[FFmpeg-devel] [PATCH] avcodec/dcaenc: Initial implementation of ADPCM encoding for DCA encoder

[Rostislav Pehlivanov]

On 6 May 2017 at 11:47, Даниил Чередник <dan.cherednik at gmail.com> wrote:

> Tiny fix: missed inline attr.
>
> On Thu, May 4, 2017 at 1:17 AM, Даниил Чередник <dan.cherednik at gmail.com>
> wrote:
>
> > All notices were fixed. Also I found issue with uninitialized subband
> > buffer - fixed. New patch attached.
> >
> > On Wed, May 3, 2017 at 7:49 AM, Rostislav Pehlivanov <
> atomnuker at gmail.com>
> > wrote:
> >
> >> On 2 May 2017 at 22:53, Даниил Чередник <dan.cherednik at gmail.com>
> wrote:
> >>
> >> > Hi.
> >> >
> >> > This patch introduces initial implementation of subband ADPCM encoding
> >> for
> >> > DCA codec.
> >> >
> >> > Some results:
> >> >
> >> > sample:
> >> >
> >> > https://yadi.sk/d/B_3sVskM3HZiWK - original
> >> >
> >> > https://yadi.sk/d/7CK47Nt63HZiWf - without adpcm
> >> >
> >> > https://yadi.sk/d/25q1JDV93HZiWq - with adpcm
> >> >
> >> > chirp tone:
> >> >
> >> > https://yadi.sk/i/tZKHoJ1d3HZk4c
> >> >
> >> > Right now this feature is disabled by default. But it is ready to try
> >> > using -dca_adpcm 1 option.
> >> >
> >> > There are some issues, should be solved before enabling this feature
> by
> >> > default:
> >> >
> >> > 1. Speed up: I am trying to find best filter in each subband. But with
> >> real
> >> > signal, usually only few subbands has significant prediction gain. The
> >> idea
> >> > is try to analyze FFT spectrum (which is already calculated), to check
> >> is
> >> > particular subband looks like tonal or noise. If subband is noise like
> >> - do
> >> > not try to find best LPC predictor.
> >> >
> >> > 2. Modify psychoacoustic to use prediction gain for bit allocation.
> >> Right
> >> > now ADPCM encoded block can get some extra bits.
> >> >
> >> > 3. Tuning the prediction gain threshold.
> >> >
> >> >
> >> > Thank you.
> >> > --
> >> > Daniil Cherednik
> >> >
> >> > _______________________________________________
> >> > ffmpeg-devel mailing list
> >> > ffmpeg-devel at ffmpeg.org
> >> > http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> >> >
> >> >
> >> >+static int64_t calc_corr(const int32_t *x, int len, int j, int k)
> >>
> >> Add inline attrib? Seems appropriate here.
> >>
> >>
> >> >+for (n = 0; n < len; n++) {
> >> >+        s += MUL64(x[n-j], x[n-k]);
> >> >+    }
> >>
> >> For loops with 1 line we leave the brackets out.
> >>
> >>
> >> >+for (i = 0; i <= DCA_ADPCM_COEFFS; i++) {
> >> >+        for (j = i; j <= DCA_ADPCM_COEFFS; j++) {
> >> >+            corr[k++] = calc_corr(in+4, len, i, j);
> >> >+        }
> >> >+    }
> >>
> >> Same
> >>
> >>
> >> >+    for (i = 0; i < len + DCA_ADPCM_COEFFS; i++) {
> >> >+        max |= FFABS(in[i]);
> >> >+    }
> >>
> >> Same
> >>
> >>
> >> >for (ch = 0; ch < c->fullband_channels; ch++) {
> >> >+        for (band = 0; band < 32; band++) {
> >> >+            if (c->prediction_mode[ch][band] >= 0) {
> >> >+                quantize_adpcm_subband(c, ch, band);
> >> >+            }
> >> >+        }
> >> >+    }
> >>
> >> Same
> >>
> >>
> >> >+    for (ch = 0; ch < c->fullband_channels; ch++) {
> >> >+        for (band = 0; band < 32; band++) {
> >> >+            if (c->prediction_mode[ch][band] == -1) {
> >> >+                for (sample = 0; sample < SUBBAND_SAMPLES; sample++) {
> >> >+                    c->quantized[ch][band][sample] =
> >> quantize_value(c->subband[ch][band][sample], c->quant[ch][band]);
> >> >+                }
> >> >+            }
> >> >+        }
> >> >+    }
> >>
> >> Same, 4 whole whitespace lines added here.
> >>
> >>
> >> >+    if (c->bitrate_index == 3) {
> >> >+        step_size = ff_dca_lossless_quant[c->abits[ch][band]];
> >> >+    } else {
> >> >+        step_size = ff_dca_lossy_quant[c->abits[ch][band]];
> >> >+    }
> >>
> >> Same
> >>
> >>
> >> >for (;;) {
> >>
> >> while (1) {
> >>
> >> >+        if (i++ == last_pos)
> >> >+            break;
> >>
> >> Better yet remove the infinite loop and just use a normal for () loop.
> >>
> >>
> >> >+static inline void ff_dca_core_dequantize(int32_t *output, const
> int32_t
> >> *input,
> >> >+                              int32_t step_size, int32_t scale, int
> >> residual, int len)
> >>
> >> Fix second line's alignment.
> >>
> >>
> >> >+struct premultiplied_coeffs {
> >> >+    int32_t aa[10];
> >> >+};
> >>
> >> I think it would be simpler to just use int32_t premultiplied_coeffs[10]
> >> instead of a struct.
> >>
> >>
> >> Apart from these style issues patch looks fine. I'll be able to test it
> in
> >> a day.
> >> _______________________________________________
> >> ffmpeg-devel mailing list
> >> ffmpeg-devel at ffmpeg.org
> >> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> >>
> >
> >
> >
> > --
> > Daniil Cherednik
> >
>
>
>
> --
> Daniil Cherednik
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
>
Tested, there's a difference though looking at the spectrum is the only way
to see it. The encoder only supports bitrates higher than ~251kbps which
isn't enough to really make a big difference. Tried to reduce it but the
bit allocation algorithm of the encoder broke.

Anyway, thanks, pushed the patch.