mirror of
https://github.com/Rockbox/rockbox.git
synced 2026-10-09 23:53:28 -04:00
wma: fix the level of noise coded high bands
Low bitrate WMA files code their upper bands as noise at a given
power. Those bands came out 12 to 25 dB too quiet, which dulled
the treble. Two things were wrong.
The exponent pointer was moved too far in blocks whose exponents
have another resolution. This is ffmpeg's r20756 (f78501b264,
"Fix apparent 10l typos introduced in r8627"); r8627 was merged
here in 2f1da8d24a but the fix for it never was.
The fixed point gain of a band lost nearly all its precision: it
was zero for three bands in four in the file traced, and the sum
for a band's power overflowed in a quarter of them. Compute the
gain in 64 bits, once a band. The factor for the noise below the
first coded coefficient (WMA v1 only) was 16 bits too large and is
corrected by the same reasoning; no file here exercises it.
Checked with perfsim (Sansa e200v1 build) against ffmpeg's decode,
as the mean difference in band levels over the whole file:
01 - Jane Austen (44.1 kHz, 32 kbps) 5.5 dB -> 0.0 dB
07.Devil.In.My.Mind (same) 6.2 dB -> 0.0 dB
beyonthepain907z (same) 6.8 dB -> 0.0 dB
moshimoashitaga (same) 2.8 dB -> 0.0 dB
Four more noise coded files, already within 0.3 dB, now match too.
The output of the 15 files without noise coding is byte-identical.
Files that use LSP exponents still differ in their top bands; that
is a separate fault in the LSP curve.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Michael Giacomelli
parent
fb2ee6c08f
commit
11b12894ad
1 changed files with 84 additions and 22 deletions
|
|
@ -777,6 +777,69 @@ static int decode_exp_vlc(WMADecodeContext *s, int ch)
|
|||
return 0;
|
||||
}
|
||||
|
||||
/* The gain of a noise coded band, as ffmpeg computes it in floating point:
|
||||
*
|
||||
* mdct_norm * sqrt(power / last_power) * 10^(value/20)
|
||||
* / (max_exponent * noise_mult)
|
||||
*
|
||||
* returned as the 16.16 factor wma_decode_block() applies to the band.
|
||||
* 1/noise_mult is 50 or 25. The terms differ too much in size for 32 bits,
|
||||
* and this runs once a band, so it is all done in 64 bits. */
|
||||
static fixed32 noise_band_gain(uint64_t power, uint64_t last_power, int value,
|
||||
fixed32 max_exponent, fixed32 mdct_norm,
|
||||
int inv_noise_mult)
|
||||
{
|
||||
uint64_t ratio, root, bit, a, b;
|
||||
int shift = 16 + 20;
|
||||
|
||||
if (power == 0 || max_exponent <= 0)
|
||||
return 0;
|
||||
|
||||
/* the power ratio in 16.16, kept small enough to take its root */
|
||||
ratio = last_power ? (power << 16) / last_power : (uint64_t)1 << 46;
|
||||
if (ratio > (uint64_t)1 << 46)
|
||||
ratio = (uint64_t)1 << 46;
|
||||
|
||||
/* its root in 16.16 */
|
||||
ratio <<= 16;
|
||||
root = 0;
|
||||
for (bit = (uint64_t)1 << 62; bit != 0; bit >>= 2) {
|
||||
if (ratio >= root + bit) {
|
||||
ratio -= root + bit;
|
||||
root = (root >> 1) + bit;
|
||||
} else {
|
||||
root >>= 1;
|
||||
}
|
||||
}
|
||||
|
||||
a = root * mdct_norm * inv_noise_mult;
|
||||
b = pow_table[value + 20];
|
||||
while (b >> 43) { /* keep the shift below in 64 bits */
|
||||
b >>= 1;
|
||||
shift--;
|
||||
}
|
||||
b = (b << 20) / max_exponent;
|
||||
|
||||
/* a * b >> shift, with both brought down to 31 bits first */
|
||||
while (a >> 31) {
|
||||
a >>= 1;
|
||||
shift--;
|
||||
}
|
||||
while (b >> 31) {
|
||||
b >>= 1;
|
||||
shift--;
|
||||
}
|
||||
a *= b;
|
||||
if (shift > 0)
|
||||
a >>= shift;
|
||||
else if (a >> (31 + shift))
|
||||
return 0x7fffffff;
|
||||
else
|
||||
a <<= -shift;
|
||||
|
||||
return a > 0x7fffffff ? 0x7fffffff : (fixed32)a;
|
||||
}
|
||||
|
||||
/* return 0 if OK. return 1 if last block of frame. return -1 if
|
||||
unrecorrable error. */
|
||||
static int wma_decode_block(WMADecodeContext *s)
|
||||
|
|
@ -1064,7 +1127,7 @@ static int wma_decode_block(WMADecodeContext *s)
|
|||
fixed64 mult1;
|
||||
fixed32 noise, temp1, temp2, mult2;
|
||||
int i, j, n, n1, last_high_band, esize;
|
||||
fixed32 exp_power[HIGH_BAND_MAX_SIZE];
|
||||
uint64_t exp_power[HIGH_BAND_MAX_SIZE];
|
||||
|
||||
//total_gain, coefs1, mdctnorm are lossless
|
||||
|
||||
|
|
@ -1098,14 +1161,15 @@ static int wma_decode_block(WMADecodeContext *s)
|
|||
for(i = 0;i < s->coefs_start; ++i)
|
||||
{
|
||||
*coefs++ = fixmul32( (fixmul32(s->noise_table[s->noise_index],
|
||||
exponents[i<<bsize>>esize])>>4),Fixed32From64(mult1)) >>2;
|
||||
exponents[i<<bsize>>esize])>>4),(fixed32)(mult1>>16)) >>2;
|
||||
s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
|
||||
}
|
||||
|
||||
n1 = s->exponent_high_sizes[bsize];
|
||||
|
||||
/* compute power of high bands */
|
||||
exponents = s->exponents[ch] +(s->high_band_start[bsize]<<bsize);
|
||||
exponents = s->exponents[ch] +
|
||||
(s->high_band_start[bsize]<<bsize>>esize);
|
||||
last_high_band = 0; /* avoid warning */
|
||||
for (j=0;j<n1;++j)
|
||||
{
|
||||
|
|
@ -1113,22 +1177,23 @@ static int wma_decode_block(WMADecodeContext *s)
|
|||
s->block_len_bits][j];
|
||||
if (s->high_band_coded[ch][j])
|
||||
{
|
||||
fixed32 e2, v;
|
||||
e2 = 0;
|
||||
uint64_t e2 = 0;
|
||||
for(i = 0;i < n; ++i)
|
||||
{
|
||||
/*v is normalized later on so its fixed format is irrelevant*/
|
||||
v = exponents[i<<bsize>>esize]>>4;
|
||||
e2 += fixmul32(v, v)>>3;
|
||||
/* Only ratios of the powers are used, so their
|
||||
format does not matter; this one cannot
|
||||
overflow. */
|
||||
uint32_t v = exponents[i<<bsize>>esize]>>6;
|
||||
e2 += (uint64_t)v * v;
|
||||
}
|
||||
exp_power[j] = e2/n; /*n is an int...*/
|
||||
exp_power[j] = e2/n;
|
||||
last_high_band = j;
|
||||
}
|
||||
exponents += n<<bsize;
|
||||
exponents += n<<bsize>>esize;
|
||||
}
|
||||
|
||||
/* main freqs and high freqs */
|
||||
exponents = s->exponents[ch] + (s->coefs_start<<bsize);
|
||||
exponents = s->exponents[ch] + (s->coefs_start<<bsize>>esize);
|
||||
for(j=-1;j<n1;++j)
|
||||
{
|
||||
if (j < 0)
|
||||
|
|
@ -1144,24 +1209,21 @@ static int wma_decode_block(WMADecodeContext *s)
|
|||
if (j >= 0 && s->high_band_coded[ch][j])
|
||||
{
|
||||
/* use noise with specified power */
|
||||
fixed32 tmp = fixdiv32(exp_power[j],exp_power[last_high_band]);
|
||||
fixed32 gain = noise_band_gain(exp_power[j],
|
||||
exp_power[last_high_band],
|
||||
s->high_band_values[ch][j],
|
||||
s->max_exponent[ch], mdct_norm,
|
||||
s->use_exp_vlc ? 50 : 25);
|
||||
|
||||
/*mult1 is 48.16, pow_table is 48.16*/
|
||||
mult1 = fixmul32(fixsqrt32(tmp),
|
||||
pow_table[s->high_band_values[ch][j]+20]) >> 16;
|
||||
|
||||
/*this step has a fairly high degree of error for some reason*/
|
||||
mult1 = fixdiv64(mult1,fixmul32(s->max_exponent[ch],s->noise_mult));
|
||||
mult1 = mult1*mdct_norm>>PRECISION;
|
||||
for(i = 0;i < n; ++i)
|
||||
{
|
||||
noise = s->noise_table[s->noise_index];
|
||||
s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
|
||||
*coefs++ = fixmul32((fixmul32(exponents[i<<bsize>>esize],noise)>>4),
|
||||
Fixed32From64(mult1)) >>2;
|
||||
gain) >>2;
|
||||
|
||||
}
|
||||
exponents += n<<bsize;
|
||||
exponents += n<<bsize>>esize;
|
||||
}
|
||||
else
|
||||
{
|
||||
|
|
@ -1176,7 +1238,7 @@ static int wma_decode_block(WMADecodeContext *s)
|
|||
temp2 = fixmul32(exponents[i<<bsize>>esize], mult>>18);
|
||||
*coefs++ = fixmul32(temp1, temp2);
|
||||
}
|
||||
exponents += n<<bsize;
|
||||
exponents += n<<bsize>>esize;
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue