void sinCos_q27(uint32_t theta, int32_t &c, int32_t &s) {
  int32_t a = (theta & ((1 << 20) - 1)) << 8; // q28
  uint32_t i = theta >> 20;
  s = ___SMMLA(a, sine2t[i + 1] - sine2t[i], sine2t[i] >> 4);
  i = (i + 1024) & 4095;
  c = ___SMMLA(a, sine2t[i + 1] - sine2t[i], sine2t[i] >> 4);
}

int32_t c, s;
{
  uint32_t theta = ((param_theta + inlet_theta) & 0x07FFFFFF) << 5;
  sinCos_q27(theta, c, s);
}

// note: read/write indexes are set BUFSIZE samples back as we will process a
// batch of BUFSIZE samples so beware of overlappings when param_delays defer by
// less than 16 samples
int32_t a_rw = (attr_delayName.writepos - (BUFSIZE - 1) - param_pos) &
               attr_delayName.LENGTHMASK;

// _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
for (int i = 0; i < BUFSIZE; i++) {
  int32_t xA = attr_delayName.array[a_rw] << (27 - 15);
  int32_t xB = inlet_in[i];
  attr_delayName.array[a_rw] =
      __SSAT(___SMMLS(xB, s, ___SMMUL(xA, c)) >> (22 - 15), 16);
  a_rw = (a_rw + 1) & attr_delayName.LENGTHMASK;
  outlet_out[i] = ___SMMLA(xB, c, ___SMMUL(xA, s)) << 5;
}