#include #include #include "common.h" #include "encoder.h" #include "mem.h" #include "fft.h" #include "psycho_1.h" #include "psycho_1_priv.h" #define DBTAB 1000 double dbtable[DBTAB]; /********************************************************************** This module implements the psychoacoustic model I for the MPEG encoder layer II. It uses simplified tonal and noise masking threshold analysis to generate SMR for the encoder bit allocation routine. **********************************************************************/ void psycho_1 (short buffer[2][1152], double scale[2][SBLIMIT], double ltmin[2][SBLIMIT], frame_info * frame) { frame_header *header = frame->header; int nch = frame->nch; int sblimit = frame->sblimit; int k, i, tone = 0, noise = 0; static char init = 0; static int off[2] = { 256, 256 }; double sample[FFT_SIZE]; double spike[2][SBLIMIT]; static D1408 *fft_buf; static mask_ptr power; static g_ptr ltg; FLOAT energy[FFT_SIZE]; /* call functions for critical boundaries, freq. */ if (!init) { /* bands, bark values, and mapping */ fft_buf = (D1408 *) mem_alloc ((long) sizeof (D1408) * 2, "fft_buf"); power = (mask_ptr) mem_alloc (sizeof (mask) * HAN_SIZE, "power"); if (header->version == MPEG_AUDIO_ID) { psycho_1_read_cbound (header->lay, header->sampling_frequency); psycho_1_read_freq_band (<g, header->lay, header->sampling_frequency); } else { psycho_1_read_cbound (header->lay, header->sampling_frequency + 4); psycho_1_read_freq_band (<g, header->lay, header->sampling_frequency + 4); } psycho_1_make_map (power, ltg); for (i = 0; i < 1408; i++) fft_buf[0][i] = fft_buf[1][i] = 0; psycho_1_init_add_db (); /* create the add_db table */ init = 1; } for (k = 0; k < nch; k++) { /* check pcm input for 3 blocks of 384 samples */ /* sami's speedup, added in 02j saves about 4% overall during an encode */ int ok = off[k] % 1408; for (i = 0; i < 1152; i++) { fft_buf[k][ok++] = (double) buffer[k][i] / SCALE; if (ok >= 1408) ok = 0; } ok = (off[k] + 1216) % 1408; for (i = 0; i < FFT_SIZE; i++) { sample[i] = fft_buf[k][ok++]; if (ok >= 1408) ok = 0; } off[k] += 1152; off[k] %= 1408; psycho_1_hann_fft_pickmax (sample, power, &spike[k][0], energy); psycho_1_tonal_label (power, &tone); psycho_1_noise_label (power, &noise, ltg, energy); //psycho_1_dump(power, &tone, &noise) ; psycho_1_subsampling (power, ltg, &tone, &noise); psycho_1_threshold (power, ltg, &tone, &noise, bitrate[header->version][header->bitrate_index] / nch); psycho_1_minimum_mask (ltg, <min[k][0], sblimit); psycho_1_smr (<min[k][0], &spike[k][0], &scale[k][0], sblimit); } } int crit_band; int *cbound; int sub_size; void psycho_1_read_cbound (int lay, int freq) /* this function reads in critical band boundaries */ { #include "critband.h" //static const int FirstCriticalBand[7][27] = {... int i, k; if ((lay < 1) || (lay > 2)) { printf ("Internal error (read_cbound())\n"); return; } if ((freq < 0) || (freq > 6) || (freq == 3)) { printf ("Internal error (read_cbound())\n"); return; } crit_band = SecondCriticalBand[freq][0]; cbound = (int *) mem_alloc (sizeof (int) * crit_band, "cbound"); for (i = 0; i < crit_band; i++) { k = SecondCriticalBand[freq][i + 1]; if (k != 0) { cbound[i] = k; } else { printf ("Internal error (read_cbound())\n"); return; } } } void psycho_1_read_freq_band (ltg, lay, freq) /* this function reads in */ int lay, freq; /* frequency bands and bark */ g_ptr *ltg; /* values */ { #include "freqtable.h" int i, k; if ((freq < 0) || (freq > 6) || (freq == 3)) { printf ("Internal error (read_freq_band())\n"); return; } /* read input for freq. subbands */ sub_size = SecondFreqEntries[freq] + 1; *ltg = (g_ptr) mem_alloc (sizeof (g_thres) * sub_size, "ltg"); (*ltg)[0].line = 0; /* initialize global masking threshold */ (*ltg)[0].bark = 0.0; (*ltg)[0].hear = 0.0; for (i = 1; i < sub_size; i++) { k = SecondFreqSubband[freq][i - 1].line; if (k != 0) { (*ltg)[i].line = k; (*ltg)[i].bark = SecondFreqSubband[freq][i - 1].bark; (*ltg)[i].hear = SecondFreqSubband[freq][i - 1].hear; } else { printf ("Internal error (read_freq_band())\n"); return; } } } void psycho_1_make_map (mask power[HAN_SIZE], g_thres * ltg) /* this function calculates the global masking threshold */ { int i, j; for (i = 1; i < sub_size; i++) for (j = ltg[i - 1].line; j <= ltg[i].line; j++) power[j].map = i; } void psycho_1_init_add_db (void) { int i; double x; for (i = 0; i < DBTAB; i++) { x = (double) i / 10.0; dbtable[i] = 10 * log10 (1 + pow (10.0, x / 10.0)) - x; } } INLINE double add_db (double a, double b) { /* MFC - if the difference between a and b is large (>99), then just return the largest one. (about 10% of the time) - For differences between 0 and 99, return the largest value, but add in a pre-calculated difference value. - the value 99 was chosen arbitarily. - maximum (a-b) i've seen is 572 */ FLOAT fdiff; int idiff; fdiff = (10.0 * (a - b)); if (fdiff > 990.0) { return a; } if (fdiff < -990.0) { return (b); } idiff = (int) fdiff; if (idiff >= 0) { return (a + dbtable[idiff]); } return (b + dbtable[-idiff]); } /**************************************************************** * Window the samples then, * Fast Fourier transform of the input samples. * * ( call the FHT-based fft() in fft.c ) * * ****************************************************************/ void psycho_1_hann_fft_pickmax (double sample[FFT_SIZE], mask power[HAN_SIZE], double spike[SBLIMIT], FLOAT energy[FFT_SIZE]) { FLOAT x_real[FFT_SIZE]; register int i, j; register double sqrt_8_over_3; static int init = 0; static double *window; double sum; if (!init) { /* calculate window function for the Fourier transform */ window = (double *) mem_alloc (sizeof (DFFT), "window"); sqrt_8_over_3 = pow (8.0 / 3.0, 0.5); for (i = 0; i < FFT_SIZE; i++) { /* Hann window formula */ window[i] = sqrt_8_over_3 * 0.5 * (1 - cos (2.0 * PI * i / (FFT_SIZE))) / FFT_SIZE; } init = 1; } for (i = 0; i < FFT_SIZE; i++) x_real[i] = (FLOAT) (sample[i] * window[i]); psycho_1_fft (x_real, energy, FFT_SIZE); for (i = 0; i < HAN_SIZE; i++) { /* calculate power density spectrum */ if (energy[i] < 1E-20) power[i].x = -200.0 + POWERNORM; else power[i].x = 10 * log10 (energy[i]) + POWERNORM; power[i].next = STOP; power[i].type = FALSE; } /* Calculate the sum of spectral component in each subband from bound 4-16 */ #define CF 1073741824 /* pow(10, 0.1*POWERNORM) */ #define DBM 1E-20 /* pow(10.0, 0.1*DBMIN */ for (i = 0; i < HAN_SIZE; spike[i >> 4] = 10.0 * log10 (sum), i += 16) { for (j = 0, sum = DBM; j < 16; j++) sum += CF * energy[i + j]; } } /**************************************************************** * * This function labels the tonal component in the power * spectrum. * ****************************************************************/ void psycho_1_tonal_label (mask power[HAN_SIZE], int *tone) /* this function extracts (tonal) sinusoidals from the spectrum */ { int i, j, last = LAST, first, run, last_but_one = LAST; /* dpwe */ double max; *tone = LAST; for (i = 2; i < HAN_SIZE - 12; i++) { if (power[i].x > power[i - 1].x && power[i].x >= power[i + 1].x) { power[i].type = TONE; power[i].next = LAST; if (last != LAST) power[last].next = i; else first = *tone = i; last = i; } } last = LAST; first = *tone; *tone = LAST; while ((first != LAST) && (first != STOP)) { /* the conditions for the tonal */ if (first < 3 || first > 500) run = 0; /* otherwise k+/-j will be out of bounds */ else if (first < 63) run = 2; /* components in layer II, which */ else if (first < 127) run = 3; /* are the boundaries for calc. */ else if (first < 255) run = 6; /* the tonal components */ else run = 12; max = power[first].x - 7; /* after calculation of tonal */ for (j = 2; j <= run; j++) /* components, set to local max */ if (max < power[first - j].x || max < power[first + j].x) { power[first].type = FALSE; break; } if (power[first].type == TONE) { /* extract tonal components */ int help = first; if (*tone == LAST) *tone = first; while ((power[help].next != LAST) && (power[help].next - first) <= run) help = power[help].next; help = power[help].next; power[first].next = help; if ((first - last) <= run) { if (last_but_one != LAST) power[last_but_one].next = first; } if (first > 1 && first < 500) { /* calculate the sum of the */ double tmp; /* powers of the components */ tmp = add_db (power[first - 1].x, power[first + 1].x); power[first].x = add_db (power[first].x, tmp); } for (j = 1; j <= run; j++) { power[first - j].x = power[first + j].x = DBMIN; power[first - j].next = power[first + j].next = STOP; power[first - j].type = power[first + j].type = FALSE; } last_but_one = last; last = first; first = power[first].next; } else { int ll; if (last == LAST); /* *tone = power[first].next; dpwe */ else power[last].next = power[first].next; ll = first; first = power[first].next; power[ll].next = STOP; } } } /**************************************************************** * * This function groups all the remaining non-tonal * spectral lines into critical band where they are replaced by * one single line. * ****************************************************************/ void psycho_1_noise_label (mask * power, int *noise, g_thres * ltg, FLOAT energy[FFT_SIZE]) { int i, j, centre, last = LAST; double index, weight, sum; /* calculate the remaining spectral */ for (i = 0; i < crit_band - 1; i++) { /* lines for non-tonal components */ for (j = cbound[i], weight = 0.0, sum = DBMIN; j < cbound[i + 1]; j++) { if (power[j].type != TONE) { if (power[j].x != DBMIN) { sum = add_db (power[j].x, sum); /* Weight is used in finding the geometric mean of the noise energy within a subband */ weight += CF * energy[j] * (double) (j - cbound[i]) / (double) (cbound[i + 1] - cbound[i]); /* correction */ power[j].x = DBMIN; } } /* check to see if the spectral line is low dB, and if */ } /* so replace the center of the critical band, which is */ /* the center freq. of the noise component */ if (sum <= DBMIN) centre = (cbound[i + 1] + cbound[i]) / 2; else { /* fprintf(stderr, "%i [%f %f] -", count++,weight/pow(10.0,0.1*sum), weight*pow(10.0,-0.1*sum)); */ index = weight * pow (10.0, -0.1 * sum); centre = cbound[i] + (int) (index * (double) (cbound[i + 1] - cbound[i])); } /* locate next non-tonal component until finished; */ /* add to list of non-tonal components */ /* Masahiro Iwadare's fix for infinite looping problem? */ if (power[centre].type == TONE) { if (power[centre + 1].type == TONE) { centre++; } else centre--; } if (last == LAST) *noise = centre; else { power[centre].next = LAST; power[last].next = centre; } power[centre].x = sum; power[centre].type = NOISE; last = centre; } } /**************************************************************** * * This function reduces the number of noise and tonal * component for further threshold analysis. * ****************************************************************/ void psycho_1_subsampling (mask power[HAN_SIZE], g_thres * ltg, int *tone, int *noise) { int i, old; i = *tone; old = STOP; /* calculate tonal components for */ while ((i != LAST) && (i != STOP)) { /* reduction of spectral lines */ if (power[i].x < ltg[power[i].map].hear) { power[i].type = FALSE; power[i].x = DBMIN; if (old == STOP) *tone = power[i].next; else power[old].next = power[i].next; } else old = i; i = power[i].next; } i = *noise; old = STOP; /* calculate non-tonal components for */ while ((i != LAST) && (i != STOP)) { /* reduction of spectral lines */ if (power[i].x < ltg[power[i].map].hear) { power[i].type = FALSE; power[i].x = DBMIN; if (old == STOP) *noise = power[i].next; else power[old].next = power[i].next; } else old = i; i = power[i].next; } i = *tone; old = STOP; while ((i != LAST) && (i != STOP)) { /* if more than one */ if (power[i].next == LAST) break; /* tonal component */ if (ltg[power[power[i].next].map].bark - /* is less than .5 */ ltg[power[i].map].bark < 0.5) { /* bark, take the */ if (power[power[i].next].x > power[i].x) { /* maximum */ if (old == STOP) *tone = power[i].next; else power[old].next = power[i].next; power[i].type = FALSE; power[i].x = DBMIN; i = power[i].next; } else { power[power[i].next].type = FALSE; power[power[i].next].x = DBMIN; power[i].next = power[power[i].next].next; old = i; } } else { old = i; i = power[i].next; } } } /**************************************************************** * * This function calculates the individual threshold and * sum with the quiet threshold to find the global threshold. * ****************************************************************/ /* mainly just changed the way range checking was done MFC Nov 1999 */ void psycho_1_threshold (mask power[HAN_SIZE], g_thres * ltg, int *tone, int *noise, int bit_rate) { int k, t; double dz, tmps, vf; for (k = 1; k < sub_size; k++) { ltg[k].x = DBMIN; t = *tone; /* calculate individual masking threshold for */ while ((t != LAST) && (t != STOP)) { /* components in order to find the global */ dz = ltg[k].bark - ltg[power[t].map].bark; /* distance of bark value */ if (dz >= -3.0 && dz < 8.0) { tmps = -1.525 - 0.275 * ltg[power[t].map].bark - 4.5 + power[t].x; /* masking function for lower & upper slopes */ if (dz < -1) vf = 17 * (dz + 1) - (0.4 * power[t].x + 6); else if (dz < 0) vf = (0.4 * power[t].x + 6) * dz; else if (dz < 1) vf = (-17 * dz); else vf = -(dz - 1) * (17 - 0.15 * power[t].x) - 17; ltg[k].x = add_db (ltg[k].x, tmps + vf); } t = power[t].next; } t = *noise; /* calculate individual masking threshold */ while ((t != LAST) && (t != STOP)) { /* for non-tonal components to find LTG */ dz = ltg[k].bark - ltg[power[t].map].bark; /* distance of bark value */ if (dz >= -3.0 && dz < 8.0) { tmps = -1.525 - 0.175 * ltg[power[t].map].bark - 0.5 + power[t].x; /* masking function for lower & upper slopes */ if (dz < -1) vf = 17 * (dz + 1) - (0.4 * power[t].x + 6); else if (dz < 0) vf = (0.4 * power[t].x + 6) * dz; else if (dz < 1) vf = (-17 * dz); else vf = -(dz - 1) * (17 - 0.15 * power[t].x) - 17; ltg[k].x = add_db (ltg[k].x, tmps + vf); } t = power[t].next; } if (bit_rate < 96) ltg[k].x = add_db (ltg[k].hear, ltg[k].x); else ltg[k].x = add_db (ltg[k].hear - 12.0, ltg[k].x); } } /**************************************************************** * * This function finds the minimum masking threshold and * return the value to the encoder. * ****************************************************************/ void psycho_1_minimum_mask (g_thres * ltg, double ltmin[SBLIMIT], int sblimit) { double min; int i, j; j = 1; for (i = 0; i < sblimit; i++) if (j >= sub_size - 1) /* check subband limit, and */ ltmin[i] = ltg[sub_size - 1].hear; /* calculate the minimum masking */ else { /* level of LTMIN for each subband */ min = ltg[j].x; while (ltg[j].line >> 4 == i && j < sub_size) { if (min > ltg[j].x) min = ltg[j].x; j++; } ltmin[i] = min; } } /***************************************************************** * * This procedure is called in musicin to pick out the * smaller of the scalefactor or threshold. * *****************************************************************/ void psycho_1_smr (double ltmin[SBLIMIT], double spike[SBLIMIT], double scale[SBLIMIT], int sblimit) { int i; double max; for (i = 0; i < sblimit; i++) { /* determine the signal */ max = 20 * log10 (scale[i] * 32768) - 10; /* level for each subband */ if (spike[i] > max) max = spike[i]; /* for the maximum scale */ max -= ltmin[i]; /* factors */ ltmin[i] = max; } } void psycho_1_dump(mask power[HAN_SIZE], int *tone, int *noise) { int t; fprintf(stdout,"1 Ton: "); t=*tone; while (t!=LAST && t!=STOP) { fprintf(stdout,"[%i] %3.0f ",t, power[t].x); t = power[t].next; } fprintf(stdout,"\n"); fprintf(stdout,"1 Nos: "); t=*noise; while (t!=LAST && t!=STOP) { fprintf(stdout,"[%i] %3.0f ",t, power[t].x); t = power[t].next; } fprintf(stdout,"\n"); }