ARM: 10l: fix large FFTs

Originally committed as revision 19846 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/arm/fft_neon.S

@@ -327,8 +327,10 @@ function ff_fft_permute_neon, export=1
 1:
         vld1.32         {d0-d1}, [r1,:128]!
         ldr             r4,  [r0], #4
-        uxtah           lr,  r3,  r4
-        uxtah           r4,  r3,  r4,  ror #16
+        uxth            lr,  r4
+        uxth            r4,  r4,  ror #16
+        add             lr,  r3,  lr,  lsl #3
+        add             r4,  r3,  r4,  lsl #3
         vst1.32         {d0}, [lr,:64]
         vst1.32         {d1}, [r4,:64]
         subs            r12, r12, #2

libavcodec/arm/mdct_neon.S

@@ -52,8 +52,10 @@ function ff_imdct_half_neon, export=1
         vmul.f32        d5,  d17, d3
         vsub.f32        d4,  d6,  d4
         vadd.f32        d5,  d5,  d7
-        uxtah           r8,  r1,  r6,  ror #16
-        uxtah           r6,  r1,  r6
+        uxth            r8,  r6,  ror #16
+        uxth            r6,  r6
+        add             r8,  r1,  r8,  lsl #3
+        add             r6,  r1,  r6,  lsl #3
         beq             1f
         vld2.32         {d16-d17},[r7,:128],r12
         vld2.32         {d0-d1},  [r2,:128]!
@@ -198,8 +200,10 @@ function ff_mdct_calc_neon, export=1
         subs            lr,  lr,  #16
         vsub.f32        d6,  d6,  d7            @ -R*c-I*s
         vadd.f32        d7,  d4,  d5            @ -R*s+I*c
-        uxtah           r10, r1,  r6,  ror #16
-        uxtah           r6,  r1,  r6
+        uxth            r10, r6,  ror #16
+        uxth            r6,  r6
+        add             r10, r1,  r10, lsl #3
+        add             r6,  r1,  r6,  lsl #3
         beq             1f
         vld2.32         {d16,d18},[r9,:128],r12 @ x,x in4d1,in4d0
         vld2.32         {d17,d19},[r8,:128],r12 @ x,x in3d1,in3d0
@@ -245,8 +249,10 @@ function ff_mdct_calc_neon, export=1
         subs            lr,  lr,  #16
         vsub.f32        d6,  d7,  d6            @ I*s-R*c
         vadd.f32        d7,  d4,  d5            @ R*s-I*c
-        uxtah           r10, r1,  r6,  ror #16
-        uxtah           r6,  r1,  r6
+        uxth            r10, r6,  ror #16
+        uxth            r6,  r6
+        add             r10, r1,  r10, lsl #3
+        add             r6,  r1,  r6,  lsl #3
         beq             1f
         vld2.32         {d16,d18},[r9,:128],r12 @ x,x in2d1,in2d0
         vld2.32         {d17,d19},[r8,:128],r12 @ x,x in1d1,in1d0

libavcodec/fft.c

@@ -64,7 +64,6 @@ av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
     float alpha, c1, s1, s2;
     int split_radix = 1;
     int av_unused has_vectors;
-    int revtab_shift = 0;
 
     if (nbits < 2 || nbits > 16)
         goto fail;
@@ -120,7 +119,6 @@ av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
     s->imdct_calc  = ff_imdct_calc_neon;
     s->imdct_half  = ff_imdct_half_neon;
     s->mdct_calc   = ff_mdct_calc_neon;
-    revtab_shift = 3;
 #endif
 
     if (split_radix) {
@@ -134,8 +132,7 @@ av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
                 tab[m/2-i] = tab[i];
         }
         for(i=0; i<n; i++)
-            s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] =
-                i << revtab_shift;
+            s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = i;
         s->tmp_buf = av_malloc(n * sizeof(FFTComplex));
     } else {
         int np, nblocks, np2, l;