Geschwindigkeitsoptimierung für die Kreuzkorrelation, MMX ist doch nicht so vorteilhaft... seltsam?

detect.cc

@@ -1,5 +1,7 @@
 
 
+#include <stdint.h>
+#include <mmintrin.h>
 #include <unistd.h>
 #include <math.h>
 #include "config.h"
@@ -240,6 +242,14 @@ void Detect::InputDetect(int *posx, int *posy) {
 #endif
 }
 
+inline float calc_vector(uint8_t a1, uint8_t a2, uint8_t b1, uint8_t b2, uint8_t c1, uint8_t c2, uint8_t d1, uint8_t d2) {
+  return (float) (a1 * a2 + b1 * b2 + c1 * c2 + d1 * d2);
+  // __m64 m1 = _mm_set_pi16 (a1, b1, c1, d1 );
+  // __m64 m2 = _mm_set_pi16 (a2, b2, c2, d2 );
+  // __m64 m3 = _mm_madd_pi16 (m1, m2);
+  // uint32_t *data = (uint32_t *) &m3;
+  // return (float) (data[0] + data[1]);
+}
 
 #define OBJSIZE 50
 #define MAXSHIFT 20
@@ -304,7 +314,12 @@ void Detect::InputDetectCrossC(int *posx, int *posy) {
 				for (x = 0; x < OBJSIZE; x++, oldi += 3, ini += 3, oldx++, inx++) {
 					if (oldx >= 0 && oldy >= 0 && oldx < oldFrame.w && oldy <= oldFrame.h &&
 						inx >= 0 && inx < inFrame.w && iny >= 0 && iny < inFrame.h) {
-							f += (float)(pxo[oldi+0])*(float)(pxi[ini+0]);
+							//f += (float)(pxo[oldi+0])*(float)(pxi[ini+0]);
+							f += calc_vector( pxo[oldi+0], pxi[ini+0],
+									  pxo[oldi+3], pxi[ini+3],
+									  pxo[oldi+6], pxi[ini+6],
+									  pxo[oldi+9], pxi[ini+9]);
+							x+=3; oldi += 9; ini += 9; oldx+=3; inx+=3;
 					}
 				}
 			}