+#if defined(CAN_COMPILE_3DNOW)
+static void Merge3DNow( void *_p_dest, const void *_p_s1, const void *_p_s2,
+ size_t i_bytes )
+{
+ uint8_t* p_dest = (uint8_t*)_p_dest;
+ const uint8_t *p_s1 = (const uint8_t *)_p_s1;
+ const uint8_t *p_s2 = (const uint8_t *)_p_s2;
+ uint8_t* p_end = p_dest + i_bytes - 8;
+ while( p_dest < p_end )
+ {
+ __asm__ __volatile__( "movq %2,%%mm1;"
+ "pavgusb %1, %%mm1;"
+ "movq %%mm1, %0" :"=m" (*p_dest):
+ "m" (*p_s1),
+ "m" (*p_s2) );
+ p_dest += 8;
+ p_s1 += 8;
+ p_s2 += 8;
+ }
+
+ p_end += 8;
+
+ while( p_dest < p_end )
+ {
+ *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
+ }
+}
+#endif
+