lpc.c

Go to the documentation of this file.

/*
 * LPC utility code
 * Copyright (c) 2006  Justin Ruggles <justin.ruggles@gmail.com>
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * FFmpeg is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */
 
#include "libavutil/common.h"
#include "libavutil/lls.h"
#include "libavutil/mem.h"
#include "libavutil/mem_internal.h"
 
#define LPC_USE_DOUBLE
#include "lpc.h"
#include "lpc_functions.h"
#include "libavutil/avassert.h"
 
/**
 * Schur recursion.
 * Produces reflection coefficients from autocorrelation data.
 */
static inline void compute_ref_coefs(const LPC_TYPE *autoc, int max_order,
                                     LPC_TYPE *ref, LPC_TYPE *error)
{
    LPC_TYPE err;
    LPC_TYPE gen0[MAX_LPC_ORDER], gen1[MAX_LPC_ORDER];
 
    for (int i = 0; i < max_order; i++)
        gen0[i] = gen1[i] = autoc[i + 1];
 
    err    = autoc[0];
    ref[0] = -gen1[0] / ((LPC_USE_FIXED || err) ? err : 1);
    err   +=  gen1[0] * ref[0];
    if (error)
        error[0] = err;
    for (int i = 1; i < max_order; i++) {
        for (int j = 0; j < max_order - i; j++) {
            gen1[j] = gen1[j + 1] + ref[i - 1] * gen0[j];
            gen0[j] = gen1[j + 1] * ref[i - 1] + gen0[j];
        }
        ref[i] = -gen1[0] / ((LPC_USE_FIXED || err) ? err : 1);
        err   +=  gen1[0] * ref[i];
        if (error)
            error[i] = err;
    }
}
 
 
/**
 * Apply Welch window function to audio block
 */
static void lpc_apply_welch_window_c(const int32_t *data, ptrdiff_t len,
                                     double *w_data)
{
    int i, n2;
    double w;
    double c;
 
    if (len == 1) {
        w_data[0] = 0.0;
        return;
    }
 
    n2 = (len >> 1);
    c = 2.0 / (len - 1.0);
 
    if (len & 1) {
        for(i=0; i<n2; i++) {
            w = c - i - 1.0;
            w = 1.0 - (w * w);
            w_data[i] = data[i] * w;
            w_data[len-1-i] = data[len-1-i] * w;
        }
        w_data[n2] = 0.0;
        return;
    }
 
    w_data+=n2;
      data+=n2;
    for(i=0; i<n2; i++) {
        w = c - n2 + i;
        w = 1.0 - (w * w);
        w_data[-i-1] = data[-i-1] * w;
        w_data[+i  ] = data[+i  ] * w;
    }
}
 
/**
 * Calculate autocorrelation data from audio samples
 * A Welch window function is applied before calculation.
 */
static void lpc_compute_autocorr_c(const double *data, ptrdiff_t len, int lag,
                                   double *autoc)
{
    int i, j;
 
    for(j=0; j<lag; j+=2){
        double sum0 = 1.0, sum1 = 1.0;
        for(i=j; i<len; i++){
            sum0 += data[i] * data[i-j];
            sum1 += data[i] * data[i-j-1];
        }
        autoc[j  ] = sum0;
        autoc[j+1] = sum1;
    }
 
    if(j==lag){
        double sum = 1.0;
        for(i=j-1; i<len; i++){
            sum += data[i] * data[i-j];
        }
        autoc[j] = sum;
    }
}
 
/**
 * Quantize LPC coefficients
 */
static void quantize_lpc_coefs(double *lpc_in, int order, int precision,
                               int32_t *lpc_out, int *shift, int min_shift,
                               int max_shift, int zero_shift)
{
    int i;
    double cmax, error;
    int32_t qmax;
    int sh;
 
    /* define maximum levels */
    qmax = (1 << (precision - 1)) - 1;
 
    /* find maximum coefficient value */
    cmax = 0.0;
    for(i=0; i<order; i++) {
        cmax= FFMAX(cmax, fabs(lpc_in[i]));
    }
 
    /* if maximum value quantizes to zero, return all zeros */
    if(cmax * (1 << max_shift) < 1.0) {
        *shift = zero_shift;
        memset(lpc_out, 0, sizeof(int32_t) * order);
        return;
    }
 
    /* calculate level shift which scales max coeff to available bits */
    sh = max_shift;
    while((cmax * (1 << sh) > qmax) && (sh > min_shift)) {
        sh--;
    }
 
    /* since negative shift values are unsupported in decoder, scale down
       coefficients instead */
    if(sh == 0 && cmax > qmax) {
        double scale = ((double)qmax) / cmax;
        for(i=0; i<order; i++) {
            lpc_in[i] *= scale;
        }
    }
 
    /* output quantized coefficients and level shift */
    error=0;
    for(i=0; i<order; i++) {
        error -= lpc_in[i] * (1 << sh);
        lpc_out[i] = av_clip(lrintf(error), -qmax, qmax);
        error -= lpc_out[i];
    }
    *shift = sh;
}
 
static int estimate_best_order(double *ref, int min_order, int max_order)
{
    int i, est;
 
    est = min_order;
    for(i=max_order-1; i>=min_order-1; i--) {
        if(ref[i] > 0.10) {
            est = i+1;
            break;
        }
    }
    return est;
}
 
int ff_lpc_calc_ref_coefs(LPCContext *s,
                          const int32_t *samples, int order, double *ref)
{
    double autoc[MAX_LPC_ORDER + 1];
 
    s->lpc_apply_welch_window(samples, s->blocksize, s->windowed_samples);
    s->lpc_compute_autocorr(s->windowed_samples, s->blocksize, order, autoc);
    compute_ref_coefs(autoc, order, ref, NULL);
 
    return order;
}
 
double ff_lpc_calc_ref_coefs_f(LPCContext *s, const float *samples, int len,
                               int order, double *ref, int apply_window)
{
    int i;
    double signal = 0.0f, avg_err = 0.0f;
    double autoc[MAX_LPC_ORDER+1] = {0}, error[MAX_LPC_ORDER+1] = {0};
    const double a = 0.5f, b = 1.0f - a;
 
    /* Apply windowing. apply_window == 0 uses a rectangular (unity) window: a Hann
     * taper zeros the edges, which over a very short region (e.g. a short-block TNS
     * region of a few dozen lines) discards most of the data and wrecks the fit. */
    for (i = 0; i <= len / 2; i++) {
        double weight = apply_window ? a - b*cos((2*M_PI*i)/(len - 1)) : 1.0;
        s->windowed_samples[i] = weight*samples[i];
        s->windowed_samples[len-1-i] = weight*samples[len-1-i];
    }
 
    s->lpc_compute_autocorr(s->windowed_samples, len, order, autoc);
    signal = autoc[0];
    compute_ref_coefs(autoc, order, ref, error);
    for (i = 0; i < order; i++)
        avg_err = (avg_err + error[i])/2.0f;
    return avg_err ? signal/avg_err : NAN;
}
 
/**
 * Calculate LPC coefficients for multiple orders
 *
 * @param lpc_type LPC method for determining coefficients,
 *                 see #FFLPCType for details
 */
int ff_lpc_calc_coefs(LPCContext *s,
                      const int32_t *samples, int blocksize, int min_order,
                      int max_order, int precision,
                      int32_t coefs[][MAX_LPC_ORDER], int *shift,
                      enum FFLPCType lpc_type, int lpc_passes,
                      int omethod, int min_shift, int max_shift, int zero_shift)
{
    double autoc[MAX_LPC_ORDER+1];
    double ref[MAX_LPC_ORDER] = { 0 };
    double lpc[MAX_LPC_ORDER][MAX_LPC_ORDER];
    int i, j, pass = 0;
    int opt_order;
 
    av_assert2(max_order >= MIN_LPC_ORDER && max_order <= MAX_LPC_ORDER &&
           lpc_type > FF_LPC_TYPE_FIXED);
    av_assert0(lpc_type == FF_LPC_TYPE_CHOLESKY || lpc_type == FF_LPC_TYPE_LEVINSON);
 
    /* reinit LPC context if parameters have changed */
    if (blocksize != s->blocksize || max_order != s->max_order ||
        lpc_type  != s->lpc_type) {
        ff_lpc_end(s);
        ff_lpc_init(s, blocksize, max_order, lpc_type);
    }
 
    if(lpc_passes <= 0)
        lpc_passes = 2;
 
    if (lpc_type == FF_LPC_TYPE_LEVINSON || (lpc_type == FF_LPC_TYPE_CHOLESKY && lpc_passes > 1)) {
        s->lpc_apply_welch_window(samples, blocksize, s->windowed_samples);
 
        s->lpc_compute_autocorr(s->windowed_samples, blocksize, max_order, autoc);
 
        compute_lpc_coefs(autoc, 0, max_order, &lpc[0][0], MAX_LPC_ORDER, 0, 1, NULL);
 
        for(i=0; i<max_order; i++)
            ref[i] = fabs(lpc[i][i]);
 
        pass++;
    }
 
    if (lpc_type == FF_LPC_TYPE_CHOLESKY) {
        LLSModel *m = s->lls_models;
        LOCAL_ALIGNED(32, double, var, [FFALIGN(MAX_LPC_ORDER+1,4)]);
        double av_uninit(weight);
        memset(var, 0, FFALIGN(MAX_LPC_ORDER+1,4)*sizeof(*var));
 
        /* Avoids initializing with an unused value when lpc_passes == 1 */
        if (lpc_passes > 1)
            for(j=0; j<max_order; j++)
                m[0].coeff[max_order-1][j] = -lpc[max_order-1][j];
 
        for(; pass<lpc_passes; pass++){
            avpriv_init_lls(&m[pass&1], max_order);
 
            weight=0;
            for(i=max_order; i<blocksize; i++){
                for(j=0; j<=max_order; j++)
                    var[j]= samples[i-j];
 
                if(pass){
                    double eval, inv, rinv;
                    eval= m[pass&1].evaluate_lls(&m[(pass-1)&1], var+1, max_order-1);
                    eval= (512>>pass) + fabs(eval - var[0]);
                    inv = 1/eval;
                    rinv = sqrt(inv);
                    for(j=0; j<=max_order; j++)
                        var[j] *= rinv;
                    weight += inv;
                }else
                    weight++;
 
                m[pass&1].update_lls(&m[pass&1], var);
            }
            avpriv_solve_lls(&m[pass&1], 0.001, 0);
        }
 
        for(i=0; i<max_order; i++){
            for(j=0; j<max_order; j++)
                lpc[i][j]=-m[(pass-1)&1].coeff[i][j];
            ref[i]= sqrt(m[(pass-1)&1].variance[i] / weight) * (blocksize - max_order) / 4000;
        }
        for(i=max_order-1; i>0; i--)
            ref[i] = ref[i-1] - ref[i];
    }
 
    opt_order = max_order;
 
    if(omethod == ORDER_METHOD_EST) {
        opt_order = estimate_best_order(ref, min_order, max_order);
        i = opt_order-1;
        quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i],
                           min_shift, max_shift, zero_shift);
    } else {
        for(i=min_order-1; i<max_order; i++) {
            quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i],
                               min_shift, max_shift, zero_shift);
        }
    }
 
    return opt_order;
}
 
av_cold int ff_lpc_init(LPCContext *s, int blocksize, int max_order,
                        enum FFLPCType lpc_type)
{
    s->blocksize = blocksize;
    s->max_order = max_order;
    s->lpc_type  = lpc_type;
 
    s->windowed_buffer = av_mallocz((blocksize + 2 + FFALIGN(max_order, 4)) *
                                    sizeof(*s->windowed_samples));
    if (!s->windowed_buffer)
        return AVERROR(ENOMEM);
    s->windowed_samples = s->windowed_buffer + FFALIGN(max_order, 4);
 
    s->lpc_apply_welch_window = lpc_apply_welch_window_c;
    s->lpc_compute_autocorr   = lpc_compute_autocorr_c;
 
#if ARCH_RISCV
    ff_lpc_init_riscv(s);
#elif ARCH_X86
    ff_lpc_init_x86(s);
#endif
 
    return 0;
}
 
av_cold void ff_lpc_end(LPCContext *s)
{
    av_freep(&s->windowed_buffer);
}