cpu_asimd.c 818 B

123456789101112131415161718192021222324252627
  1. #ifdef _MSC_VER
  2. #include <Intrin.h>
  3. #endif
  4. #include <arm_neon.h>
  5. int main(int argc, char **argv)
  6. {
  7. float *src = (float*)argv[argc-1];
  8. float32x4_t v1 = vdupq_n_f32(src[0]), v2 = vdupq_n_f32(src[1]);
  9. /* MAXMIN */
  10. int ret = (int)vgetq_lane_f32(vmaxnmq_f32(v1, v2), 0);
  11. ret += (int)vgetq_lane_f32(vminnmq_f32(v1, v2), 0);
  12. /* ROUNDING */
  13. ret += (int)vgetq_lane_f32(vrndq_f32(v1), 0);
  14. #ifdef __aarch64__
  15. {
  16. double *src2 = (double*)argv[argc-1];
  17. float64x2_t vd1 = vdupq_n_f64(src2[0]), vd2 = vdupq_n_f64(src2[1]);
  18. /* MAXMIN */
  19. ret += (int)vgetq_lane_f64(vmaxnmq_f64(vd1, vd2), 0);
  20. ret += (int)vgetq_lane_f64(vminnmq_f64(vd1, vd2), 0);
  21. /* ROUNDING */
  22. ret += (int)vgetq_lane_f64(vrndq_f64(vd1), 0);
  23. }
  24. #endif
  25. return ret;
  26. }