PPC: Altivec split-radix FFT

1.8x faster than altivec radix-2 on a G4
8% faster vorbis decoding

Patch (mostly) by Loren Merritt

Originally committed as revision 23956 to svn://svn.ffmpeg.org/ffmpeg/trunk

1 files changed, 55 insertions, 97 deletions

diff --git a/libavcodec/ppc/fft_altivec.c b/libavcodec/ppc/fft_altivec.c
index 5ef3bd3559..3f622c3aba 100644
--- a/libavcodec/ppc/fft_altivec.c
+++ b/libavcodec/ppc/fft_altivec.c
@@ -1,8 +1,7 @@
 /*
  * FFT/IFFT transforms
  * AltiVec-enabled
- * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
- * Based on code Copyright (c) 2002 Fabrice Bellard
+ * Copyright (c) 2009 Loren Merritt
  *
  * This file is part of FFmpeg.
  *
@@ -22,7 +21,8 @@
  */
 #include "libavcodec/fft.h"
 #include "util_altivec.h"
-#include "dsputil_altivec.h"
+#include "types_altivec.h"
+#include "regs.h"
 
 /**
  * Do a complex FFT with the parameters defined in ff_fft_init(). The
@@ -31,107 +31,65 @@
  * AltiVec-enabled
  * This code assumes that the 'z' pointer is 16 bytes-aligned
  * It also assumes all FFTComplex are 8 bytes-aligned pair of float
- * The code is exactly the same as the SSE version, except
- * that successive MUL + ADD/SUB have been merged into
- * fused multiply-add ('vec_madd' in altivec)
  */
-static void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z)
-{
-    register const vector float vczero = (const vector float)vec_splat_u32(0.);
-
-    int ln = s->nbits;
-    int j, np, np2;
-    int nblocks, nloops;
-    register FFTComplex *p, *q;
-    FFTComplex *cptr, *cptr1;
-    int k;
-
-    np = 1 << ln;
-
-    {
-        vector float *r, a, b, a1, c1, c2;
-
-        r = (vector float *)&z[0];
-
-        c1 = vcii(p,p,n,n);
-
-        if (s->inverse) {
-            c2 = vcii(p,p,n,p);
-        } else {
-            c2 = vcii(p,p,p,n);
-        }
-
-        j = (np >> 2);
-        do {
-            a = vec_ld(0, r);
-            a1 = vec_ld(sizeof(vector float), r);
-
-            b = vec_perm(a,a,vcprmle(1,0,3,2));
-            a = vec_madd(a,c1,b);
-            /* do the pass 0 butterfly */
-
-            b = vec_perm(a1,a1,vcprmle(1,0,3,2));
-            b = vec_madd(a1,c1,b);
-            /* do the pass 0 butterfly */
 
-            /* multiply third by -i */
-            b = vec_perm(b,b,vcprmle(2,3,1,0));
+// Pointers to functions. Not using function pointer syntax, because
+// that involves an extra level of indirection on some PPC ABIs.
+extern void *ff_fft_dispatch_altivec[2][15];
 
-            /* do the pass 1 butterfly */
-            vec_st(vec_madd(b,c2,a), 0, r);
-            vec_st(vec_nmsub(b,c2,a), sizeof(vector float), r);
-
-            r += 2;
-        } while (--j != 0);
+// Convert from simd order to C order.
+static void swizzle(vec_f *z, int n)
+{
+    int i;
+    n >>= 1;
+    for (i = 0; i < n; i += 2) {
+        vec_f re = z[i];
+        vec_f im = z[i+1];
+        z[i]   = vec_mergeh(re, im);
+        z[i+1] = vec_mergel(re, im);
     }
-    /* pass 2 .. ln-1 */
-
-    nblocks = np >> 3;
-    nloops = 1 << 2;
-    np2 = np >> 1;
-
-    cptr1 = s->exptab1;
-    do {
-        p = z;
-        q = z + nloops;
-        j = nblocks;
-        do {
-            cptr = cptr1;
-            k = nloops >> 1;
-            do {
-                vector float a,b,c,t1;
-
-                a = vec_ld(0, (float*)p);
-                b = vec_ld(0, (float*)q);
-
-                /* complex mul */
-                c = vec_ld(0, (float*)cptr);
-                /*  cre*re cim*re */
-                t1 = vec_madd(c, vec_perm(b,b,vcprmle(2,2,0,0)),vczero);
-                c = vec_ld(sizeof(vector float), (float*)cptr);
-                /*  -cim*im cre*im */
-                b = vec_madd(c, vec_perm(b,b,vcprmle(3,3,1,1)),t1);
-
-                /* butterfly */
-                vec_st(vec_add(a,b), 0, (float*)p);
-                vec_st(vec_sub(a,b), 0, (float*)q);
-
-                p += 2;
-                q += 2;
-                cptr += 4;
-            } while (--k);
+}
 
-            p += nloops;
-            q += nloops;
-        } while (--j);
-        cptr1 += nloops * 2;
-        nblocks = nblocks >> 1;
-        nloops = nloops << 1;
-    } while (nblocks != 0);
+static void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z)
+{
+    register vec_f  v14 __asm__("v14") = {0,0,0,0};
+    register vec_f  v15 __asm__("v15") = *(const vec_f*)ff_cos_16;
+    register vec_f  v16 __asm__("v16") = {0, 0.38268343, M_SQRT1_2, 0.92387953};
+    register vec_f  v17 __asm__("v17") = {-M_SQRT1_2, M_SQRT1_2, M_SQRT1_2,-M_SQRT1_2};
+    register vec_f  v18 __asm__("v18") = { M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, M_SQRT1_2};
+    register vec_u8 v19 __asm__("v19") = vcprm(s0,3,2,1);
+    register vec_u8 v20 __asm__("v20") = vcprm(0,1,s2,s1);
+    register vec_u8 v21 __asm__("v21") = vcprm(2,3,s0,s3);
+    register vec_u8 v22 __asm__("v22") = vcprm(2,s3,3,s2);
+    register vec_u8 v23 __asm__("v23") = vcprm(0,1,s0,s1);
+    register vec_u8 v24 __asm__("v24") = vcprm(2,3,s2,s3);
+    register vec_u8 v25 __asm__("v25") = vcprm(2,3,0,1);
+    register vec_u8 v26 __asm__("v26") = vcprm(1,2,s3,s0);
+    register vec_u8 v27 __asm__("v27") = vcprm(0,3,s2,s1);
+    register vec_u8 v28 __asm__("v28") = vcprm(0,2,s1,s3);
+    register vec_u8 v29 __asm__("v29") = vcprm(1,3,s0,s2);
+    register FFTSample *const*cos_tabs __asm__("r12") = ff_cos_tabs;
+    register FFTComplex *zarg __asm__("r3") = z;
+    __asm__(
+        "mtctr %0               \n"
+        "li   "r(9)", 16        \n"
+        "subi "r(1)","r(1) ",%1 \n"
+        "bctrl                  \n"
+        "addi "r(1)","r(1) ",%1 \n"
+        ::"r"(ff_fft_dispatch_altivec[1][s->nbits-2]), "i"(12*sizeof(void*)),
+          "r"(zarg), "r"(cos_tabs),
+          "v"(v14),"v"(v15),"v"(v16),"v"(v17),"v"(v18),"v"(v19),"v"(v20),"v"(v21),
+          "v"(v22),"v"(v23),"v"(v24),"v"(v25),"v"(v26),"v"(v27),"v"(v28),"v"(v29)
+        : "lr","ctr","r0","r4","r5","r6","r7","r8","r9","r10","r11",
+          "v0","v1","v2","v3","v4","v5","v6","v7","v8","v9","v10","v11","v12","v13"
+    );
+    if (s->nbits <= 4)
+        swizzle((vec_f*)z, 1<<s->nbits);
 }
 
 av_cold void ff_fft_init_altivec(FFTContext *s)
 {
-    s->fft_calc = ff_fft_calc_altivec;
-    s->split_radix = 0;
+    if (HAVE_GNU_AS)
+        s->fft_calc = ff_fft_calc_altivec;
+    s->split_radix = 1;
 }