diff options
Diffstat (limited to 'lib/rbcodec/codecs/libopus/celt/vq.c')
-rw-r--r-- | lib/rbcodec/codecs/libopus/celt/vq.c | 130 |
1 files changed, 78 insertions, 52 deletions
diff --git a/lib/rbcodec/codecs/libopus/celt/vq.c b/lib/rbcodec/codecs/libopus/celt/vq.c index b047b22774..a6b5552d69 100644 --- a/lib/rbcodec/codecs/libopus/celt/vq.c +++ b/lib/rbcodec/codecs/libopus/celt/vq.c | |||
@@ -39,10 +39,6 @@ | |||
39 | #include "rate.h" | 39 | #include "rate.h" |
40 | #include "pitch.h" | 40 | #include "pitch.h" |
41 | 41 | ||
42 | #if defined(MIPSr1_ASM) | ||
43 | #include "mips/vq_mipsr1.h" | ||
44 | #endif | ||
45 | |||
46 | #ifndef OVERRIDE_vq_exp_rotation1 | 42 | #ifndef OVERRIDE_vq_exp_rotation1 |
47 | static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_val16 s) | 43 | static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_val16 s) |
48 | { | 44 | { |
@@ -71,7 +67,7 @@ static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_ | |||
71 | } | 67 | } |
72 | #endif /* OVERRIDE_vq_exp_rotation1 */ | 68 | #endif /* OVERRIDE_vq_exp_rotation1 */ |
73 | 69 | ||
74 | static void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread) | 70 | void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread) |
75 | { | 71 | { |
76 | static const int SPREAD_FACTOR[3]={15,10,5}; | 72 | static const int SPREAD_FACTOR[3]={15,10,5}; |
77 | int i; | 73 | int i; |
@@ -162,42 +158,27 @@ static unsigned extract_collapse_mask(int *iy, int N, int B) | |||
162 | return collapse_mask; | 158 | return collapse_mask; |
163 | } | 159 | } |
164 | 160 | ||
165 | unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc | 161 | opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch) |
166 | #ifdef RESYNTH | ||
167 | , opus_val16 gain | ||
168 | #endif | ||
169 | ) | ||
170 | { | 162 | { |
171 | VARDECL(celt_norm, y); | 163 | VARDECL(celt_norm, y); |
172 | VARDECL(int, iy); | 164 | VARDECL(int, signx); |
173 | VARDECL(opus_val16, signx); | ||
174 | int i, j; | 165 | int i, j; |
175 | opus_val16 s; | ||
176 | int pulsesLeft; | 166 | int pulsesLeft; |
177 | opus_val32 sum; | 167 | opus_val32 sum; |
178 | opus_val32 xy; | 168 | opus_val32 xy; |
179 | opus_val16 yy; | 169 | opus_val16 yy; |
180 | unsigned collapse_mask; | ||
181 | SAVE_STACK; | 170 | SAVE_STACK; |
182 | 171 | ||
183 | celt_assert2(K>0, "alg_quant() needs at least one pulse"); | 172 | (void)arch; |
184 | celt_assert2(N>1, "alg_quant() needs at least two dimensions"); | ||
185 | |||
186 | ALLOC(y, N, celt_norm); | 173 | ALLOC(y, N, celt_norm); |
187 | ALLOC(iy, N, int); | 174 | ALLOC(signx, N, int); |
188 | ALLOC(signx, N, opus_val16); | ||
189 | |||
190 | exp_rotation(X, N, 1, B, K, spread); | ||
191 | 175 | ||
192 | /* Get rid of the sign */ | 176 | /* Get rid of the sign */ |
193 | sum = 0; | 177 | sum = 0; |
194 | j=0; do { | 178 | j=0; do { |
195 | if (X[j]>0) | 179 | signx[j] = X[j]<0; |
196 | signx[j]=1; | 180 | /* OPT: Make sure the compiler doesn't use a branch on ABS16(). */ |
197 | else { | 181 | X[j] = ABS16(X[j]); |
198 | signx[j]=-1; | ||
199 | X[j]=-X[j]; | ||
200 | } | ||
201 | iy[j] = 0; | 182 | iy[j] = 0; |
202 | y[j] = 0; | 183 | y[j] = 0; |
203 | } while (++j<N); | 184 | } while (++j<N); |
@@ -229,7 +210,12 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc | |||
229 | while (++j<N); | 210 | while (++j<N); |
230 | sum = QCONST16(1.f,14); | 211 | sum = QCONST16(1.f,14); |
231 | } | 212 | } |
232 | rcp = EXTRACT16(MULT16_32_Q16(K-1, celt_rcp(sum))); | 213 | #ifdef FIXED_POINT |
214 | rcp = EXTRACT16(MULT16_32_Q16(K, celt_rcp(sum))); | ||
215 | #else | ||
216 | /* Using K+e with e < 1 guarantees we cannot get more than K pulses. */ | ||
217 | rcp = EXTRACT16(MULT16_32_Q16(K+0.8f, celt_rcp(sum))); | ||
218 | #endif | ||
233 | j=0; do { | 219 | j=0; do { |
234 | #ifdef FIXED_POINT | 220 | #ifdef FIXED_POINT |
235 | /* It's really important to round *towards zero* here */ | 221 | /* It's really important to round *towards zero* here */ |
@@ -244,12 +230,12 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc | |||
244 | pulsesLeft -= iy[j]; | 230 | pulsesLeft -= iy[j]; |
245 | } while (++j<N); | 231 | } while (++j<N); |
246 | } | 232 | } |
247 | celt_assert2(pulsesLeft>=1, "Allocated too many pulses in the quick pass"); | 233 | celt_sig_assert(pulsesLeft>=0); |
248 | 234 | ||
249 | /* This should never happen, but just in case it does (e.g. on silence) | 235 | /* This should never happen, but just in case it does (e.g. on silence) |
250 | we fill the first bin with pulses. */ | 236 | we fill the first bin with pulses. */ |
251 | #ifdef FIXED_POINT_DEBUG | 237 | #ifdef FIXED_POINT_DEBUG |
252 | celt_assert2(pulsesLeft<=N+3, "Not enough pulses in the quick pass"); | 238 | celt_sig_assert(pulsesLeft<=N+3); |
253 | #endif | 239 | #endif |
254 | if (pulsesLeft > N+3) | 240 | if (pulsesLeft > N+3) |
255 | { | 241 | { |
@@ -260,12 +246,12 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc | |||
260 | pulsesLeft=0; | 246 | pulsesLeft=0; |
261 | } | 247 | } |
262 | 248 | ||
263 | s = 1; | ||
264 | for (i=0;i<pulsesLeft;i++) | 249 | for (i=0;i<pulsesLeft;i++) |
265 | { | 250 | { |
251 | opus_val16 Rxy, Ryy; | ||
266 | int best_id; | 252 | int best_id; |
267 | opus_val32 best_num = -VERY_LARGE16; | 253 | opus_val32 best_num; |
268 | opus_val16 best_den = 0; | 254 | opus_val16 best_den; |
269 | #ifdef FIXED_POINT | 255 | #ifdef FIXED_POINT |
270 | int rshift; | 256 | int rshift; |
271 | #endif | 257 | #endif |
@@ -275,10 +261,23 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc | |||
275 | best_id = 0; | 261 | best_id = 0; |
276 | /* The squared magnitude term gets added anyway, so we might as well | 262 | /* The squared magnitude term gets added anyway, so we might as well |
277 | add it outside the loop */ | 263 | add it outside the loop */ |
278 | yy = ADD32(yy, 1); | 264 | yy = ADD16(yy, 1); |
279 | j=0; | 265 | |
266 | /* Calculations for position 0 are out of the loop, in part to reduce | ||
267 | mispredicted branches (since the if condition is usually false) | ||
268 | in the loop. */ | ||
269 | /* Temporary sums of the new pulse(s) */ | ||
270 | Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[0])),rshift)); | ||
271 | /* We're multiplying y[j] by two so we don't have to do it here */ | ||
272 | Ryy = ADD16(yy, y[0]); | ||
273 | |||
274 | /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that | ||
275 | Rxy is positive because the sign is pre-computed) */ | ||
276 | Rxy = MULT16_16_Q15(Rxy,Rxy); | ||
277 | best_den = Ryy; | ||
278 | best_num = Rxy; | ||
279 | j=1; | ||
280 | do { | 280 | do { |
281 | opus_val16 Rxy, Ryy; | ||
282 | /* Temporary sums of the new pulse(s) */ | 281 | /* Temporary sums of the new pulse(s) */ |
283 | Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[j])),rshift)); | 282 | Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[j])),rshift)); |
284 | /* We're multiplying y[j] by two so we don't have to do it here */ | 283 | /* We're multiplying y[j] by two so we don't have to do it here */ |
@@ -289,8 +288,11 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc | |||
289 | Rxy = MULT16_16_Q15(Rxy,Rxy); | 288 | Rxy = MULT16_16_Q15(Rxy,Rxy); |
290 | /* The idea is to check for num/den >= best_num/best_den, but that way | 289 | /* The idea is to check for num/den >= best_num/best_den, but that way |
291 | we can do it without any division */ | 290 | we can do it without any division */ |
292 | /* OPT: Make sure to use conditional moves here */ | 291 | /* OPT: It's not clear whether a cmov is faster than a branch here |
293 | if (MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num)) | 292 | since the condition is more often false than true and using |
293 | a cmov introduces data dependencies across iterations. The optimal | ||
294 | choice may be architecture-dependent. */ | ||
295 | if (opus_unlikely(MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num))) | ||
294 | { | 296 | { |
295 | best_den = Ryy; | 297 | best_den = Ryy; |
296 | best_num = Rxy; | 298 | best_num = Rxy; |
@@ -305,23 +307,47 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc | |||
305 | 307 | ||
306 | /* Only now that we've made the final choice, update y/iy */ | 308 | /* Only now that we've made the final choice, update y/iy */ |
307 | /* Multiplying y[j] by 2 so we don't have to do it everywhere else */ | 309 | /* Multiplying y[j] by 2 so we don't have to do it everywhere else */ |
308 | y[best_id] += 2*s; | 310 | y[best_id] += 2; |
309 | iy[best_id]++; | 311 | iy[best_id]++; |
310 | } | 312 | } |
311 | 313 | ||
312 | /* Put the original sign back */ | 314 | /* Put the original sign back */ |
313 | j=0; | 315 | j=0; |
314 | do { | 316 | do { |
315 | X[j] = MULT16_16(signx[j],X[j]); | 317 | /*iy[j] = signx[j] ? -iy[j] : iy[j];*/ |
316 | if (signx[j] < 0) | 318 | /* OPT: The is more likely to be compiled without a branch than the code above |
317 | iy[j] = -iy[j]; | 319 | but has the same performance otherwise. */ |
320 | iy[j] = (iy[j]^-signx[j]) + signx[j]; | ||
318 | } while (++j<N); | 321 | } while (++j<N); |
322 | RESTORE_STACK; | ||
323 | return yy; | ||
324 | } | ||
325 | |||
326 | unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc, | ||
327 | opus_val16 gain, int resynth, int arch) | ||
328 | { | ||
329 | VARDECL(int, iy); | ||
330 | opus_val16 yy; | ||
331 | unsigned collapse_mask; | ||
332 | SAVE_STACK; | ||
333 | |||
334 | celt_assert2(K>0, "alg_quant() needs at least one pulse"); | ||
335 | celt_assert2(N>1, "alg_quant() needs at least two dimensions"); | ||
336 | |||
337 | /* Covers vectorization by up to 4. */ | ||
338 | ALLOC(iy, N+3, int); | ||
339 | |||
340 | exp_rotation(X, N, 1, B, K, spread); | ||
341 | |||
342 | yy = op_pvq_search(X, iy, K, N, arch); | ||
343 | |||
319 | encode_pulses(iy, N, K, enc); | 344 | encode_pulses(iy, N, K, enc); |
320 | 345 | ||
321 | #ifdef RESYNTH | 346 | if (resynth) |
322 | normalise_residual(iy, X, N, yy, gain); | 347 | { |
323 | exp_rotation(X, N, -1, B, K, spread); | 348 | normalise_residual(iy, X, N, yy, gain); |
324 | #endif | 349 | exp_rotation(X, N, -1, B, K, spread); |
350 | } | ||
325 | 351 | ||
326 | collapse_mask = extract_collapse_mask(iy, N, B); | 352 | collapse_mask = extract_collapse_mask(iy, N, B); |
327 | RESTORE_STACK; | 353 | RESTORE_STACK; |
@@ -350,7 +376,7 @@ unsigned alg_unquant(celt_norm *X, int N, int K, int spread, int B, | |||
350 | } | 376 | } |
351 | 377 | ||
352 | #ifndef OVERRIDE_renormalise_vector | 378 | #ifndef OVERRIDE_renormalise_vector |
353 | void renormalise_vector(celt_norm *X, int N, opus_val16 gain) | 379 | void renormalise_vector(celt_norm *X, int N, opus_val16 gain, int arch) |
354 | { | 380 | { |
355 | int i; | 381 | int i; |
356 | #ifdef FIXED_POINT | 382 | #ifdef FIXED_POINT |
@@ -360,7 +386,7 @@ void renormalise_vector(celt_norm *X, int N, opus_val16 gain) | |||
360 | opus_val16 g; | 386 | opus_val16 g; |
361 | opus_val32 t; | 387 | opus_val32 t; |
362 | celt_norm *xptr; | 388 | celt_norm *xptr; |
363 | E = EPSILON + celt_inner_prod(X, X, N); | 389 | E = EPSILON + celt_inner_prod(X, X, N, arch); |
364 | #ifdef FIXED_POINT | 390 | #ifdef FIXED_POINT |
365 | k = celt_ilog2(E)>>1; | 391 | k = celt_ilog2(E)>>1; |
366 | #endif | 392 | #endif |
@@ -377,7 +403,7 @@ void renormalise_vector(celt_norm *X, int N, opus_val16 gain) | |||
377 | } | 403 | } |
378 | #endif /* OVERRIDE_renormalise_vector */ | 404 | #endif /* OVERRIDE_renormalise_vector */ |
379 | 405 | ||
380 | int stereo_itheta(const celt_norm *X, const celt_norm *Y, int stereo, int N) | 406 | int stereo_itheta(const celt_norm *X, const celt_norm *Y, int stereo, int N, int arch) |
381 | { | 407 | { |
382 | int i; | 408 | int i; |
383 | int itheta; | 409 | int itheta; |
@@ -396,8 +422,8 @@ int stereo_itheta(const celt_norm *X, const celt_norm *Y, int stereo, int N) | |||
396 | Eside = MAC16_16(Eside, s, s); | 422 | Eside = MAC16_16(Eside, s, s); |
397 | } | 423 | } |
398 | } else { | 424 | } else { |
399 | Emid += celt_inner_prod(X, X, N); | 425 | Emid += celt_inner_prod(X, X, N, arch); |
400 | Eside += celt_inner_prod(Y, Y, N); | 426 | Eside += celt_inner_prod(Y, Y, N, arch); |
401 | } | 427 | } |
402 | mid = celt_sqrt(Emid); | 428 | mid = celt_sqrt(Emid); |
403 | side = celt_sqrt(Eside); | 429 | side = celt_sqrt(Eside); |
@@ -405,7 +431,7 @@ int stereo_itheta(const celt_norm *X, const celt_norm *Y, int stereo, int N) | |||
405 | /* 0.63662 = 2/pi */ | 431 | /* 0.63662 = 2/pi */ |
406 | itheta = MULT16_16_Q15(QCONST16(0.63662f,15),celt_atan2p(side, mid)); | 432 | itheta = MULT16_16_Q15(QCONST16(0.63662f,15),celt_atan2p(side, mid)); |
407 | #else | 433 | #else |
408 | itheta = (int)floor(.5f+16384*0.63662f*atan2(side,mid)); | 434 | itheta = (int)floor(.5f+16384*0.63662f*fast_atan2f(side,mid)); |
409 | #endif | 435 | #endif |
410 | 436 | ||
411 | return itheta; | 437 | return itheta; |