On 04/23/2011 04:33 PM, Ronald S. Bultje wrote:
> Hi,
>
> On Sun, Apr 17, 2011 at 11:39 AM, Justin Ruggles
> <[email protected]> wrote:
> [..]
>> + /* calculate energy in each band in coupling channel and each fbw
>> channel */
>> + bnd = 0;
>> + i = s->start_freq[CPL_CH];
>> + while (i < s->cpl_end_freq) {
>> + int band_size = s->cpl_band_sizes[bnd];
>> + for (ch = CPL_CH; ch <= s->fbw_channels; ch++) {
>> + for (blk = 0; blk < AC3_MAX_BLOCKS; blk++) {
>> + AC3Block *block = &s->blocks[blk];
>> + if (!block->cpl_in_use || (ch > CPL_CH &&
>> !block->channel_in_cpl[ch]))
>> + continue;
>> + for (j = 0; j < band_size; j++) {
>> + float v = block->mdct_coef[ch][i+j];
>> + energy[blk][ch][bnd] += v * v;
>
> This looks somewhat SIMD'able. (Don't bother if an existing simd
> function doesn't already exist, but sum of squares seems common enough
> to exist...)
scalarproduct_float() would work, but i would need to play around with
the alignment and size. the start position for all bands are i%4=1, and
band_size is a multiple of 12. we would likely have to modify
scalarproduct_float() to handle unaligned data. that's easy enough for
the sse version, but i don't know about the neon version.
>> + coord = 0.125f;
>> + if (energy_cpl > 0)
>> + coord = sqrtf(energy_ch / energy_cpl) / 8.0f;
>
> coord *= sqrtf(energy_ch / energy_cpl);
indeed. :)
>> @@ -1129,7 +1489,21 @@ static int compute_bit_allocation(AC3EncodeContext *s)
>> ret = cbr_bit_allocation(s);
>> while (ret) {
>> - /* fallback 1: downgrade exponents */
>> + /* fallback 1: disable channel coupling */
>> + if (s->cpl_on) {
>> + s->cpl_on = 0;
>> + apply_mdct(s);
>> + compute_coupling_strategy(s);
>> + compute_rematrixing_strategy(s);
>> + scale_coefficients(s);
>> + compute_rematrixing_strategy(s);
>> + apply_rematrixing(s);
>> + process_exponents(s);
>> + ret = compute_bit_allocation(s);
>> + continue;
>> + }
>> +
>> + /* fallback 2: downgrade exponents */
>> if (!downgrade_exponents(s)) {
>> extract_exponents(s);
>> encode_exponents(s);
>
> So if there's no bits left, first thing you do is to disable channel
> coupling. Does that always provide less quality loss per bits saved
> than changing exponents? Since you now have multiple things to try,
> you could maybe try both and calculate which provides best "bits saved
> per quality loss" scenarios.
The fallbacks are only needed at really low bitrates in order to
generate a valid stream. Either way it will sound like crap. Modifying
exponents when coupling is turned on is much more complicated.
-Justin
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel
