// SM89 kernel: write the scalar value to the supplied device pointer.
.target SM89
.kernel write_value
.param output, 8, 8
.param value, 4, 4
.registers 8, 8

// Explicit, deliberately slow scheduling for this small example.
.schedule usched_info=WAIT15_END_GROUP, req_bit_set=63
MOV R1, c[0][0x28];
MOV R2, c[0][$output];
MOV R3, c[0][$output+4];
MOV R5, c[0][$value];
ULDC.64 UR4, c[0][0x118];
STG.E [R2.64], R5;
EXIT;
.endkernel