comparison libtwamr/c4_17pf.c @ 322:fc82b113977f

libtwamr: integrate c4_17pf.c
author Mychaela Falconia <falcon@freecalypso.org>
date Thu, 18 Apr 2024 19:41:41 +0000
parents
children
comparison
equal deleted inserted replaced
321:5fc9a2d12048 322:fc82b113977f
1 /*
2 ********************************************************************************
3 *
4 * GSM AMR-NB speech codec R98 Version 7.6.0 December 12, 2001
5 * R99 Version 3.3.0
6 * REL-4 Version 4.1.0
7 *
8 ********************************************************************************
9 *
10 * File : c4_17pf.c
11 * Purpose : Searches a 17 bit algebraic codebook containing 4 pulses
12 * in a frame of 40 samples.
13 *
14 ********************************************************************************
15 */
16
17 /*
18 ********************************************************************************
19 * MODULE INCLUDE FILE AND VERSION ID
20 ********************************************************************************
21 */
22 #include "namespace.h"
23 #include "c4_17pf.h"
24
25 /*
26 ********************************************************************************
27 * INCLUDE FILES
28 ********************************************************************************
29 */
30 #include "typedef.h"
31 #include "basic_op.h"
32 #include "no_count.h"
33 #include "inv_sqrt.h"
34 #include "cnst.h"
35 #include "cor_h.h"
36 #include "set_sign.h"
37 #include "graytab.h"
38
39 /*
40 ********************************************************************************
41 * LOCAL VARIABLES AND TABLES
42 ********************************************************************************
43 */
44 #define NB_PULSE 4
45
46 /*
47 ********************************************************************************
48 * DECLARATION OF PROTOTYPES
49 ********************************************************************************
50 */
51 static void search_4i40(
52 Word16 dn[], /* i : correlation between target and h[] */
53 Word16 dn2[], /* i : maximum of corr. in each track. */
54 Word16 rr[][L_CODE],/* i : matrix of autocorrelation */
55 Word16 codvec[] /* o : algebraic codebook vector */
56 );
57
58 static Word16 build_code(
59 Word16 codvec[], /* i : algebraic codebook vector */
60 Word16 dn_sign[], /* i : sign of dn[] */
61 Word16 cod[], /* o : algebraic (fixed) codebook excitation */
62 Word16 h[], /* i : impulse response of weighted synthesis filter */
63 Word16 y[], /* o : filtered fixed codebook excitation */
64 Word16 sign[] /* o : index of 4 pulses (position+sign+ampl)*4 */
65 );
66
67
68 /*
69 ********************************************************************************
70 * PUBLIC PROGRAM CODE
71 ********************************************************************************
72 */
73 /*************************************************************************
74 *
75 * FUNCTION: code_4i40_17bits()
76 *
77 * PURPOSE: Searches a 17 bit algebraic codebook containing 4 pulses
78 * in a frame of 40 samples.
79 *
80 * DESCRIPTION:
81 * The code length is 40, containing 4 nonzero pulses: i0...i3.
82 * All pulses can have two possible amplitudes: +1 or -1.
83 * Pulse i0 to i2 can have 8 possible positions, pulse i3 can have
84 * 2x8=16 positions.
85 *
86 * i0 : 0, 5, 10, 15, 20, 25, 30, 35.
87 * i1 : 1, 6, 11, 16, 21, 26, 31, 36.
88 * i2 : 2, 7, 12, 17, 22, 27, 32, 37.
89 * i3 : 3, 8, 13, 18, 23, 28, 33, 38.
90 * 4, 9, 14, 19, 24, 29, 34, 39.
91 *
92 *************************************************************************/
93
94 Word16 code_4i40_17bits(
95 Word16 x[], /* i : target vector */
96 Word16 h[], /* i : impulse response of weighted synthesis filter */
97 /* h[-L_subfr..-1] must be set to zero. */
98 Word16 T0, /* i : Pitch lag */
99 Word16 pitch_sharp, /* i : Last quantized pitch gain */
100 Word16 code[], /* o : Innovative codebook */
101 Word16 y[], /* o : filtered fixed codebook excitation */
102 Word16 * sign /* o : Signs of 4 pulses */
103 )
104 {
105 Word16 codvec[NB_PULSE];
106 Word16 dn[L_CODE], dn2[L_CODE], dn_sign[L_CODE];
107 Word16 rr[L_CODE][L_CODE];
108 Word16 i, index, sharp;
109
110 sharp = shl(pitch_sharp, 1);
111 test ();
112 if (sub(T0, L_CODE) < 0)
113 {
114 for (i = T0; i < L_CODE; i++) {
115 h[i] = add(h[i], mult(h[i - T0], sharp)); move16 ();
116 }
117 }
118
119 cor_h_x(h, x, dn, 1);
120 set_sign(dn, dn_sign, dn2, 4);
121 cor_h(h, dn_sign, rr);
122 search_4i40(dn, dn2, rr, codvec);
123 move16 (); /* function result */
124 index = build_code(codvec, dn_sign, code, h, y, sign);
125
126 /*-----------------------------------------------------------------*
127 * Compute innovation vector gain. *
128 * Include fixed-gain pitch contribution into code[]. *
129 *-----------------------------------------------------------------*/
130
131 test ();
132 if (sub(T0, L_CODE) < 0)
133 {
134 for (i = T0; i < L_CODE; i++) {
135 code[i] = add(code[i], mult(code[i - T0], sharp)); move16 ();
136 }
137 }
138 return index;
139 }
140
141 /*
142 ********************************************************************************
143 * PRIVATE PROGRAM CODE
144 ********************************************************************************
145 */
146
147 /*************************************************************************
148 *
149 * FUNCTION search_4i40()
150 *
151 * PURPOSE: Search the best codevector; determine positions of the 4 pulses
152 * in the 40-sample frame.
153 *
154 *************************************************************************/
155
156 #define _1_2 (Word16)(32768L/2)
157 #define _1_4 (Word16)(32768L/4)
158 #define _1_8 (Word16)(32768L/8)
159 #define _1_16 (Word16)(32768L/16)
160
161 static void search_4i40(
162 Word16 dn[], /* i : correlation between target and h[] */
163 Word16 dn2[], /* i : maximum of corr. in each track. */
164 Word16 rr[][L_CODE], /* i : matrix of autocorrelation */
165 Word16 codvec[] /* o : algebraic codebook vector */
166 )
167 {
168 Word16 i0, i1, i2, i3;
169 Word16 ix = 0; /* initialization only needed to keep gcc silent */
170 Word16 ps = 0; /* initialization only needed to keep gcc silent */
171 Word16 i, pos, track, ipos[NB_PULSE];
172 Word16 psk, ps0, ps1, sq, sq1;
173 Word16 alpk, alp, alp_16;
174 Word32 s, alp0, alp1;
175
176 /* Default value */
177 psk = -1; move16 ();
178 alpk = 1; move16 ();
179 for (i = 0; i < NB_PULSE; i++)
180 {
181 codvec[i] = i; move16 ();
182 }
183
184 for (track = 3; track < 5; track++) {
185 /* fix starting position */
186
187 ipos[0] = 0; move16 ();
188 ipos[1] = 1; move16 ();
189 ipos[2] = 2; move16 ();
190 ipos[3] = track; move16 ();
191
192 /*------------------------------------------------------------------*
193 * main loop: try 4 tracks. *
194 *------------------------------------------------------------------*/
195
196 for (i = 0; i < NB_PULSE; i++)
197 {
198 /*----------------------------------------------------------------*
199 * i0 loop: try 4 positions (use position with max of corr.). *
200 *----------------------------------------------------------------*/
201
202 move16 (); /* account for ptr. init. (rr[io]) */
203 for (i0 = ipos[0]; i0 < L_CODE; i0 += STEP)
204 {
205 test (); move16 ();
206 if (dn2[i0] >= 0)
207 {
208 ps0 = dn[i0]; move16 ();
209 alp0 = L_mult(rr[i0][i0], _1_4);
210
211 /*----------------------------------------------------------------*
212 * i1 loop: 8 positions. *
213 *----------------------------------------------------------------*/
214
215 sq = -1; move16 ();
216 alp = 1; move16 ();
217 ps = 0; move16 ();
218 ix = ipos[1]; move16 ();
219
220 /* initialize 4 index for next loop. */
221 /*-------------------------------------------------------------------*
222 * These index have low complexity address computation because *
223 * they are, in fact, pointers with fixed increment. For example, *
224 * "rr[i0][i3]" is a pointer initialized to "&rr[i0][ipos[3]]" *
225 * and incremented by "STEP". *
226 *-------------------------------------------------------------------*/
227
228 move16 (); /* account for ptr. init. (rr[i1]) */
229 move16 (); /* account for ptr. init. (dn[i1]) */
230 move16 (); /* account for ptr. init. (rr[io]) */
231 for (i1 = ipos[1]; i1 < L_CODE; i1 += STEP)
232 {
233 ps1 = add(ps0, dn[i1]); /* idx increment = STEP */
234
235 /* alp1 = alp0 + rr[i0][i1] + 1/2*rr[i1][i1]; */
236
237 alp1 = L_mac(alp0, rr[i1][i1], _1_4); /* idx incr = STEP */
238 alp1 = L_mac(alp1, rr[i0][i1], _1_2); /* idx incr = STEP */
239
240 sq1 = mult(ps1, ps1);
241
242 alp_16 = round(alp1);
243
244 s = L_msu(L_mult(alp, sq1), sq, alp_16);
245
246 test ();
247 if (s > 0)
248 {
249 sq = sq1; move16 ();
250 ps = ps1; move16 ();
251 alp = alp_16; move16 ();
252 ix = i1; move16 ();
253 }
254 }
255 i1 = ix; move16 ();
256
257 /*----------------------------------------------------------------*
258 * i2 loop: 8 positions. *
259 *----------------------------------------------------------------*/
260
261 ps0 = ps; move16 ();
262 alp0 = L_mult(alp, _1_4);
263
264 sq = -1; move16 ();
265 alp = 1; move16 ();
266 ps = 0; move16 ();
267 ix = ipos[2]; move16 ();
268
269 /* initialize 4 index for next loop (see i1 loop) */
270
271 move16 (); /* account for ptr. init. (rr[i2]) */
272 move16 (); /* account for ptr. init. (rr[i1]) */
273 move16 (); /* account for ptr. init. (dn[i2]) */
274 move16 (); /* account for ptr. init. (rr[io]) */
275 for (i2 = ipos[2]; i2 < L_CODE; i2 += STEP)
276 {
277 ps1 = add(ps0, dn[i2]); /* index increment = STEP */
278
279 /* alp1 = alp0 + rr[i0][i2] + rr[i1][i2] + 1/2*rr[i2][i2]; */
280
281 alp1 = L_mac(alp0, rr[i2][i2], _1_16); /* idx incr = STEP */
282 alp1 = L_mac(alp1, rr[i1][i2], _1_8); /* idx incr = STEP */
283 alp1 = L_mac(alp1, rr[i0][i2], _1_8); /* idx incr = STEP */
284
285 sq1 = mult(ps1, ps1);
286
287 alp_16 = round(alp1);
288
289 s = L_msu(L_mult(alp, sq1), sq, alp_16);
290
291 test ();
292 if (s > 0)
293 {
294 sq = sq1; move16 ();
295 ps = ps1; move16 ();
296 alp = alp_16; move16 ();
297 ix = i2; move16 ();
298 }
299 }
300 i2 = ix; move16 ();
301
302 /*----------------------------------------------------------------*
303 * i3 loop: 8 positions. *
304 *----------------------------------------------------------------*/
305
306 ps0 = ps; move16 ();
307 alp0 = L_deposit_h(alp);
308
309 sq = -1; move16 ();
310 alp = 1; move16 ();
311 ps = 0; move16 ();
312 ix = ipos[3]; move16 ();
313
314 /* initialize 5 index for next loop (see i1 loop) */
315
316 move16 (); /* account for ptr. init. (rr[i3]) */
317 move16 (); /* account for ptr. init. (rr[i2]) */
318 move16 (); /* account for ptr. init. (rr[i1]) */
319 move16 (); /* account for ptr. init. (dn[i3]) */
320 move16 (); /* account for ptr. init. (rr[io]) */
321 for (i3 = ipos[3]; i3 < L_CODE; i3 += STEP)
322 {
323 ps1 = add(ps0, dn[i3]); /* index increment = STEP */
324
325 /* alp1 = alp0 + rr[i0][i3] + rr[i1][i3] + rr[i2][i3] + 1/2*rr[i3][i3]; */
326
327 alp1 = L_mac(alp0, rr[i3][i3], _1_16); /* idx incr = STEP */
328 alp1 = L_mac(alp1, rr[i2][i3], _1_8); /* idx incr = STEP */
329 alp1 = L_mac(alp1, rr[i1][i3], _1_8); /* idx incr = STEP */
330 alp1 = L_mac(alp1, rr[i0][i3], _1_8); /* idx incr = STEP */
331
332 sq1 = mult(ps1, ps1);
333
334 alp_16 = round(alp1);
335
336 s = L_msu(L_mult(alp, sq1), sq, alp_16);
337
338 test ();
339 if (s > 0)
340 {
341 sq = sq1; move16 ();
342 ps = ps1; move16 ();
343 alp = alp_16; move16 ();
344 ix = i3; move16 ();
345 }
346 }
347
348
349 /*----------------------------------------------------------------*
350 * memorise codevector if this one is better than the last one. *
351 *----------------------------------------------------------------*/
352
353 s = L_msu(L_mult(alpk, sq), psk, alp);
354
355 test ();
356 if (s > 0)
357 {
358 psk = sq; move16 ();
359 alpk = alp; move16 ();
360 codvec[0] = i0; move16 ();
361 codvec[1] = i1; move16 ();
362 codvec[2] = i2; move16 ();
363 codvec[3] = ix; move16 ();
364 }
365 }
366 }
367
368 /*----------------------------------------------------------------*
369 * Cyclic permutation of i0,i1,i2 and i3. *
370 *----------------------------------------------------------------*/
371
372 pos = ipos[3]; move16 ();
373 ipos[3] = ipos[2]; move16 ();
374 ipos[2] = ipos[1]; move16 ();
375 ipos[1] = ipos[0]; move16 ();
376 ipos[0] = pos; move16 ();
377 }
378 }
379
380 return;
381 }
382
383 /*************************************************************************
384 *
385 * FUNCTION: build_code()
386 *
387 * PURPOSE: Builds the codeword, the filtered codeword and index of the
388 * codevector, based on the signs and positions of 4 pulses.
389 *
390 *************************************************************************/
391
392 static Word16
393 build_code(
394 Word16 codvec[], /* i : position of pulses */
395 Word16 dn_sign[], /* i : sign of pulses */
396 Word16 cod[], /* o : innovative code vector */
397 Word16 h[], /* i : impulse response of weighted synthesis filter */
398 Word16 y[], /* o : filtered innovative code */
399 Word16 sign[] /* o : index of 4 pulses (sign+position) */
400 )
401 {
402 Word16 i, j, k, track, index, _sign[NB_PULSE], indx, rsign;
403 Word16 *p0, *p1, *p2, *p3;
404 Word32 s;
405
406 for (i = 0; i < L_CODE; i++)
407 {
408 cod[i] = 0; move16 ();
409 }
410
411 indx = 0; move16 ();
412 rsign = 0; move16 ();
413 for (k = 0; k < NB_PULSE; k++)
414 {
415 i = codvec[k]; move16 (); /* read pulse position */
416 j = dn_sign[i]; move16 (); /* read sign */
417
418 index = mult(i, 6554); /* index = pos/5 */
419 /* track = pos%5 */
420 track = sub(i, extract_l(L_shr(L_mult(index, 5), 1)));
421
422
423 index = gray[index]; move16 ();
424
425 test ();
426 if (sub(track, 1) == 0)
427 index = shl(index, 3);
428 else if (sub(track, 2) == 0)
429 {
430 test ();
431 index = shl(index, 6);
432 }
433 else if (sub(track, 3) == 0)
434 {
435 test ();test ();
436 index = shl(index, 10);
437 }
438 else if (sub(track, 4) == 0)
439 {
440 test ();test ();test ();
441 track = 3; move16 ();
442 index = add(shl(index, 10), 512);
443 }
444
445 test ();
446 if (j > 0)
447 {
448 cod[i] = 8191; move16 ();
449 _sign[k] = 32767; move16 ();
450 rsign = add(rsign, shl(1, track));
451 } else {
452 cod[i] = -8192; move16 ();
453 _sign[k] = (Word16) - 32768L; move16 ();
454 }
455
456 indx = add(indx, index);
457 }
458 *sign = rsign; move16 ();
459
460 p0 = h - codvec[0]; move16 ();
461 p1 = h - codvec[1]; move16 ();
462 p2 = h - codvec[2]; move16 ();
463 p3 = h - codvec[3]; move16 ();
464
465 for (i = 0; i < L_CODE; i++)
466 {
467 s = 0; move32 ();
468 s = L_mac(s, *p0++, _sign[0]);
469 s = L_mac(s, *p1++, _sign[1]);
470 s = L_mac(s, *p2++, _sign[2]);
471 s = L_mac(s, *p3++, _sign[3]);
472 y[i] = round(s); move16 ();
473 }
474
475 return indx;
476 }