[FFmpeg-devel] [PATCH] ALAC Encoder

[Jai Menon]

Hi,

On Monday 18 Aug 2008 3:46:23 am Michael Niedermayer wrote:
> On Mon, Aug 18, 2008 at 02:38:24AM +0530, Jai Menon wrote:
> > Hi,
> >
> > On Sunday 17 Aug 2008 5:17:52 pm Michael Niedermayer wrote:
> > > On Sun, Aug 17, 2008 at 11:17:10AM +0530, Jai Menon wrote:
>

[...]

> > @@ -79,6 +122,205 @@
> >      put_bits(&s->pbctx, 32, s->avctx->frame_size);          // No. of
> > samples in the frame }
> >
> > +static void calc_predictor_params(AlacEncodeContext *s, int ch)
> > +{
> > +    int32_t coefs[MAX_LPC_ORDER][MAX_LPC_ORDER];
> > +    int shift[MAX_LPC_ORDER];
> > +    int opt_order;
> > +
> > +    opt_order = ff_lpc_calc_coefs(&s->dspctx, s->sample_buf[ch],
> > s->avctx->frame_size, DEFAULT_MIN_PRED_ORDER, DEFAULT_MAX_PRED_ORDER, +  
> >                                 ALAC_MAX_LPC_PRECISION, coefs, shift, 1,
> > ORDER_METHOD_EST, ALAC_MAX_LPC_SHIFT, 1); +
> > +    s->lpc[ch].lpc_order = opt_order;
> > +    s->lpc[ch].lpc_quant = shift[opt_order-1];
> > +    memcpy(s->lpc[ch].lpc_coeff, coefs[opt_order-1],
> > opt_order*sizeof(int)); +}
> > +
>
> I think this should be using AVCodecContext.min/max_prediction_order
>

Done.

>
> > +    if(best == 0) {
> > +        return ALAC_CHMODE_LEFT_RIGHT;
> > +    } else if(best == 1) {
> > +        return ALAC_CHMODE_LEFT_SIDE;
> > +    } else if(best == 2) {
> > +        return ALAC_CHMODE_RIGHT_SIDE;
> > +    } else {
> > +        return ALAC_CHMODE_MID_SIDE;
> > +    }
> > +}
>
> i think best could simply be returned
>

Done.

> > +
> > +static void alac_stereo_decorrelation(AlacEncodeContext *s)
> > +{
> > +    int32_t *left = s->sample_buf[0], *right = s->sample_buf[1];
> > +    int i, mode, n = s->avctx->frame_size;
> > +
> > +    mode = estimate_stereo_mode(left, right, n);
> > +
> > +    if(mode == ALAC_CHMODE_LEFT_RIGHT) {
> > +        s->interlacing_leftweight = 0;
> > +        s->interlacing_shift = 0;
> > +        return;
> > +    }
> > +
> > +    if(mode == ALAC_CHMODE_LEFT_SIDE) {
> > +        for(i=0; i<n; i++) {
> > +            right[i] = left[i] - right[i];
> > +        }
> > +        s->interlacing_leftweight = 1;
> > +        s->interlacing_shift = 0;
> > +
> > +    } else {
> > +        int32_t tmp;
> > +        for(i=0; i<n; i++) {
> > +            tmp = left[i];
> > +            left[i] = (tmp + right[i]) >> 1;
> > +            right[i] = tmp - right[i];
> > +        }
> > +        s->interlacing_leftweight = 1;
> > +        s->interlacing_shift = 1;
> > +    }
>
> i think 1 mode is missing
> also it could be
> if
> else if
> else if
> else
>
> instead of
> if
>     return
> if
> else
>
> which would be cleaner IMHO
>

Right/Side mode added and also cleaned up the code structure.

> > +}
> > +
> > +static void alac_linear_predictor(AlacEncodeContext *s, int ch)
> > +{
> > +    int i;
> > +    LPCContext lpc = s->lpc[ch];
> > +
> > +    if(lpc.lpc_order == 31) {
> > +        s->predictor_buf[0] = s->sample_buf[ch][0];
> > +
> > +        for(i=1; i<s->avctx->frame_size; i++)
> > +            s->predictor_buf[i] = s->sample_buf[ch][i] -
> > s->sample_buf[ch][i-1]; +
> > +        return;
> > +    }
> > +
> > +    // generalised linear predictor
> > +
> > +    if(lpc.lpc_order > 0) {
> > +        int32_t *samples  = s->sample_buf[ch];
> > +        int32_t *residual = s->predictor_buf;
> > +
> > +        // generate warm-up samples
> >
> > +        i = lpc.lpc_order;
> > +        residual[0] = samples[0];
> > +        while(i > 0) {
> > +            residual[i] = samples[i] - samples[i-1];
> > +            i--;
> > +        }
>
> this also can be changed to a for() loop
> alternatively residual could be droped and the stuff could all be done
> in place in samples but this may be tricky
>

Changed to a 'for' loop. I'll look into the in-place thing after the basic 
encoder is committed ;-)

> > +        // perform lpc on remaining samples
> > +        for(i = lpc.lpc_order + 1; i < s->avctx->frame_size; i++) {
> > +            int sum = 0, res_val, j;
> > +
> > +            for (j = 0; j < lpc.lpc_order; j++) {
> > +                sum += (samples[lpc.lpc_order-j] - samples[0]) *
> > +                        lpc.lpc_coeff[j];
> > +            }
> > +            sum += (1 << (lpc.lpc_quant - 1));
> > +            sum >>= lpc.lpc_quant;
> > +            sum += samples[0];
> > +            residual[i] = samples[lpc.lpc_order+1] - sum;
> > +            res_val = residual[i];
> > +
> > +            if(res_val) {
> > +                int index = lpc.lpc_order - 1;
> > +                int neg = (res_val < 0);
> > +
> > +                while(index >= 0 && (neg ? (res_val < 0):(res_val > 0)))
> > { +                    int val = samples[0] - samples[lpc.lpc_order -
> > index]; +                    int sign = (val ? FFSIGN(val) : 0);
> > +
> > +                    if(neg)
> > +                        sign*=-1;
> > +
> > +                    lpc.lpc_coeff[index] -= sign;
> > +                    val *= sign;
> > +                    res_val -= ((val >> lpc.lpc_quant) *
> > +                            (lpc.lpc_order - index));
> > +                    index--;
> > +                }
> > +            }
> > +            samples++;
> > +        }
> > +    }
> > +}
> > +
> >
> > +static void alac_entropy_coder(AlacEncodeContext *s)
> > +{
> > +    unsigned int history = s->rc.initial_history;
> > +    int sign_modifier = 0, i, k;
> > +    int32_t *samples = s->predictor_buf;
> > +
> > +    for(i=0;i < s->avctx->frame_size;) {
> > +        int x;
> > +
> > +        k = av_log2((history >> 9) + 3);
> > +
> > +        x = -2*(*samples)-1;
> > +        x ^= (x>>31);
> > +
> > +        samples++;
> > +        i++;
> > +
> > +        encode_scalar(s, x - sign_modifier, k, s->write_sample_size);
> > +
> >
> > +        history += x * s->rc.history_mult
> > +                   - ((history * s->rc.history_mult) >> 9);
>
> not sure if its worth but this could be simplified to:
>
> history -= (((history - (x<<9)) * s->rc.history_mult) >> 9);
> (assuming things dont overflow)
>

There is no substantal change when verified using start/stop_timer. I prefer 
this as it looks clean and nicely split up :-)

> > +
> > +        sign_modifier = 0;
> > +        if(x > 0xFFFF)
> > +            history = 0xFFFF;
> > +
> > +        if((history < 128) && (i < s->avctx->frame_size)) {
> > +            unsigned int block_size = 0;
> > +
> >
> > +            sign_modifier = 1;
>
> unused
>

It is very much used actually. After we have coded a block of zeros, the next 
scalar is non_zero and this sign modifier is applied to that.

> > +
> > +    init_put_bits(pb, frame, buf_size);
> > +
> > +verbatim:
>
> the label can be moved before the init_put_bits() which makes the second
> call to it before the goto uneeded
>

Indeed. Modified accordingly.

> [...]

New patch attached. Thanks for the review and help with the decorrelation 
stuff.


Regards,

Jai Menon
-------------- next part --------------
A non-text attachment was scrubbed...
Name: alac_encoder.patch
Type: text/x-diff
Size: 8773 bytes
Desc: not available
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20080819/a1a1b350/attachment.patch>