- Bring speex codec up to date with their SVN trunk
- Speex codec should work in FIXED_POINT mode when PJ_HAS_FLOATING_POINT is set to zero.
- ulaw2linear will return zero if zero is given (this would make the VAD works better, and it also fixed click noise when call is established/hangup).
git-svn-id: https://svn.pjsip.org/repos/pjproject/trunk@628 74dad513-b988-da41-8d7b-12977e46ad98
diff --git a/pjmedia/src/pjmedia-codec/speex/_kiss_fft_guts.h b/pjmedia/src/pjmedia-codec/speex/_kiss_fft_guts.h
index abbd8b1..72acee1 100644
--- a/pjmedia/src/pjmedia-codec/speex/_kiss_fft_guts.h
+++ b/pjmedia/src/pjmedia-codec/speex/_kiss_fft_guts.h
@@ -20,7 +20,6 @@
and defines
typedef struct { kiss_fft_scalar r; kiss_fft_scalar i; }kiss_fft_cpx; */
#include "kiss_fft.h"
-#include <limits.h>
#define MAXFACTORS 32
/* e.g. an fft of length 128 has 4 factors
@@ -45,8 +44,9 @@
C_ADDTO( res , a) : res += a
* */
#ifdef FIXED_POINT
+#include "misc.h"
# define FRACBITS 15
-# define SAMPPROD int32_t
+# define SAMPPROD spx_int32_t
#define SAMP_MAX 32767
#define SAMP_MIN -SAMP_MAX
diff --git a/pjmedia/src/pjmedia-codec/speex/arch.h b/pjmedia/src/pjmedia-codec/speex/arch.h
index 5206619..0500437 100644
--- a/pjmedia/src/pjmedia-codec/speex/arch.h
+++ b/pjmedia/src/pjmedia-codec/speex/arch.h
@@ -41,18 +41,12 @@
#define ABS16(x) ((x) < 0 ? (-(x)) : (x)) /**< Absolute 16-bit value. */
#define MAX16(a,b) ((a) > (b) ? (a) : (b)) /**< Maximum 16-bit value. */
#define ABS32(x) ((x) < 0 ? (-(x)) : (x)) /**< Absolute 32-bit value. */
+#define MAX32(a,b) ((a) > (b) ? (a) : (b)) /**< Maximum 32-bit value. */
#ifdef FIXED_POINT
typedef spx_int16_t spx_word16_t;
typedef spx_int32_t spx_word32_t;
-#ifdef _MSC_VER
-typedef __int64 spx_word64_t;
-#elif defined NO_LONGLONG
-typedef double spx_word64_t;
-#else
-typedef long long spx_word64_t;
-#endif
typedef spx_word32_t spx_mem_t;
typedef spx_word16_t spx_coef_t;
typedef spx_word16_t spx_lsp_t;
@@ -103,7 +97,6 @@
typedef float spx_sig_t;
typedef float spx_word16_t;
typedef float spx_word32_t;
-typedef float spx_word64_t;
#define Q15ONE 1.0f
#define LPC_SCALING 1.f
@@ -146,7 +139,6 @@
#define SUB16(a,b) ((a)-(b))
#define ADD32(a,b) ((a)+(b))
#define SUB32(a,b) ((a)-(b))
-#define ADD64(a,b) ((a)+(b))
#define MULT16_16_16(a,b) ((a)*(b))
#define MULT16_16(a,b) ((spx_word32_t)(a)*(spx_word32_t)(b))
#define MAC16_16(c,a,b) ((c)+(spx_word32_t)(a)*(spx_word32_t)(b))
@@ -161,20 +153,25 @@
#define MAC16_16_Q11(c,a,b) ((c)+(a)*(b))
#define MAC16_16_Q13(c,a,b) ((c)+(a)*(b))
+#define MAC16_16_P13(c,a,b) ((c)+(a)*(b))
#define MULT16_16_Q11_32(a,b) ((a)*(b))
#define MULT16_16_Q13(a,b) ((a)*(b))
#define MULT16_16_Q14(a,b) ((a)*(b))
#define MULT16_16_Q15(a,b) ((a)*(b))
#define MULT16_16_P15(a,b) ((a)*(b))
+#define MULT16_16_P13(a,b) ((a)*(b))
+#define MULT16_16_P14(a,b) ((a)*(b))
-#define DIV32_16(a,b) ((a)/(b))
-#define DIV32(a,b) ((a)/(b))
+#define DIV32_16(a,b) (((spx_word32_t)(a))/(spx_word16_t)(b))
+#define PDIV32_16(a,b) (((spx_word32_t)(a))/(spx_word16_t)(b))
+#define DIV32(a,b) (((spx_word32_t)(a))/(spx_word32_t)(b))
+#define PDIV32(a,b) (((spx_word32_t)(a))/(spx_word32_t)(b))
#endif
-#ifdef CONFIG_TI_C55X
+#if defined (CONFIG_TI_C54X) || defined (CONFIG_TI_C55X)
/* 2 on TI C5x DSP */
#define BYTES_PER_CHAR 2
diff --git a/pjmedia/src/pjmedia-codec/speex/bits.c b/pjmedia/src/pjmedia-codec/speex/bits.c
index fae7a9e..376e804 100644
--- a/pjmedia/src/pjmedia-codec/speex/bits.c
+++ b/pjmedia/src/pjmedia-codec/speex/bits.c
@@ -93,28 +93,36 @@
void speex_bits_read_from(SpeexBits *bits, char *chars, int len)
{
int i;
- if (len > bits->buf_size)
+ int nchars = len / BYTES_PER_CHAR;
+ if (nchars > bits->buf_size)
{
speex_warning_int("Packet is larger than allocated buffer: ", len);
if (bits->owner)
{
- char *tmp = (char*)speex_realloc(bits->chars, len);
+ char *tmp = (char*)speex_realloc(bits->chars, nchars);
if (tmp)
{
- bits->buf_size=len;
+ bits->buf_size=nchars;
bits->chars=tmp;
} else {
- len=bits->buf_size;
+ nchars=bits->buf_size;
speex_warning("Could not resize input buffer: truncating input");
}
} else {
speex_warning("Do not own input buffer: truncating input");
- len=bits->buf_size;
+ nchars=bits->buf_size;
}
}
- for (i=0;i<len;i++)
- bits->chars[i]=chars[i];
- bits->nbBits=len<<3;
+#if (BYTES_PER_CHAR==2)
+/* Swap bytes to proper endian order (could be done externally) */
+#define HTOLS(A) ((((A) >> 8)&0xff)|(((A) & 0xff)<<8))
+#else
+#define HTOLS(A) (A)
+#endif
+ for (i=0;i<nchars;i++)
+ bits->chars[i]=HTOLS(chars[i]);
+
+ bits->nbBits=nchars<<LOG2_BITS_PER_CHAR;
bits->charPtr=0;
bits->bitPtr=0;
bits->overflow=0;
@@ -161,7 +169,7 @@
speex_bits_flush(bits);
pos=bits->nbBits>>LOG2_BITS_PER_CHAR;
for (i=0;i<nchars;i++)
- bits->chars[pos+i]=chars[i];
+ bits->chars[pos+i]=HTOLS(chars[i]);
bits->nbBits+=nchars<<LOG2_BITS_PER_CHAR;
}
@@ -182,11 +190,7 @@
if (max_nchars > ((bits->nbBits+BITS_PER_CHAR-1)>>LOG2_BITS_PER_CHAR))
max_nchars = ((bits->nbBits+BITS_PER_CHAR-1)>>LOG2_BITS_PER_CHAR);
-#if BYTES_PER_CHAR==1
-#define HTOLS(A) (A)
-#else
-#define HTOLS(A) ((((A) >> 8)&0xff)|(((A) & 0xff)<<8))
-#endif
+
for (i=0;i<max_nchars;i++)
chars[i]=HTOLS(bits->chars[i]);
return max_nchars*BYTES_PER_CHAR;
@@ -199,8 +203,8 @@
if (max_nchars > ((bits->nbBits)>>LOG2_BITS_PER_CHAR))
max_nchars = ((bits->nbBits)>>LOG2_BITS_PER_CHAR);
for (i=0;i<max_nchars;i++)
- chars[i]=bits->chars[i];
-
+ chars[i]=HTOLS(bits->chars[i]);
+
if (bits->bitPtr>0)
bits->chars[0]=bits->chars[max_nchars];
else
diff --git a/pjmedia/src/pjmedia-codec/speex/cb_search.c b/pjmedia/src/pjmedia-codec/speex/cb_search.c
index 234dc2e..b4a223b 100644
--- a/pjmedia/src/pjmedia-codec/speex/cb_search.c
+++ b/pjmedia/src/pjmedia-codec/speex/cb_search.c
@@ -70,7 +70,7 @@
for (k=0;k<=j;k++)
resj = MAC16_16(resj,shape[k],r[j-k]);
#ifdef FIXED_POINT
- res16 = EXTRACT16(SHR32(resj, 11));
+ res16 = EXTRACT16(SHR32(resj, 13));
#else
res16 = 0.03125f*resj;
#endif
@@ -88,16 +88,15 @@
static inline void target_update(spx_word16_t *t, spx_word16_t g, spx_word16_t *r, int len)
{
int n;
- int q=0;
- for (n=0;n<len;n++,q++)
- t[n] = SUB32(t[n],MULT16_16_Q11_32(g,r[q]));
+ for (n=0;n<len;n++)
+ t[n] = SUB16(t[n],PSHR32(MULT16_16(g,r[n]),13));
}
#endif
static void split_cb_search_shape_sign_N1(
-spx_sig_t target[], /* target vector */
+spx_word16_t target[], /* target vector */
spx_coef_t ak[], /* LPCs for this subframe */
spx_coef_t awk1[], /* Weighted LPCs for this subframe */
spx_coef_t awk2[], /* Weighted LPCs for this subframe */
@@ -113,9 +112,6 @@
)
{
int i,j,m,q;
-#ifndef FIXED_POINT
- int n;
-#endif
VARDECL(spx_word16_t *resp);
#ifdef _USE_SSE
VARDECL(__m128 *resp2);
@@ -158,7 +154,7 @@
/* FIXME: make that adaptive? */
for (i=0;i<nsf;i++)
- t[i]=EXTRACT16(PSHR32(target[i],6));
+ t[i]=target[i];
compute_weighted_codebook(shape_cb, r, resp, resp2, E, shape_cb_size, subvect_size, stack);
@@ -222,13 +218,10 @@
q=subvect_size-m;
#ifdef FIXED_POINT
g=sign*shape_cb[rind*subvect_size+m];
- target_update(t+subvect_size*(i+1), g, r+q, nsf-subvect_size*(i+1));
#else
g=sign*0.03125*shape_cb[rind*subvect_size+m];
- /*FIXME: I think that one too can be replaced by target_update */
- for (n=subvect_size*(i+1);n<nsf;n++,q++)
- t[n] = SUB32(t[n],g*r[q]);
#endif
+ target_update(t+subvect_size*(i+1), g, r+q, nsf-subvect_size*(i+1));
}
}
@@ -244,14 +237,14 @@
ALLOC(r2, nsf, spx_sig_t);
syn_percep_zero(e, ak, awk1, awk2, r2, nsf,p, stack);
for (j=0;j<nsf;j++)
- target[j]=SUB32(target[j],r2[j]);
+ target[j]=SUB16(target[j],EXTRACT16(PSHR32(r2[j],8)));
}
}
void split_cb_search_shape_sign(
-spx_sig_t target[], /* target vector */
+spx_word16_t target[], /* target vector */
spx_coef_t ak[], /* LPCs for this subframe */
spx_coef_t awk1[], /* Weighted LPCs for this subframe */
spx_coef_t awk2[], /* Weighted LPCs for this subframe */
@@ -356,7 +349,7 @@
/* FIXME: make that adaptive? */
for (i=0;i<nsf;i++)
- t[i]=EXTRACT16(PSHR32(target[i],6));
+ t[i]=target[i];
for (j=0;j<N;j++)
speex_move(&ot[j][0], t, nsf*sizeof(spx_word16_t));
@@ -444,13 +437,10 @@
q=subvect_size-m;
#ifdef FIXED_POINT
g=sign*shape_cb[rind*subvect_size+m];
- target_update(nt[j]+subvect_size*(i+1), g, r+q, nsf-subvect_size*(i+1));
#else
g=sign*0.03125*shape_cb[rind*subvect_size+m];
- /*FIXME: I think that one too can be replaced by target_update */
- for (n=subvect_size*(i+1);n<nsf;n++,q++)
- nt[j][n] = SUB32(nt[j][n],g*r[q]);
#endif
+ target_update(nt[j]+subvect_size*(i+1), g, r+q, nsf-subvect_size*(i+1));
}
for (q=0;q<nb_subvect;q++)
@@ -514,7 +504,7 @@
{
syn_percep_zero(e, ak, awk1, awk2, r2, nsf,p, stack);
for (j=0;j<nsf;j++)
- target[j]=SUB32(target[j],r2[j]);
+ target[j]=SUB16(target[j],EXTRACT16(PSHR32(r2[j],8)));
}
}
@@ -577,7 +567,7 @@
}
void noise_codebook_quant(
-spx_sig_t target[], /* target vector */
+spx_word16_t target[], /* target vector */
spx_coef_t ak[], /* LPCs for this subframe */
spx_coef_t awk1[], /* Weighted LPCs for this subframe */
spx_coef_t awk2[], /* Weighted LPCs for this subframe */
@@ -595,13 +585,14 @@
int i;
VARDECL(spx_sig_t *tmp);
ALLOC(tmp, nsf, spx_sig_t);
- residue_percep_zero(target, ak, awk1, awk2, tmp, nsf, p, stack);
+ for (i=0;i<nsf;i++)
+ tmp[i]=PSHR32(EXTEND32(target[i]),SIG_SHIFT);
+ residue_percep_zero(tmp, ak, awk1, awk2, tmp, nsf, p, stack);
for (i=0;i<nsf;i++)
exc[i]+=tmp[i];
for (i=0;i<nsf;i++)
target[i]=0;
-
}
@@ -613,5 +604,9 @@
char *stack
)
{
- speex_rand_vec(1, exc, nsf);
+ int i;
+ /* FIXME: This is bad, but I don't think the function ever gets called anyway */
+ spx_int32_t seed = 0;
+ for (i=0;i<nsf;i++)
+ exc[i]=SHL32(EXTEND32(speex_rand(1, &seed)),SIG_SHIFT);
}
diff --git a/pjmedia/src/pjmedia-codec/speex/cb_search.h b/pjmedia/src/pjmedia-codec/speex/cb_search.h
index 38ac077..ea8816d 100644
--- a/pjmedia/src/pjmedia-codec/speex/cb_search.h
+++ b/pjmedia/src/pjmedia-codec/speex/cb_search.h
@@ -49,7 +49,7 @@
void split_cb_search_shape_sign(
-spx_sig_t target[], /* target vector */
+spx_word16_t target[], /* target vector */
spx_coef_t ak[], /* LPCs for this subframe */
spx_coef_t awk1[], /* Weighted LPCs for this subframe */
spx_coef_t awk2[], /* Weighted LPCs for this subframe */
@@ -74,7 +74,7 @@
void noise_codebook_quant(
-spx_sig_t target[], /* target vector */
+spx_word16_t target[], /* target vector */
spx_coef_t ak[], /* LPCs for this subframe */
spx_coef_t awk1[], /* Weighted LPCs for this subframe */
spx_coef_t awk2[], /* Weighted LPCs for this subframe */
diff --git a/pjmedia/src/pjmedia-codec/speex/cb_search_bfin.h b/pjmedia/src/pjmedia-codec/speex/cb_search_bfin.h
index 7d476a3..52cc4b3 100644
--- a/pjmedia/src/pjmedia-codec/speex/cb_search_bfin.h
+++ b/pjmedia/src/pjmedia-codec/speex/cb_search_bfin.h
@@ -61,7 +61,7 @@
"A0 += R0.L*R1.L (IS) || R0 = B[P4++] (X) || R1.L = W[I1--];\n\t"
"LOOP_END inner%=;\n\t"
"R0 = A0;\n\t"
- "R0 >>>= 11;\n\t"
+ "R0 >>>= 13;\n\t"
"A1 += R0.L*R0.L (IS);\n\t"
"W[P3++] = R0;\n\t"
"P0 += 1;\n\t"
@@ -72,7 +72,8 @@
"[P4] = R1;\n\t"
:
: "m" (subvect_size), "m" (shape_cb), "m" (r), "m" (resp), "m" (E)
- : "A0", "P0", "P1", "P2", "P3", "P4", "R0", "R1", "R2", "I0", "I1", "L0", "L1", "A0", "A1", "memory"
+ : "A0", "P0", "P1", "P2", "P3", "P4", "R0", "R1", "R2", "I0", "I1", "L0",
+ "L1", "A0", "A1", "memory", "LC0", "LC1"
);
shape_cb += subvect_size;
resp += subvect_size;
@@ -83,23 +84,26 @@
#define OVERRIDE_TARGET_UPDATE
static inline void target_update(spx_word16_t *t, spx_word16_t g, spx_word16_t *r, int len)
{
+ if (!len)
+ return;
__asm__ __volatile__
(
"I0 = %0;\n\t"
"I1 = %1;\n\t"
"L0 = 0;\n\t"
"L1 = 0;\n\t"
+ "R2 = 4096;\n\t"
"LOOP tupdate%= LC0 = %3;\n\t"
"LOOP_BEGIN tupdate%=;\n\t"
"R0.L = W[I0] || R1.L = W[I1++];\n\t"
"R1 = (A1 = R1.L*%2.L) (IS);\n\t"
- "R1 >>>= 11;\n\t"
+ "R1 = R1 + R2;\n\t"
+ "R1 >>>= 13;\n\t"
"R0.L = R0.L - R1.L;\n\t"
"W[I0++] = R0.L;\n\t"
"LOOP_END tupdate%=;\n\t"
:
: "a" (t), "a" (r), "d" (g), "a" (len)
- : "R0", "R1", "A1", "I0", "I1", "L0", "L1"
+ : "R0", "R1", "R2", "A1", "I0", "I1", "L0", "L1"
);
}
-
diff --git a/pjmedia/src/pjmedia-codec/speex/config.h b/pjmedia/src/pjmedia-codec/speex/config.h
index ef3a4a0..6ab2235 100644
--- a/pjmedia/src/pjmedia-codec/speex/config.h
+++ b/pjmedia/src/pjmedia-codec/speex/config.h
@@ -1,15 +1,15 @@
-#include <pj/config.h>
+/* Check if we need to use the fixed point version */
+#if !defined(PJ_HAS_FLOATING_POINT) || PJ_HAS_FLOATING_POINT==0
+# define FIXED_POINT
+#endif
+
#define inline __inline
#define restrict
#include "misc.h"
-#if !defined(PJ_HAS_FLOATING_POINT) || PJ_HAS_FLOATING_POINT==0
-# define FIXED_POINT
-#endif
-
#ifdef _MSC_VER
# pragma warning(disable: 4100) // unreferenced formal parameter
# pragma warning(disable: 4101) // unreferenced local variable
diff --git a/pjmedia/src/pjmedia-codec/speex/fftwrap.c b/pjmedia/src/pjmedia-codec/speex/fftwrap.c
index de09356..79a1de3 100644
--- a/pjmedia/src/pjmedia-codec/speex/fftwrap.c
+++ b/pjmedia/src/pjmedia-codec/speex/fftwrap.c
@@ -42,6 +42,7 @@
#include "misc.h"
+#define MAX_FFT_SIZE 2048
#ifdef FIXED_POINT
static int maximize_range(spx_word16_t *in, spx_word16_t *out, spx_word16_t bound, int len)
@@ -225,9 +226,8 @@
#endif
-int fixed_point = 1;
#ifdef FIXED_POINT
-#include "smallft.h"
+/*#include "smallft.h"*/
void spx_fft_float(void *table, float *in, float *out)
@@ -239,13 +239,19 @@
int N = ((struct kiss_config *)table)->N;
#else
#endif
+#ifdef VAR_ARRAYS
spx_word16_t _in[N];
spx_word16_t _out[N];
+#else
+ spx_word16_t _in[MAX_FFT_SIZE];
+ spx_word16_t _out[MAX_FFT_SIZE];
+#endif
for (i=0;i<N;i++)
_in[i] = (int)floor(.5+in[i]);
spx_fft(table, _in, _out);
for (i=0;i<N;i++)
out[i] = _out[i];
+#if 0
if (!fixed_point)
{
float scale;
@@ -257,6 +263,7 @@
spx_drft_forward(&t, out);
spx_drft_clear(&t);
}
+#endif
}
void spx_ifft_float(void *table, float *in, float *out)
@@ -268,13 +275,19 @@
int N = ((struct kiss_config *)table)->N;
#else
#endif
+#ifdef VAR_ARRAYS
spx_word16_t _in[N];
spx_word16_t _out[N];
+#else
+ spx_word16_t _in[MAX_FFT_SIZE];
+ spx_word16_t _out[MAX_FFT_SIZE];
+#endif
for (i=0;i<N;i++)
_in[i] = (int)floor(.5+in[i]);
spx_ifft(table, _in, _out);
for (i=0;i<N;i++)
out[i] = _out[i];
+#if 0
if (!fixed_point)
{
int i;
@@ -285,6 +298,7 @@
spx_drft_backward(&t, out);
spx_drft_clear(&t);
}
+#endif
}
#else
diff --git a/pjmedia/src/pjmedia-codec/speex/filters.c b/pjmedia/src/pjmedia-codec/speex/filters.c
index abc8d9c..73cb391 100644
--- a/pjmedia/src/pjmedia-codec/speex/filters.c
+++ b/pjmedia/src/pjmedia-codec/speex/filters.c
@@ -75,25 +75,35 @@
}
}
-void signal_div(const spx_sig_t *x, spx_sig_t *y, spx_word32_t scale, int len)
+void signal_div(const spx_word16_t *x, spx_word16_t *y, spx_word32_t scale, int len)
{
int i;
if (scale > SHL32(EXTEND32(SIG_SCALING), 8))
{
spx_word16_t scale_1;
scale = PSHR32(scale, SIG_SHIFT);
- scale_1 = EXTRACT16(DIV32_16(SHL32(EXTEND32(SIG_SCALING),7),scale));
+ scale_1 = EXTRACT16(PDIV32_16(SHL32(EXTEND32(SIG_SCALING),7),scale));
for (i=0;i<len;i++)
{
- y[i] = SHR32(MULT16_16(scale_1, EXTRACT16(SHR32(x[i],SIG_SHIFT))),7);
+ y[i] = MULT16_16_P15(scale_1, x[i]);
}
- } else {
+ } else if (scale > SHR32(EXTEND32(SIG_SCALING), 2)) {
spx_word16_t scale_1;
scale = PSHR32(scale, SIG_SHIFT-5);
scale_1 = DIV32_16(SHL32(EXTEND32(SIG_SCALING),3),scale);
for (i=0;i<len;i++)
{
- y[i] = MULT16_16(scale_1, EXTRACT16(SHR32(x[i],SIG_SHIFT-2)));
+ y[i] = PSHR32(MULT16_16(scale_1, SHL16(x[i],2)),8);
+ }
+ } else {
+ spx_word16_t scale_1;
+ scale = PSHR32(scale, SIG_SHIFT-7);
+ if (scale < 5)
+ scale = 5;
+ scale_1 = DIV32_16(SHL32(EXTEND32(SIG_SCALING),3),scale);
+ for (i=0;i<len;i++)
+ {
+ y[i] = PSHR32(MULT16_16(scale_1, SHL16(x[i],2)),6);
}
}
}
@@ -160,9 +170,56 @@
sum = ADD32(sum,SHR32(sum2,6));
}
- return EXTRACT16(SHR32(SHL32(EXTEND32(spx_sqrt(1+DIV32(sum,len))),(sig_shift+3)),SIG_SHIFT));
+ return EXTRACT16(PSHR32(SHL32(EXTEND32(spx_sqrt(DIV32(sum,len))),(sig_shift+3)),SIG_SHIFT));
}
+spx_word16_t compute_rms16(const spx_word16_t *x, int len)
+{
+ int i;
+ spx_word16_t max_val=10;
+
+ for (i=0;i<len;i++)
+ {
+ spx_sig_t tmp = x[i];
+ if (tmp<0)
+ tmp = -tmp;
+ if (tmp > max_val)
+ max_val = tmp;
+ }
+ if (max_val>16383)
+ {
+ spx_word32_t sum=0;
+ for (i=0;i<len;i+=4)
+ {
+ spx_word32_t sum2=0;
+ sum2 = MAC16_16(sum2,PSHR16(x[i],1),PSHR16(x[i],1));
+ sum2 = MAC16_16(sum2,PSHR16(x[i+1],1),PSHR16(x[i+1],1));
+ sum2 = MAC16_16(sum2,PSHR16(x[i+2],1),PSHR16(x[i+2],1));
+ sum2 = MAC16_16(sum2,PSHR16(x[i+3],1),PSHR16(x[i+3],1));
+ sum = ADD32(sum,SHR32(sum2,6));
+ }
+ return SHL16(spx_sqrt(DIV32(sum,len)),4);
+ } else {
+ spx_word32_t sum=0;
+ int sig_shift=0;
+ if (max_val < 8192)
+ sig_shift=1;
+ if (max_val < 4096)
+ sig_shift=2;
+ if (max_val < 2048)
+ sig_shift=3;
+ for (i=0;i<len;i+=4)
+ {
+ spx_word32_t sum2=0;
+ sum2 = MAC16_16(sum2,SHL16(x[i],sig_shift),SHL16(x[i],sig_shift));
+ sum2 = MAC16_16(sum2,SHL16(x[i+1],sig_shift),SHL16(x[i+1],sig_shift));
+ sum2 = MAC16_16(sum2,SHL16(x[i+2],sig_shift),SHL16(x[i+2],sig_shift));
+ sum2 = MAC16_16(sum2,SHL16(x[i+3],sig_shift),SHL16(x[i+3],sig_shift));
+ sum = ADD32(sum,SHR32(sum2,6));
+ }
+ return SHL16(spx_sqrt(DIV32(sum,len)),3-sig_shift);
+ }
+}
#ifndef OVERRIDE_NORMALIZE16
int normalize16(const spx_sig_t *x, spx_word16_t *y, spx_sig_t max_scale, int len)
@@ -206,6 +263,10 @@
}
return sqrt(.1+sum/len);
}
+spx_word16_t compute_rms16(const spx_word16_t *x, int len)
+{
+ return compute_rms(x, len);
+}
#endif
@@ -236,6 +297,8 @@
int i,j;
spx_sig_t xi,yi,nyi;
+ for (i=0;i<ord;i++)
+ mem[i] = SHR32(mem[i],1);
for (i=0;i<N;i++)
{
xi=SATURATE(x[i],805306368);
@@ -248,10 +311,40 @@
mem[ord-1] = SUB32(MULT16_32_Q15(num[ord-1],xi), MULT16_32_Q15(den[ord-1],yi));
y[i] = yi;
}
+ for (i=0;i<ord;i++)
+ mem[i] = SHL32(mem[i],1);
}
#endif
#endif
+#ifdef FIXED_POINT
+#ifndef OVERRIDE_FILTER_MEM16
+void filter_mem16(const spx_word16_t *x, const spx_coef_t *num, const spx_coef_t *den, spx_word16_t *y, int N, int ord, spx_mem_t *mem, char *stack)
+{
+ int i,j;
+ spx_word16_t xi,yi,nyi;
+ for (i=0;i<N;i++)
+ {
+ xi= x[i];
+ yi = EXTRACT16(SATURATE(ADD32(EXTEND32(x[i]),PSHR32(mem[0],LPC_SHIFT)),32767));
+ nyi = NEG16(yi);
+ for (j=0;j<ord-1;j++)
+ {
+ mem[j] = MAC16_16(MAC16_16(mem[j+1], num[j],xi), den[j],nyi);
+ }
+ mem[ord-1] = ADD32(MULT16_16(num[ord-1],xi), MULT16_16(den[ord-1],nyi));
+ y[i] = yi;
+ }
+}
+#endif
+#else
+void filter_mem16(const spx_word16_t *x, const spx_coef_t *num, const spx_coef_t *den, spx_word16_t *y, int N, int ord, spx_mem_t *mem, char *stack)
+{
+ filter_mem2(x, num, den, y, N, ord, mem);
+}
+#endif
+
+
#ifndef OVERRIDE_IIR_MEM2
#ifdef PRECISION16
void iir_mem2(const spx_sig_t *x, const spx_coef_t *den, spx_sig_t *y, int N, int ord, spx_mem_t *mem)
@@ -277,6 +370,8 @@
int i,j;
spx_word32_t xi,yi,nyi;
+ for (i=0;i<ord;i++)
+ mem[i] = SHR32(mem[i],1);
for (i=0;i<N;i++)
{
xi=SATURATE(x[i],805306368);
@@ -289,10 +384,40 @@
mem[ord-1] = MULT16_32_Q15(den[ord-1],nyi);
y[i] = yi;
}
+ for (i=0;i<ord;i++)
+ mem[i] = SHL32(mem[i],1);
}
#endif
#endif
+#ifdef FIXED_POINT
+#ifndef OVERRIDE_IIR_MEM16
+void iir_mem16(const spx_word16_t *x, const spx_coef_t *den, spx_word16_t *y, int N, int ord, spx_mem_t *mem, char *stack)
+{
+ int i,j;
+ spx_word16_t yi,nyi;
+
+ for (i=0;i<N;i++)
+ {
+ yi = EXTRACT16(SATURATE(ADD32(EXTEND32(x[i]),PSHR32(mem[0],LPC_SHIFT)),32767));
+ nyi = NEG16(yi);
+ for (j=0;j<ord-1;j++)
+ {
+ mem[j] = MAC16_16(mem[j+1],den[j],nyi);
+ }
+ mem[ord-1] = MULT16_16(den[ord-1],nyi);
+ y[i] = yi;
+ }
+}
+#endif
+#else
+void iir_mem16(const spx_word16_t *x, const spx_coef_t *den, spx_word16_t *y, int N, int ord, spx_mem_t *mem, char *stack)
+{
+ iir_mem2(x, den, y, N, ord, mem);
+}
+#endif
+
+
#ifndef OVERRIDE_FIR_MEM2
#ifdef PRECISION16
void fir_mem2(const spx_sig_t *x, const spx_coef_t *num, spx_sig_t *y, int N, int ord, spx_mem_t *mem)
@@ -318,6 +443,8 @@
int i,j;
spx_word32_t xi,yi;
+ for (i=0;i<ord;i++)
+ mem[i] = SHR32(mem[i],1);
for (i=0;i<N;i++)
{
xi=SATURATE(x[i],805306368);
@@ -329,11 +456,38 @@
mem[ord-1] = MULT16_32_Q15(num[ord-1],xi);
y[i] = SATURATE(yi,805306368);
}
+ for (i=0;i<ord;i++)
+ mem[i] = SHL32(mem[i],1);
}
#endif
#endif
+#ifdef FIXED_POINT
+#ifndef OVERRIDE_FIR_MEM16
+void fir_mem16(const spx_word16_t *x, const spx_coef_t *num, spx_word16_t *y, int N, int ord, spx_mem_t *mem, char *stack)
+{
+ int i,j;
+ spx_word16_t xi,yi;
+ for (i=0;i<N;i++)
+ {
+ xi=x[i];
+ yi = EXTRACT16(SATURATE(ADD32(EXTEND32(x[i]),PSHR32(mem[0],LPC_SHIFT)),32767));
+ for (j=0;j<ord-1;j++)
+ {
+ mem[j] = MAC16_16(mem[j+1], num[j],xi);
+ }
+ mem[ord-1] = MULT16_16(num[ord-1],xi);
+ y[i] = yi;
+ }
+}
+#endif
+#else
+void fir_mem16(const spx_word16_t *x, const spx_coef_t *num, spx_word16_t *y, int N, int ord, spx_mem_t *mem, char *stack)
+{
+ fir_mem2(x, num, y, N, ord, mem);
+}
+#endif
@@ -382,14 +536,13 @@
i++;
for (;i<N;i++)
y[i] = VERY_SMALL;
-
for (i=0;i<ord;i++)
mem1[i] = mem2[i] = 0;
for (i=0;i<N;i++)
{
y1 = ADD16(y[i], EXTRACT16(PSHR32(mem1[0],LPC_SHIFT)));
ny1i = NEG16(y1);
- y[i] = ADD16(SHL16(y1,1), EXTRACT16(PSHR32(mem2[0],LPC_SHIFT)));
+ y[i] = PSHR32(ADD32(SHL32(EXTEND32(y1),LPC_SHIFT+1),mem2[0]),LPC_SHIFT);
ny2i = NEG16(y[i]);
for (j=0;j<ord-1;j++)
{
@@ -426,12 +579,14 @@
y2[k]=0;
for (j=0;j<M2;j++)
{
- y1[k]=ADD32(y1[k],SHR(MULT16_16(a[j],ADD16(x[i+j],x2[i-j])),1));
- y2[k]=SUB32(y2[k],SHR(MULT16_16(a[j],SUB16(x[i+j],x2[i-j])),1));
+ y1[k]=ADD32(y1[k],MULT16_16(a[j],ADD16(x[i+j],x2[i-j])));
+ y2[k]=SUB32(y2[k],MULT16_16(a[j],SUB16(x[i+j],x2[i-j])));
j++;
- y1[k]=ADD32(y1[k],SHR(MULT16_16(a[j],ADD16(x[i+j],x2[i-j])),1));
- y2[k]=ADD32(y2[k],SHR(MULT16_16(a[j],SUB16(x[i+j],x2[i-j])),1));
+ y1[k]=ADD32(y1[k],MULT16_16(a[j],ADD16(x[i+j],x2[i-j])));
+ y2[k]=ADD32(y2[k],MULT16_16(a[j],SUB16(x[i+j],x2[i-j])));
}
+ y1[k] = SHR32(y1[k],1);
+ y2[k] = SHR32(y2[k],1);
}
for (i=0;i<M-1;i++)
mem[i]=SATURATE(PSHR(xx[N-i-1],1),16383);
@@ -450,7 +605,7 @@
ALLOC(xx, M+N-1, spx_word16_t);
for (i = 0; i < N/2; i++)
- xx[2*i] = SHR(x[N/2-1-i],SIG_SHIFT+1);
+ xx[2*i] = PSHR32(x[N/2-1-i],SIG_SHIFT);
for (i = 0; i < M - 1; i += 2)
xx[N+i] = mem[i+1];
@@ -469,19 +624,19 @@
a1 = a[j+1];
x1 = xx[N-2+j-i];
- y0 = ADD32(y0,SHR(MULT16_16(a0, x1),1));
- y1 = ADD32(y1,SHR(MULT16_16(a1, x1),1));
- y2 = ADD32(y2,SHR(MULT16_16(a0, x0),1));
- y3 = ADD32(y3,SHR(MULT16_16(a1, x0),1));
+ y0 = ADD32(y0,SHR(MULT16_16(a0, x1),2));
+ y1 = ADD32(y1,SHR(MULT16_16(a1, x1),2));
+ y2 = ADD32(y2,SHR(MULT16_16(a0, x0),2));
+ y3 = ADD32(y3,SHR(MULT16_16(a1, x0),2));
a0 = a[j+2];
a1 = a[j+3];
x0 = xx[N+j-i];
- y0 = ADD32(y0,SHR(MULT16_16(a0, x0),1));
- y1 = ADD32(y1,SHR(MULT16_16(a1, x0),1));
- y2 = ADD32(y2,SHR(MULT16_16(a0, x1),1));
- y3 = ADD32(y3,SHR(MULT16_16(a1, x1),1));
+ y0 = ADD32(y0,SHR(MULT16_16(a0, x0),2));
+ y1 = ADD32(y1,SHR(MULT16_16(a1, x0),2));
+ y2 = ADD32(y2,SHR(MULT16_16(a0, x1),2));
+ y3 = ADD32(y3,SHR(MULT16_16(a1, x1),2));
}
y[i] = y0;
y[i+1] = y1;
@@ -493,113 +648,229 @@
mem[i+1] = xx[i];
}
-void comb_filter_mem_init (CombFilterMem *mem)
-{
- mem->last_pitch=0;
- mem->last_pitch_gain[0]=mem->last_pitch_gain[1]=mem->last_pitch_gain[2]=0;
- mem->smooth_gain=1;
-}
-
#ifdef FIXED_POINT
-#define COMB_STEP 32767
+#if 0
+spx_word16_t shift_filt[3][7] = {{-33, 1043, -4551, 19959, 19959, -4551, 1043},
+ {-98, 1133, -4425, 29179, 8895, -2328, 444},
+ {444, -2328, 8895, 29179, -4425, 1133, -98}};
#else
-#define COMB_STEP 1.0
+spx_word16_t shift_filt[3][7] = {{-390, 1540, -4993, 20123, 20123, -4993, 1540},
+ {-1064, 2817, -6694, 31589, 6837, -990, -209},
+ {-209, -990, 6837, 31589, -6694, 2817, -1064}};
+#endif
+#else
+#if 0
+float shift_filt[3][7] = {{-9.9369e-04, 3.1831e-02, -1.3889e-01, 6.0910e-01, 6.0910e-01, -1.3889e-01, 3.1831e-02},
+ {-0.0029937, 0.0345613, -0.1350474, 0.8904793, 0.2714479, -0.0710304, 0.0135403},
+ {0.0135403, -0.0710304, 0.2714479, 0.8904793, -0.1350474, 0.0345613, -0.0029937}};
+#else
+float shift_filt[3][7] = {{-0.011915, 0.046995, -0.152373, 0.614108, 0.614108, -0.152373, 0.046995},
+ {-0.0324855, 0.0859768, -0.2042986, 0.9640297, 0.2086420, -0.0302054, -0.0063646},
+ {-0.0063646, -0.0302054, 0.2086420, 0.9640297, -0.2042986, 0.0859768, -0.0324855}};
+#endif
#endif
-void comb_filter(
-spx_sig_t *exc, /*decoded excitation*/
-spx_sig_t *new_exc, /*enhanced excitation*/
+int interp_pitch(
+spx_word16_t *exc, /*decoded excitation*/
+spx_word16_t *interp, /*decoded excitation*/
+int pitch, /*pitch period*/
+int len
+)
+{
+ int i,j,k;
+ spx_word32_t corr[4][7];
+ spx_word32_t maxcorr;
+ int maxi, maxj;
+ for (i=0;i<7;i++)
+ {
+ corr[0][i] = inner_prod(exc, exc-pitch-3+i, len);
+ }
+ for (i=0;i<3;i++)
+ {
+ for (j=0;j<7;j++)
+ {
+ int i1, i2;
+ spx_word32_t tmp=0;
+ i1 = 3-j;
+ if (i1<0)
+ i1 = 0;
+ i2 = 10-j;
+ if (i2>7)
+ i2 = 7;
+ for (k=i1;k<i2;k++)
+ tmp += MULT16_32_Q15(shift_filt[i][k],corr[0][j+k-3]);
+ corr[i+1][j] = tmp;
+ }
+ }
+ maxi=maxj=0;
+ maxcorr = corr[0][0];
+ for (i=0;i<4;i++)
+ {
+ for (j=0;j<7;j++)
+ {
+ if (corr[i][j] > maxcorr)
+ {
+ maxcorr = corr[i][j];
+ maxi=i;
+ maxj=j;
+ }
+ }
+ }
+ for (i=0;i<len;i++)
+ {
+ spx_word32_t tmp = 0;
+ if (maxi>0)
+ {
+ for (k=0;k<7;k++)
+ {
+ tmp += MULT16_16(exc[i-(pitch-maxj+3)+k-3],shift_filt[maxi-1][k]);
+ }
+ } else {
+ tmp = SHL32(exc[i-(pitch-maxj+3)],15);
+ }
+ interp[i] = PSHR32(tmp,15);
+ }
+ return pitch-maxj+3;
+}
+
+void multicomb(
+spx_word16_t *exc, /*decoded excitation*/
+spx_word16_t *new_exc, /*enhanced excitation*/
spx_coef_t *ak, /*LPC filter coefs*/
int p, /*LPC order*/
int nsf, /*sub-frame size*/
int pitch, /*pitch period*/
-spx_word16_t *pitch_gain, /*pitch gain (3-tap)*/
+int max_pitch,
spx_word16_t comb_gain, /*gain of comb filter*/
-CombFilterMem *mem
+char *stack
)
{
- int i;
- spx_word16_t exc_energy=0, new_exc_energy=0;
- spx_word16_t gain;
- spx_word16_t step;
- spx_word16_t fact;
-
- /*Compute excitation amplitude prior to enhancement*/
- exc_energy = compute_rms(exc, nsf);
- /*for (i=0;i<nsf;i++)
- exc_energy+=((float)exc[i])*exc[i];*/
-
- /*Some gain adjustment if pitch is too high or if unvoiced*/
-#ifdef FIXED_POINT
- {
- spx_word16_t g = gain_3tap_to_1tap(pitch_gain)+gain_3tap_to_1tap(mem->last_pitch_gain);
- if (g > 166)
- comb_gain = MULT16_16_Q15(DIV32_16(SHL32(EXTEND32(165),15),g), comb_gain);
- if (g < 64)
- comb_gain = MULT16_16_Q15(SHL16(g, 9), comb_gain);
- }
-#else
- {
- float g=0;
- g = GAIN_SCALING_1*.5*(gain_3tap_to_1tap(pitch_gain)+gain_3tap_to_1tap(mem->last_pitch_gain));
- if (g>1.3)
- comb_gain*=1.3/g;
- if (g<.5)
- comb_gain*=2.*g;
- }
-#endif
- step = DIV32(COMB_STEP, nsf);
- fact=0;
-
- /*Apply pitch comb-filter (filter out noise between pitch harmonics)*/
- for (i=0;i<nsf;i++)
- {
- spx_word32_t exc1, exc2;
-
- fact = ADD16(fact,step);
-
- exc1 = SHL32(MULT16_32_Q15(SHL16(pitch_gain[0],7),exc[i-pitch+1]) +
- MULT16_32_Q15(SHL16(pitch_gain[1],7),exc[i-pitch]) +
- MULT16_32_Q15(SHL16(pitch_gain[2],7),exc[i-pitch-1]) , 2);
- exc2 = SHL32(MULT16_32_Q15(SHL16(mem->last_pitch_gain[0],7),exc[i-mem->last_pitch+1]) +
- MULT16_32_Q15(SHL16(mem->last_pitch_gain[1],7),exc[i-mem->last_pitch]) +
- MULT16_32_Q15(SHL16(mem->last_pitch_gain[2],7),exc[i-mem->last_pitch-1]),2);
-
- new_exc[i] = exc[i] + MULT16_32_Q15(comb_gain, ADD32(MULT16_32_Q15(fact,exc1), MULT16_32_Q15(SUB16(COMB_STEP,fact), exc2)));
- }
-
- mem->last_pitch_gain[0] = pitch_gain[0];
- mem->last_pitch_gain[1] = pitch_gain[1];
- mem->last_pitch_gain[2] = pitch_gain[2];
- mem->last_pitch = pitch;
-
- /*Amplitude after enhancement*/
- new_exc_energy = compute_rms(new_exc, nsf);
-
- if (exc_energy > new_exc_energy)
- exc_energy = new_exc_energy;
+ int i;
+ VARDECL(spx_word16_t *iexc);
+ spx_word16_t old_ener, new_ener;
+ int corr_pitch;
- gain = DIV32_16(SHL32(EXTEND32(exc_energy),15),ADD16(1,new_exc_energy));
+ spx_word16_t iexc0_mag, iexc1_mag, exc_mag;
+ spx_word32_t corr0, corr1;
+ spx_word16_t gain0, gain1;
+ spx_word16_t pgain1, pgain2;
+ spx_word16_t c1, c2;
+ spx_word16_t g1, g2;
+ spx_word16_t ngain;
+ spx_word16_t gg1, gg2;
-#ifdef FIXED_POINT
- if (gain < 16384)
- gain = 16384;
-#else
- if (gain < .5)
- gain=.5;
-#endif
-
-#ifdef FIXED_POINT
- for (i=0;i<nsf;i++)
+#if 0 /* Set to 1 to enable full pitch search */
+ int nol_pitch[6];
+ spx_word16_t nol_pitch_coef[6];
+ spx_word16_t ol_pitch_coef;
+ open_loop_nbest_pitch(exc, 20, 120, nsf,
+ nol_pitch, nol_pitch_coef, 6, stack);
+ corr_pitch=nol_pitch[0];
+ ol_pitch_coef = nol_pitch_coef[0];
+ /*Try to remove pitch multiples*/
+ for (i=1;i<6;i++)
{
- mem->smooth_gain = ADD16(MULT16_16_Q15(31457,mem->smooth_gain), MULT16_16_Q15(1311,gain));
- new_exc[i] = MULT16_32_Q15(mem->smooth_gain, new_exc[i]);
+#ifdef FIXED_POINT
+ if ((nol_pitch_coef[i]>MULT16_16_Q15(nol_pitch_coef[0],19661)) &&
+#else
+ if ((nol_pitch_coef[i]>.6*nol_pitch_coef[0]) &&
+#endif
+ (ABS(2*nol_pitch[i]-corr_pitch)<=2 || ABS(3*nol_pitch[i]-corr_pitch)<=3 ||
+ ABS(4*nol_pitch[i]-corr_pitch)<=4 || ABS(5*nol_pitch[i]-corr_pitch)<=5))
+ {
+ corr_pitch = nol_pitch[i];
+ }
}
#else
- for (i=0;i<nsf;i++)
- {
- mem->smooth_gain = .96*mem->smooth_gain + .04*gain;
- new_exc[i] *= mem->smooth_gain;
- }
+ corr_pitch = pitch;
#endif
+
+ ALLOC(iexc, 2*nsf, spx_word16_t);
+
+ interp_pitch(exc, iexc, corr_pitch, 80);
+ if (corr_pitch>max_pitch)
+ interp_pitch(exc, iexc+nsf, 2*corr_pitch, 80);
+ else
+ interp_pitch(exc, iexc+nsf, -corr_pitch, 80);
+
+ /*interp_pitch(exc, iexc+2*nsf, 2*corr_pitch, 80);*/
+
+ /*printf ("%d %d %f\n", pitch, corr_pitch, max_corr*ener_1);*/
+ iexc0_mag = spx_sqrt(1000+inner_prod(iexc,iexc,nsf));
+ iexc1_mag = spx_sqrt(1000+inner_prod(iexc+nsf,iexc+nsf,nsf));
+ exc_mag = spx_sqrt(1+inner_prod(exc,exc,nsf));
+ corr0 = inner_prod(iexc,exc,nsf);
+ if (corr0<0)
+ corr0=0;
+ corr1 = inner_prod(iexc+nsf,exc,nsf);
+ if (corr1<0)
+ corr1=0;
+#ifdef FIXED_POINT
+ /* Doesn't cost much to limit the ratio and it makes the rest easier */
+ if (SHL32(EXTEND32(iexc0_mag),6) < EXTEND32(exc_mag))
+ iexc0_mag = ADD16(1,PSHR16(exc_mag,6));
+ if (SHL32(EXTEND32(iexc1_mag),6) < EXTEND32(exc_mag))
+ iexc1_mag = ADD16(1,PSHR16(exc_mag,6));
+#endif
+ if (corr0 > MULT16_16(iexc0_mag,exc_mag))
+ pgain1 = QCONST16(1., 14);
+ else
+ pgain1 = PDIV32_16(SHL32(PDIV32(corr0, exc_mag),14),iexc0_mag);
+ if (corr1 > MULT16_16(iexc1_mag,exc_mag))
+ pgain2 = QCONST16(1., 14);
+ else
+ pgain2 = PDIV32_16(SHL32(PDIV32(corr1, exc_mag),14),iexc1_mag);
+ gg1 = PDIV32_16(SHL32(EXTEND32(exc_mag),8), iexc0_mag);
+ gg2 = PDIV32_16(SHL32(EXTEND32(exc_mag),8), iexc1_mag);
+ if (comb_gain>0)
+ {
+#ifdef FIXED_POINT
+ c1 = (MULT16_16_Q15(QCONST16(.4,15),comb_gain)+QCONST16(.07,15));
+ c2 = QCONST16(.5,15)+MULT16_16_Q14(QCONST16(1.72,14),(c1-QCONST16(.07,15)));
+#else
+ c1 = .4*comb_gain+.07;
+ c2 = .5+1.72*(c1-.07);
+#endif
+ } else
+ {
+ c1=c2=0;
+ }
+#ifdef FIXED_POINT
+ g1 = 32767 - MULT16_16_Q13(MULT16_16_Q15(c2, pgain1),pgain1);
+ g2 = 32767 - MULT16_16_Q13(MULT16_16_Q15(c2, pgain2),pgain2);
+#else
+ g1 = 1-c2*pgain1*pgain1;
+ g2 = 1-c2*pgain2*pgain2;
+#endif
+ if (g1<c1)
+ g1 = c1;
+ if (g2<c1)
+ g2 = c1;
+ g1 = (spx_word16_t)PDIV32_16(SHL32(EXTEND32(c1),14),(spx_word16_t)g1);
+ g2 = (spx_word16_t)PDIV32_16(SHL32(EXTEND32(c1),14),(spx_word16_t)g2);
+ if (corr_pitch>max_pitch)
+ {
+ gain0 = MULT16_16_Q15(QCONST16(.7,15),MULT16_16_Q14(g1,gg1));
+ gain1 = MULT16_16_Q15(QCONST16(.3,15),MULT16_16_Q14(g2,gg2));
+ } else {
+ gain0 = MULT16_16_Q15(QCONST16(.6,15),MULT16_16_Q14(g1,gg1));
+ gain1 = MULT16_16_Q15(QCONST16(.6,15),MULT16_16_Q14(g2,gg2));
+ }
+ for (i=0;i<nsf;i++)
+ new_exc[i] = ADD16(exc[i], EXTRACT16(PSHR32(ADD32(MULT16_16(gain0,iexc[i]), MULT16_16(gain1,iexc[i+nsf])),8)));
+ /* FIXME: compute_rms16 is currently not quite accurate enough (but close) */
+ new_ener = compute_rms16(new_exc, nsf);
+ old_ener = compute_rms16(exc, nsf);
+
+ if (old_ener < 1)
+ old_ener = 1;
+ if (new_ener < 1)
+ new_ener = 1;
+ if (old_ener > new_ener)
+ old_ener = new_ener;
+ ngain = PDIV32_16(SHL32(EXTEND32(old_ener),14),new_ener);
+
+ for (i=0;i<nsf;i++)
+ new_exc[i] = MULT16_16_Q14(ngain, new_exc[i]);
}
+
diff --git a/pjmedia/src/pjmedia-codec/speex/filters.h b/pjmedia/src/pjmedia-codec/speex/filters.h
index c86d189..35de5e1 100644
--- a/pjmedia/src/pjmedia-codec/speex/filters.h
+++ b/pjmedia/src/pjmedia-codec/speex/filters.h
@@ -38,8 +38,9 @@
#include "misc.h"
spx_word16_t compute_rms(const spx_sig_t *x, int len);
+spx_word16_t compute_rms16(const spx_word16_t *x, int len);
void signal_mul(const spx_sig_t *x, spx_sig_t *y, spx_word32_t scale, int len);
-void signal_div(const spx_sig_t *x, spx_sig_t *y, spx_word32_t scale, int len);
+void signal_div(const spx_word16_t *x, spx_word16_t *y, spx_word32_t scale, int len);
#ifdef FIXED_POINT
@@ -47,13 +48,6 @@
#endif
-/** Combined filter memory. */
-typedef struct {
- int last_pitch;
- spx_word16_t last_pitch_gain[3];
- spx_word16_t smooth_gain;
-} CombFilterMem;
-
void qmf_decomp(const spx_word16_t *xx, const spx_word16_t *aa, spx_sig_t *, spx_sig_t *y2, int N, int M, spx_word16_t *mem, char *stack);
void fir_mem_up(const spx_sig_t *x, const spx_word16_t *a, spx_sig_t *y, int N, int M, spx_word32_t *mem, char *stack);
@@ -63,6 +57,10 @@
void fir_mem2(const spx_sig_t *x, const spx_coef_t *num, spx_sig_t *y, int N, int ord, spx_mem_t *mem);
void iir_mem2(const spx_sig_t *x, const spx_coef_t *den, spx_sig_t *y, int N, int ord, spx_mem_t *mem);
+void filter_mem16(const spx_word16_t *x, const spx_coef_t *num, const spx_coef_t *den, spx_word16_t *y, int N, int ord, spx_mem_t *mem, char *stack);
+void iir_mem16(const spx_word16_t *x, const spx_coef_t *den, spx_word16_t *y, int N, int ord, spx_mem_t *mem, char *stack);
+void fir_mem16(const spx_word16_t *x, const spx_coef_t *num, spx_word16_t *y, int N, int ord, spx_mem_t *mem, char *stack);
+
/* Apply bandwidth expansion on LPC coef */
void bw_lpc(spx_word16_t , const spx_coef_t *lpc_in, spx_coef_t *lpc_out, int order);
@@ -74,19 +72,16 @@
void compute_impulse_response(const spx_coef_t *ak, const spx_coef_t *awk1, const spx_coef_t *awk2, spx_word16_t *y, int N, int ord, char *stack);
-void comb_filter_mem_init (CombFilterMem *mem);
-
-void comb_filter(
-spx_sig_t *exc, /*decoded excitation*/
-spx_sig_t *new_exc, /*enhanced excitation*/
+void multicomb(
+spx_word16_t *exc, /*decoded excitation*/
+spx_word16_t *new_exc, /*enhanced excitation*/
spx_coef_t *ak, /*LPC filter coefs*/
int p, /*LPC order*/
int nsf, /*sub-frame size*/
int pitch, /*pitch period*/
-spx_word16_t *pitch_gain, /*pitch gain (3-tap)*/
+int max_pitch, /*pitch gain (3-tap)*/
spx_word16_t comb_gain, /*gain of comb filter*/
-CombFilterMem *mem
+char *stack
);
-
#endif
diff --git a/pjmedia/src/pjmedia-codec/speex/filters_arm4.h b/pjmedia/src/pjmedia-codec/speex/filters_arm4.h
index b250364..ac4d7a9 100644
--- a/pjmedia/src/pjmedia-codec/speex/filters_arm4.h
+++ b/pjmedia/src/pjmedia-codec/speex/filters_arm4.h
@@ -101,6 +101,8 @@
int i,j;
spx_sig_t xi,yi,nyi;
+ for (i=0;i<ord;i++)
+ mem[i] = SHR32(mem[i],1);
for (i=0;i<N;i++)
{
int deadm, deadn, deadd, deadidx, x1, y1, dead1, dead2, dead3, dead4, dead5, dead6;
@@ -252,6 +254,8 @@
: "cc", "memory");
}
+ for (i=0;i<ord;i++)
+ mem[i] = SHL32(mem[i],1);
}
#define OVERRIDE_IIR_MEM2
@@ -260,6 +264,9 @@
int i,j;
spx_sig_t xi,yi,nyi;
+ for (i=0;i<ord;i++)
+ mem[i] = SHR32(mem[i],1);
+
for (i=0;i<N;i++)
{
int deadm, deadd, deadidx, dead1, dead2, dead3, dead4, dead5, dead6;
@@ -376,4 +383,7 @@
: "cc", "memory");
}
+ for (i=0;i<ord;i++)
+ mem[i] = SHL32(mem[i],1);
+
}
diff --git a/pjmedia/src/pjmedia-codec/speex/filters_bfin.h b/pjmedia/src/pjmedia-codec/speex/filters_bfin.h
index 9f7ea6a..2180ed4 100644
--- a/pjmedia/src/pjmedia-codec/speex/filters_bfin.h
+++ b/pjmedia/src/pjmedia-codec/speex/filters_bfin.h
@@ -32,14 +32,11 @@
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
-#include <stdio.h>
-
#define OVERRIDE_NORMALIZE16
int normalize16(const spx_sig_t *x, spx_word16_t *y, spx_sig_t max_scale, int len)
{
spx_sig_t max_val=1;
int sig_shift;
-
__asm__
(
"%0 = 0;\n\t"
@@ -67,18 +64,17 @@
(
"I0 = %0;\n\t"
"L0 = 0;\n\t"
- "I1 = %1;\n\t"
- "L1 = 0;\n\t"
+ "P1 = %1;\n\t"
"R0 = [I0++];\n\t"
- "LOOP norm_shift%= LC0 = %3 >> 1;\n\t"
+ "LOOP norm_shift%= LC0 = %3;\n\t"
"LOOP_BEGIN norm_shift%=;\n\t"
- "R1 = ASHIFT R0 by %2.L || R2 = [I0++];\n\t"
- "R3 = ASHIFT R2 by %2.L || R0 = [I0++];\n\t"
- "R3 = PACK(R3.L, R1.L);\n\t"
- "[I1++] = R3;\n\t"
+ "R1 = ASHIFT R0 by %2.L || R0 = [I0++];\n\t"
+ "W[P1++] = R1;\n\t"
"LOOP_END norm_shift%=;\n\t"
- : : "a" (x), "a" (y), "d" (-sig_shift), "a" (len)
- : "I0", "L0", "I1", "L1", "R0", "R1", "R2", "R3", "memory"
+ "R1 = ASHIFT R0 by %2.L;\n\t"
+ "W[P1++] = R1;\n\t"
+ : : "a" (x), "a" (y), "d" (-sig_shift), "a" (len-1)
+ : "I0", "L0", "P1", "R0", "R1", "memory"
);
return sig_shift;
}
@@ -103,26 +99,26 @@
"P0 = %3;\n\t"
"I0 = P0;\n\t"
- "B0 = P0;\n\t"
+ "B0 = P0;\n\t" /* numden */
"L0 = 0;\n\t"
- "P2 = %0;\n\t"
+ "P2 = %0;\n\t" /* Fused xy */
"I2 = P2;\n\t"
"L2 = 0;\n\t"
- "P4 = %6;\n\t"
- "P0 = %1;\n\t"
- "P1 = %2;\n\t"
+ "P4 = %6;\n\t" /* mem */
+ "P0 = %1;\n\t" /* _x */
+ "P1 = %2;\n\t" /* _y */
/* First sample */
"R1 = [P4++];\n\t"
- "R1 <<= 1;\n\t"
- "R2 = [P0++];\n\t"
+ "R1 <<= 1;\n\t" /* shift mem */
+ "R2 = [P0++];\n\t" /* load x[0] */
"R1 = R1 + R2;\n\t"
- "[P1++] = R1;\n\t"
+ "[P1++] = R1;\n\t" /* store y[0] */
"R1 <<= 2;\n\t"
"R2 <<= 2;\n\t"
- "R2 = PACK(R1.H, R2.H);\n\t"
+ "R2 = PACK(R1.H, R2.H);\n\t" /* pack x16 and y16 */
"[P2] = R2;\n\t"
/* Samples 1 to ord-1 (using memory) */
@@ -147,13 +143,13 @@
"LOOP_END filter_start_inner%=;\n\t"
"A0 += A1;\n\t"
"R4 = A0;\n\t"
- "R4 <<= 1;\n\t"
- "R2 = [P0++];\n\t"
+ "R4 <<= 1;\n\t" /* shift mem */
+ "R2 = [P0++];\n\t" /* load x */
"R4 = R4 + R2;\n\t"
- "[P1++] = R4;\n\t"
+ "[P1++] = R4;\n\t" /* store y */
"R4 <<= 2;\n\t"
"R2 <<= 2;\n\t"
- "R2 = PACK(R4.H, R2.H);\n\t"
+ "R2 = PACK(R4.H, R2.H);\n\t" /* pack x16 and y16 */
"[P2] = R2;\n\t"
"LOOP_END filter_start%=;\n\t"
@@ -161,14 +157,14 @@
/* Samples ord to N*/
"R0 = %5;\n\t"
"R0 <<= 1;\n\t"
- "I0 = B0;\n\t"
+ "I0 = B0;\n\t" /* numden */
"R0 <<= 1;\n\t"
"L0 = R0;\n\t"
- "R0 = %5;\n\t"
- "R2 = %4;\n\t"
+ "R0 = %5;\n\t" /* org */
+ "R2 = %4;\n\t" /* N */
"R2 = R2 - R0;\n\t"
- "R4 = [I0++];\n\t"
+ "R4 = [I0++];\n\t" /* numden */
"LC0 = R2;\n\t"
"P3 = R0;\n\t"
"R0 <<= 2;\n\t"
@@ -176,7 +172,7 @@
"I2 = P2;\n\t"
"M0 = R0;\n\t"
"A1 = A0 = 0;\n\t"
- "R5 = [I2--];\n\t"
+ "R5 = [I2--];\n\t" /* load xy */
"LOOP filter_mid%= LC0;\n\t"
"LOOP_BEGIN filter_mid%=;\n\t"
"LOOP filter_mid_inner%= LC1=P3;\n\t"
@@ -184,9 +180,9 @@
"A1 -= R4.H*R5.H, A0 += R4.L*R5.L (IS) || R4 = [I0++] || R5 = [I2--];\n\t"
"LOOP_END filter_mid_inner%=;\n\t"
"R0 = (A0 += A1) || I2 += M0;\n\t"
- "R0 = R0 << 1 || R5 = [P0++];\n\t"
+ "R0 = R0 << 1 || R5 = [P0++];\n\t" /* load x */
"R0 = R0 + R5;\n\t"
- "R0 = R0 << 2 || [P1++] = R0;\n\t"
+ "R0 = R0 << 2 || [P1++] = R0;\n\t" /* shift y | store y */
"R5 = R5 << 2;\n\t"
"R5 = PACK(R0.H, R5.H);\n\t"
"A1 = A0 = 0 || [I2--] = R5\n\t"
@@ -222,6 +218,150 @@
}
+#define OVERRIDE_FILTER_MEM16
+void filter_mem16(const spx_word16_t *_x, const spx_coef_t *num, const spx_coef_t *den, spx_word16_t *_y, int N, int ord, spx_mem_t *mem, char *stack)
+{
+ VARDECL(spx_word32_t *xy2);
+ VARDECL(spx_word32_t *numden_a);
+ spx_word32_t *xy;
+ spx_word16_t *numden;
+ int i;
+
+ ALLOC(xy2, (N+1), spx_word32_t);
+ ALLOC(numden_a, (2*ord+2), spx_word32_t);
+ xy = xy2+1;
+ numden = (spx_word16_t*) numden_a;
+
+ for (i=0;i<ord;i++)
+ {
+ numden[2*i] = num[i];
+ numden[2*i+1] = den[i];
+ }
+ __asm__ __volatile__
+ (
+ /* Register setup */
+ "R0 = %5;\n\t" /*ord */
+
+ "P0 = %3;\n\t"
+ "I0 = P0;\n\t"
+ "B0 = P0;\n\t" /* numden */
+ "L0 = 0;\n\t"
+
+ "P2 = %0;\n\t" /* Fused xy */
+ "I2 = P2;\n\t"
+ "L2 = 0;\n\t"
+
+ "P4 = %6;\n\t" /* mem */
+ "P0 = %1;\n\t" /* _x */
+ "P1 = %2;\n\t" /* _y */
+
+ /* First sample */
+ "R1 = [P4++];\n\t"
+ "R1 <<= 3;\n\t" /* shift mem */
+ "R1.L = R1 (RND);\n\t"
+ "R2 = W[P0++];\n\t" /* load x[0] */
+ "R1.L = R1.L + R2.L;\n\t"
+ "W[P1++] = R1;\n\t" /* store y[0] */
+ "R2 = PACK(R1.L, R2.L);\n\t" /* pack x16 and y16 */
+ "[P2] = R2;\n\t"
+
+ /* Samples 1 to ord-1 (using memory) */
+ "R0 += -1;\n\t"
+ "R3 = 0;\n\t"
+ "LC0 = R0;\n\t"
+ "LOOP filter_start%= LC0;\n\t"
+ "LOOP_BEGIN filter_start%=;\n\t"
+ "R3 += 1;\n\t"
+ "LC1 = R3;\n\t"
+
+ "R1 = [P4++];\n\t"
+ "A1 = R1;\n\t"
+ "A0 = 0;\n\t"
+ "I0 = B0;\n\t"
+ "I2 = P2;\n\t"
+ "P2 += 4;\n\t"
+ "R4 = [I0++] || R5 = [I2--];\n\t"
+ "LOOP filter_start_inner%= LC1;\n\t"
+ "LOOP_BEGIN filter_start_inner%=;\n\t"
+ "A1 -= R4.H*R5.H, A0 += R4.L*R5.L (IS) || R4 = [I0++] || R5 = [I2--];\n\t"
+ "LOOP_END filter_start_inner%=;\n\t"
+ "A0 += A1;\n\t"
+ "R4 = A0;\n\t"
+ "R4 <<= 3;\n\t" /* shift mem */
+ "R4.L = R4 (RND);\n\t"
+ "R2 = W[P0++];\n\t" /* load x */
+ "R4.L = R4.L + R2.L;\n\t"
+ "W[P1++] = R4;\n\t" /* store y */
+ //"R4 <<= 2;\n\t"
+ //"R2 <<= 2;\n\t"
+ "R2 = PACK(R4.L, R2.L);\n\t" /* pack x16 and y16 */
+ "[P2] = R2;\n\t"
+
+ "LOOP_END filter_start%=;\n\t"
+
+ /* Samples ord to N*/
+ "R0 = %5;\n\t"
+ "R0 <<= 1;\n\t"
+ "I0 = B0;\n\t" /* numden */
+ "R0 <<= 1;\n\t"
+ "L0 = R0;\n\t"
+
+ "R0 = %5;\n\t" /* org */
+ "R2 = %4;\n\t" /* N */
+ "R2 = R2 - R0;\n\t"
+ "R4 = [I0++];\n\t" /* numden */
+ "LC0 = R2;\n\t"
+ "P3 = R0;\n\t"
+ "R0 <<= 2;\n\t"
+ "R0 += 8;\n\t"
+ "I2 = P2;\n\t"
+ "M0 = R0;\n\t"
+ "A1 = A0 = 0;\n\t"
+ "R5 = [I2--];\n\t" /* load xy */
+ "LOOP filter_mid%= LC0;\n\t"
+ "LOOP_BEGIN filter_mid%=;\n\t"
+ "LOOP filter_mid_inner%= LC1=P3;\n\t"
+ "LOOP_BEGIN filter_mid_inner%=;\n\t"
+ "A1 -= R4.H*R5.H, A0 += R4.L*R5.L (IS) || R4 = [I0++] || R5 = [I2--];\n\t"
+ "LOOP_END filter_mid_inner%=;\n\t"
+ "R0 = (A0 += A1) || I2 += M0;\n\t"
+ "R0 = R0 << 3 || R5 = W[P0++];\n\t" /* load x */
+ "R0.L = R0 (RND);\n\t"
+ "R0.L = R0.L + R5.L;\n\t"
+ "R5 = PACK(R0.L, R5.L) || W[P1++] = R0;\n\t" /* shift y | store y */
+ "A1 = A0 = 0 || [I2--] = R5\n\t"
+ "LOOP_END filter_mid%=;\n\t"
+ "I2 += 4;\n\t"
+ "P2 = I2;\n\t"
+ /* Update memory */
+ "P4 = %6;\n\t"
+ "R0 = %5;\n\t"
+ "LC0 = R0;\n\t"
+ "P0 = B0;\n\t"
+ "A1 = A0 = 0;\n\t"
+ "LOOP mem_update%= LC0;\n\t"
+ "LOOP_BEGIN mem_update%=;\n\t"
+ "I2 = P2;\n\t"
+ "I0 = P0;\n\t"
+ "P0 += 4;\n\t"
+ "R0 = LC0;\n\t"
+ "LC1 = R0;\n\t"
+ "R5 = [I2--] || R4 = [I0++];\n\t"
+ "LOOP mem_accum%= LC1;\n\t"
+ "LOOP_BEGIN mem_accum%=;\n\t"
+ "A1 -= R4.H*R5.H, A0 += R4.L*R5.L (IS) || R4 = [I0++] || R5 = [I2--];\n\t"
+ "LOOP_END mem_accum%=;\n\t"
+ "R0 = (A0 += A1);\n\t"
+ "A1 = A0 = 0 || [P4++] = R0;\n\t"
+ "LOOP_END mem_update%=;\n\t"
+ "L0 = 0;\n\t"
+ : : "m" (xy), "m" (_x), "m" (_y), "m" (numden), "m" (N), "m" (ord), "m" (mem)
+ : "A0", "A1", "R0", "R1", "R2", "R3", "R4", "R5", "P0", "P1", "P2", "P3", "P4", "B0", "I0", "I2", "L0", "L2", "M0", "memory"
+ );
+
+}
+
+
#define OVERRIDE_IIR_MEM2
@@ -346,6 +486,132 @@
}
+
+#define OVERRIDE_IIR_MEM16
+void iir_mem16(const spx_word16_t *_x, const spx_coef_t *den, spx_word16_t *_y, int N, int ord, spx_mem_t *mem, char *stack)
+{
+ VARDECL(spx_word16_t *y);
+ spx_word16_t *yy;
+
+ ALLOC(y, (N+2), spx_word16_t);
+ yy = y+2;
+
+ __asm__ __volatile__
+ (
+ /* Register setup */
+ "R0 = %5;\n\t" /*ord */
+
+ "P1 = %3;\n\t"
+ "I1 = P1;\n\t"
+ "B1 = P1;\n\t"
+ "L1 = 0;\n\t"
+
+ "P3 = %0;\n\t"
+ "I3 = P3;\n\t"
+ "L3 = 0;\n\t"
+
+ "P4 = %6;\n\t"
+ "P0 = %1;\n\t"
+ "P1 = %2;\n\t"
+
+ /* First sample */
+ "R1 = [P4++];\n\t"
+ "R1 = R1 << 3 (S);\n\t"
+ "R1.L = R1 (RND);\n\t"
+ "R2 = W[P0++];\n\t"
+ "R1 = R1 + R2;\n\t"
+ "W[P1++] = R1;\n\t"
+ "W[P3] = R1;\n\t"
+
+ /* Samples 1 to ord-1 (using memory) */
+ "R0 += -1;\n\t"
+ "R3 = 0;\n\t"
+ "LC0 = R0;\n\t"
+ "LOOP filter_start%= LC0;\n\t"
+ "LOOP_BEGIN filter_start%=;\n\t"
+ "R3 += 1;\n\t"
+ "LC1 = R3;\n\t"
+
+ "R1 = [P4++];\n\t"
+ "A1 = R1;\n\t"
+ "I1 = B1;\n\t"
+ "I3 = P3;\n\t"
+ "P3 += 2;\n\t"
+ "LOOP filter_start_inner%= LC1;\n\t"
+ "LOOP_BEGIN filter_start_inner%=;\n\t"
+ "R4.L = W[I1++];\n\t"
+ "R5.L = W[I3--];\n\t"
+ "A1 -= R4.L*R5.L (IS);\n\t"
+ "LOOP_END filter_start_inner%=;\n\t"
+
+ "R1 = A1;\n\t"
+ "R1 <<= 3;\n\t"
+ "R1.L = R1 (RND);\n\t"
+ "R2 = W[P0++];\n\t"
+ "R1 = R1 + R2;\n\t"
+ "W[P1++] = R1;\n\t"
+ "W[P3] = R1;\n\t"
+ "LOOP_END filter_start%=;\n\t"
+
+ /* Samples ord to N*/
+ "R0 = %5;\n\t"
+ "R0 <<= 1;\n\t"
+ "I1 = B1;\n\t"
+ "L1 = R0;\n\t"
+
+ "R0 = %5;\n\t"
+ "R2 = %4;\n\t"
+ "R2 = R2 - R0;\n\t"
+ "R4.L = W[I1++];\n\t"
+ "LC0 = R2;\n\t"
+ "LOOP filter_mid%= LC0;\n\t"
+ "LOOP_BEGIN filter_mid%=;\n\t"
+ "LC1 = R0;\n\t"
+ "A1 = 0;\n\t"
+ "I3 = P3;\n\t"
+ "P3 += 2;\n\t"
+ "R5.L = W[I3--];\n\t"
+ "LOOP filter_mid_inner%= LC1;\n\t"
+ "LOOP_BEGIN filter_mid_inner%=;\n\t"
+ "A1 -= R4.L*R5.L (IS) || R4.L = W[I1++] || R5.L = W[I3--];\n\t"
+ "LOOP_END filter_mid_inner%=;\n\t"
+ "R1 = A1;\n\t"
+ "R1 = R1 << 3 || R2 = W[P0++];\n\t"
+ "R1.L = R1 (RND);\n\t"
+ "R1 = R1 + R2;\n\t"
+ "W[P1++] = R1;\n\t"
+ "W[P3] = R1;\n\t"
+ "LOOP_END filter_mid%=;\n\t"
+
+ /* Update memory */
+ "P4 = %6;\n\t"
+ "R0 = %5;\n\t"
+ "LC0 = R0;\n\t"
+ "P1 = B1;\n\t"
+ "LOOP mem_update%= LC0;\n\t"
+ "LOOP_BEGIN mem_update%=;\n\t"
+ "A0 = 0;\n\t"
+ "I3 = P3;\n\t"
+ "I1 = P1;\n\t"
+ "P1 += 2;\n\t"
+ "R0 = LC0;\n\t"
+ "LC1=R0;\n\t"
+ "R5.L = W[I3--] || R4.L = W[I1++];\n\t"
+ "LOOP mem_accum%= LC1;\n\t"
+ "LOOP_BEGIN mem_accum%=;\n\t"
+ "A0 -= R4.L*R5.L (IS) || R4.L = W[I1++] || R5.L = W[I3--];\n\t"
+ "LOOP_END mem_accum%=;\n\t"
+ "R0 = A0;\n\t"
+ "[P4++] = R0;\n\t"
+ "LOOP_END mem_update%=;\n\t"
+ "L1 = 0;\n\t"
+ : : "m" (yy), "m" (_x), "m" (_y), "m" (den), "m" (N), "m" (ord), "m" (mem)
+ : "A0", "A1", "R0", "R1", "R2", "R3", "R4", "R5", "P0", "P1", "P2", "P3", "P4", "B1", "I1", "I3", "L1", "L3", "memory"
+ );
+
+}
+
+
#define OVERRIDE_FIR_MEM2
void fir_mem2(const spx_sig_t *x, const spx_coef_t *num, spx_sig_t *y, int N, int ord, spx_mem_t *mem)
{
@@ -358,6 +624,18 @@
filter_mem2(x, num, den, y, N, ord, mem);
}
+#define OVERRIDE_FIR_MEM16
+void fir_mem16(const spx_word16_t *x, const spx_coef_t *num, spx_word16_t *y, int N, int ord, spx_mem_t *mem, char *stack)
+{
+ int i;
+ spx_coef_t den2[12];
+ spx_coef_t *den;
+ den = (spx_coef_t*)((((int)den2)+4)&0xfffffffc);
+ for (i=0;i<10;i++)
+ den[i] = 0;
+ filter_mem16(x, num, den, y, N, ord, mem, stack);
+}
+
#define OVERRIDE_COMPUTE_IMPULSE_RESPONSE
void compute_impulse_response(const spx_coef_t *ak, const spx_coef_t *awk1, const spx_coef_t *awk2, spx_word16_t *y, int N, int ord, char *stack)
diff --git a/pjmedia/src/pjmedia-codec/speex/fixed_bfin.h b/pjmedia/src/pjmedia-codec/speex/fixed_bfin.h
index adee43c..aa26f6a 100644
--- a/pjmedia/src/pjmedia-codec/speex/fixed_bfin.h
+++ b/pjmedia/src/pjmedia-codec/speex/fixed_bfin.h
@@ -36,11 +36,40 @@
#ifndef FIXED_BFIN_H
#define FIXED_BFIN_H
+#undef PDIV32_16
+static inline spx_word16_t PDIV32_16(spx_word32_t a, spx_word16_t b)
+{
+ spx_word32_t res, bb;
+ bb = b;
+ a += b>>1;
+ __asm__ (
+ "P0 = 15;\n\t"
+ "R0 = %1;\n\t"
+ "R1 = %2;\n\t"
+ //"R0 = R0 + R1;\n\t"
+ "R0 <<= 1;\n\t"
+ "DIVS (R0, R1);\n\t"
+ "LOOP divide%= LC0 = P0;\n\t"
+ "LOOP_BEGIN divide%=;\n\t"
+ "DIVQ (R0, R1);\n\t"
+ "LOOP_END divide%=;\n\t"
+ "R0 = R0.L;\n\t"
+ "%0 = R0;\n\t"
+ : "=m" (res)
+ : "m" (a), "m" (bb)
+ : "P0", "R0", "R1", "cc");
+ return res;
+}
+
#undef DIV32_16
static inline spx_word16_t DIV32_16(spx_word32_t a, spx_word16_t b)
{
spx_word32_t res, bb;
bb = b;
+ /* Make the roundinf consistent with the C version
+ (do we need to do that?)*/
+ if (a<0)
+ a += (b-1);
__asm__ (
"P0 = 15;\n\t"
"R0 = %1;\n\t"
@@ -79,14 +108,12 @@
spx_word32_t res;
__asm__
(
- "%1 <<= 1;\n\t"
- "A1 = %2.L*%1.L (M,IS);\n\t"
- "A1 = A1 >>> 16;\n\t"
- "R1 = (A1 += %2.L*%1.H) (IS);\n\t"
- "%0 = R1;\n\t"
- : "=&d" (res), "=&d" (b)
+ "A1 = %2.L*%1.L (M);\n\t"
+ "A1 = A1 >>> 15;\n\t"
+ "%0 = (A1 += %2.L*%1.H) ;\n\t"
+ : "=&W" (res), "=&d" (b)
: "d" (a), "1" (b)
- : "A1", "R1"
+ : "A1"
);
return res;
}
@@ -97,14 +124,13 @@
spx_word32_t res;
__asm__
(
- "%1 <<= 1;\n\t"
- "A1 = %2.L*%1.L (M,IS);\n\t"
- "A1 = A1 >>> 16;\n\t"
- "R1 = (A1 += %2.L*%1.H) (IS);\n\t"
- "%0 = R1 + %4;\n\t"
- : "=&d" (res), "=&d" (b)
+ "A1 = %2.L*%1.L (M);\n\t"
+ "A1 = A1 >>> 15;\n\t"
+ "%0 = (A1 += %2.L*%1.H);\n\t"
+ "%0 = %0 + %4;\n\t"
+ : "=&W" (res), "=&d" (b)
: "d" (a), "1" (b), "d" (c)
- : "A1", "R1"
+ : "A1"
);
return res;
}
@@ -115,14 +141,13 @@
spx_word32_t res;
__asm__
(
- "%2 <<= 2;\n\t"
- "A1 = %1.L*%2.L (M,IS);\n\t"
- "A1 = A1 >>> 16;\n\t"
- "R1 = (A1 += %1.L*%2.H) (IS);\n\t"
- "%0 = R1;\n\t"
- : "=d" (res), "=d" (a), "=d" (b)
+ "%2 <<= 1;\n\t"
+ "A1 = %1.L*%2.L (M);\n\t"
+ "A1 = A1 >>> 15;\n\t"
+ "%0 = (A1 += %1.L*%2.H);\n\t"
+ : "=W" (res), "=d" (a), "=d" (b)
: "1" (a), "2" (b)
- : "A1", "R1"
+ : "A1"
);
return res;
}
@@ -133,14 +158,14 @@
spx_word32_t res;
__asm__
(
- "%1 <<= 2;\n\t"
- "A1 = %2.L*%1.L (M,IS);\n\t"
- "A1 = A1 >>> 16;\n\t"
- "R1 = (A1 += %2.L*%1.H) (IS);\n\t"
- "%0 = R1 + %4;\n\t"
- : "=&d" (res), "=&d" (b)
+ "%1 <<= 1;\n\t"
+ "A1 = %2.L*%1.L (M);\n\t"
+ "A1 = A1 >>> 15;\n\t"
+ "%0 = (A1 += %2.L*%1.H);\n\t"
+ "%0 = %0 + %4;\n\t"
+ : "=&W" (res), "=&d" (b)
: "d" (a), "1" (b), "d" (c)
- : "A1", "R1"
+ : "A1"
);
return res;
}
diff --git a/pjmedia/src/pjmedia-codec/speex/fixed_debug.h b/pjmedia/src/pjmedia-codec/speex/fixed_debug.h
index 7403a18..65c5712 100644
--- a/pjmedia/src/pjmedia-codec/speex/fixed_debug.h
+++ b/pjmedia/src/pjmedia-codec/speex/fixed_debug.h
@@ -40,8 +40,9 @@
extern long long spx_mips;
#define MIPS_INC spx_mips++,
-#define QCONST16(x,bits) ((spx_word16_t)((x)*(1<<(bits))+(1<<((bits)-1))))
-#define QCONST32(x,bits) ((spx_word32_t)((x)*(1<<(bits))+(1<<((bits)-1))))
+#define QCONST16(x,bits) ((spx_word16_t)(.5+(x)*(((spx_word32_t)1)<<(bits))))
+#define QCONST32(x,bits) ((spx_word32_t)(.5+(x)*(((spx_word32_t)1)<<(bits))))
+
#define VERIFY_SHORT(x) ((x)<=32767&&(x)>=-32768)
#define VERIFY_INT(x) ((x)<=2147483647LL&&(x)>=-2147483648LL)
@@ -169,7 +170,7 @@
}
res = a+b;
if (!VERIFY_SHORT(res))
- fprintf (stderr, "ADD16: output is not short: %d\n", res);
+ fprintf (stderr, "ADD16: output is not short: %d+%d=%d\n", a,b,res);
spx_mips++;
return res;
}
@@ -196,7 +197,9 @@
}
res = a+b;
if (!VERIFY_INT(res))
+ {
fprintf (stderr, "ADD32: output is not int: %d\n", (int)res);
+ }
spx_mips++;
return res;
}
@@ -251,6 +254,8 @@
#define MAC16_16(c,a,b) (spx_mips--,ADD32((c),MULT16_16((a),(b))))
#define MAC16_16_Q11(c,a,b) (ADD16((c),EXTRACT16(SHR32(MULT16_16((a),(b)),11))))
#define MAC16_16_Q13(c,a,b) (ADD16((c),EXTRACT16(SHR32(MULT16_16((a),(b)),13))))
+#define MAC16_16_P13(c,a,b) (ADD32((c),SHR(ADD32(4096,MULT16_16((a),(b))),13)))
+
static inline int MULT16_32_QX(int a, long long b, int Q)
{
@@ -437,7 +442,7 @@
spx_mips+=36;
return res;
}
-
-
+#define PDIV32(a,b) DIV32(ADD32((a),(b)>>1),b)
+#define PDIV32_16(a,b) DIV32_16(ADD32((a),(b)>>1),b)
#endif
diff --git a/pjmedia/src/pjmedia-codec/speex/fixed_generic.h b/pjmedia/src/pjmedia-codec/speex/fixed_generic.h
index d4bdc15..375050c 100644
--- a/pjmedia/src/pjmedia-codec/speex/fixed_generic.h
+++ b/pjmedia/src/pjmedia-codec/speex/fixed_generic.h
@@ -35,13 +35,13 @@
#ifndef FIXED_GENERIC_H
#define FIXED_GENERIC_H
-#define QCONST16(x,bits) ((spx_word16_t)(.5+(x)*(1<<(bits))))
-#define QCONST32(x,bits) ((spx_word32_t)(.5+(x)*(1<<(bits))))
+#define QCONST16(x,bits) ((spx_word16_t)(.5+(x)*(((spx_word32_t)1)<<(bits))))
+#define QCONST32(x,bits) ((spx_word32_t)(.5+(x)*(((spx_word32_t)1)<<(bits))))
#define NEG16(x) (-(x))
#define NEG32(x) (-(x))
-#define EXTRACT16(x) ((spx_word16_t)x)
-#define EXTEND32(x) ((spx_word32_t)x)
+#define EXTRACT16(x) ((spx_word16_t)(x))
+#define EXTEND32(x) ((spx_word32_t)(x))
#define SHR16(a,shift) ((a) >> (shift))
#define SHL16(a,shift) ((a) << (shift))
#define SHR32(a,shift) ((a) >> (shift))
@@ -61,7 +61,6 @@
#define SUB16(a,b) ((spx_word16_t)(a)-(spx_word16_t)(b))
#define ADD32(a,b) ((spx_word32_t)(a)+(spx_word32_t)(b))
#define SUB32(a,b) ((spx_word32_t)(a)-(spx_word32_t)(b))
-#define ADD64(a,b) ((spx_word64_t)(a)+(spx_word64_t)(b))
/* result fits in 16 bits */
@@ -84,6 +83,7 @@
#define MAC16_16_Q11(c,a,b) (ADD32((c),SHR(MULT16_16((a),(b)),11)))
#define MAC16_16_Q13(c,a,b) (ADD32((c),SHR(MULT16_16((a),(b)),13)))
+#define MAC16_16_P13(c,a,b) (ADD32((c),SHR(ADD32(4096,MULT16_16((a),(b))),13)))
#define MULT16_16_Q11_32(a,b) (SHR(MULT16_16((a),(b)),11))
#define MULT16_16_Q13(a,b) (SHR(MULT16_16((a),(b)),13))
@@ -97,6 +97,8 @@
#define MUL_16_32_R15(a,bh,bl) ADD32(MULT16_16((a),(bh)), SHR(MULT16_16((a),(bl)),15))
#define DIV32_16(a,b) ((spx_word16_t)(((spx_word32_t)(a))/((spx_word16_t)(b))))
+#define PDIV32_16(a,b) ((spx_word16_t)(((spx_word32_t)(a)+((spx_word16_t)(b)>>1))/((spx_word16_t)(b))))
#define DIV32(a,b) (((spx_word32_t)(a))/((spx_word32_t)(b)))
+#define PDIV32(a,b) (((spx_word32_t)(a)+((spx_word16_t)(b)>>1))/((spx_word32_t)(b)))
#endif
diff --git a/pjmedia/src/pjmedia-codec/speex/gain_table.c b/pjmedia/src/pjmedia-codec/speex/gain_table.c
index 54a5407..00b8244 100644
--- a/pjmedia/src/pjmedia-codec/speex/gain_table.c
+++ b/pjmedia/src/pjmedia-codec/speex/gain_table.c
@@ -29,132 +29,132 @@
POSSIBILITY OF SUCH DAMAGE.
*/
-const signed char gain_cdbk_nb[384] = {
--32,-32,-32,
--28,-67,-5,
--42,-6,-32,
--57,-10,-54,
--16,27,-41,
-19,-19,-40,
--45,24,-21,
--8,-14,-18,
-1,14,-58,
--18,-88,-39,
--38,21,-18,
--19,20,-43,
-10,17,-48,
--52,-58,-13,
--44,-1,-11,
--12,-11,-34,
-14,0,-46,
--37,-35,-34,
--25,44,-30,
-6,-4,-63,
--31,43,-41,
--23,30,-43,
--43,26,-14,
--33,1,-13,
--13,18,-37,
--46,-73,-45,
--36,24,-25,
--36,-11,-20,
--25,12,-18,
--36,-69,-59,
--45,6,8,
--22,-14,-24,
--1,13,-44,
--39,-48,-26,
--32,31,-37,
--33,15,-46,
--24,30,-36,
--41,31,-23,
--50,22,-4,
--22,2,-21,
--17,30,-34,
--7,-60,-28,
--38,42,-28,
--44,-11,21,
--16,8,-44,
--39,-55,-43,
--11,-35,26,
--9,0,-34,
--8,121,-81,
-7,-16,-22,
--37,33,-31,
--27,-7,-36,
--34,70,-57,
--37,-11,-48,
--40,17,-1,
--33,6,-6,
--9,0,-20,
--21,69,-33,
--29,33,-31,
--55,12,-1,
--33,27,-22,
--50,-33,-47,
--50,54,51,
--1,-5,-44,
--4,22,-40,
--39,-66,-25,
--33,1,-26,
--24,-23,-25,
--11,21,-45,
--25,-45,-19,
--43,105,-16,
-5,-21,1,
--16,11,-33,
--13,-99,-4,
--37,33,-15,
--25,37,-63,
--36,24,-31,
--53,-56,-38,
--41,-4,4,
--33,13,-30,
-49,52,-94,
--5,-30,-15,
-1,38,-40,
--23,12,-36,
--17,40,-47,
--37,-41,-39,
--49,34,0,
--18,-7,-4,
--16,17,-27,
-30,5,-62,
-4,48,-68,
--43,11,-11,
--18,19,-15,
--23,-62,-39,
--42,10,-2,
--21,-13,-13,
--9,13,-47,
--23,-62,-24,
--44,60,-21,
--18,-3,-52,
--22,22,-36,
--75,57,16,
--19,3,10,
--29,23,-38,
--5,-62,-51,
--51,40,-18,
--42,13,-24,
--34,14,-20,
--56,-75,-26,
--26,32,15,
--26,17,-29,
--7,28,-52,
--12,-30,5,
--5,-48,-5,
-2,2,-43,
-21,16,16,
--25,-45,-32,
--43,18,-10,
-9,0,-1,
--1,7,-30,
-19,-48,-4,
--28,25,-29,
--22,0,-31,
--32,17,-10,
--64,-41,-62,
--52,15,16,
--30,-22,-32,
--7,9,-38};
+const signed char gain_cdbk_nb[512] = {
+-32, -32, -32, 0,
+-28, -67, -5, 33,
+-42, -6, -32, 18,
+-57, -10, -54, 35,
+-16, 27, -41, 42,
+19, -19, -40, 36,
+-45, 24, -21, 40,
+-8, -14, -18, 28,
+1, 14, -58, 53,
+-18, -88, -39, 39,
+-38, 21, -18, 37,
+-19, 20, -43, 38,
+10, 17, -48, 54,
+-52, -58, -13, 33,
+-44, -1, -11, 32,
+-12, -11, -34, 22,
+14, 0, -46, 46,
+-37, -35, -34, 5,
+-25, 44, -30, 43,
+6, -4, -63, 49,
+-31, 43, -41, 43,
+-23, 30, -43, 41,
+-43, 26, -14, 44,
+-33, 1, -13, 27,
+-13, 18, -37, 37,
+-46, -73, -45, 34,
+-36, 24, -25, 34,
+-36, -11, -20, 19,
+-25, 12, -18, 33,
+-36, -69, -59, 34,
+-45, 6, 8, 46,
+-22, -14, -24, 18,
+-1, 13, -44, 44,
+-39, -48, -26, 15,
+-32, 31, -37, 34,
+-33, 15, -46, 31,
+-24, 30, -36, 37,
+-41, 31, -23, 41,
+-50, 22, -4, 50,
+-22, 2, -21, 28,
+-17, 30, -34, 40,
+-7, -60, -28, 29,
+-38, 42, -28, 42,
+-44, -11, 21, 43,
+-16, 8, -44, 34,
+-39, -55, -43, 21,
+-11, -35, 26, 41,
+-9, 0, -34, 29,
+-8, 121, -81, 113,
+7, -16, -22, 33,
+-37, 33, -31, 36,
+-27, -7, -36, 17,
+-34, 70, -57, 65,
+-37, -11, -48, 21,
+-40, 17, -1, 44,
+-33, 6, -6, 33,
+-9, 0, -20, 34,
+-21, 69, -33, 57,
+-29, 33, -31, 35,
+-55, 12, -1, 49,
+-33, 27, -22, 35,
+-50, -33, -47, 17,
+-50, 54, 51, 94,
+-1, -5, -44, 35,
+-4, 22, -40, 45,
+-39, -66, -25, 24,
+-33, 1, -26, 20,
+-24, -23, -25, 12,
+-11, 21, -45, 44,
+-25, -45, -19, 17,
+-43, 105, -16, 82,
+5, -21, 1, 41,
+-16, 11, -33, 30,
+-13, -99, -4, 57,
+-37, 33, -15, 44,
+-25, 37, -63, 54,
+-36, 24, -31, 31,
+-53, -56, -38, 26,
+-41, -4, 4, 37,
+-33, 13, -30, 24,
+49, 52, -94, 114,
+-5, -30, -15, 23,
+1, 38, -40, 56,
+-23, 12, -36, 29,
+-17, 40, -47, 51,
+-37, -41, -39, 11,
+-49, 34, 0, 58,
+-18, -7, -4, 34,
+-16, 17, -27, 35,
+30, 5, -62, 65,
+4, 48, -68, 76,
+-43, 11, -11, 38,
+-18, 19, -15, 41,
+-23, -62, -39, 23,
+-42, 10, -2, 41,
+-21, -13, -13, 25,
+-9, 13, -47, 42,
+-23, -62, -24, 24,
+-44, 60, -21, 58,
+-18, -3, -52, 32,
+-22, 22, -36, 34,
+-75, 57, 16, 90,
+-19, 3, 10, 45,
+-29, 23, -38, 32,
+-5, -62, -51, 38,
+-51, 40, -18, 53,
+-42, 13, -24, 32,
+-34, 14, -20, 30,
+-56, -75, -26, 37,
+-26, 32, 15, 59,
+-26, 17, -29, 29,
+-7, 28, -52, 53,
+-12, -30, 5, 30,
+-5, -48, -5, 35,
+2, 2, -43, 40,
+21, 16, 16, 75,
+-25, -45, -32, 10,
+-43, 18, -10, 42,
+9, 0, -1, 52,
+-1, 7, -30, 36,
+19, -48, -4, 48,
+-28, 25, -29, 32,
+-22, 0, -31, 22,
+-32, 17, -10, 36,
+-64, -41, -62, 36,
+-52, 15, 16, 58,
+-30, -22, -32, 6,
+-7, 9, -38, 36};
diff --git a/pjmedia/src/pjmedia-codec/speex/gain_table_lbr.c b/pjmedia/src/pjmedia-codec/speex/gain_table_lbr.c
index 24357f0..3c1c3db 100644
--- a/pjmedia/src/pjmedia-codec/speex/gain_table_lbr.c
+++ b/pjmedia/src/pjmedia-codec/speex/gain_table_lbr.c
@@ -29,36 +29,36 @@
POSSIBILITY OF SUCH DAMAGE.
*/
-const signed char gain_cdbk_lbr[96] = {
--32,-32,-32,
--31,-58,-16,
--41,-24,-43,
--56,-22,-55,
--13,33,-41,
--4,-39,-9,
--41,15,-12,
--8,-15,-12,
-1,2,-44,
--22,-66,-42,
--38,28,-23,
--21,14,-37,
-0,21,-50,
--53,-71,-27,
--37,-1,-19,
--19,-5,-28,
-6,65,-44,
--33,-48,-33,
--40,57,-14,
--17,4,-45,
--31,38,-33,
--23,28,-40,
--43,29,-12,
--34,13,-23,
--16,15,-27,
--14,-82,-15,
--31,25,-32,
--21,5,-5,
--47,-63,-51,
--46,12,3,
--28,-17,-29,
--10,14,-40};
+const signed char gain_cdbk_lbr[128] = {
+-32, -32, -32, 0,
+-31, -58, -16, 22,
+-41, -24, -43, 14,
+-56, -22, -55, 29,
+-13, 33, -41, 47,
+-4, -39, -9, 29,
+-41, 15, -12, 38,
+-8, -15, -12, 31,
+1, 2, -44, 40,
+-22, -66, -42, 27,
+-38, 28, -23, 38,
+-21, 14, -37, 31,
+0, 21, -50, 52,
+-53, -71, -27, 33,
+-37, -1, -19, 25,
+-19, -5, -28, 22,
+6, 65, -44, 74,
+-33, -48, -33, 9,
+-40, 57, -14, 58,
+-17, 4, -45, 32,
+-31, 38, -33, 36,
+-23, 28, -40, 39,
+-43, 29, -12, 46,
+-34, 13, -23, 28,
+-16, 15, -27, 34,
+-14, -82, -15, 43,
+-31, 25, -32, 29,
+-21, 5, -5, 38,
+-47, -63, -51, 33,
+-46, 12, 3, 47,
+-28, -17, -29, 11,
+-10, 14, -40, 38};
diff --git a/pjmedia/src/pjmedia-codec/speex/jitter.c b/pjmedia/src/pjmedia-codec/speex/jitter.c
index 2860c10..a4c0751 100644
--- a/pjmedia/src/pjmedia-codec/speex/jitter.c
+++ b/pjmedia/src/pjmedia-codec/speex/jitter.c
@@ -36,9 +36,6 @@
#include "config.h"
#endif
-#ifndef NULL
-#define NULL 0
-#endif
#include "misc.h"
#include <speex/speex.h>
@@ -46,124 +43,162 @@
#include <speex/speex_jitter.h>
#include <stdio.h>
-#define LATE_BINS 4
+#define LATE_BINS 10
+#define MAX_MARGIN 30 /**< Number of bins in margin histogram */
-void speex_jitter_init(SpeexJitter *jitter, void *decoder, int sampling_rate)
+#define SPEEX_JITTER_MAX_BUFFER_SIZE 200 /**< Maximum number of packets in jitter buffer */
+
+
+
+#define GT32(a,b) (((spx_int32_t)((a)-(b)))>0)
+#define GE32(a,b) (((spx_int32_t)((a)-(b)))>=0)
+#define LT32(a,b) (((spx_int32_t)((a)-(b)))<0)
+#define LE32(a,b) (((spx_int32_t)((a)-(b)))<=0)
+
+/** Jitter buffer structure */
+struct JitterBuffer_ {
+ spx_uint32_t pointer_timestamp; /**< Timestamp of what we will *get* next */
+ spx_uint32_t current_timestamp; /**< Timestamp of the local clock (what we will *play* next) */
+
+ char *buf[SPEEX_JITTER_MAX_BUFFER_SIZE]; /**< Buffer of packets (NULL if slot is free) */
+ spx_uint32_t timestamp[SPEEX_JITTER_MAX_BUFFER_SIZE]; /**< Timestamp of packet */
+ int span[SPEEX_JITTER_MAX_BUFFER_SIZE]; /**< Timestamp of packet */
+ int len[SPEEX_JITTER_MAX_BUFFER_SIZE]; /**< Number of bytes in packet */
+
+ int tick_size; /**< Output granularity */
+ int reset_state; /**< True if state was just reset */
+ int buffer_margin; /**< How many frames we want to keep in the buffer (lower bound) */
+
+ int lost_count; /**< Number of consecutive lost packets */
+ float shortterm_margin[MAX_MARGIN]; /**< Short term margin histogram */
+ float longterm_margin[MAX_MARGIN]; /**< Long term margin histogram */
+ float loss_rate; /**< Average loss rate */
+};
+
+/** Initialise jitter buffer */
+JitterBuffer *jitter_buffer_init(int tick)
+{
+ JitterBuffer *jitter = speex_alloc(sizeof(JitterBuffer));
+ if (jitter)
+ {
+ int i;
+ for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
+ jitter->buf[i]=NULL;
+ jitter->tick_size = tick;
+ jitter->buffer_margin = 1;
+ jitter_buffer_reset(jitter);
+ }
+ return jitter;
+}
+
+/** Reset jitter buffer */
+void jitter_buffer_reset(JitterBuffer *jitter)
{
int i;
for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
{
- jitter->len[i]=-1;
- jitter->timestamp[i]=-1;
+ if (jitter->buf[i])
+ {
+ speex_free(jitter->buf[i]);
+ jitter->buf[i] = NULL;
+ }
}
-
- jitter->dec = decoder;
- speex_decoder_ctl(decoder, SPEEX_GET_FRAME_SIZE, &jitter->frame_size);
- jitter->frame_time = jitter->frame_size;
-
- speex_bits_init(&jitter->current_packet);
- jitter->valid_bits = 0;
-
- jitter->buffer_size = 4;
-
- jitter->pointer_timestamp = -jitter->frame_time * jitter->buffer_size;
+ /* Timestamp is actually undefined at this point */
+ jitter->pointer_timestamp = 0;
+ jitter->current_timestamp = 0;
jitter->reset_state = 1;
jitter->lost_count = 0;
jitter->loss_rate = 0;
+ for (i=0;i<MAX_MARGIN;i++)
+ {
+ jitter->shortterm_margin[i] = 0;
+ jitter->longterm_margin[i] = 0;
+ }
+ /*fprintf (stderr, "reset\n");*/
}
-void speex_jitter_destroy(SpeexJitter *jitter)
+/** Destroy jitter buffer */
+void jitter_buffer_destroy(JitterBuffer *jitter)
{
- speex_bits_destroy(&jitter->current_packet);
+ jitter_buffer_reset(jitter);
+ speex_free(jitter);
}
-
-void speex_jitter_put(SpeexJitter *jitter, char *packet, int len, int timestamp)
+/** Put one packet into the jitter buffer */
+void jitter_buffer_put(JitterBuffer *jitter, const JitterBufferPacket *packet)
{
int i,j;
- int arrival_margin;
-
+ spx_int32_t arrival_margin;
+ /*fprintf (stderr, "put packet %d %d\n", timestamp, span);*/
if (jitter->reset_state)
{
jitter->reset_state=0;
- jitter->pointer_timestamp = timestamp-jitter->frame_time * jitter->buffer_size;
- for (i=0;i<MAX_MARGIN;i++)
- {
- jitter->shortterm_margin[i] = 0;
- jitter->longterm_margin[i] = 0;
- }
- for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
- {
- jitter->len[i]=-1;
- jitter->timestamp[i]=-1;
- }
- fprintf(stderr, "reset to %d\n", timestamp);
+ jitter->pointer_timestamp = packet->timestamp;
+ jitter->current_timestamp = packet->timestamp;
+ /*fprintf(stderr, "reset to %d\n", timestamp);*/
}
/* Cleanup buffer (remove old packets that weren't played) */
for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
{
- if (jitter->timestamp[i]<jitter->pointer_timestamp)
+ if (jitter->buf[i] && LE32(jitter->timestamp[i] + jitter->span[i], jitter->pointer_timestamp))
{
- jitter->len[i]=-1;
- /*if (jitter->timestamp[i] != -1)
- fprintf (stderr, "discarding %d %d\n", jitter->timestamp[i], jitter->pointer_timestamp);*/
+ /*fprintf (stderr, "cleaned (not played)\n");*/
+ speex_free(jitter->buf[i]);
+ jitter->buf[i] = NULL;
}
}
/*Find an empty slot in the buffer*/
for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
{
- if (jitter->len[i]==-1)
+ if (jitter->buf[i]==NULL)
break;
}
/*fprintf(stderr, "%d %d %f\n", timestamp, jitter->pointer_timestamp, jitter->drift_average);*/
+ /*No place left in the buffer*/
if (i==SPEEX_JITTER_MAX_BUFFER_SIZE)
{
int earliest=jitter->timestamp[0];
i=0;
for (j=1;j<SPEEX_JITTER_MAX_BUFFER_SIZE;j++)
{
- if (jitter->timestamp[j]<earliest)
+ if (!jitter->buf[i] || LT32(jitter->timestamp[j],earliest))
{
earliest = jitter->timestamp[j];
i=j;
}
}
- /*fprintf (stderr, "Buffer is full, discarding earliest frame %d (currently at %d)\n", timestamp, jitter->pointer_timestamp);*/
- /*No place left in the buffer*/
-
- /*skip some frame(s) */
- /*return;*/
+ speex_free(jitter->buf[i]);
+ jitter->buf[i]=NULL;
+ if (jitter->lost_count>20)
+ {
+ jitter_buffer_reset(jitter);
+ }
+ /*fprintf (stderr, "Buffer is full, discarding earliest frame %d (currently at %d)\n", timestamp, jitter->pointer_timestamp);*/
}
/* Copy packet in buffer */
- if (len>SPEEX_JITTER_MAX_PACKET_SIZE)
- len=SPEEX_JITTER_MAX_PACKET_SIZE;
- for (j=0;j<len/BYTES_PER_CHAR;j++)
- jitter->buf[i][j]=packet[j];
- jitter->timestamp[i]=timestamp;
- jitter->len[i]=len;
+ jitter->buf[i]=speex_alloc(packet->len);
+ for (j=0;j<packet->len;j++)
+ jitter->buf[i][j]=packet->data[j];
+ jitter->timestamp[i]=packet->timestamp;
+ jitter->span[i]=packet->span;
+ jitter->len[i]=packet->len;
- /* Don't count late packets when adjusting the synchro (we're taking care of them elsewhere) */
- /*if (timestamp <= jitter->pointer_timestamp)
- {
- fprintf (stderr, "frame for timestamp %d arrived too late (at time %d)\n", timestamp, jitter->pointer_timestamp);
- }*/
-
/* Adjust the buffer size depending on network conditions */
- arrival_margin = (timestamp - jitter->pointer_timestamp - jitter->frame_time);
+ arrival_margin = (packet->timestamp - jitter->current_timestamp) - jitter->buffer_margin*jitter->tick_size;
- if (arrival_margin >= -LATE_BINS*jitter->frame_time)
+ if (arrival_margin >= -LATE_BINS*jitter->tick_size)
{
- int int_margin;
+ spx_int32_t int_margin;
for (i=0;i<MAX_MARGIN;i++)
{
jitter->shortterm_margin[i] *= .98;
jitter->longterm_margin[i] *= .995;
}
- int_margin = (arrival_margin + LATE_BINS*jitter->frame_time)/jitter->frame_time;
+ int_margin = LATE_BINS + arrival_margin/jitter->tick_size;
if (int_margin>MAX_MARGIN-1)
int_margin = MAX_MARGIN-1;
if (int_margin>=0)
@@ -171,21 +206,52 @@
jitter->shortterm_margin[int_margin] += .02;
jitter->longterm_margin[int_margin] += .005;
}
+ } else {
+
+ /*fprintf (stderr, "way too late = %d\n", arrival_margin);*/
+ if (jitter->lost_count>20)
+ {
+ jitter_buffer_reset(jitter);
+ }
}
-
- /*fprintf (stderr, "margin : %d %d %f %f %f %f\n", arrival_margin, jitter->buffer_size, 100*jitter->loss_rate, 100*jitter->late_ratio, 100*jitter->ontime_ratio, 100*jitter->early_ratio);*/
+#if 0 /* Enable to check how much is being buffered */
+ if (rand()%1000==0)
+ {
+ int count = 0;
+ for (j=0;j<SPEEX_JITTER_MAX_BUFFER_SIZE;j++)
+ {
+ if (jitter->buf[j])
+ count++;
+ }
+ fprintf (stderr, "buffer_size = %d\n", count);
+ }
+#endif
}
-void speex_jitter_get(SpeexJitter *jitter, short *out, int *current_timestamp)
+/** Get one packet from the jitter buffer */
+int jitter_buffer_get(JitterBuffer *jitter, JitterBufferPacket *packet, spx_uint32_t *start_offset)
{
- int i;
- int ret;
+ int i, j;
float late_ratio_short;
float late_ratio_long;
float ontime_ratio_short;
float ontime_ratio_long;
float early_ratio_short;
float early_ratio_long;
+ int chunk_size;
+ int incomplete = 0;
+
+ if (LT32(jitter->current_timestamp+jitter->tick_size, jitter->pointer_timestamp))
+ {
+ jitter->current_timestamp = jitter->pointer_timestamp;
+ speex_warning("did you forget to call jitter_buffer_tick() by any chance?");
+ }
+ /*fprintf (stderr, "get packet %d %d\n", jitter->pointer_timestamp, jitter->current_timestamp);*/
+
+ /* FIXME: This should be only what remaining of the current tick */
+ chunk_size = jitter->tick_size;
+
+ /* Compiling arrival statistics */
late_ratio_short = 0;
late_ratio_long = 0;
@@ -204,12 +270,15 @@
}
if (0&&jitter->pointer_timestamp%1000==0)
{
- fprintf (stderr, "%f %f %f %f %f %f\n", early_ratio_short, early_ratio_long, ontime_ratio_short, ontime_ratio_long, late_ratio_short, late_ratio_long);
+ /*fprintf (stderr, "%f %f %f %f %f %f\n", early_ratio_short, early_ratio_long, ontime_ratio_short, ontime_ratio_long, late_ratio_short, late_ratio_long);*/
/*fprintf (stderr, "%f %f\n", early_ratio_short + ontime_ratio_short + late_ratio_short, early_ratio_long + ontime_ratio_long + late_ratio_long);*/
}
+ /* Adjusting the buffering */
+
if (late_ratio_short > .1 || late_ratio_long > .03)
{
+ /* If too many packets are arriving late */
jitter->shortterm_margin[MAX_MARGIN-1] += jitter->shortterm_margin[MAX_MARGIN-2];
jitter->longterm_margin[MAX_MARGIN-1] += jitter->longterm_margin[MAX_MARGIN-2];
for (i=MAX_MARGIN-3;i>=0;i--)
@@ -219,18 +288,13 @@
}
jitter->shortterm_margin[0] = 0;
jitter->longterm_margin[0] = 0;
- /*fprintf (stderr, "interpolate frame\n");*/
- speex_decode_int(jitter->dec, NULL, (spx_int16_t*)out);
- if (current_timestamp)
- *current_timestamp = jitter->pointer_timestamp;
- return;
- }
-
- /* Increment timestamp */
- jitter->pointer_timestamp += jitter->frame_time;
-
- if (late_ratio_short + ontime_ratio_short < .005 && late_ratio_long + ontime_ratio_long < .01 && early_ratio_short > .8)
+ jitter->pointer_timestamp -= jitter->tick_size;
+ jitter->current_timestamp -= jitter->tick_size;
+ /*fprintf (stderr, "i");*/
+ /*fprintf (stderr, "interpolate (getting some slack)\n");*/
+ } else if (late_ratio_short + ontime_ratio_short < .005 && late_ratio_long + ontime_ratio_long < .01 && early_ratio_short > .8)
{
+ /* Many frames arriving early */
jitter->shortterm_margin[0] += jitter->shortterm_margin[1];
jitter->longterm_margin[0] += jitter->longterm_margin[1];
for (i=1;i<MAX_MARGIN-1;i++)
@@ -241,61 +305,191 @@
jitter->shortterm_margin[MAX_MARGIN-1] = 0;
jitter->longterm_margin[MAX_MARGIN-1] = 0;
/*fprintf (stderr, "drop frame\n");*/
- jitter->pointer_timestamp += jitter->frame_time;
- }
-
- if (current_timestamp)
- *current_timestamp = jitter->pointer_timestamp;
-
- /* Send zeros while we fill in the buffer */
- if (jitter->pointer_timestamp<0)
- {
- for (i=0;i<jitter->frame_size;i++)
- out[i]=0;
- return;
+ /*fprintf (stderr, "d");*/
+ jitter->pointer_timestamp += jitter->tick_size;
+ jitter->current_timestamp += jitter->tick_size;
+ /*fprintf (stderr, "dropping packet (getting more aggressive)\n");*/
}
- /* Search the buffer for a packet with the right timestamp */
+ /* Searching for the packet that fits best */
+
+ /* Search the buffer for a packet with the right timestamp and spanning the whole current chunk */
for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
{
- if (jitter->len[i]!=-1 && jitter->timestamp[i]==jitter->pointer_timestamp)
+ if (jitter->buf[i] && jitter->timestamp[i]==jitter->pointer_timestamp && GE32(jitter->timestamp[i]+jitter->span[i],jitter->pointer_timestamp+chunk_size))
break;
}
+ /* If no match, try for an "older" packet that still spans (fully) the current chunk */
if (i==SPEEX_JITTER_MAX_BUFFER_SIZE)
{
- /* No packet found */
- if (jitter->valid_bits)
+ for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
{
- /* Try decoding last received packet */
- ret = speex_decode_int(jitter->dec, &jitter->current_packet, (spx_int16_t*)out);
- if (ret == 0)
+ if (jitter->buf[i] && jitter->timestamp[i]<=jitter->pointer_timestamp && GE32(jitter->timestamp[i]+jitter->span[i],jitter->pointer_timestamp+chunk_size))
+ break;
+ }
+ }
+
+ /* If still no match, try for an "older" packet that spans part of the current chunk */
+ if (i==SPEEX_JITTER_MAX_BUFFER_SIZE)
+ {
+ for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
+ {
+ if (jitter->buf[i] && jitter->timestamp[i]<=jitter->pointer_timestamp && GT32(jitter->timestamp[i]+jitter->span[i],jitter->pointer_timestamp))
+ break;
+ }
+ }
+
+ /* If still no match, try for earliest packet possible */
+ if (i==SPEEX_JITTER_MAX_BUFFER_SIZE)
+ {
+ int found = 0;
+ spx_uint32_t best_time=0;
+ int best_span=0;
+ int besti=0;
+ for (i=0;i<SPEEX_JITTER_MAX_BUFFER_SIZE;i++)
+ {
+ /* check if packet starts within current chunk */
+ if (jitter->buf[i] && LT32(jitter->timestamp[i],jitter->pointer_timestamp+chunk_size) && GE32(jitter->timestamp[i],jitter->pointer_timestamp))
{
- jitter->lost_count = 0;
- return;
- } else {
- jitter->valid_bits = 0;
+ if (!found || LT32(jitter->timestamp[i],best_time) || (jitter->timestamp[i]==best_time && GT32(jitter->span[i],best_span)))
+ {
+ best_time = jitter->timestamp[i];
+ best_span = jitter->span[i];
+ besti = i;
+ found = 1;
+ }
}
}
-
- /*fprintf (stderr, "lost/late frame %d\n", jitter->pointer_timestamp);*/
- /*Packet is late or lost*/
- speex_decode_int(jitter->dec, NULL, (spx_int16_t*)out);
- jitter->lost_count++;
- if (jitter->lost_count>=25)
+ if (found)
{
- jitter->lost_count = 0;
- jitter->reset_state = 1;
- speex_decoder_ctl(jitter->dec, SPEEX_RESET_STATE, NULL);
+ i=besti;
+ incomplete = 1;
+ /*fprintf (stderr, "incomplete: %d %d %d %d\n", jitter->timestamp[i], jitter->pointer_timestamp, chunk_size, jitter->span[i]);*/
}
- jitter->loss_rate = .999*jitter->loss_rate + .001;
- } else {
+ }
+
+ /* If we find something */
+ if (i!=SPEEX_JITTER_MAX_BUFFER_SIZE)
+ {
+ /* We (obviously) haven't lost this packet */
jitter->lost_count = 0;
- /* Found the right packet */
- speex_bits_read_from(&jitter->current_packet, jitter->buf[i], jitter->len[i]);
- jitter->len[i]=-1;
+ jitter->loss_rate = .999*jitter->loss_rate;
+ /* Check for potential overflow */
+ packet->len = jitter->len[i];
+ /* Copy packet */
+ for (j=0;j<packet->len;j++)
+ packet->data[j] = jitter->buf[i][j];
+ /* Remove packet */
+ speex_free(jitter->buf[i]);
+ jitter->buf[i] = NULL;
+ /* Set timestamp and span (if requested) */
+ if (start_offset)
+ *start_offset = jitter->timestamp[i]-jitter->pointer_timestamp;
+ packet->timestamp = jitter->timestamp[i];
+ packet->span = jitter->span[i];
+ /* Point at the end of the current packet */
+ jitter->pointer_timestamp = jitter->timestamp[i]+jitter->span[i];
+ if (incomplete)
+ return JITTER_BUFFER_INCOMPLETE;
+ else
+ return JITTER_BUFFER_OK;
+ }
+
+
+ /* If we haven't found anything worth returning */
+ /*fprintf (stderr, "not found\n");*/
+ jitter->lost_count++;
+ /*fprintf (stderr, "m");*/
+ /*fprintf (stderr, "lost_count = %d\n", jitter->lost_count);*/
+ jitter->loss_rate = .999*jitter->loss_rate + .001;
+ if (start_offset)
+ *start_offset = 0;
+ packet->timestamp = jitter->pointer_timestamp;
+ packet->span = jitter->tick_size;
+ jitter->pointer_timestamp += chunk_size;
+ packet->len = 0;
+ return JITTER_BUFFER_MISSING;
+
+}
+
+/** Get pointer timestamp of jitter buffer */
+int jitter_buffer_get_pointer_timestamp(JitterBuffer *jitter)
+{
+ return jitter->pointer_timestamp;
+}
+
+void jitter_buffer_tick(JitterBuffer *jitter)
+{
+ jitter->current_timestamp += jitter->tick_size;
+}
+
+
+
+
+
+void speex_jitter_init(SpeexJitter *jitter, void *decoder, int sampling_rate)
+{
+ jitter->dec = decoder;
+ speex_decoder_ctl(decoder, SPEEX_GET_FRAME_SIZE, &jitter->frame_size);
+
+ jitter->packets = jitter_buffer_init(jitter->frame_size);
+
+ speex_bits_init(&jitter->current_packet);
+ jitter->valid_bits = 0;
+
+}
+
+void speex_jitter_destroy(SpeexJitter *jitter)
+{
+ jitter_buffer_destroy(jitter->packets);
+ speex_bits_destroy(&jitter->current_packet);
+}
+
+void speex_jitter_put(SpeexJitter *jitter, char *packet, int len, int timestamp)
+{
+ JitterBufferPacket p;
+ p.data = packet;
+ p.len = len;
+ p.timestamp = timestamp;
+ p.span = jitter->frame_size;
+ jitter_buffer_put(jitter->packets, &p);
+}
+
+void speex_jitter_get(SpeexJitter *jitter, short *out, int *current_timestamp)
+{
+ int i;
+ int ret;
+ char data[2048];
+ JitterBufferPacket packet;
+ packet.data = data;
+
+ if (jitter->valid_bits)
+ {
+ /* Try decoding last received packet */
+ ret = speex_decode_int(jitter->dec, &jitter->current_packet, out);
+ if (ret == 0)
+ {
+ jitter_buffer_tick(jitter->packets);
+ return;
+ } else {
+ jitter->valid_bits = 0;
+ }
+ }
+
+ ret = jitter_buffer_get(jitter->packets, &packet, NULL);
+
+ if (ret != JITTER_BUFFER_OK)
+ {
+ /* No packet found */
+
+ /*fprintf (stderr, "lost/late frame\n");*/
+ /*Packet is late or lost*/
+ speex_decode_int(jitter->dec, NULL, out);
+ } else {
+ speex_bits_read_from(&jitter->current_packet, packet.data, packet.len);
/* Decode packet */
- ret = speex_decode_int(jitter->dec, &jitter->current_packet, (spx_int16_t*)out);
+ ret = speex_decode_int(jitter->dec, &jitter->current_packet, out);
if (ret == 0)
{
jitter->valid_bits = 1;
@@ -304,13 +498,11 @@
for (i=0;i<jitter->frame_size;i++)
out[i]=0;
}
- jitter->loss_rate = .999*jitter->loss_rate;
}
-
-
+ jitter_buffer_tick(jitter->packets);
}
int speex_jitter_get_pointer_timestamp(SpeexJitter *jitter)
{
- return jitter->pointer_timestamp;
+ return jitter_buffer_get_pointer_timestamp(jitter->packets);
}
diff --git a/pjmedia/src/pjmedia-codec/speex/kiss_fft.c b/pjmedia/src/pjmedia-codec/speex/kiss_fft.c
index bea55ee..a0b3724 100644
--- a/pjmedia/src/pjmedia-codec/speex/kiss_fft.c
+++ b/pjmedia/src/pjmedia-codec/speex/kiss_fft.c
@@ -32,7 +32,7 @@
#define CHECKBUF(buf,nbuf,n) \
do { \
if ( nbuf < (size_t)(n) ) {\
- free(buf); \
+ speex_free(buf); \
buf = (kiss_fft_cpx*)KISS_FFT_MALLOC(sizeof(kiss_fft_cpx)*(n)); \
nbuf = (size_t)(n); \
} \
@@ -87,7 +87,7 @@
if (!st->inverse) {
int i;
kiss_fft_cpx *x=Fout;
- for (i=0;i<(int)(4*m);i++)
+ for (i=0;i<4*m;i++)
{
x[i].r = PSHR16(x[i].r,2);
x[i].i = PSHR16(x[i].i,2);
@@ -404,7 +404,7 @@
if (fin == fout) {
CHECKBUF(tmpbuf,ntmpbuf,st->nfft);
kf_work(tmpbuf,fin,1,in_stride, st->factors,st);
- memcpy(fout,tmpbuf,sizeof(kiss_fft_cpx)*st->nfft);
+ speex_move(fout,tmpbuf,sizeof(kiss_fft_cpx)*st->nfft);
}else{
kf_work( fout, fin, 1,in_stride, st->factors,st );
}
@@ -421,10 +421,10 @@
*/
void kiss_fft_cleanup(void)
{
- free(scratchbuf);
+ speex_free(scratchbuf);
scratchbuf = NULL;
nscratchbuf=0;
- free(tmpbuf);
+ speex_free(tmpbuf);
tmpbuf=NULL;
ntmpbuf=0;
}
diff --git a/pjmedia/src/pjmedia-codec/speex/kiss_fft.h b/pjmedia/src/pjmedia-codec/speex/kiss_fft.h
index d07b78b..54627e7 100644
--- a/pjmedia/src/pjmedia-codec/speex/kiss_fft.h
+++ b/pjmedia/src/pjmedia-codec/speex/kiss_fft.h
@@ -2,11 +2,8 @@
#define KISS_FFT_H
#include <stdlib.h>
-#include <stdio.h>
#include <math.h>
-#include <memory.h>
-//Not available in gcc MacOS X (bennylp)
-//#include <malloc.h>
+#include "misc.h"
#ifdef __cplusplus
extern "C" {
@@ -30,13 +27,13 @@
# define kiss_fft_scalar __m128
#define KISS_FFT_MALLOC(nbytes) memalign(16,nbytes)
#else
-#define KISS_FFT_MALLOC malloc
+#define KISS_FFT_MALLOC speex_alloc
#endif
#ifdef FIXED_POINT
-#include <sys/types.h>
-# define kiss_fft_scalar int16_t
+#include "misc.h"
+# define kiss_fft_scalar spx_int16_t
#else
# ifndef kiss_fft_scalar
/* default is float */
@@ -95,7 +92,7 @@
/* If kiss_fft_alloc allocated a buffer, it is one contiguous
buffer and can be simply free()d when no longer needed*/
-#define kiss_fft_free free
+#define kiss_fft_free speex_free
/*
Cleans up some memory that gets managed internally. Not necessary to call, but it might clean up
diff --git a/pjmedia/src/pjmedia-codec/speex/kiss_fftr.c b/pjmedia/src/pjmedia-codec/speex/kiss_fftr.c
index 3ac4db9..b90b725 100644
--- a/pjmedia/src/pjmedia-codec/speex/kiss_fftr.c
+++ b/pjmedia/src/pjmedia-codec/speex/kiss_fftr.c
@@ -35,7 +35,7 @@
size_t subsize, memneeded;
if (nfft & 1) {
- fprintf(stderr,"Real FFT optimization must be even.\n");
+ speex_warning("Real FFT optimization must be even.\n");
return NULL;
}
nfft >>= 1;
@@ -75,7 +75,7 @@
kiss_fft_cpx fpnk,fpk,f1k,f2k,tw,tdc;
if ( st->substate->inverse) {
- fprintf(stderr,"kiss fft usage error: improper alloc\n");
+ speex_warning("kiss fft usage error: improper alloc\n");
exit(1);
}
@@ -130,7 +130,7 @@
int k, ncfft;
if (st->substate->inverse == 0) {
- fprintf (stderr, "kiss fft usage error: improper alloc\n");
+ speex_warning ("kiss fft usage error: improper alloc\n");
exit (1);
}
diff --git a/pjmedia/src/pjmedia-codec/speex/kiss_fftr.h b/pjmedia/src/pjmedia-codec/speex/kiss_fftr.h
index 72e5a57..2e8351a 100644
--- a/pjmedia/src/pjmedia-codec/speex/kiss_fftr.h
+++ b/pjmedia/src/pjmedia-codec/speex/kiss_fftr.h
@@ -38,7 +38,7 @@
output timedata has nfft scalar points
*/
-#define kiss_fftr_free free
+#define kiss_fftr_free speex_free
#ifdef __cplusplus
}
diff --git a/pjmedia/src/pjmedia-codec/speex/lpc.c b/pjmedia/src/pjmedia-codec/speex/lpc.c
index c465fae..fd5d382 100644
--- a/pjmedia/src/pjmedia-codec/speex/lpc.c
+++ b/pjmedia/src/pjmedia-codec/speex/lpc.c
@@ -94,7 +94,7 @@
for (j = 0; j < i; j++)
rr = SUB32(rr,MULT16_16(lpc[j],ac[i - j]));
#ifdef FIXED_POINT
- r = DIV32_16(rr,ADD16(error,16));
+ r = DIV32_16(rr+PSHR32(error,1),ADD16(error,8));
#else
r = rr/(error+.003*ac[0]);
#endif
@@ -103,11 +103,11 @@
for (j = 0; j < i>>1; j++)
{
spx_word16_t tmp = lpc[j];
- lpc[j] = MAC16_16_Q13(lpc[j],r,lpc[i-1-j]);
- lpc[i-1-j] = MAC16_16_Q13(lpc[i-1-j],r,tmp);
+ lpc[j] = MAC16_16_P13(lpc[j],r,lpc[i-1-j]);
+ lpc[i-1-j] = MAC16_16_P13(lpc[i-1-j],r,tmp);
}
if (i & 1)
- lpc[j] = MAC16_16_Q13(lpc[j],lpc[j],r);
+ lpc[j] = MAC16_16_P13(lpc[j],lpc[j],r);
error = SUB16(error,MULT16_16_Q13(r,MULT16_16_Q13(error,r)));
}
diff --git a/pjmedia/src/pjmedia-codec/speex/lsp.c b/pjmedia/src/pjmedia-codec/speex/lsp.c
index f4350ae..6e7ea31 100644
--- a/pjmedia/src/pjmedia-codec/speex/lsp.c
+++ b/pjmedia/src/pjmedia-codec/speex/lsp.c
@@ -1,8 +1,6 @@
/*---------------------------------------------------------------------------*\
Original copyright
- FILE........: AKSLSPD.C
- TYPE........: Turbo C
- COMPANY.....: Voicetronix
+ FILE........: lsp.c
AUTHOR......: David Rowe
DATE CREATED: 24/2/93
@@ -44,6 +42,43 @@
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
+/*---------------------------------------------------------------------------*\
+
+ Introduction to Line Spectrum Pairs (LSPs)
+ ------------------------------------------
+
+ LSPs are used to encode the LPC filter coefficients {ak} for
+ transmission over the channel. LSPs have several properties (like
+ less sensitivity to quantisation noise) that make them superior to
+ direct quantisation of {ak}.
+
+ A(z) is a polynomial of order lpcrdr with {ak} as the coefficients.
+
+ A(z) is transformed to P(z) and Q(z) (using a substitution and some
+ algebra), to obtain something like:
+
+ A(z) = 0.5[P(z)(z+z^-1) + Q(z)(z-z^-1)] (1)
+
+ As you can imagine A(z) has complex zeros all over the z-plane. P(z)
+ and Q(z) have the very neat property of only having zeros _on_ the
+ unit circle. So to find them we take a test point z=exp(jw) and
+ evaluate P (exp(jw)) and Q(exp(jw)) using a grid of points between 0
+ and pi.
+
+ The zeros (roots) of P(z) also happen to alternate, which is why we
+ swap coefficients as we find roots. So the process of finding the
+ LSP frequencies is basically finding the roots of 5th order
+ polynomials.
+
+ The root so P(z) and Q(z) occur in symmetrical pairs at +/-w, hence
+ the name Line Spectrum Pairs (LSPs).
+
+ To convert back to ak we just evaluate (1), "clocking" an impulse
+ thru it lpcrdr times gives us the impulse response of A(z) which is
+ {ak}.
+
+\*---------------------------------------------------------------------------*/
+
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
@@ -63,8 +98,6 @@
#ifdef FIXED_POINT
-
-
#define FREQ_SCALE 16384
/*#define ANGLE2X(a) (32768*cos(((a)/8192.)))*/
@@ -73,6 +106,10 @@
/*#define X2ANGLE(x) (acos(.00006103515625*(x))*LSP_SCALING)*/
#define X2ANGLE(x) (spx_acos(x))
+#ifdef BFIN_ASM
+#include "lsp_bfin.h"
+#endif
+
#else
/*#define C1 0.99940307
@@ -88,27 +125,28 @@
/*---------------------------------------------------------------------------*\
- FUNCTION....: cheb_poly_eva()
+ FUNCTION....: cheb_poly_eva()
- AUTHOR......: David Rowe
- DATE CREATED: 24/2/93
+ AUTHOR......: David Rowe
+ DATE CREATED: 24/2/93
- This function evaluates a series of Chebyshev polynomials
+ This function evaluates a series of Chebyshev polynomials
\*---------------------------------------------------------------------------*/
#ifdef FIXED_POINT
-static inline spx_word32_t cheb_poly_eva(spx_word32_t *coef,spx_word16_t x,int m,char *stack)
-/* float coef[] coefficients of the polynomial to be evaluated */
-/* float x the point where polynomial is to be evaluated */
-/* int m order of the polynomial */
+#ifndef OVERRIDE_CHEB_POLY_EVA
+static inline spx_word32_t cheb_poly_eva(
+ spx_word16_t *coef, /* P or Q coefs in Q13 format */
+ spx_word16_t x, /* cos of freq (-1.0 to 1.0) in Q14 format */
+ int m, /* LPC order/2 */
+ char *stack
+)
{
int i;
- VARDECL(spx_word16_t *T);
+ spx_word16_t b0, b1;
spx_word32_t sum;
- int m2=m>>1;
- VARDECL(spx_word16_t *coefn);
/*Prevents overflows*/
if (x>16383)
@@ -116,73 +154,55 @@
if (x<-16383)
x = -16383;
- /* Allocate memory for Chebyshev series formulation */
- ALLOC(T, m2+1, spx_word16_t);
- ALLOC(coefn, m2+1, spx_word16_t);
-
- for (i=0;i<m2+1;i++)
- {
- coefn[i] = coef[i];
- /*printf ("%f ", coef[i]);*/
- }
- /*printf ("\n");*/
-
/* Initialise values */
- T[0]=16384;
- T[1]=x;
+ b1=16384;
+ b0=x;
- /* Evaluate Chebyshev series formulation using iterative approach */
- /* Evaluate polynomial and return value also free memory space */
- sum = ADD32(EXTEND32(coefn[m2]), EXTEND32(MULT16_16_P14(coefn[m2-1],x)));
- /*x *= 2;*/
- for(i=2;i<=m2;i++)
+ /* Evaluate Chebyshev series formulation usin g iterative approach */
+ sum = ADD32(EXTEND32(coef[m]), EXTEND32(MULT16_16_P14(coef[m-1],x)));
+ for(i=2;i<=m;i++)
{
- T[i] = SUB16(MULT16_16_Q13(x,T[i-1]), T[i-2]);
- sum = ADD32(sum, EXTEND32(MULT16_16_P14(coefn[m2-i],T[i])));
- /*printf ("%f ", sum);*/
- }
-
- /*printf ("\n");*/
- return sum;
-}
-#else
-static float cheb_poly_eva(spx_word32_t *coef,float x,int m,char *stack)
-/* float coef[] coefficients of the polynomial to be evaluated */
-/* float x the point where polynomial is to be evaluated */
-/* int m order of the polynomial */
-{
- int i;
- VARDECL(float *T);
- float sum;
- int m2=m>>1;
-
- /* Allocate memory for Chebyshev series formulation */
- ALLOC(T, m2+1, float);
-
- /* Initialise values */
- T[0]=1;
- T[1]=x;
-
- /* Evaluate Chebyshev series formulation using iterative approach */
- /* Evaluate polynomial and return value also free memory space */
- sum = coef[m2] + coef[m2-1]*x;
- x *= 2;
- for(i=2;i<=m2;i++)
- {
- T[i] = x*T[i-1] - T[i-2];
- sum += coef[m2-i] * T[i];
+ spx_word16_t tmp=b0;
+ b0 = SUB16(MULT16_16_Q13(x,b0), b1);
+ b1 = tmp;
+ sum = ADD32(sum, EXTEND32(MULT16_16_P14(coef[m-i],b0)));
}
return sum;
}
#endif
+#else
+
+static float cheb_poly_eva(spx_word32_t *coef, spx_word16_t x, int m, char *stack)
+{
+ int k;
+ float b0, b1, tmp;
+
+ /* Initial conditions */
+ b0=0; /* b_(m+1) */
+ b1=0; /* b_(m+2) */
+
+ x*=2;
+
+ /* Calculate the b_(k) */
+ for(k=m;k>0;k--)
+ {
+ tmp=b0; /* tmp holds the previous value of b0 */
+ b0=x*b0-b1+coef[m-k]; /* b0 holds its new value based on b0 and b1 */
+ b1=tmp; /* b1 holds the previous value of b0 */
+ }
+
+ return(-b1+.5*x*b0+coef[m]);
+}
+#endif
+
/*---------------------------------------------------------------------------*\
- FUNCTION....: lpc_to_lsp()
+ FUNCTION....: lpc_to_lsp()
- AUTHOR......: David Rowe
- DATE CREATED: 24/2/93
+ AUTHOR......: David Rowe
+ DATE CREATED: 24/2/93
This function converts LPC coefficients to LSP
coefficients.
@@ -210,11 +230,13 @@
int i,j,m,flag,k;
VARDECL(spx_word32_t *Q); /* ptrs for memory allocation */
VARDECL(spx_word32_t *P);
+ VARDECL(spx_word16_t *Q16); /* ptrs for memory allocation */
+ VARDECL(spx_word16_t *P16);
spx_word32_t *px; /* ptrs of respective P'(z) & Q'(z) */
spx_word32_t *qx;
spx_word32_t *p;
spx_word32_t *q;
- spx_word32_t *pt; /* ptr used for cheb_poly_eval()
+ spx_word16_t *pt; /* ptr used for cheb_poly_eval()
whether P' or Q' */
int roots=0; /* DR 8/2/94: number of roots found */
flag = 1; /* program is searching for a root when,
@@ -276,20 +298,31 @@
px = P; /* re-initialise ptrs */
qx = Q;
+ /* now that we have computed P and Q convert to 16 bits to
+ speed up cheb_poly_eval */
+
+ ALLOC(P16, m+1, spx_word16_t);
+ ALLOC(Q16, m+1, spx_word16_t);
+
+ for (i=0;i<m+1;i++)
+ {
+ P16[i] = P[i];
+ Q16[i] = Q[i];
+ }
+
/* Search for a zero in P'(z) polynomial first and then alternate to Q'(z).
Keep alternating between the two polynomials as each zero is found */
xr = 0; /* initialise xr to zero */
xl = FREQ_SCALE; /* start at point xl = 1 */
-
for(j=0;j<lpcrdr;j++){
if(j&1) /* determines whether P' or Q' is eval. */
- pt = qx;
+ pt = Q16;
else
- pt = px;
+ pt = P16;
- psuml = cheb_poly_eva(pt,xl,lpcrdr,stack); /* evals poly. at xl */
+ psuml = cheb_poly_eva(pt,xl,m,stack); /* evals poly. at xl */
flag = 1;
while(flag && (xr >= -FREQ_SCALE)){
spx_word16_t dd;
@@ -304,7 +337,7 @@
dd *= .5;
#endif
xr = SUB16(xl, dd); /* interval spacing */
- psumr = cheb_poly_eva(pt,xr,lpcrdr,stack);/* poly(xl-delta_x) */
+ psumr = cheb_poly_eva(pt,xr,m,stack);/* poly(xl-delta_x) */
temp_psumr = psumr;
temp_xr = xr;
@@ -328,7 +361,7 @@
#else
xm = .5*(xl+xr); /* bisect the interval */
#endif
- psumm=cheb_poly_eva(pt,xm,lpcrdr,stack);
+ psumm=cheb_poly_eva(pt,xm,m,stack);
/*if(psumm*psuml>0.)*/
if(!SIGN_CHANGE(psumm,psuml))
{
@@ -354,7 +387,6 @@
return(roots);
}
-
/*---------------------------------------------------------------------------*\
FUNCTION....: lsp_to_lpc()
@@ -362,8 +394,7 @@
AUTHOR......: David Rowe
DATE CREATED: 24/2/93
- lsp_to_lpc: This function converts LSP coefficients to LPC
- coefficients.
+ Converts LSP coefficients to LPC coefficients.
\*---------------------------------------------------------------------------*/
@@ -373,77 +404,119 @@
/* float *freq array of LSP frequencies in the x domain */
/* float *ak array of LPC coefficients */
/* int lpcrdr order of LPC coefficients */
-
-
{
int i,j;
- spx_word32_t xout1,xout2,xin1,xin2;
- VARDECL(spx_word32_t *Wp);
- spx_word32_t *pw,*n1,*n2,*n3,*n4=NULL;
+ spx_word32_t xout1,xout2,xin;
+ spx_word32_t mult, a;
VARDECL(spx_word16_t *freqn);
+ VARDECL(spx_word32_t **xp);
+ VARDECL(spx_word32_t *xpmem);
+ VARDECL(spx_word32_t **xq);
+ VARDECL(spx_word32_t *xqmem);
int m = lpcrdr>>1;
+
+ /*
+ Reconstruct P(z) and Q(z) by cascading second order polynomials
+ in form 1 - 2cos(w)z(-1) + z(-2), where w is the LSP frequency.
+ In the time domain this is:
+
+ y(n) = x(n) - 2cos(w)x(n-1) + x(n-2)
+
+ This is what the ALLOCS below are trying to do:
+
+ int xp[m+1][lpcrdr+1+2]; // P matrix in QIMP
+ int xq[m+1][lpcrdr+1+2]; // Q matrix in QIMP
+
+ These matrices store the output of each stage on each row. The
+ final (m-th) row has the output of the final (m-th) cascaded
+ 2nd order filter. The first row is the impulse input to the
+ system (not written as it is known).
+
+ The version below takes advantage of the fact that a lot of the
+ outputs are zero or known, for example if we put an inpulse
+ into the first section the "clock" it 10 times only the first 3
+ outputs samples are non-zero (it's an FIR filter).
+ */
+
+ ALLOC(xp, (m+1), spx_word32_t*);
+ ALLOC(xpmem, (m+1)*(lpcrdr+1+2), spx_word32_t);
+
+ ALLOC(xq, (m+1), spx_word32_t*);
+ ALLOC(xqmem, (m+1)*(lpcrdr+1+2), spx_word32_t);
+
+ for(i=0; i<=m; i++) {
+ xp[i] = xpmem + i*(lpcrdr+1+2);
+ xq[i] = xqmem + i*(lpcrdr+1+2);
+ }
+
+ /* work out 2cos terms in Q14 */
+
ALLOC(freqn, lpcrdr, spx_word16_t);
- for (i=0;i<lpcrdr;i++)
+ for (i=0;i<lpcrdr;i++)
freqn[i] = ANGLE2X(freq[i]);
- ALLOC(Wp, 4*m+2, spx_word32_t);
- pw = Wp;
+ #define QIMP 21 /* scaling for impulse */
-
- /* initialise contents of array */
-
- for(i=0;i<=4*m+1;i++){ /* set contents of buffer to 0 */
- *pw++ = 0;
+ xin = SHL32(EXTEND32(1), (QIMP-1)); /* 0.5 in QIMP format */
+
+ /* first col and last non-zero values of each row are trivial */
+
+ for(i=0;i<=m;i++) {
+ xp[i][1] = 0;
+ xp[i][2] = xin;
+ xp[i][2+2*i] = xin;
+ xq[i][1] = 0;
+ xq[i][2] = xin;
+ xq[i][2+2*i] = xin;
}
- /* Set pointers up */
+ /* 2nd row (first output row) is trivial */
- pw = Wp;
- xin1 = 1048576;
- xin2 = 1048576;
+ xp[1][3] = -MULT16_32_Q14(freqn[0],xp[0][2]);
+ xq[1][3] = -MULT16_32_Q14(freqn[1],xq[0][2]);
- /* reconstruct P(z) and Q(z) by cascading second order
- polynomials in form 1 - 2xz(-1) +z(-2), where x is the
- LSP coefficient */
+ xout1 = xout2 = 0;
- for(j=0;j<=lpcrdr;j++){
- spx_word16_t *fr=freqn;
- for(i=0;i<m;i++){
- n1 = pw+(i<<2);
- n2 = n1 + 1;
- n3 = n2 + 1;
- n4 = n3 + 1;
- xout1 = ADD32(SUB32(xin1, MULT16_32_Q14(*fr,*n1)), *n2);
- fr++;
- xout2 = ADD32(SUB32(xin2, MULT16_32_Q14(*fr,*n3)), *n4);
- fr++;
- *n2 = *n1;
- *n4 = *n3;
- *n1 = xin1;
- *n3 = xin2;
- xin1 = xout1;
- xin2 = xout2;
- }
- xout1 = xin1 + *(n4+1);
- xout2 = xin2 - *(n4+2);
- /* FIXME: perhaps apply bandwidth expansion in case of overflow? */
- if (j>0)
- {
- if (xout1 + xout2>SHL32(EXTEND32(32766),8))
- ak[j-1] = 32767;
- else if (xout1 + xout2 < -SHL32(EXTEND32(32766),8))
- ak[j-1] = -32767;
- else
- ak[j-1] = EXTRACT16(PSHR32(ADD32(xout1,xout2),8));
- } else {/*speex_warning_int("ak[0] = ", EXTRACT16(PSHR32(ADD32(xout1,xout2),8)));*/}
- *(n4+1) = xin1;
- *(n4+2) = xin2;
+ /* now generate remaining rows */
- xin1 = 0;
- xin2 = 0;
+ for(i=1;i<m;i++) {
+
+ for(j=1;j<2*(i+1)-1;j++) {
+ mult = MULT16_32_Q14(freqn[2*i],xp[i][j+1]);
+ xp[i+1][j+2] = ADD32(SUB32(xp[i][j+2], mult), xp[i][j]);
+ mult = MULT16_32_Q14(freqn[2*i+1],xq[i][j+1]);
+ xq[i+1][j+2] = ADD32(SUB32(xq[i][j+2], mult), xq[i][j]);
+ }
+
+ /* for last col xp[i][j+2] = xq[i][j+2] = 0 */
+
+ mult = MULT16_32_Q14(freqn[2*i],xp[i][j+1]);
+ xp[i+1][j+2] = SUB32(xp[i][j], mult);
+ mult = MULT16_32_Q14(freqn[2*i+1],xq[i][j+1]);
+ xq[i+1][j+2] = SUB32(xq[i][j], mult);
}
+
+ /* process last row to extra a{k} */
+
+ for(j=1;j<=lpcrdr;j++) {
+ int shift = QIMP-13;
+
+ /* final filter sections */
+ a = PSHR32(xp[m][j+2] + xout1 + xq[m][j+2] - xout2, shift);
+ xout1 = xp[m][j+2];
+ xout2 = xq[m][j+2];
+
+ /* hard limit ak's to +/- 32767 */
+
+ if (a < -32767) a = 32767;
+ if (a > 32767) a = 32767;
+ ak[j-1] = (short)a;
+
+ }
+
}
+
#else
void lsp_to_lpc(spx_lsp_t *freq,spx_coef_t *ak,int lpcrdr, char *stack)
diff --git a/pjmedia/src/pjmedia-codec/speex/ltp.c b/pjmedia/src/pjmedia-codec/speex/ltp.c
index 94189c3..9a5a295 100644
--- a/pjmedia/src/pjmedia-codec/speex/ltp.c
+++ b/pjmedia/src/pjmedia-codec/speex/ltp.c
@@ -55,7 +55,7 @@
#endif
#ifndef OVERRIDE_INNER_PROD
-static spx_word32_t inner_prod(const spx_word16_t *x, const spx_word16_t *y, int len)
+spx_word32_t inner_prod(const spx_word16_t *x, const spx_word16_t *y, int len)
{
spx_word32_t sum=0;
len >>= 2;
@@ -75,7 +75,7 @@
#ifndef OVERRIDE_PITCH_XCORR
#if 0 /* HINT: Enable this for machines with enough registers (i.e. not x86) */
-static void pitch_xcorr(const spx_word16_t *_x, const spx_word16_t *_y, spx_word32_t *corr, int len, int nb_pitch, char *stack)
+void pitch_xcorr(const spx_word16_t *_x, const spx_word16_t *_y, spx_word32_t *corr, int len, int nb_pitch, char *stack)
{
int i,j;
for (i=0;i<nb_pitch;i+=4)
@@ -138,7 +138,7 @@
}
#else
-static void pitch_xcorr(const spx_word16_t *_x, const spx_word16_t *_y, spx_word32_t *corr, int len, int nb_pitch, char *stack)
+void pitch_xcorr(const spx_word16_t *_x, const spx_word16_t *_y, spx_word32_t *corr, int len, int nb_pitch, char *stack)
{
int i;
for (i=0;i<nb_pitch;i++)
@@ -152,128 +152,122 @@
#endif
#ifndef OVERRIDE_COMPUTE_PITCH_ERROR
-static inline spx_word32_t compute_pitch_error(spx_word32_t *C, spx_word16_t *g, spx_word16_t pitch_control)
+static inline spx_word32_t compute_pitch_error(spx_word16_t *C, spx_word16_t *g, spx_word16_t pitch_control)
{
spx_word32_t sum = 0;
- sum = ADD32(sum,MULT16_32_Q15(MULT16_16_16(g[0],pitch_control),C[0]));
- sum = ADD32(sum,MULT16_32_Q15(MULT16_16_16(g[1],pitch_control),C[1]));
- sum = ADD32(sum,MULT16_32_Q15(MULT16_16_16(g[2],pitch_control),C[2]));
- sum = SUB32(sum,MULT16_32_Q15(MULT16_16_16(g[0],g[1]),C[3]));
- sum = SUB32(sum,MULT16_32_Q15(MULT16_16_16(g[2],g[1]),C[4]));
- sum = SUB32(sum,MULT16_32_Q15(MULT16_16_16(g[2],g[0]),C[5]));
- sum = SUB32(sum,MULT16_32_Q15(MULT16_16_16(g[0],g[0]),C[6]));
- sum = SUB32(sum,MULT16_32_Q15(MULT16_16_16(g[1],g[1]),C[7]));
- sum = SUB32(sum,MULT16_32_Q15(MULT16_16_16(g[2],g[2]),C[8]));
+ sum = ADD32(sum,MULT16_16(MULT16_16_16(g[0],pitch_control),C[0]));
+ sum = ADD32(sum,MULT16_16(MULT16_16_16(g[1],pitch_control),C[1]));
+ sum = ADD32(sum,MULT16_16(MULT16_16_16(g[2],pitch_control),C[2]));
+ sum = SUB32(sum,MULT16_16(MULT16_16_16(g[0],g[1]),C[3]));
+ sum = SUB32(sum,MULT16_16(MULT16_16_16(g[2],g[1]),C[4]));
+ sum = SUB32(sum,MULT16_16(MULT16_16_16(g[2],g[0]),C[5]));
+ sum = SUB32(sum,MULT16_16(MULT16_16_16(g[0],g[0]),C[6]));
+ sum = SUB32(sum,MULT16_16(MULT16_16_16(g[1],g[1]),C[7]));
+ sum = SUB32(sum,MULT16_16(MULT16_16_16(g[2],g[2]),C[8]));
return sum;
}
#endif
-void open_loop_nbest_pitch(spx_sig_t *sw, int start, int end, int len, int *pitch, spx_word16_t *gain, int N, char *stack)
+#ifndef OVERRIDE_OPEN_LOOP_NBEST_PITCH
+void open_loop_nbest_pitch(spx_word16_t *sw, int start, int end, int len, int *pitch, spx_word16_t *gain, int N, char *stack)
{
int i,j,k;
VARDECL(spx_word32_t *best_score);
+ VARDECL(spx_word32_t *best_ener);
spx_word32_t e0;
VARDECL(spx_word32_t *corr);
VARDECL(spx_word32_t *energy);
- VARDECL(spx_word32_t *score);
- VARDECL(spx_word16_t *swn2);
- spx_word16_t *swn;
ALLOC(best_score, N, spx_word32_t);
+ ALLOC(best_ener, N, spx_word32_t);
ALLOC(corr, end-start+1, spx_word32_t);
ALLOC(energy, end-start+2, spx_word32_t);
- ALLOC(score, end-start+1, spx_word32_t);
-
-#ifdef FIXED_POINT
- ALLOC(swn2, end+len, spx_word16_t);
- normalize16(sw-end, swn2, 16384, end+len);
- swn = swn2 + end;
-#else
- swn = sw;
-#endif
for (i=0;i<N;i++)
{
best_score[i]=-1;
+ best_ener[i]=0;
pitch[i]=start;
}
-
- energy[0]=inner_prod(swn-start, swn-start, len);
- e0=inner_prod(swn, swn, len);
- for (i=start;i<=end;i++)
+ energy[0]=inner_prod(sw-start, sw-start, len);
+ e0=inner_prod(sw, sw, len);
+ for (i=start;i<end;i++)
{
/* Update energy for next pitch*/
- energy[i-start+1] = SUB32(ADD32(energy[i-start],SHR32(MULT16_16(swn[-i-1],swn[-i-1]),6)), SHR32(MULT16_16(swn[-i+len-1],swn[-i+len-1]),6));
+ energy[i-start+1] = SUB32(ADD32(energy[i-start],SHR32(MULT16_16(sw[-i-1],sw[-i-1]),6)), SHR32(MULT16_16(sw[-i+len-1],sw[-i+len-1]),6));
if (energy[i-start+1] < 0)
energy[i-start+1] = 0;
}
- pitch_xcorr(swn, swn-end, corr, len, end-start+1, stack);
+ pitch_xcorr(sw, sw-end, corr, len, end-start+1, stack);
+ /* FIXME: Fixed-point and floating-point code should be merged */
#ifdef FIXED_POINT
{
VARDECL(spx_word16_t *corr16);
VARDECL(spx_word16_t *ener16);
ALLOC(corr16, end-start+1, spx_word16_t);
ALLOC(ener16, end-start+1, spx_word16_t);
- normalize16(corr, corr16, 16384, end-start+1);
- normalize16(energy, ener16, 16384, end-start+1);
+ /* Normalize to 180 so we can square it and it still fits in 16 bits */
+ normalize16(corr, corr16, 180, end-start+1);
+ normalize16(energy, ener16, 180, end-start+1);
for (i=start;i<=end;i++)
{
- spx_word16_t g;
- spx_word32_t tmp;
- tmp = corr16[i-start];
- if (tmp>0)
+ spx_word16_t tmp = MULT16_16_16(corr16[i-start],corr16[i-start]);
+ /* Instead of dividing the tmp by the energy, we multiply on the other side */
+ if (MULT16_16(tmp,best_ener[N-1])>MULT16_16(best_score[N-1],ADD16(1,ener16[i-start])))
{
- if (SHR16(corr16[i-start],4)>ener16[i-start])
- tmp = SHL32(EXTEND32(ener16[i-start]),14);
- else if (-SHR16(corr16[i-start],4)>ener16[i-start])
- tmp = -SHL32(EXTEND32(ener16[i-start]),14);
- else
- tmp = SHL32(tmp,10);
- g = DIV32_16(tmp, 8+ener16[i-start]);
- score[i-start] = MULT16_16(corr16[i-start],g);
- } else
- {
- score[i-start] = 1;
+ /* We can safely put it last and then check */
+ best_score[N-1]=tmp;
+ best_ener[N-1]=ener16[i-start]+1;
+ pitch[N-1]=i;
+ /* Check if it comes in front of others */
+ for (j=0;j<N-1;j++)
+ {
+ if (MULT16_16(tmp,best_ener[j])>MULT16_16(best_score[j],ADD16(1,ener16[i-start])))
+ {
+ for (k=N-1;k>j;k--)
+ {
+ best_score[k]=best_score[k-1];
+ best_ener[k]=best_ener[k-1];
+ pitch[k]=pitch[k-1];
+ }
+ best_score[j]=tmp;
+ best_ener[j]=ener16[i-start]+1;
+ pitch[j]=i;
+ break;
+ }
+ }
}
}
}
#else
for (i=start;i<=end;i++)
{
- float g = corr[i-start]/(1+energy[i-start]);
- if (g>16)
- g = 16;
- else if (g<-16)
- g = -16;
- score[i-start] = g*corr[i-start];
- }
-#endif
-
- /* Extract best scores */
- for (i=start;i<=end;i++)
- {
- if (score[i-start]>best_score[N-1])
+ float tmp = corr[i-start]*corr[i-start];
+ if (tmp*best_ener[N-1]>best_score[N-1]*(1+energy[i-start]))
{
for (j=0;j<N;j++)
{
- if (score[i-start] > best_score[j])
+ if (tmp*best_ener[j]>best_score[j]*(1+energy[i-start]))
{
for (k=N-1;k>j;k--)
{
best_score[k]=best_score[k-1];
+ best_ener[k]=best_ener[k-1];
pitch[k]=pitch[k-1];
}
- best_score[j]=score[i-start];
+ best_score[j]=tmp;
+ best_ener[j]=energy[i-start]+1;
pitch[j]=i;
break;
}
}
}
}
+#endif
/* Compute open-loop gain */
if (gain)
@@ -290,164 +284,131 @@
}
}
}
+#endif
+#ifndef OVERRIDE_PITCH_GAIN_SEARCH_3TAP_VQ
+static int pitch_gain_search_3tap_vq(
+ const signed char *gain_cdbk,
+ int gain_cdbk_size,
+ spx_word16_t *C16,
+ spx_word16_t max_gain
+)
+{
+ const signed char *ptr=gain_cdbk;
+ int best_cdbk=0;
+ spx_word32_t best_sum=-VERY_LARGE32;
+ spx_word32_t sum=0;
+ spx_word16_t g[3];
+ spx_word16_t pitch_control=64;
+ spx_word16_t gain_sum;
+ int i;
+
+ for (i=0;i<gain_cdbk_size;i++) {
+
+ ptr = gain_cdbk+4*i;
+ g[0]=ADD16((spx_word16_t)ptr[0],32);
+ g[1]=ADD16((spx_word16_t)ptr[1],32);
+ g[2]=ADD16((spx_word16_t)ptr[2],32);
+ gain_sum = (spx_word16_t)ptr[3];
+
+ sum = compute_pitch_error(C16, g, pitch_control);
+
+ if (sum>best_sum && gain_sum<=max_gain) {
+ best_sum=sum;
+ best_cdbk=i;
+ }
+ }
+
+ return best_cdbk;
+}
+#endif
/** Finds the best quantized 3-tap pitch predictor by analysis by synthesis */
-static spx_word64_t pitch_gain_search_3tap(
-const spx_sig_t target[], /* Target vector */
+static spx_word32_t pitch_gain_search_3tap(
+const spx_word16_t target[], /* Target vector */
const spx_coef_t ak[], /* LPCs for this subframe */
const spx_coef_t awk1[], /* Weighted LPCs #1 for this subframe */
const spx_coef_t awk2[], /* Weighted LPCs #2 for this subframe */
spx_sig_t exc[], /* Excitation */
-const void *par,
+const signed char *gain_cdbk,
+int gain_cdbk_size,
int pitch, /* Pitch value */
int p, /* Number of LPC coeffs */
int nsf, /* Number of samples in subframe */
SpeexBits *bits,
char *stack,
-const spx_sig_t *exc2,
+const spx_word16_t *exc2,
const spx_word16_t *r,
-spx_sig_t *new_target,
+spx_word16_t *new_target,
int *cdbk_index,
-int cdbk_offset,
-int plc_tuning
+int plc_tuning,
+spx_word32_t cumul_gain
)
{
int i,j;
- VARDECL(spx_sig_t *tmp1);
- VARDECL(spx_sig_t *tmp2);
- spx_sig_t *x[3];
- spx_sig_t *e[3];
+ VARDECL(spx_word16_t *tmp1);
+ VARDECL(spx_word16_t *e);
+ spx_word16_t *x[3];
spx_word32_t corr[3];
spx_word32_t A[3][3];
- int gain_cdbk_size;
- const signed char *gain_cdbk;
spx_word16_t gain[3];
- spx_word64_t err;
+ spx_word32_t err;
+ spx_word16_t max_gain=128;
+ int best_cdbk=0;
- const ltp_params *params;
- params = (const ltp_params*) par;
- gain_cdbk_size = 1<<params->gain_bits;
- gain_cdbk = params->gain_cdbk + 3*gain_cdbk_size*cdbk_offset;
- ALLOC(tmp1, 3*nsf, spx_sig_t);
- ALLOC(tmp2, 3*nsf, spx_sig_t);
+ ALLOC(tmp1, 3*nsf, spx_word16_t);
+ ALLOC(e, nsf, spx_word16_t);
+ if (cumul_gain > 262144)
+ max_gain = 31;
+
x[0]=tmp1;
x[1]=tmp1+nsf;
x[2]=tmp1+2*nsf;
- e[0]=tmp2;
- e[1]=tmp2+nsf;
- e[2]=tmp2+2*nsf;
- for (i=2;i>=0;i--)
{
- int pp=pitch+1-i;
+ VARDECL(spx_mem_t *mm);
+ int pp=pitch-1;
+ ALLOC(mm, p, spx_mem_t);
for (j=0;j<nsf;j++)
{
if (j-pp<0)
- e[i][j]=exc2[j-pp];
+ e[j]=exc2[j-pp];
else if (j-pp-pitch<0)
- e[i][j]=exc2[j-pp-pitch];
+ e[j]=exc2[j-pp-pitch];
else
- e[i][j]=0;
+ e[j]=0;
}
-
- if (i==2)
- syn_percep_zero(e[i], ak, awk1, awk2, x[i], nsf, p, stack);
- else {
- for (j=0;j<nsf-1;j++)
- x[i][j+1]=x[i+1][j];
- x[i][0]=0;
- for (j=0;j<nsf;j++)
- {
- x[i][j]=ADD32(x[i][j],SHL32(MULT16_32_Q15(r[j], e[i][0]),1));
- }
- }
+ for (j=0;j<p;j++)
+ mm[j] = 0;
+ iir_mem16(e, ak, e, nsf, p, mm, stack);
+ for (j=0;j<p;j++)
+ mm[j] = 0;
+ filter_mem16(e, awk1, awk2, e, nsf, p, mm, stack);
+ for (j=0;j<nsf;j++)
+ x[2][j] = e[j];
}
-
-#ifdef FIXED_POINT
+ for (i=1;i>=0;i--)
{
- /* If using fixed-point, we need to normalize the signals first */
- spx_word16_t *y[3];
- VARDECL(spx_word16_t *ytmp);
- VARDECL(spx_word16_t *t);
-
- spx_sig_t max_val=1;
- int sig_shift;
-
- ALLOC(ytmp, 3*nsf, spx_word16_t);
-#if 0
- ALLOC(y[0], nsf, spx_word16_t);
- ALLOC(y[1], nsf, spx_word16_t);
- ALLOC(y[2], nsf, spx_word16_t);
-#else
- y[0] = ytmp;
- y[1] = ytmp+nsf;
- y[2] = ytmp+2*nsf;
-#endif
- ALLOC(t, nsf, spx_word16_t);
- for (j=0;j<3;j++)
- {
- for (i=0;i<nsf;i++)
- {
- spx_sig_t tmp = x[j][i];
- if (tmp<0)
- tmp = -tmp;
- if (tmp > max_val)
- max_val = tmp;
- }
- }
- for (i=0;i<nsf;i++)
- {
- spx_sig_t tmp = target[i];
- if (tmp<0)
- tmp = -tmp;
- if (tmp > max_val)
- max_val = tmp;
- }
-
- sig_shift=0;
- while (max_val>16384)
- {
- sig_shift++;
- max_val >>= 1;
- }
-
- for (j=0;j<3;j++)
- {
- for (i=0;i<nsf;i++)
- {
- y[j][i] = EXTRACT16(SHR32(x[j][i],sig_shift));
- }
- }
- for (i=0;i<nsf;i++)
- {
- t[i] = EXTRACT16(SHR32(target[i],sig_shift));
- }
-
- for (i=0;i<3;i++)
- corr[i]=inner_prod(y[i],t,nsf);
-
- for (i=0;i<3;i++)
- for (j=0;j<=i;j++)
- A[i][j]=A[j][i]=inner_prod(y[i],y[j],nsf);
+ spx_word16_t e0=exc2[-pitch-1+i];
+ x[i][0]=MULT16_16_Q14(r[0], e0);
+ for (j=0;j<nsf-1;j++)
+ x[i][j+1]=ADD32(x[i+1][j],MULT16_16_P14(r[j+1], e0));
}
-#else
- {
- for (i=0;i<3;i++)
- corr[i]=inner_prod(x[i],target,nsf);
-
- for (i=0;i<3;i++)
- for (j=0;j<=i;j++)
- A[i][j]=A[j][i]=inner_prod(x[i],x[j],nsf);
- }
-#endif
+
+ for (i=0;i<3;i++)
+ corr[i]=inner_prod(x[i],target,nsf);
+ for (i=0;i<3;i++)
+ for (j=0;j<=i;j++)
+ A[i][j]=A[j][i]=inner_prod(x[i],x[j],nsf);
{
spx_word32_t C[9];
- const signed char *ptr=gain_cdbk;
- int best_cdbk=0;
- spx_word32_t best_sum=0;
+#ifdef FIXED_POINT
+ spx_word16_t C16[9];
+#else
+ spx_word16_t *C16=C;
+#endif
C[0]=corr[2];
C[1]=corr[1];
C[2]=corr[0];
@@ -461,111 +422,73 @@
/*plc_tuning *= 2;*/
if (plc_tuning<2)
plc_tuning=2;
+ if (plc_tuning>30)
+ plc_tuning=30;
#ifdef FIXED_POINT
- C[0] = MAC16_32_Q15(C[0],MULT16_16_16(plc_tuning,-327),C[0]);
- C[1] = MAC16_32_Q15(C[1],MULT16_16_16(plc_tuning,-327),C[1]);
- C[2] = MAC16_32_Q15(C[2],MULT16_16_16(plc_tuning,-327),C[2]);
C[0] = SHL32(C[0],1);
C[1] = SHL32(C[1],1);
C[2] = SHL32(C[2],1);
C[3] = SHL32(C[3],1);
C[4] = SHL32(C[4],1);
C[5] = SHL32(C[5],1);
+ C[6] = MAC16_32_Q15(C[6],MULT16_16_16(plc_tuning,655),C[6]);
+ C[7] = MAC16_32_Q15(C[7],MULT16_16_16(plc_tuning,655),C[7]);
+ C[8] = MAC16_32_Q15(C[8],MULT16_16_16(plc_tuning,655),C[8]);
+ normalize16(C, C16, 32767, 9);
#else
- C[0]*=1-.01*plc_tuning;
- C[1]*=1-.01*plc_tuning;
- C[2]*=1-.01*plc_tuning;
- C[6]*=.5*(1+.01*plc_tuning);
- C[7]*=.5*(1+.01*plc_tuning);
- C[8]*=.5*(1+.01*plc_tuning);
+ C[6]*=.5*(1+.02*plc_tuning);
+ C[7]*=.5*(1+.02*plc_tuning);
+ C[8]*=.5*(1+.02*plc_tuning);
#endif
- for (i=0;i<gain_cdbk_size;i++)
- {
- spx_word32_t sum=0;
- spx_word16_t g[3];
- spx_word16_t pitch_control=64;
- spx_word16_t gain_sum;
-
- ptr = gain_cdbk+3*i;
- g[0]=ADD16((spx_word16_t)ptr[0],32);
- g[1]=ADD16((spx_word16_t)ptr[1],32);
- g[2]=ADD16((spx_word16_t)ptr[2],32);
- /* We favor "safe" pitch values to handle packet loss better */
- gain_sum = ADD16(ADD16(g[1],MAX16(g[0], 0)),MAX16(g[2], 0));
- if (gain_sum > 64)
- {
- gain_sum = SUB16(gain_sum, 64);
- if (gain_sum > 127)
- gain_sum = 127;
+ best_cdbk = pitch_gain_search_3tap_vq(gain_cdbk, gain_cdbk_size, C16, max_gain);
+
#ifdef FIXED_POINT
- pitch_control = SUB16(64,EXTRACT16(PSHR32(MULT16_16(64,MULT16_16_16(plc_tuning, gain_sum)),10)));
-#else
- pitch_control = 64*(1.-.001*plc_tuning*gain_sum);
-#endif
- if (pitch_control < 0)
- pitch_control = 0;
- }
-
- sum = compute_pitch_error(C, g, pitch_control);
-
- if (sum>best_sum || i==0)
- {
- best_sum=sum;
- best_cdbk=i;
- }
- }
-#ifdef FIXED_POINT
- gain[0] = ADD16(32,(spx_word16_t)gain_cdbk[best_cdbk*3]);
- gain[1] = ADD16(32,(spx_word16_t)gain_cdbk[best_cdbk*3+1]);
- gain[2] = ADD16(32,(spx_word16_t)gain_cdbk[best_cdbk*3+2]);
+ gain[0] = ADD16(32,(spx_word16_t)gain_cdbk[best_cdbk*4]);
+ gain[1] = ADD16(32,(spx_word16_t)gain_cdbk[best_cdbk*4+1]);
+ gain[2] = ADD16(32,(spx_word16_t)gain_cdbk[best_cdbk*4+2]);
/*printf ("%d %d %d %d\n",gain[0],gain[1],gain[2], best_cdbk);*/
#else
- gain[0] = 0.015625*gain_cdbk[best_cdbk*3] + .5;
- gain[1] = 0.015625*gain_cdbk[best_cdbk*3+1]+ .5;
- gain[2] = 0.015625*gain_cdbk[best_cdbk*3+2]+ .5;
+ gain[0] = 0.015625*gain_cdbk[best_cdbk*4] + .5;
+ gain[1] = 0.015625*gain_cdbk[best_cdbk*4+1]+ .5;
+ gain[2] = 0.015625*gain_cdbk[best_cdbk*4+2]+ .5;
#endif
*cdbk_index=best_cdbk;
}
-#ifdef FIXED_POINT
for (i=0;i<nsf;i++)
- exc[i]=SHL32(ADD32(ADD32(MULT16_32_Q15(SHL16(gain[0],7),e[2][i]), MULT16_32_Q15(SHL16(gain[1],7),e[1][i])),
- MULT16_32_Q15(SHL16(gain[2],7),e[0][i])), 2);
-
- err=0;
+ exc[i]=0;
+ for (i=0;i<3;i++)
+ {
+ int j;
+ int tmp1, tmp3;
+ int pp=pitch+1-i;
+ tmp1=nsf;
+ if (tmp1>pp)
+ tmp1=pp;
+ for (j=0;j<tmp1;j++)
+ exc[j]=MAC16_16(exc[j],SHL16(gain[2-i],7),exc2[j-pp]);
+ tmp3=nsf;
+ if (tmp3>pp+pitch)
+ tmp3=pp+pitch;
+ for (j=tmp1;j<tmp3;j++)
+ exc[j]=MAC16_16(exc[j],SHL16(gain[2-i],7),exc2[j-pp-pitch]);
+ }
for (i=0;i<nsf;i++)
{
- spx_word16_t perr2;
- spx_sig_t tmp = SHL32(ADD32(ADD32(MULT16_32_Q15(SHL16(gain[0],7),x[2][i]),MULT16_32_Q15(SHL16(gain[1],7),x[1][i])),
- MULT16_32_Q15(SHL16(gain[2],7),x[0][i])),2);
- spx_sig_t perr=SUB32(target[i],tmp);
- new_target[i] = SUB32(target[i], tmp);
- perr2 = EXTRACT16(PSHR32(perr,15));
- err = ADD64(err,MULT16_16(perr2,perr2));
-
+ spx_word32_t tmp = ADD32(ADD32(MULT16_16(gain[0],x[2][i]),MULT16_16(gain[1],x[1][i])),
+ MULT16_16(gain[2],x[0][i]));
+ new_target[i] = SUB16(target[i], EXTRACT16(PSHR32(tmp,6)));
}
-#else
- for (i=0;i<nsf;i++)
- exc[i]=gain[0]*e[2][i]+gain[1]*e[1][i]+gain[2]*e[0][i];
-
- err=0;
- for (i=0;i<nsf;i++)
- {
- spx_sig_t tmp = gain[2]*x[0][i]+gain[1]*x[1][i]+gain[0]*x[2][i];
- new_target[i] = target[i] - tmp;
- err+=new_target[i]*new_target[i];
- }
-#endif
+ err = inner_prod(new_target, new_target, nsf);
return err;
}
-
/** Finds the best quantized 3-tap pitch predictor by analysis by synthesis */
int pitch_search_3tap(
-spx_sig_t target[], /* Target vector */
-spx_sig_t *sw,
+spx_word16_t target[], /* Target vector */
+spx_word16_t *sw,
spx_coef_t ak[], /* LPCs for this subframe */
spx_coef_t awk1[], /* Weighted LPCs #1 for this subframe */
spx_coef_t awk2[], /* Weighted LPCs #2 for this subframe */
@@ -578,24 +501,32 @@
int nsf, /* Number of samples in subframe */
SpeexBits *bits,
char *stack,
-spx_sig_t *exc2,
+spx_word16_t *exc2,
spx_word16_t *r,
int complexity,
int cdbk_offset,
-int plc_tuning
+int plc_tuning,
+spx_word32_t *cumul_gain
)
{
int i,j;
int cdbk_index, pitch=0, best_gain_index=0;
VARDECL(spx_sig_t *best_exc);
- VARDECL(spx_sig_t *new_target);
- VARDECL(spx_sig_t *best_target);
+ VARDECL(spx_word16_t *new_target);
+ VARDECL(spx_word16_t *best_target);
int best_pitch=0;
- spx_word64_t err, best_err=-1;
+ spx_word32_t err, best_err=-1;
int N;
const ltp_params *params;
+ const signed char *gain_cdbk;
+ int gain_cdbk_size;
+
VARDECL(int *nbest);
-
+
+ params = (const ltp_params*) par;
+ gain_cdbk_size = 1<<params->gain_bits;
+ gain_cdbk = params->gain_cdbk + 4*gain_cdbk_size*cdbk_offset;
+
N=complexity;
if (N>10)
N=10;
@@ -614,23 +545,24 @@
return start;
}
- ALLOC(best_exc, nsf, spx_sig_t);
- ALLOC(new_target, nsf, spx_sig_t);
- ALLOC(best_target, nsf, spx_sig_t);
-
if (N>end-start+1)
N=end-start+1;
if (end != start)
open_loop_nbest_pitch(sw, start, end, nsf, nbest, NULL, N, stack);
else
nbest[0] = start;
+
+ ALLOC(best_exc, nsf, spx_sig_t);
+ ALLOC(new_target, nsf, spx_word16_t);
+ ALLOC(best_target, nsf, spx_word16_t);
+
for (i=0;i<N;i++)
{
pitch=nbest[i];
for (j=0;j<nsf;j++)
exc[j]=0;
- err=pitch_gain_search_3tap(target, ak, awk1, awk2, exc, par, pitch, p, nsf,
- bits, stack, exc2, r, new_target, &cdbk_index, cdbk_offset, plc_tuning);
+ err=pitch_gain_search_3tap(target, ak, awk1, awk2, exc, gain_cdbk, gain_cdbk_size, pitch, p, nsf,
+ bits, stack, exc2, r, new_target, &cdbk_index, plc_tuning, *cumul_gain);
if (err<best_err || best_err<0)
{
for (j=0;j<nsf;j++)
@@ -642,10 +574,15 @@
best_gain_index=cdbk_index;
}
}
-
/*printf ("pitch: %d %d\n", best_pitch, best_gain_index);*/
speex_bits_pack(bits, best_pitch-start, params->pitch_bits);
speex_bits_pack(bits, best_gain_index, params->gain_bits);
+#ifdef FIXED_POINT
+ *cumul_gain = MULT16_32_Q13(SHL16(params->gain_cdbk[4*best_gain_index+3],8), MAX32(1024,*cumul_gain));
+#else
+ *cumul_gain = 0.03125*MAX32(1024,*cumul_gain)*params->gain_cdbk[4*best_gain_index+3];
+#endif
+ /*printf ("%f\n", cumul_gain);*/
/*printf ("encode pitch: %d %d\n", best_pitch, best_gain_index);*/
for (i=0;i<nsf;i++)
exc[i]=best_exc[i];
@@ -656,10 +593,11 @@
}
void pitch_unquant_3tap(
-spx_sig_t exc[], /* Excitation */
+spx_word16_t exc[], /* Input excitation */
+spx_word32_t exc_out[], /* Output excitation */
int start, /* Smallest pitch value allowed */
int end, /* Largest pitch value allowed */
-spx_word16_t pitch_coef, /* Voicing (pitch) coefficient */
+spx_word16_t pitch_coef, /* Voicing (pitch) coefficient */
const void *par,
int nsf, /* Number of samples in subframe */
int *pitch_val,
@@ -682,20 +620,20 @@
params = (const ltp_params*) par;
gain_cdbk_size = 1<<params->gain_bits;
- gain_cdbk = params->gain_cdbk + 3*gain_cdbk_size*cdbk_offset;
+ gain_cdbk = params->gain_cdbk + 4*gain_cdbk_size*cdbk_offset;
pitch = speex_bits_unpack_unsigned(bits, params->pitch_bits);
pitch += start;
gain_index = speex_bits_unpack_unsigned(bits, params->gain_bits);
/*printf ("decode pitch: %d %d\n", pitch, gain_index);*/
#ifdef FIXED_POINT
- gain[0] = ADD16(32,(spx_word16_t)gain_cdbk[gain_index*3]);
- gain[1] = ADD16(32,(spx_word16_t)gain_cdbk[gain_index*3+1]);
- gain[2] = ADD16(32,(spx_word16_t)gain_cdbk[gain_index*3+2]);
+ gain[0] = ADD16(32,(spx_word16_t)gain_cdbk[gain_index*4]);
+ gain[1] = ADD16(32,(spx_word16_t)gain_cdbk[gain_index*4+1]);
+ gain[2] = ADD16(32,(spx_word16_t)gain_cdbk[gain_index*4+2]);
#else
- gain[0] = 0.015625*gain_cdbk[gain_index*3]+.5;
- gain[1] = 0.015625*gain_cdbk[gain_index*3+1]+.5;
- gain[2] = 0.015625*gain_cdbk[gain_index*3+2]+.5;
+ gain[0] = 0.015625*gain_cdbk[gain_index*4]+.5;
+ gain[1] = 0.015625*gain_cdbk[gain_index*4+1]+.5;
+ gain[2] = 0.015625*gain_cdbk[gain_index*4+2]+.5;
#endif
if (count_lost && pitch > subframe_offset)
@@ -728,66 +666,36 @@
gain_val[0]=gain[0];
gain_val[1]=gain[1];
gain_val[2]=gain[2];
-
+ gain[0] = SHL16(gain[0],7);
+ gain[1] = SHL16(gain[1],7);
+ gain[2] = SHL16(gain[2],7);
+ for (i=0;i<nsf;i++)
+ exc_out[i]=0;
+ for (i=0;i<3;i++)
{
- spx_sig_t *e[3];
- VARDECL(spx_sig_t *tmp2);
- ALLOC(tmp2, 3*nsf, spx_sig_t);
- e[0]=tmp2;
- e[1]=tmp2+nsf;
- e[2]=tmp2+2*nsf;
-
- for (i=0;i<3;i++)
- {
- int j;
- int pp=pitch+1-i;
-#if 0
- for (j=0;j<nsf;j++)
- {
- if (j-pp<0)
- e[i][j]=exc[j-pp];
- else if (j-pp-pitch<0)
- e[i][j]=exc[j-pp-pitch];
- else
- e[i][j]=0;
- }
-#else
- {
- int tmp1, tmp3;
- tmp1=nsf;
- if (tmp1>pp)
- tmp1=pp;
- for (j=0;j<tmp1;j++)
- e[i][j]=exc[j-pp];
- tmp3=nsf;
- if (tmp3>pp+pitch)
- tmp3=pp+pitch;
- for (j=tmp1;j<tmp3;j++)
- e[i][j]=exc[j-pp-pitch];
- for (j=tmp3;j<nsf;j++)
- e[i][j]=0;
- }
-#endif
- }
-
-#ifdef FIXED_POINT
- {
- for (i=0;i<nsf;i++)
- exc[i]=SHL32(ADD32(ADD32(MULT16_32_Q15(SHL16(gain[0],7),e[2][i]), MULT16_32_Q15(SHL16(gain[1],7),e[1][i])),
- MULT16_32_Q15(SHL16(gain[2],7),e[0][i])), 2);
- }
-#else
- for (i=0;i<nsf;i++)
- exc[i]=VERY_SMALL+gain[0]*e[2][i]+gain[1]*e[1][i]+gain[2]*e[0][i];
-#endif
+ int j;
+ int tmp1, tmp3;
+ int pp=pitch+1-i;
+ tmp1=nsf;
+ if (tmp1>pp)
+ tmp1=pp;
+ for (j=0;j<tmp1;j++)
+ exc_out[j]=MAC16_16(exc_out[j],gain[2-i],exc[j-pp]);
+ tmp3=nsf;
+ if (tmp3>pp+pitch)
+ tmp3=pp+pitch;
+ for (j=tmp1;j<tmp3;j++)
+ exc_out[j]=MAC16_16(exc_out[j],gain[2-i],exc[j-pp-pitch]);
}
+ /*for (i=0;i<nsf;i++)
+ exc[i]=PSHR32(exc32[i],13);*/
}
/** Forced pitch delay and gain */
int forced_pitch_quant(
-spx_sig_t target[], /* Target vector */
-spx_sig_t *sw,
+spx_word16_t target[], /* Target vector */
+spx_word16_t *sw,
spx_coef_t ak[], /* LPCs for this subframe */
spx_coef_t awk1[], /* Weighted LPCs #1 for this subframe */
spx_coef_t awk2[], /* Weighted LPCs #2 for this subframe */
@@ -800,30 +708,45 @@
int nsf, /* Number of samples in subframe */
SpeexBits *bits,
char *stack,
-spx_sig_t *exc2,
+spx_word16_t *exc2,
spx_word16_t *r,
int complexity,
int cdbk_offset,
-int plc_tuning
+int plc_tuning,
+spx_word32_t *cumul_gain
)
{
int i;
- float coef = GAIN_SCALING_1*pitch_coef;
- if (coef>.99)
- coef=.99;
- for (i=0;i<nsf;i++)
+ VARDECL(spx_sig_t *res);
+ ALLOC(res, nsf, spx_sig_t);
+#ifdef FIXED_POINT
+ if (pitch_coef>63)
+ pitch_coef=63;
+#else
+ if (pitch_coef>.99)
+ pitch_coef=.99;
+#endif
+ for (i=0;i<nsf&&i<start;i++)
{
- exc[i]=exc[i-start]*coef;
+ exc[i]=MULT16_16(SHL16(pitch_coef, 7),exc2[i-start]);
}
+ for (;i<nsf;i++)
+ {
+ exc[i]=MULT16_32_Q15(SHL16(pitch_coef, 9),exc[i-start]);
+ }
+ syn_percep_zero(exc, ak, awk1, awk2, res, nsf, p, stack);
+ for (i=0;i<nsf;i++)
+ target[i]=EXTRACT16(SATURATE(SUB32(EXTEND32(target[i]),PSHR32(res[i],SIG_SHIFT-1)),32700));
return start;
}
/** Unquantize forced pitch delay and gain */
void forced_pitch_unquant(
-spx_sig_t exc[], /* Excitation */
+spx_word16_t exc[], /* Input excitation */
+spx_word32_t exc_out[], /* Output excitation */
int start, /* Smallest pitch value allowed */
int end, /* Largest pitch value allowed */
-spx_word16_t pitch_coef, /* Voicing (pitch) coefficient */
+spx_word16_t pitch_coef, /* Voicing (pitch) coefficient */
const void *par,
int nsf, /* Number of samples in subframe */
int *pitch_val,
@@ -837,12 +760,17 @@
)
{
int i;
- float coef = GAIN_SCALING_1*pitch_coef;
- if (coef>.99)
- coef=.99;
+#ifdef FIXED_POINT
+ if (pitch_coef>63)
+ pitch_coef=63;
+#else
+ if (pitch_coef>.99)
+ pitch_coef=.99;
+#endif
for (i=0;i<nsf;i++)
{
- exc[i]=exc[i-start]*coef;
+ exc_out[i]=MULT16_16(exc[i-start],SHL16(pitch_coef,7));
+ exc[i] = PSHR(exc_out[i],13);
}
*pitch_val = start;
gain_val[0]=gain_val[2]=0;
diff --git a/pjmedia/src/pjmedia-codec/speex/ltp.h b/pjmedia/src/pjmedia-codec/speex/ltp.h
index 36debbd..bc050c6 100644
--- a/pjmedia/src/pjmedia-codec/speex/ltp.h
+++ b/pjmedia/src/pjmedia-codec/speex/ltp.h
@@ -48,13 +48,16 @@
#define gain_3tap_to_1tap(g) (ABS(g[1]) + (g[0]>0 ? g[0] : -.5*g[0]) + (g[2]>0 ? g[2] : -.5*g[2]))
#endif
-void open_loop_nbest_pitch(spx_sig_t *sw, int start, int end, int len, int *pitch, spx_word16_t *gain, int N, char *stack);
+spx_word32_t inner_prod(const spx_word16_t *x, const spx_word16_t *y, int len);
+void pitch_xcorr(const spx_word16_t *_x, const spx_word16_t *_y, spx_word32_t *corr, int len, int nb_pitch, char *stack);
+
+void open_loop_nbest_pitch(spx_word16_t *sw, int start, int end, int len, int *pitch, spx_word16_t *gain, int N, char *stack);
/** Finds the best quantized 3-tap pitch predictor by analysis by synthesis */
int pitch_search_3tap(
-spx_sig_t target[], /* Target vector */
-spx_sig_t *sw,
+spx_word16_t target[], /* Target vector */
+spx_word16_t *sw,
spx_coef_t ak[], /* LPCs for this subframe */
spx_coef_t awk1[], /* Weighted LPCs #1 for this subframe */
spx_coef_t awk2[], /* Weighted LPCs #2 for this subframe */
@@ -67,19 +70,21 @@
int nsf, /* Number of samples in subframe */
SpeexBits *bits,
char *stack,
-spx_sig_t *exc2,
+spx_word16_t *exc2,
spx_word16_t *r,
int complexity,
int cdbk_offset,
-int plc_tuning
+int plc_tuning,
+spx_word32_t *cumul_gain
);
/*Unquantize adaptive codebook and update pitch contribution*/
void pitch_unquant_3tap(
-spx_sig_t exc[], /* Excitation */
+spx_word16_t exc[], /* Input excitation */
+spx_word32_t exc_out[], /* Output excitation */
int start, /* Smallest pitch value allowed */
int end, /* Largest pitch value allowed */
-spx_word16_t pitch_coef, /* Voicing (pitch) coefficient */
+spx_word16_t pitch_coef, /* Voicing (pitch) coefficient */
const void *par,
int nsf, /* Number of samples in subframe */
int *pitch_val,
@@ -94,8 +99,8 @@
/** Forced pitch delay and gain */
int forced_pitch_quant(
-spx_sig_t target[], /* Target vector */
-spx_sig_t *sw,
+spx_word16_t target[], /* Target vector */
+spx_word16_t *sw,
spx_coef_t ak[], /* LPCs for this subframe */
spx_coef_t awk1[], /* Weighted LPCs #1 for this subframe */
spx_coef_t awk2[], /* Weighted LPCs #2 for this subframe */
@@ -108,19 +113,21 @@
int nsf, /* Number of samples in subframe */
SpeexBits *bits,
char *stack,
-spx_sig_t *exc2,
+spx_word16_t *exc2,
spx_word16_t *r,
int complexity,
int cdbk_offset,
-int plc_tuning
+int plc_tuning,
+spx_word32_t *cumul_gain
);
/** Unquantize forced pitch delay and gain */
void forced_pitch_unquant(
-spx_sig_t exc[], /* Excitation */
+spx_word16_t exc[], /* Input excitation */
+spx_word32_t exc_out[], /* Output excitation */
int start, /* Smallest pitch value allowed */
int end, /* Largest pitch value allowed */
-spx_word16_t pitch_coef, /* Voicing (pitch) coefficient */
+spx_word16_t pitch_coef, /* Voicing (pitch) coefficient */
const void *par,
int nsf, /* Number of samples in subframe */
int *pitch_val,
diff --git a/pjmedia/src/pjmedia-codec/speex/ltp_arm4.h b/pjmedia/src/pjmedia-codec/speex/ltp_arm4.h
index a5a0bee..7479e8b 100644
--- a/pjmedia/src/pjmedia-codec/speex/ltp_arm4.h
+++ b/pjmedia/src/pjmedia-codec/speex/ltp_arm4.h
@@ -33,7 +33,7 @@
*/
#define OVERRIDE_INNER_PROD
-static spx_word32_t inner_prod(const spx_word16_t *x, const spx_word16_t *y, int len)
+spx_word32_t inner_prod(const spx_word16_t *x, const spx_word16_t *y, int len)
{
spx_word32_t sum1=0,sum2=0;
spx_word16_t *deadx, *deady;
@@ -84,7 +84,7 @@
}
#define OVERRIDE_PITCH_XCORR
-static void pitch_xcorr(const spx_word16_t *_x, const spx_word16_t *_y, spx_word32_t *corr, int len, int nb_pitch, char *stack)
+void pitch_xcorr(const spx_word16_t *_x, const spx_word16_t *_y, spx_word32_t *corr, int len, int nb_pitch, char *stack)
{
int i,j;
for (i=0;i<nb_pitch;i+=4)
diff --git a/pjmedia/src/pjmedia-codec/speex/ltp_bfin.h b/pjmedia/src/pjmedia-codec/speex/ltp_bfin.h
index e92dbe2..c466902 100644
--- a/pjmedia/src/pjmedia-codec/speex/ltp_bfin.h
+++ b/pjmedia/src/pjmedia-codec/speex/ltp_bfin.h
@@ -34,7 +34,7 @@
*/
#define OVERRIDE_INNER_PROD
-static spx_word32_t inner_prod(const spx_word16_t *x, const spx_word16_t *y, int len)
+spx_word32_t inner_prod(const spx_word16_t *x, const spx_word16_t *y, int len)
{
spx_word32_t sum=0;
__asm__ __volatile__ (
@@ -63,7 +63,7 @@
}
#define OVERRIDE_PITCH_XCORR
-static void pitch_xcorr(const spx_word16_t *_x, const spx_word16_t *_y, spx_word32_t *corr, int len, int nb_pitch, char *stack)
+void pitch_xcorr(const spx_word16_t *_x, const spx_word16_t *_y, spx_word32_t *corr, int len, int nb_pitch, char *stack)
{
corr += nb_pitch - 1;
__asm__ __volatile__ (
@@ -109,52 +109,41 @@
}
#define OVERRIDE_COMPUTE_PITCH_ERROR
-static inline spx_word32_t compute_pitch_error(spx_word32_t *C, spx_word16_t *g, spx_word16_t pitch_control)
+static inline spx_word32_t compute_pitch_error(spx_word16_t *C, spx_word16_t *g, spx_word16_t pitch_control)
{
spx_word32_t sum;
__asm__ __volatile__
(
- "A1 = A0 = 0;\n\t"
+ "A0 = 0;\n\t"
- "R0 = [%1++];\n\t"
+ "R0 = W[%1++];\n\t"
"R1.L = %2.L*%5.L (IS);\n\t"
- "R0 <<= 1;\n\t"
- "A1 += R1.L*R0.L (M), A0 += R1.L*R0.H (IS) || R0 = [%1++];\n\t"
+ "A0 += R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
"R1.L = %3.L*%5.L (IS);\n\t"
- "R0 <<= 1;\n\t"
- "A1 += R1.L*R0.L (M), A0 += R1.L*R0.H (IS) || R0 = [%1++];\n\t"
+ "A0 += R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
"R1.L = %4.L*%5.L (IS);\n\t"
- "R0 <<= 1;\n\t"
- "A1 += R1.L*R0.L (M), A0 += R1.L*R0.H (IS) || R0 = [%1++];\n\t"
+ "A0 += R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
"R1.L = %2.L*%3.L (IS);\n\t"
- "R0 <<= 1;\n\t"
- "A1 -= R1.L*R0.L (M), A0 -= R1.L*R0.H (IS) || R0 = [%1++];\n\t"
+ "A0 -= R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
"R1.L = %4.L*%3.L (IS);\n\t"
- "R0 <<= 1;\n\t"
- "A1 -= R1.L*R0.L (M), A0 -= R1.L*R0.H (IS) || R0 = [%1++];\n\t"
+ "A0 -= R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
"R1.L = %4.L*%2.L (IS);\n\t"
- "R0 <<= 1;\n\t"
- "A1 -= R1.L*R0.L (M), A0 -= R1.L*R0.H (IS) || R0 = [%1++];\n\t"
+ "A0 -= R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
"R1.L = %2.L*%2.L (IS);\n\t"
- "R0 <<= 1;\n\t"
- "A1 -= R1.L*R0.L (M), A0 -= R1.L*R0.H (IS) || R0 = [%1++];\n\t"
+ "A0 -= R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
"R1.L = %3.L*%3.L (IS);\n\t"
- "R0 <<= 1;\n\t"
- "A1 -= R1.L*R0.L (M), A0 -= R1.L*R0.H (IS) || R0 = [%1++];\n\t"
+ "A0 -= R1.L*R0.L (IS) || R0 = W[%1++];\n\t"
"R1.L = %4.L*%4.L (IS);\n\t"
- "R0 <<= 1;\n\t"
- "A1 -= R1.L*R0.L (M), A0 -= R1.L*R0.H (IS);\n\t"
+ "A0 -= R1.L*R0.L (IS);\n\t"
- "A1 = A1 >>> 16;\n\t"
- "A0 += A1;\n\t"
"%0 = A0;\n\t"
: "=&D" (sum), "=a" (C)
: "d" (g[0]), "d" (g[1]), "d" (g[2]), "d" (pitch_control), "1" (C)
@@ -163,3 +152,268 @@
return sum;
}
+#define OVERRIDE_OPEN_LOOP_NBEST_PITCH
+#ifdef OVERRIDE_OPEN_LOOP_NBEST_PITCH
+void open_loop_nbest_pitch(spx_word16_t *sw, int start, int end, int len, int *pitch, spx_word16_t *gain, int N, char *stack)
+{
+ int i,j,k;
+ VARDECL(spx_word32_t *best_score);
+ VARDECL(spx_word32_t *best_ener);
+ spx_word32_t e0;
+ VARDECL(spx_word32_t *corr);
+ VARDECL(spx_word32_t *energy);
+
+ ALLOC(best_score, N, spx_word32_t);
+ ALLOC(best_ener, N, spx_word32_t);
+ ALLOC(corr, end-start+1, spx_word32_t);
+ ALLOC(energy, end-start+2, spx_word32_t);
+
+ for (i=0;i<N;i++)
+ {
+ best_score[i]=-1;
+ best_ener[i]=0;
+ pitch[i]=start;
+ }
+
+ energy[0]=inner_prod(sw-start, sw-start, len);
+ e0=inner_prod(sw, sw, len);
+
+ /* energy update -------------------------------------*/
+
+ __asm__ __volatile__
+ (
+" P0 = %0;\n\t"
+" I1 = %1;\n\t"
+" L1 = 0;\n\t"
+" I2 = %2;\n\t"
+" L2 = 0;\n\t"
+" R2 = [P0++];\n\t"
+" R3 = 0;\n\t"
+" LSETUP (eu1, eu2) LC1 = %3;\n\t"
+"eu1: R1.L = W [I1--] || R0.L = W [I2--] ;\n\t"
+" R1 = R1.L * R1.L (IS);\n\t"
+" R0 = R0.L * R0.L (IS);\n\t"
+" R1 >>>= 6;\n\t"
+" R1 = R1 + R2;\n\t"
+" R0 >>>= 6;\n\t"
+" R1 = R1 - R0;\n\t"
+" R2 = MAX(R1,R3);\n\t"
+"eu2: [P0++] = R2;\n\t"
+ : : "d" (energy), "d" (&sw[-start-1]), "d" (&sw[-start+len-1]),
+ "a" (end-start)
+ : "P0", "I1", "I2", "R0", "R1", "R2", "R3"
+#if (__GNUC__ == 4)
+ , "LC1"
+#endif
+ );
+
+ pitch_xcorr(sw, sw-end, corr, len, end-start+1, stack);
+
+ /* FIXME: Fixed-point and floating-point code should be merged */
+ {
+ VARDECL(spx_word16_t *corr16);
+ VARDECL(spx_word16_t *ener16);
+ ALLOC(corr16, end-start+1, spx_word16_t);
+ ALLOC(ener16, end-start+1, spx_word16_t);
+ /* Normalize to 180 so we can square it and it still fits in 16 bits */
+ normalize16(corr, corr16, 180, end-start+1);
+ normalize16(energy, ener16, 180, end-start+1);
+
+ if (N == 1) {
+ /* optimised asm to handle N==1 case */
+ __asm__ __volatile__
+ (
+" I0 = %1;\n\t" /* I0: corr16[] */
+" L0 = 0;\n\t"
+" I1 = %2;\n\t" /* I1: energy */
+" L1 = 0;\n\t"
+" R2 = -1;\n\t" /* R2: best score */
+" R3 = 0;\n\t" /* R3: best energy */
+" P0 = %4;\n\t" /* P0: best pitch */
+" P1 = %4;\n\t" /* P1: counter */
+" LSETUP (sl1, sl2) LC1 = %3;\n\t"
+"sl1: R0.L = W [I0++] || R1.L = W [I1++];\n\t"
+" R0 = R0.L * R0.L (IS);\n\t"
+" R1 += 1;\n\t"
+" R4 = R0.L * R3.L;\n\t"
+" R5 = R2.L * R1.L;\n\t"
+" cc = R5 < R4;\n\t"
+" if cc R2 = R0;\n\t"
+" if cc R3 = R1;\n\t"
+" if cc P0 = P1;\n\t"
+"sl2: P1 += 1;\n\t"
+" %0 = P0;\n\t"
+ : "=&d" (pitch[0])
+ : "a" (corr16), "a" (ener16), "a" (end+1-start), "d" (start)
+ : "P0", "P1", "I0", "I1", "R0", "R1", "R2", "R3", "R4", "R5"
+#if (__GNUC__ == 4)
+ , "LC1"
+#endif
+ );
+
+ }
+ else {
+ for (i=start;i<=end;i++)
+ {
+ spx_word16_t tmp = MULT16_16_16(corr16[i-start],corr16[i-start]);
+ /* Instead of dividing the tmp by the energy, we multiply on the other side */
+ if (MULT16_16(tmp,best_ener[N-1])>MULT16_16(best_score[N-1],ADD16(1,ener16[i-start])))
+ {
+ /* We can safely put it last and then check */
+ best_score[N-1]=tmp;
+ best_ener[N-1]=ener16[i-start]+1;
+ pitch[N-1]=i;
+ /* Check if it comes in front of others */
+ for (j=0;j<N-1;j++)
+ {
+ if (MULT16_16(tmp,best_ener[j])>MULT16_16(best_score[j],ADD16(1,ener16[i-start])))
+ {
+ for (k=N-1;k>j;k--)
+ {
+ best_score[k]=best_score[k-1];
+ best_ener[k]=best_ener[k-1];
+ pitch[k]=pitch[k-1];
+ }
+ best_score[j]=tmp;
+ best_ener[j]=ener16[i-start]+1;
+ pitch[j]=i;
+ break;
+ }
+ }
+ }
+ }
+ }
+ }
+
+ /* Compute open-loop gain */
+ if (gain)
+ {
+ for (j=0;j<N;j++)
+ {
+ spx_word16_t g;
+ i=pitch[j];
+ g = DIV32(corr[i-start], 10+SHR32(MULT16_16(spx_sqrt(e0),spx_sqrt(energy[i-start])),6));
+ /* FIXME: g = max(g,corr/energy) */
+ if (g<0)
+ g = 0;
+ gain[j]=g;
+ }
+ }
+}
+#endif
+
+#define OVERRIDE_PITCH_GAIN_SEARCH_3TAP_VQ
+#ifdef OVERRIDE_PITCH_GAIN_SEARCH_3TAP_VQ
+static int pitch_gain_search_3tap_vq(
+ const signed char *gain_cdbk,
+ int gain_cdbk_size,
+ spx_word16_t *C16,
+ spx_word16_t max_gain
+)
+{
+ const signed char *ptr=gain_cdbk;
+ int best_cdbk=0;
+ spx_word32_t best_sum=-VERY_LARGE32;
+ spx_word32_t sum=0;
+ spx_word16_t g[3];
+ spx_word16_t pitch_control=64;
+ spx_word16_t gain_sum;
+ int i;
+
+ /* fast asm version of VQ codebook search */
+
+ __asm__ __volatile__
+ (
+
+" P0 = %2;\n\t" /* P0: ptr to gain_cdbk */
+" L1 = 0;\n\t" /* no circ addr for L1 */
+" %0 = 0;\n\t" /* %0: best_sum */
+" %1 = 0;\n\t" /* %1: best_cbdk */
+" P1 = 0;\n\t" /* P1: loop counter */
+" R5 = 64;\n\t" /* R5: pitch_control */
+
+" LSETUP (pgs1, pgs2) LC1 = %4;\n\t"
+"pgs1: R2 = B [P0++] (X);\n\t" /* R2: g[0] */
+" R3 = B [P0++] (X);\n\t" /* R3: g[1] */
+" R4 = B [P0++] (X);\n\t" /* R4: g[2] */
+" R2 += 32;\n\t"
+" R3 += 32;\n\t"
+" R4 += 32;\n\t"
+
+" R0 = B [P0++] (X);\n\t"
+" B0 = R0;\n\t" /* BO: gain_sum */
+
+ /* compute_pitch_error() -------------------------------*/
+
+" I1 = %3;\n\t" /* I1: ptr to C */
+" A0 = 0;\n\t"
+
+" R0.L = W[I1++];\n\t"
+" R1.L = R2.L*R5.L (IS);\n\t"
+" A0 += R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R3.L*R5.L (IS);\n\t"
+" A0 += R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R4.L*R5.L (IS);\n\t"
+" A0 += R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R2.L*R3.L (IS);\n\t"
+" A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R4.L*R3.L (IS);\n\t"
+" A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R4.L*R2.L (IS);\n\t"
+" A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R2.L*R2.L (IS);\n\t"
+" A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R3.L*R3.L (IS);\n\t"
+" A0 -= R1.L*R0.L (IS) || R0.L = W[I1++];\n\t"
+
+" R1.L = R4.L*R4.L (IS);\n\t"
+" R0 = (A0 -= R1.L*R0.L) (IS);\n\t"
+
+/*
+ Re-arrange the if-then to code efficiently on the Blackfin:
+
+ if (sum>best_sum && gain_sum<=max_gain) ------ (1)
+
+ if (sum>best_sum && !(gain_sum>max_gain)) ------ (2)
+
+ if (max_gain<=gain_sum) { ------ (3)
+ sum = -VERY_LARGE32;
+ }
+ if (best_sum<=sum)
+
+ The blackin cc instructions are all of the form:
+
+ cc = x < y (or cc = x <= y)
+*/
+" R1 = B0\n\t"
+" R2 = %5\n\t"
+" R3 = %6\n\t"
+" cc = R2 <= R1;\n\t"
+" if cc R0 = R3;\n\t"
+" cc = %0 <= R0;\n\t"
+" if cc %0 = R0;\n\t"
+" if cc %1 = P1;\n\t"
+
+"pgs2: P1 += 1;\n\t"
+
+ : "=&d" (best_sum), "=&d" (best_cdbk)
+ : "a" (gain_cdbk), "a" (C16), "a" (gain_cdbk_size), "a" (max_gain),
+ "b" (-VERY_LARGE32)
+ : "R0", "R1", "R2", "R3", "R4", "R5", "P0",
+ "P1", "I1", "L1", "A0", "B0"
+#if (__GNUC__ == 4)
+ , "LC1"
+#endif
+ );
+
+ return best_cdbk;
+}
+#endif
+
diff --git a/pjmedia/src/pjmedia-codec/speex/ltp_sse.h b/pjmedia/src/pjmedia-codec/speex/ltp_sse.h
index 94c0012..bed6eaa 100644
--- a/pjmedia/src/pjmedia-codec/speex/ltp_sse.h
+++ b/pjmedia/src/pjmedia-codec/speex/ltp_sse.h
@@ -35,7 +35,7 @@
#include <xmmintrin.h>
#define OVERRIDE_INNER_PROD
-static float inner_prod(const float *a, const float *b, int len)
+float inner_prod(const float *a, const float *b, int len)
{
int i;
float ret;
@@ -54,7 +54,7 @@
}
#define OVERRIDE_PITCH_XCORR
-static void pitch_xcorr(const float *_x, const float *_y, float *corr, int len, int nb_pitch, char *stack)
+void pitch_xcorr(const float *_x, const float *_y, float *corr, int len, int nb_pitch, char *stack)
{
int i, offset;
VARDECL(__m128 *x);
diff --git a/pjmedia/src/pjmedia-codec/speex/mdf.c b/pjmedia/src/pjmedia-codec/speex/mdf.c
index 0e7219c..eabf433 100644
--- a/pjmedia/src/pjmedia-codec/speex/mdf.c
+++ b/pjmedia/src/pjmedia-codec/speex/mdf.c
@@ -90,7 +90,7 @@
#endif
#ifdef FIXED_POINT
-static const spx_float_t MIN_LEAK = ((spx_float_t){16777, -24});
+static const spx_float_t MIN_LEAK = {16777, -24};
#define TOP16(x) ((x)>>16)
#else
static const spx_float_t MIN_LEAK = .001f;
@@ -140,9 +140,13 @@
spx_word16_t preemph;
spx_word16_t notch_radius;
spx_mem_t notch_mem[2];
+
+ /* NOTE: If you only use speex_echo_cancel() and want to save some memory, remove this */
+ spx_int16_t *play_buf;
+ int play_buf_pos;
};
-static inline void filter_dc_notch16(spx_int16_t *in, spx_word16_t radius, spx_word16_t *out, int len, spx_mem_t *mem)
+static inline void filter_dc_notch16(const spx_int16_t *in, spx_word16_t radius, spx_word16_t *out, int len, spx_mem_t *mem)
{
int i;
spx_word16_t den2;
@@ -166,17 +170,15 @@
}
}
-static inline spx_word32_t inner_prod(const spx_word16_t *x, const spx_word16_t *y, int len)
+static inline spx_word32_t mdf_inner_prod(const spx_word16_t *x, const spx_word16_t *y, int len)
{
spx_word32_t sum=0;
- len >>= 2;
+ len >>= 1;
while(len--)
{
spx_word32_t part=0;
part = MAC16_16(part,*x++,*y++);
part = MAC16_16(part,*x++,*y++);
- part = MAC16_16(part,*x++,*y++);
- part = MAC16_16(part,*x++,*y++);
/* HINT: If you had a 40-bit accumulator, you could shift only at the end */
sum = ADD32(sum,SHR32(part,6));
}
@@ -184,7 +186,7 @@
}
/** Compute power spectrum of a half-complex (packed) vector */
-static inline void power_spectrum(spx_word16_t *X, spx_word32_t *ps, int N)
+static inline void power_spectrum(const spx_word16_t *X, spx_word32_t *ps, int N)
{
int i, j;
ps[0]=MULT16_16(X[0],X[0]);
@@ -197,7 +199,7 @@
/** Compute cross-power spectrum of a half-complex (packed) vectors and add to acc */
#ifdef FIXED_POINT
-static inline void spectral_mul_accum(spx_word16_t *X, spx_word32_t *Y, spx_word16_t *acc, int N, int M)
+static inline void spectral_mul_accum(const spx_word16_t *X, const spx_word32_t *Y, spx_word16_t *acc, int N, int M)
{
int i,j;
spx_word32_t tmp1=0,tmp2=0;
@@ -225,7 +227,7 @@
acc[N-1] = PSHR32(tmp1,WEIGHT_SHIFT);
}
#else
-static inline void spectral_mul_accum(spx_word16_t *X, spx_word32_t *Y, spx_word16_t *acc, int N, int M)
+static inline void spectral_mul_accum(const spx_word16_t *X, const spx_word32_t *Y, spx_word16_t *acc, int N, int M)
{
int i,j;
for (i=0;i<N;i++)
@@ -246,7 +248,7 @@
#endif
/** Compute weighted cross-power spectrum of a half-complex (packed) vector with conjugate */
-static inline void weighted_spectral_mul_conj(spx_float_t *w, spx_word16_t *X, spx_word16_t *Y, spx_word32_t *prod, int N)
+static inline void weighted_spectral_mul_conj(const spx_float_t *w, const spx_word16_t *X, const spx_word16_t *Y, spx_word32_t *prod, int N)
{
int i, j;
prod[0] = FLOAT_MUL32(w[0],MULT16_16(X[0],Y[0]));
@@ -273,10 +275,10 @@
st->sum_adapt = 0;
/* FIXME: Make that an init option (new API call?) */
st->sampling_rate = 8000;
- st->spec_average = DIV32_16(SHL32(st->frame_size, 15), st->sampling_rate);
+ st->spec_average = DIV32_16(SHL32(EXTEND32(st->frame_size), 15), st->sampling_rate);
#ifdef FIXED_POINT
- st->beta0 = DIV32_16(SHL32(st->frame_size, 16), st->sampling_rate);
- st->beta_max = DIV32_16(SHL32(st->frame_size, 14), st->sampling_rate);
+ st->beta0 = DIV32_16(SHL32(EXTEND32(st->frame_size), 16), st->sampling_rate);
+ st->beta_max = DIV32_16(SHL32(EXTEND32(st->frame_size), 14), st->sampling_rate);
#else
st->beta0 = (2.0f*st->frame_size)/st->sampling_rate;
st->beta_max = (.5f*st->frame_size)/st->sampling_rate;
@@ -332,6 +334,10 @@
st->notch_mem[0] = st->notch_mem[1] = 0;
st->adapted = 0;
st->Pey = st->Pyy = FLOAT_ONE;
+
+ st->play_buf = (spx_int16_t*)speex_alloc(2*st->frame_size*sizeof(spx_int16_t));
+ st->play_buf_pos = 0;
+
return st;
}
@@ -385,12 +391,46 @@
#ifdef FIXED_POINT
speex_free(st->wtmp2);
#endif
+ speex_free(st->play_buf);
speex_free(st);
}
-extern int fixed_point;
+void speex_echo_capture(SpeexEchoState *st, const spx_int16_t *rec, spx_int16_t *out, spx_int32_t *Yout)
+{
+ int i;
+ if (st->play_buf_pos>=st->frame_size)
+ {
+ speex_echo_cancel(st, rec, st->play_buf, out, Yout);
+ st->play_buf_pos -= st->frame_size;
+ for (i=0;i<st->frame_size;i++)
+ st->play_buf[i] = st->play_buf[i+st->frame_size];
+ } else {
+ speex_warning("no playback frame available");
+ if (st->play_buf_pos!=0)
+ {
+ speex_warning("internal playback buffer corruption?");
+ st->play_buf_pos = 0;
+ }
+ for (i=0;i<st->frame_size;i++)
+ out[i] = rec[i];
+ }
+}
+
+void speex_echo_playback(SpeexEchoState *st, const spx_int16_t *play)
+{
+ if (st->play_buf_pos<=st->frame_size)
+ {
+ int i;
+ for (i=0;i<st->frame_size;i++)
+ st->play_buf[st->play_buf_pos+i] = play[i];
+ st->play_buf_pos += st->frame_size;
+ } else {
+ speex_warning("had to discard a playback frame");
+ }
+}
+
/** Performs echo cancellation on a frame */
-void speex_echo_cancel(SpeexEchoState *st, short *ref, short *echo, short *out, spx_int32_t *Yout)
+void speex_echo_cancel(SpeexEchoState *st, const spx_int16_t *ref, const spx_int16_t *echo, spx_int16_t *out, spx_int32_t *Yout)
{
int i,j;
int N,M;
@@ -402,6 +442,7 @@
spx_word16_t RER;
spx_word32_t tmp32;
spx_word16_t M_1;
+ int saturated=0;
N = st->window_size;
M = st->M;
@@ -416,18 +457,46 @@
M_1 = 1.f/M;
#endif
- filter_dc_notch16((spx_int16_t*)ref, st->notch_radius, st->d, st->frame_size, st->notch_mem);
+ filter_dc_notch16(ref, st->notch_radius, st->d, st->frame_size, st->notch_mem);
/* Copy input data to buffer */
for (i=0;i<st->frame_size;i++)
{
spx_word16_t tmp;
+ spx_word32_t tmp32;
st->x[i] = st->x[i+st->frame_size];
- st->x[i+st->frame_size] = SUB16(echo[i], MULT16_16_P15(st->preemph, st->memX));
+ tmp32 = SUB32(EXTEND32(echo[i]), EXTEND32(MULT16_16_P15(st->preemph, st->memX)));
+#ifdef FIXED_POINT
+ /*FIXME: If saturation occurs here, we need to freeze adaptation for M frames (not just one) */
+ if (tmp32 > 32767)
+ {
+ tmp32 = 32767;
+ saturated = 1;
+ }
+ if (tmp32 < -32767)
+ {
+ tmp32 = -32767;
+ saturated = 1;
+ }
+#endif
+ st->x[i+st->frame_size] = EXTRACT16(tmp32);
st->memX = echo[i];
tmp = st->d[i];
st->d[i] = st->d[i+st->frame_size];
- st->d[i+st->frame_size] = SUB16(tmp, MULT16_16_P15(st->preemph, st->memD));
+ tmp32 = SUB32(EXTEND32(tmp), EXTEND32(MULT16_16_P15(st->preemph, st->memD)));
+#ifdef FIXED_POINT
+ if (tmp32 > 32767)
+ {
+ tmp32 = 32767;
+ saturated = 1;
+ }
+ if (tmp32 < -32767)
+ {
+ tmp32 = -32767;
+ saturated = 1;
+ }
+#endif
+ st->d[i+st->frame_size] = tmp32;
st->memD = tmp;
}
@@ -465,6 +534,12 @@
else if (tmp_out<-32768)
tmp_out = -32768;
tmp_out = ADD32(tmp_out, EXTEND32(MULT16_16_P15(st->preemph, st->memE)));
+ /* This is an arbitrary test for saturation */
+ if (ref[i] <= -32000 || ref[i] >= 32000)
+ {
+ tmp_out = 0;
+ saturated = 1;
+ }
out[i] = tmp_out;
st->memE = tmp_out;
}
@@ -477,9 +552,9 @@
}
/* Compute a bunch of correlations */
- See = inner_prod(st->e+st->frame_size, st->e+st->frame_size, st->frame_size);
- See = ADD32(See, SHR32(10000,6));
- Syy = inner_prod(st->y+st->frame_size, st->y+st->frame_size, st->frame_size);
+ See = mdf_inner_prod(st->e+st->frame_size, st->e+st->frame_size, st->frame_size);
+ See = ADD32(See, SHR32(EXTEND32(10000),6));
+ Syy = mdf_inner_prod(st->y+st->frame_size, st->y+st->frame_size, st->frame_size);
/* Convert error to frequency domain */
spx_fft(st->fft_table, st->e, st->E);
@@ -544,8 +619,9 @@
st->Pey = FLOAT_MULT(MIN_LEAK,st->Pyy);
if (FLOAT_GT(st->Pey, st->Pyy))
st->Pey = st->Pyy;
- /* leak_estimate is the limear regression result */
+ /* leak_estimate is the linear regression result */
leak_estimate = FLOAT_EXTRACT16(FLOAT_SHL(FLOAT_DIVU(st->Pey, st->Pyy),14));
+ /* This looks like a stupid bug, but it's right (because we convert from Q14 to Q15) */
if (leak_estimate > 16383)
leak_estimate = 32767;
else
@@ -594,7 +670,7 @@
spx_word32_t Sxx;
spx_word16_t adapt_rate=0;
- Sxx = inner_prod(st->x+st->frame_size, st->x+st->frame_size, st->frame_size);
+ Sxx = mdf_inner_prod(st->x+st->frame_size, st->x+st->frame_size, st->frame_size);
/* Temporary adaption rate if filter is not adapted correctly */
tmp32 = MULT16_32_Q15(QCONST16(.15f, 15), Sxx);
@@ -620,12 +696,15 @@
weighted_spectral_mul_conj(st->power_1, &st->X[j*N], st->E, st->PHI+N*j, N);
}
- /* Gradient descent */
- for (i=0;i<M*N;i++)
+ if (!saturated)
{
- st->W[i] += st->PHI[i];
- /* Old value of W in PHI */
- st->PHI[i] = st->W[i] - st->PHI[i];
+ /* Gradient descent */
+ for (i=0;i<M*N;i++)
+ {
+ st->W[i] += st->PHI[i];
+ /* Old value of W in PHI */
+ st->PHI[i] = st->W[i] - st->PHI[i];
+ }
}
/* Update weight to prevent circular convolution (MDF / AUMDF) */
@@ -637,7 +716,7 @@
{
#ifdef FIXED_POINT
for (i=0;i<N;i++)
- st->wtmp2[i] = PSHR32(st->W[j*N+i],NORMALIZE_SCALEDOWN+16);
+ st->wtmp2[i] = EXTRACT16(PSHR32(st->W[j*N+i],NORMALIZE_SCALEDOWN+16));
spx_ifft(st->fft_table, st->wtmp2, st->wtmp);
for (i=0;i<st->frame_size;i++)
{
@@ -645,12 +724,12 @@
}
for (i=st->frame_size;i<N;i++)
{
- st->wtmp[i]=SHL(st->wtmp[i],NORMALIZE_SCALEUP);
+ st->wtmp[i]=SHL16(st->wtmp[i],NORMALIZE_SCALEUP);
}
spx_fft(st->fft_table, st->wtmp, st->wtmp2);
/* The "-1" in the shift is a sort of kludge that trades less efficient update speed for decrease noise */
for (i=0;i<N;i++)
- st->W[j*N+i] -= SHL32(st->wtmp2[i],16+NORMALIZE_SCALEDOWN-NORMALIZE_SCALEUP-1);
+ st->W[j*N+i] -= SHL32(EXTEND32(st->wtmp2[i]),16+NORMALIZE_SCALEDOWN-NORMALIZE_SCALEUP-1);
#else
spx_ifft(st->fft_table, &st->W[j*N], st->wtmp);
for (i=st->frame_size;i<N;i++)
@@ -715,10 +794,10 @@
break;
case SPEEX_ECHO_SET_SAMPLING_RATE:
st->sampling_rate = (*(int*)ptr);
- st->spec_average = DIV32_16(SHL32(st->frame_size, 15), st->sampling_rate);
+ st->spec_average = DIV32_16(SHL32(EXTEND32(st->frame_size), 15), st->sampling_rate);
#ifdef FIXED_POINT
- st->beta0 = DIV32_16(SHL32(st->frame_size, 16), st->sampling_rate);
- st->beta_max = DIV32_16(SHL32(st->frame_size, 14), st->sampling_rate);
+ st->beta0 = DIV32_16(SHL32(EXTEND32(st->frame_size), 16), st->sampling_rate);
+ st->beta_max = DIV32_16(SHL32(EXTEND32(st->frame_size), 14), st->sampling_rate);
#else
st->beta0 = (2.0f*st->frame_size)/st->sampling_rate;
st->beta_max = (.5f*st->frame_size)/st->sampling_rate;
diff --git a/pjmedia/src/pjmedia-codec/speex/misc.c b/pjmedia/src/pjmedia-codec/speex/misc.c
index fc2cb7c..53bdd0b 100644
--- a/pjmedia/src/pjmedia-codec/speex/misc.c
+++ b/pjmedia/src/pjmedia-codec/speex/misc.c
@@ -196,12 +196,12 @@
#endif
#ifdef FIXED_POINT
-spx_word32_t speex_rand(spx_word16_t std, spx_int32_t *seed)
+spx_word16_t speex_rand(spx_word16_t std, spx_int32_t *seed)
{
spx_word32_t res;
*seed = 1664525 * *seed + 1013904223;
res = MULT16_16(EXTRACT16(SHR32(*seed,16)),std);
- return SUB32(res, SHR(res, 3));
+ return PSHR32(SUB32(res, SHR(res, 3)),14);
}
#else
spx_word16_t speex_rand(spx_word16_t std, spx_int32_t *seed)
@@ -216,19 +216,6 @@
}
#endif
-void speex_rand_vec(float std, spx_sig_t *data, int len)
-{
- int i;
- for (i=0;i<len;i++)
- data[i]+=SIG_SCALING*3*std*((((float)rand())/RAND_MAX)-.5);
-}
-
-
-/*float speex_rand(float std)
-{
- return 3*std*((((float)rand())/RAND_MAX)-.5);
-}*/
-
#ifndef OVERRIDE_SPEEX_PUTC
void _speex_putc(int ch, void *file)
{
diff --git a/pjmedia/src/pjmedia-codec/speex/misc.h b/pjmedia/src/pjmedia-codec/speex/misc.h
index 2e69b20..95e5d0c 100644
--- a/pjmedia/src/pjmedia-codec/speex/misc.h
+++ b/pjmedia/src/pjmedia-codec/speex/misc.h
@@ -38,9 +38,34 @@
#ifndef SPEEX_VERSION
#define SPEEX_MAJOR_VERSION 1 /**< Major Speex version. */
#define SPEEX_MINOR_VERSION 1 /**< Minor Speex version. */
-#define SPEEX_MICRO_VERSION 12 /**< Micro Speex version. */
+#define SPEEX_MICRO_VERSION 13 /**< Micro Speex version. */
#define SPEEX_EXTRA_VERSION "" /**< Extra Speex version. */
-#define SPEEX_VERSION "speex-1.1.12" /**< Speex version string. */
+#define SPEEX_VERSION "speex-1.1.13" /**< Speex version string. */
+#endif
+
+/* A couple test to catch stupid option combinations */
+#ifdef FIXED_POINT
+
+#ifdef _USE_SSE
+#error SSE is only for floating-point
+#endif
+#if ((defined (ARM4_ASM)||defined (ARM4_ASM)) && defined(BFIN_ASM)) || (defined (ARM4_ASM)&&defined(ARM5E_ASM))
+#error Make up your mind. What CPU do you have?
+#endif
+#ifdef VORBIS_PSYCHO
+#error Vorbis-psy model currently not implemented in fixed-point
+#endif
+
+#else
+
+#if defined (ARM4_ASM) || defined(ARM5E_ASM) || defined(BFIN_ASM)
+#error I suppose you can have a [ARM4/ARM5E/Blackfin] that has float instructions?
+#endif
+#ifdef FIXED_POINT_DEBUG
+#error Don't you think enabling fixed-point is a good thing to do if you want to debug that?
+#endif
+
+
#endif
#include "arch.h"
@@ -88,11 +113,8 @@
/** Print warning message with integer argument to stderr */
void speex_warning_int(const char *str, int val);
-/** Generate a vector of random numbers */
-void speex_rand_vec(float std, spx_sig_t *data, int len);
-
/** Generate a random number */
-spx_word32_t speex_rand(spx_word16_t std, spx_int32_t *seed);
+spx_word16_t speex_rand(spx_word16_t std, spx_int32_t *seed);
/** Speex wrapper for putc */
void _speex_putc(int ch, void *file);
diff --git a/pjmedia/src/pjmedia-codec/speex/modes.c b/pjmedia/src/pjmedia-codec/speex/modes.c
index aa15729..1da236d 100644
--- a/pjmedia/src/pjmedia-codec/speex/modes.c
+++ b/pjmedia/src/pjmedia-codec/speex/modes.c
@@ -190,11 +190,7 @@
noise_codebook_quant,
noise_codebook_unquant,
NULL,
-#ifdef FIXED_POINT
- 22938, 22938, 0, -1,
-#else
- .7, .7, 0, -1,
-#endif
+ -1,
43
};
@@ -215,11 +211,7 @@
split_cb_search_shape_sign,
split_cb_shape_sign_unquant,
&split_cb_nb_ulbr,
-#ifdef FIXED_POINT
- 22938, 16384, 11796, 21299,
-#else
- 0.7, 0.5, .36, .65,
-#endif
+ QCONST16(.65,15),
79
};
@@ -240,11 +232,7 @@
split_cb_search_shape_sign,
split_cb_shape_sign_unquant,
&split_cb_nb_vlbr,
-#ifdef FIXED_POINT
- 22938, 16384, 11796, 18022,
-#else
- 0.7, 0.5, .36, .55,
-#endif
+ QCONST16(.55,15),
119
};
@@ -265,11 +253,7 @@
split_cb_search_shape_sign,
split_cb_shape_sign_unquant,
&split_cb_nb_lbr,
-#ifdef FIXED_POINT
- 22938, 18022, 9830, 14746,
-#else
- 0.7, 0.55, .30, .45,
-#endif
+ QCONST16(.45,15),
160
};
@@ -290,11 +274,7 @@
split_cb_search_shape_sign,
split_cb_shape_sign_unquant,
&split_cb_nb_med,
-#ifdef FIXED_POINT
- 22938, 20644, 5243, 11469,
-#else
- 0.7, 0.63, .16, .35,
-#endif
+ QCONST16(.35,15),
220
};
@@ -315,11 +295,7 @@
split_cb_search_shape_sign,
split_cb_shape_sign_unquant,
&split_cb_nb,
-#ifdef FIXED_POINT
- 22938, 21299, 3932, 8192,
-#else
- 0.7, 0.65, .12, .25,
-#endif
+ QCONST16(.2,15),
300
};
@@ -340,11 +316,7 @@
split_cb_search_shape_sign,
split_cb_shape_sign_unquant,
&split_cb_sb,
-#ifdef FIXED_POINT
- 22282, 21299, 2294, 3277,
-#else
- 0.68, 0.65, .07, .1,
-#endif
+ QCONST16(.1,15),
364
};
@@ -365,11 +337,7 @@
split_cb_search_shape_sign,
split_cb_shape_sign_unquant,
&split_cb_nb,
-#ifdef FIXED_POINT
- 21299, 21299, 0, -1,
-#else
- 0.65, 0.65, .0, -1,
-#endif
+ -1,
492
};
@@ -434,11 +402,7 @@
NULL,
NULL,
NULL,
-#ifdef FIXED_POINT
- 24576, 24576, 0, -1,
-#else
- .75, .75, .0, -1,
-#endif
+ -1,
36
};
@@ -463,11 +427,7 @@
#else
&split_cb_high_lbr,
#endif
-#ifdef FIXED_POINT
- 27853, 19661, 8192, -1,
-#else
- .85, .6, .25, -1,
-#endif
+ -1,
112
};
@@ -492,12 +452,7 @@
#else
&split_cb_high,
#endif
-
-#ifdef FIXED_POINT
- 24576, 22938, 1638, -1,
-#else
- .75, .7, .05, -1,
-#endif
+ -1,
192
};
@@ -521,11 +476,7 @@
#else
&split_cb_high,
#endif
-#ifdef FIXED_POINT
- 24576, 24576, 0, -1,
-#else
- .75, .75, .0, -1,
-#endif
+ -1,
352
};
@@ -542,8 +493,8 @@
#else
0.9, 0.6, /* gamma1, gamma2 */
#endif
- .001, /*lag_factor*/
- QCONST16(.0001,15), /*lpc_floor*/
+ .012, /*lag_factor*/
+ QCONST16(.0002,15), /*lpc_floor*/
0.9,
{NULL, &wb_submode1, &wb_submode2, &wb_submode3, &wb_submode4, NULL, NULL, NULL},
3,
@@ -588,8 +539,8 @@
#else
0.9, 0.6, /* gamma1, gamma2 */
#endif
- .002, /*lag_factor*/
- QCONST16(.0001,15), /*lpc_floor*/
+ .012, /*lag_factor*/
+ QCONST16(.0002,15), /*lpc_floor*/
0.7,
{NULL, &wb_submode1, NULL, NULL, NULL, NULL, NULL, NULL},
1,
diff --git a/pjmedia/src/pjmedia-codec/speex/modes.h b/pjmedia/src/pjmedia-codec/speex/modes.h
index 945d004..9828ee6 100644
--- a/pjmedia/src/pjmedia-codec/speex/modes.h
+++ b/pjmedia/src/pjmedia-codec/speex/modes.h
@@ -55,17 +55,17 @@
/** Long-term predictor quantization */
-typedef int (*ltp_quant_func)(spx_sig_t *, spx_sig_t *, spx_coef_t *, spx_coef_t *,
+typedef int (*ltp_quant_func)(spx_word16_t *, spx_word16_t *, spx_coef_t *, spx_coef_t *,
spx_coef_t *, spx_sig_t *, const void *, int, int, spx_word16_t,
- int, int, SpeexBits*, char *, spx_sig_t *, spx_word16_t *, int, int, int);
+ int, int, SpeexBits*, char *, spx_word16_t *, spx_word16_t *, int, int, int, spx_word32_t *);
/** Long-term un-quantize */
-typedef void (*ltp_unquant_func)(spx_sig_t *, int, int, spx_word16_t, const void *, int, int *,
+typedef void (*ltp_unquant_func)(spx_word16_t *, spx_word32_t *, int, int, spx_word16_t, const void *, int, int *,
spx_word16_t *, SpeexBits*, char*, int, int, spx_word16_t, int);
/** Innovation quantization function */
-typedef void (*innovation_quant_func)(spx_sig_t *, spx_coef_t *, spx_coef_t *, spx_coef_t *, const void *, int, int,
+typedef void (*innovation_quant_func)(spx_word16_t *, spx_coef_t *, spx_coef_t *, spx_coef_t *, const void *, int, int,
spx_sig_t *, spx_word16_t *, SpeexBits *, char *, int, int);
/** Innovation unquantization function */
@@ -84,17 +84,13 @@
/*Lont-term predictor functions*/
ltp_quant_func ltp_quant; /**< Long-term predictor (pitch) quantizer */
ltp_unquant_func ltp_unquant; /**< Long-term predictor (pitch) un-quantizer */
- const void *ltp_params; /**< Pitch parameters (options) */
+ const void *ltp_params; /**< Pitch parameters (options) */
/*Quantization of innovation*/
innovation_quant_func innovation_quant; /**< Innovation quantization */
innovation_unquant_func innovation_unquant; /**< Innovation un-quantization */
const void *innovation_params; /**< Innovation quantization parameters*/
- /*Synthesis filter enhancement*/
- spx_word16_t lpc_enh_k1; /**< Enhancer constant */
- spx_word16_t lpc_enh_k2; /**< Enhancer constant */
- spx_word16_t lpc_enh_k3; /**< Enhancer constant */
spx_word16_t comb_gain; /**< Gain of enhancer comb filter */
int bits_per_frame; /**< Number of bits per frame after encoding*/
diff --git a/pjmedia/src/pjmedia-codec/speex/nb_celp.c b/pjmedia/src/pjmedia-codec/speex/nb_celp.c
index f6f5c69..2c41649 100644
--- a/pjmedia/src/pjmedia-codec/speex/nb_celp.c
+++ b/pjmedia/src/pjmedia-codec/speex/nb_celp.c
@@ -46,6 +46,7 @@
#include <speex/speex_bits.h>
#include "vbr.h"
#include "misc.h"
+#include "math_approx.h"
#include <speex/speex_callbacks.h>
#ifdef VORBIS_PSYCHO
@@ -106,6 +107,8 @@
#define sqr(x) ((x)*(x))
+extern const spx_word16_t lpc_window[];
+
void *nb_encoder_init(const SpeexMode *m)
{
EncState *st;
@@ -125,9 +128,9 @@
st->mode=m;
st->frameSize = mode->frameSize;
- st->windowSize = st->frameSize*3/2;
st->nbSubframes=mode->frameSize/mode->subframeSize;
st->subframeSize=mode->subframeSize;
+ st->windowSize = st->frameSize+st->subframeSize;
st->lpcSize = mode->lpcSize;
st->gamma1=mode->gamma1;
st->gamma2=mode->gamma2;
@@ -149,69 +152,50 @@
st->psy = vorbis_psy_init(8000, 256);
st->curve = speex_alloc(128*sizeof(float));
st->old_curve = speex_alloc(128*sizeof(float));
+ st->psy_window = speex_alloc(256*sizeof(float));
#endif
+ st->cumul_gain = 1024;
+
/* Allocating input buffer */
- st->inBuf = speex_alloc((st->windowSize+EXTRA_BUFFER)*sizeof(spx_sig_t));
- st->frame = st->inBuf+EXTRA_BUFFER;
+ st->winBuf = speex_alloc((st->windowSize-st->frameSize)*sizeof(spx_word16_t));
/* Allocating excitation buffer */
- st->excBuf = speex_alloc((mode->frameSize+mode->pitchEnd+1)*sizeof(spx_sig_t));
- st->exc = st->excBuf + mode->pitchEnd + 1;
- st->swBuf = speex_alloc((mode->frameSize+mode->pitchEnd+1)*sizeof(spx_sig_t));
- st->sw = st->swBuf + mode->pitchEnd + 1;
+ st->excBuf = speex_alloc((mode->frameSize+mode->pitchEnd+2)*sizeof(spx_word16_t));
+ st->exc = st->excBuf + mode->pitchEnd + 2;
+ st->swBuf = speex_alloc((mode->frameSize+mode->pitchEnd+2)*sizeof(spx_word16_t));
+ st->sw = st->swBuf + mode->pitchEnd + 2;
- st->innov = speex_alloc((st->frameSize)*sizeof(spx_sig_t));
-
- /* Asymmetric "pseudo-Hamming" window */
- {
- int part1, part2;
- part1=st->frameSize - (st->subframeSize>>1);
- part2=(st->frameSize>>1) + (st->subframeSize>>1);
- st->window = speex_alloc((st->windowSize)*sizeof(spx_word16_t));
- for (i=0;i<part1;i++)
- st->window[i]=(spx_word16_t)(SIG_SCALING*(.54-.46*cos(M_PI*i/part1)));
- for (i=0;i<part2;i++)
- st->window[part1+i]=(spx_word16_t)(SIG_SCALING*(.54+.46*cos(M_PI*i/part2)));
- }
+ st->window= lpc_window;
+
/* Create the window for autocorrelation (lag-windowing) */
st->lagWindow = speex_alloc((st->lpcSize+1)*sizeof(spx_word16_t));
for (i=0;i<st->lpcSize+1;i++)
st->lagWindow[i]=16384*exp(-.5*sqr(2*M_PI*st->lag_factor*i));
- st->autocorr = speex_alloc((st->lpcSize+1)*sizeof(spx_word16_t));
-
- st->lpc = speex_alloc((st->lpcSize)*sizeof(spx_coef_t));
- st->interp_lpc = speex_alloc((st->lpcSize)*sizeof(spx_coef_t));
- st->interp_qlpc = speex_alloc((st->lpcSize)*sizeof(spx_coef_t));
- st->bw_lpc1 = speex_alloc((st->lpcSize)*sizeof(spx_coef_t));
- st->bw_lpc2 = speex_alloc((st->lpcSize)*sizeof(spx_coef_t));
-
- st->lsp = speex_alloc((st->lpcSize)*sizeof(spx_lsp_t));
- st->qlsp = speex_alloc((st->lpcSize)*sizeof(spx_lsp_t));
st->old_lsp = speex_alloc((st->lpcSize)*sizeof(spx_lsp_t));
st->old_qlsp = speex_alloc((st->lpcSize)*sizeof(spx_lsp_t));
- st->interp_lsp = speex_alloc((st->lpcSize)*sizeof(spx_lsp_t));
- st->interp_qlsp = speex_alloc((st->lpcSize)*sizeof(spx_lsp_t));
-
st->first = 1;
for (i=0;i<st->lpcSize;i++)
{
- st->lsp[i]=LSP_SCALING*(M_PI*((float)(i+1)))/(st->lpcSize+1);
+ st->old_lsp[i]=LSP_SCALING*(M_PI*((float)(i+1)))/(st->lpcSize+1);
}
st->mem_sp = speex_alloc((st->lpcSize)*sizeof(spx_mem_t));
st->mem_sw = speex_alloc((st->lpcSize)*sizeof(spx_mem_t));
st->mem_sw_whole = speex_alloc((st->lpcSize)*sizeof(spx_mem_t));
st->mem_exc = speex_alloc((st->lpcSize)*sizeof(spx_mem_t));
+ st->mem_exc2 = speex_alloc((st->lpcSize)*sizeof(spx_mem_t));
st->pi_gain = speex_alloc((st->nbSubframes)*sizeof(spx_word32_t));
-
+ st->innov_save = NULL;
+
st->pitch = speex_alloc((st->nbSubframes)*sizeof(int));
st->vbr = speex_alloc(sizeof(VBRState));
vbr_init(st->vbr);
st->vbr_quality = 8;
st->vbr_enabled = 0;
+ st->vbr_max = 0;
st->vad_enabled = 0;
st->dtx_enabled = 0;
st->abr_enabled = 0;
@@ -236,30 +220,19 @@
speex_free_scratch(st->stack);
#endif
- speex_free (st->inBuf);
+ speex_free (st->winBuf);
speex_free (st->excBuf);
- speex_free (st->innov);
- speex_free (st->interp_qlpc);
- speex_free (st->qlsp);
speex_free (st->old_qlsp);
- speex_free (st->interp_qlsp);
speex_free (st->swBuf);
- speex_free (st->window);
speex_free (st->lagWindow);
- speex_free (st->autocorr);
- speex_free (st->lpc);
- speex_free (st->lsp);
- speex_free (st->interp_lpc);
- speex_free (st->bw_lpc1);
- speex_free (st->bw_lpc2);
speex_free (st->old_lsp);
- speex_free (st->interp_lsp);
speex_free (st->mem_sp);
speex_free (st->mem_sw);
speex_free (st->mem_sw_whole);
speex_free (st->mem_exc);
+ speex_free (st->mem_exc2);
speex_free (st->pi_gain);
speex_free (st->pitch);
@@ -270,6 +243,7 @@
vorbis_psy_destroy(st->psy);
speex_free (st->curve);
speex_free (st->old_curve);
+ speex_free (st->psy_window);
#endif
/*Free state memory... should be last*/
@@ -283,12 +257,23 @@
int ol_pitch;
spx_word16_t ol_pitch_coef;
spx_word32_t ol_gain;
- VARDECL(spx_sig_t *res);
- VARDECL(spx_sig_t *target);
+ VARDECL(spx_word16_t *ringing);
+ VARDECL(spx_word16_t *target);
+ VARDECL(spx_sig_t *innov);
+ VARDECL(spx_word32_t *exc32);
VARDECL(spx_mem_t *mem);
+ VARDECL(spx_coef_t *bw_lpc1);
+ VARDECL(spx_coef_t *bw_lpc2);
+ VARDECL(spx_coef_t *lpc);
+ VARDECL(spx_lsp_t *lsp);
+ VARDECL(spx_lsp_t *qlsp);
+ VARDECL(spx_lsp_t *interp_lsp);
+ VARDECL(spx_lsp_t *interp_qlsp);
+ VARDECL(spx_coef_t *interp_lpc);
+ VARDECL(spx_coef_t *interp_qlpc);
char *stack;
VARDECL(spx_word16_t *syn_resp);
- VARDECL(spx_sig_t *real_exc);
+ VARDECL(spx_word16_t *real_exc);
#ifdef EPIC_48K
int pitch_half[2];
int ol_pitch_id=0;
@@ -298,79 +283,85 @@
st=(EncState *)state;
stack=st->stack;
- /* Copy new data in input buffer */
- speex_move(st->inBuf, st->inBuf+st->frameSize, (EXTRA_BUFFER+st->windowSize-st->frameSize)*sizeof(spx_sig_t));
- for (i=0;i<st->frameSize;i++)
- st->inBuf[st->windowSize-st->frameSize+i+EXTRA_BUFFER] = SHL32(EXTEND32(in[i]), SIG_SHIFT);
+ ALLOC(lpc, st->lpcSize, spx_coef_t);
+ ALLOC(bw_lpc1, st->lpcSize, spx_coef_t);
+ ALLOC(bw_lpc2, st->lpcSize, spx_coef_t);
+ ALLOC(lsp, st->lpcSize, spx_lsp_t);
+ ALLOC(qlsp, st->lpcSize, spx_lsp_t);
+ ALLOC(interp_lsp, st->lpcSize, spx_lsp_t);
+ ALLOC(interp_qlsp, st->lpcSize, spx_lsp_t);
+ ALLOC(interp_lpc, st->lpcSize, spx_coef_t);
+ ALLOC(interp_qlpc, st->lpcSize, spx_coef_t);
/* Move signals 1 frame towards the past */
- speex_move(st->excBuf, st->excBuf+st->frameSize, (st->max_pitch+1)*sizeof(spx_sig_t));
- speex_move(st->swBuf, st->swBuf+st->frameSize, (st->max_pitch+1)*sizeof(spx_sig_t));
+ speex_move(st->excBuf, st->excBuf+st->frameSize, (st->max_pitch+2)*sizeof(spx_word16_t));
+ speex_move(st->swBuf, st->swBuf+st->frameSize, (st->max_pitch+2)*sizeof(spx_word16_t));
{
VARDECL(spx_word16_t *w_sig);
+ VARDECL(spx_word16_t *autocorr);
ALLOC(w_sig, st->windowSize, spx_word16_t);
+ ALLOC(autocorr, st->lpcSize+1, spx_word16_t);
/* Window for analysis */
- for (i=0;i<st->windowSize;i++)
- w_sig[i] = EXTRACT16(SHR32(MULT16_16(EXTRACT16(SHR32(st->frame[i],SIG_SHIFT)),st->window[i]),SIG_SHIFT));
-
+ for (i=0;i<st->windowSize-st->frameSize;i++)
+ w_sig[i] = EXTRACT16(SHR32(MULT16_16(st->winBuf[i],st->window[i]),SIG_SHIFT));
+ for (;i<st->windowSize;i++)
+ w_sig[i] = EXTRACT16(SHR32(MULT16_16(in[i-st->windowSize+st->frameSize],st->window[i]),SIG_SHIFT));
/* Compute auto-correlation */
- _spx_autocorr(w_sig, st->autocorr, st->lpcSize+1, st->windowSize);
- }
- st->autocorr[0] = ADD16(st->autocorr[0],MULT16_16_Q15(st->autocorr[0],st->lpc_floor)); /* Noise floor in auto-correlation domain */
+ _spx_autocorr(w_sig, autocorr, st->lpcSize+1, st->windowSize);
+ autocorr[0] = ADD16(autocorr[0],MULT16_16_Q15(autocorr[0],st->lpc_floor)); /* Noise floor in auto-correlation domain */
- /* Lag windowing: equivalent to filtering in the power-spectrum domain */
- for (i=0;i<st->lpcSize+1;i++)
- st->autocorr[i] = MULT16_16_Q14(st->autocorr[i],st->lagWindow[i]);
+ /* Lag windowing: equivalent to filtering in the power-spectrum domain */
+ for (i=0;i<st->lpcSize+1;i++)
+ autocorr[i] = MULT16_16_Q14(autocorr[i],st->lagWindow[i]);
- /* Levinson-Durbin */
- _spx_lpc(st->lpc, st->autocorr, st->lpcSize);
-
- /* LPC to LSPs (x-domain) transform */
- roots=lpc_to_lsp (st->lpc, st->lpcSize, st->lsp, 15, LSP_DELTA1, stack);
- /* Check if we found all the roots */
- if (roots!=st->lpcSize)
- {
- /* Search again if we can afford it */
- if (st->complexity>1)
- roots = lpc_to_lsp (st->lpc, st->lpcSize, st->lsp, 11, LSP_DELTA2, stack);
- if (roots!=st->lpcSize)
+ /* Levinson-Durbin */
+ _spx_lpc(lpc, autocorr, st->lpcSize);
+ /* LPC to LSPs (x-domain) transform */
+ roots=lpc_to_lsp (lpc, st->lpcSize, lsp, 10, LSP_DELTA1, stack);
+ /* Check if we found all the roots */
+ if (roots!=st->lpcSize)
{
/*If we can't find all LSP's, do some damage control and use previous filter*/
for (i=0;i<st->lpcSize;i++)
{
- st->lsp[i]=st->old_lsp[i];
+ lsp[i]=st->old_lsp[i];
}
}
}
+
/* Whole frame analysis (open-loop estimation of pitch and excitation gain) */
{
if (st->first)
for (i=0;i<st->lpcSize;i++)
- st->interp_lsp[i] = st->lsp[i];
+ interp_lsp[i] = lsp[i];
else
- lsp_interpolate(st->old_lsp, st->lsp, st->interp_lsp, st->lpcSize, st->nbSubframes, st->nbSubframes<<1);
+ lsp_interpolate(st->old_lsp, lsp, interp_lsp, st->lpcSize, st->nbSubframes, st->nbSubframes<<1);
- lsp_enforce_margin(st->interp_lsp, st->lpcSize, LSP_MARGIN);
+ lsp_enforce_margin(interp_lsp, st->lpcSize, LSP_MARGIN);
/* Compute interpolated LPCs (unquantized) for whole frame*/
- lsp_to_lpc(st->interp_lsp, st->interp_lpc, st->lpcSize,stack);
+ lsp_to_lpc(interp_lsp, interp_lpc, st->lpcSize,stack);
/*Open-loop pitch*/
- if (!st->submodes[st->submodeID] || st->vbr_enabled || st->vad_enabled || SUBMODE(forced_pitch_gain) ||
+ if (st->complexity>2 || !st->submodes[st->submodeID] || st->vbr_enabled || st->vad_enabled || SUBMODE(forced_pitch_gain) ||
SUBMODE(lbr_pitch) != -1)
{
int nol_pitch[6];
spx_word16_t nol_pitch_coef[6];
- bw_lpc(st->gamma1, st->interp_lpc, st->bw_lpc1, st->lpcSize);
- bw_lpc(st->gamma2, st->interp_lpc, st->bw_lpc2, st->lpcSize);
+ bw_lpc(st->gamma1, interp_lpc, bw_lpc1, st->lpcSize);
+ bw_lpc(st->gamma2, interp_lpc, bw_lpc2, st->lpcSize);
- filter_mem2(st->frame, st->bw_lpc1, st->bw_lpc2, st->sw, st->frameSize, st->lpcSize, st->mem_sw_whole);
+ for (i=0;i<st->windowSize-st->frameSize;i++)
+ st->sw[i] = st->winBuf[i];
+ for (;i<st->frameSize;i++)
+ st->sw[i] = in[i-st->windowSize+st->frameSize];
+ filter_mem16(st->sw, bw_lpc1, bw_lpc2, st->sw, st->frameSize, st->lpcSize, st->mem_sw_whole, stack);
open_loop_nbest_pitch(st->sw, st->min_pitch, st->max_pitch, st->frameSize,
nol_pitch, nol_pitch_coef, 6, stack);
@@ -412,8 +403,13 @@
ol_pitch=0;
ol_pitch_coef=0;
}
+
/*Compute "real" excitation*/
- fir_mem2(st->frame, st->interp_lpc, st->exc, st->frameSize, st->lpcSize, st->mem_exc);
+ for (i=0;i<st->windowSize-st->frameSize;i++)
+ st->exc[i] = st->winBuf[i];
+ for (;i<st->frameSize;i++)
+ st->exc[i] = in[i-st->windowSize+st->frameSize];
+ fir_mem16(st->exc, interp_lpc, st->exc, st->frameSize, st->lpcSize, st->mem_exc, stack);
/* Compute open-loop excitation gain */
#ifdef EPIC_48K
@@ -421,8 +417,8 @@
{
float ol1=0,ol2=0;
float ol_gain2;
- ol1 = compute_rms(st->exc, st->frameSize>>1);
- ol2 = compute_rms(st->exc+(st->frameSize>>1), st->frameSize>>1);
+ ol1 = compute_rms16(st->exc, st->frameSize>>1);
+ ol2 = compute_rms16(st->exc+(st->frameSize>>1), st->frameSize>>1);
ol1 *= ol1*(st->frameSize>>1);
ol2 *= ol2*(st->frameSize>>1);
@@ -433,16 +429,24 @@
ol_gain=SHR(sqrt(1+ol_gain2/st->frameSize),SIG_SHIFT);
- } else {
+ } else
#endif
- ol_gain = SHL32(EXTEND32(compute_rms(st->exc, st->frameSize)),SIG_SHIFT);
-#ifdef EPIC_48K
+ {
+ spx_word16_t g = compute_rms16(st->exc, st->frameSize);
+ if (ol_pitch>0)
+ ol_gain = MULT16_16(g, MULT16_16_Q14(QCONST16(1.1,14),
+ spx_sqrt(QCONST32(1.,28)-MULT16_32_Q15(QCONST16(.8,15),SHL32(MULT16_16(ol_pitch_coef,ol_pitch_coef),16)))));
+ else
+ ol_gain = SHL32(EXTEND32(g),SIG_SHIFT);
}
-#endif
}
#ifdef VORBIS_PSYCHO
- compute_curve(st->psy, st->frame-16, st->curve);
+ for(i=0;i<256-st->frameSize;i++)
+ st->psy_window[i] = st->psy_window[i+st->frameSize];
+ for(i=0;i<st->frameSize;i++)
+ st->psy_window[256-st->frameSize+i] = in[i];
+ compute_curve(st->psy, st->psy_window, st->curve);
/*print_vec(st->curve, 128, "curve");*/
if (st->first)
for (i=0;i<128;i++)
@@ -454,7 +458,7 @@
{
float lsp_dist=0;
for (i=0;i<st->lpcSize;i++)
- lsp_dist += (st->old_lsp[i] - st->lsp[i])*(st->old_lsp[i] - st->lsp[i]);
+ lsp_dist += (st->old_lsp[i] - lsp[i])*(st->old_lsp[i] - lsp[i]);
lsp_dist /= LSP_SCALING*LSP_SCALING;
if (st->abr_enabled)
@@ -518,7 +522,17 @@
}
speex_encoder_ctl(state, SPEEX_SET_MODE, &mode);
-
+ if (st->vbr_max>0)
+ {
+ spx_int32_t rate;
+ speex_encoder_ctl(state, SPEEX_GET_BITRATE, &rate);
+ if (rate > st->vbr_max)
+ {
+ rate = st->vbr_max;
+ speex_encoder_ctl(state, SPEEX_SET_BITRATE, &rate);
+ }
+ }
+
if (st->abr_enabled)
{
int bitrate;
@@ -580,13 +594,11 @@
st->first=1;
st->bounded_pitch = 1;
- /* Final signal synthesis from excitation */
- iir_mem2(st->exc, st->interp_qlpc, st->frame, st->frameSize, st->lpcSize, st->mem_sp);
+ speex_move(st->winBuf, in+2*st->frameSize-st->windowSize, (st->windowSize-st->frameSize)*sizeof(spx_word16_t));
-#ifdef RESYNTH
- for (i=0;i<st->frameSize;i++)
- in[i]=st->frame[i];
-#endif
+ /* Clear memory (no need to really compute it) */
+ for (i=0;i<st->lpcSize;i++)
+ st->mem_sp[i] = 0;
return 0;
}
@@ -595,16 +607,16 @@
if (st->first)
{
for (i=0;i<st->lpcSize;i++)
- st->old_lsp[i] = st->lsp[i];
+ st->old_lsp[i] = lsp[i];
}
/*Quantize LSPs*/
#if 1 /*0 for unquantized*/
- SUBMODE(lsp_quant)(st->lsp, st->qlsp, st->lpcSize, bits);
+ SUBMODE(lsp_quant)(lsp, qlsp, st->lpcSize, bits);
#else
for (i=0;i<st->lpcSize;i++)
- st->qlsp[i]=st->lsp[i];
+ qlsp[i]=lsp[i];
#endif
#ifdef EPIC_48K
@@ -685,22 +697,25 @@
if (st->first)
{
for (i=0;i<st->lpcSize;i++)
- st->old_qlsp[i] = st->qlsp[i];
+ st->old_qlsp[i] = qlsp[i];
}
- /* Filter response */
- ALLOC(res, st->subframeSize, spx_sig_t);
/* Target signal */
- ALLOC(target, st->subframeSize, spx_sig_t);
+ ALLOC(target, st->subframeSize, spx_word16_t);
+ ALLOC(innov, st->subframeSize, spx_sig_t);
+ ALLOC(exc32, st->subframeSize, spx_word32_t);
+ ALLOC(ringing, st->subframeSize, spx_word16_t);
ALLOC(syn_resp, st->subframeSize, spx_word16_t);
- ALLOC(real_exc, st->subframeSize, spx_sig_t);
+ ALLOC(real_exc, st->subframeSize, spx_word16_t);
ALLOC(mem, st->lpcSize, spx_mem_t);
/* Loop on sub-frames */
for (sub=0;sub<st->nbSubframes;sub++)
{
int offset;
- spx_sig_t *sp, *sw, *exc;
+ spx_word16_t *sw;
+ spx_word16_t *exc;
+ spx_sig_t *innov_save = NULL;
int pitch;
int response_bound = st->subframeSize;
#ifdef EPIC_48K
@@ -715,25 +730,26 @@
/* Offset relative to start of frame */
offset = st->subframeSize*sub;
- /* Original signal */
- sp=st->frame+offset;
/* Excitation */
exc=st->exc+offset;
/* Weighted signal */
sw=st->sw+offset;
-
+ /* Pointer for saving innovation */
+ if (st->innov_save)
+ innov_save = st->innov_save+offset;
+
/* LSP interpolation (quantized and unquantized) */
- lsp_interpolate(st->old_lsp, st->lsp, st->interp_lsp, st->lpcSize, sub, st->nbSubframes);
- lsp_interpolate(st->old_qlsp, st->qlsp, st->interp_qlsp, st->lpcSize, sub, st->nbSubframes);
+ lsp_interpolate(st->old_lsp, lsp, interp_lsp, st->lpcSize, sub, st->nbSubframes);
+ lsp_interpolate(st->old_qlsp, qlsp, interp_qlsp, st->lpcSize, sub, st->nbSubframes);
/* Make sure the filters are stable */
- lsp_enforce_margin(st->interp_lsp, st->lpcSize, LSP_MARGIN);
- lsp_enforce_margin(st->interp_qlsp, st->lpcSize, LSP_MARGIN);
+ lsp_enforce_margin(interp_lsp, st->lpcSize, LSP_MARGIN);
+ lsp_enforce_margin(interp_qlsp, st->lpcSize, LSP_MARGIN);
/* Compute interpolated LPCs (quantized and unquantized) */
- lsp_to_lpc(st->interp_lsp, st->interp_lpc, st->lpcSize,stack);
+ lsp_to_lpc(interp_lsp, interp_lpc, st->lpcSize,stack);
- lsp_to_lpc(st->interp_qlsp, st->interp_qlpc, st->lpcSize, stack);
+ lsp_to_lpc(interp_qlsp, interp_qlpc, st->lpcSize, stack);
/* Compute analysis filter gain at w=pi (for use in SB-CELP) */
{
@@ -741,7 +757,7 @@
for (i=0;i<st->lpcSize;i+=2)
{
/*pi_g += -st->interp_qlpc[i] + st->interp_qlpc[i+1];*/
- pi_g = ADD32(pi_g, SUB32(st->interp_qlpc[i+1],st->interp_qlpc[i]));
+ pi_g = ADD32(pi_g, SUB32(EXTEND32(interp_qlpc[i+1]),EXTEND32(interp_qlpc[i])));
}
st->pi_gain[sub] = pi_g;
}
@@ -752,56 +768,66 @@
float fact = ((float)sub+1.0f)/st->nbSubframes;
for (i=0;i<128;i++)
curr_curve[i] = (1.0f-fact)*st->old_curve[i] + fact*st->curve[i];
- curve_to_lpc(st->psy, curr_curve, st->bw_lpc1, st->bw_lpc2, 10);
+ curve_to_lpc(st->psy, curr_curve, bw_lpc1, bw_lpc2, 10);
}
#else
/* Compute bandwidth-expanded (unquantized) LPCs for perceptual weighting */
- bw_lpc(st->gamma1, st->interp_lpc, st->bw_lpc1, st->lpcSize);
+ bw_lpc(st->gamma1, interp_lpc, bw_lpc1, st->lpcSize);
if (st->gamma2>=0)
- bw_lpc(st->gamma2, st->interp_lpc, st->bw_lpc2, st->lpcSize);
+ bw_lpc(st->gamma2, interp_lpc, bw_lpc2, st->lpcSize);
else
{
- st->bw_lpc2[0]=1;
+ bw_lpc2[0]=1;
for (i=1;i<=st->lpcSize;i++)
- st->bw_lpc2[i]=0;
+ bw_lpc2[i]=0;
}
/*print_vec(st->bw_lpc1, 10, "bw_lpc");*/
#endif
- for (i=0;i<st->subframeSize;i++)
- real_exc[i] = exc[i];
+ {
+ /*FIXME: This will break if we change the window size */
+ if (st->windowSize-st->frameSize != st->subframeSize)
+ speex_error("windowSize-frameSize != subframeSize");
+ if (sub==0)
+ {
+ for (i=0;i<st->subframeSize;i++)
+ real_exc[i] = sw[i] = st->winBuf[i];
+ } else {
+ for (i=0;i<st->subframeSize;i++)
+ real_exc[i] = sw[i] = in[i+((sub-1)*st->subframeSize)];
+ }
+ }
+ fir_mem16(real_exc, interp_qlpc, real_exc, st->subframeSize, st->lpcSize, st->mem_exc2, stack);
if (st->complexity==0)
response_bound >>= 1;
- compute_impulse_response(st->interp_qlpc, st->bw_lpc1, st->bw_lpc2, syn_resp, response_bound, st->lpcSize, stack);
+ compute_impulse_response(interp_qlpc, bw_lpc1, bw_lpc2, syn_resp, response_bound, st->lpcSize, stack);
for (i=response_bound;i<st->subframeSize;i++)
syn_resp[i]=VERY_SMALL;
- /* Reset excitation */
- for (i=0;i<st->subframeSize;i++)
- exc[i]=VERY_SMALL;
-
/* Compute zero response of A(z/g1) / ( A(z/g2) * A(z) ) */
for (i=0;i<st->lpcSize;i++)
- mem[i]=st->mem_sp[i];
+ mem[i]=SHL32(st->mem_sp[i],1);
+ for (i=0;i<st->subframeSize;i++)
+ ringing[i] = VERY_SMALL;
#ifdef SHORTCUTS2
- iir_mem2(exc, st->interp_qlpc, exc, response_bound, st->lpcSize, mem);
+ iir_mem16(ringing, interp_qlpc, ringing, response_bound, st->lpcSize, mem, stack);
for (i=0;i<st->lpcSize;i++)
- mem[i]=st->mem_sw[i];
- filter_mem2(exc, st->bw_lpc1, st->bw_lpc2, res, response_bound, st->lpcSize, mem);
+ mem[i]=SHL32(st->mem_sw[i],1);
+ filter_mem16(ringing, st->bw_lpc1, st->bw_lpc2, ringing, response_bound, st->lpcSize, mem, stack);
for (i=response_bound;i<st->subframeSize;i++)
- res[i]=0;
+ ringing[i]=0;
#else
- iir_mem2(exc, st->interp_qlpc, exc, st->subframeSize, st->lpcSize, mem);
+ iir_mem16(ringing, interp_qlpc, ringing, st->subframeSize, st->lpcSize, mem, stack);
for (i=0;i<st->lpcSize;i++)
- mem[i]=st->mem_sw[i];
- filter_mem2(exc, st->bw_lpc1, st->bw_lpc2, res, st->subframeSize, st->lpcSize, mem);
+ mem[i]=SHL32(st->mem_sw[i],1);
+ filter_mem16(ringing, bw_lpc1, bw_lpc2, ringing, st->subframeSize, st->lpcSize, mem, stack);
#endif
/* Compute weighted signal */
for (i=0;i<st->lpcSize;i++)
mem[i]=st->mem_sw[i];
- filter_mem2(sp, st->bw_lpc1, st->bw_lpc2, sw, st->subframeSize, st->lpcSize, mem);
+ filter_mem16(sw, bw_lpc1, bw_lpc2, sw, st->subframeSize, st->lpcSize, mem, stack);
if (st->complexity==0)
for (i=0;i<st->lpcSize;i++)
@@ -809,8 +835,9 @@
/* Compute target signal */
for (i=0;i<st->subframeSize;i++)
- target[i]=sw[i]-res[i];
+ target[i]=SUB16(sw[i],PSHR32(ringing[i],1));
+ /* Reset excitation */
for (i=0;i<st->subframeSize;i++)
exc[i]=0;
@@ -847,18 +874,18 @@
#ifdef EPIC_48K
if (st->lbr_48k)
{
- pitch = SUBMODE(ltp_quant)(target, sw, st->interp_qlpc, st->bw_lpc1, st->bw_lpc2,
- exc, SUBMODE(ltp_params), pit_min, pit_max, ol_pitch_coef,
+ pitch = SUBMODE(ltp_quant)(target, sw, interp_qlpc, bw_lpc1, bw_lpc2,
+ exc32, SUBMODE(ltp_params), pit_min, pit_max, ol_pitch_coef,
st->lpcSize, st->subframeSize, bits, stack,
- exc, syn_resp, st->complexity, ol_pitch_id, st->plc_tuning);
+ exc, syn_resp, st->complexity, ol_pitch_id, st->plc_tuning, &st->cumul_gain);
} else {
#endif
/* Perform pitch search */
- pitch = SUBMODE(ltp_quant)(target, sw, st->interp_qlpc, st->bw_lpc1, st->bw_lpc2,
- exc, SUBMODE(ltp_params), pit_min, pit_max, ol_pitch_coef,
+ pitch = SUBMODE(ltp_quant)(target, sw, interp_qlpc, bw_lpc1, bw_lpc2,
+ exc32, SUBMODE(ltp_params), pit_min, pit_max, ol_pitch_coef,
st->lpcSize, st->subframeSize, bits, stack,
- exc, syn_resp, st->complexity, 0, st->plc_tuning);
+ exc, syn_resp, st->complexity, 0, st->plc_tuning, &st->cumul_gain);
#ifdef EPIC_48K
}
#endif
@@ -870,30 +897,28 @@
/* Quantization of innovation */
{
- spx_sig_t *innov;
spx_word32_t ener=0;
spx_word16_t fine_gain;
- innov = st->innov+sub*st->subframeSize;
for (i=0;i<st->subframeSize;i++)
innov[i]=0;
for (i=0;i<st->subframeSize;i++)
- real_exc[i] = SUB32(real_exc[i], exc[i]);
+ real_exc[i] = SUB16(real_exc[i], PSHR32(exc32[i],SIG_SHIFT-1));
- ener = SHL32(EXTEND32(compute_rms(real_exc, st->subframeSize)),SIG_SHIFT);
+ ener = SHL32(EXTEND32(compute_rms16(real_exc, st->subframeSize)),SIG_SHIFT);
/*FIXME: Should use DIV32_16 and make sure result fits in 16 bits */
#ifdef FIXED_POINT
{
- spx_word32_t f = DIV32(ener,PSHR32(ol_gain,SIG_SHIFT));
+ spx_word32_t f = PDIV32(ener,PSHR32(ol_gain,SIG_SHIFT));
if (f<=32767)
fine_gain = f;
else
fine_gain = 32767;
}
#else
- fine_gain = DIV32_16(ener,PSHR32(ol_gain,SIG_SHIFT));
+ fine_gain = PDIV32_16(ener,PSHR32(ol_gain,SIG_SHIFT));
#endif
/* Calculate gain correction for the sub-frame (if any) */
if (SUBMODE(have_subframe_gain))
@@ -922,7 +947,7 @@
if (SUBMODE(innovation_quant))
{
/* Codebook search */
- SUBMODE(innovation_quant)(target, st->interp_qlpc, st->bw_lpc1, st->bw_lpc2,
+ SUBMODE(innovation_quant)(target, interp_qlpc, bw_lpc1, bw_lpc2,
SUBMODE(innovation_params), st->lpcSize, st->subframeSize,
innov, syn_resp, bits, stack, st->complexity, SUBMODE(double_codebook));
@@ -930,11 +955,16 @@
signal_mul(innov, innov, ener, st->subframeSize);
for (i=0;i<st->subframeSize;i++)
- exc[i] = ADD32(exc[i],innov[i]);
+ exc[i] = EXTRACT16(PSHR32(ADD32(SHL32(exc32[i],1),innov[i]),SIG_SHIFT));
} else {
speex_error("No fixed codebook");
}
+ if (innov_save)
+ {
+ for (i=0;i<st->subframeSize;i++)
+ innov_save[i] = innov[i];
+ }
/* In some (rare) modes, we do a second search (more bits) to reduce noise even more */
if (SUBMODE(double_codebook)) {
char *tmp_stack=stack;
@@ -943,24 +973,31 @@
for (i=0;i<st->subframeSize;i++)
innov2[i]=0;
for (i=0;i<st->subframeSize;i++)
- target[i]*=2.2;
- SUBMODE(innovation_quant)(target, st->interp_qlpc, st->bw_lpc1, st->bw_lpc2,
+ target[i]=MULT16_16_P13(QCONST16(2.2,13), target[i]);
+ SUBMODE(innovation_quant)(target, interp_qlpc, bw_lpc1, bw_lpc2,
SUBMODE(innovation_params), st->lpcSize, st->subframeSize,
innov2, syn_resp, bits, stack, st->complexity, 0);
- signal_mul(innov2, innov2, (spx_word32_t) (ener*(1.f/2.2f)), st->subframeSize);
+ signal_mul(innov2, innov2, MULT16_32_Q15(QCONST16(0.454545,15),ener), st->subframeSize);
for (i=0;i<st->subframeSize;i++)
- exc[i] = ADD32(exc[i],innov2[i]);
+ exc[i] = ADD32(exc[i],PSHR32(innov2[i],SIG_SHIFT));
+ if (innov_save)
+ {
+ for (i=0;i<st->subframeSize;i++)
+ innov_save[i] = ADD32(innov_save[i],innov2[i]);
+ }
stack = tmp_stack;
}
}
+ for (i=0;i<st->subframeSize;i++)
+ sw[i] = exc[i];
/* Final signal synthesis from excitation */
- iir_mem2(exc, st->interp_qlpc, sp, st->subframeSize, st->lpcSize, st->mem_sp);
+ iir_mem16(sw, interp_qlpc, sw, st->subframeSize, st->lpcSize, st->mem_sp, stack);
/* Compute weighted signal again, from synthesized speech (not sure it's the right thing) */
if (st->complexity!=0)
- filter_mem2(sp, st->bw_lpc1, st->bw_lpc2, sw, st->subframeSize, st->lpcSize, st->mem_sw);
+ filter_mem16(sw, bw_lpc1, bw_lpc2, sw, st->subframeSize, st->lpcSize, st->mem_sw, stack);
}
@@ -968,9 +1005,9 @@
if (st->submodeID>=1)
{
for (i=0;i<st->lpcSize;i++)
- st->old_lsp[i] = st->lsp[i];
+ st->old_lsp[i] = lsp[i];
for (i=0;i<st->lpcSize;i++)
- st->old_qlsp[i] = st->qlsp[i];
+ st->old_qlsp[i] = qlsp[i];
}
#ifdef VORBIS_PSYCHO
@@ -991,19 +1028,7 @@
/* The next frame will not be the first (Duh!) */
st->first = 0;
-
-#ifdef RESYNTH
- /* Replace input by synthesized speech */
- for (i=0;i<st->frameSize;i++)
- {
- spx_word32_t sig = PSHR32(st->frame[i],SIG_SHIFT);
- if (sig>32767)
- sig = 32767;
- if (sig<-32767)
- sig = -32767;
- in[i]=sig;
- }
-#endif
+ speex_move(st->winBuf, in+2*st->frameSize-st->windowSize, (st->windowSize-st->frameSize)*sizeof(spx_word16_t));
if (SUBMODE(innovation_quant) == noise_codebook_quant || st->submodeID==0)
st->bounded_pitch = 1;
@@ -1013,7 +1038,6 @@
return 1;
}
-
void *nb_decoder_init(const SpeexMode *m)
{
DecState *st;
@@ -1050,27 +1074,16 @@
st->submodes=mode->submodes;
st->submodeID=mode->defaultSubmode;
- st->lpc_enh_enabled=0;
+ st->lpc_enh_enabled=1;
-
- st->inBuf = speex_alloc((st->frameSize)*sizeof(spx_sig_t));
- st->frame = st->inBuf;
- st->excBuf = speex_alloc((st->frameSize + st->max_pitch + 1)*sizeof(spx_sig_t));
- st->exc = st->excBuf + st->max_pitch + 1;
- for (i=0;i<st->frameSize;i++)
- st->inBuf[i]=0;
+ st->excBuf = speex_alloc((st->frameSize + 2*st->max_pitch + st->subframeSize + 12)*sizeof(spx_word16_t));
+ st->exc = st->excBuf + 2*st->max_pitch + st->subframeSize + 6;
for (i=0;i<st->frameSize + st->max_pitch + 1;i++)
st->excBuf[i]=0;
- st->innov = speex_alloc((st->frameSize)*sizeof(spx_sig_t));
st->interp_qlpc = speex_alloc(st->lpcSize*sizeof(spx_coef_t));
- st->qlsp = speex_alloc(st->lpcSize*sizeof(spx_lsp_t));
st->old_qlsp = speex_alloc(st->lpcSize*sizeof(spx_lsp_t));
- st->interp_qlsp = speex_alloc(st->lpcSize*sizeof(spx_lsp_t));
- st->mem_sp = speex_alloc((5*st->lpcSize)*sizeof(spx_mem_t));
- st->comb_mem = speex_alloc(sizeof(CombFilterMem));
- comb_filter_mem_init (st->comb_mem);
-
+ st->mem_sp = speex_alloc(st->lpcSize*sizeof(spx_mem_t));
st->pi_gain = speex_alloc((st->nbSubframes)*sizeof(spx_word32_t));
st->last_pitch = 40;
st->count_lost=0;
@@ -1104,15 +1117,10 @@
speex_free_scratch(st->stack);
#endif
- speex_free (st->inBuf);
speex_free (st->excBuf);
- speex_free (st->innov);
speex_free (st->interp_qlpc);
- speex_free (st->qlsp);
speex_free (st->old_qlsp);
- speex_free (st->interp_qlsp);
speex_free (st->mem_sp);
- speex_free (st->comb_mem);
speex_free (st->pi_gain);
speex_free(state);
@@ -1131,9 +1139,6 @@
{
int i, sub;
int pitch_val;
- VARDECL(spx_coef_t *awk1);
- VARDECL(spx_coef_t *awk2);
- VARDECL(spx_coef_t *awk3);
spx_word16_t pitch_gain;
spx_word16_t fact;
spx_word16_t gain_med;
@@ -1162,48 +1167,27 @@
pitch_gain = MULT16_16_Q15(fact,pitch_gain) + VERY_SMALL;
/* Shift all buffers by one frame */
- /*speex_move(st->inBuf, st->inBuf+st->frameSize, (st->bufSize-st->frameSize)*sizeof(spx_sig_t));*/
- speex_move(st->excBuf, st->excBuf+st->frameSize, (st->max_pitch + 1)*sizeof(spx_sig_t));
-
- ALLOC(awk1, (st->lpcSize+1), spx_coef_t);
- ALLOC(awk2, (st->lpcSize+1), spx_coef_t);
- ALLOC(awk3, (st->lpcSize+1), spx_coef_t);
-
+ speex_move(st->excBuf, st->excBuf+st->frameSize, (2*st->max_pitch + st->subframeSize + 12)*sizeof(spx_word16_t));
for (sub=0;sub<st->nbSubframes;sub++)
{
int offset;
- spx_sig_t *sp, *exc;
+ spx_word16_t *sp;
+ spx_word16_t *exc;
/* Offset relative to start of frame */
offset = st->subframeSize*sub;
/* Original signal */
- sp=st->frame+offset;
+ sp=out+offset;
/* Excitation */
exc=st->exc+offset;
/* Excitation after post-filter*/
-
- /* Calculate perceptually enhanced LPC filter */
- if (st->lpc_enh_enabled)
- {
- spx_word16_t k1,k2,k3;
- if (st->submodes[st->submodeID] != NULL)
- {
- k1=SUBMODE(lpc_enh_k1);
- k2=SUBMODE(lpc_enh_k2);
- k3=SUBMODE(lpc_enh_k3);
- } else {
- k1=k2=.7*GAMMA_SCALING;
- k3=.0;
- }
- bw_lpc(k1, st->interp_qlpc, awk1, st->lpcSize);
- bw_lpc(k2, st->interp_qlpc, awk2, st->lpcSize);
- bw_lpc(k3, st->interp_qlpc, awk3, st->lpcSize);
- }
/* Make up a plausible excitation */
/* FIXME: THIS CAN BE IMPROVED */
/*if (pitch_gain>.95)
pitch_gain=.95;*/
- innov_gain = compute_rms(st->innov, st->frameSize);
+
+ /* FIXME: This was rms of innovation (not exc) */
+ innov_gain = compute_rms16(st->exc, st->frameSize);
pitch_val = st->last_pitch + SHR32((spx_int32_t)speex_rand(1+st->count_lost, &st->seed),SIG_SHIFT);
if (pitch_val > st->max_pitch)
pitch_val = st->max_pitch;
@@ -1211,36 +1195,16 @@
pitch_val = st->min_pitch;
for (i=0;i<st->subframeSize;i++)
{
- exc[i]= MULT16_32_Q15(pitch_gain, (exc[i-pitch_val]+VERY_SMALL)) +
- MULT16_32_Q15(fact, MULT16_32_Q15(SHL(Q15ONE,15)-SHL(MULT16_16(pitch_gain,pitch_gain),1),speex_rand(innov_gain, &st->seed)));
+ /* FIXME: Second term need to be 16-bit */
+ exc[i]= MULT16_16_Q15(pitch_gain, (exc[i-pitch_val]+VERY_SMALL)) +
+ MULT16_16_Q15(fact, MULT16_16_Q15(SHL(Q15ONE,15)-SHL(MULT16_16(pitch_gain,pitch_gain),1),speex_rand(innov_gain, &st->seed)));
}
-
for (i=0;i<st->subframeSize;i++)
- sp[i]=exc[i];
-
- /* Signal synthesis */
- if (st->lpc_enh_enabled)
- {
- filter_mem2(sp, awk2, awk1, sp, st->subframeSize, st->lpcSize,
- st->mem_sp+st->lpcSize);
- filter_mem2(sp, awk3, st->interp_qlpc, sp, st->subframeSize, st->lpcSize,
- st->mem_sp);
- } else {
- for (i=0;i<st->lpcSize;i++)
- st->mem_sp[st->lpcSize+i] = 0;
- iir_mem2(sp, st->interp_qlpc, sp, st->subframeSize, st->lpcSize,
- st->mem_sp);
- }
- }
+ sp[i]=exc[i-st->subframeSize];
+ iir_mem16(sp, st->interp_qlpc, sp, st->subframeSize, st->lpcSize,
+ st->mem_sp, stack);
- for (i=0;i<st->frameSize;i++)
- {
- spx_word32_t sig = PSHR32(st->frame[i],SIG_SHIFT);
- if (sig>32767)
- sig = 32767;
- if (sig<-32767)
- sig = -32767;
- out[i]=sig;
+ bw_lpc(QCONST16(.98,15), st->interp_qlpc, st->interp_qlpc, st->lpcSize);
}
st->first = 0;
@@ -1250,6 +1214,7 @@
st->pitch_gain_buf_idx = 0;
}
+
int nb_decode(void *state, SpeexBits *bits, void *vout)
{
DecState *st;
@@ -1264,15 +1229,17 @@
int wideband;
int m;
char *stack;
- VARDECL(spx_coef_t *awk1);
- VARDECL(spx_coef_t *awk2);
- VARDECL(spx_coef_t *awk3);
+ VARDECL(spx_sig_t *innov);
+ VARDECL(spx_word32_t *exc32);
+ VARDECL(spx_coef_t *ak);
+ VARDECL(spx_lsp_t *qlsp);
spx_word16_t pitch_average=0;
#ifdef EPIC_48K
int pitch_half[2];
int ol_pitch_id=0;
#endif
spx_word16_t *out = vout;
+ VARDECL(spx_lsp_t *interp_qlsp);
st=(DecState*)state;
stack=st->stack;
@@ -1373,7 +1340,7 @@
}
/* Shift all buffers by one frame */
- speex_move(st->excBuf, st->excBuf+st->frameSize, (st->max_pitch + 1)*sizeof(spx_sig_t));
+ speex_move(st->excBuf, st->excBuf+st->frameSize, (2*st->max_pitch + st->subframeSize + 12)*sizeof(spx_word16_t));
/* If null mode (no transmission), just set a couple things to zero*/
if (st->submodes[st->submodeID] == NULL)
@@ -1386,34 +1353,28 @@
float pgain=GAIN_SCALING_1*st->last_pitch_gain;
if (pgain>.6)
pgain=.6;
- innov_gain = compute_rms(st->innov, st->frameSize);
+ /* FIXME: This was innov, not exc */
+ innov_gain = compute_rms16(st->exc, st->frameSize);
for (i=0;i<st->frameSize;i++)
- st->exc[i]=VERY_SMALL;
- speex_rand_vec(innov_gain, st->exc, st->frameSize);
+ st->exc[i]=speex_rand(innov_gain, &st->seed);
}
st->first=1;
- /* Final signal synthesis from excitation */
- iir_mem2(st->exc, lpc, st->frame, st->frameSize, st->lpcSize, st->mem_sp);
-
for (i=0;i<st->frameSize;i++)
- {
- spx_word32_t sig = PSHR32(st->frame[i],SIG_SHIFT);
- if (sig>32767)
- sig = 32767;
- if (sig<-32767)
- sig = -32767;
- out[i]=sig;
- }
+ out[i] = st->exc[i];
+ /* Final signal synthesis from excitation */
+ iir_mem16(out, lpc, out, st->frameSize, st->lpcSize, st->mem_sp, stack);
st->count_lost=0;
return 0;
}
+ ALLOC(qlsp, st->lpcSize, spx_lsp_t);
+
/* Unquantize LSPs */
- SUBMODE(lsp_unquant)(st->qlsp, st->lpcSize, bits);
+ SUBMODE(lsp_unquant)(qlsp, st->lpcSize, bits);
/*Damp memory if a frame was lost and the LSP changed too much*/
if (st->count_lost)
@@ -1421,13 +1382,13 @@
spx_word16_t fact;
spx_word32_t lsp_dist=0;
for (i=0;i<st->lpcSize;i++)
- lsp_dist = ADD32(lsp_dist, EXTEND32(ABS(st->old_qlsp[i] - st->qlsp[i])));
+ lsp_dist = ADD32(lsp_dist, EXTEND32(ABS(st->old_qlsp[i] - qlsp[i])));
#ifdef FIXED_POINT
fact = SHR16(19661,SHR32(lsp_dist,LSP_SHIFT+2));
#else
fact = .6*exp(-.2*lsp_dist);
#endif
- for (i=0;i<2*st->lpcSize;i++)
+ for (i=0;i<st->lpcSize;i++)
st->mem_sp[i] = MULT16_32_Q15(fact,st->mem_sp[i]);
}
@@ -1436,7 +1397,7 @@
if (st->first || st->count_lost)
{
for (i=0;i<st->lpcSize;i++)
- st->old_qlsp[i] = st->qlsp[i];
+ st->old_qlsp[i] = qlsp[i];
}
#ifdef EPIC_48K
@@ -1483,9 +1444,9 @@
}
#endif
- ALLOC(awk1, st->lpcSize+1, spx_coef_t);
- ALLOC(awk2, st->lpcSize+1, spx_coef_t);
- ALLOC(awk3, st->lpcSize+1, spx_coef_t);
+ ALLOC(ak, st->lpcSize, spx_coef_t);
+ ALLOC(innov, st->subframeSize, spx_sig_t);
+ ALLOC(exc32, st->subframeSize, spx_word32_t);
if (st->submodeID==1)
{
@@ -1504,7 +1465,9 @@
for (sub=0;sub<st->nbSubframes;sub++)
{
int offset;
- spx_sig_t *sp, *exc;
+ spx_word16_t *exc;
+ spx_word16_t *sp;
+ spx_sig_t *innov_save = NULL;
spx_word16_t tmp;
#ifdef EPIC_48K
@@ -1519,40 +1482,13 @@
/* Offset relative to start of frame */
offset = st->subframeSize*sub;
- /* Original signal */
- sp=st->frame+offset;
/* Excitation */
exc=st->exc+offset;
- /* Excitation after post-filter*/
+ /* Original signal */
+ sp=out+offset;
+ if (st->innov_save)
+ innov_save = st->innov_save+offset;
- /* LSP interpolation (quantized and unquantized) */
- lsp_interpolate(st->old_qlsp, st->qlsp, st->interp_qlsp, st->lpcSize, sub, st->nbSubframes);
-
- /* Make sure the LSP's are stable */
- lsp_enforce_margin(st->interp_qlsp, st->lpcSize, LSP_MARGIN);
-
-
- /* Compute interpolated LPCs (unquantized) */
- lsp_to_lpc(st->interp_qlsp, st->interp_qlpc, st->lpcSize, stack);
-
- /* Compute enhanced synthesis filter */
- if (st->lpc_enh_enabled)
- {
- bw_lpc(SUBMODE(lpc_enh_k1), st->interp_qlpc, awk1, st->lpcSize);
- bw_lpc(SUBMODE(lpc_enh_k2), st->interp_qlpc, awk2, st->lpcSize);
- bw_lpc(SUBMODE(lpc_enh_k3), st->interp_qlpc, awk3, st->lpcSize);
- }
-
- /* Compute analysis filter at w=pi */
- {
- spx_word32_t pi_g=LPC_SCALING;
- for (i=0;i<st->lpcSize;i+=2)
- {
- /*pi_g += -st->interp_qlpc[i] + st->interp_qlpc[i+1];*/
- pi_g = ADD32(pi_g, SUB32(st->interp_qlpc[i+1],st->interp_qlpc[i]));
- }
- st->pi_gain[sub] = pi_g;
- }
/* Reset excitation */
for (i=0;i<st->subframeSize;i++)
@@ -1595,13 +1531,13 @@
#ifdef EPIC_48K
if (st->lbr_48k)
{
- SUBMODE(ltp_unquant)(exc, pit_min, pit_max, ol_pitch_coef, SUBMODE(ltp_params),
+ SUBMODE(ltp_unquant)(exc, exc32, pit_min, pit_max, ol_pitch_coef, SUBMODE(ltp_params),
st->subframeSize, &pitch, &pitch_gain[0], bits, stack,
st->count_lost, offset, st->last_pitch_gain, ol_pitch_id);
} else {
#endif
- SUBMODE(ltp_unquant)(exc, pit_min, pit_max, ol_pitch_coef, SUBMODE(ltp_params),
+ SUBMODE(ltp_unquant)(exc, exc32, pit_min, pit_max, ol_pitch_coef, SUBMODE(ltp_params),
st->subframeSize, &pitch, &pitch_gain[0], bits, stack,
st->count_lost, offset, st->last_pitch_gain, 0);
@@ -1609,25 +1545,16 @@
}
#endif
-
- /* If we had lost frames, check energy of last received frame */
- if (st->count_lost && ol_gain < st->last_ol_gain)
- {
- /*float fact = (float)ol_gain/(st->last_ol_gain+1);
- for (i=0;i<st->subframeSize;i++)
- exc[i]*=fact;*/
- spx_word16_t fact = DIV32_16(SHL32(EXTEND32(ol_gain),15),st->last_ol_gain+1);
- for (i=0;i<st->subframeSize;i++)
- exc[i] = MULT16_32_Q15(fact, exc[i]);
- }
-
tmp = gain_3tap_to_1tap(pitch_gain);
pitch_average += tmp;
- if (tmp>best_pitch_gain)
+ if ((tmp>best_pitch_gain&&ABS(2*best_pitch-pitch)>=3&&ABS(3*best_pitch-pitch)>=4&&ABS(4*best_pitch-pitch)>=5)
+ || (tmp>MULT16_16_Q15(QCONST16(.6,15),best_pitch_gain)&&(ABS(best_pitch-2*pitch)<3||ABS(best_pitch-3*pitch)<4||ABS(best_pitch-4*pitch)<5))
+ || (MULT16_16_Q15(QCONST16(.67,15),tmp)>best_pitch_gain&&(ABS(2*best_pitch-pitch)<3||ABS(3*best_pitch-pitch)<4||ABS(4*best_pitch-pitch)<5)) )
{
best_pitch = pitch;
- best_pitch_gain = tmp;
+ if (tmp > best_pitch_gain)
+ best_pitch_gain = tmp;
}
} else {
speex_error("No pitch prediction, what's wrong");
@@ -1637,9 +1564,7 @@
{
int q_energy;
spx_word32_t ener;
- spx_sig_t *innov;
- innov = st->innov+sub*st->subframeSize;
for (i=0;i<st->subframeSize;i++)
innov[i]=0;
@@ -1681,7 +1606,7 @@
while (st->voc_offset<st->subframeSize)
{
if (st->voc_offset>=0)
- exc[st->voc_offset]=SIG_SCALING*sqrt(1.0*ol_pitch);
+ exc[st->voc_offset]=sqrt(1.0*ol_pitch);
st->voc_offset+=ol_pitch;
}
st->voc_offset -= st->subframeSize;
@@ -1693,8 +1618,9 @@
g=1;
for (i=0;i<st->subframeSize;i++)
{
- float exci=exc[i];
- exc[i]=.8*g*exc[i]*ol_gain/SIG_SCALING + .6*g*st->voc_m1*ol_gain/SIG_SCALING + .5*g*innov[i] - .5*g*st->voc_m2 + (1-g)*innov[i];
+ spx_word16_t exci=exc[i];
+ /* FIXME: cleanup the innov[i]/SIG_SCALING */
+ exc[i]=.8*g*exc[i]*PSHR32(ol_gain,SIG_SHIFT) + .6*g*st->voc_m1*PSHR32(ol_gain,SIG_SHIFT) + (1-.5*g)*PSHR32(innov[i],SIG_SHIFT) - .5*g*PSHR32(st->voc_m2,SIG_SHIFT);
st->voc_m1 = exci;
st->voc_m2=innov[i];
st->voc_mean = .95*st->voc_mean + .05*exc[i];
@@ -1702,9 +1628,14 @@
}
} else {
for (i=0;i<st->subframeSize;i++)
- exc[i]=ADD32(exc[i],innov[i]);
+ exc[i]=PSHR32(ADD32(SHL32(exc32[i],1),innov[i]),SIG_SHIFT);
/*print_vec(exc, 40, "innov");*/
}
+ if (innov_save)
+ {
+ for (i=0;i<st->subframeSize;i++)
+ innov_save[i] = innov[i];
+ }
/* Decode second codebook (only for some modes) */
if (SUBMODE(double_codebook))
{
@@ -1714,68 +1645,93 @@
for (i=0;i<st->subframeSize;i++)
innov2[i]=0;
SUBMODE(innovation_unquant)(innov2, SUBMODE(innovation_params), st->subframeSize, bits, stack);
- signal_mul(innov2, innov2, (spx_word32_t) (ener*(1/2.2)), st->subframeSize);
+ signal_mul(innov2, innov2, MULT16_32_Q15(QCONST16(0.454545,15),ener), st->subframeSize);
for (i=0;i<st->subframeSize;i++)
- exc[i] = ADD32(exc[i],innov2[i]);
+ exc[i] = ADD16(exc[i],PSHR32(innov2[i],SIG_SHIFT));
+ if (innov_save)
+ {
+ for (i=0;i<st->subframeSize;i++)
+ innov_save[i] = ADD32(innov_save[i],innov2[i]);
+ }
stack = tmp_stack;
}
-
- }
-
- /* If the last packet was lost, re-scale the excitation to obtain the same energy as encoded in ol_gain */
- if (st->count_lost)
- {
- spx_word16_t exc_ener;
- spx_word32_t gain32;
- spx_word16_t gain;
- exc_ener = compute_rms (exc, st->subframeSize);
- gain32 = DIV32(ol_gain, ADD16(exc_ener,1));
-#ifdef FIXED_POINT
- if (gain32 > 32768)
- gain32 = 32768;
- gain = EXTRACT16(gain32);
-#else
- if (gain32 > 2)
- gain32=2;
- gain = gain32;
-#endif
- for (i=0;i<st->subframeSize;i++)
- exc[i] = MULT16_32_Q14(gain, exc[i]);
- }
-
- for (i=0;i<st->subframeSize;i++)
- sp[i]=exc[i];
-
- /* Signal synthesis */
- if (st->lpc_enh_enabled && SUBMODE(comb_gain)>0)
- comb_filter(exc, sp, st->interp_qlpc, st->lpcSize, st->subframeSize,
- pitch, pitch_gain, SUBMODE(comb_gain), st->comb_mem);
-
- if (st->lpc_enh_enabled)
- {
- /* Use enhanced LPC filter */
- filter_mem2(sp, awk2, awk1, sp, st->subframeSize, st->lpcSize,
- st->mem_sp+st->lpcSize);
- filter_mem2(sp, awk3, st->interp_qlpc, sp, st->subframeSize, st->lpcSize,
- st->mem_sp);
- } else {
- /* Use regular filter */
- for (i=0;i<st->lpcSize;i++)
- st->mem_sp[st->lpcSize+i] = 0;
- iir_mem2(sp, st->interp_qlpc, sp, st->subframeSize, st->lpcSize,
- st->mem_sp);
}
}
- /*Copy output signal*/
- for (i=0;i<st->frameSize;i++)
+ ALLOC(interp_qlsp, st->lpcSize, spx_lsp_t);
+
+ if (st->lpc_enh_enabled && SUBMODE(comb_gain)>0 && !st->count_lost)
{
- spx_word32_t sig = PSHR32(st->frame[i],SIG_SHIFT);
- if (sig>32767)
- sig = 32767;
- if (sig<-32767)
- sig = -32767;
- out[i]=sig;
+ multicomb(st->exc-st->subframeSize, out, st->interp_qlpc, st->lpcSize, 2*st->subframeSize, best_pitch, 40, SUBMODE(comb_gain), stack);
+ multicomb(st->exc+st->subframeSize, out+2*st->subframeSize, st->interp_qlpc, st->lpcSize, 2*st->subframeSize, best_pitch, 40, SUBMODE(comb_gain), stack);
+ } else {
+ for (i=0;i<st->frameSize;i++)
+ out[i]=st->exc[i-st->subframeSize];
+ }
+
+ /* If the last packet was lost, re-scale the excitation to obtain the same energy as encoded in ol_gain */
+ if (st->count_lost)
+ {
+ spx_word16_t exc_ener;
+ spx_word32_t gain32;
+ spx_word16_t gain;
+ exc_ener = compute_rms16 (st->exc, st->frameSize);
+ gain32 = PDIV32(ol_gain, ADD16(exc_ener,1));
+#ifdef FIXED_POINT
+ if (gain32 > 32768)
+ gain32 = 32768;
+ gain = EXTRACT16(gain32);
+#else
+ if (gain32 > 2)
+ gain32=2;
+ gain = gain32;
+#endif
+ for (i=0;i<st->frameSize;i++)
+ {
+ st->exc[i] = MULT16_16_Q14(gain, st->exc[i]);
+ out[i]=st->exc[i-st->subframeSize];
+ }
+ }
+
+ /*Loop on subframes */
+ for (sub=0;sub<st->nbSubframes;sub++)
+ {
+ int offset;
+ spx_word16_t *sp;
+ spx_word16_t *exc;
+ /* Offset relative to start of frame */
+ offset = st->subframeSize*sub;
+ /* Original signal */
+ sp=out+offset;
+ /* Excitation */
+ exc=st->exc+offset;
+
+ /* LSP interpolation (quantized and unquantized) */
+ lsp_interpolate(st->old_qlsp, qlsp, interp_qlsp, st->lpcSize, sub, st->nbSubframes);
+
+ /* Make sure the LSP's are stable */
+ lsp_enforce_margin(interp_qlsp, st->lpcSize, LSP_MARGIN);
+
+ /* Compute interpolated LPCs (unquantized) */
+ lsp_to_lpc(interp_qlsp, ak, st->lpcSize, stack);
+
+ /* Compute analysis filter at w=pi */
+ {
+ spx_word32_t pi_g=LPC_SCALING;
+ for (i=0;i<st->lpcSize;i+=2)
+ {
+ /*pi_g += -st->interp_qlpc[i] + st->interp_qlpc[i+1];*/
+ pi_g = ADD32(pi_g, SUB32(EXTEND32(st->interp_qlpc[i+1]),EXTEND32(st->interp_qlpc[i])));
+ }
+ st->pi_gain[sub] = pi_g;
+ }
+
+ iir_mem16(sp, st->interp_qlpc, sp, st->subframeSize, st->lpcSize,
+ st->mem_sp, stack);
+
+ for (i=0;i<st->lpcSize;i++)
+ st->interp_qlpc[i] = ak[i];
+
}
/*for (i=0;i<st->frameSize;i++)
@@ -1783,7 +1739,7 @@
/* Store the LSPs for interpolation in the next frame */
for (i=0;i<st->lpcSize;i++)
- st->old_qlsp[i] = st->qlsp[i];
+ st->old_qlsp[i] = qlsp[i];
/* The next frame will not be the first (Duh!) */
st->first = 0;
@@ -1839,12 +1795,14 @@
(*(int*)ptr) = st->dtx_enabled;
break;
case SPEEX_SET_ABR:
- st->abr_enabled = (*(int*)ptr);
- st->vbr_enabled = 1;
+ st->abr_enabled = (*(spx_int32_t*)ptr);
+ st->vbr_enabled = st->abr_enabled!=0;
+ if (st->vbr_enabled)
{
- int i=10, rate, target;
+ int i=10;
+ spx_int32_t rate, target;
float vbr_qual;
- target = (*(int*)ptr);
+ target = (*(spx_int32_t*)ptr);
while (i>=0)
{
speex_encoder_ctl(st, SPEEX_SET_QUALITY, &i);
@@ -1864,7 +1822,7 @@
break;
case SPEEX_GET_ABR:
- (*(int*)ptr) = st->abr_enabled;
+ (*(spx_int32_t*)ptr) = st->abr_enabled;
break;
case SPEEX_SET_VBR_QUALITY:
st->vbr_quality = (*(float*)ptr);
@@ -1888,12 +1846,13 @@
st->complexity=0;
break;
case SPEEX_GET_COMPLEXITY:
- (*(int*)ptr) = st->complexity;
+ (*(spx_int32_t*)ptr) = st->complexity;
break;
case SPEEX_SET_BITRATE:
{
- int i=10, rate, target;
- target = (*(int*)ptr);
+ int i=10;
+ spx_int32_t rate, target;
+ target = (*(spx_int32_t*)ptr);
while (i>=0)
{
speex_encoder_ctl(st, SPEEX_SET_QUALITY, &i);
@@ -1906,15 +1865,15 @@
break;
case SPEEX_GET_BITRATE:
if (st->submodes[st->submodeID])
- (*(int*)ptr) = st->sampling_rate*SUBMODE(bits_per_frame)/st->frameSize;
+ (*(spx_int32_t*)ptr) = st->sampling_rate*SUBMODE(bits_per_frame)/st->frameSize;
else
- (*(int*)ptr) = st->sampling_rate*(NB_SUBMODE_BITS+1)/st->frameSize;
+ (*(spx_int32_t*)ptr) = st->sampling_rate*(NB_SUBMODE_BITS+1)/st->frameSize;
break;
case SPEEX_SET_SAMPLING_RATE:
- st->sampling_rate = (*(int*)ptr);
+ st->sampling_rate = (*(spx_int32_t*)ptr);
break;
case SPEEX_GET_SAMPLING_RATE:
- (*(int*)ptr)=st->sampling_rate;
+ (*(spx_int32_t*)ptr)=st->sampling_rate;
break;
case SPEEX_RESET_STATE:
{
@@ -1922,13 +1881,13 @@
st->bounded_pitch = 1;
st->first = 1;
for (i=0;i<st->lpcSize;i++)
- st->lsp[i]=(M_PI*((float)(i+1)))/(st->lpcSize+1);
+ st->old_lsp[i]=(M_PI*((float)(i+1)))/(st->lpcSize+1);
for (i=0;i<st->lpcSize;i++)
st->mem_sw[i]=st->mem_sw_whole[i]=st->mem_sp[i]=st->mem_exc[i]=0;
for (i=0;i<st->frameSize+st->max_pitch+1;i++)
st->excBuf[i]=st->swBuf[i]=0;
- for (i=0;i<st->windowSize;i++)
- st->inBuf[i]=0;
+ for (i=0;i<st->windowSize-st->frameSize;i++)
+ st->winBuf[i]=0;
}
break;
case SPEEX_SET_SUBMODE_ENCODING:
@@ -1948,6 +1907,15 @@
case SPEEX_GET_PLC_TUNING:
(*(int*)ptr)=(st->plc_tuning);
break;
+ case SPEEX_SET_VBR_MAX_BITRATE:
+ st->vbr_max = (*(spx_int32_t*)ptr);
+ break;
+ case SPEEX_GET_VBR_MAX_BITRATE:
+ (*(spx_int32_t*)ptr) = st->vbr_max;
+ break;
+
+
+ /* This is all internal stuff past this point */
case SPEEX_GET_PI_GAIN:
{
int i;
@@ -1959,22 +1927,17 @@
case SPEEX_GET_EXC:
{
int i;
- spx_sig_t *e = (spx_sig_t*)ptr;
+ spx_word16_t *e = (spx_word16_t*)ptr;
for (i=0;i<st->frameSize;i++)
e[i]=st->exc[i];
}
break;
- case SPEEX_GET_INNOV:
- {
- int i;
- spx_sig_t *e = (spx_sig_t*)ptr;
- for (i=0;i<st->frameSize;i++)
- e[i]=st->innov[i];
- }
- break;
case SPEEX_GET_RELATIVE_QUALITY:
(*(float*)ptr)=st->relative_quality;
break;
+ case SPEEX_SET_INNOVATION_SAVE:
+ st->innov_save = ptr;
+ break;
default:
speex_warning_int("Unknown nb_ctl request: ", request);
return -1;
@@ -2007,15 +1970,15 @@
break;
case SPEEX_GET_BITRATE:
if (st->submodes[st->submodeID])
- (*(int*)ptr) = st->sampling_rate*SUBMODE(bits_per_frame)/st->frameSize;
+ (*(spx_int32_t*)ptr) = st->sampling_rate*SUBMODE(bits_per_frame)/st->frameSize;
else
- (*(int*)ptr) = st->sampling_rate*(NB_SUBMODE_BITS+1)/st->frameSize;
+ (*(spx_int32_t*)ptr) = st->sampling_rate*(NB_SUBMODE_BITS+1)/st->frameSize;
break;
case SPEEX_SET_SAMPLING_RATE:
- st->sampling_rate = (*(int*)ptr);
+ st->sampling_rate = (*(spx_int32_t*)ptr);
break;
case SPEEX_GET_SAMPLING_RATE:
- (*(int*)ptr)=st->sampling_rate;
+ (*(spx_int32_t*)ptr)=st->sampling_rate;
break;
case SPEEX_SET_HANDLER:
{
@@ -2036,12 +1999,10 @@
case SPEEX_RESET_STATE:
{
int i;
- for (i=0;i<2*st->lpcSize;i++)
+ for (i=0;i<st->lpcSize;i++)
st->mem_sp[i]=0;
for (i=0;i<st->frameSize + st->max_pitch + 1;i++)
st->excBuf[i]=0;
- for (i=0;i<st->frameSize;i++)
- st->inBuf[i] = 0;
}
break;
case SPEEX_SET_SUBMODE_ENCODING:
@@ -2050,6 +2011,9 @@
case SPEEX_GET_SUBMODE_ENCODING:
(*(int*)ptr) = st->encode_submode;
break;
+ case SPEEX_GET_LOOKAHEAD:
+ (*(int*)ptr)=st->subframeSize;
+ break;
case SPEEX_GET_PI_GAIN:
{
int i;
@@ -2061,22 +2025,17 @@
case SPEEX_GET_EXC:
{
int i;
- spx_sig_t *e = (spx_sig_t*)ptr;
+ spx_word16_t *e = (spx_word16_t*)ptr;
for (i=0;i<st->frameSize;i++)
e[i]=st->exc[i];
}
break;
- case SPEEX_GET_INNOV:
- {
- int i;
- spx_sig_t *e = (spx_sig_t*)ptr;
- for (i=0;i<st->frameSize;i++)
- e[i]=st->innov[i];
- }
- break;
case SPEEX_GET_DTX_STATUS:
*((int*)ptr) = st->dtx_enabled;
break;
+ case SPEEX_SET_INNOVATION_SAVE:
+ st->innov_save = ptr;
+ break;
default:
speex_warning_int("Unknown nb_ctl request: ", request);
return -1;
diff --git a/pjmedia/src/pjmedia-codec/speex/nb_celp.h b/pjmedia/src/pjmedia-codec/speex/nb_celp.h
index c9fb2b3..92028cb 100644
--- a/pjmedia/src/pjmedia-codec/speex/nb_celp.h
+++ b/pjmedia/src/pjmedia-codec/speex/nb_celp.h
@@ -48,20 +48,20 @@
/**Structure representing the full state of the narrowband encoder*/
typedef struct EncState {
- const SpeexMode *mode; /**< Mode corresponding to the state */
- int first; /**< Is this the first frame? */
- int frameSize; /**< Size of frames */
- int subframeSize; /**< Size of sub-frames */
- int nbSubframes; /**< Number of sub-frames */
- int windowSize; /**< Analysis (LPC) window length */
- int lpcSize; /**< LPC order */
- int min_pitch; /**< Minimum pitch value allowed */
- int max_pitch; /**< Maximum pitch value allowed */
+ const SpeexMode *mode; /**< Mode corresponding to the state */
+ int first; /**< Is this the first frame? */
+ int frameSize; /**< Size of frames */
+ int subframeSize; /**< Size of sub-frames */
+ int nbSubframes; /**< Number of sub-frames */
+ int windowSize; /**< Analysis (LPC) window length */
+ int lpcSize; /**< LPC order */
+ int min_pitch; /**< Minimum pitch value allowed */
+ int max_pitch; /**< Maximum pitch value allowed */
- int safe_pitch; /**< Don't use too large values for pitch (in case we lose a packet) */
- int bounded_pitch; /**< Next frame should not rely on previous frames for pitch */
- int ol_pitch; /**< Open-loop pitch */
- int ol_voiced; /**< Open-loop voiced/non-voiced decision */
+ spx_word32_t cumul_gain; /**< Product of previously used pitch gains (Q10) */
+ int bounded_pitch; /**< Next frame should not rely on previous frames for pitch */
+ int ol_pitch; /**< Open-loop pitch */
+ int ol_voiced; /**< Open-loop voiced/non-voiced decision */
int *pitch;
#ifdef EPIC_48K
@@ -70,111 +70,100 @@
#ifdef VORBIS_PSYCHO
VorbisPsy *psy;
+ float *psy_window;
float *curve;
float *old_curve;
#endif
spx_word16_t gamma1; /**< Perceptual filter: A(z/gamma1) */
spx_word16_t gamma2; /**< Perceptual filter: A(z/gamma2) */
- float lag_factor; /**< Lag windowing Gaussian width */
+ float lag_factor; /**< Lag windowing Gaussian width */
spx_word16_t lpc_floor; /**< Noise floor multiplier for A[0] in LPC analysis*/
- char *stack; /**< Pseudo-stack allocation for temporary memory */
- spx_sig_t *inBuf; /**< Input buffer (original signal) */
- spx_sig_t *frame; /**< Start of original frame */
- spx_sig_t *excBuf; /**< Excitation buffer */
- spx_sig_t *exc; /**< Start of excitation frame */
- spx_sig_t *swBuf; /**< Weighted signal buffer */
- spx_sig_t *sw; /**< Start of weighted signal frame */
- spx_sig_t *innov; /**< Innovation for the frame */
- spx_word16_t *window; /**< Temporary (Hanning) window */
- spx_word16_t *autocorr; /**< auto-correlation */
+ char *stack; /**< Pseudo-stack allocation for temporary memory */
+ spx_word16_t *winBuf; /**< Input buffer (original signal) */
+ spx_word16_t *excBuf; /**< Excitation buffer */
+ spx_word16_t *exc; /**< Start of excitation frame */
+ spx_word16_t *swBuf; /**< Weighted signal buffer */
+ spx_word16_t *sw; /**< Start of weighted signal frame */
+ const spx_word16_t *window; /**< Temporary (Hanning) window */
spx_word16_t *lagWindow; /**< Window applied to auto-correlation */
- spx_coef_t *lpc; /**< LPCs for current frame */
- spx_lsp_t *lsp; /**< LSPs for current frame */
- spx_lsp_t *qlsp; /**< Quantized LSPs for current frame */
- spx_lsp_t *old_lsp; /**< LSPs for previous frame */
- spx_lsp_t *old_qlsp; /**< Quantized LSPs for previous frame */
- spx_lsp_t *interp_lsp; /**< Interpolated LSPs */
- spx_lsp_t *interp_qlsp; /**< Interpolated quantized LSPs */
- spx_coef_t *interp_lpc; /**< Interpolated LPCs */
- spx_coef_t *interp_qlpc; /**< Interpolated quantized LPCs */
- spx_coef_t *bw_lpc1; /**< LPCs after bandwidth expansion by gamma1 for perceptual weighting*/
- spx_coef_t *bw_lpc2; /**< LPCs after bandwidth expansion by gamma2 for perceptual weighting*/
- spx_mem_t *mem_sp; /**< Filter memory for signal synthesis */
- spx_mem_t *mem_sw; /**< Filter memory for perceptually-weighted signal */
- spx_mem_t *mem_sw_whole; /**< Filter memory for perceptually-weighted signal (whole frame)*/
- spx_mem_t *mem_exc; /**< Filter memory for excitation (whole frame) */
+ spx_lsp_t *old_lsp; /**< LSPs for previous frame */
+ spx_lsp_t *old_qlsp; /**< Quantized LSPs for previous frame */
+ spx_mem_t *mem_sp; /**< Filter memory for signal synthesis */
+ spx_mem_t *mem_sw; /**< Filter memory for perceptually-weighted signal */
+ spx_mem_t *mem_sw_whole; /**< Filter memory for perceptually-weighted signal (whole frame)*/
+ spx_mem_t *mem_exc; /**< Filter memory for excitation (whole frame) */
+ spx_mem_t *mem_exc2; /**< Filter memory for excitation (whole frame) */
spx_word32_t *pi_gain; /**< Gain of LPC filter at theta=pi (fe/2) */
-
- VBRState *vbr; /**< State of the VBR data */
- float vbr_quality; /**< Quality setting for VBR encoding */
- float relative_quality; /**< Relative quality that will be needed by VBR */
- int vbr_enabled; /**< 1 for enabling VBR, 0 otherwise */
- int vad_enabled; /**< 1 for enabling VAD, 0 otherwise */
- int dtx_enabled; /**< 1 for enabling DTX, 0 otherwise */
- int dtx_count; /**< Number of consecutive DTX frames */
- int abr_enabled; /**< ABR setting (in bps), 0 if off */
+ spx_sig_t *innov_save; /**< If non-NULL, innovation is copied here */
+
+ VBRState *vbr; /**< State of the VBR data */
+ float vbr_quality; /**< Quality setting for VBR encoding */
+ float relative_quality; /**< Relative quality that will be needed by VBR */
+ int vbr_enabled; /**< 1 for enabling VBR, 0 otherwise */
+ spx_int32_t vbr_max; /**< Max bit-rate allowed in VBR mode */
+ int vad_enabled; /**< 1 for enabling VAD, 0 otherwise */
+ int dtx_enabled; /**< 1 for enabling DTX, 0 otherwise */
+ int dtx_count; /**< Number of consecutive DTX frames */
+ spx_int32_t abr_enabled; /**< ABR setting (in bps), 0 if off */
float abr_drift;
float abr_drift2;
float abr_count;
- int complexity; /**< Complexity setting (0-10 from least complex to most complex) */
- int sampling_rate;
+ int complexity; /**< Complexity setting (0-10 from least complex to most complex) */
+ spx_int32_t sampling_rate;
int plc_tuning;
int encode_submode;
const SpeexSubmode * const *submodes; /**< Sub-mode data */
- int submodeID; /**< Activated sub-mode */
- int submodeSelect; /**< Mode chosen by the user (may differ from submodeID if VAD is on) */
+ int submodeID; /**< Activated sub-mode */
+ int submodeSelect; /**< Mode chosen by the user (may differ from submodeID if VAD is on) */
} EncState;
/**Structure representing the full state of the narrowband decoder*/
typedef struct DecState {
const SpeexMode *mode; /**< Mode corresponding to the state */
- int first; /**< Is this the first frame? */
- int count_lost; /**< Was the last frame lost? */
- int frameSize; /**< Size of frames */
- int subframeSize; /**< Size of sub-frames */
- int nbSubframes; /**< Number of sub-frames */
- int lpcSize; /**< LPC order */
- int min_pitch; /**< Minimum pitch value allowed */
- int max_pitch; /**< Maximum pitch value allowed */
- int sampling_rate;
+ int first; /**< Is this the first frame? */
+ int count_lost; /**< Was the last frame lost? */
+ int frameSize; /**< Size of frames */
+ int subframeSize; /**< Size of sub-frames */
+ int nbSubframes; /**< Number of sub-frames */
+ int lpcSize; /**< LPC order */
+ int min_pitch; /**< Minimum pitch value allowed */
+ int max_pitch; /**< Maximum pitch value allowed */
+ spx_int32_t sampling_rate;
#ifdef EPIC_48K
int lbr_48k;
#endif
- spx_word16_t last_ol_gain; /**< Open-loop gain for previous frame */
+ spx_word16_t last_ol_gain; /**< Open-loop gain for previous frame */
- char *stack; /**< Pseudo-stack allocation for temporary memory */
- spx_sig_t *inBuf; /**< Input buffer (original signal) */
- spx_sig_t *frame; /**< Start of original frame */
- spx_sig_t *excBuf; /**< Excitation buffer */
- spx_sig_t *exc; /**< Start of excitation frame */
- spx_sig_t *innov; /**< Innovation for the frame */
- spx_lsp_t *qlsp; /**< Quantized LSPs for current frame */
- spx_lsp_t *old_qlsp; /**< Quantized LSPs for previous frame */
- spx_lsp_t *interp_qlsp; /**< Interpolated quantized LSPs */
- spx_coef_t *interp_qlpc; /**< Interpolated quantized LPCs */
- spx_mem_t *mem_sp; /**< Filter memory for synthesis signal */
- spx_word32_t *pi_gain; /**< Gain of LPC filter at theta=pi (fe/2) */
- int last_pitch; /**< Pitch of last correctly decoded frame */
+ char *stack; /**< Pseudo-stack allocation for temporary memory */
+ spx_word16_t *excBuf; /**< Excitation buffer */
+ spx_word16_t *exc; /**< Start of excitation frame */
+ spx_lsp_t *old_qlsp; /**< Quantized LSPs for previous frame */
+ spx_coef_t *interp_qlpc; /**< Interpolated quantized LPCs */
+ spx_mem_t *mem_sp; /**< Filter memory for synthesis signal */
+ spx_word32_t *pi_gain; /**< Gain of LPC filter at theta=pi (fe/2) */
+ spx_sig_t *innov_save; /** If non-NULL, innovation is copied here */
+
+ /* This is used in packet loss concealment */
+ int last_pitch; /**< Pitch of last correctly decoded frame */
spx_word16_t last_pitch_gain; /**< Pitch gain of last correctly decoded frame */
- spx_word16_t pitch_gain_buf[3]; /**< Pitch gain of last decoded frames */
- int pitch_gain_buf_idx; /**< Tail of the buffer */
- spx_int32_t seed; /** Seed used for random number generation */
+ spx_word16_t pitch_gain_buf[3]; /**< Pitch gain of last decoded frames */
+ int pitch_gain_buf_idx; /**< Tail of the buffer */
+ spx_int32_t seed; /** Seed used for random number generation */
int encode_submode;
const SpeexSubmode * const *submodes; /**< Sub-mode data */
- int submodeID; /**< Activated sub-mode */
- int lpc_enh_enabled; /**< 1 when LPC enhancer is on, 0 otherwise */
- CombFilterMem *comb_mem;
+ int submodeID; /**< Activated sub-mode */
+ int lpc_enh_enabled; /**< 1 when LPC enhancer is on, 0 otherwise */
SpeexCallback speex_callbacks[SPEEX_MAX_CALLBACKS];
SpeexCallback user_callback;
/*Vocoder data*/
- float voc_m1;
- float voc_m2;
+ spx_word16_t voc_m1;
+ spx_word32_t voc_m2;
float voc_mean;
int voc_offset;
diff --git a/pjmedia/src/pjmedia-codec/speex/pseudofloat.h b/pjmedia/src/pjmedia-codec/speex/pseudofloat.h
index 8642bf0..e85f60e 100644
--- a/pjmedia/src/pjmedia-codec/speex/pseudofloat.h
+++ b/pjmedia/src/pjmedia-codec/speex/pseudofloat.h
@@ -45,9 +45,9 @@
spx_int16_t e;
} spx_float_t;
-#define FLOAT_ZERO ((spx_float_t){0,0})
-#define FLOAT_ONE ((spx_float_t){16384,-14})
-#define FLOAT_HALF ((spx_float_t){16384,-15})
+static const spx_float_t FLOAT_ZERO = {0,0};
+static const spx_float_t FLOAT_ONE = {16384,-14};
+static const spx_float_t FLOAT_HALF = {16384,-15};
#define MIN(a,b) ((a)<(b)?(a):(b))
static inline spx_float_t PSEUDOFLOAT(spx_int32_t x)
@@ -60,7 +60,10 @@
x = -x;
}
if (x==0)
- return (spx_float_t) {0,0};
+ {
+ spx_float_t r = {0,0};
+ return r;
+ }
while (x>32767)
{
x >>= 1;
@@ -74,9 +77,19 @@
e--;
}
if (sign)
- return (spx_float_t) {-x,e};
+ {
+ spx_float_t r;
+ r.m = -x;
+ r.e = e;
+ return r;
+ }
else
- return (spx_float_t) {x,e};
+ {
+ spx_float_t r;
+ r.m = x;
+ r.e = e;
+ return r;
+ }
}
@@ -87,7 +100,16 @@
return b;
else if (b.m==0)
return a;
- r = (a).e > (b).e ? (spx_float_t) {((a).m>>1) + ((b).m>>MIN(15,(a).e-(b).e+1)),(a).e+1} : (spx_float_t) {((b).m>>1) + ((a).m>>MIN(15,(b).e-(a).e+1)),(b).e+1};
+ if ((a).e > (b).e)
+ {
+ r.m = ((a).m>>1) + ((b).m>>MIN(15,(a).e-(b).e+1));
+ r.e = (a).e+1;
+ }
+ else
+ {
+ r.m = ((b).m>>1) + ((a).m>>MIN(15,(b).e-(a).e+1));
+ r.e = (b).e+1;
+ }
if (r.m>0)
{
if (r.m<16384)
@@ -113,7 +135,16 @@
return b;
else if (b.m==0)
return a;
- r = (a).e > (b).e ? (spx_float_t) {((a).m>>1) - ((b).m>>MIN(15,(a).e-(b).e+1)),(a).e+1} : (spx_float_t) {((a).m>>MIN(15,(b).e-(a).e+1)) - ((b).m>>1) ,(b).e+1};
+ if ((a).e > (b).e)
+ {
+ r.m = ((a).m>>1) - ((b).m>>MIN(15,(a).e-(b).e+1));
+ r.e = (a).e+1;
+ }
+ else
+ {
+ r.m = ((a).m>>MIN(15,(b).e-(a).e+1)) - ((b).m>>1);
+ r.e = (b).e+1;
+ }
if (r.m>0)
{
if (r.m<16384)
@@ -152,7 +183,9 @@
static inline spx_float_t FLOAT_MULT(spx_float_t a, spx_float_t b)
{
- spx_float_t r = (spx_float_t) {(spx_int16_t)((spx_int32_t)(a).m*(b).m>>15), (a).e+(b).e+15};
+ spx_float_t r;
+ r.m = (spx_int16_t)((spx_int32_t)(a).m*(b).m>>15);
+ r.e = (a).e+(b).e+15;
if (r.m>0)
{
if (r.m<16384)
@@ -174,13 +207,16 @@
static inline spx_float_t FLOAT_SHL(spx_float_t a, int b)
{
- return (spx_float_t) {a.m,a.e+b};
+ spx_float_t r;
+ r.m = a.m;
+ r.e = a.e+b;
+ return r;
}
static inline spx_int16_t FLOAT_EXTRACT16(spx_float_t a)
{
if (a.e<0)
- return (a.m+(1<<(-a.e-1)))>>-a.e;
+ return EXTRACT16((EXTEND32(a.m)+(1<<(-a.e-1)))>>-a.e);
else
return a.m<<a.e;
}
@@ -196,9 +232,12 @@
static inline spx_float_t FLOAT_MUL32U(spx_word32_t a, spx_word32_t b)
{
int e=0;
+ spx_float_t r;
/* FIXME: Handle the sign */
if (a==0)
- return (spx_float_t) {0,0};
+ {
+ return FLOAT_ZERO;
+ }
while (a>32767)
{
a >>= 1;
@@ -219,35 +258,45 @@
b <<= 1;
e--;
}
- return (spx_float_t) {MULT16_16_Q15(a,b),e+15};
+ r.m = MULT16_16_Q15(a,b);
+ r.e = e+15;
+ return r;
}
static inline spx_float_t FLOAT_DIV32_FLOAT(spx_word32_t a, spx_float_t b)
{
int e=0;
+ spx_float_t r;
/* FIXME: Handle the sign */
if (a==0)
- return (spx_float_t) {0,0};
- while (a<SHL32(b.m,14))
+ {
+ return FLOAT_ZERO;
+ }
+ while (a<SHL32(EXTEND32(b.m),14))
{
a <<= 1;
e--;
}
- while (a>=SHL32(b.m-1,15))
+ while (a>=SHL32(EXTEND32(b.m-1),15))
{
a >>= 1;
e++;
}
- return (spx_float_t) {DIV32_16(a,b.m),e-b.e};
+ r.m = DIV32_16(a,b.m);
+ r.e = e-b.e;
+ return r;
}
static inline spx_float_t FLOAT_DIV32(spx_word32_t a, spx_word32_t b)
{
int e=0;
+ spx_float_t r;
/* FIXME: Handle the sign */
if (a==0)
- return (spx_float_t) {0,0};
+ {
+ return FLOAT_ZERO;
+ }
while (b>32767)
{
b >>= 1;
@@ -263,13 +312,16 @@
a >>= 1;
e++;
}
- return (spx_float_t) {DIV32_16(a,b),e};
+ r.m = DIV32_16(a,b);
+ r.e = e;
+ return r;
}
static inline spx_float_t FLOAT_DIVU(spx_float_t a, spx_float_t b)
{
int e=0;
spx_int32_t num;
+ spx_float_t r;
num = a.m;
while (a.m >= b.m)
{
@@ -277,7 +329,9 @@
a.m >>= 1;
}
num = num << (15-e);
- return (spx_float_t) {DIV32_16(num,b.m),a.e-b.e-15+e};
+ r.m = DIV32_16(num,b.m);
+ r.e = a.e-b.e-15+e;
+ return r;
}
#else
diff --git a/pjmedia/src/pjmedia-codec/speex/quant_lsp.c b/pjmedia/src/pjmedia-codec/speex/quant_lsp.c
index 7bd0b91..bfca587 100644
--- a/pjmedia/src/pjmedia-codec/speex/quant_lsp.c
+++ b/pjmedia/src/pjmedia-codec/speex/quant_lsp.c
@@ -40,9 +40,12 @@
#define M_PI 3.14159265358979323846
#endif
-
#include "misc.h"
+#ifdef BFIN_ASM
+#include "quant_lsp_bfin.h"
+#endif
+
#ifdef FIXED_POINT
#define LSP_LINEAR(i) (SHL16(i+1,11))
@@ -90,12 +93,13 @@
}
/* Note: x is modified*/
+#ifndef OVERRIDE_LSP_QUANT
static int lsp_quant(spx_word16_t *x, const signed char *cdbk, int nbVec, int nbDim)
{
int i,j;
spx_word32_t dist;
spx_word16_t tmp;
- spx_word32_t best_dist=0;
+ spx_word32_t best_dist=VERY_LARGE32;
int best_id=0;
const signed char *ptr=cdbk;
for (i=0;i<nbVec;i++)
@@ -105,8 +109,8 @@
{
tmp=SUB16(x[j],SHL16((spx_word16_t)*ptr++,5));
dist=MAC16_16(dist,tmp,tmp);
- }
- if (dist<best_dist || i==0)
+ }
+ if (dist<best_dist)
{
best_dist=dist;
best_id=i;
@@ -118,14 +122,16 @@
return best_id;
}
+#endif
/* Note: x is modified*/
+#ifndef OVERRIDE_LSP_WEIGHT_QUANT
static int lsp_weight_quant(spx_word16_t *x, spx_word16_t *weight, const signed char *cdbk, int nbVec, int nbDim)
{
int i,j;
spx_word32_t dist;
spx_word16_t tmp;
- spx_word32_t best_dist=0;
+ spx_word32_t best_dist=VERY_LARGE32;
int best_id=0;
const signed char *ptr=cdbk;
for (i=0;i<nbVec;i++)
@@ -136,7 +142,7 @@
tmp=SUB16(x[j],SHL16((spx_word16_t)*ptr++,5));
dist=MAC16_32_Q15(dist,weight[j],MULT16_16(tmp,tmp));
}
- if (dist<best_dist || i==0)
+ if (dist<best_dist)
{
best_dist=dist;
best_id=i;
@@ -147,7 +153,7 @@
x[j] = SUB16(x[j],SHL16((spx_word16_t)cdbk[best_id*nbDim+j],5));
return best_id;
}
-
+#endif
void lsp_quant_nb(spx_lsp_t *lsp, spx_lsp_t *qlsp, int order, SpeexBits *bits)
{
diff --git a/pjmedia/src/pjmedia-codec/speex/sb_celp.c b/pjmedia/src/pjmedia-codec/speex/sb_celp.c
index 5adf44d..465a538 100644
--- a/pjmedia/src/pjmedia-codec/speex/sb_celp.c
+++ b/pjmedia/src/pjmedia-codec/speex/sb_celp.c
@@ -201,6 +201,8 @@
};
#endif
+extern const spx_word16_t lpc_window[];
+
static void mix_and_saturate(spx_word32_t *x0, spx_word32_t *x1, spx_word16_t *out, int len)
{
int i;
@@ -244,7 +246,7 @@
st->frame_size = mode->frameSize;
st->subframeSize = mode->subframeSize;
st->nbSubframes = mode->frameSize/mode->subframeSize;
- st->windowSize = st->frame_size*3/2;
+ st->windowSize = st->frame_size+st->subframeSize;
st->lpcSize=mode->lpcSize;
st->bufSize=mode->bufSize;
@@ -277,18 +279,7 @@
st->res=speex_alloc((st->frame_size)*sizeof(spx_sig_t));
st->sw=speex_alloc((st->frame_size)*sizeof(spx_sig_t));
- st->target=speex_alloc((st->frame_size)*sizeof(spx_sig_t));
- /*Asymmetric "pseudo-Hamming" window*/
- {
- int part1, part2;
- part1 = st->subframeSize*7/2;
- part2 = st->subframeSize*5/2;
- st->window = speex_alloc((st->windowSize)*sizeof(spx_word16_t));
- for (i=0;i<part1;i++)
- st->window[i]=(spx_word16_t)(SIG_SCALING*(.54-.46*cos(M_PI*i/part1)));
- for (i=0;i<part2;i++)
- st->window[part1+i]=(spx_word16_t)(SIG_SCALING*(.54+.46*cos(M_PI*i/part2)));
- }
+ st->window= lpc_window;
st->lagWindow = speex_alloc((st->lpcSize+1)*sizeof(spx_word16_t));
for (i=0;i<st->lpcSize+1;i++)
@@ -307,13 +298,18 @@
st->interp_lpc = speex_alloc(st->lpcSize*sizeof(spx_coef_t));
st->interp_qlpc = speex_alloc(st->lpcSize*sizeof(spx_coef_t));
st->pi_gain = speex_alloc((st->nbSubframes)*sizeof(spx_word32_t));
-
+ st->low_innov = speex_alloc((st->frame_size)*sizeof(spx_word32_t));
+ speex_encoder_ctl(st->st_low, SPEEX_SET_INNOVATION_SAVE, st->low_innov);
+ st->innov_save = NULL;
+
st->mem_sp = speex_alloc((st->lpcSize)*sizeof(spx_mem_t));
st->mem_sp2 = speex_alloc((st->lpcSize)*sizeof(spx_mem_t));
st->mem_sw = speex_alloc((st->lpcSize)*sizeof(spx_mem_t));
st->vbr_quality = 8;
st->vbr_enabled = 0;
+ st->vbr_max = 0;
+ st->vbr_max_high = 20000; /* We just need a big value here */
st->vad_enabled = 0;
st->abr_enabled = 0;
st->relative_quality=0;
@@ -350,8 +346,6 @@
speex_free(st->excBuf);
speex_free(st->res);
speex_free(st->sw);
- speex_free(st->target);
- speex_free(st->window);
speex_free(st->lagWindow);
speex_free(st->autocorr);
@@ -384,10 +378,10 @@
char *stack;
VARDECL(spx_mem_t *mem);
VARDECL(spx_sig_t *innov);
+ VARDECL(spx_word16_t *target);
VARDECL(spx_word16_t *syn_resp);
VARDECL(spx_word32_t *low_pi_gain);
- VARDECL(spx_sig_t *low_exc);
- VARDECL(spx_sig_t *low_innov);
+ VARDECL(spx_word16_t *low_exc);
const SpeexSBMode *mode;
int dtx;
spx_word16_t *in = vin;
@@ -422,11 +416,9 @@
ALLOC(low_pi_gain, st->nbSubframes, spx_word32_t);
- ALLOC(low_exc, st->frame_size, spx_sig_t);
- ALLOC(low_innov, st->frame_size, spx_sig_t);
+ ALLOC(low_exc, st->frame_size, spx_word16_t);
speex_encoder_ctl(st->st_low, SPEEX_GET_PI_GAIN, low_pi_gain);
speex_encoder_ctl(st->st_low, SPEEX_GET_EXC, low_exc);
- speex_encoder_ctl(st->st_low, SPEEX_GET_INNOV, low_innov);
speex_encoder_ctl(st->st_low, SPEEX_GET_LOW_MODE, &dtx);
@@ -455,15 +447,15 @@
_spx_lpc(st->lpc, st->autocorr, st->lpcSize);
/* LPC to LSPs (x-domain) transform */
- roots=lpc_to_lsp (st->lpc, st->lpcSize, st->lsp, 15, LSP_DELTA1, stack);
+ roots=lpc_to_lsp (st->lpc, st->lpcSize, st->lsp, 10, LSP_DELTA1, stack);
if (roots!=st->lpcSize)
{
- roots = lpc_to_lsp (st->lpc, st->lpcSize, st->lsp, 11, LSP_DELTA2, stack);
+ roots = lpc_to_lsp (st->lpc, st->lpcSize, st->lsp, 10, LSP_DELTA2, stack);
if (roots!=st->lpcSize) {
/*If we can't find all LSP's, do some damage control and use a flat filter*/
for (i=0;i<st->lpcSize;i++)
{
- st->lsp[i]=M_PI*((float)(i+1))/(st->lpcSize+1);
+ st->lsp[i]=LSP_SCALING*M_PI*((float)(i+1))/(st->lpcSize+1);
}
}
}
@@ -521,7 +513,7 @@
else
thresh = (st->vbr_quality-v1) * mode->vbr_thresh[modeid][v1+1] +
(1+v1-st->vbr_quality) * mode->vbr_thresh[modeid][v1];
- if (st->relative_quality >= thresh)
+ if (st->relative_quality >= thresh && st->sampling_rate*st->submodes[modeid]->bits_per_frame/st->full_frame_size <= st->vbr_max_high)
break;
modeid--;
}
@@ -601,10 +593,11 @@
ALLOC(mem, st->lpcSize, spx_mem_t);
ALLOC(syn_resp, st->subframeSize, spx_word16_t);
ALLOC(innov, st->subframeSize, spx_sig_t);
+ ALLOC(target, st->subframeSize, spx_word16_t);
for (sub=0;sub<st->nbSubframes;sub++)
{
- spx_sig_t *exc, *sp, *res, *target, *sw;
+ spx_sig_t *exc, *sp, *res, *sw, *innov_save=NULL;
spx_word16_t filter_ratio;
int offset;
spx_word32_t rl, rh;
@@ -614,8 +607,14 @@
sp=st->high+offset;
exc=st->exc+offset;
res=st->res+offset;
- target=st->target+offset;
sw=st->sw+offset;
+ /* Pointer for saving innovation */
+ if (st->innov_save)
+ {
+ innov_save = st->innov_save+2*offset;
+ for (i=0;i<2*st->subframeSize;i++)
+ innov_save[i]=0;
+ }
/* LSP interpolation (quantized and unquantized) */
lsp_interpolate(st->old_lsp, st->lsp, st->interp_lsp, st->lpcSize, sub, st->nbSubframes);
@@ -642,7 +641,7 @@
rl = low_pi_gain[sub];
#ifdef FIXED_POINT
- filter_ratio=DIV32_16(SHL(rl+82,2),SHR(82+rh,5));
+ filter_ratio=PDIV32_16(SHL(rl+82,2),SHR(82+rh,5));
#else
filter_ratio=(rl+.01)/(rh+.01);
#endif
@@ -656,10 +655,10 @@
if (!SUBMODE(innovation_quant)) {/* 1 for spectral folding excitation, 0 for stochastic */
float g;
spx_word16_t el;
- el = compute_rms(low_innov+offset, st->subframeSize);
+ el = compute_rms(st->low_innov+offset, st->subframeSize);
/* Gain to use if we want to use the low-band excitation for high-band */
- g=eh/(.01+el);
+ g=eh/(1.+el);
#if 0
{
@@ -669,7 +668,7 @@
ALLOC(tmp_sig, st->subframeSize, spx_sig_t);
for (i=0;i<st->lpcSize;i++)
mem[i]=st->mem_sp[i];
- iir_mem2(low_innov+offset, st->interp_qlpc, tmp_sig, st->subframeSize, st->lpcSize, mem);
+ iir_mem2(st->low_innov+offset, st->interp_qlpc, tmp_sig, st->subframeSize, st->lpcSize, mem);
g2 = compute_rms(sp, st->subframeSize)/(.01+compute_rms(tmp_sig, st->subframeSize));
/*fprintf (stderr, "gains: %f %f\n", g, g2);*/
g = g2;
@@ -698,9 +697,9 @@
spx_word16_t gc;
spx_word32_t scale;
spx_word16_t el;
- el = compute_rms(low_exc+offset, st->subframeSize);
+ el = compute_rms16(low_exc+offset, st->subframeSize);
- gc = DIV32_16(MULT16_16(filter_ratio,1+eh),1+el);
+ gc = PDIV32_16(MULT16_16(filter_ratio,1+eh),1+el);
/* This is a kludge that cleans up a historical bug */
if (st->subframeSize==80)
@@ -726,7 +725,7 @@
if (st->subframeSize==80)
gc *= 1.4142;
- scale = SHL(MULT16_16(DIV32_16(SHL(gc,SIG_SHIFT-4),filter_ratio),(1+el)),4);
+ scale = SHL32(MULT16_16(PDIV32_16(SHL32(EXTEND32(gc),SIG_SHIFT-6),filter_ratio),(1+el)),6);
compute_impulse_response(st->interp_qlpc, st->bw_lpc1, st->bw_lpc2, syn_resp, st->subframeSize, st->lpcSize, stack);
@@ -751,7 +750,7 @@
/* Compute target signal */
for (i=0;i<st->subframeSize;i++)
- target[i]=sw[i]-res[i];
+ target[i]=PSHR32(sw[i]-res[i],SIG_SHIFT);
for (i=0;i<st->subframeSize;i++)
exc[i]=0;
@@ -773,6 +772,12 @@
for (i=0;i<st->subframeSize;i++)
exc[i] = ADD32(exc[i], innov[i]);
+ if (st->innov_save)
+ {
+ for (i=0;i<st->subframeSize;i++)
+ innov_save[2*i]=innov[i];
+ }
+
if (SUBMODE(double_codebook)) {
char *tmp_stack=stack;
VARDECL(spx_sig_t *innov2);
@@ -871,6 +876,7 @@
st->g1_mem=speex_alloc((QMF_ORDER)*sizeof(spx_word32_t));
st->exc=speex_alloc((st->frame_size)*sizeof(spx_sig_t));
+ st->excBuf=speex_alloc((st->subframeSize)*sizeof(spx_sig_t));
st->qlsp = speex_alloc((st->lpcSize)*sizeof(spx_lsp_t));
st->old_qlsp = speex_alloc((st->lpcSize)*sizeof(spx_lsp_t));
@@ -880,6 +886,11 @@
st->pi_gain = speex_alloc((st->nbSubframes)*sizeof(spx_word32_t));
st->mem_sp = speex_alloc((2*st->lpcSize)*sizeof(spx_mem_t));
+ st->low_innov = speex_alloc((st->frame_size)*sizeof(spx_word32_t));
+ speex_decoder_ctl(st->st_low, SPEEX_SET_INNOVATION_SAVE, st->low_innov);
+ st->innov_save = NULL;
+
+
st->lpc_enh_enabled=0;
st->seed = 1000;
@@ -906,6 +917,7 @@
speex_free(st->g0_mem);
speex_free(st->g1_mem);
speex_free(st->exc);
+ speex_free(st->excBuf);
speex_free(st->qlsp);
speex_free(st->old_qlsp);
speex_free(st->interp_qlsp);
@@ -919,9 +931,6 @@
static void sb_decode_lost(SBDecState *st, spx_word16_t *out, int dtx, char *stack)
{
int i;
- VARDECL(spx_coef_t *awk1);
- VARDECL(spx_coef_t *awk2);
- VARDECL(spx_coef_t *awk3);
int saved_modeid=0;
if (dtx)
@@ -934,28 +943,6 @@
st->first=1;
- ALLOC(awk1, st->lpcSize+1, spx_coef_t);
- ALLOC(awk2, st->lpcSize+1, spx_coef_t);
- ALLOC(awk3, st->lpcSize+1, spx_coef_t);
-
- if (st->lpc_enh_enabled)
- {
- spx_word16_t k1,k2,k3;
- if (st->submodes[st->submodeID] != NULL)
- {
- k1=SUBMODE(lpc_enh_k1);
- k2=SUBMODE(lpc_enh_k2);
- k3=SUBMODE(lpc_enh_k3);
- } else {
- k1=k2=.7*GAMMA_SCALING;
- k3 = 0;
- }
- bw_lpc(k1, st->interp_qlpc, awk1, st->lpcSize);
- bw_lpc(k2, st->interp_qlpc, awk2, st->lpcSize);
- bw_lpc(k3, st->interp_qlpc, awk3, st->lpcSize);
- /*fprintf (stderr, "%f %f %f\n", k1, k2, k3);*/
- }
-
/* Final signal synthesis from excitation */
if (!dtx)
@@ -969,22 +956,9 @@
for (i=0;i<st->frame_size;i++)
st->high[i]=st->exc[i];
- if (st->lpc_enh_enabled)
- {
- /* Use enhanced LPC filter */
- filter_mem2(st->high, awk2, awk1, st->high, st->frame_size, st->lpcSize,
- st->mem_sp+st->lpcSize);
- filter_mem2(st->high, awk3, st->interp_qlpc, st->high, st->frame_size, st->lpcSize,
- st->mem_sp);
- } else {
- /* Use regular filter */
- for (i=0;i<st->lpcSize;i++)
- st->mem_sp[st->lpcSize+i] = 0;
- iir_mem2(st->high, st->interp_qlpc, st->high, st->frame_size, st->lpcSize,
- st->mem_sp);
- }
+ iir_mem2(st->high, st->interp_qlpc, st->high, st->frame_size, st->lpcSize,
+ st->mem_sp);
- /*iir_mem2(st->exc, st->interp_qlpc, st->high, st->frame_size, st->lpcSize, st->mem_sp);*/
/* Reconstruct the original */
fir_mem_up(st->x0d, h0, st->y0, st->full_frame_size, QMF_ORDER, st->g0_mem, stack);
@@ -1008,11 +982,8 @@
int ret;
char *stack;
VARDECL(spx_word32_t *low_pi_gain);
- VARDECL(spx_sig_t *low_exc);
- VARDECL(spx_sig_t *low_innov);
- VARDECL(spx_coef_t *awk1);
- VARDECL(spx_coef_t *awk2);
- VARDECL(spx_coef_t *awk3);
+ VARDECL(spx_word16_t *low_exc);
+ VARDECL(spx_coef_t *ak);
int dtx;
const SpeexSBMode *mode;
spx_word16_t *out = vout;
@@ -1101,11 +1072,9 @@
st->exc[i]=0;
ALLOC(low_pi_gain, st->nbSubframes, spx_word32_t);
- ALLOC(low_exc, st->frame_size, spx_sig_t);
- ALLOC(low_innov, st->frame_size, spx_sig_t);
+ ALLOC(low_exc, st->frame_size, spx_word16_t);
speex_decoder_ctl(st->st_low, SPEEX_GET_PI_GAIN, low_pi_gain);
speex_decoder_ctl(st->st_low, SPEEX_GET_EXC, low_exc);
- speex_decoder_ctl(st->st_low, SPEEX_GET_INNOV, low_innov);
SUBMODE(lsp_unquant)(st->qlsp, st->lpcSize, bits);
@@ -1115,13 +1084,11 @@
st->old_qlsp[i] = st->qlsp[i];
}
- ALLOC(awk1, st->lpcSize+1, spx_coef_t);
- ALLOC(awk2, st->lpcSize+1, spx_coef_t);
- ALLOC(awk3, st->lpcSize+1, spx_coef_t);
+ ALLOC(ak, st->lpcSize, spx_coef_t);
for (sub=0;sub<st->nbSubframes;sub++)
{
- spx_sig_t *exc, *sp;
+ spx_sig_t *exc, *sp, *innov_save=NULL;
spx_word16_t filter_ratio;
spx_word16_t el=0;
int offset;
@@ -1130,6 +1097,13 @@
offset = st->subframeSize*sub;
sp=st->high+offset;
exc=st->exc+offset;
+ /* Pointer for saving innovation */
+ if (st->innov_save)
+ {
+ innov_save = st->innov_save+2*offset;
+ for (i=0;i<2*st->subframeSize;i++)
+ innov_save[i]=0;
+ }
/* LSP interpolation */
lsp_interpolate(st->old_qlsp, st->qlsp, st->interp_qlsp, st->lpcSize, sub, st->nbSubframes);
@@ -1137,21 +1111,7 @@
lsp_enforce_margin(st->interp_qlsp, st->lpcSize, LSP_MARGIN);
/* LSP to LPC */
- lsp_to_lpc(st->interp_qlsp, st->interp_qlpc, st->lpcSize, stack);
-
-
- if (st->lpc_enh_enabled)
- {
- spx_word16_t k1,k2,k3;
- k1=SUBMODE(lpc_enh_k1);
- k2=SUBMODE(lpc_enh_k2);
- k3=SUBMODE(lpc_enh_k3);
- bw_lpc(k1, st->interp_qlpc, awk1, st->lpcSize);
- bw_lpc(k2, st->interp_qlpc, awk2, st->lpcSize);
- bw_lpc(k3, st->interp_qlpc, awk3, st->lpcSize);
- /*fprintf (stderr, "%f %f %f\n", k1, k2, k3);*/
- }
-
+ lsp_to_lpc(st->interp_qlsp, ak, st->lpcSize, stack);
/* Calculate reponse ratio between the low and high filter in the middle
of the band (4000 Hz) */
@@ -1166,7 +1126,7 @@
rl = low_pi_gain[sub];
#ifdef FIXED_POINT
- filter_ratio=DIV32_16(SHL(rl+82,2),SHR(82+rh,5));
+ filter_ratio=PDIV32_16(SHL(rl+82,2),SHR(82+rh,5));
#else
filter_ratio=(rl+.01)/(rh+.01);
#endif
@@ -1190,7 +1150,7 @@
#if 0
for (i=0;i<st->subframeSize;i++)
- exc[i]=mode->folding_gain*g*low_innov[offset+i];
+ exc[i]=mode->folding_gain*g*st->low_innov[offset+i];
#else
{
float tmp=1;
@@ -1199,7 +1159,7 @@
el = compute_rms(low_innov+offset, st->subframeSize);*/
for (i=0;i<st->subframeSize;i++)
{
- float e=tmp*g*mode->folding_gain*low_innov[offset+i];
+ float e=tmp*g*mode->folding_gain*st->low_innov[offset+i];
tmp *= -1;
exc[i] = e;
/*float r = speex_rand(g*el,&seed);
@@ -1210,14 +1170,13 @@
}
- /*speex_rand_vec(mode->folding_gain*g*el, exc, st->subframeSize);*/
#endif
} else {
spx_word16_t gc;
spx_word32_t scale;
int qgc = speex_bits_unpack_unsigned(bits, 4);
- el = compute_rms(low_exc+offset, st->subframeSize);
+ el = compute_rms16(low_exc+offset, st->subframeSize);
#ifdef FIXED_POINT
gc = MULT16_32_Q15(28626,gc_quant_bound[qgc]);
@@ -1228,7 +1187,7 @@
if (st->subframeSize==80)
gc *= 1.4142;
- scale = SHL(MULT16_16(DIV32_16(SHL(gc,SIG_SHIFT-4),filter_ratio),(1+el)),4);
+ scale = SHL(MULT16_16(PDIV32_16(SHL(gc,SIG_SHIFT-6),filter_ratio),(1+el)),6);
SUBMODE(innovation_unquant)(exc, SUBMODE(innovation_params), st->subframeSize,
bits, stack);
@@ -1251,24 +1210,21 @@
}
}
-
- for (i=0;i<st->subframeSize;i++)
- sp[i]=exc[i];
- if (st->lpc_enh_enabled)
+
+ if (st->innov_save)
{
- /* Use enhanced LPC filter */
- filter_mem2(sp, awk2, awk1, sp, st->subframeSize, st->lpcSize,
- st->mem_sp+st->lpcSize);
- filter_mem2(sp, awk3, st->interp_qlpc, sp, st->subframeSize, st->lpcSize,
- st->mem_sp);
- } else {
- /* Use regular filter */
- for (i=0;i<st->lpcSize;i++)
- st->mem_sp[st->lpcSize+i] = 0;
- iir_mem2(sp, st->interp_qlpc, sp, st->subframeSize, st->lpcSize,
- st->mem_sp);
+ for (i=0;i<st->subframeSize;i++)
+ innov_save[2*i]=exc[i];
}
- /*iir_mem2(exc, st->interp_qlpc, sp, st->subframeSize, st->lpcSize, st->mem_sp);*/
+
+ for (i=0;i<st->subframeSize;i++)
+ sp[i]=st->excBuf[i];
+ iir_mem2(sp, st->interp_qlpc, sp, st->subframeSize, st->lpcSize,
+ st->mem_sp);
+ for (i=0;i<st->subframeSize;i++)
+ st->excBuf[i]=exc[i];
+ for (i=0;i<st->lpcSize;i++)
+ st->interp_qlpc[i] = ak[i];
}
@@ -1345,13 +1301,14 @@
(*(float*)ptr) = st->vbr_quality;
break;
case SPEEX_SET_ABR:
- st->abr_enabled = (*(int*)ptr);
- st->vbr_enabled = 1;
+ st->abr_enabled = (*(spx_int32_t*)ptr);
+ st->vbr_enabled = st->abr_enabled!=0;
speex_encoder_ctl(st->st_low, SPEEX_SET_VBR, &st->vbr_enabled);
+ if (st->vbr_enabled)
{
int i=10, rate, target;
float vbr_qual;
- target = (*(int*)ptr);
+ target = (*(spx_int32_t*)ptr);
while (i>=0)
{
speex_encoder_ctl(st, SPEEX_SET_QUALITY, &i);
@@ -1371,7 +1328,7 @@
break;
case SPEEX_GET_ABR:
- (*(int*)ptr) = st->abr_enabled;
+ (*(spx_int32_t*)ptr) = st->abr_enabled;
break;
case SPEEX_SET_QUALITY:
{
@@ -1397,8 +1354,9 @@
break;
case SPEEX_SET_BITRATE:
{
- int i=10, rate, target;
- target = (*(int*)ptr);
+ int i=10;
+ spx_int32_t rate, target;
+ target = (*(spx_int32_t*)ptr);
while (i>=0)
{
speex_encoder_ctl(st, SPEEX_SET_QUALITY, &i);
@@ -1413,21 +1371,21 @@
speex_encoder_ctl(st->st_low, request, ptr);
/*fprintf (stderr, "before: %d\n", (*(int*)ptr));*/
if (st->submodes[st->submodeID])
- (*(int*)ptr) += st->sampling_rate*SUBMODE(bits_per_frame)/st->full_frame_size;
+ (*(spx_int32_t*)ptr) += st->sampling_rate*SUBMODE(bits_per_frame)/st->full_frame_size;
else
- (*(int*)ptr) += st->sampling_rate*(SB_SUBMODE_BITS+1)/st->full_frame_size;
+ (*(spx_int32_t*)ptr) += st->sampling_rate*(SB_SUBMODE_BITS+1)/st->full_frame_size;
/*fprintf (stderr, "after: %d\n", (*(int*)ptr));*/
break;
case SPEEX_SET_SAMPLING_RATE:
{
- int tmp=(*(int*)ptr);
+ spx_int32_t tmp=(*(spx_int32_t*)ptr);
st->sampling_rate = tmp;
tmp>>=1;
speex_encoder_ctl(st->st_low, SPEEX_SET_SAMPLING_RATE, &tmp);
}
break;
case SPEEX_GET_SAMPLING_RATE:
- (*(int*)ptr)=st->sampling_rate;
+ (*(spx_int32_t*)ptr)=st->sampling_rate;
break;
case SPEEX_RESET_STATE:
{
@@ -1454,6 +1412,45 @@
speex_encoder_ctl(st->st_low, SPEEX_GET_LOOKAHEAD, ptr);
(*(int*)ptr) = 2*(*(int*)ptr) + QMF_ORDER - 1;
break;
+ case SPEEX_SET_PLC_TUNING:
+ speex_encoder_ctl(st->st_low, SPEEX_SET_PLC_TUNING, ptr);
+ break;
+ case SPEEX_GET_PLC_TUNING:
+ speex_encoder_ctl(st->st_low, SPEEX_GET_PLC_TUNING, ptr);
+ break;
+ case SPEEX_SET_VBR_MAX_BITRATE:
+ {
+ st->vbr_max = (*(spx_int32_t*)ptr);
+ if (SPEEX_SET_VBR_MAX_BITRATE<1)
+ {
+ speex_encoder_ctl(st->st_low, SPEEX_SET_VBR_MAX_BITRATE, &st->vbr_max);
+ st->vbr_max_high = 17600;
+ } else {
+ spx_int32_t low_rate;
+ /* FIXME: Need to adapt that to ultra-wideband */
+ if (st->vbr_max >= 42200)
+ {
+ st->vbr_max_high = 17600;
+ } else if (st->vbr_max >= 27800)
+ {
+ st->vbr_max_high = 9600;
+ } else if (st->vbr_max > 20600)
+ {
+ st->vbr_max_high = 5600;
+ } else {
+ st->vbr_max_high = 1800;
+ }
+ low_rate = st->vbr_max - st->vbr_max_high;
+ speex_encoder_ctl(st->st_low, SPEEX_SET_VBR_MAX_BITRATE, &low_rate);
+ }
+ }
+ break;
+ case SPEEX_GET_VBR_MAX_BITRATE:
+ (*(spx_int32_t*)ptr) = st->vbr_max;
+ break;
+
+
+ /* This is all internal stuff past this point */
case SPEEX_GET_PI_GAIN:
{
int i;
@@ -1485,6 +1482,9 @@
case SPEEX_GET_RELATIVE_QUALITY:
(*(float*)ptr)=st->relative_quality;
break;
+ case SPEEX_SET_INNOVATION_SAVE:
+ st->innov_save = ptr;
+ break;
default:
speex_warning_int("Unknown nb_ctl request: ", request);
return -1;
@@ -1534,20 +1534,20 @@
case SPEEX_GET_BITRATE:
speex_decoder_ctl(st->st_low, request, ptr);
if (st->submodes[st->submodeID])
- (*(int*)ptr) += st->sampling_rate*SUBMODE(bits_per_frame)/st->full_frame_size;
+ (*(spx_int32_t*)ptr) += st->sampling_rate*SUBMODE(bits_per_frame)/st->full_frame_size;
else
- (*(int*)ptr) += st->sampling_rate*(SB_SUBMODE_BITS+1)/st->full_frame_size;
+ (*(spx_int32_t*)ptr) += st->sampling_rate*(SB_SUBMODE_BITS+1)/st->full_frame_size;
break;
case SPEEX_SET_SAMPLING_RATE:
{
- int tmp=(*(int*)ptr);
+ spx_int32_t tmp=(*(spx_int32_t*)ptr);
st->sampling_rate = tmp;
tmp>>=1;
speex_decoder_ctl(st->st_low, SPEEX_SET_SAMPLING_RATE, &tmp);
}
break;
case SPEEX_GET_SAMPLING_RATE:
- (*(int*)ptr)=st->sampling_rate;
+ (*(spx_int32_t*)ptr)=st->sampling_rate;
break;
case SPEEX_SET_HANDLER:
speex_decoder_ctl(st->st_low, SPEEX_SET_HANDLER, ptr);
@@ -1571,6 +1571,10 @@
case SPEEX_GET_SUBMODE_ENCODING:
(*(int*)ptr) = st->encode_submode;
break;
+ case SPEEX_GET_LOOKAHEAD:
+ speex_decoder_ctl(st->st_low, SPEEX_GET_LOOKAHEAD, ptr);
+ (*(int*)ptr) = 2*(*(int*)ptr);
+ break;
case SPEEX_GET_PI_GAIN:
{
int i;
@@ -1602,6 +1606,9 @@
case SPEEX_GET_DTX_STATUS:
speex_decoder_ctl(st->st_low, SPEEX_GET_DTX_STATUS, ptr);
break;
+ case SPEEX_SET_INNOVATION_SAVE:
+ st->innov_save = ptr;
+ break;
default:
speex_warning_int("Unknown nb_ctl request: ", request);
return -1;
diff --git a/pjmedia/src/pjmedia-codec/speex/sb_celp.h b/pjmedia/src/pjmedia-codec/speex/sb_celp.h
index c283109..194cdd9 100644
--- a/pjmedia/src/pjmedia-codec/speex/sb_celp.h
+++ b/pjmedia/src/pjmedia-codec/speex/sb_celp.h
@@ -42,60 +42,63 @@
/**Structure representing the full state of the sub-band encoder*/
typedef struct SBEncState {
- const SpeexMode *mode; /**< Pointer to the mode (containing for vtable info) */
- void *st_low; /**< State of the low-band (narrowband) encoder */
- int full_frame_size; /**< Length of full-band frames*/
- int frame_size; /**< Length of high-band frames*/
- int subframeSize; /**< Length of high-band sub-frames*/
- int nbSubframes; /**< Number of high-band sub-frames*/
- int windowSize; /**< Length of high-band LPC window*/
- int lpcSize; /**< Order of high-band LPC analysis */
- int bufSize; /**< Buffer size */
- int first; /**< First frame? */
- float lag_factor; /**< Lag-windowing control parameter */
- spx_word16_t lpc_floor; /**< Controls LPC analysis noise floor */
- spx_word16_t gamma1; /**< Perceptual weighting coef 1 */
- spx_word16_t gamma2; /**< Perceptual weighting coef 2 */
+ const SpeexMode *mode; /**< Pointer to the mode (containing for vtable info) */
+ void *st_low; /**< State of the low-band (narrowband) encoder */
+ int full_frame_size; /**< Length of full-band frames*/
+ int frame_size; /**< Length of high-band frames*/
+ int subframeSize; /**< Length of high-band sub-frames*/
+ int nbSubframes; /**< Number of high-band sub-frames*/
+ int windowSize; /**< Length of high-band LPC window*/
+ int lpcSize; /**< Order of high-band LPC analysis */
+ int bufSize; /**< Buffer size */
+ int first; /**< First frame? */
+ float lag_factor; /**< Lag-windowing control parameter */
+ spx_word16_t lpc_floor; /**< Controls LPC analysis noise floor */
+ spx_word16_t gamma1; /**< Perceptual weighting coef 1 */
+ spx_word16_t gamma2; /**< Perceptual weighting coef 2 */
- char *stack; /**< Temporary allocation stack */
- spx_sig_t *x0d, *x1d; /**< QMF filter signals*/
- spx_sig_t *high; /**< High-band signal (buffer) */
- spx_sig_t *y0, *y1; /**< QMF synthesis signals */
+ char *stack; /**< Temporary allocation stack */
+ spx_sig_t *x0d, *x1d; /**< QMF filter signals*/
+ spx_sig_t *high; /**< High-band signal (buffer) */
+ spx_sig_t *y0, *y1; /**< QMF synthesis signals */
spx_word16_t *h0_mem, *h1_mem;
spx_word32_t *g0_mem, *g1_mem; /**< QMF memories */
- spx_sig_t *excBuf; /**< High-band excitation */
- spx_sig_t *exc; /**< High-band excitation (for QMF only)*/
- spx_sig_t *res; /**< Zero-input response (ringing) */
- spx_sig_t *sw; /**< Perceptually weighted signal */
- spx_sig_t *target; /**< Weighted target signal (analysis by synthesis) */
- spx_word16_t *window; /**< LPC analysis window */
- spx_word16_t *lagWindow; /**< Auto-correlation window */
- spx_word16_t *autocorr; /**< Auto-correlation (for LPC analysis) */
- spx_coef_t *lpc; /**< LPC coefficients */
- spx_lsp_t *lsp; /**< LSP coefficients */
- spx_lsp_t *qlsp; /**< Quantized LSPs */
- spx_lsp_t *old_lsp; /**< LSPs of previous frame */
- spx_lsp_t *old_qlsp; /**< Quantized LSPs of previous frame */
- spx_lsp_t *interp_lsp; /**< Interpolated LSPs for current sub-frame */
- spx_lsp_t *interp_qlsp; /**< Interpolated quantized LSPs for current sub-frame */
- spx_coef_t *interp_lpc; /**< Interpolated LPCs for current sub-frame */
- spx_coef_t *interp_qlpc; /**< Interpolated quantized LPCs for current sub-frame */
- spx_coef_t *bw_lpc1; /**< Bandwidth-expanded version of LPCs (#1) */
- spx_coef_t *bw_lpc2; /**< Bandwidth-expanded version of LPCs (#2) */
+ spx_sig_t *excBuf; /**< High-band excitation */
+ spx_sig_t *exc; /**< High-band excitation (for QMF only)*/
+ spx_sig_t *res; /**< Zero-input response (ringing) */
+ spx_sig_t *sw; /**< Perceptually weighted signal */
+ const spx_word16_t *window; /**< LPC analysis window */
+ spx_word16_t *lagWindow; /**< Auto-correlation window */
+ spx_word16_t *autocorr; /**< Auto-correlation (for LPC analysis) */
+ spx_coef_t *lpc; /**< LPC coefficients */
+ spx_lsp_t *lsp; /**< LSP coefficients */
+ spx_lsp_t *qlsp; /**< Quantized LSPs */
+ spx_lsp_t *old_lsp; /**< LSPs of previous frame */
+ spx_lsp_t *old_qlsp; /**< Quantized LSPs of previous frame */
+ spx_lsp_t *interp_lsp; /**< Interpolated LSPs for current sub-frame */
+ spx_lsp_t *interp_qlsp; /**< Interpolated quantized LSPs for current sub-frame */
+ spx_coef_t *interp_lpc; /**< Interpolated LPCs for current sub-frame */
+ spx_coef_t *interp_qlpc; /**< Interpolated quantized LPCs for current sub-frame */
+ spx_coef_t *bw_lpc1; /**< Bandwidth-expanded version of LPCs (#1) */
+ spx_coef_t *bw_lpc2; /**< Bandwidth-expanded version of LPCs (#2) */
- spx_mem_t *mem_sp; /**< Synthesis signal memory */
+ spx_mem_t *mem_sp; /**< Synthesis signal memory */
spx_mem_t *mem_sp2;
- spx_mem_t *mem_sw; /**< Perceptual signal memory */
+ spx_mem_t *mem_sw; /**< Perceptual signal memory */
spx_word32_t *pi_gain;
+ spx_sig_t *innov_save; /**< If non-NULL, innovation is copied here */
+ spx_sig_t *low_innov; /**< Lower-band innovation is copied here magically */
- float vbr_quality; /**< Quality setting for VBR encoding */
- int vbr_enabled; /**< 1 for enabling VBR, 0 otherwise */
- int abr_enabled; /**< ABR setting (in bps), 0 if off */
+ float vbr_quality; /**< Quality setting for VBR encoding */
+ int vbr_enabled; /**< 1 for enabling VBR, 0 otherwise */
+ spx_int32_t vbr_max; /**< Max bit-rate allowed in VBR mode (total) */
+ spx_int32_t vbr_max_high; /**< Max bit-rate allowed in VBR mode for the high-band */
+ spx_int32_t abr_enabled; /**< ABR setting (in bps), 0 if off */
float abr_drift;
float abr_drift2;
float abr_count;
- int vad_enabled; /**< 1 for enabling VAD, 0 otherwise */
+ int vad_enabled; /**< 1 for enabling VAD, 0 otherwise */
float relative_quality;
int encode_submode;
@@ -103,7 +106,7 @@
int submodeID;
int submodeSelect;
int complexity;
- int sampling_rate;
+ spx_int32_t sampling_rate;
} SBEncState;
@@ -118,7 +121,7 @@
int nbSubframes;
int lpcSize;
int first;
- int sampling_rate;
+ spx_int32_t sampling_rate;
int lpc_enh_enabled;
char *stack;
@@ -128,6 +131,7 @@
spx_word32_t *g0_mem, *g1_mem;
spx_sig_t *exc;
+ spx_sig_t *excBuf;
spx_lsp_t *qlsp;
spx_lsp_t *old_qlsp;
spx_lsp_t *interp_qlsp;
@@ -135,6 +139,9 @@
spx_mem_t *mem_sp;
spx_word32_t *pi_gain;
+ spx_sig_t *innov_save; /** If non-NULL, innovation is copied here */
+ spx_sig_t *low_innov; /** Lower-band innovation is copied here magically */
+
spx_int32_t seed;
int encode_submode;
diff --git a/pjmedia/src/pjmedia-codec/speex/speex.h b/pjmedia/src/pjmedia-codec/speex/speex.h
index 0eb2b8a..c7f7547 100644
--- a/pjmedia/src/pjmedia-codec/speex/speex.h
+++ b/pjmedia/src/pjmedia-codec/speex/speex.h
@@ -141,6 +141,11 @@
/** Gets tuning for PLC */
#define SPEEX_GET_PLC_TUNING 41
+/** Sets the max bit-rate allowed in VBR mode */
+#define SPEEX_SET_VBR_MAX_BITRATE 42
+/** Gets the max bit-rate allowed in VBR mode */
+#define SPEEX_GET_VBR_MAX_BITRATE 43
+
/* Used internally, not to be used in applications */
/** Used internally*/
#define SPEEX_GET_PI_GAIN 100
@@ -150,6 +155,8 @@
#define SPEEX_GET_INNOV 102
/** Used internally*/
#define SPEEX_GET_DTX_STATUS 103
+/** Used internally*/
+#define SPEEX_SET_INNOVATION_SAVE 104
/* Preserving compatibility:*/
diff --git a/pjmedia/src/pjmedia-codec/speex/speex_echo.h b/pjmedia/src/pjmedia-codec/speex/speex_echo.h
index 1962a56..4813b5a 100644
--- a/pjmedia/src/pjmedia-codec/speex/speex_echo.h
+++ b/pjmedia/src/pjmedia-codec/speex/speex_echo.h
@@ -61,7 +61,13 @@
void speex_echo_state_destroy(SpeexEchoState *st);
/** Performs echo cancellation a frame */
-void speex_echo_cancel(SpeexEchoState *st, short *ref, short *echo, short *out, spx_int32_t *Y);
+void speex_echo_cancel(SpeexEchoState *st, const spx_int16_t *rec, const spx_int16_t *play, spx_int16_t *out, spx_int32_t *Yout);
+
+/** Perform echo cancellation using internal playback buffer */
+void speex_echo_capture(SpeexEchoState *st, const spx_int16_t *rec, spx_int16_t *out, spx_int32_t *Yout);
+
+/** Let the echo canceller know that a frame was just played */
+void speex_echo_playback(SpeexEchoState *st, const spx_int16_t *play);
/** Reset the echo canceller state */
void speex_echo_state_reset(SpeexEchoState *st);
diff --git a/pjmedia/src/pjmedia-codec/speex/speex_jitter.h b/pjmedia/src/pjmedia-codec/speex/speex_jitter.h
index 31b5c53..34043b3 100644
--- a/pjmedia/src/pjmedia-codec/speex/speex_jitter.h
+++ b/pjmedia/src/pjmedia-codec/speex/speex_jitter.h
@@ -43,32 +43,55 @@
extern "C" {
#endif
-#define SPEEX_JITTER_MAX_PACKET_SIZE 1500 /**< Maximum number of bytes per packet */
-#define SPEEX_JITTER_MAX_BUFFER_SIZE 20 /**< Maximum number of packets in jitter buffer */
+struct JitterBuffer_;
-#define MAX_MARGIN 12 /**< Number of bins in margin histogram */
+typedef struct JitterBuffer_ JitterBuffer;
+
+typedef struct _JitterBufferPacket JitterBufferPacket;
+
+struct _JitterBufferPacket {
+ char *data;
+ spx_uint32_t len;
+ spx_uint32_t timestamp;
+ spx_uint32_t span;
+};
+
+
+#define JITTER_BUFFER_OK 0
+#define JITTER_BUFFER_MISSING 1
+#define JITTER_BUFFER_INCOMPLETE 2
+#define JITTER_BUFFER_INTERNAL_ERROR -1
+#define JITTER_BUFFER_BAD_ARGUMENT -2
+
+/** Initialise jitter buffer */
+JitterBuffer *jitter_buffer_init(int tick);
+
+/** Reset jitter buffer */
+void jitter_buffer_reset(JitterBuffer *jitter);
+
+/** Destroy jitter buffer */
+void jitter_buffer_destroy(JitterBuffer *jitter);
+
+/** Put one packet into the jitter buffer */
+void jitter_buffer_put(JitterBuffer *jitter, const JitterBufferPacket *packet);
+
+/** Get one packet from the jitter buffer */
+int jitter_buffer_get(JitterBuffer *jitter, JitterBufferPacket *packet, spx_uint32_t *current_timestamp);
+
+/** Get pointer timestamp of jitter buffer */
+int jitter_buffer_get_pointer_timestamp(JitterBuffer *jitter);
+
+/** Advance by one tick */
+void jitter_buffer_tick(JitterBuffer *jitter);
+
/** Speex jitter-buffer state. */
typedef struct SpeexJitter {
- int buffer_size; /**< Buffer size */
- int pointer_timestamp; /**< Pointer timestamp */
-
SpeexBits current_packet; /**< Current Speex packet */
int valid_bits; /**< True if Speex bits are valid */
-
- char buf[SPEEX_JITTER_MAX_BUFFER_SIZE][SPEEX_JITTER_MAX_PACKET_SIZE]; /**< Buffer of packets */
- int timestamp[SPEEX_JITTER_MAX_BUFFER_SIZE]; /**< Timestamp of packet */
- int len[SPEEX_JITTER_MAX_BUFFER_SIZE]; /**< Number of bytes in packet */
-
+ JitterBuffer *packets;
void *dec; /**< Pointer to Speex decoder */
int frame_size; /**< Frame size of Speex decoder */
- int frame_time; /**< Frame time in [ms] of Speex decoder */
- int reset_state; /**< True if Speex state was reset */
-
- int lost_count; /**< Number of lost packets */
- float shortterm_margin[MAX_MARGIN]; /**< Short term margins */
- float longterm_margin[MAX_MARGIN]; /**< Long term margins */
- float loss_rate; /**< Loss rate */
} SpeexJitter;
/** Initialise jitter buffer */
@@ -81,7 +104,7 @@
void speex_jitter_put(SpeexJitter *jitter, char *packet, int len, int timestamp);
/** Get one packet from the jitter buffer */
-void speex_jitter_get(SpeexJitter *jitter, short *out, int *current_timestamp);
+void speex_jitter_get(SpeexJitter *jitter, spx_int16_t *out, int *start_offset);
/** Get pointer timestamp of jitter buffer */
int speex_jitter_get_pointer_timestamp(SpeexJitter *jitter);
diff --git a/pjmedia/src/pjmedia-codec/speex/speex_stereo.h b/pjmedia/src/pjmedia-codec/speex/speex_stereo.h
index 0b70021..6ccaa31 100644
--- a/pjmedia/src/pjmedia-codec/speex/speex_stereo.h
+++ b/pjmedia/src/pjmedia-codec/speex/speex_stereo.h
@@ -53,7 +53,7 @@
} SpeexStereoState;
/** Initialization value for a stereo state */
-#define SPEEX_STEREO_STATE_INIT {1,.5,1,1}
+#define SPEEX_STEREO_STATE_INIT {1,.5,1,1,0,0}
/** Transforms a stereo frame into a mono frame and stores intensity stereo info in 'bits' */
void speex_encode_stereo(float *data, int frame_size, SpeexBits *bits);
diff --git a/pjmedia/src/pjmedia-codec/speex/speex_types.h b/pjmedia/src/pjmedia-codec/speex/speex_types.h
index b67c74f..a6ebf0c 100644
--- a/pjmedia/src/pjmedia-codec/speex/speex_types.h
+++ b/pjmedia/src/pjmedia-codec/speex/speex_types.h
@@ -26,7 +26,6 @@
# if defined(__CYGWIN__)
# include <_G_config.h>
- typedef _G_int64_t spx_int64_t;
typedef _G_int32_t spx_int32_t;
typedef _G_uint32_t spx_uint32_t;
typedef _G_int16_t spx_int16_t;
@@ -36,17 +35,13 @@
typedef unsigned short spx_uint16_t;
typedef int spx_int32_t;
typedef unsigned int spx_uint32_t;
- typedef long long spx_int64_t;
- typedef unsigned long long spx_uint64_t;
# elif defined(__MWERKS__)
- typedef long long spx_int64_t;
typedef int spx_int32_t;
typedef unsigned int spx_uint32_t;
typedef short spx_int16_t;
typedef unsigned short spx_uint16_t;
# else
/* MSVC/Borland */
- typedef __int64 spx_int64_t;
typedef __int32 spx_int32_t;
typedef unsigned __int32 spx_uint32_t;
typedef __int16 spx_int16_t;
@@ -60,7 +55,6 @@
typedef UInt16 spx_uint16_t;
typedef SInt32 spx_int32_t;
typedef UInt32 spx_uint32_t;
- typedef SInt64 spx_int64_t;
#elif defined(__MACOSX__) /* MacOS X Framework build */
@@ -69,7 +63,6 @@
typedef u_int16_t spx_uint16_t;
typedef int32_t spx_int32_t;
typedef u_int32_t spx_uint32_t;
- typedef int64_t spx_int64_t;
#elif defined(__BEOS__)
@@ -79,7 +72,6 @@
typedef u_int16_t spx_uint16_t;
typedef int32_t spx_int32_t;
typedef u_int32_t spx_uint32_t;
- typedef int64_t spx_int64_t;
#elif defined (__EMX__)
@@ -88,7 +80,6 @@
typedef unsigned short spx_uint16_t;
typedef int spx_int32_t;
typedef unsigned int spx_uint32_t;
- typedef long long spx_int64_t;
#elif defined (DJGPP)
@@ -96,12 +87,10 @@
typedef short spx_int16_t;
typedef int spx_int32_t;
typedef unsigned int spx_uint32_t;
- typedef long long spx_int64_t;
#elif defined(R5900)
/* PS2 EE */
- typedef long spx_int64_t;
typedef int spx_int32_t;
typedef unsigned spx_uint32_t;
typedef short spx_int16_t;
@@ -113,7 +102,6 @@
typedef unsigned short spx_uint16_t;
typedef signed int spx_int32_t;
typedef unsigned int spx_uint32_t;
- typedef long long int spx_int64_t;
#elif defined(CONFIG_TI_C54X) || defined (CONFIG_TI_C55X)
@@ -122,7 +110,7 @@
typedef long spx_int32_t;
typedef unsigned long spx_uint32_t;
-#elif defined(CONFIG_TI_C5X)
+#elif defined(CONFIG_TI_C6X)
typedef short spx_int16_t;
typedef unsigned short spx_uint16_t;
diff --git a/pjmedia/src/pjmedia-codec/speex/stack_alloc.h b/pjmedia/src/pjmedia-codec/speex/stack_alloc.h
index 6270d12..cb048fa 100644
--- a/pjmedia/src/pjmedia-codec/speex/stack_alloc.h
+++ b/pjmedia/src/pjmedia-codec/speex/stack_alloc.h
@@ -114,7 +114,7 @@
#define ALLOC(var, size, type) type var[size]
#elif defined(USE_ALLOCA)
#define VARDECL(var) var
-#define ALLOC(var, size, type) var = alloca(sizeof(type)*size)
+#define ALLOC(var, size, type) var = alloca(sizeof(type)*(size))
#else
#define VARDECL(var) var
#define ALLOC(var, size, type) var = PUSH(stack, size, type)
diff --git a/pjmedia/src/pjmedia-codec/speex/window.c b/pjmedia/src/pjmedia-codec/speex/window.c
new file mode 100644
index 0000000..3748f65
--- /dev/null
+++ b/pjmedia/src/pjmedia-codec/speex/window.c
@@ -0,0 +1,94 @@
+/* Copyright (C) 2006 Jean-Marc Valin
+ File: window.c
+
+ Redistribution and use in source and binary forms, with or without
+ modification, are permitted provided that the following conditions
+ are met:
+
+ - Redistributions of source code must retain the above copyright
+ notice, this list of conditions and the following disclaimer.
+
+ - Redistributions in binary form must reproduce the above copyright
+ notice, this list of conditions and the following disclaimer in the
+ documentation and/or other materials provided with the distribution.
+
+ - Neither the name of the Xiph.org Foundation nor the names of its
+ contributors may be used to endorse or promote products derived from
+ this software without specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR
+ CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+ EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+ PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+ LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+ NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "misc.h"
+
+#ifdef FIXED_POINT
+const spx_word16_t lpc_window[200] = {
+1310, 1313, 1321, 1333, 1352, 1375, 1403, 1436,
+1475, 1518, 1567, 1621, 1679, 1743, 1811, 1884,
+1962, 2044, 2132, 2224, 2320, 2421, 2526, 2636,
+2750, 2868, 2990, 3116, 3246, 3380, 3518, 3659,
+3804, 3952, 4104, 4259, 4417, 4578, 4742, 4909,
+5079, 5251, 5425, 5602, 5781, 5963, 6146, 6331,
+6518, 6706, 6896, 7087, 7280, 7473, 7668, 7863,
+8059, 8256, 8452, 8650, 8847, 9044, 9241, 9438,
+9635, 9831, 10026, 10220, 10414, 10606, 10797, 10987,
+11176, 11363, 11548, 11731, 11912, 12091, 12268, 12443,
+12615, 12785, 12952, 13116, 13277, 13435, 13590, 13742,
+13890, 14035, 14176, 14314, 14448, 14578, 14704, 14826,
+14944, 15058, 15168, 15273, 15374, 15470, 15562, 15649,
+15732, 15810, 15883, 15951, 16015, 16073, 16127, 16175,
+16219, 16257, 16291, 16319, 16342, 16360, 16373, 16381,
+16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384,
+16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384,
+16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384,
+16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384,
+16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384,
+16384, 16384, 16384, 16384, 16384, 16384, 16384, 16384,
+16384, 16384, 16384, 16361, 16294, 16183, 16028, 15830,
+15588, 15304, 14979, 14613, 14207, 13763, 13282, 12766,
+12215, 11631, 11016, 10373, 9702, 9007, 8289, 7551,
+6797, 6028, 5251, 4470, 3695, 2943, 2248, 1696
+};
+#else
+const spx_word16_t lpc_window[200] = {
+ 0.080000, 0.080158, 0.080630, 0.081418, 0.082520, 0.083935, 0.085663, 0.087703,
+ 0.090052, 0.092710, 0.095674, 0.098943, 0.102514, 0.106385, 0.110553, 0.115015,
+ 0.119769, 0.124811, 0.130137, 0.135744, 0.141628, 0.147786, 0.154212, 0.160902,
+ 0.167852, 0.175057, 0.182513, 0.190213, 0.198153, 0.206328, 0.214731, 0.223357,
+ 0.232200, 0.241254, 0.250513, 0.259970, 0.269619, 0.279453, 0.289466, 0.299651,
+ 0.310000, 0.320507, 0.331164, 0.341965, 0.352901, 0.363966, 0.375151, 0.386449,
+ 0.397852, 0.409353, 0.420943, 0.432615, 0.444361, 0.456172, 0.468040, 0.479958,
+ 0.491917, 0.503909, 0.515925, 0.527959, 0.540000, 0.552041, 0.564075, 0.576091,
+ 0.588083, 0.600042, 0.611960, 0.623828, 0.635639, 0.647385, 0.659057, 0.670647,
+ 0.682148, 0.693551, 0.704849, 0.716034, 0.727099, 0.738035, 0.748836, 0.759493,
+ 0.770000, 0.780349, 0.790534, 0.800547, 0.810381, 0.820030, 0.829487, 0.838746,
+ 0.847800, 0.856643, 0.865269, 0.873672, 0.881847, 0.889787, 0.897487, 0.904943,
+ 0.912148, 0.919098, 0.925788, 0.932214, 0.938372, 0.944256, 0.949863, 0.955189,
+ 0.960231, 0.964985, 0.969447, 0.973615, 0.977486, 0.981057, 0.984326, 0.987290,
+ 0.989948, 0.992297, 0.994337, 0.996065, 0.997480, 0.998582, 0.999370, 0.999842,
+ 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+ 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+ 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+ 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+ 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+ 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
+ 1.000000, 1.000000, 1.000000, 0.998640, 0.994566, 0.987787, 0.978324, 0.966203,
+ 0.951458, 0.934131, 0.914270, 0.891931, 0.867179, 0.840084, 0.810723, 0.779182,
+ 0.745551, 0.709930, 0.672424, 0.633148, 0.592223, 0.549781, 0.505964, 0.460932,
+ 0.414863, 0.367968, 0.320511, 0.272858, 0.225569, 0.179655, 0.137254, 0.103524
+};
+#endif
diff --git a/pjmedia/src/pjmedia-codec/speex_codec.c b/pjmedia/src/pjmedia-codec/speex_codec.c
index 6a64c77..b434ee8 100644
--- a/pjmedia/src/pjmedia-codec/speex_codec.c
+++ b/pjmedia/src/pjmedia-codec/speex_codec.c
@@ -683,9 +683,7 @@
struct pjmedia_frame *output)
{
struct spx_private *spx;
- float tmp[642]; /* 20ms at 32KHz + 2 */
- pj_int16_t *samp_in;
- unsigned i, samp_count, sz;
+ unsigned sz;
int tx;
spx = (struct spx_private*) codec->codec_data;
@@ -698,19 +696,11 @@
return PJ_SUCCESS;
}
- /* Copy frame to float buffer. */
- samp_count = input->size / 2;
- pj_assert(samp_count <= PJ_ARRAY_SIZE(tmp));
- samp_in = input->buf;
- for (i=0; i<samp_count; ++i) {
- tmp[i] = samp_in[i];
- }
-
/* Flush all the bits in the struct so we can encode a new frame */
speex_bits_reset(&spx->enc_bits);
/* Encode the frame */
- tx = speex_encode(spx->enc, tmp, &spx->enc_bits);
+ tx = speex_encode_int(spx->enc, input->buf, &spx->enc_bits);
/* Check if we need not to transmit the frame (DTX) */
if (tx == 0) {
@@ -743,9 +733,6 @@
struct pjmedia_frame *output)
{
struct spx_private *spx;
- float tmp[642]; /* 20ms at 32KHz + 2 */
- pj_int16_t *dst_buf;
- unsigned i, count, sz;
spx = (struct spx_private*) codec->codec_data;
@@ -764,21 +751,11 @@
speex_bits_read_from(&spx->dec_bits, input->buf, input->size);
/* Decode the data */
- speex_decode(spx->dec, &spx->dec_bits, tmp);
+ speex_decode_int(spx->dec, &spx->dec_bits, output->buf);
- /* Check size. */
- sz = speex_bits_nbytes(&spx->enc_bits);
- pj_assert(sz <= output_buf_len);
-
- /* Copy from float to short samples. */
- count = spx_factory.speex_param[spx->param_id].clock_rate * 20 / 1000;
- pj_assert((count <= output_buf_len/2) && count <= PJ_ARRAY_SIZE(tmp));
- dst_buf = output->buf;
- for (i=0; i<count; ++i) {
- dst_buf[i] = (pj_int16_t)tmp[i];
- }
output->type = PJMEDIA_FRAME_TYPE_AUDIO;
- output->size = count * 2;
+ output->size = speex_bits_nbytes(&spx->dec_bits);
+ pj_assert(output->size <= (int)output_buf_len);
output->timestamp.u64 = input->timestamp.u64;
@@ -793,9 +770,7 @@
struct pjmedia_frame *output)
{
struct spx_private *spx;
- float tmp[642]; /* 20ms at 32KHz + 2 */
- pj_int16_t *dst_buf;
- unsigned i, count;
+ unsigned count;
/* output_buf_len is unreferenced when building in Release mode */
PJ_UNUSED_ARG(output_buf_len);
@@ -803,16 +778,11 @@
spx = (struct spx_private*) codec->codec_data;
count = spx_factory.speex_param[spx->param_id].clock_rate * 20 / 1000;
- pj_assert((count <= output_buf_len/2) && count <= PJ_ARRAY_SIZE(tmp));
+ pj_assert(count <= output_buf_len/2);
/* Recover packet loss */
- speex_decode(spx->dec, NULL, tmp);
+ speex_decode_int(spx->dec, NULL, output->buf);
- /* Copy from float to short samples. */
- dst_buf = output->buf;
- for (i=0; i<count; ++i) {
- dst_buf[i] = (pj_int16_t)tmp[i];
- }
output->size = count * 2;
return PJ_SUCCESS;