You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

275 lines
7.8KB

  1. /*
  2. * LPC utility code
  3. * Copyright (c) 2006 Justin Ruggles <justin.ruggles@gmail.com>
  4. *
  5. * This file is part of FFmpeg.
  6. *
  7. * FFmpeg is free software; you can redistribute it and/or
  8. * modify it under the terms of the GNU Lesser General Public
  9. * License as published by the Free Software Foundation; either
  10. * version 2.1 of the License, or (at your option) any later version.
  11. *
  12. * FFmpeg is distributed in the hope that it will be useful,
  13. * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15. * Lesser General Public License for more details.
  16. *
  17. * You should have received a copy of the GNU Lesser General Public
  18. * License along with FFmpeg; if not, write to the Free Software
  19. * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20. */
  21. #include "libavutil/lls.h"
  22. #define LPC_USE_DOUBLE
  23. #include "lpc.h"
  24. #include "libavutil/avassert.h"
  25. /**
  26. * Apply Welch window function to audio block
  27. */
  28. static void lpc_apply_welch_window_c(const int32_t *data, int len,
  29. double *w_data)
  30. {
  31. int i, n2;
  32. double w;
  33. double c;
  34. /* The optimization in commit fa4ed8c does not support odd len.
  35. * If someone wants odd len extend that change. */
  36. av_assert2(!(len & 1));
  37. n2 = (len >> 1);
  38. c = 2.0 / (len - 1.0);
  39. w_data+=n2;
  40. data+=n2;
  41. for(i=0; i<n2; i++) {
  42. w = c - n2 + i;
  43. w = 1.0 - (w * w);
  44. w_data[-i-1] = data[-i-1] * w;
  45. w_data[+i ] = data[+i ] * w;
  46. }
  47. }
  48. /**
  49. * Calculate autocorrelation data from audio samples
  50. * A Welch window function is applied before calculation.
  51. */
  52. static void lpc_compute_autocorr_c(const double *data, int len, int lag,
  53. double *autoc)
  54. {
  55. int i, j;
  56. for(j=0; j<lag; j+=2){
  57. double sum0 = 1.0, sum1 = 1.0;
  58. for(i=j; i<len; i++){
  59. sum0 += data[i] * data[i-j];
  60. sum1 += data[i] * data[i-j-1];
  61. }
  62. autoc[j ] = sum0;
  63. autoc[j+1] = sum1;
  64. }
  65. if(j==lag){
  66. double sum = 1.0;
  67. for(i=j-1; i<len; i+=2){
  68. sum += data[i ] * data[i-j ]
  69. + data[i+1] * data[i-j+1];
  70. }
  71. autoc[j] = sum;
  72. }
  73. }
  74. /**
  75. * Quantize LPC coefficients
  76. */
  77. static void quantize_lpc_coefs(double *lpc_in, int order, int precision,
  78. int32_t *lpc_out, int *shift, int max_shift, int zero_shift)
  79. {
  80. int i;
  81. double cmax, error;
  82. int32_t qmax;
  83. int sh;
  84. /* define maximum levels */
  85. qmax = (1 << (precision - 1)) - 1;
  86. /* find maximum coefficient value */
  87. cmax = 0.0;
  88. for(i=0; i<order; i++) {
  89. cmax= FFMAX(cmax, fabs(lpc_in[i]));
  90. }
  91. /* if maximum value quantizes to zero, return all zeros */
  92. if(cmax * (1 << max_shift) < 1.0) {
  93. *shift = zero_shift;
  94. memset(lpc_out, 0, sizeof(int32_t) * order);
  95. return;
  96. }
  97. /* calculate level shift which scales max coeff to available bits */
  98. sh = max_shift;
  99. while((cmax * (1 << sh) > qmax) && (sh > 0)) {
  100. sh--;
  101. }
  102. /* since negative shift values are unsupported in decoder, scale down
  103. coefficients instead */
  104. if(sh == 0 && cmax > qmax) {
  105. double scale = ((double)qmax) / cmax;
  106. for(i=0; i<order; i++) {
  107. lpc_in[i] *= scale;
  108. }
  109. }
  110. /* output quantized coefficients and level shift */
  111. error=0;
  112. for(i=0; i<order; i++) {
  113. error -= lpc_in[i] * (1 << sh);
  114. lpc_out[i] = av_clip(lrintf(error), -qmax, qmax);
  115. error -= lpc_out[i];
  116. }
  117. *shift = sh;
  118. }
  119. static int estimate_best_order(double *ref, int min_order, int max_order)
  120. {
  121. int i, est;
  122. est = min_order;
  123. for(i=max_order-1; i>=min_order-1; i--) {
  124. if(ref[i] > 0.10) {
  125. est = i+1;
  126. break;
  127. }
  128. }
  129. return est;
  130. }
  131. /**
  132. * Calculate LPC coefficients for multiple orders
  133. *
  134. * @param lpc_type LPC method for determining coefficients,
  135. * see #FFLPCType for details
  136. */
  137. int ff_lpc_calc_coefs(LPCContext *s,
  138. const int32_t *samples, int blocksize, int min_order,
  139. int max_order, int precision,
  140. int32_t coefs[][MAX_LPC_ORDER], int *shift,
  141. enum FFLPCType lpc_type, int lpc_passes,
  142. int omethod, int max_shift, int zero_shift)
  143. {
  144. double autoc[MAX_LPC_ORDER+1];
  145. double ref[MAX_LPC_ORDER];
  146. double lpc[MAX_LPC_ORDER][MAX_LPC_ORDER];
  147. int i, j, pass;
  148. int opt_order;
  149. av_assert2(max_order >= MIN_LPC_ORDER && max_order <= MAX_LPC_ORDER &&
  150. lpc_type > FF_LPC_TYPE_FIXED);
  151. /* reinit LPC context if parameters have changed */
  152. if (blocksize != s->blocksize || max_order != s->max_order ||
  153. lpc_type != s->lpc_type) {
  154. ff_lpc_end(s);
  155. ff_lpc_init(s, blocksize, max_order, lpc_type);
  156. }
  157. if (lpc_type == FF_LPC_TYPE_LEVINSON) {
  158. double *windowed_samples = s->windowed_samples + max_order;
  159. s->lpc_apply_welch_window(samples, blocksize, windowed_samples);
  160. s->lpc_compute_autocorr(windowed_samples, blocksize, max_order, autoc);
  161. compute_lpc_coefs(autoc, max_order, &lpc[0][0], MAX_LPC_ORDER, 0, 1);
  162. for(i=0; i<max_order; i++)
  163. ref[i] = fabs(lpc[i][i]);
  164. } else if (lpc_type == FF_LPC_TYPE_CHOLESKY) {
  165. LLSModel m[2];
  166. double var[MAX_LPC_ORDER+1], av_uninit(weight);
  167. if(lpc_passes <= 0)
  168. lpc_passes = 2;
  169. for(pass=0; pass<lpc_passes; pass++){
  170. av_init_lls(&m[pass&1], max_order);
  171. weight=0;
  172. for(i=max_order; i<blocksize; i++){
  173. for(j=0; j<=max_order; j++)
  174. var[j]= samples[i-j];
  175. if(pass){
  176. double eval, inv, rinv;
  177. eval= av_evaluate_lls(&m[(pass-1)&1], var+1, max_order-1);
  178. eval= (512>>pass) + fabs(eval - var[0]);
  179. inv = 1/eval;
  180. rinv = sqrt(inv);
  181. for(j=0; j<=max_order; j++)
  182. var[j] *= rinv;
  183. weight += inv;
  184. }else
  185. weight++;
  186. av_update_lls(&m[pass&1], var, 1.0);
  187. }
  188. av_solve_lls(&m[pass&1], 0.001, 0);
  189. }
  190. for(i=0; i<max_order; i++){
  191. for(j=0; j<max_order; j++)
  192. lpc[i][j]=-m[(pass-1)&1].coeff[i][j];
  193. ref[i]= sqrt(m[(pass-1)&1].variance[i] / weight) * (blocksize - max_order) / 4000;
  194. }
  195. for(i=max_order-1; i>0; i--)
  196. ref[i] = ref[i-1] - ref[i];
  197. }
  198. opt_order = max_order;
  199. if(omethod == ORDER_METHOD_EST) {
  200. opt_order = estimate_best_order(ref, min_order, max_order);
  201. i = opt_order-1;
  202. quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i], max_shift, zero_shift);
  203. } else {
  204. for(i=min_order-1; i<max_order; i++) {
  205. quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i], max_shift, zero_shift);
  206. }
  207. }
  208. return opt_order;
  209. }
  210. av_cold int ff_lpc_init(LPCContext *s, int blocksize, int max_order,
  211. enum FFLPCType lpc_type)
  212. {
  213. s->blocksize = blocksize;
  214. s->max_order = max_order;
  215. s->lpc_type = lpc_type;
  216. if (lpc_type == FF_LPC_TYPE_LEVINSON) {
  217. s->windowed_samples = av_mallocz((blocksize + max_order + 2) *
  218. sizeof(*s->windowed_samples));
  219. if (!s->windowed_samples)
  220. return AVERROR(ENOMEM);
  221. } else {
  222. s->windowed_samples = NULL;
  223. }
  224. s->lpc_apply_welch_window = lpc_apply_welch_window_c;
  225. s->lpc_compute_autocorr = lpc_compute_autocorr_c;
  226. if (HAVE_MMX)
  227. ff_lpc_init_x86(s);
  228. return 0;
  229. }
  230. av_cold void ff_lpc_end(LPCContext *s)
  231. {
  232. av_freep(&s->windowed_samples);
  233. }