summaryrefslogtreecommitdiff
path: root/libs/libopus/celt/celt_lpc.c
diff options
context:
space:
mode:
Diffstat (limited to 'libs/libopus/celt/celt_lpc.c')
-rw-r--r--libs/libopus/celt/celt_lpc.c124
1 files changed, 77 insertions, 47 deletions
diff --git a/libs/libopus/celt/celt_lpc.c b/libs/libopus/celt/celt_lpc.c
index b410a21c5..242e6df55 100644
--- a/libs/libopus/celt/celt_lpc.c
+++ b/libs/libopus/celt/celt_lpc.c
@@ -50,17 +50,21 @@ int p
#endif
OPUS_CLEAR(lpc, p);
+#ifdef FIXED_POINT
if (ac[0] != 0)
+#else
+ if (ac[0] > 1e-10f)
+#endif
{
for (i = 0; i < p; i++) {
/* Sum up this iteration's reflection coefficient */
opus_val32 rr = 0;
for (j = 0; j < i; j++)
rr += MULT32_32_Q31(lpc[j],ac[i - j]);
- rr += SHR32(ac[i + 1],3);
- r = -frac_div32(SHL32(rr,3), error);
+ rr += SHR32(ac[i + 1],6);
+ r = -frac_div32(SHL32(rr,6), error);
/* Update LPC coefficients and total error */
- lpc[i] = SHR32(r,3);
+ lpc[i] = SHR32(r,6);
for (j = 0; j < (i+1)>>1; j++)
{
opus_val32 tmp1, tmp2;
@@ -73,74 +77,100 @@ int p
error = error - MULT32_32_Q31(MULT32_32_Q31(r,r),error);
/* Bail out once we get 30 dB gain */
#ifdef FIXED_POINT
- if (error<SHR32(ac[0],10))
+ if (error<=SHR32(ac[0],10))
break;
#else
- if (error<.001f*ac[0])
+ if (error<=.001f*ac[0])
break;
#endif
}
}
#ifdef FIXED_POINT
- for (i=0;i<p;i++)
- _lpc[i] = ROUND16(lpc[i],16);
+ {
+ /* Convert the int32 lpcs to int16 and ensure there are no wrap-arounds.
+ This reuses the logic in silk_LPC_fit() and silk_bwexpander_32(). Any bug
+ fixes should also be applied there. */
+ int iter, idx = 0;
+ opus_val32 maxabs, absval, chirp_Q16, chirp_minus_one_Q16;
+
+ for (iter = 0; iter < 10; iter++) {
+ maxabs = 0;
+ for (i = 0; i < p; i++) {
+ absval = ABS32(lpc[i]);
+ if (absval > maxabs) {
+ maxabs = absval;
+ idx = i;
+ }
+ }
+ maxabs = PSHR32(maxabs, 13); /* Q25->Q12 */
+
+ if (maxabs > 32767) {
+ maxabs = MIN32(maxabs, 163838);
+ chirp_Q16 = QCONST32(0.999, 16) - DIV32(SHL32(maxabs - 32767, 14),
+ SHR32(MULT32_32_32(maxabs, idx + 1), 2));
+ chirp_minus_one_Q16 = chirp_Q16 - 65536;
+
+ /* Apply bandwidth expansion. */
+ for (i = 0; i < p - 1; i++) {
+ lpc[i] = MULT32_32_Q16(chirp_Q16, lpc[i]);
+ chirp_Q16 += PSHR32(MULT32_32_32(chirp_Q16, chirp_minus_one_Q16), 16);
+ }
+ lpc[p - 1] = MULT32_32_Q16(chirp_Q16, lpc[p - 1]);
+ } else {
+ break;
+ }
+ }
+
+ if (iter == 10) {
+ /* If the coeffs still do not fit into the 16 bit range after 10 iterations,
+ fall back to the A(z)=1 filter. */
+ OPUS_CLEAR(lpc, p);
+ _lpc[0] = 4096; /* Q12 */
+ } else {
+ for (i = 0; i < p; i++) {
+ _lpc[i] = EXTRACT16(PSHR32(lpc[i], 13)); /* Q25->Q12 */
+ }
+ }
+ }
#endif
}
void celt_fir_c(
- const opus_val16 *_x,
+ const opus_val16 *x,
const opus_val16 *num,
- opus_val16 *_y,
+ opus_val16 *y,
int N,
int ord,
- opus_val16 *mem,
int arch)
{
int i,j;
VARDECL(opus_val16, rnum);
- VARDECL(opus_val16, x);
SAVE_STACK;
-
+ celt_assert(x != y);
ALLOC(rnum, ord, opus_val16);
- ALLOC(x, N+ord, opus_val16);
for(i=0;i<ord;i++)
rnum[i] = num[ord-i-1];
- for(i=0;i<ord;i++)
- x[i] = mem[ord-i-1];
- for (i=0;i<N;i++)
- x[i+ord]=_x[i];
- for(i=0;i<ord;i++)
- mem[i] = _x[N-i-1];
-#ifdef SMALL_FOOTPRINT
- (void)arch;
- for (i=0;i<N;i++)
- {
- opus_val32 sum = SHL32(EXTEND32(_x[i]), SIG_SHIFT);
- for (j=0;j<ord;j++)
- {
- sum = MAC16_16(sum,rnum[j],x[i+j]);
- }
- _y[i] = SATURATE16(PSHR32(sum, SIG_SHIFT));
- }
-#else
for (i=0;i<N-3;i+=4)
{
- opus_val32 sum[4]={0,0,0,0};
- xcorr_kernel(rnum, x+i, sum, ord, arch);
- _y[i ] = SATURATE16(ADD32(EXTEND32(_x[i ]), PSHR32(sum[0], SIG_SHIFT)));
- _y[i+1] = SATURATE16(ADD32(EXTEND32(_x[i+1]), PSHR32(sum[1], SIG_SHIFT)));
- _y[i+2] = SATURATE16(ADD32(EXTEND32(_x[i+2]), PSHR32(sum[2], SIG_SHIFT)));
- _y[i+3] = SATURATE16(ADD32(EXTEND32(_x[i+3]), PSHR32(sum[3], SIG_SHIFT)));
+ opus_val32 sum[4];
+ sum[0] = SHL32(EXTEND32(x[i ]), SIG_SHIFT);
+ sum[1] = SHL32(EXTEND32(x[i+1]), SIG_SHIFT);
+ sum[2] = SHL32(EXTEND32(x[i+2]), SIG_SHIFT);
+ sum[3] = SHL32(EXTEND32(x[i+3]), SIG_SHIFT);
+ xcorr_kernel(rnum, x+i-ord, sum, ord, arch);
+ y[i ] = ROUND16(sum[0], SIG_SHIFT);
+ y[i+1] = ROUND16(sum[1], SIG_SHIFT);
+ y[i+2] = ROUND16(sum[2], SIG_SHIFT);
+ y[i+3] = ROUND16(sum[3], SIG_SHIFT);
}
for (;i<N;i++)
{
- opus_val32 sum = 0;
+ opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT);
for (j=0;j<ord;j++)
- sum = MAC16_16(sum,rnum[j],x[i+j]);
- _y[i] = SATURATE16(ADD32(EXTEND32(_x[i]), PSHR32(sum, SIG_SHIFT)));
+ sum = MAC16_16(sum,rnum[j],x[i+j-ord]);
+ y[i] = ROUND16(sum, SIG_SHIFT);
}
-#endif
RESTORE_STACK;
}
@@ -166,7 +196,7 @@ void celt_iir(const opus_val32 *_x,
{
mem[j]=mem[j-1];
}
- mem[0] = ROUND16(sum,SIG_SHIFT);
+ mem[0] = SROUND16(sum, SIG_SHIFT);
_y[i] = sum;
}
#else
@@ -195,20 +225,20 @@ void celt_iir(const opus_val32 *_x,
xcorr_kernel(rden, y+i, sum, ord, arch);
/* Patch up the result to compensate for the fact that this is an IIR */
- y[i+ord ] = -ROUND16(sum[0],SIG_SHIFT);
+ y[i+ord ] = -SROUND16(sum[0],SIG_SHIFT);
_y[i ] = sum[0];
sum[1] = MAC16_16(sum[1], y[i+ord ], den[0]);
- y[i+ord+1] = -ROUND16(sum[1],SIG_SHIFT);
+ y[i+ord+1] = -SROUND16(sum[1],SIG_SHIFT);
_y[i+1] = sum[1];
sum[2] = MAC16_16(sum[2], y[i+ord+1], den[0]);
sum[2] = MAC16_16(sum[2], y[i+ord ], den[1]);
- y[i+ord+2] = -ROUND16(sum[2],SIG_SHIFT);
+ y[i+ord+2] = -SROUND16(sum[2],SIG_SHIFT);
_y[i+2] = sum[2];
sum[3] = MAC16_16(sum[3], y[i+ord+2], den[0]);
sum[3] = MAC16_16(sum[3], y[i+ord+1], den[1]);
sum[3] = MAC16_16(sum[3], y[i+ord ], den[2]);
- y[i+ord+3] = -ROUND16(sum[3],SIG_SHIFT);
+ y[i+ord+3] = -SROUND16(sum[3],SIG_SHIFT);
_y[i+3] = sum[3];
}
for (;i<N;i++)
@@ -216,7 +246,7 @@ void celt_iir(const opus_val32 *_x,
opus_val32 sum = _x[i];
for (j=0;j<ord;j++)
sum -= MULT16_16(rden[j],y[i+j]);
- y[i+ord] = ROUND16(sum,SIG_SHIFT);
+ y[i+ord] = SROUND16(sum,SIG_SHIFT);
_y[i] = sum;
}
for(i=0;i<ord;i++)
generated by cgit v1.2.3 (git 2.26.2) at 2024-11-09 17:12:14 +0000