view libfaad2/filtbank.c @ 17588:79081ba52e00

Move the v{Y,C}CoeffsBank vectors into the SwsContext, filling them in just once when the scaler is initialized, instead of building them and freeing them over and over. This gives massive performance improvements. patch by Alan Curry, pacman*at*TheWorld*dot*com

/*
** FAAD2 - Freeware Advanced Audio (AAC) Decoder including SBR decoding
** Copyright (C) 2003-2004 M. Bakker, Ahead Software AG, http://www.nero.com
**  
** This program is free software; you can redistribute it and/or modify
** it under the terms of the GNU General Public License as published by
** the Free Software Foundation; either version 2 of the License, or
** (at your option) any later version.
** 
** This program is distributed in the hope that it will be useful,
** but WITHOUT ANY WARRANTY; without even the implied warranty of
** MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
** GNU General Public License for more details.
** 
** You should have received a copy of the GNU General Public License
** along with this program; if not, write to the Free Software 
** Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
**
** Any non-GPL usage of this software or parts of this software is strictly
** forbidden.
**
** Commercial non-GPL licensing of this software is possible.
** For more info contact Ahead Software through [e-mail].
**
** $Id: filtbank.c,v 1.38 2004/06/30 12:45:56 menno Exp $
**/

#include "common.h"
#include "structs.h"

#include <stdlib.h>
#include <string.h>
#ifdef _WIN32_WCE
#define assert(x)
#else
#include <assert.h>
#endif

#include "filtbank.h"
#include "decoder.h"
#include "syntax.h"
#include "kbd_win.h"
#include "sine_win.h"
#include "mdct.h"


fb_info *filter_bank_init(uint16_t frame_len)
{
    uint16_t nshort = frame_len/8;
#ifdef LD_DEC
    uint16_t frame_len_ld = frame_len/2;
#endif

    fb_info *fb = (fb_info*)faad_malloc(sizeof(fb_info));
    memset(fb, 0, sizeof(fb_info));

    /* normal */
    fb->mdct256 = faad_mdct_init(2*nshort);
    fb->mdct2048 = faad_mdct_init(2*frame_len);
#ifdef LD_DEC
    /* LD */
    fb->mdct1024 = faad_mdct_init(2*frame_len_ld);
#endif

#ifdef ALLOW_SMALL_FRAMELENGTH
    if (frame_len == 1024)
    {
#endif
        fb->long_window[0]  = sine_long_1024;
        fb->short_window[0] = sine_short_128;
        fb->long_window[1]  = kbd_long_1024;
        fb->short_window[1] = kbd_short_128;
#ifdef LD_DEC
        fb->ld_window[0] = sine_mid_512;
        fb->ld_window[1] = ld_mid_512;
#endif
#ifdef ALLOW_SMALL_FRAMELENGTH
    } else /* (frame_len == 960) */ {
        fb->long_window[0]  = sine_long_960;
        fb->short_window[0] = sine_short_120;
        fb->long_window[1]  = kbd_long_960;
        fb->short_window[1] = kbd_short_120;
#ifdef LD_DEC
        fb->ld_window[0] = sine_mid_480;
        fb->ld_window[1] = ld_mid_480;
#endif
    }
#endif

#ifdef USE_SSE
    if (cpu_has_sse())
    {
        fb->if_func = ifilter_bank_sse;
    } else {
        fb->if_func = ifilter_bank;
    }
#endif

    return fb;
}

void filter_bank_end(fb_info *fb)
{
    if (fb != NULL)
    {
#ifdef PROFILE
        printf("FB:                 %I64d cycles\n", fb->cycles);
#endif

        faad_mdct_end(fb->mdct256);
        faad_mdct_end(fb->mdct2048);
#ifdef LD_DEC
        faad_mdct_end(fb->mdct1024);
#endif

        faad_free(fb);
    }
}

static INLINE void imdct_long(fb_info *fb, real_t *in_data, real_t *out_data, uint16_t len)
{
#ifdef LD_DEC
    mdct_info *mdct = NULL;

    switch (len)
    {
    case 2048:
    case 1920:
        mdct = fb->mdct2048;
        break;
    case 1024:
    case 960:
        mdct = fb->mdct1024;
        break;
    }

    faad_imdct(mdct, in_data, out_data);
#else
    faad_imdct(fb->mdct2048, in_data, out_data);
#endif
}

#ifdef USE_SSE
static INLINE void imdct_long_sse(fb_info *fb, real_t *in_data, real_t *out_data, uint16_t len)
{
#ifdef LD_DEC
    mdct_info *mdct = NULL;

    switch (len)
    {
    case 2048:
    case 1920:
        mdct = fb->mdct2048;
        break;
    case 1024:
    case 960:
        mdct = fb->mdct1024;
        break;
    }

    faad_imdct_sse(mdct, in_data, out_data);
#else
    faad_imdct_sse(fb->mdct2048, in_data, out_data);
#endif
}
#endif

#ifdef LTP_DEC
static INLINE void mdct(fb_info *fb, real_t *in_data, real_t *out_data, uint16_t len)
{
    mdct_info *mdct = NULL;

    switch (len)
    {
    case 2048:
    case 1920:
        mdct = fb->mdct2048;
        break;
    case 256:
    case 240:
        mdct = fb->mdct256;
        break;
#ifdef LD_DEC
    case 1024:
    case 960:
        mdct = fb->mdct1024;
        break;
#endif
    }

    faad_mdct(mdct, in_data, out_data);
}
#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
}

#ifdef USE_SSE
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
}
#endif

#ifdef LTP_DEC
/* only works for LTP -> no overlapping, no short blocks */
void filter_bank_ltp(fb_info *fb, uint8_t window_sequence, uint8_t window_shape,
                     uint8_t window_shape_prev, real_t *in_data, real_t *out_mdct,
                     uint8_t object_type, uint16_t frame_len)
{
    int16_t i;
    ALIGN real_t windowed_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 nflat_ls = (nlong-nshort)/2;

    assert(window_sequence != EIGHT_SHORT_SEQUENCE);

#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:
        for (i = nlong-1; i >= 0; i--)
        {
            windowed_buf[i] = MUL_F(in_data[i], window_long_prev[i]);
            windowed_buf[i+nlong] = MUL_F(in_data[i+nlong], window_long[nlong-1-i]);
        }
        mdct(fb, windowed_buf, out_mdct, 2*nlong);
        break;

    case LONG_START_SEQUENCE:
        for (i = 0; i < nlong; i++)
            windowed_buf[i] = MUL_F(in_data[i], window_long_prev[i]);
        for (i = 0; i < nflat_ls; i++)
            windowed_buf[i+nlong] = in_data[i+nlong];
        for (i = 0; i < nshort; i++)
            windowed_buf[i+nlong+nflat_ls] = MUL_F(in_data[i+nlong+nflat_ls], window_short[nshort-1-i]);
        for (i = 0; i < nflat_ls; i++)
            windowed_buf[i+nlong+nflat_ls+nshort] = 0;
        mdct(fb, windowed_buf, out_mdct, 2*nlong);
        break;

    case LONG_STOP_SEQUENCE:
        for (i = 0; i < nflat_ls; i++)
            windowed_buf[i] = 0;
        for (i = 0; i < nshort; i++)
            windowed_buf[i+nflat_ls] = MUL_F(in_data[i+nflat_ls], window_short_prev[i]);
        for (i = 0; i < nflat_ls; i++)
            windowed_buf[i+nflat_ls+nshort] = in_data[i+nflat_ls+nshort];
        for (i = 0; i < nlong; i++)
            windowed_buf[i+nlong] = MUL_F(in_data[i+nlong], window_long[nlong-1-i]);
        mdct(fb, windowed_buf, out_mdct, 2*nlong);
        break;
    }
}
#endif