Faster SSE FFT/MDCT, patch by Zuxy Meng %zuxy P meng A gmail P com%

unrolls some loops, utilizing all 8 xmm registers. fft-test
shows ~10% speed up in (I)FFT and ~8% speed up in (I)MDCT on Dothan

Originally committed as revision 9017 to svn://svn.ffmpeg.org/ffmpeg/trunk

1 files changed, 66 insertions, 14 deletions

diff --git a/libavcodec/i386/fft_sse.c b/libavcodec/i386/fft_sse.c
index 0dc0c61c10..a5ae9c8c4a 100644
--- a/libavcodec/i386/fft_sse.c
+++ b/libavcodec/i386/fft_sse.c
@@ -100,20 +100,33 @@ void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
             i = nloops*8;
             asm volatile(
                 "1: \n\t"
-                "sub $16, %0 \n\t"
+                "sub $32, %0 \n\t"
                 "movaps    (%2,%0), %%xmm1 \n\t"
                 "movaps    (%1,%0), %%xmm0 \n\t"
+                "movaps  16(%2,%0), %%xmm5 \n\t"
+                "movaps  16(%1,%0), %%xmm4 \n\t"
                 "movaps     %%xmm1, %%xmm2 \n\t"
+                "movaps     %%xmm5, %%xmm6 \n\t"
                 "shufps      $0xA0, %%xmm1, %%xmm1 \n\t"
                 "shufps      $0xF5, %%xmm2, %%xmm2 \n\t"
+                "shufps      $0xA0, %%xmm5, %%xmm5 \n\t"
+                "shufps      $0xF5, %%xmm6, %%xmm6 \n\t"
                 "mulps   (%3,%0,2), %%xmm1 \n\t" //  cre*re cim*re
                 "mulps 16(%3,%0,2), %%xmm2 \n\t" // -cim*im cre*im
+                "mulps 32(%3,%0,2), %%xmm5 \n\t" //  cre*re cim*re
+                "mulps 48(%3,%0,2), %%xmm6 \n\t" // -cim*im cre*im
                 "addps      %%xmm2, %%xmm1 \n\t"
+                "addps      %%xmm6, %%xmm5 \n\t"
                 "movaps     %%xmm0, %%xmm3 \n\t"
+                "movaps     %%xmm4, %%xmm7 \n\t"
                 "addps      %%xmm1, %%xmm0 \n\t"
                 "subps      %%xmm1, %%xmm3 \n\t"
+                "addps      %%xmm5, %%xmm4 \n\t"
+                "subps      %%xmm5, %%xmm7 \n\t"
                 "movaps     %%xmm0, (%1,%0) \n\t"
                 "movaps     %%xmm3, (%2,%0) \n\t"
+                "movaps     %%xmm4, 16(%1,%0) \n\t"
+                "movaps     %%xmm7, 16(%2,%0) \n\t"
                 "jg 1b \n\t"
                 :"+r"(i)
                 :"r"(p), "r"(p + nloops), "r"(cptr)
@@ -141,67 +154,106 @@ void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output,
     n4 = n >> 2;
     n8 = n >> 3;
 
-    asm volatile ("movaps %0, %%xmm7\n\t"::"m"(*p1m1p1m1));
+#ifdef ARCH_X86_64
+    asm volatile ("movaps %0, %%xmm8\n\t"::"m"(*p1m1p1m1));
+#define P1M1P1M1 "%%xmm8"
+#else
+#define P1M1P1M1 "%4"
+#endif
 
     /* pre rotation */
     in1 = input;
     in2 = input + n2 - 4;
 
-    /* Complex multiplication
-       Two complex products per iteration, we could have 4 with 8 xmm
-       registers, 8 with 16 xmm registers.
-       Maybe we should unroll more.
-    */
-    for (k = 0; k < n4; k += 2) {
+    /* Complex multiplication */
+    for (k = 0; k < n4; k += 4) {
         asm volatile (
             "movaps          %0, %%xmm0 \n\t"   // xmm0 = r0 X  r1 X : in2
             "movaps          %1, %%xmm3 \n\t"   // xmm3 = X  i1 X  i0: in1
+            "movaps      -16+%0, %%xmm4 \n\t"   // xmm4 = r0 X  r1 X : in2
+            "movaps       16+%1, %%xmm7 \n\t"   // xmm7 = X  i1 X  i0: in1
             "movlps          %2, %%xmm1 \n\t"   // xmm1 = X  X  R1 R0: tcos
             "movlps          %3, %%xmm2 \n\t"   // xmm2 = X  X  I1 I0: tsin
+            "movlps        8+%2, %%xmm5 \n\t"   // xmm5 = X  X  R1 R0: tcos
+            "movlps        8+%3, %%xmm6 \n\t"   // xmm6 = X  X  I1 I0: tsin
             "shufps $95, %%xmm0, %%xmm0 \n\t"   // xmm0 = r1 r1 r0 r0
             "shufps $160,%%xmm3, %%xmm3 \n\t"   // xmm3 = i1 i1 i0 i0
+            "shufps $95, %%xmm4, %%xmm4 \n\t"   // xmm4 = r1 r1 r0 r0
+            "shufps $160,%%xmm7, %%xmm7 \n\t"   // xmm7 = i1 i1 i0 i0
             "unpcklps    %%xmm2, %%xmm1 \n\t"   // xmm1 = I1 R1 I0 R0
+            "unpcklps    %%xmm6, %%xmm5 \n\t"   // xmm5 = I1 R1 I0 R0
             "movaps      %%xmm1, %%xmm2 \n\t"   // xmm2 = I1 R1 I0 R0
-            "xorps       %%xmm7, %%xmm2 \n\t"   // xmm2 = -I1 R1 -I0 R0
+            "movaps      %%xmm5, %%xmm6 \n\t"   // xmm6 = I1 R1 I0 R0
+            "xorps   "P1M1P1M1", %%xmm2 \n\t"   // xmm2 = -I1 R1 -I0 R0
+            "xorps   "P1M1P1M1", %%xmm6 \n\t"   // xmm6 = -I1 R1 -I0 R0
             "mulps       %%xmm1, %%xmm0 \n\t"   // xmm0 = rI rR rI rR
+            "mulps       %%xmm5, %%xmm4 \n\t"   // xmm4 = rI rR rI rR
             "shufps $177,%%xmm2, %%xmm2 \n\t"   // xmm2 = R1 -I1 R0 -I0
+            "shufps $177,%%xmm6, %%xmm6 \n\t"   // xmm6 = R1 -I1 R0 -I0
             "mulps       %%xmm2, %%xmm3 \n\t"   // xmm3 = Ri -Ii Ri -Ii
+            "mulps       %%xmm6, %%xmm7 \n\t"   // xmm7 = Ri -Ii Ri -Ii
             "addps       %%xmm3, %%xmm0 \n\t"   // xmm0 = result
+            "addps       %%xmm7, %%xmm4 \n\t"   // xmm4 = result
             ::"m"(in2[-2*k]), "m"(in1[2*k]),
               "m"(tcos[k]), "m"(tsin[k])
+#ifndef ARCH_X86_64
+              ,"m"(*p1m1p1m1)
+#endif
         );
         /* Should be in the same block, hack for gcc2.95 & gcc3 */
         asm (
             "movlps      %%xmm0, %0     \n\t"
             "movhps      %%xmm0, %1     \n\t"
-            :"=m"(z[revtab[k]]), "=m"(z[revtab[k + 1]])
+            "movlps      %%xmm4, %2     \n\t"
+            "movhps      %%xmm4, %3     \n\t"
+            :"=m"(z[revtab[k]]), "=m"(z[revtab[k + 1]]),
+             "=m"(z[revtab[k + 2]]), "=m"(z[revtab[k + 3]])
         );
     }
 
     ff_fft_calc_sse(&s->fft, z);
 
-    /* Not currently needed, added for safety */
-    asm volatile ("movaps %0, %%xmm7\n\t"::"m"(*p1m1p1m1));
+#ifndef ARCH_X86_64
+#undef P1M1P1M1
+#define P1M1P1M1 "%3"
+#endif
 
     /* post rotation + reordering */
-    for (k = 0; k < n4; k += 2) {
+    for (k = 0; k < n4; k += 4) {
         asm (
             "movaps          %0, %%xmm0 \n\t"   // xmm0 = i1 r1 i0 r0: z
+            "movaps       16+%0, %%xmm4 \n\t"   // xmm4 = i1 r1 i0 r0: z
             "movlps          %1, %%xmm1 \n\t"   // xmm1 = X  X  R1 R0: tcos
+            "movlps        8+%1, %%xmm5 \n\t"   // xmm5 = X  X  R1 R0: tcos
             "movaps      %%xmm0, %%xmm3 \n\t"   // xmm3 = i1 r1 i0 r0
+            "movaps      %%xmm4, %%xmm7 \n\t"   // xmm7 = i1 r1 i0 r0
             "movlps          %2, %%xmm2 \n\t"   // xmm2 = X  X  I1 I0: tsin
+            "movlps        8+%2, %%xmm6 \n\t"   // xmm6 = X  X  I1 I0: tsin
             "shufps $160,%%xmm0, %%xmm0 \n\t"   // xmm0 = r1 r1 r0 r0
             "shufps $245,%%xmm3, %%xmm3 \n\t"   // xmm3 = i1 i1 i0 i0
+            "shufps $160,%%xmm4, %%xmm4 \n\t"   // xmm4 = r1 r1 r0 r0
+            "shufps $245,%%xmm7, %%xmm7 \n\t"   // xmm7 = i1 i1 i0 i0
             "unpcklps    %%xmm2, %%xmm1 \n\t"   // xmm1 = I1 R1 I0 R0
+            "unpcklps    %%xmm6, %%xmm5 \n\t"   // xmm5 = I1 R1 I0 R0
             "movaps      %%xmm1, %%xmm2 \n\t"   // xmm2 = I1 R1 I0 R0
-            "xorps       %%xmm7, %%xmm2 \n\t"   // xmm2 = -I1 R1 -I0 R0
+            "movaps      %%xmm5, %%xmm6 \n\t"   // xmm6 = I1 R1 I0 R0
+            "xorps   "P1M1P1M1", %%xmm2 \n\t"   // xmm2 = -I1 R1 -I0 R0
             "mulps       %%xmm1, %%xmm0 \n\t"   // xmm0 = rI rR rI rR
+            "xorps   "P1M1P1M1", %%xmm6 \n\t"   // xmm6 = -I1 R1 -I0 R0
+            "mulps       %%xmm5, %%xmm4 \n\t"   // xmm4 = rI rR rI rR
             "shufps $177,%%xmm2, %%xmm2 \n\t"   // xmm2 = R1 -I1 R0 -I0
+            "shufps $177,%%xmm6, %%xmm6 \n\t"   // xmm6 = R1 -I1 R0 -I0
             "mulps       %%xmm2, %%xmm3 \n\t"   // xmm3 = Ri -Ii Ri -Ii
+            "mulps       %%xmm6, %%xmm7 \n\t"   // xmm7 = Ri -Ii Ri -Ii
             "addps       %%xmm3, %%xmm0 \n\t"   // xmm0 = result
+            "addps       %%xmm7, %%xmm4 \n\t"   // xmm4 = result
             "movaps      %%xmm0, %0     \n\t"
+            "movaps      %%xmm4, 16+%0  \n\t"
             :"+m"(z[k])
             :"m"(tcos[k]), "m"(tsin[k])
+#ifndef ARCH_X86_64
+             ,"m"(*p1m1p1m1)
+#endif
         );
     }