split-radix FFT

c is 1.9x faster than previous c (on various x86 cpus), sse is 1.6x faster than previous sse.

Originally committed as revision 14698 to svn://svn.ffmpeg.org/ffmpeg/trunk

1 files changed, 16 insertions, 95 deletions

diff --git a/libavcodec/i386/fft_3dn2.c b/libavcodec/i386/fft_3dn2.c
index 9068dff24b..635bc76658 100644
--- a/libavcodec/i386/fft_3dn2.c
+++ b/libavcodec/i386/fft_3dn2.c
@@ -23,105 +23,26 @@
 #include "libavutil/x86_cpu.h"
 #include "libavcodec/dsputil.h"
 
-static const int p1m1[2] __attribute__((aligned(8))) =
-    { 0, 1 << 31 };
+#ifdef EMULATE_3DNOWEXT
+#define ff_fft_calc_3dn2 ff_fft_calc_3dn
+#define ff_fft_dispatch_3dn2 ff_fft_dispatch_3dn
+#define ff_fft_dispatch_interleave_3dn2 ff_fft_dispatch_interleave_3dn
+#define ff_imdct_calc_3dn2 ff_imdct_calc_3dn
+#define ff_imdct_half_3dn2 ff_imdct_half_3dn
+#endif
 
-static const int m1p1[2] __attribute__((aligned(8))) =
-    { 1 << 31, 0 };
+void ff_fft_dispatch_3dn2(FFTComplex *z, int nbits);
+void ff_fft_dispatch_interleave_3dn2(FFTComplex *z, int nbits);
 
 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z)
 {
-    int ln = s->nbits;
-    long j;
-    x86_reg i;
-    long nblocks, nloops;
-    FFTComplex *p, *cptr;
-
-    asm volatile(
-        /* FEMMS is not a must here but recommended by AMD */
-        "femms \n\t"
-        "movq %0, %%mm7 \n\t"
-        ::"m"(*(s->inverse ? m1p1 : p1m1))
-    );
-
-    i = 8 << ln;
-    asm volatile(
-        "1: \n\t"
-        "sub $32, %0 \n\t"
-        "movq    (%0,%1), %%mm0 \n\t"
-        "movq  16(%0,%1), %%mm1 \n\t"
-        "movq   8(%0,%1), %%mm2 \n\t"
-        "movq  24(%0,%1), %%mm3 \n\t"
-        "movq      %%mm0, %%mm4 \n\t"
-        "movq      %%mm1, %%mm5 \n\t"
-        "pfadd     %%mm2, %%mm0 \n\t"
-        "pfadd     %%mm3, %%mm1 \n\t"
-        "pfsub     %%mm2, %%mm4 \n\t"
-        "pfsub     %%mm3, %%mm5 \n\t"
-        "movq      %%mm0, %%mm2 \n\t"
-        "pswapd    %%mm5, %%mm5 \n\t"
-        "movq      %%mm4, %%mm3 \n\t"
-        "pxor      %%mm7, %%mm5 \n\t"
-        "pfadd     %%mm1, %%mm0 \n\t"
-        "pfadd     %%mm5, %%mm4 \n\t"
-        "pfsub     %%mm1, %%mm2 \n\t"
-        "pfsub     %%mm5, %%mm3 \n\t"
-        "movq      %%mm0,   (%0,%1) \n\t"
-        "movq      %%mm4,  8(%0,%1) \n\t"
-        "movq      %%mm2, 16(%0,%1) \n\t"
-        "movq      %%mm3, 24(%0,%1) \n\t"
-        "jg 1b \n\t"
-        :"+r"(i)
-        :"r"(z)
-    );
-    /* pass 2 .. ln-1 */
-
-    nblocks = 1 << (ln-3);
-    nloops = 1 << 2;
-    cptr = s->exptab1;
-    do {
-        p = z;
-        j = nblocks;
-        do {
-            i = nloops*8;
-            asm volatile(
-                "1: \n\t"
-                "sub $16, %0 \n\t"
-                "movq    (%1,%0), %%mm0 \n\t"
-                "movq   8(%1,%0), %%mm1 \n\t"
-                "movq    (%2,%0), %%mm2 \n\t"
-                "movq   8(%2,%0), %%mm3 \n\t"
-                "movq  (%3,%0,2), %%mm4 \n\t"
-                "movq 8(%3,%0,2), %%mm5 \n\t"
-                "pswapd    %%mm4, %%mm6 \n\t" // no need for cptr[2] & cptr[3]
-                "pswapd    %%mm5, %%mm7 \n\t"
-                "pfmul     %%mm2, %%mm4 \n\t" // cre*re cim*im
-                "pfmul     %%mm3, %%mm5 \n\t"
-                "pfmul     %%mm2, %%mm6 \n\t" // cim*re cre*im
-                "pfmul     %%mm3, %%mm7 \n\t"
-                "pfpnacc   %%mm6, %%mm4 \n\t" // cre*re-cim*im cim*re+cre*im
-                "pfpnacc   %%mm7, %%mm5 \n\t"
-                "movq      %%mm0, %%mm2 \n\t"
-                "movq      %%mm1, %%mm3 \n\t"
-                "pfadd     %%mm4, %%mm0 \n\t"
-                "pfadd     %%mm5, %%mm1 \n\t"
-                "pfsub     %%mm4, %%mm2 \n\t"
-                "pfsub     %%mm5, %%mm3 \n\t"
-                "movq      %%mm0,  (%1,%0) \n\t"
-                "movq      %%mm1, 8(%1,%0) \n\t"
-                "movq      %%mm2,  (%2,%0) \n\t"
-                "movq      %%mm3, 8(%2,%0) \n\t"
-                "jg 1b \n\t"
-                :"+r"(i)
-                :"r"(p), "r"(p + nloops), "r"(cptr)
-            );
-            p += nloops*2;
-        } while (--j);
-        cptr += nloops*2;
-        nblocks >>= 1;
-        nloops <<= 1;
-    } while (nblocks != 0);
+    int n = 1<<s->nbits;
+    int i;
+    ff_fft_dispatch_interleave_3dn2(z, s->nbits);
     asm volatile("femms");
+    if(n <= 8)
+        for(i=0; i<n; i+=2)
+            FFSWAP(FFTSample, z[i].im, z[i+1].re);
 }
 
 static void imdct_3dn2(MDCTContext *s, const FFTSample *input, FFTSample *tmp)
@@ -162,7 +83,7 @@ static void imdct_3dn2(MDCTContext *s, const FFTSample *input, FFTSample *tmp)
         );
     }
 
-    ff_fft_calc(&s->fft, z);
+    ff_fft_calc_3dn2(&s->fft, z);
 
     /* post rotation + reordering */
     for(k = 0; k < n4; k++) {