comparison aacdec.c @ 12259:97db3c60fb9c libavcodec

aacdec: 4% faster main profile decoding.

12258:86d6c00756cc

    pun.f = pf;
    pun.i &= 0xFFFF0000U;
    return pun.f;
}

static av_always_inline void predict(AACContext *ac, PredictorState *ps, float *coef,
                    int output_enable)
{
    const float a     = 0.953125; // 61.0 / 64
    const float alpha = 0.90625;  // 29.0 / 32
    float e0, e1;

    k1 = ps->var0 > 1 ? ps->cor0 * flt16_even(a / ps->var0) : 0;
    k2 = ps->var1 > 1 ? ps->cor1 * flt16_even(a / ps->var1) : 0;

    pv = flt16_round(k1 * ps->r0 + k2 * ps->r1);
    if (output_enable)
        *coef += pv * ac->sf_scale;

    e0 = *coef / ac->sf_scale;
    e1 = e0 - k1 * ps->r0;

    ps->cor1 = flt16_trunc(alpha * ps->cor1 + ps->r1 * e1);
    ps->var1 = flt16_trunc(alpha * ps->var1 + 0.5f * (ps->r1 * ps->r1 + e1 * e1));
    ps->cor0 = flt16_trunc(alpha * ps->cor0 + ps->r0 * e0);

 * Apply AAC-Main style frequency domain prediction.
 */
static void apply_prediction(AACContext *ac, SingleChannelElement *sce)
{
    int sfb, k;

    if (!sce->ics.predictor_initialized) {
        reset_all_predictors(sce->predictor_state);
        sce->ics.predictor_initialized = 1;
    }

    if (sce->ics.window_sequence[0] != EIGHT_SHORT_SEQUENCE) {
        for (sfb = 0; sfb < ff_aac_pred_sfb_max[ac->m4ac.sampling_index]; sfb++) {
            for (k = sce->ics.swb_offset[sfb]; k < sce->ics.swb_offset[sfb + 1]; k++) {
                predict(ac, &sce->predictor_state[k], &sce->coeffs[k],
                        sce->ics.predictor_present && sce->ics.prediction_used[sfb]);
            }
        }
        if (sce->ics.predictor_reset_group)
            reset_predictor_group(sce->predictor_state, sce->ics.predictor_reset_group);

12259:97db3c60fb9c

    pun.f = pf;
    pun.i &= 0xFFFF0000U;
    return pun.f;
}

static av_always_inline void predict(PredictorState *ps, float *coef,
                                     float sf_scale, float inv_sf_scale,
                    int output_enable)
{
    const float a     = 0.953125; // 61.0 / 64
    const float alpha = 0.90625;  // 29.0 / 32
    float e0, e1;

    k1 = ps->var0 > 1 ? ps->cor0 * flt16_even(a / ps->var0) : 0;
    k2 = ps->var1 > 1 ? ps->cor1 * flt16_even(a / ps->var1) : 0;

    pv = flt16_round(k1 * ps->r0 + k2 * ps->r1);
    if (output_enable)
        *coef += pv * sf_scale;

    e0 = *coef * inv_sf_scale;
    e1 = e0 - k1 * ps->r0;

    ps->cor1 = flt16_trunc(alpha * ps->cor1 + ps->r1 * e1);
    ps->var1 = flt16_trunc(alpha * ps->var1 + 0.5f * (ps->r1 * ps->r1 + e1 * e1));
    ps->cor0 = flt16_trunc(alpha * ps->cor0 + ps->r0 * e0);

 * Apply AAC-Main style frequency domain prediction.
 */
static void apply_prediction(AACContext *ac, SingleChannelElement *sce)
{
    int sfb, k;
    float sf_scale = ac->sf_scale, inv_sf_scale = 1 / ac->sf_scale;

    if (!sce->ics.predictor_initialized) {
        reset_all_predictors(sce->predictor_state);
        sce->ics.predictor_initialized = 1;
    }

    if (sce->ics.window_sequence[0] != EIGHT_SHORT_SEQUENCE) {
        for (sfb = 0; sfb < ff_aac_pred_sfb_max[ac->m4ac.sampling_index]; sfb++) {
            for (k = sce->ics.swb_offset[sfb]; k < sce->ics.swb_offset[sfb + 1]; k++) {
                predict(&sce->predictor_state[k], &sce->coeffs[k],
                        sf_scale, inv_sf_scale,
                        sce->ics.predictor_present && sce->ics.prediction_used[sfb]);
            }
        }
        if (sce->ics.predictor_reset_group)
            reset_predictor_group(sce->predictor_state, sce->ics.predictor_reset_group);