blob: 15956e3f052662640ece04ad16a8194f6fa11ae7
1 | /* |
2 | * AAC Spectral Band Replication decoding functions |
3 | * Copyright (c) 2008-2009 Robert Swain ( rob opendot cl ) |
4 | * Copyright (c) 2009-2010 Alex Converse <alex.converse@gmail.com> |
5 | * |
6 | * This file is part of FFmpeg. |
7 | * |
8 | * FFmpeg is free software; you can redistribute it and/or |
9 | * modify it under the terms of the GNU Lesser General Public |
10 | * License as published by the Free Software Foundation; either |
11 | * version 2.1 of the License, or (at your option) any later version. |
12 | * |
13 | * FFmpeg is distributed in the hope that it will be useful, |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
16 | * Lesser General Public License for more details. |
17 | * |
18 | * You should have received a copy of the GNU Lesser General Public |
19 | * License along with FFmpeg; if not, write to the Free Software |
20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
21 | */ |
22 | |
23 | /** |
24 | * @file |
25 | * AAC Spectral Band Replication decoding functions |
26 | * @author Robert Swain ( rob opendot cl ) |
27 | */ |
28 | #define USE_FIXED 0 |
29 | |
30 | #include "aac.h" |
31 | #include "sbr.h" |
32 | #include "aacsbr.h" |
33 | #include "aacsbrdata.h" |
34 | #include "aacsbr_tablegen.h" |
35 | #include "fft.h" |
36 | #include "internal.h" |
37 | #include "aacps.h" |
38 | #include "sbrdsp.h" |
39 | #include "libavutil/internal.h" |
40 | #include "libavutil/libm.h" |
41 | #include "libavutil/avassert.h" |
42 | |
43 | #include <stdint.h> |
44 | #include <float.h> |
45 | #include <math.h> |
46 | |
47 | #if ARCH_MIPS |
48 | #include "mips/aacsbr_mips.h" |
49 | #endif /* ARCH_MIPS */ |
50 | |
51 | static VLC vlc_sbr[10]; |
52 | static void aacsbr_func_ptr_init(AACSBRContext *c); |
53 | |
54 | static void make_bands(int16_t* bands, int start, int stop, int num_bands) |
55 | { |
56 | int k, previous, present; |
57 | float base, prod; |
58 | |
59 | base = powf((float)stop / start, 1.0f / num_bands); |
60 | prod = start; |
61 | previous = start; |
62 | |
63 | for (k = 0; k < num_bands-1; k++) { |
64 | prod *= base; |
65 | present = lrintf(prod); |
66 | bands[k] = present - previous; |
67 | previous = present; |
68 | } |
69 | bands[num_bands-1] = stop - previous; |
70 | } |
71 | |
72 | /// Dequantization and stereo decoding (14496-3 sp04 p203) |
73 | static void sbr_dequant(SpectralBandReplication *sbr, int id_aac) |
74 | { |
75 | int k, e; |
76 | int ch; |
77 | static const double exp2_tab[2] = {1, M_SQRT2}; |
78 | if (id_aac == TYPE_CPE && sbr->bs_coupling) { |
79 | int pan_offset = sbr->data[0].bs_amp_res ? 12 : 24; |
80 | for (e = 1; e <= sbr->data[0].bs_num_env; e++) { |
81 | for (k = 0; k < sbr->n[sbr->data[0].bs_freq_res[e]]; k++) { |
82 | float temp1, temp2, fac; |
83 | if (sbr->data[0].bs_amp_res) { |
84 | temp1 = ff_exp2fi(sbr->data[0].env_facs_q[e][k] + 7); |
85 | temp2 = ff_exp2fi(pan_offset - sbr->data[1].env_facs_q[e][k]); |
86 | } |
87 | else { |
88 | temp1 = ff_exp2fi((sbr->data[0].env_facs_q[e][k]>>1) + 7) * |
89 | exp2_tab[sbr->data[0].env_facs_q[e][k] & 1]; |
90 | temp2 = ff_exp2fi((pan_offset - sbr->data[1].env_facs_q[e][k])>>1) * |
91 | exp2_tab[(pan_offset - sbr->data[1].env_facs_q[e][k]) & 1]; |
92 | } |
93 | if (temp1 > 1E20) { |
94 | av_log(NULL, AV_LOG_ERROR, "envelope scalefactor overflow in dequant\n"); |
95 | temp1 = 1; |
96 | } |
97 | fac = temp1 / (1.0f + temp2); |
98 | sbr->data[0].env_facs[e][k] = fac; |
99 | sbr->data[1].env_facs[e][k] = fac * temp2; |
100 | } |
101 | } |
102 | for (e = 1; e <= sbr->data[0].bs_num_noise; e++) { |
103 | for (k = 0; k < sbr->n_q; k++) { |
104 | float temp1 = ff_exp2fi(NOISE_FLOOR_OFFSET - sbr->data[0].noise_facs_q[e][k] + 1); |
105 | float temp2 = ff_exp2fi(12 - sbr->data[1].noise_facs_q[e][k]); |
106 | float fac; |
107 | av_assert0(temp1 <= 1E20); |
108 | fac = temp1 / (1.0f + temp2); |
109 | sbr->data[0].noise_facs[e][k] = fac; |
110 | sbr->data[1].noise_facs[e][k] = fac * temp2; |
111 | } |
112 | } |
113 | } else { // SCE or one non-coupled CPE |
114 | for (ch = 0; ch < (id_aac == TYPE_CPE) + 1; ch++) { |
115 | for (e = 1; e <= sbr->data[ch].bs_num_env; e++) |
116 | for (k = 0; k < sbr->n[sbr->data[ch].bs_freq_res[e]]; k++){ |
117 | if (sbr->data[ch].bs_amp_res) |
118 | sbr->data[ch].env_facs[e][k] = ff_exp2fi(sbr->data[ch].env_facs_q[e][k] + 6); |
119 | else |
120 | sbr->data[ch].env_facs[e][k] = ff_exp2fi((sbr->data[ch].env_facs_q[e][k]>>1) + 6) |
121 | * exp2_tab[sbr->data[ch].env_facs_q[e][k] & 1]; |
122 | if (sbr->data[ch].env_facs[e][k] > 1E20) { |
123 | av_log(NULL, AV_LOG_ERROR, "envelope scalefactor overflow in dequant\n"); |
124 | sbr->data[ch].env_facs[e][k] = 1; |
125 | } |
126 | } |
127 | |
128 | for (e = 1; e <= sbr->data[ch].bs_num_noise; e++) |
129 | for (k = 0; k < sbr->n_q; k++) |
130 | sbr->data[ch].noise_facs[e][k] = |
131 | ff_exp2fi(NOISE_FLOOR_OFFSET - sbr->data[ch].noise_facs_q[e][k]); |
132 | } |
133 | } |
134 | } |
135 | |
136 | /** High Frequency Generation (14496-3 sp04 p214+) and Inverse Filtering |
137 | * (14496-3 sp04 p214) |
138 | * Warning: This routine does not seem numerically stable. |
139 | */ |
140 | static void sbr_hf_inverse_filter(SBRDSPContext *dsp, |
141 | float (*alpha0)[2], float (*alpha1)[2], |
142 | const float X_low[32][40][2], int k0) |
143 | { |
144 | int k; |
145 | for (k = 0; k < k0; k++) { |
146 | LOCAL_ALIGNED_16(float, phi, [3], [2][2]); |
147 | float dk; |
148 | |
149 | dsp->autocorrelate(X_low[k], phi); |
150 | |
151 | dk = phi[2][1][0] * phi[1][0][0] - |
152 | (phi[1][1][0] * phi[1][1][0] + phi[1][1][1] * phi[1][1][1]) / 1.000001f; |
153 | |
154 | if (!dk) { |
155 | alpha1[k][0] = 0; |
156 | alpha1[k][1] = 0; |
157 | } else { |
158 | float temp_real, temp_im; |
159 | temp_real = phi[0][0][0] * phi[1][1][0] - |
160 | phi[0][0][1] * phi[1][1][1] - |
161 | phi[0][1][0] * phi[1][0][0]; |
162 | temp_im = phi[0][0][0] * phi[1][1][1] + |
163 | phi[0][0][1] * phi[1][1][0] - |
164 | phi[0][1][1] * phi[1][0][0]; |
165 | |
166 | alpha1[k][0] = temp_real / dk; |
167 | alpha1[k][1] = temp_im / dk; |
168 | } |
169 | |
170 | if (!phi[1][0][0]) { |
171 | alpha0[k][0] = 0; |
172 | alpha0[k][1] = 0; |
173 | } else { |
174 | float temp_real, temp_im; |
175 | temp_real = phi[0][0][0] + alpha1[k][0] * phi[1][1][0] + |
176 | alpha1[k][1] * phi[1][1][1]; |
177 | temp_im = phi[0][0][1] + alpha1[k][1] * phi[1][1][0] - |
178 | alpha1[k][0] * phi[1][1][1]; |
179 | |
180 | alpha0[k][0] = -temp_real / phi[1][0][0]; |
181 | alpha0[k][1] = -temp_im / phi[1][0][0]; |
182 | } |
183 | |
184 | if (alpha1[k][0] * alpha1[k][0] + alpha1[k][1] * alpha1[k][1] >= 16.0f || |
185 | alpha0[k][0] * alpha0[k][0] + alpha0[k][1] * alpha0[k][1] >= 16.0f) { |
186 | alpha1[k][0] = 0; |
187 | alpha1[k][1] = 0; |
188 | alpha0[k][0] = 0; |
189 | alpha0[k][1] = 0; |
190 | } |
191 | } |
192 | } |
193 | |
194 | /// Chirp Factors (14496-3 sp04 p214) |
195 | static void sbr_chirp(SpectralBandReplication *sbr, SBRData *ch_data) |
196 | { |
197 | int i; |
198 | float new_bw; |
199 | static const float bw_tab[] = { 0.0f, 0.75f, 0.9f, 0.98f }; |
200 | |
201 | for (i = 0; i < sbr->n_q; i++) { |
202 | if (ch_data->bs_invf_mode[0][i] + ch_data->bs_invf_mode[1][i] == 1) { |
203 | new_bw = 0.6f; |
204 | } else |
205 | new_bw = bw_tab[ch_data->bs_invf_mode[0][i]]; |
206 | |
207 | if (new_bw < ch_data->bw_array[i]) { |
208 | new_bw = 0.75f * new_bw + 0.25f * ch_data->bw_array[i]; |
209 | } else |
210 | new_bw = 0.90625f * new_bw + 0.09375f * ch_data->bw_array[i]; |
211 | ch_data->bw_array[i] = new_bw < 0.015625f ? 0.0f : new_bw; |
212 | } |
213 | } |
214 | |
215 | /** |
216 | * Calculation of levels of additional HF signal components (14496-3 sp04 p219) |
217 | * and Calculation of gain (14496-3 sp04 p219) |
218 | */ |
219 | static void sbr_gain_calc(AACContext *ac, SpectralBandReplication *sbr, |
220 | SBRData *ch_data, const int e_a[2]) |
221 | { |
222 | int e, k, m; |
223 | // max gain limits : -3dB, 0dB, 3dB, inf dB (limiter off) |
224 | static const float limgain[4] = { 0.70795, 1.0, 1.41254, 10000000000 }; |
225 | |
226 | for (e = 0; e < ch_data->bs_num_env; e++) { |
227 | int delta = !((e == e_a[1]) || (e == e_a[0])); |
228 | for (k = 0; k < sbr->n_lim; k++) { |
229 | float gain_boost, gain_max; |
230 | float sum[2] = { 0.0f, 0.0f }; |
231 | for (m = sbr->f_tablelim[k] - sbr->kx[1]; m < sbr->f_tablelim[k + 1] - sbr->kx[1]; m++) { |
232 | const float temp = sbr->e_origmapped[e][m] / (1.0f + sbr->q_mapped[e][m]); |
233 | sbr->q_m[e][m] = sqrtf(temp * sbr->q_mapped[e][m]); |
234 | sbr->s_m[e][m] = sqrtf(temp * ch_data->s_indexmapped[e + 1][m]); |
235 | if (!sbr->s_mapped[e][m]) { |
236 | sbr->gain[e][m] = sqrtf(sbr->e_origmapped[e][m] / |
237 | ((1.0f + sbr->e_curr[e][m]) * |
238 | (1.0f + sbr->q_mapped[e][m] * delta))); |
239 | } else { |
240 | sbr->gain[e][m] = sqrtf(sbr->e_origmapped[e][m] * sbr->q_mapped[e][m] / |
241 | ((1.0f + sbr->e_curr[e][m]) * |
242 | (1.0f + sbr->q_mapped[e][m]))); |
243 | } |
244 | } |
245 | for (m = sbr->f_tablelim[k] - sbr->kx[1]; m < sbr->f_tablelim[k + 1] - sbr->kx[1]; m++) { |
246 | sum[0] += sbr->e_origmapped[e][m]; |
247 | sum[1] += sbr->e_curr[e][m]; |
248 | } |
249 | gain_max = limgain[sbr->bs_limiter_gains] * sqrtf((FLT_EPSILON + sum[0]) / (FLT_EPSILON + sum[1])); |
250 | gain_max = FFMIN(100000.f, gain_max); |
251 | for (m = sbr->f_tablelim[k] - sbr->kx[1]; m < sbr->f_tablelim[k + 1] - sbr->kx[1]; m++) { |
252 | float q_m_max = sbr->q_m[e][m] * gain_max / sbr->gain[e][m]; |
253 | sbr->q_m[e][m] = FFMIN(sbr->q_m[e][m], q_m_max); |
254 | sbr->gain[e][m] = FFMIN(sbr->gain[e][m], gain_max); |
255 | } |
256 | sum[0] = sum[1] = 0.0f; |
257 | for (m = sbr->f_tablelim[k] - sbr->kx[1]; m < sbr->f_tablelim[k + 1] - sbr->kx[1]; m++) { |
258 | sum[0] += sbr->e_origmapped[e][m]; |
259 | sum[1] += sbr->e_curr[e][m] * sbr->gain[e][m] * sbr->gain[e][m] |
260 | + sbr->s_m[e][m] * sbr->s_m[e][m] |
261 | + (delta && !sbr->s_m[e][m]) * sbr->q_m[e][m] * sbr->q_m[e][m]; |
262 | } |
263 | gain_boost = sqrtf((FLT_EPSILON + sum[0]) / (FLT_EPSILON + sum[1])); |
264 | gain_boost = FFMIN(1.584893192f, gain_boost); |
265 | for (m = sbr->f_tablelim[k] - sbr->kx[1]; m < sbr->f_tablelim[k + 1] - sbr->kx[1]; m++) { |
266 | sbr->gain[e][m] *= gain_boost; |
267 | sbr->q_m[e][m] *= gain_boost; |
268 | sbr->s_m[e][m] *= gain_boost; |
269 | } |
270 | } |
271 | } |
272 | } |
273 | |
274 | /// Assembling HF Signals (14496-3 sp04 p220) |
275 | static void sbr_hf_assemble(float Y1[38][64][2], |
276 | const float X_high[64][40][2], |
277 | SpectralBandReplication *sbr, SBRData *ch_data, |
278 | const int e_a[2]) |
279 | { |
280 | int e, i, j, m; |
281 | const int h_SL = 4 * !sbr->bs_smoothing_mode; |
282 | const int kx = sbr->kx[1]; |
283 | const int m_max = sbr->m[1]; |
284 | static const float h_smooth[5] = { |
285 | 0.33333333333333, |
286 | 0.30150283239582, |
287 | 0.21816949906249, |
288 | 0.11516383427084, |
289 | 0.03183050093751, |
290 | }; |
291 | float (*g_temp)[48] = ch_data->g_temp, (*q_temp)[48] = ch_data->q_temp; |
292 | int indexnoise = ch_data->f_indexnoise; |
293 | int indexsine = ch_data->f_indexsine; |
294 | |
295 | if (sbr->reset) { |
296 | for (i = 0; i < h_SL; i++) { |
297 | memcpy(g_temp[i + 2*ch_data->t_env[0]], sbr->gain[0], m_max * sizeof(sbr->gain[0][0])); |
298 | memcpy(q_temp[i + 2*ch_data->t_env[0]], sbr->q_m[0], m_max * sizeof(sbr->q_m[0][0])); |
299 | } |
300 | } else if (h_SL) { |
301 | for (i = 0; i < 4; i++) { |
302 | memcpy(g_temp[i + 2 * ch_data->t_env[0]], |
303 | g_temp[i + 2 * ch_data->t_env_num_env_old], |
304 | sizeof(g_temp[0])); |
305 | memcpy(q_temp[i + 2 * ch_data->t_env[0]], |
306 | q_temp[i + 2 * ch_data->t_env_num_env_old], |
307 | sizeof(q_temp[0])); |
308 | } |
309 | } |
310 | |
311 | for (e = 0; e < ch_data->bs_num_env; e++) { |
312 | for (i = 2 * ch_data->t_env[e]; i < 2 * ch_data->t_env[e + 1]; i++) { |
313 | memcpy(g_temp[h_SL + i], sbr->gain[e], m_max * sizeof(sbr->gain[0][0])); |
314 | memcpy(q_temp[h_SL + i], sbr->q_m[e], m_max * sizeof(sbr->q_m[0][0])); |
315 | } |
316 | } |
317 | |
318 | for (e = 0; e < ch_data->bs_num_env; e++) { |
319 | for (i = 2 * ch_data->t_env[e]; i < 2 * ch_data->t_env[e + 1]; i++) { |
320 | LOCAL_ALIGNED_16(float, g_filt_tab, [48]); |
321 | LOCAL_ALIGNED_16(float, q_filt_tab, [48]); |
322 | float *g_filt, *q_filt; |
323 | |
324 | if (h_SL && e != e_a[0] && e != e_a[1]) { |
325 | g_filt = g_filt_tab; |
326 | q_filt = q_filt_tab; |
327 | for (m = 0; m < m_max; m++) { |
328 | const int idx1 = i + h_SL; |
329 | g_filt[m] = 0.0f; |
330 | q_filt[m] = 0.0f; |
331 | for (j = 0; j <= h_SL; j++) { |
332 | g_filt[m] += g_temp[idx1 - j][m] * h_smooth[j]; |
333 | q_filt[m] += q_temp[idx1 - j][m] * h_smooth[j]; |
334 | } |
335 | } |
336 | } else { |
337 | g_filt = g_temp[i + h_SL]; |
338 | q_filt = q_temp[i]; |
339 | } |
340 | |
341 | sbr->dsp.hf_g_filt(Y1[i] + kx, X_high + kx, g_filt, m_max, |
342 | i + ENVELOPE_ADJUSTMENT_OFFSET); |
343 | |
344 | if (e != e_a[0] && e != e_a[1]) { |
345 | sbr->dsp.hf_apply_noise[indexsine](Y1[i] + kx, sbr->s_m[e], |
346 | q_filt, indexnoise, |
347 | kx, m_max); |
348 | } else { |
349 | int idx = indexsine&1; |
350 | int A = (1-((indexsine+(kx & 1))&2)); |
351 | int B = (A^(-idx)) + idx; |
352 | float *out = &Y1[i][kx][idx]; |
353 | float *in = sbr->s_m[e]; |
354 | for (m = 0; m+1 < m_max; m+=2) { |
355 | out[2*m ] += in[m ] * A; |
356 | out[2*m+2] += in[m+1] * B; |
357 | } |
358 | if(m_max&1) |
359 | out[2*m ] += in[m ] * A; |
360 | } |
361 | indexnoise = (indexnoise + m_max) & 0x1ff; |
362 | indexsine = (indexsine + 1) & 3; |
363 | } |
364 | } |
365 | ch_data->f_indexnoise = indexnoise; |
366 | ch_data->f_indexsine = indexsine; |
367 | } |
368 | |
369 | #include "aacsbr_template.c" |
370 |