changeset 10204:db033d1fbf44 libavcodec

Allow arch-specific mdct code to request interleaving of cos/sin tables

--- a/dsputil.h	Mon Sep 21 01:40:08 2009 +0000
+++ b/dsputil.h	Mon Sep 21 02:56:06 2009 +0000
@@ -687,6 +687,9 @@
     void (*imdct_half)(struct FFTContext *s, FFTSample *output, const FFTSample *input);
     void (*mdct_calc)(struct FFTContext *s, FFTSample *output, const FFTSample *input);
     int split_radix;
+    int permutation;
+#define FF_MDCT_PERM_NONE       0
+#define FF_MDCT_PERM_INTERLEAVE 1
 } FFTContext;
 
 extern FFTSample* const ff_cos_tabs[13];

--- a/mdct.c	Mon Sep 21 01:40:08 2009 +0000
+++ b/mdct.c	Mon Sep 21 02:56:06 2009 +0000
@@ -76,32 +76,45 @@
 {
     int n, n4, i;
     double alpha, theta;
+    int tstep;
 
     memset(s, 0, sizeof(*s));
     n = 1 << nbits;
     s->mdct_bits = nbits;
     s->mdct_size = n;
     n4 = n >> 2;
-    s->tcos = av_malloc(n4 * sizeof(FFTSample));
+    s->permutation = FF_MDCT_PERM_NONE;
+
+    if (ff_fft_init(s, s->mdct_bits - 2, inverse) < 0)
+        goto fail;
+
+    s->tcos = av_malloc(n/2 * sizeof(FFTSample));
     if (!s->tcos)
         goto fail;
-    s->tsin = av_malloc(n4 * sizeof(FFTSample));
-    if (!s->tsin)
+
+    switch (s->permutation) {
+    case FF_MDCT_PERM_NONE:
+        s->tsin = s->tcos + n4;
+        tstep = 1;
+        break;
+    case FF_MDCT_PERM_INTERLEAVE:
+        s->tsin = s->tcos + 1;
+        tstep = 2;
+        break;
+    default:
         goto fail;
+    }
 
     theta = 1.0 / 8.0 + (scale < 0 ? n4 : 0);
     scale = sqrt(fabs(scale));
     for(i=0;i<n4;i++) {
         alpha = 2 * M_PI * (i + theta) / n;
-        s->tcos[i] = -cos(alpha) * scale;
-        s->tsin[i] = -sin(alpha) * scale;
+        s->tcos[i*tstep] = -cos(alpha) * scale;
+        s->tsin[i*tstep] = -sin(alpha) * scale;
     }
-    if (ff_fft_init(s, s->mdct_bits - 2, inverse) < 0)
-        goto fail;
     return 0;
  fail:
-    av_freep(&s->tcos);
-    av_freep(&s->tsin);
+    ff_mdct_end(s);
     return -1;
 }
 
@@ -229,6 +242,5 @@
 av_cold void ff_mdct_end(FFTContext *s)
 {
     av_freep(&s->tcos);
-    av_freep(&s->tsin);
     ff_fft_end(s);
 }