Advertisement
Not a member of Pastebin yet?
Sign Up,
it unlocks many cool features!
- LLVM:
- disasm:
- main:
- BB19_0:
- s_mov_b64 exec, -1 ; BEFE04C1
- s_barrier ; BF8A0000
- v_mbcnt_lo_u32_b32_e64 v0, -1, 0 ; D7650000 000100C1
- s_bfe_u32 s2, s3, 0x80008 ; 9382FF03 00080008
- v_mbcnt_hi_u32_b32_e64 v0, -1, v0 ; D7660000 000200C1
- v_cmp_gt_u32_e32 vcc, s2, v0 ; 7D880002
- s_and_saveexec_b64 s[2:3], vcc ; BE82246A
- s_cbranch_execz BB19_3 ; BF880000
- BB19_1:
- s_movk_i32 s9, 0x8000 ; B0098000
- v_and_b32_e32 v0, 0xff, v1 ; 360002FF 000000FF
- v_and_b32_e32 v1, 0x1f00, v1 ; 360202FF 00001F00
- s_load_dwordx4 s[8:11], s[8:9], 0x0 ; F4080204 FA000000
- v_lshlrev_b32_e32 v2, 5, v0 ; 34040085
- v_cmp_eq_u32_e32 vcc, 0, v1 ; 7D840280
- s_waitcnt lgkmcnt(0) ; BF8CC07F
- buffer_load_dwordx4 v[3:6], off, s[8:11], 0 ; E0380000 80020300
- buffer_load_dwordx2 v[7:8], off, s[8:11], 0 offset:16 ; E0340010 80020700
- s_waitcnt vmcnt(1) ; BF8C3F71
- ds_write2_b32 v2, v3, v4 offset0:4 offset1:5 ; D8380504 00040302
- ds_write2_b32 v2, v5, v6 offset1:1 ; D8380100 00060502
- s_waitcnt vmcnt(0) ; BF8C3F70
- ds_write2_b32 v2, v7, v8 offset0:2 offset1:3 ; D8380302 00080702
- s_waitcnt lgkmcnt(0) ; BF8CC07F
- s_barrier ; BF8A0000
- v_nop ; 7E000000
- s_and_b64 exec, exec, vcc ; 87FE6A7E
- s_cbranch_execz BB19_3 ; BF880000
- BB19_2:
- ds_read_b64 v[3:4], v2 ; D9D80000 03000002
- s_load_dwordx4 s[0:3], s[0:1], 0x50 ; F4080000 FA000050
- ds_read2_b32 v[5:6], v2 offset0:2 offset1:4 ; D8DC0402 05000002
- v_lshlrev_b32_e32 v0, 4, v0 ; 34000084
- s_waitcnt lgkmcnt(0) ; BF8CC07F
- buffer_store_dwordx4 v[3:6], v0, s[0:3], s4 offen glc ; E0785000 04000300
- BB19_3:
- s_endpgm ; BF810000
- ACO:
- disasm:
- BB0:
- s_mov_b64 exec, -1 ; befe04c1
- s_barrier ; bf8a0000
- s_bfe_u32 s2, s3, 0x80008 ; 9382ff03 00080008
- v_mbcnt_lo_u32_b32_e64 v0, -1, 0 ; d7650000 000100c1
- v_mbcnt_hi_u32_b32_e64 v0, -1, v0 ; d7660000 000200c1
- v_cmp_gt_u32_e32 vcc, s2, v0 ; 7d880002
- s_and_saveexec_b64 s[2:3], vcc ; be82246a
- s_cbranch_execz BB7 ; bf880034
- BB1:
- s_movk_i32 s9, 0x8000 ; b0098000
- s_load_dwordx4 s[8:11], s[8:9], 0x0 ; f4080204 fa000000
- v_and_b32_e32 v0, 0xff, v1 ; 360002ff 000000ff
- v_lshlrev_b32_e32 v2, 5, v0 ; 34040085
- s_movk_i32 m0, 0xffff ; b07cffff
- s_waitcnt lgkmcnt(0) ; bf8cc07f
- s_buffer_load_dwordx4 s[12:15], s[8:11], 0x0 ; f4280304 fa000000
- s_buffer_load_dwordx2 s[6:7], s[8:11], 0x10 ; f4240184 fa000010
- s_waitcnt lgkmcnt(0) ; bf8cc07f
- v_mov_b32_e32 v3, s12 ; 7e06020c
- ds_write_b32 v2, v3 offset:4 ; d8340004 00000302
- v_mov_b32_e32 v3, s13 ; 7e06020d
- v_mov_b32_e32 v4, s14 ; 7e08020e
- v_mov_b32_e32 v5, s15 ; 7e0a020f
- v_mov_b32_e32 v6, s6 ; 7e0c0206
- v_mov_b32_e32 v7, s7 ; 7e0e0207
- ds_write_b32 v2, v3 offset:5 ; d8340005 00000302
- ds_write_b32 v2, v4 ; d8340000 00000402
- ds_write_b32 v2, v5 offset:1 ; d8340001 00000502
- ds_write_b32 v2, v6 offset:2 ; d8340002 00000602
- ds_write_b32 v2, v7 offset:3 ; d8340003 00000702
- s_waitcnt lgkmcnt(0) ; bf8cc07f
- s_barrier ; bf8a0000
- v_bfe_u32 v1, v1, 8, 5 ; d5480001 02151101
- v_cmp_eq_u32_e32 vcc, 0, v1 ; 7d840280
- s_and_saveexec_b64 s[6:7], vcc ; be86246a
- s_cbranch_execz BB7 ; bf88000e
- BB2:
- ds_read_b32 v3, v2 ; d8d80000 03000002
- ds_read_b32 v4, v2 offset:1 ; d8d80001 04000002
- ds_read_b32 v5, v2 offset:2 ; d8d80002 05000002
- s_load_dwordx4 s[8:11], s[0:1], 0x50 ; f4080200 fa000050
- ds_read_b32 v6, v2 offset:4 ; d8d80004 06000002
- v_lshlrev_b32_e32 v0, 4, v0 ; 34000084
- s_waitcnt lgkmcnt(0) ; bf8cc07f
- buffer_store_dwordx4 v[3:6], v0, s[8:11], s4 offen glc ; e0785000 04020300
- BB7:
- BB12:
- v_nop ; 7e000000
- s_waitcnt_vscnt null, 0x0 ; bbfd0000
- s_mov_b64 exec, s[2:3] ; befe0402
- s_endpgm ; bf810000
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement