32702, st->tsnr), L_mult(66, xt)));
}
}
/* Quantize the long-term SNR in 3 dB steps, limit to 0 <= tsnrq <= 19 */
tsnrq = shr(mult(st->tsnr, 10923), 8);
/* tsnrq = min(19, max(0, tsnrq)); */ test(),test();
if (sub(tsnrq, 19) > 0)
{
tsnrq = 19; move16();
}
else if (tsnrq < 0)
{
tsnrq = 0; move16();
}
/* Calculate the negative SNR sensitivity bias */
test();
if (xt < 0)
{
/* negSNRvar = 0.99*negSNRvar + 0.01*xt*xt; */
/* xt scaled as 7,8 => xt*xt scaled as 14,17, shift to 7,8 and round */
tmp = round(L_shl(L_mult(xt, xt), 7));
st->negSNRvar = round(L_add(L_mult(32440, st->negSNRvar), L_mult(328, tmp)));
/* if (negSNRvar > 4.0) negSNRvar = 4.0; */ test();
if (sub(st->negSNRvar, 1024) > 0)
{
st->negSNRvar = 1024; move16();
}
/* negSNRbias = max(12.0*(negSNRvar - 0.65), 0.0); */
tmp = mult_r(shl(sub(st->negSNRvar, 166), 4), 24576); test();
if (tmp < 0)
{
st->negSNRbias = 0; move16();
}
else
{
st->negSNRbias = shr(tmp, 8);
}
}
/* Determine VAD as a function of the voice metric sum and quantized SNR */
tmp = add(vm_threshold_table[tsnrq], st->negSNRbias); test();
if (sub(vm_sum, tmp) > 0)
{
ivad = 1; move16();
st->burstcount = add(st->burstcount, 1); test();
if (sub(st->burstcount, burstcount_table[tsnrq]) > 0)
{
st->hangover = hangover_table[tsnrq]; move16();
}
}
else
{
st->burstcount = 0; move16();
st->hangover = sub(st->hangover, 1); test();
if (st->hangover <= 0)
{
ivad = 0; move16();
st->hangover = 0; move16();
}
else
{
ivad = 1; move16();
}
}
/* Calculate log spectral deviation */
ch_enrg_dev = 0; move16();
test();
if (L_sub(st->Lframe_cnt, 1) == 0)
{
for (i = LO_CHAN; i <= HI_CHAN; i++)
{
st->ch_enrg_long_db[i] = ch_enrg_db[i]; move16();
}
}
else
{
for (i = LO_CHAN; i <= HI_CHAN; i++)
{
tmp = abs_s(sub(st->ch_enrg_long_db[i], ch_enrg_db[i]));
ch_enrg_dev = add(ch_enrg_dev, tmp);
}
}
/*
* Calculate long term integration constant as a function of instantaneous SNR
* (i.e., high SNR (tsnr dB) -> slower integration (alpha = HIGH_ALPHA),
* low SNR (0 dB) -> faster integration (alpha = LOW_ALPHA)
*/
/* alpha = HIGH_ALPHA - ALPHA_RANGE * (tsnr - xt) / tsnr, low <= alpha <= high */
tmp = sub(st->tsnr, xt); test(),logic16(),test(),test();
if (tmp <= 0 || st->tsnr <= 0)
{
alpha = HIGH_ALPHA; move16();
one_m_alpha = 32768L-HIGH_ALPHA; move16();
}
else if (sub(tmp, st->tsnr) > 0)
{
alpha = LOW_ALPHA; move16();
one_m_alpha = 32768L-LOW_ALPHA; move16();
}
else
{
tmp = div_s(tmp, st->tsnr);
alpha = sub(HIGH_ALPHA, mult(ALPHA_RANGE, tmp));
one_m_alpha = sub(32767, alpha);
}
/* Calc long term log spectral energy */
for (i = LO_CHAN; i <= HI_CHAN; i++)
{
Ltmp1 = L_mult(one_m_alpha, ch_enrg_db[i]);
Ltmp2 = L_mult(alpha, st->ch_enrg_long_db[i]);
st->ch_enrg_long_db[i] = round(L_add(Ltmp1, Ltmp2));
}
/* Set or clear the noise update flags */
update_flag = FALSE; move16();
st->fupdate_flag = FALSE; move16();
test(),test();
if (sub(vm_sum, UPDATE_THLD) <= 0)
{ test();
if (st->burstcount == 0)
{
update_flag = TRUE; move16();
st->update_cnt = 0; move16();
}
}
else if (L_sub(Ltce, noise_floor_chan[st->shift_state]) > 0)
{ test();
if (sub(ch_enrg_dev, DEV_THLD) < 0)
{ test();
if (p2a_flag == FALSE)
{ test();
if (st->LTP_flag == FALSE)
{
st->update_cnt = add(st->update_cnt, 1); test();
if (sub(st->update_cnt, UPDATE_CNT_THLD) >= 0)
{
update_flag = TRUE; move16();
st->fupdate_flag = TRUE; move16();
}
}
}
}
}
test();
if (sub(st->update_cnt, st->last_update_cnt) == 0)
{
st->hyster_cnt = add(st->hyster_cnt, 1);
}
else
{
st->hyster_cnt = 0; move16();
}
st->last_update_cnt = st->update_cnt; move16();
test();
if (sub(st->hyster_cnt, HYSTER_CNT_THLD) > 0)
{
st->update_cnt = 0; move16();
}
/* Conditionally update the channel noise estimates */
test();
if (update_flag == TRUE)
{
/* Check shift state */ test();
if (st->shift_state == 1)
{
/* get factor to shift ch_enrg[] from state 1 to 0 (noise always state 0) */
tmp = state_change_shift_r[0]; move16();
}
else
{
/* No shift if already state 0 */
tmp = 0; move16();
}
/* Update noise energy estimate */
for (i = LO_CHAN; i <= HI_CHAN; i++)
{ test();
/* integrate over time: en[i] = (1-alpha)*en[i] + alpha*e[n] */
/* (extract with shift compensation for state 1) */
L_Extract (L_shr(st->Lch_enrg[i], tmp), &hi1, &lo1);
Ltmp = Mpy_32_16(hi1, lo1, CNE_SM_FAC);
L_Extract (st->Lch_noise[i], &hi1, &lo1);
st->Lch_noise[i] = L_add(Ltmp, Mpy_32_16(hi1, lo1, ONE_MINUS_CNE_SM_FAC)); move32();
/* Limit low level noise */ test();
if (L_sub(st->Lch_noise[i], MIN_NOISE_ENRG_0) < 0)
{
st->Lch_noise[i] = MIN_NOISE_ENRG_0; move32();
}
}
}
return(ivad);
} /* end of vad2 () */
void A_Refl(
Word16 a[], /* i : Directform coefficients */
Word16 refl[], /* o : Reflection coefficients */
Flag *pOverflow
)
{
/* local variables */
Word16 i;
Word16 j;
Word16 aState[M];
Word16 bState[M];
Word16 normShift;
Word16 normProd;
Word32 L_acc;
Word16 scale;
Word32 L_temp;
Word16 temp;
Word16 mult;
/* initialize states */
for (i = 0; i < M; i++)
{
aState[i] = a[i];
}
/* backward Levinson recursion */
for (i = M - 1; i >= 0; i--)
{
if (abs_s(aState[i]) >= 4096)
{
for (i = 0; i < M; i++)
{
refl[i] = 0;
}
break;
}
refl[i] = shl(aState[i], 3, pOverflow);
L_temp = L_mult(refl[i], refl[i], pOverflow);
L_acc = L_sub(MAX_32, L_temp, pOverflow);
normShift = norm_l(L_acc);
scale = sub(15, normShift, pOverflow);
L_acc = L_shl(L_acc, normShift, pOverflow);
normProd = pv_round(L_acc, pOverflow);
mult = div_s(16384, normProd);
for (j = 0; j < i; j++)
{
L_acc = L_deposit_h(aState[j]);
L_acc = L_msu(L_acc, refl[i], aState[i-j-1], pOverflow);
temp = pv_round(L_acc, pOverflow);
L_temp = L_mult(mult, temp, pOverflow);
L_temp = L_shr_r(L_temp, scale, pOverflow);
if (L_abs(L_temp) > 32767)
{
for (i = 0; i < M; i++)
{
refl[i] = 0;
}
break;
}
bState[j] = extract_l(L_temp);
}
for (j = 0; j < i; j++)
{
aState[j] = bState[j];
}
}
return;
}
static Word16 D4i40_17_fast(/*(o) : Index of pulses positions. */
Word16 dn[], /* (i) : Correlations between h[] and Xn[]. */
Word16 rr[], /* (i) : Correlations of impulse response h[]. */
Word16 h[], /* (i) Q12: Impulse response of filters. */
Word16 cod[], /* (o) Q13: Selected algebraic codeword. */
Word16 y[], /* (o) Q12: Filtered algebraic codeword. */
Word16 *sign /* (o) : Signs of 4 pulses. */
)
{
Word16 i0, i1, i2, i3, ip0, ip1, ip2, ip3;
Word16 i, j, ix, iy, track, trk, max;
Word16 prev_i0, i1_offset;
Word16 psk, ps, ps0, ps1, ps2, sq, sq2;
Word16 alpk, alp, alp_16;
Word32 s, alp0, alp1, alp2;
Word16 *p0, *p1, *p2, *p3, *p4;
Word16 sign_dn[L_SUBFR], sign_dn_inv[L_SUBFR], *psign;
Word16 tmp_vect[NB_POS];
Word16 *rri0i0, *rri1i1, *rri2i2, *rri3i3, *rri4i4;
Word16 *rri0i1, *rri0i2, *rri0i3, *rri0i4;
Word16 *rri1i2, *rri1i3, *rri1i4;
Word16 *rri2i3, *rri2i4;
Word16 *ptr_rri0i3_i4;
Word16 *ptr_rri1i3_i4;
Word16 *ptr_rri2i3_i4;
Word16 *ptr_rri3i3_i4;
/* Init pointers */
rri0i0 = rr;
rri1i1 = rri0i0 + NB_POS;
rri2i2 = rri1i1 + NB_POS;
rri3i3 = rri2i2 + NB_POS;
rri4i4 = rri3i3 + NB_POS;
rri0i1 = rri4i4 + NB_POS;
rri0i2 = rri0i1 + MSIZE;
rri0i3 = rri0i2 + MSIZE;
rri0i4 = rri0i3 + MSIZE;
rri1i2 = rri0i4 + MSIZE;
rri1i3 = rri1i2 + MSIZE;
rri1i4 = rri1i3 + MSIZE;
rri2i3 = rri1i4 + MSIZE;
rri2i4 = rri2i3 + MSIZE;
/*-----------------------------------------------------------------------*
* Chose the sign of the impulse. *
*-----------------------------------------------------------------------*/
for (i=0; i<L_SUBFR; i++)
{
if (dn[i] >= 0)
{
sign_dn[i] = MAX_16;
sign_dn_inv[i] = MIN_16;
}
else
{
sign_dn[i] = MIN_16;
sign_dn_inv[i] = MAX_16;
dn[i] = negate(dn[i]);
}
}
/*-------------------------------------------------------------------*
* Modification of rrixiy[] to take signs into account. *
*-------------------------------------------------------------------*/
p0 = rri0i1;
p1 = rri0i2;
p2 = rri0i3;
p3 = rri0i4;
for(i0=0; i0<L_SUBFR; i0+=STEP)
{
psign = sign_dn;
if (psign[i0] < 0) psign = sign_dn_inv;
for(i1=1; i1<L_SUBFR; i1+=STEP)
{
*p0++ = mult(*p0, psign[i1]);
*p1++ = mult(*p1, psign[i1+1]);
*p2++ = mult(*p2, psign[i1+2]);
*p3++ = mult(*p3, psign[i1+3]);
}
}
p0 = rri1i2;
p1 = rri1i3;
p2 = rri1i4;
for(i1=1; i1<L_SUBFR; i1+=STEP)
{
psign = sign_dn;
if (psign[i1] < 0) psign = sign_dn_inv;
for(i2=2; i2<L_SUBFR; i2+=STEP)
{
*p0++ = mult(*p0, psign[i2]);
*p1++ = mult(*p1, psign[i2+1]);
*p2++ = mult(*p2, psign[i2+2]);
}
}
p0 = rri2i3;
p1 = rri2i4;
for(i2=2; i2<L_SUBFR; i2+=STEP)
{
psign = sign_dn;
if (psign[i2] < 0) psign = sign_dn_inv;
for(i3=3; i3<L_SUBFR; i3+=STEP)
{
*p0++ = mult(*p0, psign[i3]);
*p1++ = mult(*p1, psign[i3+1]);
}
}
/*-------------------------------------------------------------------*
* Search the optimum positions of the four pulses which maximize *
* square(correlation) / energy *
*-------------------------------------------------------------------*/
psk = -1;
alpk = 1;
ptr_rri0i3_i4 = rri0i3;
ptr_rri1i3_i4 = rri1i3;
ptr_rri2i3_i4 = rri2i3;
ptr_rri3i3_i4 = rri3i3;
/* Initializations only to remove warning from some compilers */
ip0=0; ip1=1; ip2=2; ip3=3; ix=0; iy=0; ps=0;
/* search 2 times: track 3 and 4 */
for (track=3, trk=0; track<5; track++, trk++)
{
/*------------------------------------------------------------------*
* depth first search 3, phase A: track 2 and 3/4. *
*------------------------------------------------------------------*/
sq = -1;
alp = 1;
/* i0 loop: 2 positions in track 2 */
prev_i0 = -1;
for (i=0; i<2; i++)
{
max = -1;
/* search "dn[]" maximum position in track 2 */
for (j=2; j<L_SUBFR; j+=STEP)
{
if ((sub(dn[j], max) > 0) && (sub(prev_i0,j) != 0))
{
max = dn[j];
i0 = j;
}
}
prev_i0 = i0;
j = mult(i0, 6554); /* j = i0/5 */
p0 = rri2i2 + j;
ps1 = dn[i0];
alp1 = L_mult(*p0, _1_4);
/* i1 loop: 8 positions in track 2 */
p0 = ptr_rri2i3_i4 + shl(j, 3);
p1 = ptr_rri3i3_i4;
for (i1=track; i1<L_SUBFR; i1+=STEP)
{
ps2 = add(ps1, dn[i1]); /* index increment = STEP */
/* alp1 = alp0 + rr[i0][i1] + 1/2*rr[i1][i1]; */
alp2 = L_mac(alp1, *p0++, _1_2);
alp2 = L_mac(alp2, *p1++, _1_4);
sq2 = mult(ps2, ps2);
alp_16 = round(alp2);
s = L_msu(L_mult(alp,sq2),sq,alp_16);
if (s > 0)
{
sq = sq2;
ps = ps2;
alp = alp_16;
ix = i0;
iy = i1;
}
}
}
i0 = ix;
i1 = iy;
i1_offset = shl(mult(i1, 6554), 3); /* j = 8*(i1/5) */
/*------------------------------------------------------------------*
* depth first search 3, phase B: track 0 and 1. *
*------------------------------------------------------------------*/
ps0 = ps;
alp0 = L_mult(alp, _1_4);
sq = -1;
alp = 1;
/* build vector for next loop to decrease complexity */
p0 = rri1i2 + mult(i0, 6554);
p1 = ptr_rri1i3_i4 + mult(i1, 6554);
p2 = rri1i1;
p3 = tmp_vect;
for (i3=1; i3<L_SUBFR; i3+=STEP)
{
/* rrv[i3] = rr[i3][i3] + rr[i0][i3] + rr[i1][i3]; */
s = L_mult(*p0, _1_4); p0 += NB_POS;
s = L_mac(s, *p1, _1_4); p1 += NB_POS;
s = L_mac(s, *p2++, _1_8);
*p3++ = round(s);
}
/* i2 loop: 8 positions in track 0 */
p0 = rri0i2 + mult(i0, 6554);
p1 = ptr_rri0i3_i4 + mult(i1, 6554);
p2 = rri0i0;
p3 = rri0i1;
for (i2=0; i2<L_SUBFR; i2+=STEP)
{
ps1 = add(ps0, dn[i2]); /* index increment = STEP */
/* alp1 = alp0 + rr[i0][i2] + rr[i1][i2] + 1/2*rr[i2][i2]; */
alp1 = L_mac(alp0, *p0, _1_8); p0 += NB_POS;
alp1 = L_mac(alp1, *p1, _1_8); p1 += NB_POS;
alp1 = L_mac(alp1, *p2++, _1_16);
/* i3 loop: 8 positions in track 1 */
p4 = tmp_vect;
for (i3=1; i3<L_SUBFR; i3+=STEP)
{
ps2 = add(ps1, dn[i3]); /* index increment = STEP */
/* alp1 = alp0 + rr[i0][i3] + rr[i1][i3] + rr[i2][i3] + 1/2*rr[i3][i3]; */
alp2 = L_mac(alp1, *p3++, _1_8);
alp2 = L_mac(alp2, *p4++, _1_2);
sq2 = mult(ps2, ps2);
alp_16 = round(alp2);
s = L_msu(L_mult(alp,sq2),sq,alp_16);
if (s > 0)
{
sq = sq2;
alp = alp_16;
ix = i2;
iy = i3;
}
}
}
/*----------------------------------------------------------------*
* depth first search 3: compare codevector with the best case. *
*----------------------------------------------------------------*/
s = L_msu(L_mult(alpk,sq),psk,alp);
if (s > 0)
{
psk = sq;
alpk = alp;
ip2 = i0;
ip3 = i1;
ip0 = ix;
ip1 = iy;
}
/*------------------------------------------------------------------*
* depth first search 4, phase A: track 3 and 0. *
*------------------------------------------------------------------*/
sq = -1;
alp = 1;
/* i0 loop: 2 positions in track 3/4 */
prev_i0 = -1;
for (i=0; i<2; i++)
{
max = -1;
/* search "dn[]" maximum position in track 3/4 */
for (j=track; j<L_SUBFR; j+=STEP)
{
if ((sub(dn[j], max) > 0) && (sub(prev_i0,j) != 0))
{
max = dn[j];
i0 = j;
}
}
prev_i0 = i0;
j = mult(i0, 6554); /* j = i0/5 */
p0 = ptr_rri3i3_i4 + j;
ps1 = dn[i0];
alp1 = L_mult(*p0, _1_4);
/* i1 loop: 8 positions in track 0 */
p0 = ptr_rri0i3_i4 + j;
p1 = rri0i0;
for (i1=0; i1<L_SUBFR; i1+=STEP)
{
ps2 = add(ps1, dn[i1]); /* index increment = STEP */
/* alp1 = alp0 + rr[i0][i1] + 1/2*rr[i1][i1]; */
alp2 = L_mac(alp1, *p0, _1_2); p0 += NB_POS;
alp2 = L_mac(alp2, *p1++, _1_4);
sq2 = mult(ps2, ps2);
alp_16 = round(alp2);
s = L_msu(L_mult(alp,sq2),sq,alp_16);
if (s > 0)
{
sq = sq2;
ps = ps2;
alp = alp_16;
ix = i0;
iy = i1;
}
}
}
i0 = ix;
i1 = iy;
i1_offset = shl(mult(i1, 6554), 3); /* j = 8*(i1/5) */
/*------------------------------------------------------------------*
* depth first search 4, phase B: track 1 and 2. *
*------------------------------------------------------------------*/
ps0 = ps;
alp0 = L_mult(alp, _1_4);
sq = -1;
alp = 1;
/* build vector for next loop to decrease complexity */
p0 = ptr_rri2i3_i4 + mult(i0, 6554);
p1 = rri0i2 + i1_offset;
p2 = rri2i2;
p3 = tmp_vect;
for (i3=2; i3<L_SUBFR; i3+=STEP)
{
/* rrv[i3] = rr[i3][i3] + rr[i0][i3] + rr[i1][i3]; */
s = L_mult(*p0, _1_4); p0 += NB_POS;
s = L_mac(s, *p1++, _1_4);
s = L_mac(s, *p2++, _1_8);
*p3++ = round(s);
}
/* i2 loop: 8 positions in track 1 */
p0 = ptr_rri1i3_i4 + mult(i0, 6554);
p1 = rri0i1 + i1_offset;
p2 = rri1i1;
p3 = rri1i2;
for (i2=1; i2<L_SUBFR; i2+=STEP)
{
ps1 = add(ps0, dn[i2]); /* index increment = STEP */
/* alp1 = alp0 + rr[i0][i2] + rr[i1][i2] + 1/2*rr[i2][i2]; */
alp1 = L_mac(alp0, *p0, _1_8); p0 += NB_POS;
alp1 = L_mac(alp1, *p1++, _1_8);
alp1 = L_mac(alp1, *p2++, _1_16);
/* i3 loop: 8 positions in track 2 */
p4 = tmp_vect;
for (i3=2; i3<L_SUBFR; i3+=STEP)
{
ps2 = add(ps1, dn[i3]); /* index increment = STEP */
/* alp1 = alp0 + rr[i0][i3] + rr[i1][i3] + rr[i2][i3] + 1/2*rr[i3][i3]; */
alp2 = L_mac(alp1, *p3++, _1_8);
alp2 = L_mac(alp2, *p4++, _1_2);
sq2 = mult(ps2, ps2);
alp_16 = round(alp2);
s = L_msu(L_mult(alp,sq2),sq,alp_16);
if (s > 0)
{
sq = sq2;
alp = alp_16;
ix = i2;
iy = i3;
}
}
}
/*----------------------------------------------------------------*
* depth first search 1: compare codevector with the best case. *
*----------------------------------------------------------------*/
s = L_msu(L_mult(alpk,sq),psk,alp);
if (s > 0)
{
psk = sq;
alpk = alp;
ip3 = i0;
ip0 = i1;
ip1 = ix;
ip2 = iy;
}
ptr_rri0i3_i4 = rri0i4;
ptr_rri1i3_i4 = rri1i4;
ptr_rri2i3_i4 = rri2i4;
ptr_rri3i3_i4 = rri4i4;
}
/* Set the sign of impulses */
i0 = sign_dn[ip0];
i1 = sign_dn[ip1];
i2 = sign_dn[ip2];
i3 = sign_dn[ip3];
/* Find the codeword corresponding to the selected positions */
for(i=0; i<L_SUBFR; i++) {
cod[i] = 0;
}
cod[ip0] = shr(i0, 2); /* From Q15 to Q13 */
cod[ip1] = shr(i1, 2);
cod[ip2] = shr(i2, 2);
cod[ip3] = shr(i3, 2);
/* find the filtered codeword */
for (i = 0; i < ip0; i++) y[i] = 0;
if(i0 > 0)
for(i=ip0, j=0; i<L_SUBFR; i++, j++) y[i] = h[j];
else
for(i=ip0, j=0; i<L_SUBFR; i++, j++) y[i] = negate(h[j]);
if(i1 > 0)
for(i=ip1, j=0; i<L_SUBFR; i++, j++) y[i] = add(y[i], h[j]);
else
for(i=ip1, j=0; i<L_SUBFR; i++, j++) y[i] = sub(y[i], h[j]);
if(i2 > 0)
for(i=ip2, j=0; i<L_SUBFR; i++, j++) y[i] = add(y[i], h[j]);
else
for(i=ip2, j=0; i<L_SUBFR; i++, j++) y[i] = sub(y[i], h[j]);
if(i3 > 0)
for(i=ip3, j=0; i<L_SUBFR; i++, j++) y[i] = add(y[i], h[j]);
else
for(i=ip3, j=0; i<L_SUBFR; i++, j++) y[i] = sub(y[i], h[j]);
/* find codebook index; 17-bit address */
i = 0;
if(i0 > 0) i = add(i, 1);
if(i1 > 0) i = add(i, 2);
if(i2 > 0) i = add(i, 4);
if(i3 > 0) i = add(i, 8);
*sign = i;
ip0 = mult(ip0, 6554); /* ip0/5 */
ip1 = mult(ip1, 6554); /* ip1/5 */
ip2 = mult(ip2, 6554); /* ip2/5 */
i = mult(ip3, 6554); /* ip3/5 */
j = add(i, shl(i, 2)); /* j = i*5 */
j = sub(ip3, add(j, 3)); /* j= ip3%5 -3 */
ip3 = add(shl(i, 1), j);
i = add(ip0, shl(ip1, 3));
i = add(i , shl(ip2, 6));
i = add(i , shl(ip3, 9));
return i;
}
static Word16 D2i40_11( /* (o) : Index of pulses positions. */
Word16 Dn[], /* (i) : Correlations between h[] and Xn[]. */
Word16 rr[], /* (i) : Correlations of impulse response h[]. */
Word16 h[], /* (i) : Impulse response of filters. */
Word16 code[], /* (o) : Selected algebraic codeword. */
Word16 y[], /* (o) : Filtered algebraic codeword. */
Word16 *sign, /* (o) : Signs of 4 pulses. */
Word16 i_subfr /* (i) : subframe flag */
)
{
Word16 i0, i1, ip0, ip1, p0, p1;
Word16 i, j, index, tmp, swap;
Word16 ps0, ps1, alp, alp0;
Word32 alp1;
Word16 ps1c, psc, alpha;
Word32 L_temp;
Word16 posIndex[2], signIndex[2];
Word16 m0_bestPos, m1_bestPos;
Word16 p_sign[L_SUBFR];
Word16 *rri0i0, *rri1i1, *rri2i2, *rri3i3, *rri4i4;
Word16 *rri0i1, *RRi1i1, *rri0i3, *RRi3i4;
Word16 *rri1i2, *rri1i3, *rri1i4;
Word16 *rri2i3;
Word16 *ptr_ri0i0, *ptr_ri1i1;
Word16 *ptr_ri0i1, *ptr_Ri0i2, *ptr_ri0i3, *ptr_Ri3i4;
Word16 *ptr_ri1i2, *ptr_ri1i3, *ptr_ri1i4;
Word16 *ptr_ri2i3;
Word16 *outPtr_ri1i1; /* Outside loop pointer */
/* Init pointers */
rri0i0 = rr;
rri1i1 = rri0i0 + NB_POS;
rri2i2 = rri1i1 + NB_POS;
rri3i3 = rri2i2 + NB_POS;
rri4i4 = rri3i3 + NB_POS;
rri0i1 = rri4i4 + NB_POS;
RRi1i1 = rri0i1 + MSIZE; /* Special for 6.4 kbps */
rri0i3 = RRi1i1 + MSIZE;
RRi3i4 = rri0i3 + MSIZE; /* Special for 6.4 kbps */
rri1i2 = RRi3i4 + MSIZE;
rri1i3 = rri1i2 + MSIZE;
rri1i4 = rri1i3 + MSIZE;
rri2i3 = rri1i4 + MSIZE;
/*-----------------------------------------------------------------------*
* Chose the sign of the impulse. *
*-----------------------------------------------------------------------*/
for (i=0; i<L_SUBFR; i++)
{
if( Dn[i] >= 0)
{
p_sign[i] = 0x7fff;
}
else
{
p_sign[i] = (Word16)0x8000;
Dn[i] = negate(Dn[i]);
}
}
/*-------------------------------------------------------------------*
* Modification of rrixiy[] to take signs into account. *
*-------------------------------------------------------------------*/
ptr_ri0i1 = rri0i1;
ptr_ri0i3 = rri0i3;
for(i0=0; i0<L_SUBFR; i0+=STEP) {
for(i1=1; i1<L_SUBFR; i1+=STEP) {
*ptr_ri0i1 = mult(*ptr_ri0i1, mult(p_sign[i0], p_sign[i1]));
ptr_ri0i1++;
*ptr_ri0i3 = mult(*ptr_ri0i3, mult(p_sign[i0], p_sign[i1+2]));
ptr_ri0i3++;
}
}
ptr_ri1i2 = rri1i2;
ptr_ri1i3 = rri1i3;
ptr_ri1i4 = rri1i4;
for(i0=1; i0<L_SUBFR; i0+=STEP) {
for(i1=2; i1<L_SUBFR; i1+=STEP) {
*ptr_ri1i2 = mult(*ptr_ri1i2, mult(p_sign[i0], p_sign[i1]));
ptr_ri1i2++;
*ptr_ri1i3 = mult(*ptr_ri1i3, mult(p_sign[i0], p_sign[i1+1]));
ptr_ri1i3++;
*ptr_ri1i4 = mult(*ptr_ri1i4, mult(p_sign[i0], p_sign[i1+2]));
ptr_ri1i4++;
}
}
ptr_ri2i3 = rri2i3;
ptr_Ri3i4 = RRi3i4;
for(i0=2; i0<L_SUBFR; i0+=STEP) {
for(i1=3; i1<L_SUBFR; i1+=STEP) {
*ptr_ri2i3 = mult(*ptr_ri2i3, mult(p_sign[i0], p_sign[i1]));
ptr_ri2i3++;
*ptr_Ri3i4 = mult(*ptr_Ri3i4, mult(p_sign[i0+1], p_sign[i1+1]));
ptr_Ri3i4++;
}
}
ptr_Ri0i2 = RRi1i1;
for(i0=1; i0<L_SUBFR; i0+=STEP) {
for(i1=1; i1<L_SUBFR; i1+=STEP) {
*ptr_Ri0i2 = mult(*ptr_Ri0i2, mult(p_sign[i0], p_sign[i1]));
ptr_Ri0i2++;
}
}
/*-------------------------------------------------------------------*
* The actual search. *
*-------------------------------------------------------------------*/
ip0 = 1; /* Set to any valid pulse position */
ip1 = 0; /* Set to any valid pulse position */
psc = 0;
alpha = MAX_16;
ptr_ri0i1 = rri0i1;
outPtr_ri1i1 = rri1i1; /* Initial values for tripple loop below */
p0=0; /* Search i0,sub0 vs. i1,sub0 */
p1=1;
ptr_ri0i0 = rri0i0;
for (i = 0; i<9; i++) {
if (i == 4) i++; /* To get right exchange sequence */
swap = i & 1;
if (i == 1) p0=1; /* Search i0,sub1 vs. i1,sub0 */
else if (i == 2) { /* Search i0,sub0 vs. i1,sub1 */
outPtr_ri1i1 = rri3i3;
p0=0;
p1=3;
ptr_ri0i0 = rri0i0;
}
else if (i == 3) { /* Search i0,sub3 vs. i1,sub1 */
outPtr_ri1i1 = rri4i4;
p0=3;
p1=4;
ptr_ri0i0 = rri3i3;
}
else if (i == 5) { /* Search i0,sub2 vs. i1,sub0 */
outPtr_ri1i1 = rri2i2;
p0=1;
p1=2;
ptr_ri0i0 = rri1i1;
}
else if (i == 6) { /* Search i0,sub1 vs. i1,sub1 */
outPtr_ri1i1 = rri3i3;
p1=3;
ptr_ri0i0 = rri1i1;
}
else if (i == 7) { /* Search i0,sub3 vs. i1,sub0 */
outPtr_ri1i1 = rri4i4;
p1=4;
ptr_ri0i0 = rri1i1;
}
else if (i == 8) { /* Search i0,sub2 vs. i1,sub1 */
outPtr_ri1i1 = rri3i3;
p0=2;
p1=3;
}
for (i0 = p0; i0<40; i0+=STEP) {
ptr_ri1i1 = outPtr_ri1i1;
ps0 = Dn[i0];
alp0 = *ptr_ri0i0++;
for (i1 = p1; i1<40; i1+=STEP) {
ps1 = add(ps0, Dn[i1]);
alp1 = L_mult(alp0, 1);
alp1 = L_mac(alp1, *ptr_ri1i1++, 1);
alp1 = L_mac(alp1, *ptr_ri0i1++, 2);
alp = extract_l(L_shr(alp1, 5));
ps1c = mult(ps1, ps1);
L_temp = L_mult(ps1c, alpha);
L_temp = L_msu(L_temp, psc, alp);
if (L_temp > 0L) {
psc = ps1c;
alpha = alp;
ip0 = i1;
ip1 = i0;
if ( swap ) {
ip0 = i0;
ip1 = i1;
}
}
}
}
}
/* convert from position to table entry index */
for (i0=0; i0<16; i0++)
if (ip0 == trackTable0[i0]) break;
ip0=i0;
for (i1=0; i1<32; i1++)
if (ip1 == trackTable1[i1]) break;
ip1=i1;
m0_bestPos = trackTable0[ip0];
m1_bestPos = trackTable1[ip1];
posIndex[0] = grayEncode[ip0];
posIndex[1] = grayEncode[ip1];
if (p_sign[m0_bestPos] > 0)
signIndex[0] = 1;
else
signIndex[0] = 0;
if (p_sign[m1_bestPos] > 0)
signIndex[1] = 1;
else
signIndex[1] = 0;
/* build innovation vector */
for (i = 0; i < L_SUBFR; i++) code[i] = 0;
code[m0_bestPos] = shr(p_sign[m0_bestPos], 2);
code[m1_bestPos] = add(code[m1_bestPos], shr(p_sign[m1_bestPos], 2));
*sign = add(signIndex[1], signIndex[1]);
*sign = add(*sign, signIndex[0]);
tmp = shl(posIndex[1], 4);
index = add(posIndex[0], tmp);
/* compute filtered cbInnovation */
for (i = 0; i < L_SUBFR; i++) y[i] = 0;
if(signIndex[0] == 0)
for(i=m0_bestPos, j=0; i<L_SUBFR; i++, j++) y[i] = negate(h[j]);
else
for(i=m0_bestPos, j=0; i<L_SUBFR; i++, j++) y[i] = h[j];
if(signIndex[1] == 0)
for(i=m1_bestPos, j=0; i<L_SUBFR; i++, j++) y[i] = sub(y[i], h[j]);
else
for(i=m1_bestPos, j=0; i<L_SUBFR; i++, j++) y[i] = add(y[i], h[j]);
return index;
}
static void Cor_h(
Word16 *H, /* (i) Q12 :Impulse response of filters */
Word16 *rr /* (o) :Correlations of H[] */
)
{
Word16 *rri0i0, *rri1i1, *rri2i2, *rri3i3, *rri4i4;
Word16 *rri0i1, *rri0i2, *rri0i3, *rri0i4;
Word16 *rri1i2, *rri1i3, *rri1i4;
Word16 *rri2i3, *rri2i4;
Word16 *p0, *p1, *p2, *p3, *p4;
Word16 *ptr_hd, *ptr_hf, *ptr_h1, *ptr_h2;
Word32 cor;
Word16 i, k, ldec, l_fin_sup, l_fin_inf;
Word16 h[L_SUBFR];
/* Scaling h[] for maximum precision */
cor = 0;
for(i=0; i<L_SUBFR; i++)
cor = L_mac(cor, H[i], H[i]);
if(sub(extract_h(cor),32000) > 0)
{
for(i=0; i<L_SUBFR; i++) {
h[i] = shr(H[i], 1);
}
}
else
{
k = norm_l(cor);
k = shr(k, 1);
for(i=0; i<L_SUBFR; i++) {
h[i] = shl(H[i], k);
}
}
/*------------------------------------------------------------*
* Compute rri0i0[], rri1i1[], rri2i2[], rri3i3 and rri4i4[] *
*------------------------------------------------------------*/
/* Init pointers */
rri0i0 = rr;
rri1i1 = rri0i0 + NB_POS;
rri2i2 = rri1i1 + NB_POS;
rri3i3 = rri2i2 + NB_POS;
rri4i4 = rri3i3 + NB_POS;
rri0i1 = rri4i4 + NB_POS;
rri0i2 = rri0i1 + MSIZE;
rri0i3 = rri0i2 + MSIZE;
rri0i4 = rri0i3 + MSIZE;
rri1i2 = rri0i4 + MSIZE;
rri1i3 = rri1i2 + MSIZE;
rri1i4 = rri1i3 + MSIZE;
rri2i3 = rri1i4 + MSIZE;
rri2i4 = rri2i3 + MSIZE;
p0 = rri0i0 + NB_POS-1; /* Init pointers to last position of rrixix[] */
p1 = rri1i1 + NB_POS-1;
p2 = rri2i2 + NB_POS-1;
p3 = rri3i3 + NB_POS-1;
p4 = rri4i4 + NB_POS-1;
ptr_h1 = h;
cor = 0;
for(i=0; i<NB_POS; i++)
{
cor = L_mac(cor, *ptr_h1, *ptr_h1); ptr_h1++;
*p4-- = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h1); ptr_h1++;
*p3-- = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h1); ptr_h1++;
*p2-- = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h1); ptr_h1++;
*p1-- = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h1); ptr_h1++;
*p0-- = extract_h(cor);
}
/*-----------------------------------------------------------------*
* Compute elements of: rri2i3[], rri1i2[], rri0i1[] and rri0i4[] *
*-----------------------------------------------------------------*/
l_fin_sup = MSIZE-1;
l_fin_inf = l_fin_sup-(Word16)1;
ldec = NB_POS+1;
ptr_hd = h;
ptr_hf = ptr_hd + 1;
for(k=0; k<NB_POS; k++) {
p3 = rri2i3 + l_fin_sup;
p2 = rri1i2 + l_fin_sup;
p1 = rri0i1 + l_fin_sup;
p0 = rri0i4 + l_fin_inf;
cor = 0;
ptr_h1 = ptr_hd;
ptr_h2 = ptr_hf;
for(i=k+(Word16)1; i<NB_POS; i++ ) {
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p2 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p1 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p0 = extract_h(cor);
p3 -= ldec;
p2 -= ldec;
p1 -= ldec;
p0 -= ldec;
}
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p2 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p1 = extract_h(cor);
l_fin_sup -= NB_POS;
l_fin_inf--;
ptr_hf += STEP;
}
/*---------------------------------------------------------------------*
* Compute elements of: rri2i4[], rri1i3[], rri0i2[], rri1i4[], rri0i3 *
*---------------------------------------------------------------------*/
ptr_hd = h;
ptr_hf = ptr_hd + 2;
l_fin_sup = MSIZE-1;
l_fin_inf = l_fin_sup-(Word16)1;
for(k=0; k<NB_POS; k++) {
p4 = rri2i4 + l_fin_sup;
p3 = rri1i3 + l_fin_sup;
p2 = rri0i2 + l_fin_sup;
p1 = rri1i4 + l_fin_inf;
p0 = rri0i3 + l_fin_inf;
cor = 0;
ptr_h1 = ptr_hd;
ptr_h2 = ptr_hf;
for(i=k+(Word16)1; i<NB_POS; i++ ) {
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p4 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p2 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p1 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p0 = extract_h(cor);
p4 -= ldec;
p3 -= ldec;
p2 -= ldec;
p1 -= ldec;
p0 -= ldec;
}
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p4 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p2 = extract_h(cor);
l_fin_sup -= NB_POS;
l_fin_inf--;
ptr_hf += STEP;
}
/*----------------------------------------------------------------------*
* Compute elements of: rri1i4[], rri0i3[], rri2i4[], rri1i3[], rri0i2 *
*----------------------------------------------------------------------*/
ptr_hd = h;
ptr_hf = ptr_hd + 3;
l_fin_sup = MSIZE-1;
l_fin_inf = l_fin_sup-(Word16)1;
for(k=0; k<NB_POS; k++) {
p4 = rri1i4 + l_fin_sup;
p3 = rri0i3 + l_fin_sup;
p2 = rri2i4 + l_fin_inf;
p1 = rri1i3 + l_fin_inf;
p0 = rri0i2 + l_fin_inf;
ptr_h1 = ptr_hd;
ptr_h2 = ptr_hf;
cor = 0;
for(i=k+(Word16)1; i<NB_POS; i++ ) {
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p4 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p2 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p1 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p0 = extract_h(cor);
p4 -= ldec;
p3 -= ldec;
p2 -= ldec;
p1 -= ldec;
p0 -= ldec;
}
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p4 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p3 = extract_h(cor);
l_fin_sup -= NB_POS;
l_fin_inf--;
ptr_hf += STEP;
}
/*----------------------------------------------------------------------*
* Compute elements of: rri0i4[], rri2i3[], rri1i2[], rri0i1[] *
*----------------------------------------------------------------------*/
ptr_hd = h;
ptr_hf = ptr_hd + 4;
l_fin_sup = MSIZE-1;
l_fin_inf = l_fin_sup-(Word16)1;
for(k=0; k<NB_POS; k++) {
p3 = rri0i4 + l_fin_sup;
p2 = rri2i3 + l_fin_inf;
p1 = rri1i2 + l_fin_inf;
p0 = rri0i1 + l_fin_inf;
ptr_h1 = ptr_hd;
ptr_h2 = ptr_hf;
cor = 0;
for(i=k+(Word16)1; i<NB_POS; i++ ) {
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p2 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p1 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p0 = extract_h(cor);
p3 -= ldec;
p2 -= ldec;
p1 -= ldec;
p0 -= ldec;
}
cor = L_mac(cor, *ptr_h1, *ptr_h2); ptr_h1++; ptr_h2++;
*p3 = extract_h(cor);
l_fin_sup -= NB_POS;
l_fin_inf--;
ptr_hf += STEP;
}
return;
}
static void Cor_h_D(
Word16 *H, /* (i) Q12 :Impulse response of filters */
Word16 *rr /* (o) :Correlations of H[] */
)
{
Word16 *rri0i0, *rri1i1, *rri2i2, *rri3i3, *rri4i4;
Word16 *rri0i1, *rri0i2, *rri0i3, *rri0i4;
Word16 *rri1i2, *rri1i3, *rri1i4;
Word16 *rri2i3, *rri2i4;
Word16 *p0, *p1, *p2, *p3, *p4;
Word16 *ptr_hd, *ptr_hf, *ptr_h1, *ptr_h2;
Word32 cor;
Word16 i, k, ldec, l_fin_sup, l_fin_inf;
Word16 h[L_SUBFR];
Word32 L_tmp;
Word16 lsym;
/* Scaling h[] for maximum precision */
cor = 0;
for(i=0; i<L_SUBFR; i++)
cor = L_mac(cor, H[i], H[i]);
L_tmp = L_sub(extract_h(cor),32000);
if(L_tmp>0L )
{
for(i=0; i<L_SUBFR; i++) {
h[i] = shr(H[i], 1);
}
}
else
{
k = norm_l(cor);
k = shr(k, 1);
for(i=0; i<L_SUBFR; i++) {
h[i] = shl(H[i], k);
}
}
/*-----------------------------------------------------------------*
* In case of G729 mode, nine cross correlations has to be *
* calculated, namely the following: *
* *
* rri0i1[], *
* rri0i2[], rri1i2[], *
* rri0i3[], rri1i3[], rri2i3[], *
* rri0i4[], rri1i4[], rri2i4[], *
* *
* In case of G729 on 6.4 kbps mode, three of the above nine cross *
* correlations are not needed for the codebook search, namely *
* rri0i2[], rri0i4[] and rri2i4[]. Two of these three 64-element *
* positions are instead used by two cross correlations needed *
* only by the 6.4 kbps mode (see D2i40_11() for details). *
*-----------------------------------------------------------------*/
/* Init pointers */
rri0i0 = rr;
rri1i1 = rri0i0 + NB_POS;
rri2i2 = rri1i1 + NB_POS;
rri3i3 = rri2i2 + NB_POS;
rri4i4 = rri3i3 + NB_POS;
rri0i1 = rri4i4 + NB_POS;
rri0i2 = rri0i1 + MSIZE; /* Holds RRi1i1[] in 6.4 kbps mode */
rri0i3 = rri0i2 + MSIZE;
rri0i4 = rri0i3 + MSIZE; /* Holds RRi3i4[] in 6.4 kbps mode */
rri1i2 = rri0i4 + MSIZE;
rri1i3 = rri1i2 + MSIZE;
rri1i4 = rri1i3 + MSIZE;
rri2i3 = rri1i4 + MSIZE;
rri2i4 = rri2i3 + MSIZE; /* Not used in 6.4 kbps mode */
/*------------------------------------------------------------*
* Compute rri0i0[], rri1i1[], rri2i2[], rri3i3 and rri4i4[] *
*------------------------------------------------------------*/
p0 = rri0i0 + NB_POS-1; /* Init pointers to last position of rrixix[] */
p1 = rri1i1 + NB_POS-1;
p2 = rri2i2 + NB_POS-1;
p3 = rri3i3 + NB_POS-1;
p4 = rri4i4 + NB_POS-1;
ptr_h1 = h;
cor = 0;
for(i=0; i<NB_POS; i++)
{
cor = L_mac(cor, *ptr_h1, *ptr_h1);
ptr_h1++;
*p4-- = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h1);
ptr_h1++;
*p3-- = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h1);
ptr_h1++;
*p2-- = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h1);
ptr_h1++;
*p1-- = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h1);
ptr_h1++;
*p0-- = extract_h(cor);
}
/*-----------------------------------------------------------------*
* Compute elements of: rri2i3[], rri1i2[], rri0i1[] and rri0i4[] *
*-----------------------------------------------------------------*/
l_fin_sup = MSIZE-1;
l_fin_inf = l_fin_sup-(Word16)1;
ldec = NB_POS+1;
ptr_hd = h;
ptr_hf = ptr_hd + 1;
for(k=0; k<NB_POS; k++) {
p4 = rri0i4 + l_fin_sup;
p3 = rri2i3 + l_fin_sup;
p2 = rri1i2 + l_fin_sup;
p1 = rri0i1 + l_fin_sup;
p0 = rri0i4 + l_fin_inf;
cor = 0;
ptr_h1 = ptr_hd;
ptr_h2 = ptr_hf;
for(i=k+(Word16)1; i<NB_POS; i++ ) {
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
if (sub(CODEC_MODE, 1) == 0) *p4 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p2 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p1 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
if (sub(CODEC_MODE, 2) == 0) *p0 = extract_h(cor);
p4 -= ldec;
p3 -= ldec;
p2 -= ldec;
p1 -= ldec;
p0 -= ldec;
}
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
if (sub(CODEC_MODE, 1) == 0) *p4 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p2 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p1 = extract_h(cor);
l_fin_sup -= NB_POS;
l_fin_inf--;
ptr_hf += STEP;
}
/*---------------------------------------------------------------------*
* Compute elements of: rri2i4[], rri1i3[], rri0i2[], rri1i4[], rri0i3 *
*---------------------------------------------------------------------*/
ptr_hd = h;
ptr_hf = ptr_hd + 2;
l_fin_sup = MSIZE-1;
l_fin_inf = l_fin_sup-(Word16)1;
for(k=0; k<NB_POS; k++) {
p4 = rri2i4 + l_fin_sup;
p3 = rri1i3 + l_fin_sup;
p2 = rri0i2 + l_fin_sup;
p1 = rri1i4 + l_fin_inf;
p0 = rri0i3 + l_fin_inf;
cor = 0;
ptr_h1 = ptr_hd;
ptr_h2 = ptr_hf;
for(i=k+(Word16)1; i<NB_POS; i++ ) {
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p4 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p2 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p1 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p0 = extract_h(cor);
p4 -= ldec;
p3 -= ldec;
p2 -= ldec;
p1 -= ldec;
p0 -= ldec;
}
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p4 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p2 = extract_h(cor);
l_fin_sup -= NB_POS;
l_fin_inf--;
ptr_hf += STEP;
}
/*----------------------------------------------------------------------*
* Compute elements of: rri1i4[], rri0i3[], rri2i4[], rri1i3[], rri0i2 *
*----------------------------------------------------------------------*/
ptr_hd = h;
ptr_hf = ptr_hd + 3;
l_fin_sup = MSIZE-1;
l_fin_inf = l_fin_sup-(Word16)1;
for(k=0; k<NB_POS; k++) {
p4 = rri1i4 + l_fin_sup;
p3 = rri0i3 + l_fin_sup;
p2 = rri2i4 + l_fin_inf;
p1 = rri1i3 + l_fin_inf;
p0 = rri0i2 + l_fin_inf;
ptr_h1 = ptr_hd;
ptr_h2 = ptr_hf;
cor = 0;
for(i=k+(Word16)1; i<NB_POS; i++ ) {
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p4 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p2 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p1 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p0 = extract_h(cor);
p4 -= ldec;
p3 -= ldec;
p2 -= ldec;
p1 -= ldec;
p0 -= ldec;
}
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p4 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p3 = extract_h(cor);
l_fin_sup -= NB_POS;
l_fin_inf--;
ptr_hf += STEP;
}
/*----------------------------------------------------------------------*
* Compute elements of: rri0i4[], rri2i3[], rri1i2[], rri0i1[] *
*----------------------------------------------------------------------*/
ptr_hd = h;
ptr_hf = ptr_hd + 4;
l_fin_sup = MSIZE-1;
l_fin_inf = l_fin_sup-(Word16)1;
for(k=0; k<NB_POS; k++) {
if (sub(CODEC_MODE, 2) == 0)
p3 = rri0i4 + l_fin_sup;
if (sub(CODEC_MODE, 1) == 0)
p3 = rri0i4 + l_fin_inf;
p2 = rri2i3 + l_fin_inf;
p1 = rri1i2 + l_fin_inf;
p0 = rri0i1 + l_fin_inf;
ptr_h1 = ptr_hd;
ptr_h2 = ptr_hf;
cor = 0;
for(i=k+(Word16)1; i<NB_POS; i++ ) {
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
if (sub(CODEC_MODE, 2) == 0) *p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
if (sub(CODEC_MODE, 1) == 0) *p3 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p2 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p1 = extract_h(cor);
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p0 = extract_h(cor);
p3 -= ldec;
p2 -= ldec;
p1 -= ldec;
p0 -= ldec;
}
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
if (sub(CODEC_MODE, 2) == 0) *p3 = extract_h(cor);
l_fin_sup -= NB_POS;
l_fin_inf--;
ptr_hf += STEP;
}
if (sub(CODEC_MODE, 1) == 0) {
/*-----------------------------------------------------------------*
* Compute elements of RRi1i1[] *
*-----------------------------------------------------------------*/
p0 = rri0i2;
for (k=0; k<NB_POS; k++) {
*p0 = *rri1i1;
rri1i1++;
p0 += ldec;
}
ptr_hd = h;
ptr_hf = ptr_hd + 5;
l_fin_sup = MSIZE-1;
l_fin_inf = l_fin_sup-NB_POS;
lsym = NB_POS - (Word16)1;
for(k=(Word16)1; k<NB_POS; k++) {
p0 = rri0i2 + l_fin_inf;
ptr_h1 = ptr_hd;
ptr_h2 = ptr_hf;
cor = 0;
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p0 = extract_h(cor);
*(p0+lsym) = extract_h(cor);
p0 -= ldec;
for(i=k+(Word16)1; i<NB_POS; i++ ) {
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
cor = L_mac(cor, *ptr_h1, *ptr_h2);
ptr_h1++;
ptr_h2++;
*p0 = extract_h(cor);
*(p0+lsym) = extract_h(cor);
p0 -= ldec;
}
l_fin_inf -= NB_POS;
ptr_hf += STEP;
lsym += NB_POS - (Word16)1;
}
}
return;
}
/*
********************************************************************************
* PUBLIC PROGRAM CODE
********************************************************************************
*/
Word16 hp_max (
Word32 corr[], /* i : correlation vector. */
Word16 scal_sig[], /* i : scaled signal. */
Word16 L_frame, /* i : length of frame to compute pitch */
Word16 lag_max, /* i : maximum lag */
Word16 lag_min, /* i : minimum lag */
Word16 *cor_hp_max) /* o : max high-pass filtered norm. correlation */
{
Word16 i;
Word16 *p, *p1;
Word32 max, t0, t1;
Word16 max16, t016, cor_max;
Word16 shift, shift1, shift2;
max = MIN_32; move32 ();
t0 = 0L; move32 ();
for (i = lag_max-1; i > lag_min; i--)
{
/* high-pass filtering */
t0 = L_sub (L_sub(L_shl(corr[-i], 1), corr[-i-1]), corr[-i+1]);
t0 = L_abs (t0);
test ();
if (L_sub (t0, max) >= 0)
{
max = t0; move32 ();
}
}
/* compute energy */
p = scal_sig; move16 ();
p1 = &scal_sig[0]; move16 ();
t0 = 0L; move32 ();
for (i = 0; i < L_frame; i++, p++, p1++)
{
t0 = L_mac (t0, *p, *p1);
}
p = scal_sig; move16 ();
p1 = &scal_sig[-1]; move16 ();
t1 = 0L; move32 ();
for (i = 0; i < L_frame; i++, p++, p1++)
{
t1 = L_mac (t1, *p, *p1);
}
/* high-pass filtering */
t0 = L_sub(L_shl(t0, 1), L_shl(t1, 1));
t0 = L_abs (t0);
/* max/t0 */
shift1 = sub(norm_l(max), 1);
max16 = extract_h(L_shl(max, shift1));
shift2 = norm_l(t0);
t016 = extract_h(L_shl(t0, shift2));
test ();
if (t016 != 0)
{
cor_max = div_s(max16, t016);
}
else
{
cor_max = 0; move16 ();
}
shift = sub(shift1, shift2);
test ();
if (shift >= 0)
{
*cor_hp_max = shr(cor_max, shift); move16 (); /* Q15 */
}
else
{
*cor_hp_max = shl(cor_max, negate(shift)); move16 (); /* Q15 */
}
return 0;
}
static Word16 D4i40_17( /* (o) : Index of pulses positions. */
Word16 Dn[], /* (i) : Correlations between h[] and Xn[]. */
Word16 rr[], /* (i) : Correlations of impulse response h[]. */
Word16 h[], /* (i) Q12: Impulse response of filters. */
Word16 cod[], /* (o) Q13: Selected algebraic codeword. */
Word16 y[], /* (o) Q12: Filtered algebraic codeword. */
Word16 *sign, /* (o) : Signs of 4 pulses. */
Word16 i_subfr /* (i) : subframe flag */
)
{
Word16 i0, i1, i2, i3, ip0, ip1, ip2, ip3;
Word16 i, j, time;
Word16 ps0, ps1, ps2, ps3, alp, alp0;
Word32 alp1, alp2, alp3, L32;
Word16 ps3c, psc, alpha;
Word16 average, max0, max1, max2, thres;
Word32 L_temp;
Word16 *rri0i0, *rri1i1, *rri2i2, *rri3i3, *rri4i4;
Word16 *rri0i1, *rri0i2, *rri0i3, *rri0i4;
Word16 *rri1i2, *rri1i3, *rri1i4;
Word16 *rri2i3, *rri2i4;
Word16 *ptr_ri0i0, *ptr_ri1i1, *ptr_ri2i2, *ptr_ri3i3, *ptr_ri4i4;
Word16 *ptr_ri0i1, *ptr_ri0i2, *ptr_ri0i3, *ptr_ri0i4;
Word16 *ptr_ri1i2, *ptr_ri1i3, *ptr_ri1i4;
Word16 *ptr_ri2i3, *ptr_ri2i4;
Word16 p_sign[L_SUBFR];
/* Init pointers */
rri0i0 = rr;
rri1i1 = rri0i0 + NB_POS;
rri2i2 = rri1i1 + NB_POS;
rri3i3 = rri2i2 + NB_POS;
rri4i4 = rri3i3 + NB_POS;
rri0i1 = rri4i4 + NB_POS;
rri0i2 = rri0i1 + MSIZE;
rri0i3 = rri0i2 + MSIZE;
rri0i4 = rri0i3 + MSIZE;
rri1i2 = rri0i4 + MSIZE;
rri1i3 = rri1i2 + MSIZE;
rri1i4 = rri1i3 + MSIZE;
rri2i3 = rri1i4 + MSIZE;
rri2i4 = rri2i3 + MSIZE;
/*-----------------------------------------------------------------------*
* Reset max_time for 1st subframe. *
*-----------------------------------------------------------------------*/
if (i_subfr == 0) {
extra = 30;
}
/*-----------------------------------------------------------------------*
* Chose the sign of the impulse. *
*-----------------------------------------------------------------------*/
for (i=0; i<L_SUBFR; i++)
{
if( Dn[i] >= 0)
{
p_sign[i] = 0x7fff;
}
else
{
p_sign[i] = (Word16)0x8000;
Dn[i] = negate(Dn[i]);
}
}
/*-------------------------------------------------------------------*
* - Compute the search threshold after three pulses *
*-------------------------------------------------------------------*/
/* Find maximum of Dn[i0]+Dn[i1]+Dn[i2] */
max0 = Dn[0];
max1 = Dn[1];
max2 = Dn[2];
for (i = 5; i < L_SUBFR; i+=STEP)
{
if (sub(Dn[i] , max0) > 0) {
max0 = Dn[i];
}
if (sub(Dn[i+1], max1) > 0) {
max1 = Dn[i+1];
}
if (sub(Dn[i+2], max2) > 0) {
max2 = Dn[i+2];
}
}
max0 = add(max0, max1);
max0 = add(max0, max2);
/* Find average of Dn[i0]+Dn[i1]+Dn[i2] */
L32 = 0;
for (i = 0; i < L_SUBFR; i+=STEP)
{
L32 = L_mac(L32, Dn[i], 1);
L32 = L_mac(L32, Dn[i+1], 1);
L32 = L_mac(L32, Dn[i+2], 1);
}
average =extract_l( L_shr(L32, 4)); /* 1/8 of sum */
/* thres = average + (max0-average)*THRESHFCB; */
thres = sub(max0, average);
thres = mult(thres, THRESHFCB);
thres = add(thres, average);
/*-------------------------------------------------------------------*
* Modification of rrixiy[] to take signs into account. *
*-------------------------------------------------------------------*/
ptr_ri0i1 = rri0i1;
ptr_ri0i2 = rri0i2;
ptr_ri0i3 = rri0i3;
ptr_ri0i4 = rri0i4;
for(i0=0; i0<L_SUBFR; i0+=STEP)
{
for(i1=1; i1<L_SUBFR; i1+=STEP)
{
*ptr_ri0i1 = mult(*ptr_ri0i1, mult(p_sign[i0], p_sign[i1]));
ptr_ri0i1++;
*ptr_ri0i2 = mult(*ptr_ri0i2, mult(p_sign[i0], p_sign[i1+1]));
ptr_ri0i2++;
*ptr_ri0i3 = mult(*ptr_ri0i3, mult(p_sign[i0], p_sign[i1+2]));
ptr_ri0i3++;
*ptr_ri0i4 = mult(*ptr_ri0i4, mult(p_sign[i0], p_sign[i1+3]));
ptr_ri0i4++;
}
}
ptr_ri1i2 = rri1i2;
ptr_ri1i3 = rri1i3;
ptr_ri1i4 = rri1i4;
for(i1=1; i1<L_SUBFR; i1+=STEP)
{
for(i2=2; i2<L_SUBFR; i2+=STEP)
{
*ptr_ri1i2 = mult(*ptr_ri1i2, mult(p_sign[i1], p_sign[i2]));
ptr_ri1i2++;
*ptr_ri1i3 = mult(*ptr_ri1i3, mult(p_sign[i1], p_sign[i2+1]));
ptr_ri1i3++;
*ptr_ri1i4 = mult(*ptr_ri1i4, mult(p_sign[i1], p_sign[i2+2]));
ptr_ri1i4++;
}
}
ptr_ri2i3 = rri2i3;
ptr_ri2i4 = rri2i4;
for(i2=2; i2<L_SUBFR; i2+=STEP)
{
for(i3=3; i3<L_SUBFR; i3+=STEP)
{
*ptr_ri2i3 = mult(*ptr_ri2i3, mult(p_sign[i2], p_sign[i3]));
ptr_ri2i3++;
*ptr_ri2i4 = mult(*ptr_ri2i4, mult(p_sign[i2], p_sign[i3+1]));
ptr_ri2i4++;
}
}
/*-------------------------------------------------------------------*
* Search the optimum positions of the four pulses which maximize *
* square(correlation) / energy *
* The search is performed in four nested loops. At each loop, one *
* pulse contribution is added to the correlation and energy. *
* *
* The fourth loop is entered only if the correlation due to the *
* contribution of the first three pulses exceeds the preset *
* threshold. *
*-------------------------------------------------------------------*/
/* Default values */
ip0 = 0;
ip1 = 1;
ip2 = 2;
ip3 = 3;
psc = 0;
alpha = MAX_16;
time = add(MAX_TIME, extra);
/* Four loops to search innovation code. */
ptr_ri0i0 = rri0i0; /* Init. pointers that depend on first loop */
ptr_ri0i1 = rri0i1;
ptr_ri0i2 = rri0i2;
ptr_ri0i3 = rri0i3;
ptr_ri0i4 = rri0i4;
for (i0 = 0; i0 < L_SUBFR; i0 += STEP) /* first pulse loop */
{
ps0 = Dn[i0];
alp0 = *ptr_ri0i0++;
ptr_ri1i1 = rri1i1; /* Init. pointers that depend on second loop */
ptr_ri1i2 = rri1i2;
ptr_ri1i3 = rri1i3;
ptr_ri1i4 = rri1i4;
for (i1 = 1; i1 < L_SUBFR; i1 += STEP) /* second pulse loop */
{
ps1 = add(ps0, Dn[i1]);
/* alp1 = alp0 + *ptr_ri1i1++ + 2.0 * ( *ptr_ri0i1++); */
alp1 = L_mult(alp0, 1);
alp1 = L_mac(alp1, *ptr_ri1i1++, 1);
alp1 = L_mac(alp1, *ptr_ri0i1++, 2);
ptr_ri2i2 = rri2i2; /* Init. pointers that depend on third loop */
ptr_ri2i3 = rri2i3;
ptr_ri2i4 = rri2i4;
for (i2 = 2; i2 < L_SUBFR; i2 += STEP) /* third pulse loop */
{
ps2 = add(ps1, Dn[i2]);
/* alp2 = alp1 + *ptr_ri2i2++ + 2.0 * (*ptr_ri0i2++ + *ptr_ri1i2++); */
alp2 = L_mac(alp1, *ptr_ri2i2++, 1);
alp2 = L_mac(alp2, *ptr_ri0i2++, 2);
alp2 = L_mac(alp2, *ptr_ri1i2++, 2);
/* Test threshold */
if ( sub(ps2, thres) > 0)
{
ptr_ri3i3 = rri3i3; /* Init. pointers that depend on 4th loop */
for (i3 = 3; i3 < L_SUBFR; i3 += STEP) /* 4th pulse loop */
{
ps3 = add(ps2, Dn[i3]);
/* alp3 = alp2 + *ptr_ri3i3++ */
/* + 2.0*( *ptr_ri0i3++ + *ptr_ri1i3++ + *ptr_ri2i3++); */
alp3 = L_mac(alp2, *ptr_ri3i3++, 1);
alp3 = L_mac(alp3, *ptr_ri0i3++, 2);
alp3 = L_mac(alp3, *ptr_ri1i3++, 2);
alp3 = L_mac(alp3, *ptr_ri2i3++, 2);
alp = extract_l(L_shr(alp3, 5));
ps3c = mult(ps3, ps3);
L_temp = L_mult(ps3c, alpha);
L_temp = L_msu(L_temp, psc, alp);
if( L_temp > 0L )
{
psc = ps3c;
alpha = alp;
ip0 = i0;
ip1 = i1;
ip2 = i2;
ip3 = i3;
}
} /* end of for i3 = */
ptr_ri0i3 -= NB_POS;
ptr_ri1i3 -= NB_POS;
ptr_ri4i4 = rri4i4; /* Init. pointers that depend on 4th loop */
for (i3 = 4; i3 < L_SUBFR; i3 += STEP) /* 4th pulse loop */
{
ps3 = add(ps2, Dn[i3]);
/* alp3 = alp2 + *ptr_ri4i4++ */
/* + 2.0*( *ptr_ri0i4++ + *ptr_ri1i4++ + *ptr_ri2i4++); */
alp3 = L_mac(alp2, *ptr_ri4i4++, 1);
alp3 = L_mac(alp3, *ptr_ri0i4++, 2);
alp3 = L_mac(alp3, *ptr_ri1i4++, 2);
alp3 = L_mac(alp3, *ptr_ri2i4++, 2);
alp = extract_l(L_shr(alp3, 5));
ps3c = mult(ps3, ps3);
L_temp = L_mult(ps3c, alpha);
L_temp = L_msu(L_temp, psc, alp);
if( L_temp > 0L )
{
psc = ps3c;
alpha = alp;
ip0 = i0;
ip1 = i1;
ip2 = i2;
ip3 = i3;
}
} /* end of for i3 = */
ptr_ri0i4 -= NB_POS;
ptr_ri1i4 -= NB_POS;
time = sub(time, 1);
if(time <= 0 ) goto end_search; /* Maximum time finish */
} /* end of if >thres */
else
{
ptr_ri2i3 += NB_POS;
ptr_ri2i4 += NB_POS;
}
} /* end of for i2 = */
ptr_ri0i2 -= NB_POS;
ptr_ri1i3 += NB_POS;
ptr_ri1i4 += NB_POS;
} /* end of for i1 = */
ptr_ri0i2 += NB_POS;
ptr_ri0i3 += NB_POS;
ptr_ri0i4 += NB_POS;
} /* end of for i0 = */
end_search:
extra = time;
/* Set the sign of impulses */
i0 = p_sign[ip0];
i1 = p_sign[ip1];
i2 = p_sign[ip2];
i3 = p_sign[ip3];
/* Find the codeword corresponding to the selected positions */
for(i=0; i<L_SUBFR; i++) {
cod[i] = 0;
}
cod[ip0] = shr(i0, 2); /* From Q15 to Q13 */
cod[ip1] = shr(i1, 2);
cod[ip2] = shr(i2, 2);
cod[ip3] = shr(i3, 2);
/* find the filtered codeword */
for (i = 0; i < L_SUBFR; i++) {
y[i] = 0;
}
if(i0 > 0)
for(i=ip0, j=0; i<L_SUBFR; i++, j++) {
y[i] = add(y[i], h[j]);
}
else
for(i=ip0, j=0; i<L_SUBFR; i++, j++) {
y[i] = sub(y[i], h[j]);
}
if(i1 > 0)
for(i=ip1, j=0; i<L_SUBFR; i++, j++) {
y[i] = add(y[i], h[j]);
}
else
for(i=ip1, j=0; i<L_SUBFR; i++, j++) {
y[i] = sub(y[i], h[j]);
}
if(i2 > 0)
for(i=ip2, j=0; i<L_SUBFR; i++, j++) {
y[i] = add(y[i], h[j]);
}
else
for(i=ip2, j=0; i<L_SUBFR; i++, j++) {
y[i] = sub(y[i], h[j]);
}
if(i3 > 0)
for(i=ip3, j=0; i<L_SUBFR; i++, j++) {
y[i] = add(y[i], h[j]);
}
else
for(i=ip3, j=0; i<L_SUBFR; i++, j++) {
y[i] = sub(y[i], h[j]);
}
/* find codebook index; 17-bit address */
i = 0;
if(i0 > 0) i = add(i, 1);
if(i1 > 0) i = add(i, 2);
if(i2 > 0) i = add(i, 4);
if(i3 > 0) i = add(i, 8);
*sign = i;
ip0 = mult(ip0, 6554); /* ip0/5 */
ip1 = mult(ip1, 6554); /* ip1/5 */
ip2 = mult(ip2, 6554); /* ip2/5 */
i = mult(ip3, 6554); /* ip3/5 */
j = add(i, shl(i, 2)); /* j = i*5 */
j = sub(ip3, add(j, 3)); /* j= ip3%5 -3 */
ip3 = add(shl(i, 1), j);
i = add(ip0, shl(ip1, 3));
i = add(i , shl(ip2, 6));
i = add(i , shl(ip3, 9));
return i;
}
/***************************************************************************
Function: comp_powercat_and_catbalance
Syntax: void comp_powercat_and_catbalance(Int16 *power_categories,
Int16 *category_balances,
Int16 *rms_index,
Int16 number_of_available_bits,
Int16 num_categorization_control_possibilities,
Int16 offset)
inputs: *rms_index
number_of_available_bits
num_categorization_control_possibilities
offset
outputs: *power_categories
*category_balances
Description: Computes the power_categories and the category balances
WMOPS: 7kHz | 24kbit | 32kbit
-------|--------------|----------------
AVG | 0.10 | 0.10
-------|--------------|----------------
MAX | 0.11 | 0.11
-------|--------------|----------------
14kHz | 24kbit | 32kbit | 48kbit
-------|--------------|----------------|----------------
AVG | 0.32 | 0.35 | 0.38
-------|--------------|----------------|----------------
MAX | 0.38 | 0.42 | 0.43
-------|--------------|----------------|----------------
***************************************************************************/
void comp_powercat_and_catbalance(Int16 *power_categories,
Int16 *category_balances,
Int16 *rms_index,
Int16 number_of_available_bits,
Int16 offset)
{
Int16 region;
Int16 j;
Int16 max_rate_categories[MAX_NUMBER_OF_REGIONS];
Int16 min_rate_categories[MAX_NUMBER_OF_REGIONS];
Int16 temp_category_balances[2*MAX_NUM_CATEGORIZATION_CONTROL_POSSIBILITIES];
Int16 raw_max, raw_min;
Int16 raw_max_index, raw_min_index;
Int16 max_rate_pointer, min_rate_pointer;
Int16 max, min;
Int16 itemp0;
Int16 itemp1;
Int16 min_plus_max;
max = 0;
for (region = 0; region < gNumber_of_regions; region++)
{
j = sub(offset, rms_index[region]);
/* make sure j is between 0 and NUM_CAT-1 */
if (j < 0) j = 0;
else j = j >> 1;
if(j >= NUM_CATEGORIES)
j = NUM_CATEGORIES - 1;
/* compute the power categories based on the uniform offset */
power_categories[region] = j;
/* can't overflow */
max += expected_bits_table[j];
max_rate_categories[region] = j;
min_rate_categories[region] = j;
}
min = max;
max_rate_pointer = gNum_categorization_control_possibilities;
min_rate_pointer = gNum_categorization_control_possibilities;
for (j = 0; j < gNum_categorization_control_possibilities - 1; j++)
{
min_plus_max = add(max, min);
itemp0 = shl(number_of_available_bits, 1);
if (min_plus_max <= itemp0)
{
raw_min = 99;
/* Search from lowest freq regions to highest for best */
/* region to reassign to a higher bit rate category. */
for (region = 0; region < gNumber_of_regions; region++)
{
if (max_rate_categories[region] > 0)
{
itemp0 = max_rate_categories[region] << 1;
itemp1 = sub(offset, rms_index[region]);
itemp0 = sub(itemp1, itemp0);
if (itemp0 < raw_min)
{
raw_min = itemp0;
raw_min_index = region;
}
}
}
max_rate_pointer--;
temp_category_balances[max_rate_pointer] = raw_min_index;
max = sub(max, expected_bits_table[max_rate_categories[raw_min_index]]);
max_rate_categories[raw_min_index]--;
max = add(max, expected_bits_table[max_rate_categories[raw_min_index]]);
}
else
{
raw_max = -99;
/* Search from highest freq regions to lowest for best region to
reassign to a lower bit rate category. */
for (region = gNumber_of_regions; region--; )
{
if (min_rate_categories[region] < (NUM_CATEGORIES-1))
{
itemp0 = min_rate_categories[region] << 1;
itemp1 = sub(offset, rms_index[region]);
itemp0 = sub(itemp1, itemp0);
if (itemp0 > raw_max)
{
raw_max = itemp0;
raw_max_index = region;
}
}
}
temp_category_balances[min_rate_pointer] = raw_max_index;
min_rate_pointer++;
min = sub(min, expected_bits_table[min_rate_categories[raw_max_index]]);
min_rate_categories[raw_max_index]++;
min = add(min, expected_bits_table[min_rate_categories[raw_max_index]]);
}
}
for (region = gNumber_of_regions; region--; )
power_categories[region] = max_rate_categories[region];
for (j = 0; j < gNum_categorization_control_possibilities - 1; j++)
category_balances[j] = temp_category_balances[max_rate_pointer++];
}
Word16 Pitch_ol_fast( /* output: open loop pitch lag */
Word16 signal[], /* input : signal used to compute the open loop pitch */
/* signal[-pit_max] to signal[-1] should be known */
Word16 pit_max, /* input : maximum pitch lag */
Word16 L_frame /* input : length of frame to compute pitch */
)
{
Word16 i, j;
Word16 max1, max2, max3;
Word16 max_h, max_l, ener_h, ener_l;
Word16 T1, T2, T3;
Word16 *p, *p1;
Word32 max, sum, L_temp;
/* Scaled signal */
Word16 scaled_signal[L_FRAME+PIT_MAX];
Word16 *scal_sig;
scal_sig = &scaled_signal[pit_max];
/*--------------------------------------------------------*
* Verification for risk of overflow. *
*--------------------------------------------------------*/
Overflow = 0;
sum = 0;
for(i= -pit_max; i< L_frame; i+=2)
sum = L_mac(sum, signal[i], signal[i]);
/*--------------------------------------------------------*
* Scaling of input signal. *
* *
* if Overflow -> scal_sig[i] = signal[i]>>3 *
* else if sum < 1^20 -> scal_sig[i] = signal[i]<<3 *
* else -> scal_sig[i] = signal[i] *
*--------------------------------------------------------*/
if(Overflow == 1)
{
for(i=-pit_max; i<L_frame; i++)
{
scal_sig[i] = shr(signal[i], 3);
}
}
else {
L_temp = L_sub(sum, (Word32)1048576L);
if ( L_temp < (Word32)0 ) /* if (sum < 2^20) */
{
for(i=-pit_max; i<L_frame; i++)
{
scal_sig[i] = shl(signal[i], 3);
}
}
else
{
for(i=-pit_max; i<L_frame; i++)
{
scal_sig[i] = signal[i];
}
}
}
/*--------------------------------------------------------------------*
* The pitch lag search is divided in three sections. *
* Each section cannot have a pitch multiple. *
* We find a maximum for each section. *
* We compare the maxima of each section by favoring small lag. *
* *
* First section: lag delay = 20 to 39 *
* Second section: lag delay = 40 to 79 *
* Third section: lag delay = 80 to 143 *
*--------------------------------------------------------------------*/
/* First section */
max = MIN_32;
T1 = 20; /* Only to remove warning from some compilers */
for (i = 20; i < 40; i++) {
p = scal_sig;
p1 = &scal_sig[-i];
sum = 0;
for (j=0; j<L_frame; j+=2, p+=2, p1+=2)
sum = L_mac(sum, *p, *p1);
L_temp = L_sub(sum, max);
if (L_temp > 0) { max = sum; T1 = i; }
}
/* compute energy of maximum */
sum = 1; /* to avoid division by zero */
p = &scal_sig[-T1];
for(i=0; i<L_frame; i+=2, p+=2)
sum = L_mac(sum, *p, *p);
/* max1 = max/sqrt(energy) */
/* This result will always be on 16 bits !! */
sum = Inv_sqrt(sum); /* 1/sqrt(energy), result in Q30 */
L_Extract(max, &max_h, &max_l);
L_Extract(sum, &ener_h, &ener_l);
sum = Mpy_32(max_h, max_l, ener_h, ener_l);
max1 = extract_l(sum);
/* Second section */
max = MIN_32;
T2 = 40; /* Only to remove warning from some compilers */
for (i = 40; i < 80; i++) {
p = scal_sig;
p1 = &scal_sig[-i];
sum = 0;
for (j=0; j<L_frame; j+=2, p+=2, p1+=2)
sum = L_mac(sum, *p, *p1);
L_temp = L_sub(sum, max);
if (L_temp > 0) { max = sum; T2 = i; }
}
/* compute energy of maximum */
sum = 1; /* to avoid division by zero */
p = &scal_sig[-T2];
for(i=0; i<L_frame; i+=2, p+=2)
sum = L_mac(sum, *p, *p);
/* max2 = max/sqrt(energy) */
/* This result will always be on 16 bits !! */
sum = Inv_sqrt(sum); /* 1/sqrt(energy), result in Q30 */
L_Extract(max, &max_h, &max_l);
L_Extract(sum, &ener_h, &ener_l);
sum = Mpy_32(max_h, max_l, ener_h, ener_l);
max2 = extract_l(sum);
/* Third section */
max = MIN_32;
T3 = 80; /* Only to remove warning from some compilers */
for (i = 80; i < 143; i+=2) {
p = scal_sig;
p1 = &scal_sig[-i];
sum = 0;
for (j=0; j<L_frame; j+=2, p+=2, p1+=2)
sum = L_mac(sum, *p, *p1);
L_temp = L_sub(sum, max);
if (L_temp > 0) { max = sum; T3 = i; }
}
/* Test around max3 */
i = T3;
p = scal_sig;
p1 = &scal_sig[-(i+1)];
sum = 0;
for (j=0; j<L_frame; j+=2, p+=2, p1+=2)
sum = L_mac(sum, *p, *p1);
L_temp = L_sub(sum, max);
if (L_temp > 0) { max = sum; T3 = i+(Word16)1; }
p = scal_sig;
p1 = &scal_sig[-(i-1)];
sum = 0;
for (j=0; j<L_frame; j+=2, p+=2, p1+=2)
sum = L_mac(sum, *p, *p1);
L_temp = L_sub(sum, max);
if (L_temp > 0) { max = sum; T3 = i-(Word16)1; }
/* compute energy of maximum */
sum = 1; /* to avoid division by zero */
p = &scal_sig[-T3];
for(i=0; i<L_frame; i+=2, p+=2)
sum = L_mac(sum, *p, *p);
/* max1 = max/sqrt(energy) */
/* This result will always be on 16 bits !! */
sum = Inv_sqrt(sum); /* 1/sqrt(energy), result in Q30 */
L_Extract(max, &max_h, &max_l);
L_Extract(sum, &ener_h, &ener_l);
sum = Mpy_32(max_h, max_l, ener_h, ener_l);
max3 = extract_l(sum);
/*-----------------------*
* Test for multiple. *
*-----------------------*/
/* if( abs(T2*2 - T3) < 5) */
/* max2 += max3 * 0.25; */
i = sub(shl(T2,1), T3);
j = sub(abs_s(i), 5);
if(j < 0)
max2 = add(max2, shr(max3, 2));
/* if( abs(T2*3 - T3) < 7) */
/* max2 += max3 * 0.25; */
i = add(i, T2);
j = sub(abs_s(i), 7);
if(j < 0)
max2 = add(max2, shr(max3, 2));
/* if( abs(T1*2 - T2) < 5) */
/* max1 += max2 * 0.20; */
i = sub(shl(T1,1), T2);
j = sub(abs_s(i), 5);
if(j < 0)
max1 = add(max1, mult(max2, 6554));
/* if( abs(T1*3 - T2) < 7) */
/* max1 += max2 * 0.20; */
i = add(i, T1);
j = sub(abs_s(i), 7);
if(j < 0)
max1 = add(max1, mult(max2, 6554));
/*--------------------------------------------------------------------*
* Compare the 3 sections maxima. *
*--------------------------------------------------------------------*/
if( sub(max1, max2) < 0 ) {max1 = max2; T1 = T2; }
if( sub(max1, max3) <0 ) {T1 = T3; }
return T1;
}
void Relspwede(
Word16 lsp[], /* (i) Q13 : unquantized LSP parameters */
Word16 wegt[], /* (i) norm: weighting coefficients */
Word16 lspq[], /* (o) Q13 : quantized LSP parameters */
Word16 lspcb1[][M], /* (i) Q13 : first stage LSP codebook */
Word16 lspcb2[][M], /* (i) Q13 : Second stage LSP codebook */
Word16 fg[MODE][MA_NP][M], /* (i) Q15 : MA prediction coefficients */
Word16 freq_prev[MA_NP][M], /* (i) Q13 : previous LSP vector */
Word16 fg_sum[MODE][M], /* (i) Q15 : present MA prediction coef.*/
Word16 fg_sum_inv[MODE][M], /* (i) Q12 : inverse coef. */
Word16 code_ana[], /* (o) : codes of the selected LSP */
Word16 freq_cur[] /* (o) Q13 : current LSP MA vector */
)
{
Word16 mode, j;
Word16 index, mode_index;
Word16 cand[MODE], cand_cur;
Word16 tindex1[MODE], tindex2[MODE];
Word32 L_tdist[MODE]; /* Q26 */
Word16 rbuf[M]; /* Q13 */
Word16 buf[M]; /* Q13 */
for(mode = 0; mode<MODE; mode++) {
Lsp_prev_extract(lsp, rbuf, fg[mode], freq_prev, fg_sum_inv[mode]);
Lsp_pre_select(rbuf, lspcb1, &cand_cur );
cand[mode] = cand_cur;
Lsp_select_1(rbuf, lspcb1[cand_cur], wegt, lspcb2, &index);
tindex1[mode] = index;
for( j = 0 ; j < NC ; j++ )
buf[j] = add( lspcb1[cand_cur][j], lspcb2[index][j] );
Lsp_expand_1(buf, GAP1);
Lsp_select_2(rbuf, lspcb1[cand_cur], wegt, lspcb2, &index);
tindex2[mode] = index;
for( j = NC ; j < M ; j++ )
buf[j] = add( lspcb1[cand_cur][j], lspcb2[index][j] );
Lsp_expand_2(buf, GAP1);
Lsp_expand_1_2(buf, GAP2);
Lsp_get_tdist(wegt, buf, &L_tdist[mode], rbuf, fg_sum[mode]);
}
Lsp_last_select(L_tdist, &mode_index);
code_ana[0] = shl( mode_index,NC0_B ) | cand[mode_index];
code_ana[1] = shl( tindex1[mode_index],NC1_B ) | tindex2[mode_index];
Lsp_get_quante(lspcb1, lspcb2, cand[mode_index],
tindex1[mode_index], tindex2[mode_index],
fg[mode_index], freq_prev, lspq, fg_sum[mode_index], freq_cur) ;
return;
}
Word16 Cb_gain_average(
Cb_gain_averageState *st, /* i/o : State variables for CB gain averaging */
enum Mode mode, /* i : AMR mode */
Word16 gain_code, /* i : CB gain Q1 */
Word16 lsp[], /* i : The LSP for the current frame Q15 */
Word16 lspAver[], /* i : The average of LSP for 8 frames Q15 */
Word16 bfi, /* i : bad frame indication flag */
Word16 prev_bf, /* i : previous bad frame indication flag */
Word16 pdfi, /* i : potential degraded bad frame ind flag */
Word16 prev_pdf, /* i : prev pot. degraded bad frame ind flag */
Word16 inBackgroundNoise, /* i : background noise decision */
Word16 voicedHangover, /* i : # of frames after last voiced frame */
Flag *pOverflow
)
{
Word16 i;
Word16 cbGainMix;
Word16 diff;
Word16 tmp_diff;
Word16 bgMix;
Word16 cbGainMean;
Word32 L_sum;
Word16 tmp[M];
Word16 tmp1;
Word16 tmp2;
Word16 shift1;
Word16 shift2;
Word16 shift;
/*---------------------------------------------------------*
* Compute mixed cb gain, used to make cb gain more *
* smooth in background noise for modes 5.15, 5.9 and 6.7 *
* states that needs to be updated by all *
*---------------------------------------------------------*/
/* set correct cbGainMix for MR74, MR795, MR122 */
cbGainMix = gain_code;
/*-------------------------------------------------------*
* Store list of CB gain needed in the CB gain *
* averaging *
*-------------------------------------------------------*/
for (i = 0; i < (L_CBGAINHIST - 1); i++)
{
st->cbGainHistory[i] = st->cbGainHistory[i+1];
}
st->cbGainHistory[L_CBGAINHIST-1] = gain_code;
diff = 0;
/* compute lsp difference */
for (i = 0; i < M; i++)
{
tmp1 = abs_s(sub(*(lspAver + i), *(lsp + i), pOverflow));
/* Q15 */
shift1 = sub(norm_s(tmp1), 1, pOverflow); /* Qn */
tmp1 = shl(tmp1, shift1, pOverflow); /* Q15+Qn */
shift2 = norm_s(*(lspAver + i)); /* Qm */
tmp2 = shl(*(lspAver + i), shift2, pOverflow); /* Q15+Qm */
tmp[i] = div_s(tmp1, tmp2); /* Q15+(Q15+Qn)-(Q15+Qm) */
shift = 2 + shift1 - shift2;
if (shift >= 0)
{
*(tmp + i) = shr(*(tmp + i), shift, pOverflow);
/* Q15+Qn-Qm-Qx=Q13 */
}
else
{
*(tmp + i) = shl(*(tmp + i), negate(shift), pOverflow);
/* Q15+Qn-Qm-Qx=Q13 */
}
diff = add(diff, *(tmp + i), pOverflow); /* Q13 */
}
/* Compute hangover */
if (diff > 5325) /* 0.65 in Q11 */
{
st->hangVar += 1;
}
else
{
st->hangVar = 0;
}
if (st->hangVar > 10)
{
/* Speech period, reset hangover variable */
st->hangCount = 0;
}
/* Compute mix constant (bgMix) */
bgMix = 8192; /* 1 in Q13 */
if ((mode <= MR67) || (mode == MR102))
/* MR475, MR515, MR59, MR67, MR102 */
{
/* if errors and presumed noise make smoothing probability stronger */
if (((((pdfi != 0) && (prev_pdf != 0)) || (bfi != 0) ||
(prev_bf != 0))
&& (voicedHangover > 1)
&& (inBackgroundNoise != 0)
&& ((mode == MR475) || (mode == MR515) ||
(mode == MR59))))
{
/* bgMix = min(0.25, max(0.0, diff-0.55)) / 0.25; */
tmp_diff = sub(diff, 4506, pOverflow); /* 0.55 in Q13 */
}
else
{
/* bgMix = min(0.25, max(0.0, diff-0.40)) / 0.25; */
tmp_diff = sub(diff, 3277, pOverflow); /* 0.4 in Q13 */
}
/* max(0.0, diff-0.55) or */
/* max(0.0, diff-0.40) */
if (tmp_diff > 0)
{
tmp1 = tmp_diff;
}
else
{
tmp1 = 0;
}
/* min(0.25, tmp1) */
if (2048 < tmp1)
{
bgMix = 8192;
}
else
{
bgMix = shl(tmp1, 2, pOverflow);
}
if ((st->hangCount < 40) || (diff > 5325)) /* 0.65 in Q13 */
{
/* disable mix if too short time since */
bgMix = 8192;
}
/* Smoothen the cb gain trajectory */
/* smoothing depends on mix constant bgMix */
L_sum = L_mult(6554, st->cbGainHistory[2], pOverflow);
/* 0.2 in Q15; L_sum in Q17 */
for (i = 3; i < L_CBGAINHIST; i++)
{
L_sum = L_mac(L_sum, 6554, st->cbGainHistory[i], pOverflow);
}
cbGainMean = pv_round(L_sum, pOverflow); /* Q1 */
/* more smoothing in error and bg noise (NB no DFI used here) */
if (((bfi != 0) || (prev_bf != 0)) && (inBackgroundNoise != 0)
&& ((mode == MR475) || (mode == MR515)
|| (mode == MR59)))
{
/* 0.143 in Q15; L_sum in Q17 */
L_sum = L_mult(4681, st->cbGainHistory[0], pOverflow);
for (i = 1; i < L_CBGAINHIST; i++)
{
L_sum =
L_mac(L_sum, 4681, st->cbGainHistory[i], pOverflow);
}
cbGainMean = pv_round(L_sum, pOverflow); /* Q1 */
}
/* cbGainMix = bgMix*cbGainMix + (1-bgMix)*cbGainMean; */
/* L_sum in Q15 */
L_sum = L_mult(bgMix, cbGainMix, pOverflow);
L_sum = L_mac(L_sum, 8192, cbGainMean, pOverflow);
L_sum = L_msu(L_sum, bgMix, cbGainMean, pOverflow);
cbGainMix = pv_round(L_shl(L_sum, 2, pOverflow), pOverflow); /* Q1 */
}
st->hangCount += 1;
return (cbGainMix);
}
/*************************************************************************
*
* FUNCTION: Pitch_ol
*
* PURPOSE: Compute the open loop pitch lag.
*
* DESCRIPTION:
* The open-loop pitch lag is determined based on the perceptually
* weighted speech signal. This is done in the following steps:
* - find three maxima of the correlation <sw[n],sw[n-T]>,
* dividing the search range into three parts:
* pit_min ... 2*pit_min-1
* 2*pit_min ... 4*pit_min-1
* 4*pit_min ... pit_max
* - divide each maximum by <sw[n-t], sw[n-t]> where t is the delay at
* that maximum correlation.
* - select the delay of maximum normalized correlation (among the
* three candidates) while favoring the lower delay ranges.
*
*************************************************************************/
Word16 Pitch_ol ( /* o : open loop pitch lag */
vadState *vadSt, /* i/o : VAD state struct */
enum Mode mode, /* i : coder mode */
Word16 signal[], /* i : signal used to compute the open loop pitch */
/* signal[-pit_max] to signal[-1] should be known */
Word16 pit_min, /* i : minimum pitch lag */
Word16 pit_max, /* i : maximum pitch lag */
Word16 L_frame, /* i : length of frame to compute pitch */
Word16 idx, /* i : frame index */
Flag dtx /* i : dtx flag; use dtx=1, do not use dtx=0 */
)
{
Word16 i, j;
Word16 max1, max2, max3;
Word16 p_max1, p_max2, p_max3;
Word16 scal_flag = 0;
Word32 t0;
#ifdef VAD2
Word32 r01, r02, r03;
Word32 rmax1, rmax2, rmax3;
#else
Word16 corr_hp_max;
#endif
Word32 corr[PIT_MAX+1], *corr_ptr;
/* Scaled signal */
Word16 scaled_signal[L_FRAME + PIT_MAX];
Word16 *scal_sig, scal_fac;
#ifndef VAD2
if (dtx)
{ /* no test() call since this if is only in simulation env */
/* update tone detection */
test(); test();
if ((sub(mode, MR475) == 0) || (sub(mode, MR515) == 0))
{
vad_tone_detection_update (vadSt, 1);
}
else
{
vad_tone_detection_update (vadSt, 0);
}
}
#endif
scal_sig = &scaled_signal[pit_max]; move16 ();
t0 = 0L; move32 ();
for (i = -pit_max; i < L_frame; i++)
{
t0 = L_mac (t0, signal[i], signal[i]);
}
/*--------------------------------------------------------*
* Scaling of input signal. *
* *
* if Overflow -> scal_sig[i] = signal[i]>>3 *
* else if t0 < 1^20 -> scal_sig[i] = signal[i]<<3 *
* else -> scal_sig[i] = signal[i] *
*--------------------------------------------------------*/
/*--------------------------------------------------------*
* Verification for risk of overflow. *
*--------------------------------------------------------*/
test ();
if (L_sub (t0, MAX_32) == 0L) /* Test for overflow */
{
for (i = -pit_max; i < L_frame; i++)
{
scal_sig[i] = shr (signal[i], 3); move16 ();
}
scal_fac = 3; move16 ();
}
else if (L_sub (t0, (Word32) 1048576L) < (Word32) 0)
/* if (t0 < 2^20) */
{
test ();
for (i = -pit_max; i < L_frame; i++)
{
scal_sig[i] = shl (signal[i], 3); move16 ();
}
scal_fac = -3; move16 ();
}
else
{
test ();
for (i = -pit_max; i < L_frame; i++)
{
scal_sig[i] = signal[i]; move16 ();
}
scal_fac = 0; move16 ();
}
/* calculate all coreelations of scal_sig, from pit_min to pit_max */
corr_ptr = &corr[pit_max]; move32 ();
comp_corr (scal_sig, L_frame, pit_max, pit_min, corr_ptr);
/*--------------------------------------------------------------------*
* The pitch lag search is divided in three sections. *
* Each section cannot have a pitch multiple. *
* We find a maximum for each section. *
* We compare the maximum of each section by favoring small lags. *
* *
* First section: lag delay = pit_max downto 4*pit_min *
* Second section: lag delay = 4*pit_min-1 downto 2*pit_min *
* Third section: lag delay = 2*pit_min-1 downto pit_min *
*--------------------------------------------------------------------*/
/* mode dependent scaling in Lag_max */
test ();
if (sub(mode, MR122) == 0)
{
scal_flag = 1; move16 ();
}
else
{
scal_flag = 0; move16 ();
}
#ifdef VAD2
j = shl (pit_min, 2);
p_max1 = Lag_max (corr_ptr, scal_sig, scal_fac, scal_flag, L_frame,
pit_max, j, &max1, &rmax1, &r01, dtx);
move16 (); /* function result */
i = sub (j, 1);
j = shl (pit_min, 1);
p_max2 = Lag_max (corr_ptr, scal_sig, scal_fac, scal_flag, L_frame,
i, j, &max2, &rmax2, &r02, dtx);
move16 (); /* function result */
i = sub (j, 1);
p_max3 = Lag_max (corr_ptr, scal_sig, scal_fac, scal_flag, L_frame,
i, pit_min, &max3, &rmax3, &r03, dtx);
move16 (); /* function result */
#else
j = shl (pit_min, 2);
p_max1 = Lag_max (vadSt, corr_ptr, scal_sig, scal_fac, scal_flag, L_frame,
pit_max, j, &max1, dtx); move16 (); /* function result */
i = sub (j, 1);
j = shl (pit_min, 1);
p_max2 = Lag_max (vadSt, corr_ptr, scal_sig, scal_fac, scal_flag, L_frame,
i, j, &max2, dtx); move16 (); /* function result */
i = sub (j, 1);
p_max3 = Lag_max (vadSt, corr_ptr, scal_sig, scal_fac, scal_flag, L_frame,
i, pit_min, &max3, dtx); move16 (); /* function result */
if (dtx)
{ /* no test() call since this if is only in simulation env */
test ();
if (sub(idx, 1) == 0)
{
/* calculate max high-passed filtered correlation of all lags */
hp_max (corr_ptr, scal_sig, L_frame, pit_max, pit_min, &corr_hp_max);
/* update complex background detector */
vad_complex_detection_update(vadSt, corr_hp_max);
}
}
#endif
/*--------------------------------------------------------------------*
* Compare the 3 sections maximum, and favor small lag. *
*--------------------------------------------------------------------*/
test ();
if (sub (mult (max1, THRESHOLD), max2) < 0)
{
max1 = max2; move16 ();
p_max1 = p_max2; move16 ();
#ifdef VAD2
if (dtx)
{
rmax1 = rmax2; move32 ();
r01 = r02; move32 ();
}
#endif
}
test ();
if (sub (mult (max1, THRESHOLD), max3) < 0)
{
p_max1 = p_max3; move16 ();
#ifdef VAD2
if (dtx)
{
rmax1 = rmax3; move32 ();
r01 = r03; move32 ();
}
#endif
}
#ifdef VAD2
if (dtx)
{
vadSt->L_Rmax = L_add(vadSt->L_Rmax, rmax1); /* Save max correlation */
vadSt->L_R0 = L_add(vadSt->L_R0, r01); /* Save max energy */
}
#endif
return (p_max1);
}
void GPUDrawScanlineCodeGenerator::Init()
{
mov(eax, dword[esp + _top]);
// uint16* fb = (uint16*)m_global.vm + (top << (10 + sel.scalex)) + left;
mov(edi, eax);
shl(edi, 10 + m_sel.scalex);
add(edi, edx);
lea(edi, ptr[edi * 2 + (size_t)m_local.gd->vm]);
// int steps = pixels - 8;
sub(ecx, 8);
if(m_sel.dtd)
{
// dither = GSVector4i::load<false>(&m_dither[top & 3][left & 3]);
and(eax, 3);
shl(eax, 5);
and(edx, 3);
shl(edx, 1);
movdqu(xmm0, ptr[eax + edx + (size_t)m_dither]);
movdqa(ptr[&m_local.temp.dither], xmm0);
}
mov(edx, dword[esp + _v]);
if(m_sel.tme)
{
mov(esi, dword[&m_local.gd->tex]);
// GSVector4i vt = GSVector4i(v.t).xxzzl();
cvttps2dq(xmm4, ptr[edx + offsetof(GSVertexSW, t)]);
pshuflw(xmm4, xmm4, _MM_SHUFFLE(2, 2, 0, 0));
// s = vt.xxxx().add16(m_local.d.s);
// t = vt.yyyy().add16(m_local.d.t);
pshufd(xmm2, xmm4, _MM_SHUFFLE(0, 0, 0, 0));
pshufd(xmm3, xmm4, _MM_SHUFFLE(1, 1, 1, 1));
paddw(xmm2, ptr[&m_local.d.s]);
if(!m_sel.sprite)
{
paddw(xmm3, ptr[&m_local.d.t]);
}
else
{
if(m_sel.ltf)
{
movdqa(xmm0, xmm3);
psllw(xmm0, 8);
psrlw(xmm0, 1);
movdqa(ptr[&m_local.temp.vf], xmm0);
}
}
movdqa(ptr[&m_local.temp.s], xmm2);
movdqa(ptr[&m_local.temp.t], xmm3);
}
if(m_sel.tfx != 3) // != decal
{
// GSVector4i vc = GSVector4i(v.c).xxzzlh();
cvttps2dq(xmm6, ptr[edx + offsetof(GSVertexSW, c)]);
pshuflw(xmm6, xmm6, _MM_SHUFFLE(2, 2, 0, 0));
pshufhw(xmm6, xmm6, _MM_SHUFFLE(2, 2, 0, 0));
// r = vc.xxxx();
// g = vc.yyyy();
// b = vc.zzzz();
pshufd(xmm4, xmm6, _MM_SHUFFLE(0, 0, 0, 0));
pshufd(xmm5, xmm6, _MM_SHUFFLE(1, 1, 1, 1));
pshufd(xmm6, xmm6, _MM_SHUFFLE(2, 2, 2, 2));
if(m_sel.iip)
{
// r = r.add16(m_local.d.r);
// g = g.add16(m_local.d.g);
// b = b.add16(m_local.d.b);
paddw(xmm4, ptr[&m_local.d.r]);
paddw(xmm5, ptr[&m_local.d.g]);
paddw(xmm6, ptr[&m_local.d.b]);
}
movdqa(ptr[&m_local.temp.r], xmm4);
movdqa(ptr[&m_local.temp.g], xmm5);
movdqa(ptr[&m_local.temp.b], xmm6);
}
}
/*---------------------------------------------------------------------------*
* Function Qua_gain *
* ~~~~~~~~~~~~~~~~~~ *
* Inputs: *
* code[] :Innovative codebook. *
* g_coeff[] :Correlations compute for pitch. *
* L_subfr :Subframe length. *
* *
* Outputs: *
* gain_pit :Quantized pitch gain. *
* gain_cod :Quantized code gain. *
* *
* Return: *
* Index of quantization. *
* *
*--------------------------------------------------------------------------*/
Word16 Qua_gain(
Word16 code[], /* (i) Q13 :Innovative vector. */
Word16 g_coeff[], /* (i) :Correlations <xn y1> -2<y1 y1> */
/* <y2,y2>, -2<xn,y2>, 2<y1,y2> */
Word16 exp_coeff[], /* (i) :Q-Format g_coeff[] */
Word16 L_subfr, /* (i) :Subframe length. */
Word16 *gain_pit, /* (o) Q14 :Pitch gain. */
Word16 *gain_cod, /* (o) Q1 :Code gain. */
Word16 tameflag /* (i) : set to 1 if taming is needed */
)
{
Word16 i, j, index1, index2;
Word16 cand1, cand2;
Word16 exp, gcode0, exp_gcode0, gcode0_org, e_min ;
Word16 nume, denom, inv_denom;
Word16 exp1,exp2,exp_nume,exp_denom,exp_inv_denom,sft,tmp;
Word16 g_pitch, g2_pitch, g_code, g2_code, g_pit_cod;
Word16 coeff[5], coeff_lsf[5];
Word16 exp_min[5];
Word32 L_gbk12;
Word32 L_tmp, L_dist_min, L_tmp1, L_tmp2, L_acc, L_accb;
Word16 best_gain[2];
/* Gain predictor, Past quantized energies = -14.0 in Q10 */
static Word16 past_qua_en[4] = { -14336, -14336, -14336, -14336 };
/*---------------------------------------------------*
*- energy due to innovation -*
*- predicted energy -*
*- predicted codebook gain => gcode0[exp_gcode0] -*
*---------------------------------------------------*/
Gain_predict( past_qua_en, code, L_subfr, &gcode0, &exp_gcode0 );
/*-----------------------------------------------------------------*
* pre-selection *
*-----------------------------------------------------------------*/
/*-----------------------------------------------------------------*
* calculate best gain *
* *
* tmp = -1./(4.*coeff[0]*coeff[2]-coeff[4]*coeff[4]) ; *
* best_gain[0] = (2.*coeff[2]*coeff[1]-coeff[3]*coeff[4])*tmp ; *
* best_gain[1] = (2.*coeff[0]*coeff[3]-coeff[1]*coeff[4])*tmp ; *
* gbk_presel(best_gain,&cand1,&cand2,gcode0) ; *
* *
*-----------------------------------------------------------------*/
/*-----------------------------------------------------------------*
* tmp = -1./(4.*coeff[0]*coeff[2]-coeff[4]*coeff[4]) ; *
*-----------------------------------------------------------------*/
L_tmp1 = L_mult( g_coeff[0], g_coeff[2] );
exp1 = add( add( exp_coeff[0], exp_coeff[2] ), 1-2 );
L_tmp2 = L_mult( g_coeff[4], g_coeff[4] );
exp2 = add( add( exp_coeff[4], exp_coeff[4] ), 1 );
//if( sub(exp1, exp2)>0 ){
if( exp1 > exp2 ){
L_tmp = L_sub( L_shr( L_tmp1, sub(exp1,exp2) ), L_tmp2 );
exp = exp2;
}
else{
L_tmp = L_sub( L_tmp1, L_shr( L_tmp2, sub(exp2,exp1) ) );
exp = exp1;
}
sft = norm_l_g729( L_tmp );
denom = extract_h( L_shl(L_tmp, sft) );
exp_denom = sub( add( exp, sft ), 16 );
inv_denom = div_s_g729(16384,denom);
inv_denom = negate( inv_denom );
exp_inv_denom = sub( 14+15, exp_denom );
/*-----------------------------------------------------------------*
* best_gain[0] = (2.*coeff[2]*coeff[1]-coeff[3]*coeff[4])*tmp ; *
*-----------------------------------------------------------------*/
L_tmp1 = L_mult( g_coeff[2], g_coeff[1] );
exp1 = add( exp_coeff[2], exp_coeff[1] );
L_tmp2 = L_mult( g_coeff[3], g_coeff[4] );
exp2 = add( add( exp_coeff[3], exp_coeff[4] ), 1 );
//if( sub(exp1, exp2)>0 ){
if (exp1 > exp2){
L_tmp = L_sub( L_shr( L_tmp1, add(sub(exp1,exp2),1 )), L_shr( L_tmp2,1 ) );
exp = sub(exp2,1);
}
else{
L_tmp = L_sub( L_shr( L_tmp1,1 ), L_shr( L_tmp2, add(sub(exp2,exp1),1 )) );
exp = sub(exp1,1);
}
sft = norm_l_g729( L_tmp );
nume = extract_h( L_shl(L_tmp, sft) );
exp_nume = sub( add( exp, sft ), 16 );
sft = sub( add( exp_nume, exp_inv_denom ), (9+16-1) );
L_acc = L_shr( L_mult( nume,inv_denom ), sft );
best_gain[0] = extract_h( L_acc ); /*-- best_gain[0]:Q9 --*/
if (tameflag == 1){
//if(sub(best_gain[0], GPCLIP2) > 0) best_gain[0] = GPCLIP2;
if(best_gain[0] > GPCLIP2) best_gain[0] = GPCLIP2;
}
/*-----------------------------------------------------------------*
* best_gain[1] = (2.*coeff[0]*coeff[3]-coeff[1]*coeff[4])*tmp ; *
*-----------------------------------------------------------------*/
L_tmp1 = L_mult( g_coeff[0], g_coeff[3] );
exp1 = add( exp_coeff[0], exp_coeff[3] ) ;
L_tmp2 = L_mult( g_coeff[1], g_coeff[4] );
exp2 = add( add( exp_coeff[1], exp_coeff[4] ), 1 );
//if( sub(exp1, exp2)>0 ){
if( exp1 > exp2 ){
L_tmp = L_sub( L_shr( L_tmp1, add(sub(exp1,exp2),1) ), L_shr( L_tmp2,1 ) );
exp = sub(exp2,1);
//exp = exp2--;
}
else{
L_tmp = L_sub( L_shr( L_tmp1,1 ), L_shr( L_tmp2, add(sub(exp2,exp1),1) ) );
exp = sub(exp1,1);
//exp = exp1--;
}
sft = norm_l_g729( L_tmp );
nume = extract_h( L_shl(L_tmp, sft) );
exp_nume = sub( add( exp, sft ), 16 );
sft = sub( add( exp_nume, exp_inv_denom ), (2+16-1) );
L_acc = L_shr( L_mult( nume,inv_denom ), sft );
best_gain[1] = extract_h( L_acc ); /*-- best_gain[1]:Q2 --*/
/*--- Change Q-format of gcode0 ( Q[exp_gcode0] -> Q4 ) ---*/
//if( sub(exp_gcode0,4) >= 0 ){
if (exp_gcode0 >=4) {
gcode0_org = shr_g729( gcode0, sub(exp_gcode0,4) );
}
else{
L_acc = L_deposit_l_g729( gcode0 );
L_acc = L_shl( L_acc, sub( (4+16), exp_gcode0 ) );
gcode0_org = extract_h( L_acc ); /*-- gcode0_org:Q4 --*/
}
/*----------------------------------------------*
* - presearch for gain codebook - *
*----------------------------------------------*/
Gbk_presel(best_gain, &cand1, &cand2, gcode0_org );
/*---------------------------------------------------------------------------*
* *
* Find the best quantizer. *
* *
* dist_min = MAX_32; *
* for ( i=0 ; i<NCAN1 ; i++ ){ *
* for ( j=0 ; j<NCAN2 ; j++ ){ *
* g_pitch = gbk1[cand1+i][0] + gbk2[cand2+j][0]; *
* g_code = gcode0 * (gbk1[cand1+i][1] + gbk2[cand2+j][1]); *
* dist = g_pitch*g_pitch * coeff[0] *
* + g_pitch * coeff[1] *
* + g_code*g_code * coeff[2] *
* + g_code * coeff[3] *
* + g_pitch*g_code * coeff[4] ; *
* *
* if (dist < dist_min){ *
* dist_min = dist; *
* indice1 = cand1 + i ; *
* indice2 = cand2 + j ; *
* } *
* } *
* } *
* *
* g_pitch = Q13 *
* g_pitch*g_pitch = Q11:(13+13+1-16) *
* g_code = Q[exp_gcode0-3]:(exp_gcode0+(13-1)+1-16) *
* g_code*g_code = Q[2*exp_gcode0-21]:(exp_gcode0-3+exp_gcode0-3+1-16) *
* g_pitch*g_code = Q[exp_gcode0-5]:(13+exp_gcode0-3+1-16) *
* *
* term 0: g_pitch*g_pitch*coeff[0] ;exp_min0 = 13 +exp_coeff[0] *
* term 1: g_pitch *coeff[1] ;exp_min1 = 14 +exp_coeff[1] *
* term 2: g_code*g_code *coeff[2] ;exp_min2 = 2*exp_gcode0-21+exp_coeff[2] *
* term 3: g_code *coeff[3] ;exp_min3 = exp_gcode0 - 3+exp_coeff[3] *
* term 4: g_pitch*g_code *coeff[4] ;exp_min4 = exp_gcode0 - 4+exp_coeff[4] *
* *
*---------------------------------------------------------------------------*/
exp_min[0] = add( exp_coeff[0], 13 );
exp_min[1] = add( exp_coeff[1], 14 );
exp_min[2] = add( exp_coeff[2], sub( shl( exp_gcode0, 1 ), 21 ) );
exp_min[3] = add( exp_coeff[3], sub( exp_gcode0, 3 ) );
exp_min[4] = add( exp_coeff[4], sub( exp_gcode0, 4 ) );
e_min = exp_min[0];
for(i=1; i<5; i++){
//if( sub(exp_min[i], e_min) < 0 ){
if (exp_min[i] < e_min) {
e_min = exp_min[i];
}
}
/* align coeff[] and save in special 32 bit double precision */
for(i=0; i<5; i++){
j = sub( exp_min[i], e_min );
L_tmp = (Word32)g_coeff[i] << 16;
L_tmp = L_shr( L_tmp, j ); /* L_tmp:Q[exp_g_coeff[i]+16-j] */
L_Extract( L_tmp, &coeff[i], &coeff_lsf[i] ); /* DPF */
}
/* Codebook search */
L_dist_min = MAX_32;
/* initialization used only to suppress Microsoft Visual C++ warnings */
index1 = cand1;
index2 = cand2;
if(tameflag == 1){
for(i=0; i<NCAN1; i++){
for(j=0; j<NCAN2; j++){
g_pitch = add( gbk1[cand1+i][0], gbk2[cand2+j][0] ); /* Q14 */
if(g_pitch < GP0999) {
L_acc = L_deposit_l_g729( gbk1[cand1+i][1] );
L_accb = L_deposit_l_g729( gbk2[cand2+j][1] ); /* Q13 */
L_tmp = L_add( L_acc,L_accb );
tmp = extract_l( L_shr( L_tmp,1 ) ); /* Q12 */
g_code = mult( gcode0, tmp ); /* Q[exp_gcode0+12-15] */
g2_pitch = mult(g_pitch, g_pitch); /* Q13 */
g2_code = mult(g_code, g_code); /* Q[2*exp_gcode0-6-15] */
g_pit_cod= mult(g_code, g_pitch); /* Q[exp_gcode0-3+14-15] */
L_tmp = Mpy_32_16(coeff[0], coeff_lsf[0], g2_pitch);
//L_tmp = L_add(L_tmp, Mpy_32_16(coeff[1], coeff_lsf[1], g_pitch) );
//L_tmp = L_add(L_tmp, Mpy_32_16(coeff[2], coeff_lsf[2], g2_code) );
//L_tmp = L_add(L_tmp, Mpy_32_16(coeff[3], coeff_lsf[3], g_code) );
//L_tmp = L_add(L_tmp, Mpy_32_16(coeff[4], coeff_lsf[4], g_pit_cod) );
L_tmp += Mpy_32_16(coeff[1], coeff_lsf[1], g_pitch);
L_tmp += Mpy_32_16(coeff[2], coeff_lsf[2], g2_code);
L_tmp += Mpy_32_16(coeff[3], coeff_lsf[3], g_code);
L_tmp += Mpy_32_16(coeff[4], coeff_lsf[4], g_pit_cod);
//L_temp = L_sub(L_tmp, L_dist_min);
//if( L_temp < 0L ){
if( L_tmp < L_dist_min ){
L_dist_min = L_tmp;
index1 = add(cand1,i);
index2 = add(cand2,j);
}
}
}
}
}
else{
for(i=0; i<NCAN1; i++){
for(j=0; j<NCAN2; j++){
g_pitch = add( gbk1[cand1+i][0], gbk2[cand2+j][0] ); /* Q14 */
L_acc = L_deposit_l_g729( gbk1[cand1+i][1] );
L_accb = L_deposit_l_g729( gbk2[cand2+j][1] ); /* Q13 */
L_tmp = L_add( L_acc,L_accb );
tmp = extract_l( L_shr( L_tmp,1 ) ); /* Q12 */
g_code = mult( gcode0, tmp ); /* Q[exp_gcode0+12-15] */
g2_pitch = mult(g_pitch, g_pitch); /* Q13 */
g2_code = mult(g_code, g_code); /* Q[2*exp_gcode0-6-15] */
g_pit_cod= mult(g_code, g_pitch); /* Q[exp_gcode0-3+14-15] */
L_tmp = Mpy_32_16(coeff[0], coeff_lsf[0], g2_pitch);
//L_tmp = L_add(L_tmp, Mpy_32_16(coeff[1], coeff_lsf[1], g_pitch) );
//L_tmp = L_add(L_tmp, Mpy_32_16(coeff[2], coeff_lsf[2], g2_code) );
//L_tmp = L_add(L_tmp, Mpy_32_16(coeff[3], coeff_lsf[3], g_code) );
//L_tmp = L_add(L_tmp, Mpy_32_16(coeff[4], coeff_lsf[4], g_pit_cod) );
L_tmp += Mpy_32_16(coeff[1], coeff_lsf[1], g_pitch);
L_tmp += Mpy_32_16(coeff[2], coeff_lsf[2], g2_code);
L_tmp += Mpy_32_16(coeff[3], coeff_lsf[3], g_code);
L_tmp += Mpy_32_16(coeff[4], coeff_lsf[4], g_pit_cod);
if( L_tmp < L_dist_min ){
L_dist_min = L_tmp;
index1 = add(cand1,i);
index2 = add(cand2,j);
}
}
}
}
/* Read the quantized gain */
/*-----------------------------------------------------------------*
* *gain_pit = gbk1[indice1][0] + gbk2[indice2][0]; *
*-----------------------------------------------------------------*/
*gain_pit = add( gbk1[index1][0], gbk2[index2][0] ); /* Q14 */
/*-----------------------------------------------------------------*
* *gain_code = (gbk1[indice1][1]+gbk2[indice2][1]) * gcode0; *
*-----------------------------------------------------------------*/
L_gbk12 = (Word32)gbk1[index1][1] + (Word32)gbk2[index2][1]; /* Q13 */
tmp = extract_l( L_shr( L_gbk12,1 ) ); /* Q12 */
L_acc = L_mult(tmp, gcode0); /* Q[exp_gcode0+12+1] */
L_acc = L_shl(L_acc, add( negate(exp_gcode0),(-12-1+1+16) ));
*gain_cod = extract_h( L_acc ); /* Q1 */
/*----------------------------------------------*
* update table of past quantized energies *
*----------------------------------------------*/
Gain_update( past_qua_en, L_gbk12 );
return( add( map1[index1]*(Word16)NCODE2, map2[index2] ) );
}
void GPUDrawScanlineCodeGenerator::Generate()
{
push(esi);
push(edi);
Init();
align(16);
L("loop");
// GSVector4i test = m_test[7 + (steps & (steps >> 31))];
mov(edx, ecx);
sar(edx, 31);
and(edx, ecx);
shl(edx, 4);
movdqa(xmm7, ptr[edx + (size_t)&m_test[7]]);
// movdqu(xmm1, ptr[edi]);
movq(xmm1, qword[edi]);
movhps(xmm1, qword[edi + 8]);
// ecx = steps
// esi = tex (tme)
// edi = fb
// xmm1 = fd
// xmm2 = s
// xmm3 = t
// xmm4 = r
// xmm5 = g
// xmm6 = b
// xmm7 = test
TestMask();
SampleTexture();
// xmm1 = fd
// xmm3 = a
// xmm4 = r
// xmm5 = g
// xmm6 = b
// xmm7 = test
// xmm0, xmm2 = free
ColorTFX();
AlphaBlend();
Dither();
WriteFrame();
L("step");
// if(steps <= 0) break;
test(ecx, ecx);
jle("exit", T_NEAR);
Step();
jmp("loop", T_NEAR);
L("exit");
pop(edi);
pop(esi);
ret(8);
}
//------------------------------------------------------------------------------------------------------------------------
// Call stubs are used to call Java from C
//
// GR_I0 - call wrapper address : address
// GR_I1 - result : intptr_t*
// GR_I2 - result type : BasicType
// GR_I3 - method : methodOop
// GR_I4 - interpreter entry point : address
// GR_I5 - parameter block : intptr_t*
// GR_I6 - parameter count in words : int
// GR_I7 - thread : Thread*
//
address generate_call_stub(address& return_address) {
StubCodeMark mark(this, "StubRoutines", "call_stub");
const Register result = GR_I1;
const Register type = GR_I2;
const Register method = GR_I3;
const Register entry_ptr = GR_I4;
const Register parms = GR_I5;
const Register parm_count = GR_I6;
const Register thread = GR_I7;
const Register parm_size = GR31_SCRATCH;
const Register entry = GR30_SCRATCH;
const Register arg = GR29_SCRATCH;
const Register out_tos = GR49; // Equivalent of GR_Otos
const Register out_parms = GR50; // Equivalent of GR_Olocals (unused)
const BranchRegister entry_br = BR6_SCRATCH;
const PredicateRegister no_args = PR6_SCRATCH;
address start = __ emit_fd();
// Must allocate 8 output registers in case we go thru an i2c
// and the callee needs 8 input registers
__ alloc(GR_Lsave_PFS, 8, 9, 8, 0); // save AR_PFS
__ sxt4(parm_count, parm_count); // # of parms
__ mov(GR_Lsave_SP, SP); // save caller's SP
__ mov(GR_entry_frame_GR5, GR5_poll_page_addr);
__ mov(GR_entry_frame_GR6, GR6_caller_BSP);
__ mov(GR_entry_frame_GR7, GR7_reg_stack_limit);
// We can not tolerate an eager RSE cpu. Itanium-1 & 2 do not support
// this feature but we turn it off anyway
const Register RSC = GR2_SCRATCH;
__ mov(RSC, AR_RSC);
__ and3(RSC, -4, RSC); // Turn off two low bits
__ mov(AR_RSC, RSC); // enforced lazy mode
__ shladd(parm_size, parm_count, Interpreter::logStackElementSize(), GR0); // size of stack space for the parms
__ mov(GR_Lsave_RP, RP); // save return address
__ add(parm_size, parm_size, 15); // round up to multiple of 16 bytes. we use
// caller's 16-byte scratch area for params,
// so no need to add 16 to the current frame size.
__ mov(GR_Lsave_LC, AR_LC); // save AR_LC
__ add(out_parms, SP, Interpreter::stackElementSize()); // caller's SP+8 is 1st parm addr == target method locals addr
__ and3(parm_size, parm_size, -16);
__ cmp4(PR0, no_args, 0, parm_count, Assembler::less); // any parms?
__ mov(GR_entry_frame_GR4, GR4_thread); // save GR4_thread: it's a preserved register
__ sub(SP, SP, parm_size); // allocate the space for args + scratch
__ mov(entry_br, entry_ptr);
__ mov(GR27_method, method); // load method
__ mov(GR4_thread, thread); // load thread
if (TaggedStackInterpreter) __ shl(parm_count, parm_count, 1); // 2x tags
__ sub(parm_count, parm_count, 1); // cloop counts down to zero
// Initialize the register and memory stack limits for stack checking in compiled code
__ add(GR7_reg_stack_limit, thread_(register_stack_limit));
__ mov(GR6_caller_BSP, AR_BSP); // load register SP
__ movl(GR5_poll_page_addr, (intptr_t) os::get_polling_page() );
__ ld8(GR7_reg_stack_limit, GR7_reg_stack_limit); // load register stack limit
Label exit;
__ mov(AR_LC, parm_count);
__ mov(out_tos, out_parms); // out_tos = &out_parms[0]
__ br(no_args, exit, Assembler::dpnt);
// Reverse argument list and set up sender tos
Label copy_word;
__ bind(copy_word);
__ ld8(arg, parms, BytesPerWord); // load *parms++
__ st8(out_tos, arg, -BytesPerWord); // store *out_tos--
__ cloop(copy_word, Assembler::sptk, Assembler::few);
// Bias stack for tags.
if (TaggedStackInterpreter) __ st8(out_tos, GR0, -BytesPerWord);
__ bind(exit);
__ mov(GR_entry_frame_TOS, out_tos); // so entry_frame_argument_at can find TOS
// call interpreter frame manager
// Remember the senderSP so we interpreter can pop c2i arguments off of the stack
// when called via a c2i.
__ mov(GR28_sender_SP, SP);
__ call(entry_br);
return_address = __ pc();
// Store result depending on type. Everything that is not
// T_OBJECT, T_LONG, T_FLOAT, or T_DOUBLE is treated as T_INT.
const PredicateRegister is_obj = PR6_SCRATCH;
const PredicateRegister is_flt = PR7_SCRATCH;
const PredicateRegister is_dbl = PR8_SCRATCH;
const PredicateRegister is_lng = PR9_SCRATCH;
__ cmp4(is_obj, PR0, T_OBJECT, type, Assembler::equal);
__ cmp4(is_flt, PR0, T_FLOAT, type, Assembler::equal);
__ st4( result, GR_RET);
__ st8( is_obj, result, GR_RET);
__ stfs(is_flt, result, FR_RET);
__ cmp4(is_dbl, PR0, T_DOUBLE, type, Assembler::equal);
__ stfd(is_dbl, result, FR_RET);
__ cmp4(is_lng, PR0, T_LONG, type, Assembler::equal);
__ mov(RP, GR_Lsave_RP);
__ st8( is_lng, result, GR_RET);
__ mov(GR4_thread, GR_entry_frame_GR4);
__ mov(GR6_caller_BSP, GR_entry_frame_GR6);
__ mov(GR7_reg_stack_limit, GR_entry_frame_GR7);
__ mov(GR5_poll_page_addr, GR_entry_frame_GR5);
__ mov(AR_PFS, GR_Lsave_PFS);
__ mov(AR_LC, GR_Lsave_LC);
__ mov(SP, GR_Lsave_SP);
__ ret();
return start;
}
void Az_lsp(
Word16 a[], /* (i) Q12 : predictor coefficients */
Word16 lsp[], /* (o) Q15 : line spectral pairs */
Word16 old_lsp[] /* (i) : old lsp[] (in case not found 10 roots) */
)
{
Word16 i, j, nf, ip;
Word16 xlow, ylow, xhigh, yhigh, xmid, ymid, xint;
Word16 x, y, sign, exp;
Word16 *coef;
Word16 f1[M/2+1], f2[M/2+1];
Word32 t0, L_temp;
Flag ovf_coef;
Word16 (*pChebps)(Word16 x, Word16 f[], Word16 n);
/*-------------------------------------------------------------*
* find the sum and diff. pol. F1(z) and F2(z) *
* F1(z) <--- F1(z)/(1+z**-1) & F2(z) <--- F2(z)/(1-z**-1) *
* *
* f1[0] = 1.0; *
* f2[0] = 1.0; *
* *
* for (i = 0; i< NC; i++) *
* { *
* f1[i+1] = a[i+1] + a[M-i] - f1[i] ; *
* f2[i+1] = a[i+1] - a[M-i] + f2[i] ; *
* } *
*-------------------------------------------------------------*/
ovf_coef = 0;
pChebps = Chebps_11;
f1[0] = 2048; /* f1[0] = 1.0 is in Q11 */
f2[0] = 2048; /* f2[0] = 1.0 is in Q11 */
for (i = 0; i< NC; i++)
{
Overflow = 0;
t0 = L_mult(a[i+1], 16384); /* x = (a[i+1] + a[M-i]) >> 1 */
t0 = L_mac(t0, a[M-i], 16384); /* -> From Q12 to Q11 */
x = extract_h(t0);
if ( Overflow ) {
ovf_coef = 1; }
Overflow = 0;
f1[i+1] = sub(x, f1[i]); /* f1[i+1] = a[i+1] + a[M-i] - f1[i] */
if ( Overflow ) {
ovf_coef = 1; }
Overflow = 0;
t0 = L_mult(a[i+1], 16384); /* x = (a[i+1] - a[M-i]) >> 1 */
t0 = L_msu(t0, a[M-i], 16384); /* -> From Q12 to Q11 */
x = extract_h(t0);
if ( Overflow ) {
ovf_coef = 1; }
Overflow = 0;
f2[i+1] = add(x, f2[i]); /* f2[i+1] = a[i+1] - a[M-i] + f2[i] */
if ( Overflow ) {
ovf_coef = 1; }
}
if ( ovf_coef ) {
/*printf("===== OVF ovf_coef =====\n");*/
pChebps = Chebps_10;
f1[0] = 1024; /* f1[0] = 1.0 is in Q10 */
f2[0] = 1024; /* f2[0] = 1.0 is in Q10 */
for (i = 0; i< NC; i++)
{
t0 = L_mult(a[i+1], 8192); /* x = (a[i+1] + a[M-i]) >> 1 */
t0 = L_mac(t0, a[M-i], 8192); /* -> From Q11 to Q10 */
x = extract_h(t0);
f1[i+1] = sub(x, f1[i]); /* f1[i+1] = a[i+1] + a[M-i] - f1[i] */
t0 = L_mult(a[i+1], 8192); /* x = (a[i+1] - a[M-i]) >> 1 */
t0 = L_msu(t0, a[M-i], 8192); /* -> From Q11 to Q10 */
x = extract_h(t0);
f2[i+1] = add(x, f2[i]); /* f2[i+1] = a[i+1] - a[M-i] + f2[i] */
}
}
/*-------------------------------------------------------------*
* find the LSPs using the Chebichev pol. evaluation *
*-------------------------------------------------------------*/
nf=0; /* number of found frequencies */
ip=0; /* indicator for f1 or f2 */
coef = f1;
xlow = grid[0];
ylow = (*pChebps)(xlow, coef, NC);
j = 0;
while ( (nf < M) && (j < GRID_POINTS) )
{
j =add(j,1);
xhigh = xlow;
yhigh = ylow;
xlow = grid[j];
ylow = (*pChebps)(xlow,coef,NC);
L_temp = L_mult(ylow ,yhigh);
if ( L_temp <= (Word32)0)
{
/* divide 4 times the interval */
for (i = 0; i < 4; i++)
{
xmid = add( shr(xlow, 1) , shr(xhigh, 1)); /* xmid = (xlow + xhigh)/2 */
ymid = (*pChebps)(xmid,coef,NC);
L_temp = L_mult(ylow,ymid);
if ( L_temp <= (Word32)0)
{
yhigh = ymid;
xhigh = xmid;
}
else
{
ylow = ymid;
xlow = xmid;
}
}
/*-------------------------------------------------------------*
* Linear interpolation *
* xint = xlow - ylow*(xhigh-xlow)/(yhigh-ylow); *
*-------------------------------------------------------------*/
x = sub(xhigh, xlow);
y = sub(yhigh, ylow);
if(y == 0)
{
xint = xlow;
}
else
{
sign= y;
y = abs_s(y);
exp = norm_s(y);
y = shl(y, exp);
y = div_s( (Word16)16383, y);
t0 = L_mult(x, y);
t0 = L_shr(t0, sub(20, exp) );
y = extract_l(t0); /* y= (xhigh-xlow)/(yhigh-ylow) in Q11 */
if(sign < 0) y = negate(y);
t0 = L_mult(ylow, y); /* result in Q26 */
t0 = L_shr(t0, 11); /* result in Q15 */
xint = sub(xlow, extract_l(t0)); /* xint = xlow - ylow*y */
}
lsp[nf] = xint;
xlow = xint;
nf =add(nf,1);
if(ip == 0)
{
ip = 1;
coef = f2;
}
else
{
ip = 0;
coef = f1;
}
ylow = (*pChebps)(xlow,coef,NC);
}
}
/* Check if M roots found */
if( sub(nf, M) < 0)
{
for(i=0; i<M; i++)
{
lsp[i] = old_lsp[i];
}
/* printf("\n !!Not 10 roots found in Az_lsp()!!!\n"); */
}
return;
}
