blob: f8da1e12667773cb394297bde8fba41d07959251
| 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 | |
| 22 | #include "libavutil/common.h" |
| 23 | #include "libavutil/lls.h" |
| 24 | |
| 25 | #define LPC_USE_DOUBLE |
| 26 | #include "lpc.h" |
| 27 | #include "libavutil/avassert.h" |
| 28 | |
| 29 | |
| 30 | /** |
| 31 | * Apply Welch window function to audio block |
| 32 | */ |
| 33 | static void lpc_apply_welch_window_c(const int32_t *data, int len, |
| 34 | double *w_data) |
| 35 | { |
| 36 | int i, n2; |
| 37 | double w; |
| 38 | double c; |
| 39 | |
| 40 | n2 = (len >> 1); |
| 41 | c = 2.0 / (len - 1.0); |
| 42 | |
| 43 | if (len & 1) { |
| 44 | for(i=0; i<n2; i++) { |
| 45 | w = c - i - 1.0; |
| 46 | w = 1.0 - (w * w); |
| 47 | w_data[i] = data[i] * w; |
| 48 | w_data[len-1-i] = data[len-1-i] * w; |
| 49 | } |
| 50 | return; |
| 51 | } |
| 52 | |
| 53 | w_data+=n2; |
| 54 | data+=n2; |
| 55 | for(i=0; i<n2; i++) { |
| 56 | w = c - n2 + i; |
| 57 | w = 1.0 - (w * w); |
| 58 | w_data[-i-1] = data[-i-1] * w; |
| 59 | w_data[+i ] = data[+i ] * w; |
| 60 | } |
| 61 | } |
| 62 | |
| 63 | /** |
| 64 | * Calculate autocorrelation data from audio samples |
| 65 | * A Welch window function is applied before calculation. |
| 66 | */ |
| 67 | static void lpc_compute_autocorr_c(const double *data, int len, int lag, |
| 68 | double *autoc) |
| 69 | { |
| 70 | int i, j; |
| 71 | |
| 72 | for(j=0; j<lag; j+=2){ |
| 73 | double sum0 = 1.0, sum1 = 1.0; |
| 74 | for(i=j; i<len; i++){ |
| 75 | sum0 += data[i] * data[i-j]; |
| 76 | sum1 += data[i] * data[i-j-1]; |
| 77 | } |
| 78 | autoc[j ] = sum0; |
| 79 | autoc[j+1] = sum1; |
| 80 | } |
| 81 | |
| 82 | if(j==lag){ |
| 83 | double sum = 1.0; |
| 84 | for(i=j-1; i<len; i+=2){ |
| 85 | sum += data[i ] * data[i-j ] |
| 86 | + data[i+1] * data[i-j+1]; |
| 87 | } |
| 88 | autoc[j] = sum; |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | /** |
| 93 | * Quantize LPC coefficients |
| 94 | */ |
| 95 | static void quantize_lpc_coefs(double *lpc_in, int order, int precision, |
| 96 | int32_t *lpc_out, int *shift, int min_shift, |
| 97 | int max_shift, int zero_shift) |
| 98 | { |
| 99 | int i; |
| 100 | double cmax, error; |
| 101 | int32_t qmax; |
| 102 | int sh; |
| 103 | |
| 104 | /* define maximum levels */ |
| 105 | qmax = (1 << (precision - 1)) - 1; |
| 106 | |
| 107 | /* find maximum coefficient value */ |
| 108 | cmax = 0.0; |
| 109 | for(i=0; i<order; i++) { |
| 110 | cmax= FFMAX(cmax, fabs(lpc_in[i])); |
| 111 | } |
| 112 | |
| 113 | /* if maximum value quantizes to zero, return all zeros */ |
| 114 | if(cmax * (1 << max_shift) < 1.0) { |
| 115 | *shift = zero_shift; |
| 116 | memset(lpc_out, 0, sizeof(int32_t) * order); |
| 117 | return; |
| 118 | } |
| 119 | |
| 120 | /* calculate level shift which scales max coeff to available bits */ |
| 121 | sh = max_shift; |
| 122 | while((cmax * (1 << sh) > qmax) && (sh > min_shift)) { |
| 123 | sh--; |
| 124 | } |
| 125 | |
| 126 | /* since negative shift values are unsupported in decoder, scale down |
| 127 | coefficients instead */ |
| 128 | if(sh == 0 && cmax > qmax) { |
| 129 | double scale = ((double)qmax) / cmax; |
| 130 | for(i=0; i<order; i++) { |
| 131 | lpc_in[i] *= scale; |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | /* output quantized coefficients and level shift */ |
| 136 | error=0; |
| 137 | for(i=0; i<order; i++) { |
| 138 | error -= lpc_in[i] * (1 << sh); |
| 139 | lpc_out[i] = av_clip(lrintf(error), -qmax, qmax); |
| 140 | error -= lpc_out[i]; |
| 141 | } |
| 142 | *shift = sh; |
| 143 | } |
| 144 | |
| 145 | static int estimate_best_order(double *ref, int min_order, int max_order) |
| 146 | { |
| 147 | int i, est; |
| 148 | |
| 149 | est = min_order; |
| 150 | for(i=max_order-1; i>=min_order-1; i--) { |
| 151 | if(ref[i] > 0.10) { |
| 152 | est = i+1; |
| 153 | break; |
| 154 | } |
| 155 | } |
| 156 | return est; |
| 157 | } |
| 158 | |
| 159 | int ff_lpc_calc_ref_coefs(LPCContext *s, |
| 160 | const int32_t *samples, int order, double *ref) |
| 161 | { |
| 162 | double autoc[MAX_LPC_ORDER + 1]; |
| 163 | |
| 164 | s->lpc_apply_welch_window(samples, s->blocksize, s->windowed_samples); |
| 165 | s->lpc_compute_autocorr(s->windowed_samples, s->blocksize, order, autoc); |
| 166 | compute_ref_coefs(autoc, order, ref, NULL); |
| 167 | |
| 168 | return order; |
| 169 | } |
| 170 | |
| 171 | double ff_lpc_calc_ref_coefs_f(LPCContext *s, const float *samples, int len, |
| 172 | int order, double *ref) |
| 173 | { |
| 174 | int i; |
| 175 | double signal = 0.0f, avg_err = 0.0f; |
| 176 | double autoc[MAX_LPC_ORDER+1] = {0}, error[MAX_LPC_ORDER+1] = {0}; |
| 177 | const double a = 0.5f, b = 1.0f - a; |
| 178 | |
| 179 | /* Apply windowing */ |
| 180 | for (i = 0; i <= len / 2; i++) { |
| 181 | double weight = a - b*cos((2*M_PI*i)/(len - 1)); |
| 182 | s->windowed_samples[i] = weight*samples[i]; |
| 183 | s->windowed_samples[len-1-i] = weight*samples[len-1-i]; |
| 184 | } |
| 185 | |
| 186 | s->lpc_compute_autocorr(s->windowed_samples, len, order, autoc); |
| 187 | signal = autoc[0]; |
| 188 | compute_ref_coefs(autoc, order, ref, error); |
| 189 | for (i = 0; i < order; i++) |
| 190 | avg_err = (avg_err + error[i])/2.0f; |
| 191 | return signal/avg_err; |
| 192 | } |
| 193 | |
| 194 | /** |
| 195 | * Calculate LPC coefficients for multiple orders |
| 196 | * |
| 197 | * @param lpc_type LPC method for determining coefficients, |
| 198 | * see #FFLPCType for details |
| 199 | */ |
| 200 | int ff_lpc_calc_coefs(LPCContext *s, |
| 201 | const int32_t *samples, int blocksize, int min_order, |
| 202 | int max_order, int precision, |
| 203 | int32_t coefs[][MAX_LPC_ORDER], int *shift, |
| 204 | enum FFLPCType lpc_type, int lpc_passes, |
| 205 | int omethod, int min_shift, int max_shift, int zero_shift) |
| 206 | { |
| 207 | double autoc[MAX_LPC_ORDER+1]; |
| 208 | double ref[MAX_LPC_ORDER] = { 0 }; |
| 209 | double lpc[MAX_LPC_ORDER][MAX_LPC_ORDER]; |
| 210 | int i, j, pass = 0; |
| 211 | int opt_order; |
| 212 | |
| 213 | av_assert2(max_order >= MIN_LPC_ORDER && max_order <= MAX_LPC_ORDER && |
| 214 | lpc_type > FF_LPC_TYPE_FIXED); |
| 215 | av_assert0(lpc_type == FF_LPC_TYPE_CHOLESKY || lpc_type == FF_LPC_TYPE_LEVINSON); |
| 216 | |
| 217 | /* reinit LPC context if parameters have changed */ |
| 218 | if (blocksize != s->blocksize || max_order != s->max_order || |
| 219 | lpc_type != s->lpc_type) { |
| 220 | ff_lpc_end(s); |
| 221 | ff_lpc_init(s, blocksize, max_order, lpc_type); |
| 222 | } |
| 223 | |
| 224 | if(lpc_passes <= 0) |
| 225 | lpc_passes = 2; |
| 226 | |
| 227 | if (lpc_type == FF_LPC_TYPE_LEVINSON || (lpc_type == FF_LPC_TYPE_CHOLESKY && lpc_passes > 1)) { |
| 228 | s->lpc_apply_welch_window(samples, blocksize, s->windowed_samples); |
| 229 | |
| 230 | s->lpc_compute_autocorr(s->windowed_samples, blocksize, max_order, autoc); |
| 231 | |
| 232 | compute_lpc_coefs(autoc, max_order, &lpc[0][0], MAX_LPC_ORDER, 0, 1); |
| 233 | |
| 234 | for(i=0; i<max_order; i++) |
| 235 | ref[i] = fabs(lpc[i][i]); |
| 236 | |
| 237 | pass++; |
| 238 | } |
| 239 | |
| 240 | if (lpc_type == FF_LPC_TYPE_CHOLESKY) { |
| 241 | LLSModel *m = s->lls_models; |
| 242 | LOCAL_ALIGNED(32, double, var, [FFALIGN(MAX_LPC_ORDER+1,4)]); |
| 243 | double av_uninit(weight); |
| 244 | memset(var, 0, FFALIGN(MAX_LPC_ORDER+1,4)*sizeof(*var)); |
| 245 | |
| 246 | for(j=0; j<max_order; j++) |
| 247 | m[0].coeff[max_order-1][j] = -lpc[max_order-1][j]; |
| 248 | |
| 249 | for(; pass<lpc_passes; pass++){ |
| 250 | avpriv_init_lls(&m[pass&1], max_order); |
| 251 | |
| 252 | weight=0; |
| 253 | for(i=max_order; i<blocksize; i++){ |
| 254 | for(j=0; j<=max_order; j++) |
| 255 | var[j]= samples[i-j]; |
| 256 | |
| 257 | if(pass){ |
| 258 | double eval, inv, rinv; |
| 259 | eval= m[pass&1].evaluate_lls(&m[(pass-1)&1], var+1, max_order-1); |
| 260 | eval= (512>>pass) + fabs(eval - var[0]); |
| 261 | inv = 1/eval; |
| 262 | rinv = sqrt(inv); |
| 263 | for(j=0; j<=max_order; j++) |
| 264 | var[j] *= rinv; |
| 265 | weight += inv; |
| 266 | }else |
| 267 | weight++; |
| 268 | |
| 269 | m[pass&1].update_lls(&m[pass&1], var); |
| 270 | } |
| 271 | avpriv_solve_lls(&m[pass&1], 0.001, 0); |
| 272 | } |
| 273 | |
| 274 | for(i=0; i<max_order; i++){ |
| 275 | for(j=0; j<max_order; j++) |
| 276 | lpc[i][j]=-m[(pass-1)&1].coeff[i][j]; |
| 277 | ref[i]= sqrt(m[(pass-1)&1].variance[i] / weight) * (blocksize - max_order) / 4000; |
| 278 | } |
| 279 | for(i=max_order-1; i>0; i--) |
| 280 | ref[i] = ref[i-1] - ref[i]; |
| 281 | } |
| 282 | |
| 283 | opt_order = max_order; |
| 284 | |
| 285 | if(omethod == ORDER_METHOD_EST) { |
| 286 | opt_order = estimate_best_order(ref, min_order, max_order); |
| 287 | i = opt_order-1; |
| 288 | quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i], |
| 289 | min_shift, max_shift, zero_shift); |
| 290 | } else { |
| 291 | for(i=min_order-1; i<max_order; i++) { |
| 292 | quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i], |
| 293 | min_shift, max_shift, zero_shift); |
| 294 | } |
| 295 | } |
| 296 | |
| 297 | return opt_order; |
| 298 | } |
| 299 | |
| 300 | av_cold int ff_lpc_init(LPCContext *s, int blocksize, int max_order, |
| 301 | enum FFLPCType lpc_type) |
| 302 | { |
| 303 | s->blocksize = blocksize; |
| 304 | s->max_order = max_order; |
| 305 | s->lpc_type = lpc_type; |
| 306 | |
| 307 | s->windowed_buffer = av_mallocz((blocksize + 2 + FFALIGN(max_order, 4)) * |
| 308 | sizeof(*s->windowed_samples)); |
| 309 | if (!s->windowed_buffer) |
| 310 | return AVERROR(ENOMEM); |
| 311 | s->windowed_samples = s->windowed_buffer + FFALIGN(max_order, 4); |
| 312 | |
| 313 | s->lpc_apply_welch_window = lpc_apply_welch_window_c; |
| 314 | s->lpc_compute_autocorr = lpc_compute_autocorr_c; |
| 315 | |
| 316 | if (ARCH_X86) |
| 317 | ff_lpc_init_x86(s); |
| 318 | |
| 319 | return 0; |
| 320 | } |
| 321 | |
| 322 | av_cold void ff_lpc_end(LPCContext *s) |
| 323 | { |
| 324 | av_freep(&s->windowed_buffer); |
| 325 | } |
| 326 |