#define PRECISION 16
#define fixmul32(x, y)  \
    ({ int32_t __hi;  \
       uint32_t __lo;  \
       int32_t __result;  \
       asm ("smull   %0, %1, %3, %4\n\t"  \
            "movs    %0, %0, lsr %5\n\t"  \
            "adc    %2, %0, %1, lsl %6"  \
            : "=&r" (__lo), "=&r" (__hi), "=r" (__result)  \
            : "%r" (x), "r" (y),  \
              "M" (PRECISION), "M" (32 - PRECISION)  \
            : "cc");  \
       __result;  \
    })



tmp = fixmul32(output[i].r,  output[i].r) + fixmul32(output[i].i, output[i].i);