Custom renderscript is slower than intrinsic

Question

I have a simple custom renderscript to convert NV21 to RGB bitmap:

#pragma rs_fp_relaxed

rs_allocation NV21;
uint32_t Yline