.version 6.5
.target sm_30
.address_size 64

.visible .entry ntid(
	.param .u64 input,
	.param .u64 output
)
{
	.reg .u64 	    in_addr;
    .reg .u64 	    out_addr;
    .reg .u32 	    in_val;
    .reg .u32 	    global_count;

	ld.param.u64 	in_addr, [input];
    ld.param.u64 	out_addr, [output];

    ld.u32          in_val, [in_addr];
    mov.u32         global_count, %ntid.x;
	add.u32		    in_val, in_val, global_count;
    st.u32          [out_addr], in_val;
	ret;
}