1@localsize 16, 1, 1 2@buf 128 (c2.x) ; c2.xy 3@invocationid(r0.x) ; r0.xyz 4mov.u32u32 r0.y, r0.x 5mov.u32u32 r1.x, c2.x 6mov.u32u32 r1.y, c2.y 7mov.u32u32 r2.x, 0xff 8(rpt5)nop 9stg.a.u32 g[r1.x+r0.y<<4+2<<2], r2.x, 1 10nop(sy) 11ldg.a.u32 r4.x, g[r1.x+r0.y<<4+2<<2], 1 12nop(sy) 13add.u r4.x, r4.x, 1 14(rpt3)nop 15stg.a.u32 g[r1.x+r0.y<<4+1<<2], r4.x, 1 16end 17nop 18