FFT: factor a shuffle out of the inner loop and merge it into fft_permute.

6% faster SSE FFT on Conroe, 2.5% on Penryn.

Signed-off-by: Janne Grunau <janne-ffmpeg@jannau.net>
(cherry picked from commit e6b1ed693ae4098e6b9eabf938fc31ec0b09b120)

6 files changed, 45 insertions, 38 deletions

diff --git a/libavcodec/arm/fft_init_arm.c b/libavcodec/arm/fft_init_arm.c
index 3f2d554413..dff0689566 100644
--- a/libavcodec/arm/fft_init_arm.c
+++ b/libavcodec/arm/fft_init_arm.c
@@ -44,7 +44,7 @@ av_cold void ff_fft_init_arm(FFTContext *s)
         s->imdct_calc   = ff_imdct_calc_neon;
         s->imdct_half   = ff_imdct_half_neon;
         s->mdct_calc    = ff_mdct_calc_neon;
-        s->permutation  = FF_MDCT_PERM_INTERLEAVE;
+        s->mdct_permutation = FF_MDCT_PERM_INTERLEAVE;
     }
 }
 
diff --git a/libavcodec/fft.c b/libavcodec/fft.c
index 3fd4d279cb..eade76a516 100644
--- a/libavcodec/fft.c
+++ b/libavcodec/fft.c
@@ -97,6 +97,7 @@ av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
     if (!s->tmp_buf)
         goto fail;
     s->inverse = inverse;
+    s->fft_permutation = FF_FFT_PERM_DEFAULT;
 
     s->fft_permute = ff_fft_permute_c;
     s->fft_calc    = ff_fft_calc_c;
@@ -113,8 +114,12 @@ av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse)
     for(j=4; j<=nbits; j++) {
         ff_init_ff_cos_tabs(j);
     }
-    for(i=0; i<n; i++)
-        s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = i;
+    for(i=0; i<n; i++) {
+        int j = i;
+        if (s->fft_permutation == FF_FFT_PERM_SWAP_LSBS)
+            j = (j&~3) | ((j>>1)&1) | ((j<<1)&2);
+        s->revtab[-split_radix_permutation(i, n, s->inverse) & (n-1)] = j;
+    }
 
     return 0;
  fail:
diff --git a/libavcodec/fft.h b/libavcodec/fft.h
index 37cbfbf6a1..58a7f30ad1 100644
--- a/libavcodec/fft.h
+++ b/libavcodec/fft.h
@@ -44,7 +44,10 @@ struct FFTContext {
     void (*imdct_calc)(struct FFTContext *s, FFTSample *output, const FFTSample *input);
     void (*imdct_half)(struct FFTContext *s, FFTSample *output, const FFTSample *input);
     void (*mdct_calc)(struct FFTContext *s, FFTSample *output, const FFTSample *input);
-    int permutation;
+    int fft_permutation;
+#define FF_FFT_PERM_DEFAULT   0
+#define FF_FFT_PERM_SWAP_LSBS 1
+    int mdct_permutation;
 #define FF_MDCT_PERM_NONE       0
 #define FF_MDCT_PERM_INTERLEAVE 1
 };
diff --git a/libavcodec/mdct.c b/libavcodec/mdct.c
index c511188d22..819f618115 100644
--- a/libavcodec/mdct.c
+++ b/libavcodec/mdct.c
@@ -71,7 +71,7 @@ av_cold int ff_mdct_init(FFTContext *s, int nbits, int inverse, double scale)
     s->mdct_bits = nbits;
     s->mdct_size = n;
     n4 = n >> 2;
-    s->permutation = FF_MDCT_PERM_NONE;
+    s->mdct_permutation = FF_MDCT_PERM_NONE;
 
     if (ff_fft_init(s, s->mdct_bits - 2, inverse) < 0)
         goto fail;
@@ -80,7 +80,7 @@ av_cold int ff_mdct_init(FFTContext *s, int nbits, int inverse, double scale)
     if (!s->tcos)
         goto fail;
 
-    switch (s->permutation) {
+    switch (s->mdct_permutation) {
     case FF_MDCT_PERM_NONE:
         s->tsin = s->tcos + n4;
         tstep = 1;
diff --git a/libavcodec/x86/fft.c b/libavcodec/x86/fft.c
index 771b1e6649..5ca341d5e4 100644
--- a/libavcodec/x86/fft.c
+++ b/libavcodec/x86/fft.c
@@ -30,6 +30,7 @@ av_cold void ff_fft_init_mmx(FFTContext *s)
         s->imdct_half  = ff_imdct_half_sse;
         s->fft_permute = ff_fft_permute_sse;
         s->fft_calc    = ff_fft_calc_sse;
+        s->fft_permutation = FF_FFT_PERM_SWAP_LSBS;
     } else if (has_vectors & AV_CPU_FLAG_3DNOWEXT && HAVE_AMD3DNOWEXT) {
         /* 3DNowEx for K7 */
         s->imdct_calc = ff_imdct_calc_3dn2;
diff --git a/libavcodec/x86/fft_mmx.asm b/libavcodec/x86/fft_mmx.asm
index 1dcd629184..68e20df3b0 100644
--- a/libavcodec/x86/fft_mmx.asm
+++ b/libavcodec/x86/fft_mmx.asm
@@ -51,6 +51,7 @@ SECTION_RODATA
 %define M_SQRT1_2 0.70710678118654752440
 ps_root2: times 4 dd M_SQRT1_2
 ps_root2mppm: dd -M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, -M_SQRT1_2
+ps_p1p1m1p1: dd 0, 0, 1<<31, 0
 ps_m1p1: dd 1<<31, 0
 
 %assign i 16
@@ -95,54 +96,51 @@ section .text align=16
     SWAP     %3, %6
 %endmacro
 
-; in:  %1={r0,i0,r1,i1} %2={r2,i2,r3,i3}
+; in:  %1={r0,i0,r2,i2} %2={r1,i1,r3,i3}
 ; out: %1={r0,r1,r2,r3} %2={i0,i1,i2,i3}
 %macro T4_SSE 3
     mova     %3, %1
-    shufps   %1, %2, 0x64 ; {r0,i0,r3,i2}
-    shufps   %3, %2, 0xce ; {r1,i1,r2,i3}
+    addps    %1, %2       ; {t1,t2,t6,t5}
+    subps    %3, %2       ; {t3,t4,-t8,t7}
+    xorps    %3, [ps_p1p1m1p1]
     mova     %2, %1
-    addps    %1, %3       ; {t1,t2,t6,t5}
-    subps    %2, %3       ; {t3,t4,t8,t7}
+    shufps   %1, %3, 0x44 ; {t1,t2,t3,t4}
+    shufps   %2, %3, 0xbe ; {t6,t5,t7,t8}
     mova     %3, %1
-    shufps   %1, %2, 0x44 ; {t1,t2,t3,t4}
-    shufps   %3, %2, 0xbe ; {t6,t5,t7,t8}
+    addps    %1, %2       ; {r0,i0,r1,i1}
+    subps    %3, %2       ; {r2,i2,r3,i3}
     mova     %2, %1
-    addps    %1, %3       ; {r0,i0,r1,i1}
-    subps    %2, %3       ; {r2,i2,r3,i3}
-    mova     %3, %1
-    shufps   %1, %2, 0x88 ; {r0,r1,r2,r3}
-    shufps   %3, %2, 0xdd ; {i0,i1,i2,i3}
-    SWAP     %2, %3
+    shufps   %1, %3, 0x88 ; {r0,r1,r2,r3}
+    shufps   %2, %3, 0xdd ; {i0,i1,i2,i3}
 %endmacro
 
-%macro T8_SSE 6 ; r0,i0,r1,i1,t0,t1
-    mova     %5, %3
-    shufps   %3, %4, 0x44 ; {r4,i4,r6,i6}
-    shufps   %5, %4, 0xee ; {r5,i5,r7,i7}
+; in:  %1={r0,r1,r2,r3} %2={i0,i1,i2,i3} %3={r4,i4,r6,i6} %4={r5,i5,r7,i7}
+; out: %1={r0,r1,r2,r3} %2={i0,i1,i2,i3} %1={r4,r5,r6,r7} %2={i4,i5,i6,i7}
+%macro T8_SSE 6
     mova     %6, %3
-    subps    %3, %5       ; {r5,i5,r7,i7}
-    addps    %6, %5       ; {t1,t2,t3,t4}
-    mova     %5, %3
-    shufps   %5, %5, 0xb1 ; {i5,r5,i7,r7}
+    subps    %3, %4       ; {r5,i5,r7,i7}
+    addps    %6, %4       ; {t1,t2,t3,t4}
+    mova     %4, %3
+    shufps   %4, %4, 0xb1 ; {i5,r5,i7,r7}
     mulps    %3, [ps_root2mppm] ; {-r5,i5,r7,-i7}
-    mulps    %5, [ps_root2]
-    addps    %3, %5       ; {t8,t7,ta,t9}
-    mova     %5, %6
+    mulps    %4, [ps_root2]
+    addps    %3, %4       ; {t8,t7,ta,t9}
+    mova     %4, %6
     shufps   %6, %3, 0x36 ; {t3,t2,t9,t8}
-    shufps   %5, %3, 0x9c ; {t1,t4,t7,ta}
+    shufps   %4, %3, 0x9c ; {t1,t4,t7,ta}
     mova     %3, %6
-    addps    %6, %5       ; {t1,t2,t9,ta}
-    subps    %3, %5       ; {t6,t5,tc,tb}
-    mova     %5, %6
+    addps    %6, %4       ; {t1,t2,t9,ta}
+    subps    %3, %4       ; {t6,t5,tc,tb}
+    mova     %4, %6
     shufps   %6, %3, 0xd8 ; {t1,t9,t5,tb}
-    shufps   %5, %3, 0x8d ; {t2,ta,t6,tc}
+    shufps   %4, %3, 0x8d ; {t2,ta,t6,tc}
     mova     %3, %1
-    mova     %4, %2
+    mova     %5, %2
     addps    %1, %6       ; {r0,r1,r2,r3}
-    addps    %2, %5       ; {i0,i1,i2,i3}
+    addps    %2, %4       ; {i0,i1,i2,i3}
     subps    %3, %6       ; {r4,r5,r6,r7}
-    subps    %4, %5       ; {i4,i5,i6,i7}
+    subps    %5, %4       ; {i4,i5,i6,i7}
+    SWAP     %4, %5
 %endmacro
 
 ; scheduled for cpu-bound sizes