shader main
type(CS)

user_sgpr_count(9)   // 2 for the buffer resource + 5 for thread/thread group parameters

//s[0:1] the mmeory address for the buffer resource
//s2 x
//s3 x*y
//s4 x*y*z
//s5 X	
//s6 X*Y
//s7 output offset
//s8 loop

tgid_x_en(1)       //s_tgid_x s9
tgid_y_en(1)       //s_tgid_y s10
tgid_z_en(1)       //s_tgid_z s11

//vo for tid_x
//v1 for tid_y
//v2 for tid_z

s_mov_b32 s16, s2

//SPI may touch v0,v1,v2 before shader is run

//store it 10 times
v_mov_b32 v10, v1
v_mov_b32 v11, v2
v_mov_b32 v12, v1
v_mov_b32 v13, v2
v_mov_b32 v14, v1
v_mov_b32 v15, v2
v_mov_b32 v16, v1
v_mov_b32 v17, v2
v_mov_b32 v18, v1
v_mov_b32 v19, v0

// read them back
v_mov_b32 v29, v10
v_mov_b32 v28, v11
v_mov_b32 v27, v12
v_mov_b32 v26, v13
v_mov_b32 v25, v14
v_mov_b32 v24, v15
v_mov_b32 v23, v16
v_mov_b32 v22, v17
v_mov_b32 v21, v18
v_mov_b32 v20, v19

s_store_dword s16, s[0:1], 0x0 glc

s_endpgm
end