Faster 16/32bit 2D Morton Code encode/decode functions

morton.h

uint16_t encode16_morton2(uint8_t x_, uint8_t y_)
{
  uint32_t res=x_|(uint32_t(y_)<<16);
  res=(res|(res<<4))&0x0f0f0f0f;
  res=(res|(res<<2))&0x33333333;
  res=(res|(res<<1))&0x55555555;
  return uint16_t(res|(res>>15));
}
//----

uint32_t encode32_morton2(uint16_t x_, uint16_t y_)
{
  uint64_t res=x_|(uint64_t(y_)<<32);
  res=(res|(res<<8))&0x00ff00ff00ff00ff;
  res=(res|(res<<4))&0x0f0f0f0f0f0f0f0f;
  res=(res|(res<<2))&0x3333333333333333;
  res=(res|(res<<1))&0x5555555555555555;
  return uint32_t(res|(res>>31));
}
//----

void decode16_morton2(uint8_t &x_, uint8_t &y_, uint16_t mc_)
{
  uint32_t res=(mc_|(uint32_t(mc_)<<15))&0x55555555;
  res=(res|(res>>1))&0x33333333;
  res=(res|(res>>2))&0x0f0f0f0f;
  res=res|(res>>4);
  x_=uint8_t(res);
  y_=uint8_t(res>>16);
}
//----

void decode32_morton2(uint16_t &x_, uint16_t &y_, uint32_t mc_)
{
  uint64_t res=(mc_|(uint64_t(mc_)<<31))&0x5555555555555555;
  res=(res|(res>>1))&0x3333333333333333;
  res=(res|(res>>2))&0x0f0f0f0f0f0f0f0f;
  res=(res|(res>>4))&0x00ff00ff00ff00ff;
  res=res|(res>>8);
  x_=uint16_t(res);
  y_=uint16_t(res>>32);
}

Even better:
return _pdep_u32(x, 0x55555555) | _pdep_u32(y, 0xaaaaaaaa);

See: https://twitter.com/JarkkoPFC/status/1443462656669663235