void *output_to_PCM(NeAACDecHandle hDecoder,
real_t **input, void *sample_buffer, uint8_t channels,
uint16_t frame_len, uint8_t format)
{
int16_t *short_sample_buffer = (int16_t*)sample_buffer;
int32_t *int_sample_buffer = (int32_t*)sample_buffer;
float32_t *float_sample_buffer = (float32_t*)sample_buffer;
double *double_sample_buffer = (double*)sample_buffer;
#ifdef PROFILE
int64_t count = faad_get_ts();
#endif
/* Copy output to a standard PCM buffer */
switch (format)
{
case FAAD_FMT_16BIT:
to_PCM_16bit(hDecoder, input, channels, frame_len, &short_sample_buffer);
break;
case FAAD_FMT_24BIT:
to_PCM_24bit(hDecoder, input, channels, frame_len, &int_sample_buffer);
break;
case FAAD_FMT_32BIT:
to_PCM_32bit(hDecoder, input, channels, frame_len, &int_sample_buffer);
break;
case FAAD_FMT_FLOAT:
to_PCM_float(hDecoder, input, channels, frame_len, &float_sample_buffer);
break;
case FAAD_FMT_DOUBLE:
to_PCM_double(hDecoder, input, channels, frame_len, &double_sample_buffer);
break;
}
#ifdef PROFILE
count = faad_get_ts() - count;
hDecoder->output_cycles += count;
#endif
return sample_buffer;
}
void faad_imdct(mdct_info *mdct, real_t *X_in, real_t *X_out)
{
uint16_t k;
complex_t x;
#ifdef ALLOW_SMALL_FRAMELENGTH
#ifdef FIXED_POINT
real_t scale, b_scale = 0;
#endif
#endif
ALIGN complex_t Z1[512];
complex_t *sincos = mdct->sincos;
uint16_t N = mdct->N;
uint16_t N2 = N >> 1;
uint16_t N4 = N >> 2;
uint16_t N8 = N >> 3;
#ifdef PROFILE
int64_t count1, count2 = faad_get_ts();
#endif
#ifdef ALLOW_SMALL_FRAMELENGTH
#ifdef FIXED_POINT
/* detect non-power of 2 */
if (N & (N-1))
{
/* adjust scale for non-power of 2 MDCT */
/* 2048/1920 */
b_scale = 1;
scale = COEF_CONST(1.0666666666666667);
}
#endif
#endif
/* pre-IFFT complex multiplication */
for (k = 0; k < N4; k++)
{
ComplexMult(&IM(Z1[k]), &RE(Z1[k]),
X_in[2*k], X_in[N2 - 1 - 2*k], RE(sincos[k]), IM(sincos[k]));
}
#ifdef PROFILE
count1 = faad_get_ts();
#endif
/* complex IFFT, any non-scaling FFT can be used here */
cfftb(mdct->cfft, Z1);
#ifdef PROFILE
count1 = faad_get_ts() - count1;
#endif
/* post-IFFT complex multiplication */
for (k = 0; k < N4; k++)
{
RE(x) = RE(Z1[k]);
IM(x) = IM(Z1[k]);
ComplexMult(&IM(Z1[k]), &RE(Z1[k]),
IM(x), RE(x), RE(sincos[k]), IM(sincos[k]));
#ifdef ALLOW_SMALL_FRAMELENGTH
#ifdef FIXED_POINT
/* non-power of 2 MDCT scaling */
if (b_scale)
{
RE(Z1[k]) = MUL_C(RE(Z1[k]), scale);
IM(Z1[k]) = MUL_C(IM(Z1[k]), scale);
}
#endif
#endif
}
/* reordering */
for (k = 0; k < N8; k+=2)
{
X_out[ 2*k] = IM(Z1[N8 + k]);
X_out[ 2 + 2*k] = IM(Z1[N8 + 1 + k]);
X_out[ 1 + 2*k] = -RE(Z1[N8 - 1 - k]);
X_out[ 3 + 2*k] = -RE(Z1[N8 - 2 - k]);
X_out[N4 + 2*k] = RE(Z1[ k]);
X_out[N4 + + 2 + 2*k] = RE(Z1[ 1 + k]);
X_out[N4 + 1 + 2*k] = -IM(Z1[N4 - 1 - k]);
X_out[N4 + 3 + 2*k] = -IM(Z1[N4 - 2 - k]);
X_out[N2 + 2*k] = RE(Z1[N8 + k]);
X_out[N2 + + 2 + 2*k] = RE(Z1[N8 + 1 + k]);
X_out[N2 + 1 + 2*k] = -IM(Z1[N8 - 1 - k]);
X_out[N2 + 3 + 2*k] = -IM(Z1[N8 - 2 - k]);
X_out[N2 + N4 + 2*k] = -IM(Z1[ k]);
X_out[N2 + N4 + 2 + 2*k] = -IM(Z1[ 1 + k]);
X_out[N2 + N4 + 1 + 2*k] = RE(Z1[N4 - 1 - k]);
X_out[N2 + N4 + 3 + 2*k] = RE(Z1[N4 - 2 - k]);
}
#ifdef PROFILE
count2 = faad_get_ts() - count2;
mdct->fft_cycles += count1;
mdct->cycles += (count2 - count1);
#endif
}
uint8_t reconstruct_single_channel(NeAACDecHandle hDecoder, ic_stream *ics,
element *sce, int16_t *spec_data)
{
uint8_t retval, output_channels;
#ifdef PROFILE
int64_t count = faad_get_ts();
#endif
/* always allocate 2 channels, PS can always "suddenly" turn up */
#if (defined(PS_DEC) || defined(DRM_PS))
output_channels = 2;
#else
output_channels = 1;
#endif
if (hDecoder->element_output_channels[hDecoder->fr_ch_ele] == 0)
{
/* element_output_channels not set yet */
hDecoder->element_output_channels[hDecoder->fr_ch_ele] = output_channels;
} else if (hDecoder->element_output_channels[hDecoder->fr_ch_ele] != output_channels) {
/* element inconsistency */
return 21;
}
if (hDecoder->element_alloced[hDecoder->fr_ch_ele] == 0)
{
retval = allocate_single_channel(hDecoder, sce->channel, output_channels);
if (retval > 0)
return retval;
hDecoder->element_alloced[hDecoder->fr_ch_ele] = 1;
}
/* dequantisation and scaling */
retval = quant_to_spec(hDecoder, ics, spec_data, spec_coef1, hDecoder->frameLength);
if (retval > 0)
return retval;
#ifdef PROFILE
count = faad_get_ts() - count;
hDecoder->requant_cycles += count;
#endif
/* pns decoding */
pns_decode(ics, NULL, spec_coef1, NULL, hDecoder->frameLength, 0, hDecoder->object_type);
#ifdef MAIN_DEC
/* MAIN object type prediction */
if (hDecoder->object_type == MAIN)
{
/* intra channel prediction */
ic_prediction(ics, spec_coef1, hDecoder->pred_stat[sce->channel], hDecoder->frameLength,
hDecoder->sf_index);
/* In addition, for scalefactor bands coded by perceptual
noise substitution the predictors belonging to the
corresponding spectral coefficients are reset.
*/
pns_reset_pred_state(ics, hDecoder->pred_stat[sce->channel]);
}
#endif
#ifdef LTP_DEC
if (is_ltp_ot(hDecoder->object_type))
{
#ifdef LD_DEC
if (hDecoder->object_type == LD)
{
if (ics->ltp.data_present)
{
if (ics->ltp.lag_update)
hDecoder->ltp_lag[sce->channel] = ics->ltp.lag;
}
ics->ltp.lag = hDecoder->ltp_lag[sce->channel];
}
#endif
/* long term prediction */
lt_prediction(ics, &(ics->ltp), spec_coef1, hDecoder->lt_pred_stat[sce->channel], hDecoder->fb,
ics->window_shape, hDecoder->window_shape_prev[sce->channel],
hDecoder->sf_index, hDecoder->object_type, hDecoder->frameLength);
}
#endif
/* tns decoding */
tns_decode_frame(ics, &(ics->tns), hDecoder->sf_index, hDecoder->object_type,
spec_coef1, hDecoder->frameLength);
/* drc decoding */
if (hDecoder->drc->present)
{
if (!hDecoder->drc->exclude_mask[sce->channel] || !hDecoder->drc->excluded_chns_present)
drc_decode(hDecoder->drc, spec_coef1);
}
/* filter bank */
#ifdef SSR_DEC
if (hDecoder->object_type != SSR)
{
#endif
ifilter_bank(ics->window_sequence,ics->window_shape,
hDecoder->window_shape_prev[sce->channel],spec_coef1,
hDecoder->time_out[sce->channel], hDecoder->fb_intermed[sce->channel],
hDecoder->object_type, hDecoder->frameLength);
#ifdef SSR_DEC
} else {
ssr_decode(&(ics->ssr), hDecoder->fb, ics->window_sequence, ics->window_shape,
hDecoder->window_shape_prev[sce->channel], spec_coef1, hDecoder->time_out[sce->channel],
hDecoder->ssr_overlap[sce->channel], hDecoder->ipqf_buffer[sce->channel], hDecoder->prev_fmd[sce->channel],
hDecoder->frameLength);
}
#endif
/* save window shape for next frame */
hDecoder->window_shape_prev[sce->channel] = ics->window_shape;
#ifdef LTP_DEC
if (is_ltp_ot(hDecoder->object_type))
{
lt_update_state(hDecoder->lt_pred_stat[sce->channel], hDecoder->time_out[sce->channel],
hDecoder->fb_intermed[sce->channel], hDecoder->frameLength, hDecoder->object_type);
}
#endif
#ifdef SBR_DEC
if (((hDecoder->sbr_present_flag == 1) || (hDecoder->forceUpSampling == 1))
&& hDecoder->sbr_alloced[hDecoder->fr_ch_ele])
{
uint8_t ele = hDecoder->fr_ch_ele;
uint8_t ch = sce->channel;
/* following case can happen when forceUpSampling == 1 */
if (hDecoder->sbr[ele] == NULL)
{
hDecoder->sbr[ele] = sbrDecodeInit(hDecoder->frameLength,
hDecoder->element_id[ele], 2*get_sample_rate(hDecoder->sf_index),
hDecoder->downSampledSBR
#ifdef DRM
, 0
#endif
);
}
if (sce->ics1.window_sequence == EIGHT_SHORT_SEQUENCE)
hDecoder->sbr[ele]->maxAACLine = 8*sce->ics1.swb_offset[max(sce->ics1.max_sfb-1, 0)];
else
hDecoder->sbr[ele]->maxAACLine = sce->ics1.swb_offset[max(sce->ics1.max_sfb-1, 0)];
/* check if any of the PS tools is used */
#if (defined(PS_DEC) || defined(DRM_PS))
if (hDecoder->ps_used[ele] == 0)
{
#endif
retval = sbrDecodeSingleFrame(hDecoder->sbr[ele], hDecoder->time_out[ch],
hDecoder->postSeekResetFlag, hDecoder->downSampledSBR);
#if (defined(PS_DEC) || defined(DRM_PS))
} else {
retval = sbrDecodeSingleFramePS(hDecoder->sbr[ele], hDecoder->time_out[ch],
hDecoder->time_out[ch+1], hDecoder->postSeekResetFlag,
hDecoder->downSampledSBR);
}
#endif
if (retval > 0)
return retval;
} else if (((hDecoder->sbr_present_flag == 1) || (hDecoder->forceUpSampling == 1))
&& !hDecoder->sbr_alloced[hDecoder->fr_ch_ele])
{
return 23;
}
#endif
/* copy L to R when no PS is used */
#if (defined(PS_DEC) || defined(DRM_PS))
if ((hDecoder->ps_used[hDecoder->fr_ch_ele] == 0))
{
uint8_t ele = hDecoder->fr_ch_ele;
uint8_t ch = sce->channel;
uint16_t frame_size = (hDecoder->sbr_alloced[ele]) ? 2 : 1;
frame_size *= hDecoder->frameLength*sizeof(real_t);
memcpy(hDecoder->time_out[ch+1], hDecoder->time_out[ch], frame_size);
}
#endif
return 0;
}
uint8_t reconstruct_channel_pair(NeAACDecHandle hDecoder, ic_stream *ics1, ic_stream *ics2,
element *cpe, int16_t *spec_data1, int16_t *spec_data2)
{
uint8_t retval;
#ifdef PROFILE
int64_t count = faad_get_ts();
#endif
if (hDecoder->element_alloced[hDecoder->fr_ch_ele] == 0)
{
retval = allocate_channel_pair(hDecoder, cpe->channel, (uint8_t)cpe->paired_channel);
if (retval > 0)
return retval;
hDecoder->element_alloced[hDecoder->fr_ch_ele] = 1;
}
/* dequantisation and scaling */
retval = quant_to_spec(hDecoder, ics1, spec_data1, spec_coef1, hDecoder->frameLength);
if (retval > 0)
return retval;
retval = quant_to_spec(hDecoder, ics2, spec_data2, spec_coef2, hDecoder->frameLength);
if (retval > 0)
return retval;
#ifdef PROFILE
count = faad_get_ts() - count;
hDecoder->requant_cycles += count;
#endif
/* pns decoding */
if (ics1->ms_mask_present)
{
pns_decode(ics1, ics2, spec_coef1, spec_coef2, hDecoder->frameLength, 1, hDecoder->object_type);
} else {
pns_decode(ics1, NULL, spec_coef1, NULL, hDecoder->frameLength, 0, hDecoder->object_type);
pns_decode(ics2, NULL, spec_coef2, NULL, hDecoder->frameLength, 0, hDecoder->object_type);
}
/* mid/side decoding */
ms_decode(ics1, ics2, spec_coef1, spec_coef2, hDecoder->frameLength);
#if 0
{
int i;
for (i = 0; i < 1024; i++)
{
//printf("%d\n", spec_coef1[i]);
printf("0x%.8X\n", spec_coef1[i]);
}
for (i = 0; i < 1024; i++)
{
//printf("%d\n", spec_coef2[i]);
printf("0x%.8X\n", spec_coef2[i]);
}
}
#endif
/* intensity stereo decoding */
is_decode(ics1, ics2, spec_coef1, spec_coef2, hDecoder->frameLength);
#if 0
{
int i;
for (i = 0; i < 1024; i++)
{
printf("%d\n", spec_coef1[i]);
//printf("0x%.8X\n", spec_coef1[i]);
}
for (i = 0; i < 1024; i++)
{
printf("%d\n", spec_coef2[i]);
//printf("0x%.8X\n", spec_coef2[i]);
}
}
#endif
#ifdef MAIN_DEC
/* MAIN object type prediction */
if (hDecoder->object_type == MAIN)
{
/* intra channel prediction */
ic_prediction(ics1, spec_coef1, hDecoder->pred_stat[cpe->channel], hDecoder->frameLength,
hDecoder->sf_index);
ic_prediction(ics2, spec_coef2, hDecoder->pred_stat[cpe->paired_channel], hDecoder->frameLength,
hDecoder->sf_index);
/* In addition, for scalefactor bands coded by perceptual
noise substitution the predictors belonging to the
corresponding spectral coefficients are reset.
*/
pns_reset_pred_state(ics1, hDecoder->pred_stat[cpe->channel]);
pns_reset_pred_state(ics2, hDecoder->pred_stat[cpe->paired_channel]);
}
#endif
#ifdef LTP_DEC
if (is_ltp_ot(hDecoder->object_type))
{
ltp_info *ltp1 = &(ics1->ltp);
ltp_info *ltp2 = (cpe->common_window) ? &(ics2->ltp2) : &(ics2->ltp);
#ifdef LD_DEC
if (hDecoder->object_type == LD)
{
if (ltp1->data_present)
{
if (ltp1->lag_update)
hDecoder->ltp_lag[cpe->channel] = ltp1->lag;
}
ltp1->lag = hDecoder->ltp_lag[cpe->channel];
if (ltp2->data_present)
{
if (ltp2->lag_update)
hDecoder->ltp_lag[cpe->paired_channel] = ltp2->lag;
}
ltp2->lag = hDecoder->ltp_lag[cpe->paired_channel];
}
#endif
/* long term prediction */
lt_prediction(ics1, ltp1, spec_coef1, hDecoder->lt_pred_stat[cpe->channel], hDecoder->fb,
ics1->window_shape, hDecoder->window_shape_prev[cpe->channel],
hDecoder->sf_index, hDecoder->object_type, hDecoder->frameLength);
lt_prediction(ics2, ltp2, spec_coef2, hDecoder->lt_pred_stat[cpe->paired_channel], hDecoder->fb,
ics2->window_shape, hDecoder->window_shape_prev[cpe->paired_channel],
hDecoder->sf_index, hDecoder->object_type, hDecoder->frameLength);
}
#endif
/* tns decoding */
tns_decode_frame(ics1, &(ics1->tns), hDecoder->sf_index, hDecoder->object_type,
spec_coef1, hDecoder->frameLength);
tns_decode_frame(ics2, &(ics2->tns), hDecoder->sf_index, hDecoder->object_type,
spec_coef2, hDecoder->frameLength);
/* drc decoding */
if (hDecoder->drc->present)
{
if (!hDecoder->drc->exclude_mask[cpe->channel] || !hDecoder->drc->excluded_chns_present)
drc_decode(hDecoder->drc, spec_coef1);
if (!hDecoder->drc->exclude_mask[cpe->paired_channel] || !hDecoder->drc->excluded_chns_present)
drc_decode(hDecoder->drc, spec_coef2);
}
/* filter bank */
#ifdef SSR_DEC
if (hDecoder->object_type != SSR)
{
#endif
ifilter_bank(ics1->window_sequence,ics1->window_shape,
hDecoder->window_shape_prev[cpe->channel],spec_coef1,
hDecoder->time_out[cpe->channel], hDecoder->fb_intermed[cpe->channel],
hDecoder->object_type, hDecoder->frameLength);
ifilter_bank(ics2->window_sequence,ics2->window_shape,
hDecoder->window_shape_prev[cpe->paired_channel], spec_coef2,
hDecoder->time_out[cpe->paired_channel], hDecoder->fb_intermed[cpe->paired_channel],
hDecoder->object_type, hDecoder->frameLength);
#ifdef SSR_DEC
} else {
ssr_decode(&(ics1->ssr), hDecoder->fb, ics1->window_sequence, ics1->window_shape,
hDecoder->window_shape_prev[cpe->channel], spec_coef1, hDecoder->time_out[cpe->channel],
hDecoder->ssr_overlap[cpe->channel], hDecoder->ipqf_buffer[cpe->channel],
hDecoder->prev_fmd[cpe->channel], hDecoder->frameLength);
ssr_decode(&(ics2->ssr), hDecoder->fb, ics2->window_sequence, ics2->window_shape,
hDecoder->window_shape_prev[cpe->paired_channel], spec_coef2, hDecoder->time_out[cpe->paired_channel],
hDecoder->ssr_overlap[cpe->paired_channel], hDecoder->ipqf_buffer[cpe->paired_channel],
hDecoder->prev_fmd[cpe->paired_channel], hDecoder->frameLength);
}
#endif
/* save window shape for next frame */
hDecoder->window_shape_prev[cpe->channel] = ics1->window_shape;
hDecoder->window_shape_prev[cpe->paired_channel] = ics2->window_shape;
#ifdef LTP_DEC
if (is_ltp_ot(hDecoder->object_type))
{
lt_update_state(hDecoder->lt_pred_stat[cpe->channel], hDecoder->time_out[cpe->channel],
hDecoder->fb_intermed[cpe->channel], hDecoder->frameLength, hDecoder->object_type);
lt_update_state(hDecoder->lt_pred_stat[cpe->paired_channel], hDecoder->time_out[cpe->paired_channel],
hDecoder->fb_intermed[cpe->paired_channel], hDecoder->frameLength, hDecoder->object_type);
}
#endif
#ifdef SBR_DEC
if (((hDecoder->sbr_present_flag == 1) || (hDecoder->forceUpSampling == 1))
&& hDecoder->sbr_alloced[hDecoder->fr_ch_ele])
{
uint8_t ele = hDecoder->fr_ch_ele;
uint8_t ch0 = cpe->channel;
uint8_t ch1 = cpe->paired_channel;
/* following case can happen when forceUpSampling == 1 */
if (hDecoder->sbr[ele] == NULL)
{
hDecoder->sbr[ele] = sbrDecodeInit(hDecoder->frameLength,
hDecoder->element_id[ele], 2*get_sample_rate(hDecoder->sf_index),
hDecoder->downSampledSBR
#ifdef DRM
, 0
#endif
);
}
if (cpe->ics1.window_sequence == EIGHT_SHORT_SEQUENCE)
hDecoder->sbr[ele]->maxAACLine = 8*cpe->ics1.swb_offset[max(cpe->ics1.max_sfb-1, 0)];
else
hDecoder->sbr[ele]->maxAACLine = cpe->ics1.swb_offset[max(cpe->ics1.max_sfb-1, 0)];
retval = sbrDecodeCoupleFrame(hDecoder->sbr[ele],
hDecoder->time_out[ch0], hDecoder->time_out[ch1],
hDecoder->postSeekResetFlag, hDecoder->downSampledSBR);
if (retval > 0)
return retval;
} else if (((hDecoder->sbr_present_flag == 1) || (hDecoder->forceUpSampling == 1))
&& !hDecoder->sbr_alloced[hDecoder->fr_ch_ele])
{
return 23;
}
#endif
return 0;
}
void ifilter_bank_sse(fb_info *fb, uint8_t window_sequence, uint8_t window_shape,
uint8_t window_shape_prev, real_t *freq_in,
real_t *time_out, uint8_t object_type, uint16_t frame_len)
{
int16_t i;
ALIGN real_t transf_buf[2*1024] = {0};
const real_t *window_long = NULL;
const real_t *window_long_prev = NULL;
const real_t *window_short = NULL;
const real_t *window_short_prev = NULL;
uint16_t nlong = frame_len;
uint16_t nshort = frame_len/8;
uint16_t trans = nshort/2;
uint16_t nflat_ls = (nlong-nshort)/2;
#ifdef PROFILE
int64_t count = faad_get_ts();
#endif
#ifdef LD_DEC
if (object_type == LD)
{
window_long = fb->ld_window[window_shape];
window_long_prev = fb->ld_window[window_shape_prev];
} else {
#endif
window_long = fb->long_window[window_shape];
window_long_prev = fb->long_window[window_shape_prev];
window_short = fb->short_window[window_shape];
window_short_prev = fb->short_window[window_shape_prev];
#ifdef LD_DEC
}
#endif
switch (window_sequence)
{
case ONLY_LONG_SEQUENCE:
imdct_long_sse(fb, freq_in, transf_buf, 2*nlong);
for (i = 0; i < nlong; i+=4)
{
__m128 m1, m2, m3, m4, m5, m6, m7, m8;
m1 = _mm_load_ps(&transf_buf[i]);
m2 = _mm_load_ps(&window_long_prev[i]);
m6 = _mm_load_ps(&window_long[nlong-4-i]);
m3 = _mm_load_ps(&time_out[nlong+i]);
m5 = _mm_load_ps(&transf_buf[nlong+i]);
m4 = _mm_mul_ps(m1, m2);
m7 = _mm_shuffle_ps(m6, m6, _MM_SHUFFLE(0, 1, 2, 3));
m4 = _mm_add_ps(m4, m3);
m8 = _mm_mul_ps(m5, m7);
_mm_store_ps(&time_out[i], m4);
_mm_store_ps(&time_out[nlong+i], m8);
}
break;
case LONG_START_SEQUENCE:
imdct_long_sse(fb, freq_in, transf_buf, 2*nlong);
for (i = 0; i < nlong; i+=4)
{
__m128 m1 = _mm_load_ps(&transf_buf[i]);
__m128 m2 = _mm_load_ps(&window_long_prev[i]);
__m128 m3 = _mm_load_ps(&time_out[nlong+i]);
__m128 m4 = _mm_mul_ps(m1, m2);
m4 = _mm_add_ps(m4, m3);
_mm_store_ps(&time_out[i], m4);
}
for (i = 0; i < nflat_ls; i+=4)
{
__m128 m1 = _mm_load_ps(&transf_buf[nlong+i]);
_mm_store_ps(&time_out[nlong+i], m1);
}
for (i = 0; i < nshort; i+=4)
{
__m128 m1 = _mm_load_ps(&transf_buf[nlong+nflat_ls+i]);
__m128 m2 = _mm_load_ps(&window_short[nshort-4-i]);
__m128 m3, m4;
m3 = _mm_shuffle_ps(m2, m2, _MM_SHUFFLE(0, 1, 2, 3));
m4 = _mm_mul_ps(m1, m3);
_mm_store_ps(&time_out[nlong+nflat_ls+i], m4);
}
for (i = 0; i < nflat_ls; i+=4)
{
__m128 m1 = _mm_setzero_ps();
_mm_store_ps(&time_out[nlong+nflat_ls+nshort+i], m1);
}
break;
case EIGHT_SHORT_SEQUENCE:
faad_imdct_sse(fb->mdct256, &freq_in[0*nshort], &transf_buf[2*nshort*0]);
faad_imdct_sse(fb->mdct256, &freq_in[1*nshort], &transf_buf[2*nshort*1]);
faad_imdct_sse(fb->mdct256, &freq_in[2*nshort], &transf_buf[2*nshort*2]);
faad_imdct_sse(fb->mdct256, &freq_in[3*nshort], &transf_buf[2*nshort*3]);
faad_imdct_sse(fb->mdct256, &freq_in[4*nshort], &transf_buf[2*nshort*4]);
faad_imdct_sse(fb->mdct256, &freq_in[5*nshort], &transf_buf[2*nshort*5]);
faad_imdct_sse(fb->mdct256, &freq_in[6*nshort], &transf_buf[2*nshort*6]);
faad_imdct_sse(fb->mdct256, &freq_in[7*nshort], &transf_buf[2*nshort*7]);
for (i = 0; i < nflat_ls; i+=4)
{
__m128 m1 = _mm_load_ps(&time_out[nlong+i]);
_mm_store_ps(&time_out[i], m1);
}
for (i = 0; i < nshort; i+=4)
{
__m128 m1 = _mm_load_ps(&transf_buf[nshort*0+i]);
__m128 m2 = _mm_load_ps(&window_short_prev[i]);
__m128 m3 = _mm_load_ps(&time_out[nlong+nflat_ls+i]);
__m128 m4 = _mm_mul_ps(m1, m2);
m4 = _mm_add_ps(m4, m3);
_mm_store_ps(&time_out[nflat_ls+i], m4);
}
for (i = 0; i < nshort; i+=4)
{
__m128 m1, m2, m3, m4, m5, m6, m7, m8;
m1 = _mm_load_ps(&transf_buf[nshort*1+i]);
m2 = _mm_load_ps(&window_short[nshort-4-i]);
m3 = _mm_load_ps(&time_out[nlong+nflat_ls+nshort*1+i]);
m6 = _mm_load_ps(&transf_buf[nshort*2+i]);
m7 = _mm_load_ps(&window_short[i]);
m5 = _mm_shuffle_ps(m2, m2, _MM_SHUFFLE(0, 1, 2, 3));
m4 = _mm_mul_ps(m1, m5);
m8 = _mm_mul_ps(m6, m7);
m4 = _mm_add_ps(m4, m3);
m4 = _mm_add_ps(m4, m8);
_mm_store_ps(&time_out[nflat_ls+1*nshort+i], m4);
}
for (i = 0; i < nshort; i+=4)
{
__m128 m1, m2, m3, m4, m5, m6, m7, m8;
m1 = _mm_load_ps(&transf_buf[nshort*3+i]);
m2 = _mm_load_ps(&window_short[nshort-4-i]);
m3 = _mm_load_ps(&time_out[nlong+nflat_ls+nshort*2+i]);
m6 = _mm_load_ps(&transf_buf[nshort*4+i]);
m7 = _mm_load_ps(&window_short[i]);
m5 = _mm_shuffle_ps(m2, m2, _MM_SHUFFLE(0, 1, 2, 3));
m4 = _mm_mul_ps(m1, m5);
m8 = _mm_mul_ps(m6, m7);
m4 = _mm_add_ps(m4, m3);
m4 = _mm_add_ps(m4, m8);
_mm_store_ps(&time_out[nflat_ls+2*nshort+i], m4);
}
for (i = 0; i < nshort; i+=4)
{
__m128 m1, m2, m3, m4, m5, m6, m7, m8;
m1 = _mm_load_ps(&transf_buf[nshort*5+i]);
m2 = _mm_load_ps(&window_short[nshort-4-i]);
m3 = _mm_load_ps(&time_out[nlong+nflat_ls+nshort*3+i]);
m6 = _mm_load_ps(&transf_buf[nshort*6+i]);
m7 = _mm_load_ps(&window_short[i]);
m5 = _mm_shuffle_ps(m2, m2, _MM_SHUFFLE(0, 1, 2, 3));
m4 = _mm_mul_ps(m1, m5);
m8 = _mm_mul_ps(m6, m7);
m4 = _mm_add_ps(m4, m3);
m4 = _mm_add_ps(m4, m8);
_mm_store_ps(&time_out[nflat_ls+3*nshort+i], m4);
}
for(i = 0; i < trans; i+=4)
{
__m128 m1, m2, m3, m4, m5, m6, m7, m8;
m1 = _mm_load_ps(&transf_buf[nshort*7+i]);
m2 = _mm_load_ps(&window_short[nshort-4-i]);
m3 = _mm_load_ps(&time_out[nlong+nflat_ls+nshort*4+i]);
m6 = _mm_load_ps(&transf_buf[nshort*8+i]);
m7 = _mm_load_ps(&window_short[i]);
m5 = _mm_shuffle_ps(m2, m2, _MM_SHUFFLE(0, 1, 2, 3));
m4 = _mm_mul_ps(m1, m5);
m8 = _mm_mul_ps(m6, m7);
m4 = _mm_add_ps(m4, m3);
m4 = _mm_add_ps(m4, m8);
_mm_store_ps(&time_out[nflat_ls+4*nshort+i], m4);
}
for (i = trans; i < nshort; i+=4)
{
__m128 m1, m2, m3, m4, m5, m6, m7, m8;
m1 = _mm_load_ps(&transf_buf[nshort*7+i]);
m2 = _mm_load_ps(&window_short[nshort-4-i]);
m6 = _mm_load_ps(&transf_buf[nshort*8+i]);
m7 = _mm_load_ps(&window_short[i]);
m5 = _mm_shuffle_ps(m2, m2, _MM_SHUFFLE(0, 1, 2, 3));
m4 = _mm_mul_ps(m1, m5);
m8 = _mm_mul_ps(m6, m7);
m3 = _mm_add_ps(m4, m8);
_mm_store_ps(&time_out[nflat_ls+4*nshort+i], m3);
}
for (i = 0; i < nshort; i+=4)
{
__m128 m1, m2, m3, m4, m5, m6, m7, m8;
m1 = _mm_load_ps(&transf_buf[nshort*9+i]);
m2 = _mm_load_ps(&window_short[nshort-4-i]);
m6 = _mm_load_ps(&transf_buf[nshort*10+i]);
m7 = _mm_load_ps(&window_short[i]);
m5 = _mm_shuffle_ps(m2, m2, _MM_SHUFFLE(0, 1, 2, 3));
m4 = _mm_mul_ps(m1, m5);
m8 = _mm_mul_ps(m6, m7);
m3 = _mm_add_ps(m4, m8);
_mm_store_ps(&time_out[nflat_ls+5*nshort+i], m3);
}
for (i = 0; i < nshort; i+=4)
{
__m128 m1, m2, m3, m4, m5, m6, m7, m8;
m1 = _mm_load_ps(&transf_buf[nshort*11+i]);
m2 = _mm_load_ps(&window_short[nshort-4-i]);
m6 = _mm_load_ps(&transf_buf[nshort*12+i]);
m7 = _mm_load_ps(&window_short[i]);
m5 = _mm_shuffle_ps(m2, m2, _MM_SHUFFLE(0, 1, 2, 3));
m4 = _mm_mul_ps(m1, m5);
m8 = _mm_mul_ps(m6, m7);
m3 = _mm_add_ps(m4, m8);
_mm_store_ps(&time_out[nflat_ls+6*nshort+i], m3);
}
for (i = 0; i < nshort; i+=4)
{
__m128 m1, m2, m3, m4, m5, m6, m7, m8;
m1 = _mm_load_ps(&transf_buf[nshort*13+i]);
m2 = _mm_load_ps(&window_short[nshort-4-i]);
m6 = _mm_load_ps(&transf_buf[nshort*14+i]);
m7 = _mm_load_ps(&window_short[i]);
m5 = _mm_shuffle_ps(m2, m2, _MM_SHUFFLE(0, 1, 2, 3));
m4 = _mm_mul_ps(m1, m5);
m8 = _mm_mul_ps(m6, m7);
m3 = _mm_add_ps(m4, m8);
_mm_store_ps(&time_out[nflat_ls+7*nshort+i], m3);
}
for (i = 0; i < nshort; i+=4)
{
__m128 m1, m2, m3, m5;
m1 = _mm_load_ps(&transf_buf[nshort*15+i]);
m2 = _mm_load_ps(&window_short[nshort-4-i]);
m5 = _mm_shuffle_ps(m2, m2, _MM_SHUFFLE(0, 1, 2, 3));
m3 = _mm_mul_ps(m1, m5);
_mm_store_ps(&time_out[nflat_ls+8*nshort+i], m3);
}
for (i = 0; i < nflat_ls; i+=4)
{
__m128 m1 = _mm_setzero_ps();
_mm_store_ps(&time_out[nlong+nflat_ls+nshort+i], m1);
}
break;
case LONG_STOP_SEQUENCE:
imdct_long_sse(fb, freq_in, transf_buf, 2*nlong);
for (i = 0; i < nflat_ls; i+=4)
{
__m128 m1 = _mm_load_ps(&time_out[nlong+i]);
_mm_store_ps(&time_out[i], m1);
}
for (i = 0; i < nshort; i+=4)
{
__m128 m1 = _mm_load_ps(&transf_buf[nflat_ls+i]);
__m128 m2 = _mm_load_ps(&window_short_prev[i]);
__m128 m3 = _mm_load_ps(&time_out[nlong+nflat_ls+i]);
__m128 m4 = _mm_mul_ps(m1, m2);
m4 = _mm_add_ps(m4, m3);
_mm_store_ps(&time_out[nflat_ls+i], m4);
}
for (i = 0; i < nflat_ls; i+=4)
{
__m128 m1 = _mm_load_ps(&transf_buf[nflat_ls+nshort+i]);
__m128 m2 = _mm_load_ps(&time_out[nlong+nflat_ls+nshort+i]);
__m128 m3 = _mm_add_ps(m1, m2);
_mm_store_ps(&time_out[nflat_ls+nshort+i], m3);
}
for (i = 0; i < nlong; i+=4)
{
__m128 m1 = _mm_load_ps(&transf_buf[nlong+i]);
__m128 m2 = _mm_load_ps(&window_long[nlong-4-i]);
__m128 m3, m4;
m3 = _mm_shuffle_ps(m2, m2, _MM_SHUFFLE(0, 1, 2, 3));
m4 = _mm_mul_ps(m1, m3);
_mm_store_ps(&time_out[nlong+i], m4);
}
break;
}
#ifdef PROFILE
count = faad_get_ts() - count;
fb->cycles += count;
#endif
}
void ifilter_bank(fb_info *fb, uint8_t window_sequence, uint8_t window_shape,
uint8_t window_shape_prev, real_t *freq_in,
real_t *time_out, real_t *overlap,
uint8_t object_type, uint16_t frame_len)
{
int16_t i;
ALIGN real_t transf_buf[2*1024] = {0};
const real_t *window_long = NULL;
const real_t *window_long_prev = NULL;
const real_t *window_short = NULL;
const real_t *window_short_prev = NULL;
uint16_t nlong = frame_len;
uint16_t nshort = frame_len/8;
uint16_t trans = nshort/2;
uint16_t nflat_ls = (nlong-nshort)/2;
#ifdef PROFILE
int64_t count = faad_get_ts();
#endif
#ifdef LD_DEC
if (object_type == LD)
{
window_long = fb->ld_window[window_shape];
window_long_prev = fb->ld_window[window_shape_prev];
} else {
#endif
window_long = fb->long_window[window_shape];
window_long_prev = fb->long_window[window_shape_prev];
window_short = fb->short_window[window_shape];
window_short_prev = fb->short_window[window_shape_prev];
#ifdef LD_DEC
}
#endif
switch (window_sequence)
{
case ONLY_LONG_SEQUENCE:
imdct_long(fb, freq_in, transf_buf, 2*nlong);
for (i = 0; i < nlong; i+=4)
{
time_out[i] = overlap[i] + MUL_F(transf_buf[i],window_long_prev[i]);
time_out[i+1] = overlap[i+1] + MUL_F(transf_buf[i+1],window_long_prev[i+1]);
time_out[i+2] = overlap[i+2] + MUL_F(transf_buf[i+2],window_long_prev[i+2]);
time_out[i+3] = overlap[i+3] + MUL_F(transf_buf[i+3],window_long_prev[i+3]);
}
for (i = 0; i < nlong; i+=4)
{
overlap[i] = MUL_F(transf_buf[nlong+i],window_long[nlong-1-i]);
overlap[i+1] = MUL_F(transf_buf[nlong+i+1],window_long[nlong-2-i]);
overlap[i+2] = MUL_F(transf_buf[nlong+i+2],window_long[nlong-3-i]);
overlap[i+3] = MUL_F(transf_buf[nlong+i+3],window_long[nlong-4-i]);
}
break;
case LONG_START_SEQUENCE:
imdct_long(fb, freq_in, transf_buf, 2*nlong);
for (i = 0; i < nlong; i+=4)
{
time_out[i] = overlap[i] + MUL_F(transf_buf[i],window_long_prev[i]);
time_out[i+1] = overlap[i+1] + MUL_F(transf_buf[i+1],window_long_prev[i+1]);
time_out[i+2] = overlap[i+2] + MUL_F(transf_buf[i+2],window_long_prev[i+2]);
time_out[i+3] = overlap[i+3] + MUL_F(transf_buf[i+3],window_long_prev[i+3]);
}
for (i = 0; i < nflat_ls; i++)
overlap[i] = transf_buf[nlong+i];
for (i = 0; i < nshort; i++)
overlap[nflat_ls+i] = MUL_F(transf_buf[nlong+nflat_ls+i],window_short[nshort-i-1]);
for (i = 0; i < nflat_ls; i++)
overlap[nflat_ls+nshort+i] = 0;
break;
case EIGHT_SHORT_SEQUENCE:
faad_imdct(fb->mdct256, freq_in+0*nshort, transf_buf+2*nshort*0);
faad_imdct(fb->mdct256, freq_in+1*nshort, transf_buf+2*nshort*1);
faad_imdct(fb->mdct256, freq_in+2*nshort, transf_buf+2*nshort*2);
faad_imdct(fb->mdct256, freq_in+3*nshort, transf_buf+2*nshort*3);
faad_imdct(fb->mdct256, freq_in+4*nshort, transf_buf+2*nshort*4);
faad_imdct(fb->mdct256, freq_in+5*nshort, transf_buf+2*nshort*5);
faad_imdct(fb->mdct256, freq_in+6*nshort, transf_buf+2*nshort*6);
faad_imdct(fb->mdct256, freq_in+7*nshort, transf_buf+2*nshort*7);
for (i = 0; i < nflat_ls; i++)
time_out[i] = overlap[i];
for(i = 0; i < nshort; i++)
{
time_out[nflat_ls+ i] = overlap[nflat_ls+ i] + MUL_F(transf_buf[nshort*0+i],window_short_prev[i]);
time_out[nflat_ls+1*nshort+i] = overlap[nflat_ls+nshort*1+i] + MUL_F(transf_buf[nshort*1+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*2+i],window_short[i]);
time_out[nflat_ls+2*nshort+i] = overlap[nflat_ls+nshort*2+i] + MUL_F(transf_buf[nshort*3+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*4+i],window_short[i]);
time_out[nflat_ls+3*nshort+i] = overlap[nflat_ls+nshort*3+i] + MUL_F(transf_buf[nshort*5+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*6+i],window_short[i]);
if (i < trans)
time_out[nflat_ls+4*nshort+i] = overlap[nflat_ls+nshort*4+i] + MUL_F(transf_buf[nshort*7+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*8+i],window_short[i]);
}
for(i = 0; i < nshort; i++)
{
if (i >= trans)
overlap[nflat_ls+4*nshort+i-nlong] = MUL_F(transf_buf[nshort*7+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*8+i],window_short[i]);
overlap[nflat_ls+5*nshort+i-nlong] = MUL_F(transf_buf[nshort*9+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*10+i],window_short[i]);
overlap[nflat_ls+6*nshort+i-nlong] = MUL_F(transf_buf[nshort*11+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*12+i],window_short[i]);
overlap[nflat_ls+7*nshort+i-nlong] = MUL_F(transf_buf[nshort*13+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*14+i],window_short[i]);
overlap[nflat_ls+8*nshort+i-nlong] = MUL_F(transf_buf[nshort*15+i],window_short[nshort-1-i]);
}
for (i = 0; i < nflat_ls; i++)
overlap[nflat_ls+nshort+i] = 0;
break;
case LONG_STOP_SEQUENCE:
imdct_long(fb, freq_in, transf_buf, 2*nlong);
for (i = 0; i < nflat_ls; i++)
time_out[i] = overlap[i];
for (i = 0; i < nshort; i++)
time_out[nflat_ls+i] = overlap[nflat_ls+i] + MUL_F(transf_buf[nflat_ls+i],window_short_prev[i]);
for (i = 0; i < nflat_ls; i++)
time_out[nflat_ls+nshort+i] = overlap[nflat_ls+nshort+i] + transf_buf[nflat_ls+nshort+i];
for (i = 0; i < nlong; i++)
overlap[i] = MUL_F(transf_buf[nlong+i],window_long[nlong-1-i]);
break;
}
#ifdef PROFILE
count = faad_get_ts() - count;
fb->cycles += count;
#endif
}
void ifilter_bank(fb_info *fb, uint8_t window_sequence, uint8_t window_shape,
uint8_t window_shape_prev, real_t *freq_in,
real_t *time_out, real_t *overlap,
uint8_t object_type, uint16_t frame_len)
{
int16_t i;
ALIGN real_t transf_buf[2*1024] = {0};
const real_t *window_long = NULL;
const real_t *window_long_prev = NULL;
const real_t *window_short = NULL;
const real_t *window_short_prev = NULL;
uint16_t nlong = frame_len;
uint16_t nshort = frame_len/8;
uint16_t trans = nshort/2;
uint16_t nflat_ls = (nlong-nshort)/2;
#ifdef PROFILE
int64_t count = faad_get_ts();
#endif
/* select windows of current frame and previous frame (Sine or KBD) */
#ifdef LD_DEC
if (object_type == LD)
{
window_long = fb->ld_window[window_shape];
window_long_prev = fb->ld_window[window_shape_prev];
} else {
#endif
window_long = fb->long_window[window_shape];
window_long_prev = fb->long_window[window_shape_prev];
window_short = fb->short_window[window_shape];
window_short_prev = fb->short_window[window_shape_prev];
#ifdef LD_DEC
}
#endif
#if 0
for (i = 0; i < 1024; i++)
{
printf("%d\n", freq_in[i]);
}
#endif
#if 0
printf("%d %d\n", window_sequence, window_shape);
#endif
switch (window_sequence)
{
case ONLY_LONG_SEQUENCE:
/* perform iMDCT */
imdct_long(fb, freq_in, transf_buf, 2*nlong);
/* add second half output of previous frame to windowed output of current frame */
for (i = 0; i < nlong; i+=4)
{
time_out[i] = overlap[i] + MUL_F(transf_buf[i],window_long_prev[i]);
time_out[i+1] = overlap[i+1] + MUL_F(transf_buf[i+1],window_long_prev[i+1]);
time_out[i+2] = overlap[i+2] + MUL_F(transf_buf[i+2],window_long_prev[i+2]);
time_out[i+3] = overlap[i+3] + MUL_F(transf_buf[i+3],window_long_prev[i+3]);
}
/* window the second half and save as overlap for next frame */
for (i = 0; i < nlong; i+=4)
{
overlap[i] = MUL_F(transf_buf[nlong+i],window_long[nlong-1-i]);
overlap[i+1] = MUL_F(transf_buf[nlong+i+1],window_long[nlong-2-i]);
overlap[i+2] = MUL_F(transf_buf[nlong+i+2],window_long[nlong-3-i]);
overlap[i+3] = MUL_F(transf_buf[nlong+i+3],window_long[nlong-4-i]);
}
break;
case LONG_START_SEQUENCE:
/* perform iMDCT */
imdct_long(fb, freq_in, transf_buf, 2*nlong);
/* add second half output of previous frame to windowed output of current frame */
for (i = 0; i < nlong; i+=4)
{
time_out[i] = overlap[i] + MUL_F(transf_buf[i],window_long_prev[i]);
time_out[i+1] = overlap[i+1] + MUL_F(transf_buf[i+1],window_long_prev[i+1]);
time_out[i+2] = overlap[i+2] + MUL_F(transf_buf[i+2],window_long_prev[i+2]);
time_out[i+3] = overlap[i+3] + MUL_F(transf_buf[i+3],window_long_prev[i+3]);
}
/* window the second half and save as overlap for next frame */
/* construct second half window using padding with 1's and 0's */
for (i = 0; i < nflat_ls; i++)
overlap[i] = transf_buf[nlong+i];
for (i = 0; i < nshort; i++)
overlap[nflat_ls+i] = MUL_F(transf_buf[nlong+nflat_ls+i],window_short[nshort-i-1]);
for (i = 0; i < nflat_ls; i++)
overlap[nflat_ls+nshort+i] = 0;
break;
case EIGHT_SHORT_SEQUENCE:
/* perform iMDCT for each short block */
faad_imdct(fb->mdct256, freq_in+0*nshort, transf_buf+2*nshort*0);
faad_imdct(fb->mdct256, freq_in+1*nshort, transf_buf+2*nshort*1);
faad_imdct(fb->mdct256, freq_in+2*nshort, transf_buf+2*nshort*2);
faad_imdct(fb->mdct256, freq_in+3*nshort, transf_buf+2*nshort*3);
faad_imdct(fb->mdct256, freq_in+4*nshort, transf_buf+2*nshort*4);
faad_imdct(fb->mdct256, freq_in+5*nshort, transf_buf+2*nshort*5);
faad_imdct(fb->mdct256, freq_in+6*nshort, transf_buf+2*nshort*6);
faad_imdct(fb->mdct256, freq_in+7*nshort, transf_buf+2*nshort*7);
/* add second half output of previous frame to windowed output of current frame */
for (i = 0; i < nflat_ls; i++)
time_out[i] = overlap[i];
for(i = 0; i < nshort; i++)
{
time_out[nflat_ls+ i] = overlap[nflat_ls+ i] + MUL_F(transf_buf[nshort*0+i],window_short_prev[i]);
time_out[nflat_ls+1*nshort+i] = overlap[nflat_ls+nshort*1+i] + MUL_F(transf_buf[nshort*1+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*2+i],window_short[i]);
time_out[nflat_ls+2*nshort+i] = overlap[nflat_ls+nshort*2+i] + MUL_F(transf_buf[nshort*3+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*4+i],window_short[i]);
time_out[nflat_ls+3*nshort+i] = overlap[nflat_ls+nshort*3+i] + MUL_F(transf_buf[nshort*5+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*6+i],window_short[i]);
if (i < trans)
time_out[nflat_ls+4*nshort+i] = overlap[nflat_ls+nshort*4+i] + MUL_F(transf_buf[nshort*7+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*8+i],window_short[i]);
}
/* window the second half and save as overlap for next frame */
for(i = 0; i < nshort; i++)
{
if (i >= trans)
overlap[nflat_ls+4*nshort+i-nlong] = MUL_F(transf_buf[nshort*7+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*8+i],window_short[i]);
overlap[nflat_ls+5*nshort+i-nlong] = MUL_F(transf_buf[nshort*9+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*10+i],window_short[i]);
overlap[nflat_ls+6*nshort+i-nlong] = MUL_F(transf_buf[nshort*11+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*12+i],window_short[i]);
overlap[nflat_ls+7*nshort+i-nlong] = MUL_F(transf_buf[nshort*13+i],window_short[nshort-1-i]) + MUL_F(transf_buf[nshort*14+i],window_short[i]);
overlap[nflat_ls+8*nshort+i-nlong] = MUL_F(transf_buf[nshort*15+i],window_short[nshort-1-i]);
}
for (i = 0; i < nflat_ls; i++)
overlap[nflat_ls+nshort+i] = 0;
break;
case LONG_STOP_SEQUENCE:
/* perform iMDCT */
imdct_long(fb, freq_in, transf_buf, 2*nlong);
/* add second half output of previous frame to windowed output of current frame */
/* construct first half window using padding with 1's and 0's */
for (i = 0; i < nflat_ls; i++)
time_out[i] = overlap[i];
for (i = 0; i < nshort; i++)
time_out[nflat_ls+i] = overlap[nflat_ls+i] + MUL_F(transf_buf[nflat_ls+i],window_short_prev[i]);
for (i = 0; i < nflat_ls; i++)
time_out[nflat_ls+nshort+i] = overlap[nflat_ls+nshort+i] + transf_buf[nflat_ls+nshort+i];
/* window the second half and save as overlap for next frame */
for (i = 0; i < nlong; i++)
overlap[i] = MUL_F(transf_buf[nlong+i],window_long[nlong-1-i]);
break;
}
#if 0
for (i = 0; i < 1024; i++)
{
printf("%d\n", time_out[i]);
//printf("0x%.8X\n", time_out[i]);
}
#endif
#ifdef PROFILE
count = faad_get_ts() - count;
fb->cycles += count;
#endif
}
void faad_imdct_sse(mdct_info *mdct, real_t *X_in, real_t *X_out)
{
uint16_t k;
ALIGN complex_t Z1[512];
complex_t *sincos = mdct->sincos;
uint16_t N = mdct->N;
uint16_t N2 = N >> 1;
uint16_t N4 = N >> 2;
uint16_t N8 = N >> 3;
#ifdef PROFILE
int64_t count1, count2 = faad_get_ts();
#endif
/* pre-IFFT complex multiplication */
for (k = 0; k < N4; k+=4)
{
__m128 m12, m13, m14, m0, m1, m2, m3, m4, m5, m6, m7, m8, m9, m10, m11;
__m128 n12, n13, n14, n0, n1, n2, n3, n4, n5, n6, n7, n8, n9, n10, n11;
n12 = _mm_load_ps(&X_in[N2 - 2*k - 8]);
m12 = _mm_load_ps(&X_in[N2 - 2*k - 4]);
m13 = _mm_load_ps(&X_in[2*k]);
n13 = _mm_load_ps(&X_in[2*k + 4]);
m1 = _mm_load_ps(&RE(sincos[k]));
n1 = _mm_load_ps(&RE(sincos[k+2]));
m0 = _mm_shuffle_ps(m12, m13, _MM_SHUFFLE(2,0,1,3));
m2 = _mm_shuffle_ps(m1, m1, _MM_SHUFFLE(2,3,0,1));
m14 = _mm_shuffle_ps(m0, m0, _MM_SHUFFLE(3,1,2,0));
n0 = _mm_shuffle_ps(n12, n13, _MM_SHUFFLE(2,0,1,3));
n2 = _mm_shuffle_ps(n1, n1, _MM_SHUFFLE(2,3,0,1));
n14 = _mm_shuffle_ps(n0, n0, _MM_SHUFFLE(3,1,2,0));
m3 = _mm_mul_ps(m14, m1);
n3 = _mm_mul_ps(n14, n1);
m4 = _mm_mul_ps(m14, m2);
n4 = _mm_mul_ps(n14, n2);
m5 = _mm_shuffle_ps(m3, m4, _MM_SHUFFLE(2,0,2,0));
n5 = _mm_shuffle_ps(n3, n4, _MM_SHUFFLE(2,0,2,0));
m6 = _mm_shuffle_ps(m3, m4, _MM_SHUFFLE(3,1,3,1));
n6 = _mm_shuffle_ps(n3, n4, _MM_SHUFFLE(3,1,3,1));
m7 = _mm_add_ps(m5, m6);
n7 = _mm_add_ps(n5, n6);
m8 = _mm_sub_ps(m5, m6);
n8 = _mm_sub_ps(n5, n6);
m9 = _mm_shuffle_ps(m7, m7, _MM_SHUFFLE(3,2,3,2));
n9 = _mm_shuffle_ps(n7, n7, _MM_SHUFFLE(3,2,3,2));
m10 = _mm_shuffle_ps(m8, m8, _MM_SHUFFLE(1,0,1,0));
n10 = _mm_shuffle_ps(n8, n8, _MM_SHUFFLE(1,0,1,0));
m11 = _mm_unpacklo_ps(m10, m9);
n11 = _mm_unpacklo_ps(n10, n9);
_mm_store_ps(&RE(Z1[k]), m11);
_mm_store_ps(&RE(Z1[k+2]), n11);
}
#ifdef PROFILE
count1 = faad_get_ts();
#endif
/* complex IFFT, any non-scaling FFT can be used here */
cfftb_sse(mdct->cfft, Z1);
#ifdef PROFILE
count1 = faad_get_ts() - count1;
#endif
/* post-IFFT complex multiplication */
for (k = 0; k < N4; k+=4)
{
__m128 m0, m1, m2, m3, m4, m5, m6, m7, m8, m9, m10, m11;
__m128 n0, n1, n2, n3, n4, n5, n6, n7, n8, n9, n10, n11;
m0 = _mm_load_ps(&RE(Z1[k]));
n0 = _mm_load_ps(&RE(Z1[k+2]));
m1 = _mm_load_ps(&RE(sincos[k]));
n1 = _mm_load_ps(&RE(sincos[k+2]));
m2 = _mm_shuffle_ps(m1, m1, _MM_SHUFFLE(2,3,0,1));
n2 = _mm_shuffle_ps(n1, n1, _MM_SHUFFLE(2,3,0,1));
m3 = _mm_mul_ps(m0, m1);
n3 = _mm_mul_ps(n0, n1);
m4 = _mm_mul_ps(m0, m2);
n4 = _mm_mul_ps(n0, n2);
m5 = _mm_shuffle_ps(m3, m4, _MM_SHUFFLE(2,0,2,0));
n5 = _mm_shuffle_ps(n3, n4, _MM_SHUFFLE(2,0,2,0));
m6 = _mm_shuffle_ps(m3, m4, _MM_SHUFFLE(3,1,3,1));
n6 = _mm_shuffle_ps(n3, n4, _MM_SHUFFLE(3,1,3,1));
m7 = _mm_add_ps(m5, m6);
n7 = _mm_add_ps(n5, n6);
m8 = _mm_sub_ps(m5, m6);
n8 = _mm_sub_ps(n5, n6);
m9 = _mm_shuffle_ps(m7, m7, _MM_SHUFFLE(3,2,3,2));
n9 = _mm_shuffle_ps(n7, n7, _MM_SHUFFLE(3,2,3,2));
m10 = _mm_shuffle_ps(m8, m8, _MM_SHUFFLE(1,0,1,0));
n10 = _mm_shuffle_ps(n8, n8, _MM_SHUFFLE(1,0,1,0));
m11 = _mm_unpacklo_ps(m10, m9);
n11 = _mm_unpacklo_ps(n10, n9);
_mm_store_ps(&RE(Z1[k]), m11);
_mm_store_ps(&RE(Z1[k+2]), n11);
}
/* reordering */
for (k = 0; k < N8; k+=2)
{
__m128 m0, m1, m2, m3, m4, m5, m6, m7, m8, m9, m10, m11, m13;
__m128 n4, n5, n6, n7, n8, n9;
__m128 neg1 = _mm_set_ps(-1.0, 1.0, -1.0, 1.0);
__m128 neg2 = _mm_set_ps(-1.0, -1.0, -1.0, -1.0);
m0 = _mm_load_ps(&RE(Z1[k]));
m1 = _mm_load_ps(&RE(Z1[N8 - 2 - k]));
m2 = _mm_load_ps(&RE(Z1[N8 + k]));
m3 = _mm_load_ps(&RE(Z1[N4 - 2 - k]));
m10 = _mm_mul_ps(m0, neg1);
m11 = _mm_mul_ps(m1, neg2);
m13 = _mm_mul_ps(m3, neg1);
m5 = _mm_shuffle_ps(m2, m2, _MM_SHUFFLE(3,1,2,0));
n4 = _mm_shuffle_ps(m10, m10, _MM_SHUFFLE(3,1,2,0));
m4 = _mm_shuffle_ps(m11, m11, _MM_SHUFFLE(3,1,2,0));
n5 = _mm_shuffle_ps(m13, m13, _MM_SHUFFLE(3,1,2,0));
m6 = _mm_shuffle_ps(m4, m5, _MM_SHUFFLE(3,2,1,0));
n6 = _mm_shuffle_ps(n4, n5, _MM_SHUFFLE(3,2,1,0));
m7 = _mm_shuffle_ps(m5, m4, _MM_SHUFFLE(3,2,1,0));
n7 = _mm_shuffle_ps(n5, n4, _MM_SHUFFLE(3,2,1,0));
m8 = _mm_shuffle_ps(m6, m6, _MM_SHUFFLE(0,3,1,2));
n8 = _mm_shuffle_ps(n6, n6, _MM_SHUFFLE(2,1,3,0));
m9 = _mm_shuffle_ps(m7, m7, _MM_SHUFFLE(2,1,3,0));
n9 = _mm_shuffle_ps(n7, n7, _MM_SHUFFLE(0,3,1,2));
_mm_store_ps(&X_out[2*k], m8);
_mm_store_ps(&X_out[N4 + 2*k], n8);
_mm_store_ps(&X_out[N2 + 2*k], m9);
_mm_store_ps(&X_out[N2 + N4 + 2*k], n9);
}
#ifdef PROFILE
count2 = faad_get_ts() - count2;
mdct->fft_cycles += count1;
mdct->cycles += (count2 - count1);
#endif
}
void faad_imdct(mdct_info *mdct, real_t *X_in, real_t *X_out)
{
uint16_t k;
complex_t x;
ALIGN complex_t Z1[512];
complex_t *sincos = mdct->sincos;
uint16_t N = mdct->N;
uint16_t N2 = N >> 1;
uint16_t N4 = N >> 2;
uint16_t N8 = N >> 3;
#ifdef PROFILE
int64_t count1, count2 = faad_get_ts();
#endif
/* pre-IFFT complex multiplication */
for (k = 0; k < N4; k++)
{
ComplexMult(&IM(Z1[k]), &RE(Z1[k]),
X_in[2*k], X_in[N2 - 1 - 2*k], RE(sincos[k]), IM(sincos[k]));
}
#ifdef PROFILE
count1 = faad_get_ts();
#endif
/* complex IFFT, any non-scaling FFT can be used here */
cfftb(mdct->cfft, Z1);
#ifdef PROFILE
count1 = faad_get_ts() - count1;
#endif
/* post-IFFT complex multiplication */
for (k = 0; k < N4; k++)
{
RE(x) = RE(Z1[k]);
IM(x) = IM(Z1[k]);
ComplexMult(&IM(Z1[k]), &RE(Z1[k]),
IM(x), RE(x), RE(sincos[k]), IM(sincos[k]));
}
/* reordering */
for (k = 0; k < N8; k+=2)
{
X_out[ 2*k] = IM(Z1[N8 + k]);
X_out[ 2 + 2*k] = IM(Z1[N8 + 1 + k]);
X_out[ 1 + 2*k] = -RE(Z1[N8 - 1 - k]);
X_out[ 3 + 2*k] = -RE(Z1[N8 - 2 - k]);
X_out[N4 + 2*k] = RE(Z1[ k]);
X_out[N4 + + 2 + 2*k] = RE(Z1[ 1 + k]);
X_out[N4 + 1 + 2*k] = -IM(Z1[N4 - 1 - k]);
X_out[N4 + 3 + 2*k] = -IM(Z1[N4 - 2 - k]);
X_out[N2 + 2*k] = RE(Z1[N8 + k]);
X_out[N2 + + 2 + 2*k] = RE(Z1[N8 + 1 + k]);
X_out[N2 + 1 + 2*k] = -IM(Z1[N8 - 1 - k]);
X_out[N2 + 3 + 2*k] = -IM(Z1[N8 - 2 - k]);
X_out[N2 + N4 + 2*k] = -IM(Z1[ k]);
X_out[N2 + N4 + 2 + 2*k] = -IM(Z1[ 1 + k]);
X_out[N2 + N4 + 1 + 2*k] = RE(Z1[N4 - 1 - k]);
X_out[N2 + N4 + 3 + 2*k] = RE(Z1[N4 - 2 - k]);
}
#ifdef PROFILE
count2 = faad_get_ts() - count2;
mdct->fft_cycles += count1;
mdct->cycles += (count2 - count1);
#endif
}
