Skip to content

Commit a4854af

Browse files
jmvalinjmvalin
committed
Preserving 24-bit accuracy for fixed-point decoder
Convert to 16 bits only at the very end
1 parent 4141c4d commit a4854af

14 files changed

+207
-152
lines changed

‎celt/arch.h

+35-4
Original file line numberDiff line numberDiff line change
@@ -106,6 +106,14 @@ void celt_fatal(const char *str, const char *file, int line)
106106
#define UADD32(a,b) ((a)+(b))
107107
#define USUB32(a,b) ((a)-(b))
108108

109+
/* Throughout the code, we use the following scaling for signals:
110+
FLOAT: used for float API, normalized to +/-1.
111+
INT16: used for 16-bit API, normalized to +/- 32768
112+
RES: internal Opus resolution, defined as +/-1. in float builds, or either 16-bit or 24-bit int for fixed-point builds
113+
SIG: internal CELT resolution: defined as +/- 32768. in float builds, or Q27 in fixed-point builds (int16 shifted by 12)
114+
*/
115+
116+
109117
/* Set this if opus_int64 is a native type of the CPU. */
110118
/* Assume that all LP64 architectures have fast 64-bit types; also x86_64
111119
(which can be ILP32 for x32) and Win64 (which is LLP64). */
@@ -127,6 +135,26 @@ typedef opus_val32 celt_sig;
127135
typedef opus_val16 celt_norm;
128136
typedef opus_val32 celt_ener;
129137

138+
#ifdef ENABLE_RES24
139+
typedef opus_val32 opus_res;
140+
#define RES_SHIFT 8
141+
#define SCALEIN(a) (a)
142+
#define SIG2RES(a) PSHR32(a, SIG_SHIFT-RES_SHIFT)
143+
#define RES2INT16(a) SAT16(PSHR32(a, RES_SHIFT))
144+
#define RES2FLOAT(a) ((1.f/32768.f/256.)*(a))
145+
#define INT16TORES(a) SHL32(EXTEND32(a), RES_SHIFT)
146+
#define ADD_RES(a, b) ADD32(a, b)
147+
#else
148+
typedef opus_val16 opus_res;
149+
#define RES_SHIFT 0
150+
#define SCALEIN(a) (a)
151+
#define SIG2RES(a) SIG2WORD16(a)
152+
#define RES2INT16(a) (a)
153+
#define RES2FLOAT(a) ((1.f/32768.f)*(a))
154+
#define INT16TORES(a) (a)
155+
#define ADD_RES(a, b) SAT16(ADD32((a), (b)));
156+
#endif
157+
130158
#define celt_isnan(x) 0
131159

132160
#define Q15ONE 32767
@@ -150,8 +178,6 @@ typedef opus_val32 celt_ener;
150178
#define VERY_LARGE16 ((opus_val16)32767)
151179
#define Q15_ONE ((opus_val16)32767)
152180

153-
#define SCALEIN(a) (a)
154-
#define SCALEOUT(a) (a)
155181

156182
#define ABS16(x) ((x) < 0 ? (-(x)) : (x))
157183
#define ABS32(x) ((x) < 0 ? (-(x)) : (x))
@@ -192,6 +218,8 @@ typedef float celt_sig;
192218
typedef float celt_norm;
193219
typedef float celt_ener;
194220

221+
typedef float opus_res;
222+
195223
#ifdef FLOAT_APPROX
196224
/* This code should reliably detect NaN/inf even when -ffast-math is used.
197225
Assumes IEEE 754 format. */
@@ -279,9 +307,12 @@ static OPUS_INLINE int celt_isnan(float x)
279307
#define DIV32(a,b) (((opus_val32)(a))/(opus_val32)(b))
280308

281309
#define SCALEIN(a) ((a)*CELT_SIG_SCALE)
282-
#define SCALEOUT(a) ((a)*(1/CELT_SIG_SCALE))
283310

284-
#define SIG2WORD16(x) (x)
311+
#define SIG2RES(a) ((1/CELT_SIG_SCALE)*(a))
312+
#define RES2INT16(a) FLOAT2INT16(a)
313+
#define RES2FLOAT(a) (a)
314+
#define INT16TORES(a) ((a)*(1/CELT_SIG_SCALE))
315+
#define ADD_RES(a, b) ADD32(a, b)
285316

286317
#endif /* !FIXED_POINT */
287318

‎celt/celt.h

+3-3
Original file line numberDiff line numberDiff line change
@@ -154,14 +154,14 @@ int celt_decoder_get_size(int channels);
154154
int celt_decoder_init(CELTDecoder *st, opus_int32 sampling_rate, int channels);
155155

156156
int celt_decode_with_ec_dred(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data,
157-
int len, opus_val16 * OPUS_RESTRICT pcm, int frame_size, ec_dec *dec, int accum
157+
int len, opus_res * OPUS_RESTRICT pcm, int frame_size, ec_dec *dec, int accum
158158
#ifdef ENABLE_DEEP_PLC
159159
,LPCNetPLCState *lpcnet
160160
#endif
161161
);
162162

163163
int celt_decode_with_ec(OpusCustomDecoder * OPUS_RESTRICT st, const unsigned char *data,
164-
int len, opus_val16 * OPUS_RESTRICT pcm, int frame_size, ec_dec *dec, int accum);
164+
int len, opus_res * OPUS_RESTRICT pcm, int frame_size, ec_dec *dec, int accum);
165165

166166
#define celt_encoder_ctl opus_custom_encoder_ctl
167167
#define celt_decoder_ctl opus_custom_decoder_ctl
@@ -239,7 +239,7 @@ void comb_filter(opus_val32 *y, opus_val32 *x, int T0, int T1, int N,
239239
void init_caps(const CELTMode *m,int *cap,int LM,int C);
240240

241241
#ifdef RESYNTH
242-
void deemphasis(celt_sig *in[], opus_val16 *pcm, int N, int C, int downsample, const opus_val16 *coef, celt_sig *mem, int accum);
242+
void deemphasis(celt_sig *in[], opus_res *pcm, int N, int C, int downsample, const opus_val16 *coef, celt_sig *mem, int accum);
243243
void celt_synthesis(const CELTMode *mode, celt_norm *X, celt_sig * out_syn[],
244244
opus_val16 *oldBandE, int start, int effEnd, int C, int CC, int isTransient,
245245
int LM, int downsample, int silence, int arch);

‎celt/celt_decoder.c

+39-23
Original file line numberDiff line numberDiff line change
@@ -246,7 +246,7 @@ void opus_custom_decoder_destroy(CELTDecoder *st)
246246
/* Special case for stereo with no downsampling and no accumulation. This is
247247
quite common and we can make it faster by processing both channels in the
248248
same loop, reducing overhead due to the dependency loop in the IIR filter. */
249-
static void deemphasis_stereo_simple(celt_sig *in[], opus_val16 *pcm, int N, const opus_val16 coef0,
249+
static void deemphasis_stereo_simple(celt_sig *in[], opus_res *pcm, int N, const opus_val16 coef0,
250250
celt_sig *mem)
251251
{
252252
celt_sig * OPUS_RESTRICT x0;
@@ -265,8 +265,8 @@ static void deemphasis_stereo_simple(celt_sig *in[], opus_val16 *pcm, int N, con
265265
tmp1 = SATURATE(x1[j] + VERY_SMALL + m1, SIG_SAT);
266266
m0 = MULT16_32_Q15(coef0, tmp0);
267267
m1 = MULT16_32_Q15(coef0, tmp1);
268-
pcm[2*j ] = SCALEOUT(SIG2WORD16(tmp0));
269-
pcm[2*j+1] = SCALEOUT(SIG2WORD16(tmp1));
268+
pcm[2*j ] = SIG2RES(tmp0);
269+
pcm[2*j+1] = SIG2RES(tmp1);
270270
}
271271
mem[0] = m0;
272272
mem[1] = m1;
@@ -276,7 +276,7 @@ static void deemphasis_stereo_simple(celt_sig *in[], opus_val16 *pcm, int N, con
276276
#ifndef RESYNTH
277277
static
278278
#endif
279-
void deemphasis(celt_sig *in[], opus_val16 *pcm, int N, int C, int downsample, const opus_val16 *coef,
279+
void deemphasis(celt_sig *in[], opus_res *pcm, int N, int C, int downsample, const opus_val16 *coef,
280280
celt_sig *mem, int accum)
281281
{
282282
int c;
@@ -292,18 +292,14 @@ void deemphasis(celt_sig *in[], opus_val16 *pcm, int N, int C, int downsample, c
292292
deemphasis_stereo_simple(in, pcm, N, coef[0], mem);
293293
return;
294294
}
295-
#endif
296-
#ifndef FIXED_POINT
297-
(void)accum;
298-
celt_assert(accum==0);
299295
#endif
300296
ALLOC(scratch, N, celt_sig);
301297
coef0 = coef[0];
302298
Nd = N/downsample;
303299
c=0; do {
304300
int j;
305301
celt_sig * OPUS_RESTRICT x;
306-
opus_val16 * OPUS_RESTRICT y;
302+
opus_res * OPUS_RESTRICT y;
307303
celt_sig m = mem[c];
308304
x =in[c];
309305
y = pcm+c;
@@ -335,23 +331,21 @@ void deemphasis(celt_sig *in[], opus_val16 *pcm, int N, int C, int downsample, c
335331
apply_downsampling=1;
336332
} else {
337333
/* Shortcut for the standard (non-custom modes) case */
338-
#ifdef FIXED_POINT
339334
if (accum)
340335
{
341336
for (j=0;j<N;j++)
342337
{
343338
celt_sig tmp = SATURATE(x[j] + m + VERY_SMALL, SIG_SAT);
344339
m = MULT16_32_Q15(coef0, tmp);
345-
y[j*C] = SAT16(ADD32(y[j*C], SCALEOUT(SIG2WORD16(tmp))));
340+
y[j*C] = ADD_RES(y[j*C], SIG2RES(tmp));
346341
}
347342
} else
348-
#endif
349343
{
350344
for (j=0;j<N;j++)
351345
{
352346
celt_sig tmp = SATURATE(x[j] + VERY_SMALL + m, SIG_SAT);
353347
m = MULT16_32_Q15(coef0, tmp);
354-
y[j*C] = SCALEOUT(SIG2WORD16(tmp));
348+
y[j*C] = SIG2RES(tmp);
355349
}
356350
}
357351
}
@@ -360,16 +354,14 @@ void deemphasis(celt_sig *in[], opus_val16 *pcm, int N, int C, int downsample, c
360354
if (apply_downsampling)
361355
{
362356
/* Perform down-sampling */
363-
#ifdef FIXED_POINT
364357
if (accum)
365358
{
366359
for (j=0;j<Nd;j++)
367-
y[j*C] = SAT16(ADD32(y[j*C], SCALEOUT(SIG2WORD16(scratch[j*downsample]))));
360+
y[j*C] = ADD_RES(y[j*C], SIG2RES(scratch[j*downsample]));
368361
} else
369-
#endif
370362
{
371363
for (j=0;j<Nd;j++)
372-
y[j*C] = SCALEOUT(SIG2WORD16(scratch[j*downsample]));
364+
y[j*C] = SIG2RES(scratch[j*downsample]);
373365
}
374366
}
375367
} while (++c<C);
@@ -968,7 +960,7 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM
968960
}
969961

970962
int celt_decode_with_ec_dred(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data,
971-
int len, opus_val16 * OPUS_RESTRICT pcm, int frame_size, ec_dec *dec, int accum
963+
int len, opus_res * OPUS_RESTRICT pcm, int frame_size, ec_dec *dec, int accum
972964
#ifdef ENABLE_DEEP_PLC
973965
,LPCNetPLCState *lpcnet
974966
#endif
@@ -1369,7 +1361,7 @@ int celt_decode_with_ec_dred(CELTDecoder * OPUS_RESTRICT st, const unsigned char
13691361
}
13701362

13711363
int celt_decode_with_ec(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data,
1372-
int len, opus_val16 * OPUS_RESTRICT pcm, int frame_size, ec_dec *dec, int accum)
1364+
int len, opus_res * OPUS_RESTRICT pcm, int frame_size, ec_dec *dec, int accum)
13731365
{
13741366
return celt_decode_with_ec_dred(st, data, len, pcm, frame_size, dec, accum
13751367
#ifdef ENABLE_DEEP_PLC
@@ -1381,16 +1373,40 @@ int celt_decode_with_ec(CELTDecoder * OPUS_RESTRICT st, const unsigned char *dat
13811373
#ifdef CUSTOM_MODES
13821374

13831375
#ifdef FIXED_POINT
1376+
#ifdef ENABLE_RES24
1377+
int opus_custom_decode(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data, int len, opus_int16 * OPUS_RESTRICT pcm, int frame_size)
1378+
{
1379+
int j, ret, C, N;
1380+
VARDECL(opus_res, out);
1381+
ALLOC_STACK;
1382+
1383+
if (pcm==NULL)
1384+
return OPUS_BAD_ARG;
1385+
1386+
C = st->channels;
1387+
N = frame_size;
1388+
1389+
ALLOC(out, C*N, opus_res);
1390+
ret = celt_decode_with_ec(st, data, len, out, frame_size, NULL, 0);
1391+
if (ret>0)
1392+
for (j=0;j<C*ret;j++)
1393+
pcm[j]=RES2INT16(out[j]);
1394+
1395+
RESTORE_STACK;
1396+
return ret;
1397+
}
1398+
#else
13841399
int opus_custom_decode(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data, int len, opus_int16 * OPUS_RESTRICT pcm, int frame_size)
13851400
{
13861401
return celt_decode_with_ec(st, data, len, pcm, frame_size, NULL, 0);
13871402
}
1403+
#endif
13881404

13891405
#ifndef DISABLE_FLOAT_API
13901406
int opus_custom_decode_float(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data, int len, float * OPUS_RESTRICT pcm, int frame_size)
13911407
{
13921408
int j, ret, C, N;
1393-
VARDECL(opus_int16, out);
1409+
VARDECL(opus_res, out);
13941410
ALLOC_STACK;
13951411

13961412
if (pcm==NULL)
@@ -1399,11 +1415,11 @@ int opus_custom_decode_float(CELTDecoder * OPUS_RESTRICT st, const unsigned char
13991415
C = st->channels;
14001416
N = frame_size;
14011417

1402-
ALLOC(out, C*N, opus_int16);
1418+
ALLOC(out, C*N, opus_res);
14031419
ret=celt_decode_with_ec(st, data, len, out, frame_size, NULL, 0);
14041420
if (ret>0)
14051421
for (j=0;j<C*ret;j++)
1406-
pcm[j]=out[j]*(1.f/32768.f);
1422+
pcm[j]=RES2FLOAT(out[j]);
14071423

14081424
RESTORE_STACK;
14091425
return ret;
@@ -1434,7 +1450,7 @@ int opus_custom_decode(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data
14341450

14351451
if (ret>0)
14361452
for (j=0;j<C*ret;j++)
1437-
pcm[j] = FLOAT2INT16 (out[j]);
1453+
pcm[j] = RES2INT16 (out[j]);
14381454

14391455
RESTORE_STACK;
14401456
return ret;

‎celt/celt_encoder.c

+1-1
Original file line numberDiff line numberDiff line change
@@ -2591,7 +2591,7 @@ int opus_custom_encode(CELTEncoder * OPUS_RESTRICT st, const opus_int16 * pcm, i
25912591
N=frame_size;
25922592
ALLOC(in, C*N, celt_sig);
25932593
for (j=0;j<C*N;j++) {
2594-
in[j] = SCALEOUT(pcm[j]);
2594+
in[j] = (1.0f/32768)*pcm[j];
25952595
}
25962596

25972597
ret = celt_encode_with_ec(st,in,frame_size,compressed,nbCompressedBytes, NULL);

‎configure.ac

+8
Original file line numberDiff line numberDiff line change
@@ -151,6 +151,14 @@ AS_IF([test "$enable_float_api" = "no"],[
151151
AC_DEFINE([DISABLE_FLOAT_API], [1], [Do not build the float API])
152152
])
153153

154+
AC_ARG_ENABLE([fixed-res24],
155+
[AS_HELP_STRING([--enable-fixed-res24], [Use 24-bit internal resolution for fixed-point implementation])],,
156+
[enable_fixed_res24=no])
157+
158+
AS_IF([test "$enable_fixed_res24" = "yes"],[
159+
AC_DEFINE([ENABLE_RES24], [1], [24-bit internal resolution for fixed-point])
160+
])
161+
154162
AC_ARG_ENABLE([custom-modes],
155163
[AS_HELP_STRING([--enable-custom-modes], [enable non-Opus modes, e.g. 44.1 kHz & 2^n frames])],,
156164
[enable_custom_modes=no])

‎silk/API.h

+1-1
Original file line numberDiff line numberDiff line change
@@ -129,7 +129,7 @@ opus_int silk_Decode( /* O Returns error co
129129
opus_int lostFlag, /* I 0: no loss, 1 loss, 2 decode fec */
130130
opus_int newPacketFlag, /* I Indicates first decoder call for this packet */
131131
ec_dec *psRangeDec, /* I/O Compressor data structure */
132-
opus_int16 *samplesOut, /* O Decoded output speech vector */
132+
opus_res *samplesOut, /* O Decoded output speech vector */
133133
opus_int32 *nSamplesOut, /* O Number of samples decoded */
134134
#ifdef ENABLE_DEEP_PLC
135135
LPCNetPLCState *lpcnet,

0 commit comments

Comments
 (0)