update to the 2.0 release of faad, patch by adland

git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@12528 b3059339-0415-0410-9bf9-f77b7e298cf2

1 files changed, 348 insertions, 94 deletions

diff --git a/libfaad2/sbr_qmf.c b/libfaad2/sbr_qmf.c
index d7708979ed..114e7b1a2d 100644
--- a/libfaad2/sbr_qmf.c
+++ b/libfaad2/sbr_qmf.c
@@ -1,6 +1,6 @@
 /*
 ** FAAD2 - Freeware Advanced Audio (AAC) Decoder including SBR decoding
-** Copyright (C) 2003 M. Bakker, Ahead Software AG, http://www.nero.com
+** 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
@@ -22,7 +22,7 @@
 ** Commercial non-GPL licensing of this software is possible.
 ** For more info contact Ahead Software through Mpeg4AAClicense@nero.com.
 **
-** $Id: sbr_qmf.c,v 1.13 2003/09/30 12:43:05 menno Exp $
+** $Id: sbr_qmf.c,v 1.2 2003/10/03 22:22:27 alex Exp $
 **/
 
 #include "common.h"
@@ -41,8 +41,8 @@
 
 qmfa_info *qmfa_init(uint8_t channels)
 {
-    qmfa_info *qmfa = (qmfa_info*)malloc(sizeof(qmfa_info));
-    qmfa->x = (real_t*)malloc(channels * 10 * sizeof(real_t));
+    qmfa_info *qmfa = (qmfa_info*)faad_malloc(sizeof(qmfa_info));
+    qmfa->x = (real_t*)faad_malloc(channels * 10 * sizeof(real_t));
     memset(qmfa->x, 0, channels * 10 * sizeof(real_t));
 
     qmfa->channels = channels;
@@ -54,22 +54,22 @@ void qmfa_end(qmfa_info *qmfa)
 {
     if (qmfa)
     {
-        if (qmfa->x) free(qmfa->x);
-        free(qmfa);
+        if (qmfa->x) faad_free(qmfa->x);
+        faad_free(qmfa);
     }
 }
 
 void sbr_qmf_analysis_32(sbr_info *sbr, qmfa_info *qmfa, const real_t *input,
-                         qmf_t *X, uint8_t offset, uint8_t kx)
+                         qmf_t X[MAX_NTSRHFG][32], uint8_t offset, uint8_t kx)
 {
-    uint8_t l;
-    real_t u[64];
+    ALIGN real_t u[64];
 #ifndef SBR_LOW_POWER
-    real_t x[64], y[64];
+    ALIGN real_t x[64], y[64];
 #else
-    real_t y[32];
+    ALIGN real_t y[32];
 #endif
-    const real_t *inptr = input;
+    uint16_t in = 0;
+    uint8_t l;
 
     /* qmf subsample l */
     for (l = 0; l < sbr->numTimeSlotsRate; l++)
@@ -83,20 +83,20 @@ void sbr_qmf_analysis_32(sbr_info *sbr, qmfa_info *qmfa, const real_t *input,
         for (n = 32 - 1; n >= 0; n--)
         {
 #ifdef FIXED_POINT
-            qmfa->x[n] = (*inptr++) >> 5;
+            qmfa->x[n] = (input[in++]) >> 5;
 #else
-            qmfa->x[n] = *inptr++;
+            qmfa->x[n] = input[in++];
 #endif
         }
 
         /* window and summation to create array u */
         for (n = 0; n < 64; n++)
         {
-            u[n] = MUL_R_C(qmfa->x[n], qmf_c[2*n]) +
-                MUL_R_C(qmfa->x[n + 64], qmf_c[2*(n + 64)]) +
-                MUL_R_C(qmfa->x[n + 128], qmf_c[2*(n + 128)]) +
-                MUL_R_C(qmfa->x[n + 192], qmf_c[2*(n + 192)]) +
-                MUL_R_C(qmfa->x[n + 256], qmf_c[2*(n + 256)]);
+            u[n] = MUL_F(qmfa->x[n], qmf_c[2*n]) +
+                MUL_F(qmfa->x[n + 64], qmf_c[2*(n + 64)]) +
+                MUL_F(qmfa->x[n + 128], qmf_c[2*(n + 128)]) +
+                MUL_F(qmfa->x[n + 192], qmf_c[2*(n + 192)]) +
+                MUL_F(qmfa->x[n + 256], qmf_c[2*(n + 256)]);
         }
 
         /* calculate 32 subband samples by introducing X */
@@ -114,39 +114,39 @@ void sbr_qmf_analysis_32(sbr_info *sbr, qmfa_info *qmfa, const real_t *input,
             if (n < kx)
             {
 #ifdef FIXED_POINT
-            QMF_RE(X[((l + offset)<<5) + n]) = u[n] << 1;
+                QMF_RE(X[l + offset][n]) = u[n] << 1;
 #else
-            QMF_RE(X[((l + offset)<<5) + n]) = 2. * u[n];
+                QMF_RE(X[l + offset][n]) = 2. * u[n];
 #endif
             } else {
-                QMF_RE(X[((l + offset)<<5) + n]) = 0;
+                QMF_RE(X[l + offset][n]) = 0;
             }
         }
 #else
         x[0] = u[0];
-        x[63] = u[32];
-        for (n = 2; n < 64; n += 2)
+        for (n = 0; n < 31; n++)
         {
-            x[n-1] = u[(n>>1)];
-            x[n] = -u[64-(n>>1)];
+            x[2*n+1] = u[n+1] + u[63-n];
+            x[2*n+2] = u[n+1] - u[63-n];
         }
+        x[63] = u[32];
 
-        DCT4_64(y, x);
+        DCT4_64_kernel(y, x);
 
         for (n = 0; n < 32; n++)
         {
             if (n < kx)
             {
 #ifdef FIXED_POINT
-            QMF_RE(X[((l + offset)<<5) + n]) = y[n] << 1;
-            QMF_IM(X[((l + offset)<<5) + n]) = -y[63-n] << 1;
+                QMF_RE(X[l + offset][n]) = y[n] << 1;
+                QMF_IM(X[l + offset][n]) = -y[63-n] << 1;
 #else
-            QMF_RE(X[((l + offset)<<5) + n]) = 2. * y[n];
-            QMF_IM(X[((l + offset)<<5) + n]) = -2. * y[63-n];
+                QMF_RE(X[l + offset][n]) = 2. * y[n];
+                QMF_IM(X[l + offset][n]) = -2. * y[63-n];
 #endif
             } else {
-                QMF_RE(X[((l + offset)<<5) + n]) = 0;
-                QMF_IM(X[((l + offset)<<5) + n]) = 0;
+                QMF_RE(X[l + offset][n]) = 0;
+                QMF_IM(X[l + offset][n]) = 0;
             }
         }
 #endif
@@ -155,18 +155,32 @@ void sbr_qmf_analysis_32(sbr_info *sbr, qmfa_info *qmfa, const real_t *input,
 
 qmfs_info *qmfs_init(uint8_t channels)
 {
-    int size = 0;
-    qmfs_info *qmfs = (qmfs_info*)malloc(sizeof(qmfs_info));
+    qmfs_info *qmfs = (qmfs_info*)faad_malloc(sizeof(qmfs_info));
 
-    qmfs->v[0] = (real_t*)malloc(channels * 10 * sizeof(real_t));
+#ifndef SBR_LOW_POWER
+    qmfs->v[0] = (real_t*)faad_malloc(channels * 10 * sizeof(real_t));
     memset(qmfs->v[0], 0, channels * 10 * sizeof(real_t));
-    qmfs->v[1] = (real_t*)malloc(channels * 10 * sizeof(real_t));
+    qmfs->v[1] = (real_t*)faad_malloc(channels * 10 * sizeof(real_t));
     memset(qmfs->v[1], 0, channels * 10 * sizeof(real_t));
+#else
+    qmfs->v[0] = (real_t*)faad_malloc(channels * 20 * sizeof(real_t));
+    memset(qmfs->v[0], 0, channels * 20 * sizeof(real_t));
+    qmfs->v[1] = NULL;
+#endif
 
     qmfs->v_index = 0;
 
     qmfs->channels = channels;
 
+#ifdef USE_SSE
+    if (cpu_has_sse())
+    {
+        qmfs->qmf_func = sbr_qmf_synthesis_64_sse;
+    } else {
+        qmfs->qmf_func = sbr_qmf_synthesis_64;
+    }
+#endif
+
     return qmfs;
 }
 
@@ -174,84 +188,195 @@ void qmfs_end(qmfs_info *qmfs)
 {
     if (qmfs)
     {
-        if (qmfs->v[0]) free(qmfs->v[0]);
-        if (qmfs->v[1]) free(qmfs->v[1]);
-        free(qmfs);
+        if (qmfs->v[0]) faad_free(qmfs->v[0]);
+#ifndef SBR_LOW_POWER
+        if (qmfs->v[1]) faad_free(qmfs->v[1]);
+#endif
+        faad_free(qmfs);
     }
 }
 
 #ifdef SBR_LOW_POWER
-void sbr_qmf_synthesis_64(sbr_info *sbr, qmfs_info *qmfs, const qmf_t *X,
+void sbr_qmf_synthesis_64(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][64],
                           real_t *output)
 {
+    ALIGN real_t x[64];
+    ALIGN real_t y[64];
+    int16_t n, k, out = 0;
     uint8_t l;
-    int16_t n, k;
-    real_t x[64];
-    real_t *outptr = output;
 
 
     /* qmf subsample l */
     for (l = 0; l < sbr->numTimeSlotsRate; l++)
+    {
+        //real_t *v0, *v1;
+
+        /* shift buffers */
+        //memmove(qmfs->v[0] + 64, qmfs->v[0], (640-64)*sizeof(real_t));
+        //memmove(qmfs->v[1] + 64, qmfs->v[1], (640-64)*sizeof(real_t));
+        memmove(qmfs->v[0] + 128, qmfs->v[0], (1280-128)*sizeof(real_t));
+
+        //v0 = qmfs->v[qmfs->v_index];
+        //v1 = qmfs->v[(qmfs->v_index + 1) & 0x1];
+        //qmfs->v_index = (qmfs->v_index + 1) & 0x1;
+
+        /* calculate 128 samples */
+        for (k = 0; k < 64; k++)
         {
-        real_t *v0, *v1;
+#ifdef FIXED_POINT
+            x[k] = QMF_RE(X[l][k]);
+#else
+            x[k] = QMF_RE(X[l][k]) / 32.;
+#endif
+        }
+
+        for (n = 0; n < 32; n++)
+        {
+            y[2*n]   = -x[2*n];
+            y[2*n+1] =  x[2*n+1];
+        }
+
+        DCT2_64_unscaled(x, x);
+
+        for (n = 0; n < 64; n++)
+        {
+            qmfs->v[0][n+32] = x[n];
+        }
+        for (n = 0; n < 32; n++)
+        {
+            qmfs->v[0][31 - n] = x[n + 1];
+        }
+        DST2_64_unscaled(x, y);
+        qmfs->v[0][96] = 0;
+        for (n = 1; n < 32; n++)
+        {
+            qmfs->v[0][n + 96] = x[n-1];
+        }
+
+        /* calculate 64 output samples and window */
+        for (k = 0; k < 64; k++)
+        {
+#if 1
+             output[out++] = MUL_F(qmfs->v[0][k], qmf_c[k]) +
+                 MUL_F(qmfs->v[0][192 + k], qmf_c[64 + k]) +
+                 MUL_F(qmfs->v[0][256 + k], qmf_c[128 + k]) +
+                 MUL_F(qmfs->v[0][256 + 192 + k], qmf_c[128 + 64 + k]) +
+                 MUL_F(qmfs->v[0][512 + k], qmf_c[256 + k]) +
+                 MUL_F(qmfs->v[0][512 + 192 + k], qmf_c[256 + 64 + k]) +
+                 MUL_F(qmfs->v[0][768 + k], qmf_c[384 + k]) +
+                 MUL_F(qmfs->v[0][768 + 192 + k], qmf_c[384 + 64 + k]) +
+                 MUL_F(qmfs->v[0][1024 + k], qmf_c[512 + k]) +
+                 MUL_F(qmfs->v[0][1024 + 192 + k], qmf_c[512 + 64 + k]);
+#else
+            output[out++] = MUL_F(v0[k], qmf_c[k]) +
+                MUL_F(v0[64 + k], qmf_c[64 + k]) +
+                MUL_F(v0[128 + k], qmf_c[128 + k]) +
+                MUL_F(v0[192 + k], qmf_c[192 + k]) +
+                MUL_F(v0[256 + k], qmf_c[256 + k]) +
+                MUL_F(v0[320 + k], qmf_c[320 + k]) +
+                MUL_F(v0[384 + k], qmf_c[384 + k]) +
+                MUL_F(v0[448 + k], qmf_c[448 + k]) +
+                MUL_F(v0[512 + k], qmf_c[512 + k]) +
+                MUL_F(v0[576 + k], qmf_c[576 + k]);
+#endif
+        }
+    }
+}
+
+void sbr_qmf_synthesis_64_sse(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][64],
+                              real_t *output)
+{
+    ALIGN real_t x[64];
+    ALIGN real_t y[64];
+    ALIGN real_t y2[64];
+    int16_t n, k, out = 0;
+    uint8_t l;
+
+    /* qmf subsample l */
+    for (l = 0; l < sbr->numTimeSlotsRate; l++)
+    {
+        //real_t *v0, *v1;
 
         /* shift buffers */
-        memmove(qmfs->v[0] + 64, qmfs->v[0], (640-64)*sizeof(real_t));
-        memmove(qmfs->v[1] + 64, qmfs->v[1], (640-64)*sizeof(real_t));
+        //memmove(qmfs->v[0] + 64, qmfs->v[0], (640-64)*sizeof(real_t));
+        //memmove(qmfs->v[1] + 64, qmfs->v[1], (640-64)*sizeof(real_t));
+        memmove(qmfs->v[0] + 128, qmfs->v[0], (1280-128)*sizeof(real_t));
 
-        v0 = qmfs->v[qmfs->v_index];
-        v1 = qmfs->v[(qmfs->v_index + 1) & 0x1];
-        qmfs->v_index = (qmfs->v_index + 1) & 0x1;
+        //v0 = qmfs->v[qmfs->v_index];
+        //v1 = qmfs->v[(qmfs->v_index + 1) & 0x1];
+        //qmfs->v_index = (qmfs->v_index + 1) & 0x1;
 
         /* calculate 128 samples */
         for (k = 0; k < 64; k++)
         {
 #ifdef FIXED_POINT
-            x[k] = QMF_RE(X[(l<<6) + k]);
+            x[k] = QMF_RE(X[l][k]);
 #else
-            x[k] = QMF_RE(X[(l<<6) + k]) / 32.;
+            x[k] = QMF_RE(X[l][k]) / 32.;
 #endif
         }
 
+        for (n = 0; n < 32; n++)
+        {
+            y[2*n]   = -x[2*n];
+            y[2*n+1] =  x[2*n+1];
+        }
+
         DCT2_64_unscaled(x, x);
 
+        for (n = 0; n < 64; n++)
+        {
+            qmfs->v[0][n+32] = x[n];
+        }
         for (n = 0; n < 32; n++)
         {
-            v0[n+32] = x[n];
-            v1[n] = x[n+32];
+            qmfs->v[0][31 - n] = x[n + 1];
         }
-        v0[0] = v1[0];
+
+        DST2_64_unscaled(x, y);
+        qmfs->v[0][96] = 0;
         for (n = 1; n < 32; n++)
-            {
-            v0[32 - n] =  v0[n + 32];
-            v1[n + 32] = -v1[32 - n];
-            }
-        v1[32] = 0;
+        {
+            qmfs->v[0][n + 96] = x[n-1];
+        }
 
         /* calculate 64 output samples and window */
         for (k = 0; k < 64; k++)
         {
-            *outptr++ = MUL_R_C(v0[k], qmf_c[k]) +
-                MUL_R_C(v0[64 + k], qmf_c[64 + k]) +
-                MUL_R_C(v0[128 + k], qmf_c[128 + k]) +
-                MUL_R_C(v0[192 + k], qmf_c[192 + k]) +
-                MUL_R_C(v0[256 + k], qmf_c[256 + k]) +
-                MUL_R_C(v0[320 + k], qmf_c[320 + k]) +
-                MUL_R_C(v0[384 + k], qmf_c[384 + k]) +
-                MUL_R_C(v0[448 + k], qmf_c[448 + k]) +
-                MUL_R_C(v0[512 + k], qmf_c[512 + k]) +
-                MUL_R_C(v0[576 + k], qmf_c[576 + k]);
+#if 1
+             output[out++] = MUL_F(qmfs->v[0][k], qmf_c[k]) +
+                 MUL_F(qmfs->v[0][192 + k], qmf_c[64 + k]) +
+                 MUL_F(qmfs->v[0][256 + k], qmf_c[128 + k]) +
+                 MUL_F(qmfs->v[0][256 + 192 + k], qmf_c[128 + 64 + k]) +
+                 MUL_F(qmfs->v[0][512 + k], qmf_c[256 + k]) +
+                 MUL_F(qmfs->v[0][512 + 192 + k], qmf_c[256 + 64 + k]) +
+                 MUL_F(qmfs->v[0][768 + k], qmf_c[384 + k]) +
+                 MUL_F(qmfs->v[0][768 + 192 + k], qmf_c[384 + 64 + k]) +
+                 MUL_F(qmfs->v[0][1024 + k], qmf_c[512 + k]) +
+                 MUL_F(qmfs->v[0][1024 + 192 + k], qmf_c[512 + 64 + k]);
+#else
+            output[out++] = MUL_F(v0[k], qmf_c[k]) +
+                MUL_F(v0[64 + k], qmf_c[64 + k]) +
+                MUL_F(v0[128 + k], qmf_c[128 + k]) +
+                MUL_F(v0[192 + k], qmf_c[192 + k]) +
+                MUL_F(v0[256 + k], qmf_c[256 + k]) +
+                MUL_F(v0[320 + k], qmf_c[320 + k]) +
+                MUL_F(v0[384 + k], qmf_c[384 + k]) +
+                MUL_F(v0[448 + k], qmf_c[448 + k]) +
+                MUL_F(v0[512 + k], qmf_c[512 + k]) +
+                MUL_F(v0[576 + k], qmf_c[576 + k]);
+#endif
         }
     }
 }
 #else
-void sbr_qmf_synthesis_64(sbr_info *sbr, qmfs_info *qmfs, const qmf_t *X,
+void sbr_qmf_synthesis_64(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][64],
                           real_t *output)
 {
+    ALIGN real_t x1[64], x2[64];
+    real_t scale = 1.f/64.f;
+    int16_t n, k, out = 0;
     uint8_t l;
-    int16_t n, k;
-    real_t x1[64], x2[64];
-    real_t *outptr = output;
 
 
     /* qmf subsample l */
@@ -268,39 +393,168 @@ void sbr_qmf_synthesis_64(sbr_info *sbr, qmfs_info *qmfs, const qmf_t *X,
         qmfs->v_index = (qmfs->v_index + 1) & 0x1;
 
         /* calculate 128 samples */
-        for (k = 0; k < 64; k++)
+        x1[0] = scale*QMF_RE(X[l][0]);
+        x2[63] = scale*QMF_IM(X[l][0]);
+        for (k = 0; k < 31; k++)
         {
-            x1[k] = QMF_RE(X[(l<<6) + k])/64.;
-            x2[63 - k] = QMF_IM(X[(l<<6) + k])/64.;
+            x1[2*k+1] = scale*(QMF_RE(X[l][2*k+1]) - QMF_RE(X[l][2*k+2]));
+            x1[2*k+2] = scale*(QMF_RE(X[l][2*k+1]) + QMF_RE(X[l][2*k+2]));
+
+            x2[61 - 2*k] = scale*(QMF_IM(X[l][2*k+2]) - QMF_IM(X[l][2*k+1]));
+            x2[62 - 2*k] = scale*(QMF_IM(X[l][2*k+2]) + QMF_IM(X[l][2*k+1]));
         }
+        x1[63] = scale*QMF_RE(X[l][63]);
+        x2[0] = scale*QMF_IM(X[l][63]);
 
-        DCT4_64(x1, x1);
-        DCT4_64(x2, x2);
+        DCT4_64_kernel(x1, x1);
+        DCT4_64_kernel(x2, x2);
 
-        for (n = 0; n < 64; n+=2)
+        for (n = 0; n < 32; n++)
         {
-            v0[n]      =  x2[n]   - x1[n];
-            v0[n+1]    = -x2[n+1] - x1[n+1];
-            v1[63-n]   =  x2[n]   + x1[n];
-            v1[63-n-1] = -x2[n+1] + x1[n+1];
+            v0[   2*n]   =  x2[2*n]   - x1[2*n];
+            v1[63-2*n]   =  x2[2*n]   + x1[2*n];
+            v0[   2*n+1] = -x2[2*n+1] - x1[2*n+1];
+            v1[62-2*n]   = -x2[2*n+1] + x1[2*n+1];
         }
 
         /* calculate 64 output samples and window */
         for (k = 0; k < 64; k++)
         {
-            *outptr++ = MUL_R_C(v0[k], qmf_c[k]) +
-                MUL_R_C(v0[64 + k], qmf_c[64 + k]) +
-                MUL_R_C(v0[128 + k], qmf_c[128 + k]) +
-                MUL_R_C(v0[192 + k], qmf_c[192 + k]) +
-                MUL_R_C(v0[256 + k], qmf_c[256 + k]) +
-                MUL_R_C(v0[320 + k], qmf_c[320 + k]) +
-                MUL_R_C(v0[384 + k], qmf_c[384 + k]) +
-                MUL_R_C(v0[448 + k], qmf_c[448 + k]) +
-                MUL_R_C(v0[512 + k], qmf_c[512 + k]) +
-                MUL_R_C(v0[576 + k], qmf_c[576 + k]);
+            output[out++] = MUL_F(v0[k], qmf_c[k]) +
+                MUL_F(v0[64 + k], qmf_c[64 + k]) +
+                MUL_F(v0[128 + k], qmf_c[128 + k]) +
+                MUL_F(v0[192 + k], qmf_c[192 + k]) +
+                MUL_F(v0[256 + k], qmf_c[256 + k]) +
+                MUL_F(v0[320 + k], qmf_c[320 + k]) +
+                MUL_F(v0[384 + k], qmf_c[384 + k]) +
+                MUL_F(v0[448 + k], qmf_c[448 + k]) +
+                MUL_F(v0[512 + k], qmf_c[512 + k]) +
+                MUL_F(v0[576 + k], qmf_c[576 + k]);
         }
     }
 }
+
+#ifdef USE_SSE
+void memmove_sse_576(real_t *out, const real_t *in)
+{
+    __m128 m[144];
+    uint16_t i;
+
+    for (i = 0; i < 144; i++)
+    {
+        m[i] = _mm_load_ps(&in[i*4]);
+    }
+    for (i = 0; i < 144; i++)
+    {
+        _mm_store_ps(&out[i*4], m[i]);
+    }
+}
+
+void sbr_qmf_synthesis_64_sse(sbr_info *sbr, qmfs_info *qmfs, qmf_t X[MAX_NTSRHFG][64],
+                              real_t *output)
+{
+    ALIGN real_t x1[64], x2[64];
+    real_t scale = 1.f/64.f;
+    int16_t n, k, out = 0;
+    uint8_t l;
+
+
+    /* qmf subsample l */
+    for (l = 0; l < sbr->numTimeSlotsRate; l++)
+    {
+        real_t *v0, *v1;
+
+        /* shift buffers */
+        memmove_sse_576(qmfs->v[0] + 64, qmfs->v[0]);
+        memmove_sse_576(qmfs->v[1] + 64, qmfs->v[1]);
+
+        v0 = qmfs->v[qmfs->v_index];
+        v1 = qmfs->v[(qmfs->v_index + 1) & 0x1];
+        qmfs->v_index = (qmfs->v_index + 1) & 0x1;
+
+        /* calculate 128 samples */
+        x1[0] = scale*QMF_RE(X[l][0]);
+        x2[63] = scale*QMF_IM(X[l][0]);
+        for (k = 0; k < 31; k++)
+        {
+            x1[2*k+1] = scale*(QMF_RE(X[l][2*k+1]) - QMF_RE(X[l][2*k+2]));
+            x1[2*k+2] = scale*(QMF_RE(X[l][2*k+1]) + QMF_RE(X[l][2*k+2]));
+
+            x2[61 - 2*k] = scale*(QMF_IM(X[l][2*k+2]) - QMF_IM(X[l][2*k+1]));
+            x2[62 - 2*k] = scale*(QMF_IM(X[l][2*k+2]) + QMF_IM(X[l][2*k+1]));
+        }
+        x1[63] = scale*QMF_RE(X[l][63]);
+        x2[0] = scale*QMF_IM(X[l][63]);
+
+        DCT4_64_kernel(x1, x1);
+        DCT4_64_kernel(x2, x2);
+
+        for (n = 0; n < 32; n++)
+        {
+            v0[    2*n   ] =  x2[2*n]   - x1[2*n];
+            v1[63- 2*n   ] =  x2[2*n]   + x1[2*n];
+            v0[    2*n+1 ] = -x2[2*n+1] - x1[2*n+1];
+            v1[63-(2*n+1)] = -x2[2*n+1] + x1[2*n+1];
+        }
+
+        /* calculate 64 output samples and window */
+        for (k = 0; k < 64; k+=4)
+        {
+            __m128 m0, m1, m2, m3, m4, m5, m6, m7, m8, m9;
+            __m128 c0, c1, c2, c3, c4, c5, c6, c7, c8, c9;
+            __m128 s1, s2, s3, s4, s5, s6, s7, s8, s9;
+
+            m0 = _mm_load_ps(&v0[k]);
+            m1 = _mm_load_ps(&v0[k + 64]);
+            m2 = _mm_load_ps(&v0[k + 128]);
+            m3 = _mm_load_ps(&v0[k + 192]);
+            m4 = _mm_load_ps(&v0[k + 256]);
+            c0 = _mm_load_ps(&qmf_c[k]);
+            c1 = _mm_load_ps(&qmf_c[k + 64]);
+            c2 = _mm_load_ps(&qmf_c[k + 128]);
+            c3 = _mm_load_ps(&qmf_c[k + 192]);
+            c4 = _mm_load_ps(&qmf_c[k + 256]);
+
+            m0 = _mm_mul_ps(m0, c0);
+            m1 = _mm_mul_ps(m1, c1);
+            m2 = _mm_mul_ps(m2, c2);
+            m3 = _mm_mul_ps(m3, c3);
+            m4 = _mm_mul_ps(m4, c4);
+
+            s1 = _mm_add_ps(m0, m1);
+            s2 = _mm_add_ps(m2, m3);
+            s6 = _mm_add_ps(s1, s2);
+
+            m5 = _mm_load_ps(&v0[k + 320]);
+            m6 = _mm_load_ps(&v0[k + 384]);
+            m7 = _mm_load_ps(&v0[k + 448]);
+            m8 = _mm_load_ps(&v0[k + 512]);
+            m9 = _mm_load_ps(&v0[k + 576]);
+            c5 = _mm_load_ps(&qmf_c[k + 320]);
+            c6 = _mm_load_ps(&qmf_c[k + 384]);
+            c7 = _mm_load_ps(&qmf_c[k + 448]);
+            c8 = _mm_load_ps(&qmf_c[k + 512]);
+            c9 = _mm_load_ps(&qmf_c[k + 576]);
+
+            m5 = _mm_mul_ps(m5, c5);
+            m6 = _mm_mul_ps(m6, c6);
+            m7 = _mm_mul_ps(m7, c7);
+            m8 = _mm_mul_ps(m8, c8);
+            m9 = _mm_mul_ps(m9, c9);
+
+            s3 = _mm_add_ps(m4, m5);
+            s4 = _mm_add_ps(m6, m7);
+            s5 = _mm_add_ps(m8, m9);
+            s7 = _mm_add_ps(s3, s4);
+            s8 = _mm_add_ps(s5, s6);
+            s9 = _mm_add_ps(s7, s8);
+
+            _mm_store_ps(&output[out], s9);
+            out += 4;
+        }
+    }
+}
+#endif
 #endif
 
 #endif