sbrdsp: Unroll and use integer operations

This patch can be controversial, by assuming floats are IEEE-754 and
particular behaviour of the FPU will get in the way.
Timing on Arrandale and Win32 (thus, x87 FPU is used in the reference).

sbr_qmf_pre_shuffle_c: 115 to 76
sbr_neg_odd_64_c: 84 to 55
sbr_qmf_post_shuffle_c: 112 to 83
Signed-off-by: Diego Biurrun <diego@biurrun.de>

libavcodec/sbrdsp.c

@@ -22,6 +22,7 @@
 
 #include "config.h"
 #include "libavutil/attributes.h"
+#include "libavutil/intfloat.h"
 #include "sbrdsp.h"
 
 static void sbr_sum64x5_c(float *z)
@@ -51,37 +52,51 @@ static float sbr_sum_square_c(float (*x)[2], int n)
 
 static void sbr_neg_odd_64_c(float *x)
 {
+    union av_intfloat32 *xi = (union av_intfloat32*) x;
     int i;
-    for (i = 1; i < 64; i += 2)
-        x[i] = -x[i];
+    for (i = 1; i < 64; i += 4) {
+        xi[i + 0].i ^= 1U << 31;
+        xi[i + 2].i ^= 1U << 31;
+    }
 }
 
 static void sbr_qmf_pre_shuffle_c(float *z)
 {
+    union av_intfloat32 *zi = (union av_intfloat32*) z;
     int k;
-    z[64] = z[0];
-    z[65] = z[1];
-    for (k = 1; k < 32; k++) {
-        z[64+2*k  ] = -z[64 - k];
-        z[64+2*k+1] =  z[ k + 1];
+    zi[64].i = zi[0].i;
+    zi[65].i = zi[1].i;
+    for (k = 1; k < 31; k += 2) {
+        zi[64 + 2 * k + 0].i = zi[64 - k].i ^ (1U << 31);
+        zi[64 + 2 * k + 1].i = zi[ k + 1].i;
+        zi[64 + 2 * k + 2].i = zi[63 - k].i ^ (1U << 31);
+        zi[64 + 2 * k + 3].i = zi[ k + 2].i;
     }
+    zi[64 + 2 * 31 + 0].i = zi[64 - 31].i ^ (1U << 31);
+    zi[64 + 2 * 31 + 1].i = zi[31 +  1].i;
 }
 
 static void sbr_qmf_post_shuffle_c(float W[32][2], const float *z)
 {
+    const union av_intfloat32 *zi = (const union av_intfloat32*) z;
+    union av_intfloat32 *Wi       = (union av_intfloat32*) W;
     int k;
-    for (k = 0; k < 32; k++) {
-        W[k][0] = -z[63-k];
-        W[k][1] = z[k];
+    for (k = 0; k < 32; k += 2) {
+        Wi[2 * k + 0].i = zi[63 - k].i ^ (1U << 31);
+        Wi[2 * k + 1].i = zi[ k + 0].i;
+        Wi[2 * k + 2].i = zi[62 - k].i ^ (1U << 31);
+        Wi[2 * k + 3].i = zi[ k + 1].i;
     }
 }
 
 static void sbr_qmf_deint_neg_c(float *v, const float *src)
 {
+    const union av_intfloat32 *si = (const union av_intfloat32*)src;
+    union av_intfloat32 *vi = (union av_intfloat32*)v;
     int i;
     for (i = 0; i < 32; i++) {
-        v[     i] =  src[63 - 2*i    ];
-        v[63 - i] = -src[63 - 2*i - 1];
+        vi[     i].i = si[63 - 2 * i    ].i;
+        vi[63 - i].i = si[63 - 2 * i - 1].i ^ (1U << 31);
     }
 }