/* * only when len>6, memory access can be reduced by half. */
opus_int32 silk_inner_prod_aligned(
const opus_int16 *const inVec1, /* I input vector 1 */
const opus_int16 *const inVec2, /* I input vector 2 */
const opus_int len, /* I vector lengths */
int arch /* I Run-time architecture */
)
{
#ifdef FIXED_POINT
return celt_inner_prod(inVec1, inVec2, len, arch);
#else
opus_int i;
opus_int32 sum = 0;
for(i = 0; i < len; i++) {
sum = silk_SMLABB(sum, inVec1[ i ], inVec2[ i ]);
}
return sum;
#endif
}
void surround_analysis(const CELTMode *celt_mode, const void *pcm, opus_val16 *bandLogE, opus_val32 *mem, opus_val32 *preemph_mem,
int len, int overlap, int channels, int rate, opus_copy_channel_in_func copy_channel_in, int arch
)
{
int c;
int i;
int LM;
int pos[8] = {0};
int upsample;
int frame_size;
int freq_size;
opus_val16 channel_offset;
opus_val32 bandE[21];
opus_val16 maskLogE[3][21];
VARDECL(opus_val32, in);
VARDECL(opus_val16, x);
VARDECL(opus_val32, freq);
SAVE_STACK;
upsample = resampling_factor(rate);
frame_size = len*upsample;
freq_size = IMIN(960, frame_size);
/* LM = log2(frame_size / 120) */
for (LM=0;LM<celt_mode->maxLM;LM++)
if (celt_mode->shortMdctSize<<LM==frame_size)
break;
ALLOC(in, frame_size+overlap, opus_val32);
ALLOC(x, len, opus_val16);
ALLOC(freq, freq_size, opus_val32);
channel_pos(channels, pos);
for (c=0;c<3;c++)
for (i=0;i<21;i++)
maskLogE[c][i] = -QCONST16(28.f, DB_SHIFT);
for (c=0;c<channels;c++)
{
int frame;
int nb_frames = frame_size/freq_size;
celt_assert(nb_frames*freq_size == frame_size);
OPUS_COPY(in, mem+c*overlap, overlap);
(*copy_channel_in)(x, 1, pcm, channels, c, len);
celt_preemphasis(x, in+overlap, frame_size, 1, upsample, celt_mode->preemph, preemph_mem+c, 0);
#ifndef FIXED_POINT
{
opus_val32 sum;
sum = celt_inner_prod(in, in, frame_size+overlap, 0);
/* This should filter out both NaNs and ridiculous signals that could
cause NaNs further down. */
if (!(sum < 1e9f) || celt_isnan(sum))
{
OPUS_CLEAR(in, frame_size+overlap);
preemph_mem[c] = 0;
}
}
#endif
OPUS_CLEAR(bandE, 21);
for (frame=0;frame<nb_frames;frame++)
{
opus_val32 tmpE[21];
clt_mdct_forward(&celt_mode->mdct, in+960*frame, freq, celt_mode->window,
overlap, celt_mode->maxLM-LM, 1, arch);
if (upsample != 1)
{
int bound = freq_size/upsample;
for (i=0;i<bound;i++)
freq[i] *= upsample;
for (;i<freq_size;i++)
freq[i] = 0;
}
compute_band_energies(celt_mode, freq, tmpE, 21, 1, LM);
/* If we have multiple frames, take the max energy. */
for (i=0;i<21;i++)
bandE[i] = MAX32(bandE[i], tmpE[i]);
}
amp2Log2(celt_mode, 21, 21, bandE, bandLogE+21*c, 1);
/* Apply spreading function with -6 dB/band going up and -12 dB/band going down. */
for (i=1;i<21;i++)
bandLogE[21*c+i] = MAX16(bandLogE[21*c+i], bandLogE[21*c+i-1]-QCONST16(1.f, DB_SHIFT));
for (i=19;i>=0;i--)
bandLogE[21*c+i] = MAX16(bandLogE[21*c+i], bandLogE[21*c+i+1]-QCONST16(2.f, DB_SHIFT));
if (pos[c]==1)
{
for (i=0;i<21;i++)
maskLogE[0][i] = logSum(maskLogE[0][i], bandLogE[21*c+i]);
} else if (pos[c]==3)
{
for (i=0;i<21;i++)
maskLogE[2][i] = logSum(maskLogE[2][i], bandLogE[21*c+i]);
} else if (pos[c]==2)
{
for (i=0;i<21;i++)
{
maskLogE[0][i] = logSum(maskLogE[0][i], bandLogE[21*c+i]-QCONST16(.5f, DB_SHIFT));
maskLogE[2][i] = logSum(maskLogE[2][i], bandLogE[21*c+i]-QCONST16(.5f, DB_SHIFT));
}
}
#if 0
for (i=0;i<21;i++)
printf("%f ", bandLogE[21*c+i]);
float sum=0;
for (i=0;i<21;i++)
sum += bandLogE[21*c+i];
printf("%f ", sum/21);
#endif
OPUS_COPY(mem+c*overlap, in+frame_size, overlap);
}
for (i=0;i<21;i++)
maskLogE[1][i] = MIN32(maskLogE[0][i],maskLogE[2][i]);
channel_offset = HALF16(celt_log2(QCONST32(2.f,14)/(channels-1)));
for (c=0;c<3;c++)
for (i=0;i<21;i++)
maskLogE[c][i] += channel_offset;
#if 0
for (c=0;c<3;c++)
{
for (i=0;i<21;i++)
printf("%f ", maskLogE[c][i]);
}
#endif
for (c=0;c<channels;c++)
{
opus_val16 *mask;
if (pos[c]!=0)
{
mask = &maskLogE[pos[c]-1][0];
for (i=0;i<21;i++)
bandLogE[21*c+i] = bandLogE[21*c+i] - mask[i];
} else {
for (i=0;i<21;i++)
bandLogE[21*c+i] = 0;
}
#if 0
for (i=0;i<21;i++)
printf("%f ", bandLogE[21*c+i]);
printf("\n");
#endif
#if 0
float sum=0;
for (i=0;i<21;i++)
sum += bandLogE[21*c+i];
printf("%f ", sum/(float)QCONST32(21.f, DB_SHIFT));
printf("\n");
#endif
}
RESTORE_STACK;
}
