opus/dnn/dred_encoder.c

/* Copyright (c) 2022 Amazon
   Written by Jan Buethe */
/*
   Redistribution and use in source and binary forms, with or without
   modification, are permitted provided that the following conditions
   are met:

   - Redistributions of source code must retain the above copyright
   notice, this list of conditions and the following disclaimer.

   - Redistributions in binary form must reproduce the above copyright
   notice, this list of conditions and the following disclaimer in the
   documentation and/or other materials provided with the distribution.

   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#include <string.h>

#if 0
#include <stdio.h>
#include <math.h>
#endif

#include "dred_encoder.h"
#include "dred_coding.h"
#include "celt/entenc.h"

#include "dred_decoder.h"
#include "float_cast.h"
#include "os_support.h"
#include "celt/laplace.h"
#include "dred_rdovae_stats_data.h"


static void DRED_rdovae_init_encoder(RDOVAEEncState *enc_state)
{
    memset(enc_state, 0, sizeof(*enc_state));
}

int dred_encoder_load_model(DREDEnc* enc, const void *data, int len)
{
    WeightArray *list;
    int ret;
    parse_weights(&list, data, len);
    ret = init_rdovaeenc(&enc->model, list);
    opus_free(list);
    if (ret == 0) {
      ret = lpcnet_encoder_load_model(&enc->lpcnet_enc_state, data, len);
    }
    if (ret == 0) enc->loaded = 1;
    return (ret == 0) ? OPUS_OK : OPUS_BAD_ARG;
}

void dred_encoder_reset(DREDEnc* enc)
{
    OPUS_CLEAR((char*)&enc->DREDENC_RESET_START,
              sizeof(DREDEnc)-
              ((char*)&enc->DREDENC_RESET_START - (char*)enc));
    enc->input_buffer_fill = DRED_SILK_ENCODER_DELAY;
    lpcnet_encoder_init(&enc->lpcnet_enc_state);
    DRED_rdovae_init_encoder(&enc->rdovae_enc);
}

void dred_encoder_init(DREDEnc* enc, opus_int32 Fs, int channels)
{
    enc->Fs = Fs;
    enc->channels = channels;
    enc->loaded = 0;
#ifndef USE_WEIGHTS_FILE
    if (init_rdovaeenc(&enc->model, rdovaeenc_arrays) == 0) enc->loaded = 1;
#endif
    dred_encoder_reset(enc);
}

static void dred_process_frame(DREDEnc *enc, int arch)
{
    float feature_buffer[2 * 36];
    float input_buffer[2*DRED_NUM_FEATURES] = {0};

    celt_assert(enc->loaded);
    /* shift latents buffer */
    OPUS_MOVE(enc->latents_buffer + DRED_LATENT_DIM, enc->latents_buffer, (DRED_MAX_FRAMES - 1) * DRED_LATENT_DIM);
    OPUS_MOVE(enc->state_buffer + DRED_STATE_DIM, enc->state_buffer, (DRED_MAX_FRAMES - 1) * DRED_STATE_DIM);

    /* calculate LPCNet features */
    lpcnet_compute_single_frame_features_float(&enc->lpcnet_enc_state, enc->input_buffer, feature_buffer, arch);
    lpcnet_compute_single_frame_features_float(&enc->lpcnet_enc_state, enc->input_buffer + DRED_FRAME_SIZE, feature_buffer + 36, arch);

    /* prepare input buffer (discard LPC coefficients) */
    OPUS_COPY(input_buffer, feature_buffer, DRED_NUM_FEATURES);
    OPUS_COPY(input_buffer + DRED_NUM_FEATURES, feature_buffer + 36, DRED_NUM_FEATURES);

    /* run RDOVAE encoder */
    dred_rdovae_encode_dframe(&enc->rdovae_enc, &enc->model, enc->latents_buffer, enc->state_buffer, input_buffer, arch);
    enc->latents_buffer_fill = IMIN(enc->latents_buffer_fill+1, DRED_NUM_REDUNDANCY_FRAMES);
}

void filter_df2t(const float *in, float *out, int len, float b0, const float *b, const float *a, int order, float *mem)
{
    int i;
    for (i=0;i<len;i++) {
        int j;
        float xi, yi, nyi;
        xi = in[i];
        yi = xi*b0 + mem[0];
        nyi = -yi;
        for (j=0;j<order;j++)
        {
           mem[j] = mem[j+1] + b[j]*xi + a[j]*nyi;
        }
        out[i] = yi;
        /*fprintf(stdout, "%f\n", out[i]);*/
    }
}

#define MAX_DOWNMIX_BUFFER (960*2)
static void dred_convert_to_16k(DREDEnc *enc, const float *in, int in_len, float *out, int out_len)
{
    float downmix[MAX_DOWNMIX_BUFFER];
    int i;
    int up;
    celt_assert(enc->channels*in_len <= MAX_DOWNMIX_BUFFER);
    celt_assert(in_len * (opus_int32)16000 == out_len * enc->Fs);
    switch(enc->Fs) {
        case 8000:
            up = 2;
            break;
        case 12000:
            up = 4;
            break;
        case 16000:
            up = 1;
            break;
        case 24000:
            up = 2;
            break;
        case 48000:
            up = 1;
            break;
        default:
            celt_assert(0);
    }
    OPUS_CLEAR(downmix, up*in_len);
    if (enc->channels == 1) {
        for (i=0;i<in_len;i++) downmix[up*i] = FLOAT2INT16(up*in[i]);
    } else {
        for (i=0;i<in_len;i++) downmix[up*i] = FLOAT2INT16(.5*up*(in[2*i]+in[2*i+1]));
    }
    if (enc->Fs == 16000) {
        OPUS_COPY(out, downmix, out_len);
    } else if (enc->Fs == 48000 || enc->Fs == 24000) {
        /* ellip(7, .2, 70, 7750/24000) */

        static const float filter_b[8] = { 0.005873358047f,  0.012980854831f, 0.014531340042f,  0.014531340042f, 0.012980854831f,  0.005873358047f, 0.004523418224f, 0.f};
        static const float filter_a[8] = {-3.878718597768f, 7.748834257468f, -9.653651699533f, 8.007342726666f, -4.379450178552f, 1.463182111810f, -0.231720677804f, 0.f};
        float b0 = 0.004523418224f;
        filter_df2t(downmix, downmix, up*in_len, b0, filter_b, filter_a, RESAMPLING_ORDER, enc->resample_mem);
        for (i=0;i<out_len;i++) out[i] = downmix[3*i];
    } else if (enc->Fs == 12000) {
        /* ellip(7, .2, 70, 7750/24000) */
        static const float filter_b[8] = {-0.001017101081f,  0.003673127243f,   0.001009165267f,  0.001009165267f,  0.003673127243f, -0.001017101081f,  0.002033596776f, 0.f};
        static const float filter_a[8] = {-4.930414411612f, 11.291643096504f, -15.322037343815f, 13.216403930898f, -7.220409219553f,  2.310550142771f, -0.334338618782f, 0.f};
        float b0 = 0.002033596776f;
        filter_df2t(downmix, downmix, up*in_len, b0, filter_b, filter_a, RESAMPLING_ORDER, enc->resample_mem);
        for (i=0;i<out_len;i++) out[i] = downmix[3*i];
    } else if (enc->Fs == 8000) {
        /* ellip(7, .2, 70, 3900/8000) */
        static const float filter_b[8] = { 0.081670120929f, 0.180401598565f,  0.259391051971f, 0.259391051971f,  0.180401598565f, 0.081670120929f,  0.020109185709f, 0.f};
        static const float filter_a[8] = {-1.393651933659f, 2.609789872676f, -2.403541968806f, 2.056814957331f, -1.148908574570f, 0.473001413788f, -0.110359852412f, 0.f};
        float b0 = 0.020109185709f;
        filter_df2t(downmix, out, out_len, b0, filter_b, filter_a, RESAMPLING_ORDER, enc->resample_mem);
    } else {
        celt_assert(0);
    }
}

void dred_compute_latents(DREDEnc *enc, const float *pcm, int frame_size, int extra_delay, int arch)
{
    int curr_offset16k;
    int frame_size16k = frame_size * 16000 / enc->Fs;
    celt_assert(enc->loaded);
    curr_offset16k = 40 + extra_delay*16000/enc->Fs - enc->input_buffer_fill;
    enc->dred_offset = (int)floor((curr_offset16k+20.f)/40.f);
    enc->latent_offset = 0;
    while (frame_size16k > 0) {
        int process_size16k;
        int process_size;
        process_size16k = IMIN(2*DRED_FRAME_SIZE, frame_size16k);
        process_size = process_size16k * enc->Fs / 16000;
        dred_convert_to_16k(enc, pcm, process_size, &enc->input_buffer[enc->input_buffer_fill], process_size16k);
        enc->input_buffer_fill += process_size16k;
        if (enc->input_buffer_fill >= 2*DRED_FRAME_SIZE)
        {
            curr_offset16k += 320;
            dred_process_frame(enc, arch);
            enc->input_buffer_fill -= 2*DRED_FRAME_SIZE;
            OPUS_MOVE(&enc->input_buffer[0], &enc->input_buffer[2*DRED_FRAME_SIZE], enc->input_buffer_fill);
            /* 15 ms (6*2.5 ms) is the ideal offset for DRED because it corresponds to our vocoder look-ahead. */
            if (enc->dred_offset < 6) {
                enc->dred_offset += 8;
            } else {
                enc->latent_offset++;
            }
        }

        pcm += process_size;
        frame_size16k -= process_size16k;
    }
}

static void dred_encode_latents(ec_enc *enc, const float *x, const opus_uint8 *scale, const opus_uint8 *dzone, const opus_uint8 *r, const opus_uint8 *p0, int dim, int arch) {
    int i;
    int q[IMAX(DRED_LATENT_DIM,DRED_STATE_DIM)];
    float xq[IMAX(DRED_LATENT_DIM,DRED_STATE_DIM)];
    float delta[IMAX(DRED_LATENT_DIM,DRED_STATE_DIM)];
    float deadzone[IMAX(DRED_LATENT_DIM,DRED_STATE_DIM)];
    float eps = .1f;
    /* This is split into multiple loops (with temporary arrays) so that the compiler
       can vectorize all of it, and so we can call the vector tanh(). */
    for (i=0;i<dim;i++) {
        delta[i] = dzone[i]*(1.f/256.f);
        xq[i] = x[i]*scale[i]*(1.f/256.f);
        deadzone[i] = xq[i]/(delta[i]+eps);
    }
    compute_activation(deadzone, deadzone, dim, ACTIVATION_TANH, arch);
    for (i=0;i<dim;i++) {
        xq[i] = xq[i] - delta[i]*deadzone[i];
        q[i] = (int)floor(.5f+xq[i]);
    }
    for (i=0;i<dim;i++) {
        /* Make the impossible actually impossible. */
        if (r[i] == 0 || p0[i] == 255) q[i] = 0;
        else ec_laplace_encode_p0(enc, q[i], p0[i]<<7, r[i]<<7);
    }
}

static int dred_voice_active(const unsigned char *activity_mem, int offset) {
    int i;
    for (i=0;i<16;i++) {
        if (activity_mem[8*offset + i] == 1) return 1;
    }
    return 0;
}

int dred_encode_silk_frame(DREDEnc *enc, unsigned char *buf, int max_chunks, int max_bytes, int q0, int dQ, int qmax, unsigned char *activity_mem, int arch) {
    ec_enc ec_encoder;

    int q_level;
    int i;
    int offset;
    int ec_buffer_fill;
    int state_qoffset;
    ec_enc ec_bak;
    int prev_active=0;
    int latent_offset;
    int extra_dred_offset=0;
    int dred_encoded=0;
    int delayed_dred=0;
    int total_offset;

    latent_offset = enc->latent_offset;
    /* Delaying new DRED data when just out of silence because we already have the
       main Opus payload for that frame. */
    if (activity_mem[0] && enc->last_extra_dred_offset>0) {
        latent_offset = enc->last_extra_dred_offset;
        delayed_dred = 1;
        enc->last_extra_dred_offset = 0;
    }
    while (latent_offset < enc->latents_buffer_fill && !dred_voice_active(activity_mem, latent_offset)) {
       latent_offset++;
       extra_dred_offset++;
    }
    if (!delayed_dred) enc->last_extra_dred_offset = extra_dred_offset;

    /* entropy coding of state and latents */
    ec_enc_init(&ec_encoder, buf, max_bytes);
    ec_enc_uint(&ec_encoder, q0, 16);
    ec_enc_uint(&ec_encoder, dQ, 8);
    total_offset = 16 - (enc->dred_offset - extra_dred_offset*8);
    celt_assert(total_offset>=0);
    if (total_offset > 31) {
       ec_enc_uint(&ec_encoder, 1, 2);
       ec_enc_uint(&ec_encoder, total_offset>>5, 256);
       ec_enc_uint(&ec_encoder, total_offset&31, 32);
    } else {
       ec_enc_uint(&ec_encoder, 0, 2);
       ec_enc_uint(&ec_encoder, total_offset, 32);
    }
    celt_assert(qmax >= q0);
    if (q0 < 14 && dQ > 0) {
      int nvals;
      /* If you want to use qmax == q0, you should have set dQ = 0. */
      celt_assert(qmax > q0);
      nvals = 15 - (q0 + 1);
      ec_encode(&ec_encoder, qmax >= 15 ? 0 : nvals + qmax - (q0 + 1),
        qmax >= 15 ? nvals : nvals + qmax - q0, 2*nvals);
    }
    state_qoffset = q0*DRED_STATE_DIM;
    dred_encode_latents(
        &ec_encoder,
        &enc->state_buffer[latent_offset*DRED_STATE_DIM],
        dred_state_quant_scales_q8 + state_qoffset,
        dred_state_dead_zone_q8 + state_qoffset,
        dred_state_r_q8 + state_qoffset,
        dred_state_p0_q8 + state_qoffset,
        DRED_STATE_DIM,
        arch);
    if (ec_tell(&ec_encoder) > 8*max_bytes) {
      return 0;
    }
    ec_bak = ec_encoder;
    for (i = 0; i < IMIN(2*max_chunks, enc->latents_buffer_fill-latent_offset-1); i += 2)
    {
        int active;
        q_level = compute_quantizer(q0, dQ, qmax, i/2);
        offset = q_level * DRED_LATENT_DIM;

        dred_encode_latents(
            &ec_encoder,
            enc->latents_buffer + (i+latent_offset) * DRED_LATENT_DIM,
            dred_latent_quant_scales_q8 + offset,
            dred_latent_dead_zone_q8 + offset,
            dred_latent_r_q8 + offset,
            dred_latent_p0_q8 + offset,
            DRED_LATENT_DIM,
            arch
        );
        if (ec_tell(&ec_encoder) > 8*max_bytes) {
          /* If we haven't been able to code one chunk, give up on DRED completely. */
          if (i==0) return 0;
          break;
        }
        active = dred_voice_active(activity_mem, i+latent_offset);
        if (active || prev_active) {
           ec_bak = ec_encoder;
           dred_encoded = i+2;
        }
        prev_active = active;
    }
    /* Avoid sending empty DRED packets. */
    if (dred_encoded==0 || (dred_encoded<=2 && extra_dred_offset)) return 0;
    ec_encoder = ec_bak;

    ec_buffer_fill = (ec_tell(&ec_encoder)+7)/8;
    ec_enc_shrink(&ec_encoder, ec_buffer_fill);
    ec_enc_done(&ec_encoder);
    return ec_buffer_fill;
}