ATTENTION: default value of option vblank_mode overridden by environment. radeonsi: Compiling shader 4 radeonsi: Compiling shader 5 SHADER KEY part.ps.prolog.color_two_side = 0 part.ps.prolog.flatshade_colors = 0 part.ps.prolog.poly_stipple = 0 part.ps.prolog.force_persp_sample_interp = 0 part.ps.prolog.force_linear_sample_interp = 0 part.ps.prolog.force_persp_center_interp = 0 part.ps.prolog.force_linear_center_interp = 0 part.ps.prolog.bc_optimize_for_persp = 0 part.ps.prolog.bc_optimize_for_linear = 0 part.ps.epilog.spi_shader_col_format = 0x44444444 part.ps.epilog.color_is_int8 = 0x0 part.ps.epilog.last_cbuf = 0 part.ps.epilog.alpha_func = 7 part.ps.epilog.alpha_to_one = 0 part.ps.epilog.poly_line_smoothing = 0 part.ps.epilog.clamp_color = 0 Pixel Shader: Shader main disassembly: s_mov_b32 s84, SCRATCH_RSRC_DWORD0 ; BED400FF 00000000 s_mov_b32 s85, SCRATCH_RSRC_DWORD1 ; BED500FF 00000000 s_mov_b32 s86, -1 ; BED600C1 s_mov_b32 s89, s12 ; BED9000C s_mov_b32 s87, 0xe80000 ; BED700FF 00E80000 s_add_u32 m0, s89, 0x100 ; 807CFF59 00000100 s_buffer_store_dword exec_lo, s[84:87], m0 ; C0601FAA 0000007C s_add_u32 m0, s89, 0x200 ; 807CFF59 00000200 s_buffer_store_dword exec_hi, s[84:87], m0 ; C0601FEA 0000007C v_mov_b32_e32 v16, v13 ; 7E20030D s_waitcnt lgkmcnt(0) ; BF8C007F s_wqm_b64 exec, exec ; BEFE077E s_mov_b32 m0, s89 ; BEFC0059 s_mov_b32 s19, s11 ; BE93000B s_buffer_store_dword s10, s[84:87], m0 ; C06002AA 0000007C s_waitcnt lgkmcnt(0) ; BF8C007F s_load_dwordx8 s[8:15], s[4:5], 0x140 ; C00E0202 00000140 s_mov_b32 m0, s19 ; BEFC0013 v_interp_p1_f32 v4, v2, 0, 0, [m0] ; D4100002 v_interp_p2_f32 v4, [v4], v3, 0, 0, [m0] ; D4110003 v_interp_p1_f32 v5, v2, 1, 0, [m0] ; D4140102 v_interp_p2_f32 v5, [v5], v3, 1, 0, [m0] ; D4150103 s_mov_b32 vcc_hi, m0 ; BEEB007C s_add_u32 m0, s89, 0x300 ; 807CFF59 00000300 s_mov_b64 s[6:7], s[2:3] ; BE860102 s_waitcnt lgkmcnt(0) ; BF8C007F s_buffer_store_dword s8, s[84:87], m0 ; C060022A 0000007C s_add_u32 m0, s89, 0x400 ; 807CFF59 00000400 s_load_dwordx4 s[36:39], s[6:7], 0x40 ; C00A0903 00000040 s_load_dwordx4 s[0:3], s[4:5], 0x70 ; C00A0002 00000070 s_load_dwordx4 s[24:27], s[4:5], 0x130 ; C00A0602 00000130 s_load_dwordx8 s[52:59], s[4:5], 0x100 ; C00E0D02 00000100 v_mul_f32_e32 v6, 0, v4 ; 0A0C0880 s_buffer_store_dword s9, s[84:87], m0 ; C060026A 0000007C s_add_u32 m0, s89, 0x500 ; 807CFF59 00000500 s_buffer_store_dword s10, s[84:87], m0 ; C06002AA 0000007C s_add_u32 m0, s89, 0x600 ; 807CFF59 00000600 s_buffer_store_dword s11, s[84:87], m0 ; C06002EA 0000007C s_waitcnt lgkmcnt(0) ; BF8C007F s_load_dwordx4 s[8:11], s[6:7], 0x0 ; C00A0203 00000000 s_add_u32 m0, s89, 0x700 ; 807CFF59 00000700 s_buffer_store_dword s12, s[84:87], m0 ; C060032A 0000007C s_add_u32 m0, s89, 0x800 ; 807CFF59 00000800 s_buffer_store_dword s13, s[84:87], m0 ; C060036A 0000007C s_add_u32 m0, s89, 0x900 ; 807CFF59 00000900 s_buffer_load_dword s7, s[36:39], 0x52c ; C02201D2 0000052C s_buffer_load_dword s6, s[36:39], 0x528 ; C0220192 00000528 s_waitcnt lgkmcnt(0) ; BF8C007F s_buffer_load_dword s22, s[8:11], 0x70 ; C0220584 00000070 v_mov_b32_e32 v20, v7 ; 7E280307 s_buffer_store_dword s14, s[84:87], m0 ; C06003AA 0000007C s_add_u32 m0, s89, 0xa00 ; 807CFF59 00000A00 s_buffer_store_dword s15, s[84:87], m0 ; C06003EA 0000007C s_waitcnt lgkmcnt(0) ; BF8C007F s_load_dwordx8 s[8:15], s[4:5], 0x40 ; C00E0202 00000040 v_mov_b32_e32 v18, v5 ; 7E240305 v_add_f32_e32 v13, s7, v5 ; 021A0A07 v_mov_b32_e32 v11, v5 ; 7E160305 v_mov_b32_e32 v12, v6 ; 7E180306 v_mov_b32_e32 v19, v6 ; 7E260306 v_mov_b32_e32 v17, v4 ; 7E220304 v_mov_b32_e32 v7, 0xc2000000 ; 7E0E02FF C2000000 v_add_f32_e32 v10, s6, v4 ; 02140806 v_mov_b32_e32 v18, v13 ; 7E24030D s_waitcnt lgkmcnt(0) ; BF8C007F image_sample_l v14, v[4:7], s[8:15], s[0:3] dmask:0x1 ; F0900100 00020E04 s_nop 0 ; BF800000 image_sample_l v[0:1], v[4:7], s[52:59], s[24:27] dmask:0x3 ; F0900300 00CD0004 s_nop 0 ; BF800000 image_sample_l v15, v[10:13], s[8:15], s[0:3] dmask:0x1 ; F0900100 00020F0A s_mov_b32 s16, 0x40ff0000 ; BE9000FF 40FF0000 image_sample_l v17, v[17:20], s[8:15], s[0:3] dmask:0x1 ; F0900100 00021111 v_mov_b32_e32 v11, v13 ; 7E16030D image_sample_l v6, v[10:13], s[8:15], s[0:3] dmask:0x1 ; F0900100 0002060A s_waitcnt vmcnt(1) ; BF8C0F71 v_mul_f32_e32 v12, s16, v17 ; 0A182210 v_floor_f32_e32 v12, v12 ; 7E183F0C v_mad_f32 v12, v17, s16, -v12 ; D1C1000C 84302111 v_mad_f32 v13, v17, s16, -v12 ; D1C1000D 84302111 v_mul_f32_e32 v12, s16, v15 ; 0A181E10 v_floor_f32_e32 v12, v12 ; 7E183F0C v_mul_f32_e32 v10, s16, v14 ; 0A141C10 v_mad_f32 v12, v15, s16, -v12 ; D1C1000C 8430210F v_mad_f32 v24, v15, s16, -v12 ; D1C10018 8430210F v_floor_f32_e32 v10, v10 ; 7E143F0A s_waitcnt vmcnt(0) ; BF8C0F70 v_mul_f32_e32 v12, s16, v6 ; 0A180C10 s_buffer_load_dword s28, s[36:39], 0x1a4 ; C0220712 000001A4 s_buffer_load_dword s0, s[36:39], 0x1a0 ; C0220012 000001A0 v_mad_f32 v10, v14, s16, -v10 ; D1C1000A 8428210E v_floor_f32_e32 v12, v12 ; 7E183F0C s_buffer_load_dword s48, s[36:39], 0x1a8 ; C0220C12 000001A8 s_buffer_load_dword s35, s[36:39], 0x1ac ; C02208D2 000001AC v_mad_f32 v10, v14, s16, -v10 ; D1C1000A 8428210E v_mov_b32_e32 v8, 0x437f0000 ; 7E1002FF 437F0000 v_mul_f32_e32 v25, v7, v24 ; 0A323107 s_load_dwordx4 s[44:47], s[4:5], 0x1b0 ; C00A0B02 000001B0 v_mad_f32 v12, v6, s16, -v12 ; D1C1000C 84302106 s_load_dwordx8 s[52:59], s[4:5], 0x180 ; C00E0D02 00000180 v_mul_f32_e32 v11, v7, v10 ; 0A161507 v_mac_f32_e32 v25, v8, v15 ; 2C321F08 v_mad_f32 v15, v6, s16, -v12 ; D1C1000F 84302106 s_buffer_load_dword s17, s[36:39], 0x580 ; C0220452 00000580 v_mul_f32_e32 v23, v7, v13 ; 0A2E1B07 v_mul_f32_e32 v40, v7, v15 ; 0A501F07 v_mac_f32_e32 v11, v8, v14 ; 2C161D08 v_mac_f32_e32 v23, v8, v17 ; 2C2E2308 v_mac_f32_e32 v40, v8, v6 ; 2C500D08 v_add_f32_e32 v6, v11, v11 ; 020C170B s_waitcnt lgkmcnt(0) ; BF8C007F v_mov_b32_e32 v17, s0 ; 7E220200 v_mov_b32_e32 v18, s28 ; 7E24021C s_mov_b32 m0, vcc_hi ; BEFC006B v_mov_b32_e32 v9, 0 ; 7E120280 v_mad_f32 v7, s48, v6, v17 ; D1C10007 04460C30 v_mad_f32 v8, s35, v6, v18 ; D1C10008 044A0C23 image_sample_l v[19:22], v[7:10], s[52:59], s[44:47] dmask:0xf ; F0900F00 016D1307 v_interp_p1_f32 v6, v2, 2, 1, [m0] ; D4180602 v_interp_p2_f32 v6, [v6], v3, 2, 1, [m0] ; D4190603 v_add_f32_e64 v11, s17, s17 ; D101000B 00002211 v_mul_f32_e32 v6, v6, v11 ; 0A0C1706 s_brev_b32 s34, 1 ; BEA20881 v_sub_f32_e32 v11, 2.0, v13 ; 04161AF4 v_cmp_gt_f32_e64 vcc, |v11|, s34 ; D044016A 0000450B v_sub_f32_e32 v11, 1.0, v10 ; 041614F2 v_sub_f32_e32 v12, 2.0, v10 ; 041814F4 v_sub_f32_e32 v10, 1.0, v13 ; 04141AF2 v_cmp_gt_f32_e64 s[10:11], |v11|, s34 ; D044010A 0000450B v_interp_p1_f32 v11, v2, 0, 1, [m0] ; D42C0402 v_cmp_gt_f32_e64 s[28:29], |v10|, s34 ; D044011C 0000450A v_sub_f32_e32 v10, 1.0, v24 ; 041430F2 v_interp_p2_f32 v11, [v11], v3, 0, 1, [m0] ; D42D0403 v_cmp_gt_f32_e64 s[6:7], |v10|, s34 ; D0440106 0000450A v_interp_p1_f32 v10, v2, 1, 1, [m0] ; D4280502 v_mul_f32_e32 v42, -0.5, v6 ; 0A540CF1 v_mad_f32 v6, v1, -2.0, 1.0 ; D1C10006 03C9EB01 v_interp_p2_f32 v10, [v10], v3, 1, 1, [m0] ; D4290503 v_mul_f32_e32 v41, s17, v11 ; 0A521611 v_mul_f32_e32 v11, s17, v10 ; 0A161411 v_mad_f32 v10, 2.0, v0, -1.0 ; D1C1000A 03CE00F4 v_cmp_nlt_f32_e64 s[8:9], s34, v6 ; D04E0008 00020C22 v_cmp_gt_f32_e64 s[12:13], 0, v10 ; D044000C 00021480 v_cndmask_b32_e64 v14, -1.0, 1.0, s[8:9] ; D100000E 0021E4F3 v_cndmask_b32_e64 v13, 1.0, -1.0, s[12:13] ; D100000D 0031E6F2 v_mul_f32_e32 v45, v14, v41 ; 0A5A530E v_xor_b32_e32 v43, 0x80000000, v11 ; 2A5616FF 80000000 v_mul_f32_e32 v44, v13, v11 ; 0A58170D v_cmp_gt_f32_e64 s[12:13], |v12|, s34 ; D044010C 0000450C v_cndmask_b32_e64 v11, v45, v41, s[10:11] ; D100000B 002A532D v_cndmask_b32_e64 v26, v44, v11, s[12:13] ; D100001A 0032172C v_cndmask_b32_e64 v11, v42, v43, s[10:11] ; D100000B 002A572A v_cndmask_b32_e64 v27, v42, v11, s[12:13] ; D100001B 0032172A v_cndmask_b32_e64 v11, v45, v41, s[28:29] ; D100000B 0072532D v_cndmask_b32_e32 v28, v44, v11, vcc ; 0038172C v_cndmask_b32_e64 v11, v42, v43, s[28:29] ; D100000B 0072572A v_cndmask_b32_e32 v29, v42, v11, vcc ; 003A172A v_mbcnt_lo_u32_b32_e64 v11, -1, 0 ; D28C000B 000100C1 v_mbcnt_hi_u32_b32_e64 v11, -1, v11 ; D28D000B 000216C1 v_and_b32_e32 v11, 60, v11 ; 261616BC v_lshlrev_b32_e32 v46, 2, v11 ; 245C1682 ds_bpermute_b32 v11, v46, v26 ; D87E0000 0B001A2E ds_bpermute_b32 v30, v46, v26 offset:4 ; D87E0004 1E001A2E ds_bpermute_b32 v31, v46, v27 offset:4 ; D87E0004 1F001B2E ds_bpermute_b32 v32, v46, v28 offset:4 ; D87E0004 20001C2E s_buffer_load_dword s88, s[36:39], 0x294 ; C0221612 00000294 s_buffer_load_dword s49, s[36:39], 0x1b0 ; C0220C52 000001B0 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v11, v11, v30 ; 06163D0B ds_bpermute_b32 v30, v46, v27 ; D87E0000 1E001B2E s_buffer_load_dword s21, s[36:39], 0x290 ; C0220552 00000290 s_buffer_load_dword s50, s[36:39], 0x1b8 ; C0220C92 000001B8 v_mul_f32_e32 v35, s22, v27 ; 0A463616 v_add_f32_e32 v47, v25, v25 ; 025E3319 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v30, v30, v31 ; 063C3F1E ds_bpermute_b32 v31, v46, v28 ; D87E0000 1F001C2E v_mov_b32_e32 v48, 0x7effffff ; 7E6002FF 7EFFFFFF s_buffer_load_dword s51, s[36:39], 0x298 ; C0220CD2 00000298 v_mov_b32_e32 v50, 0x3f600000 ; 7E6402FF 3F600000 v_mov_b32_e32 v51, 0xff5fffff ; 7E6602FF FF5FFFFF s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v37, v31, v32 ; 064A411F ds_bpermute_b32 v31, v46, v29 ; D87E0000 1F001D2E ds_bpermute_b32 v32, v46, v29 offset:4 ; D87E0004 20001D2E v_mul_f32_e32 v36, s22, v28 ; 0A483816 s_load_dwordx8 s[60:67], s[4:5], 0x80 ; C00E0F02 00000080 s_load_dwordx4 s[80:83], s[4:5], 0xb0 ; C00A1402 000000B0 s_load_dwordx8 s[68:75], s[4:5], 0xc0 ; C00E1102 000000C0 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v38, v31, v32 ; 064C411F s_waitcnt vmcnt(0) ; BF8C0F70 v_mul_f32_e32 v32, s88, v21 ; 0A402A58 v_mul_f32_e32 v32, s49, v32 ; 0A404031 v_mul_f32_e32 v31, s21, v21 ; 0A3E2A15 v_mul_f32_e32 v31, s50, v31 ; 0A3E3E32 v_mul_f32_e32 v34, v32, v22 ; 0A442D20 v_mul_f32_e32 v30, v30, v34 ; 0A3C451E v_mul_f32_e32 v33, v31, v22 ; 0A422D1F v_mul_f32_e32 v11, v11, v33 ; 0A16430B v_mul_f32_e32 v30, v30, v30 ; 0A3C3D1E v_mac_f32_e32 v30, v11, v11 ; 2C3C170B v_mul_f32_e32 v11, s22, v26 ; 0A163416 ds_bpermute_b32 v39, v46, v11 ; D87E0000 27000B2E ds_bpermute_b32 v11, v46, v11 offset:8 ; D87E0008 0B000B2E v_rcp_f32_e32 v25, v32 ; 7E324520 v_cmp_eq_f32_e64 s[12:13], 0, v32 ; D042000C 00024080 v_cmp_neq_f32_e64 s[8:9], 0, v31 ; D04D0008 00023E80 s_load_dwordx4 s[76:79], s[4:5], 0xf0 ; C00A1302 000000F0 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v11, v39, v11 ; 06161727 ds_bpermute_b32 v39, v46, v35 ; D87E0000 2700232E ds_bpermute_b32 v35, v46, v35 offset:8 ; D87E0008 2300232E v_mul_f32_e32 v11, v33, v11 ; 0A161721 v_mul_f32_e32 v25, 0.5, v25 ; 0A3232F0 v_sub_f32_e32 v52, 2.0, v24 ; 046830F4 s_buffer_load_dword s23, s[36:39], 0x524 ; C02205D2 00000524 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v35, v39, v35 ; 06464727 v_mul_f32_e32 v33, v34, v35 ; 0A424722 v_mul_f32_e32 v33, v33, v33 ; 0A424321 v_mac_f32_e32 v33, v11, v11 ; 2C42170B v_max_f32_e32 v11, v33, v30 ; 16163D21 v_add_f32_e32 v30, v23, v23 ; 023C2F17 v_rcp_f32_e32 v23, v31 ; 7E2E451F v_rsq_f32_e64 v32, |v11| ; D1640120 0000010B v_cndmask_b32_e64 v31, v25, v48, s[12:13] ; D100001F 00326119 v_mul_f32_e32 v25, s22, v29 ; 0A323A16 v_mul_f32_e32 v23, 0.5, v23 ; 0A2E2EF0 v_cndmask_b32_e64 v23, v48, v23, s[8:9] ; D1000017 00222F30 v_cmp_neq_f32_e64 s[8:9], 0, v11 ; D04D0008 00021680 v_mov_b32_e32 v11, 0x7f7fffff ; 7E1602FF 7F7FFFFF v_cndmask_b32_e64 v32, v11, v32, s[8:9] ; D1000020 0022410B v_rcp_f32_e32 v34, v32 ; 7E444520 v_cmp_eq_f32_e64 s[8:9], 0, v32 ; D0420008 00024080 ds_bpermute_b32 v33, v46, v25 ; D87E0000 2100192E ds_bpermute_b32 v25, v46, v25 offset:8 ; D87E0008 1900192E v_cndmask_b32_e64 v32, v34, v11, s[8:9] ; D1000020 00221722 v_log_f32_e64 v34, |v32| ; D1610122 00000120 v_cmp_eq_f32_e64 s[8:9], 0, v32 ; D0420008 00024080 v_mul_f32_e32 v35, v27, v22 ; 0A462D1B s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v49, v33, v25 ; 06623321 v_mul_f32_e32 v25, v50, v34 ; 0A324532 v_cndmask_b32_e64 v25, v25, v51, s[8:9] ; D1000019 00226719 v_min_f32_e32 v25, s51, v25 ; 14323233 v_max_f32_e32 v25, 0, v25 ; 16323280 v_floor_f32_e64 v32, -v25 ; D15F0020 20000119 v_sub_f32_e64 v32, -v25, v32 ; D1020020 20024119 v_add_f32_e32 v32, v32, v25 ; 02403320 v_exp_f32_e32 v32, v32 ; 7E404120 v_mul_f32_e32 v33, v32, v23 ; 0A422F20 v_mul_f32_e32 v34, v32, v31 ; 0A443F20 v_mad_f32 v23, -v23, v32, 1.0 ; D1C10017 23CA4117 v_mad_f32 v31, -v31, v32, 1.0 ; D1C1001F 23CA411F v_mul_f32_e32 v32, v26, v22 ; 0A402D1A v_floor_f32_e32 v32, v32 ; 7E403F20 ds_bpermute_b32 v39, v46, v36 ; D87E0000 2700242E ds_bpermute_b32 v36, v46, v36 offset:8 ; D87E0008 2400242E v_mad_f32 v32, v22, v26, -v32 ; D1C10020 84823516 v_floor_f32_e32 v26, v35 ; 7E343F23 v_mad_f32 v22, v22, v27, -v26 ; D1C10016 846A3716 v_max_f32_e32 v22, v34, v22 ; 162C2D22 v_add_f32_e32 v26, s48, v7 ; 02340E30 v_max_f32_e32 v7, v33, v32 ; 160E4121 v_add_f32_e32 v27, s35, v8 ; 02361023 v_mul_f32_e32 v8, s50, v21 ; 0A102A32 v_min_f32_e32 v7, v7, v23 ; 140E2F07 v_mul_f32_e32 v21, s49, v21 ; 0A2A2A31 v_min_f32_e32 v22, v22, v31 ; 142C3F16 v_mad_f32 v23, v8, v7, v19 ; D1C10017 044E0F08 v_mad_f32 v24, v21, v22, v20 ; D1C10018 04522D15 v_mad_f32 v7, s48, v30, v17 ; D1C10007 04463C30 v_mad_f32 v8, s35, v30, v18 ; D1C10008 044A3C23 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v39, v39, v36 ; 064E4927 image_sample_l v[31:33], v[23:26], s[60:67], s[80:83] dmask:0x7 ; F0900700 028F1F17 s_nop 0 ; BF800000 image_sample_l v[34:36], v[23:26], s[68:75], s[76:79] dmask:0xb ; F0900B00 02712217 s_nop 0 ; BF800000 image_sample_l v[19:22], v[7:10], s[52:59], s[44:47] dmask:0xf ; F0900F00 016D1307 s_waitcnt vmcnt(0) ; BF8C0F70 v_mul_f32_e32 v24, s88, v21 ; 0A302A58 v_mul_f32_e32 v23, s21, v21 ; 0A2E2A15 v_mul_f32_e32 v24, s49, v24 ; 0A303031 v_mul_f32_e32 v23, s50, v23 ; 0A2E2E32 v_mul_f32_e32 v30, v24, v22 ; 0A3C2D18 v_rcp_f32_e32 v25, v23 ; 7E324517 v_cmp_neq_f32_e64 s[12:13], 0, v23 ; D04D000C 00022E80 v_mul_f32_e32 v23, v23, v22 ; 0A2E2D17 v_mul_f32_e32 v38, v30, v38 ; 0A4C4D1E v_mul_f32_e32 v30, v30, v49 ; 0A3C631E v_mul_f32_e32 v37, v23, v37 ; 0A4A4B17 v_mul_f32_e32 v23, v23, v39 ; 0A2E4F17 v_mul_f32_e32 v30, v30, v30 ; 0A3C3D1E v_mac_f32_e32 v30, v23, v23 ; 2C3C2F17 v_mul_f32_e32 v23, v38, v38 ; 0A2E4D26 v_mac_f32_e32 v23, v37, v37 ; 2C2E4B25 v_max_f32_e32 v23, v30, v23 ; 162E2F1E v_rsq_f32_e64 v30, |v23| ; D164011E 00000117 v_cmp_eq_f32_e64 s[16:17], 0, v23 ; D0420010 00022E80 v_mul_f32_e32 v25, 0.5, v25 ; 0A3232F0 v_cndmask_b32_e64 v25, v48, v25, s[12:13] ; D1000019 00323330 v_cndmask_b32_e64 v23, v30, v11, s[16:17] ; D1000017 0042171E v_rcp_f32_e32 v30, v23 ; 7E3C4517 v_cmp_eq_f32_e64 s[12:13], 0, v23 ; D042000C 00022E80 v_cmp_eq_f32_e64 s[14:15], 0, v24 ; D042000E 00023080 v_rcp_f32_e32 v24, v24 ; 7E304518 v_cndmask_b32_e64 v23, v30, v11, s[12:13] ; D1000017 0032171E v_log_f32_e64 v30, |v23| ; D161011E 00000117 v_cmp_eq_f32_e64 s[12:13], 0, v23 ; D042000C 00022E80 v_mul_f32_e32 v24, 0.5, v24 ; 0A3030F0 v_cndmask_b32_e64 v24, v24, v48, s[14:15] ; D1000018 003A6118 v_mul_f32_e32 v23, v50, v30 ; 0A2E3D32 v_cndmask_b32_e64 v23, v23, v51, s[12:13] ; D1000017 00326717 v_min_f32_e32 v23, s51, v23 ; 142E2E33 v_max_f32_e32 v30, 0, v23 ; 163C2E80 v_floor_f32_e64 v23, -v30 ; D15F0017 2000011E v_sub_f32_e64 v23, -v30, v23 ; D1020017 20022F1E v_add_f32_e32 v23, v23, v30 ; 022E3D17 v_exp_f32_e32 v23, v23 ; 7E2E4117 v_mul_f32_e32 v37, v23, v25 ; 0A4A3317 v_mul_f32_e32 v38, v23, v24 ; 0A4C3117 v_mad_f32 v25, -v25, v23, 1.0 ; D1C10019 23CA2F19 v_mad_f32 v23, -v24, v23, 1.0 ; D1C10017 23CA2F18 v_mul_f32_e32 v24, v22, v28 ; 0A303916 v_floor_f32_e32 v24, v24 ; 7E303F18 v_mad_f32 v24, v28, v22, -v24 ; D1C10018 84622D1C v_mul_f32_e32 v28, v22, v29 ; 0A383B16 v_floor_f32_e32 v28, v28 ; 7E383F1C v_mad_f32 v22, v29, v22, -v28 ; D1C10016 84722D1D v_max_f32_e32 v22, v38, v22 ; 162C2D26 v_cmp_gt_f32_e64 s[8:9], |v52|, s34 ; D0440108 00004534 v_cndmask_b32_e64 v28, v45, v41, s[6:7] ; D100001C 001A532D v_cndmask_b32_e64 v29, v42, v43, s[6:7] ; D100001D 001A572A v_cndmask_b32_e64 v49, v44, v28, s[8:9] ; D1000031 0022392C v_mul_f32_e32 v28, s50, v21 ; 0A382A32 v_mul_f32_e32 v21, s49, v21 ; 0A2A2A31 v_min_f32_e32 v22, v22, v23 ; 142C2F16 v_max_f32_e32 v24, v37, v24 ; 16303125 v_cndmask_b32_e64 v53, v42, v29, s[8:9] ; D1000035 00223B2A v_mad_f32 v29, v21, v22, v20 ; D1C1001D 04522D15 v_add_f32_e32 v20, s48, v7 ; 02280E30 v_mul_f32_e32 v7, s23, v5 ; 0A0E0A17 v_min_f32_e32 v24, v24, v25 ; 14303318 v_floor_f32_e32 v7, v7 ; 7E0E3F07 v_mad_f32 v54, s23, v5, -v7 ; D1C10036 841E0A17 v_mad_f32 v28, v28, v24, v19 ; D1C1001C 044E311C image_sample_l v[22:24], v[28:31], s[60:67], s[80:83] dmask:0x7 ; F0900700 028F161C v_mad_f32 v55, -v54, v31, v31 ; D1C10037 247E3F36 s_waitcnt vmcnt(0) ; BF8C0F70 v_mac_f32_e32 v55, v54, v22 ; 2C6E2D36 ds_bpermute_b32 v19, v46, v49 ; D87E0000 1300312E ds_bpermute_b32 v22, v46, v49 offset:4 ; D87E0004 1600312E v_mad_f32 v56, -v54, v32, v32 ; D1C10038 24824136 v_mac_f32_e32 v56, v54, v23 ; 2C702F36 ds_bpermute_b32 v23, v46, v53 offset:4 ; D87E0004 1700352E v_mad_f32 v57, -v54, v33, v33 ; D1C10039 24864336 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v19, v19, v22 ; 06262D13 ds_bpermute_b32 v22, v46, v53 ; D87E0000 1600352E v_add_f32_e32 v21, s35, v8 ; 022A1023 v_mad_f32 v7, s48, v47, v17 ; D1C10007 04465E30 v_mad_f32 v8, s35, v47, v18 ; D1C10008 044A5E23 image_sample_l v[37:39], v[28:31], s[68:75], s[76:79] dmask:0xb ; F0900B00 0271251C v_mac_f32_e32 v57, v54, v24 ; 2C723136 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v28, v22, v23 ; 06382F16 image_sample_l v[22:25], v[7:10], s[52:59], s[44:47] dmask:0xf ; F0900F00 016D1607 s_waitcnt vmcnt(0) ; BF8C0F70 v_mul_f32_e32 v30, s88, v24 ; 0A3C3058 v_mul_f32_e32 v29, s21, v24 ; 0A3A3015 v_mul_f32_e32 v31, s49, v30 ; 0A3E3C31 v_mul_f32_e32 v29, s50, v29 ; 0A3A3A32 v_mul_f32_e32 v32, v31, v25 ; 0A40331F v_mul_f32_e32 v28, v28, v32 ; 0A38411C v_mul_f32_e32 v30, v29, v25 ; 0A3C331D v_mul_f32_e32 v19, v19, v30 ; 0A263D13 v_mul_f32_e32 v28, v28, v28 ; 0A38391C v_mac_f32_e32 v28, v19, v19 ; 2C382713 v_mul_f32_e32 v19, s22, v49 ; 0A266216 ds_bpermute_b32 v47, v46, v19 ; D87E0000 2F00132E ds_bpermute_b32 v19, v46, v19 offset:8 ; D87E0008 1300132E v_mul_f32_e32 v33, s22, v53 ; 0A426A16 v_cmp_eq_f32_e64 s[12:13], 0, v29 ; D042000C 00023A80 v_add_f32_e32 v7, s48, v7 ; 020E0E30 v_add_f32_e32 v8, s35, v8 ; 02101023 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v19, v47, v19 ; 0626272F ds_bpermute_b32 v47, v46, v33 ; D87E0000 2F00212E ds_bpermute_b32 v33, v46, v33 offset:8 ; D87E0008 2100212E v_mul_f32_e32 v19, v30, v19 ; 0A26271E s_buffer_load_dword s33, s[36:39], 0x520 ; C0220852 00000520 v_cmp_ngt_f32_e64 s[0:1], |v52|, s34 ; D04B0100 00004534 s_load_dwordx4 s[40:43], s[4:5], 0x30 ; C00A0A02 00000030 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v33, v47, v33 ; 0642432F v_mul_f32_e32 v30, v32, v33 ; 0A3C4320 v_mul_f32_e32 v30, v30, v30 ; 0A3C3D1E v_mac_f32_e32 v30, v19, v19 ; 2C3C2713 v_max_f32_e32 v19, v30, v28 ; 1626391E v_cmp_neq_f32_e64 s[8:9], 0, v19 ; D04D0008 00022680 v_rsq_f32_e64 v19, |v19| ; D1640113 00000113 v_rcp_f32_e32 v28, v29 ; 7E38451D v_mul_f32_e32 v33, v25, v49 ; 0A426319 v_floor_f32_e32 v33, v33 ; 7E423F21 v_cndmask_b32_e64 v19, v11, v19, s[8:9] ; D1000013 0022270B v_cmp_eq_f32_e64 s[8:9], 0, v19 ; D0420008 00022680 v_rcp_f32_e32 v19, v19 ; 7E264513 v_mul_f32_e32 v28, 0.5, v28 ; 0A3838F0 v_cndmask_b32_e64 v28, v28, v48, s[12:13] ; D100001C 0032611C v_mad_f32 v33, v49, v25, -v33 ; D1C10021 84863331 v_cndmask_b32_e64 v19, v19, v11, s[8:9] ; D1000013 00221713 v_cmp_eq_f32_e64 s[8:9], 0, v19 ; D0420008 00022680 v_log_f32_e64 v19, |v19| ; D1610113 00000113 v_mul_f32_e32 v29, s50, v24 ; 0A3A3032 v_add_f32_e32 v0, v0, v0 ; 02000100 v_mul_f32_e32 v1, -2.0, v1 ; 0A0202F5 v_mul_f32_e32 v19, v50, v19 ; 0A262732 v_cndmask_b32_e64 v19, v19, v51, s[8:9] ; D1000013 00226713 v_min_f32_e32 v19, s51, v19 ; 14262633 v_max_f32_e32 v30, 0, v19 ; 163C2680 v_floor_f32_e64 v19, -v30 ; D15F0013 2000011E v_sub_f32_e64 v19, -v30, v19 ; D1020013 2002271E v_add_f32_e32 v19, v19, v30 ; 02263D13 v_exp_f32_e32 v19, v19 ; 7E264113 v_mul_f32_e32 v32, v19, v28 ; 0A403913 v_mad_f32 v28, -v28, v19, 1.0 ; D1C1001C 23CA271C v_max_f32_e32 v32, v32, v33 ; 16404320 v_min_f32_e32 v28, v32, v28 ; 14383920 v_mad_f32 v28, v29, v28, v22 ; D1C1001C 045A391D v_sub_f32_e32 v22, 1.0, v15 ; 042C1EF2 v_cmp_gt_f32_e64 s[8:9], |v22|, s34 ; D0440108 00004516 v_sub_f32_e32 v15, 2.0, v15 ; 041E1EF4 v_cndmask_b32_e64 v22, v45, v41, s[8:9] ; D1000016 0022532D v_cmp_gt_f32_e64 s[12:13], |v15|, s34 ; D044010C 0000450F v_cndmask_b32_e64 v32, v44, v22, s[12:13] ; D1000020 00322D2C v_cndmask_b32_e64 v22, v42, v43, s[8:9] ; D1000016 0022572A v_cndmask_b32_e64 v47, v42, v22, s[12:13] ; D100002F 00322D2A v_add_f32_e32 v22, v40, v40 ; 022C5128 v_mac_f32_e32 v17, s48, v22 ; 2C222C30 v_mac_f32_e32 v18, s35, v22 ; 2C242C23 ds_bpermute_b32 v22, v46, v32 ; D87E0000 1600202E ds_bpermute_b32 v29, v46, v32 offset:4 ; D87E0004 1D00202E v_cmp_eq_f32_e64 s[12:13], 0, v31 ; D042000C 00023E80 s_buffer_load_dword s31, s[36:39], 0xc0 ; C02207D2 000000C0 s_buffer_load_dword s32, s[36:39], 0xc4 ; C0220812 000000C4 s_buffer_load_dword s30, s[36:39], 0xc8 ; C0220792 000000C8 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v33, v22, v29 ; 06423B16 ds_bpermute_b32 v22, v46, v47 ; D87E0000 16002F2E ds_bpermute_b32 v29, v46, v47 offset:4 ; D87E0004 1D002F2E s_buffer_load_dword s20, s[36:39], 0xd0 ; C0220512 000000D0 s_buffer_load_dword s18, s[36:39], 0xd4 ; C0220492 000000D4 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v49, v22, v29 ; 06623B16 v_rcp_f32_e32 v22, v31 ; 7E2C451F v_mul_f32_e32 v22, 0.5, v22 ; 0A2C2CF0 v_cndmask_b32_e64 v22, v22, v48, s[12:13] ; D1000016 00326116 v_mul_f32_e32 v29, v19, v22 ; 0A3A2D13 v_mad_f32 v19, -v22, v19, 1.0 ; D1C10013 23CA2716 v_mul_f32_e32 v22, v25, v53 ; 0A2C6B19 v_floor_f32_e32 v22, v22 ; 7E2C3F16 v_mad_f32 v22, v53, v25, -v22 ; D1C10016 845A3335 v_max_f32_e32 v22, v29, v22 ; 162C2D1D v_min_f32_e32 v19, v22, v19 ; 14262716 v_mul_f32_e32 v22, s49, v24 ; 0A2C3031 v_mad_f32 v29, v22, v19, v23 ; D1C1001D 045E2716 v_mul_f32_e32 v19, s22, v32 ; 0A264016 ds_bpermute_b32 v23, v46, v19 ; D87E0000 1700132E ds_bpermute_b32 v19, v46, v19 offset:8 ; D87E0008 1300132E v_mul_f32_e32 v22, s22, v47 ; 0A2C5E16 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v31, v23, v19 ; 063E2717 ds_bpermute_b32 v19, v46, v22 ; D87E0000 1300162E ds_bpermute_b32 v22, v46, v22 offset:8 ; D87E0008 1600162E image_sample_l v[40:42], v[28:31], s[60:67], s[80:83] dmask:0x7 ; F0900700 028F281C s_nop 0 ; BF800000 image_sample_l v[43:45], v[28:31], s[68:75], s[76:79] dmask:0xb ; F0900B00 02712B1C v_add_f32_e32 v29, s48, v17 ; 023A2230 v_add_f32_e32 v30, s35, v18 ; 023C2423 s_waitcnt lgkmcnt(0) ; BF8C007F v_subrev_f32_e32 v46, v19, v22 ; 065C2D13 v_mov_b32_e32 v19, v9 ; 7E260309 image_sample_l v[22:25], v[17:20], s[52:59], s[44:47] dmask:0xf ; F0900F00 016D1611 s_waitcnt vmcnt(0) ; BF8C0F70 v_mul_f32_e32 v17, s21, v24 ; 0A223015 v_mul_f32_e32 v18, s88, v24 ; 0A243058 v_mul_f32_e32 v17, s50, v17 ; 0A222232 v_mul_f32_e32 v18, s49, v18 ; 0A242431 v_rcp_f32_e32 v19, v17 ; 7E264511 v_rcp_f32_e32 v28, v18 ; 7E384512 v_cmp_eq_f32_e64 s[12:13], 0, v17 ; D042000C 00022280 v_cmp_eq_f32_e64 s[14:15], 0, v18 ; D042000E 00022480 v_mul_f32_e32 v17, v17, v25 ; 0A223311 v_mul_f32_e32 v18, v18, v25 ; 0A243312 v_mul_f32_e32 v33, v17, v33 ; 0A424311 v_mul_f32_e32 v17, v17, v31 ; 0A223F11 v_mul_f32_e32 v31, v18, v49 ; 0A3E6312 v_mul_f32_e32 v18, v18, v46 ; 0A245D12 v_mul_f32_e32 v18, v18, v18 ; 0A242512 v_mac_f32_e32 v18, v17, v17 ; 2C242311 v_mul_f32_e32 v17, v31, v31 ; 0A223F1F v_mac_f32_e32 v17, v33, v33 ; 2C224321 v_max_f32_e32 v17, v18, v17 ; 16222312 v_rsq_f32_e64 v18, |v17| ; D1640112 00000111 v_mul_f32_e32 v19, 0.5, v19 ; 0A2626F0 v_cndmask_b32_e64 v19, v19, v48, s[12:13] ; D1000013 00326113 v_cmp_eq_f32_e64 s[12:13], 0, v17 ; D042000C 00022280 v_cndmask_b32_e64 v17, v18, v11, s[12:13] ; D1000011 00321712 v_rcp_f32_e32 v18, v17 ; 7E244511 v_cmp_eq_f32_e64 s[12:13], 0, v17 ; D042000C 00022280 v_mul_f32_e32 v28, 0.5, v28 ; 0A3838F0 v_cndmask_b32_e64 v28, v28, v48, s[14:15] ; D100001C 003A611C v_cndmask_b32_e64 v17, v18, v11, s[12:13] ; D1000011 00321712 v_log_f32_e64 v18, |v17| ; D1610112 00000111 v_cmp_eq_f32_e64 s[12:13], 0, v17 ; D042000C 00022280 v_mul_f32_e32 v17, v50, v18 ; 0A222532 v_cndmask_b32_e64 v17, v17, v51, s[12:13] ; D1000011 00326711 v_min_f32_e32 v17, s51, v17 ; 14222233 v_max_f32_e32 v33, 0, v17 ; 16422280 v_floor_f32_e64 v17, -v33 ; D15F0011 20000121 v_sub_f32_e64 v17, -v33, v17 ; D1020011 20022321 v_add_f32_e32 v17, v17, v33 ; 02224311 v_exp_f32_e32 v17, v17 ; 7E224111 v_mul_f32_e32 v18, v17, v19 ; 0A242711 v_mad_f32 v19, -v19, v17, 1.0 ; D1C10013 23CA2313 v_mul_f32_e32 v31, v17, v28 ; 0A3E3911 v_mad_f32 v17, -v28, v17, 1.0 ; D1C10011 23CA231C v_mul_f32_e32 v28, v25, v32 ; 0A384119 v_floor_f32_e32 v28, v28 ; 7E383F1C v_mad_f32 v28, v32, v25, -v28 ; D1C1001C 84723320 v_mul_f32_e32 v32, v25, v47 ; 0A405F19 v_floor_f32_e32 v32, v32 ; 7E403F20 v_mad_f32 v25, v47, v25, -v32 ; D1C10019 8482332F v_max_f32_e32 v18, v18, v28 ; 16243912 v_max_f32_e32 v25, v31, v25 ; 1632331F v_min_f32_e32 v18, v18, v19 ; 14242712 v_mul_f32_e32 v19, s50, v24 ; 0A263032 v_min_f32_e32 v17, v25, v17 ; 14222319 v_mul_f32_e32 v24, s49, v24 ; 0A303031 v_mad_f32 v31, v19, v18, v22 ; D1C1001F 045A2513 v_mad_f32 v32, v24, v17, v23 ; D1C10020 045E2318 image_sample_l v[17:19], v[31:34], s[60:67], s[80:83] dmask:0x7 ; F0900700 028F111F s_nop 0 ; BF800000 image_sample_l v[22:24], v[31:34], s[68:75], s[76:79] dmask:0xb ; F0900B00 0271161F v_mad_f32 v32, -v54, v44, v44 ; D1C10020 24B25936 v_mad_f32 v25, -v54, v43, v43 ; D1C10019 24AE5736 s_waitcnt vmcnt(0) ; BF8C0F70 v_mac_f32_e32 v32, v54, v23 ; 2C402F36 v_mad_f32 v23, -v54, v45, v45 ; D1C10017 24B65B36 v_mad_f32 v33, -v54, v41, v41 ; D1C10021 24A65336 v_mac_f32_e32 v25, v54, v22 ; 2C322D36 v_mac_f32_e32 v23, v54, v24 ; 2C2E3136 v_mad_f32 v24, -v54, v40, v40 ; D1C10018 24A25136 v_mad_f32 v40, -v54, v42, v42 ; D1C10028 24AA5536 v_mov_b32_e32 v28, v9 ; 7E380309 v_mov_b32_e32 v22, v9 ; 7E2C0309 v_mac_f32_e32 v24, v54, v17 ; 2C302336 v_mac_f32_e32 v33, v54, v18 ; 2C422536 image_sample_l v[17:18], v[26:29], s[52:59], s[44:47] dmask:0x3 ; F0900300 016D111A v_mac_f32_e32 v40, v54, v19 ; 2C502736 image_sample_l v[19:20], v[20:23], s[52:59], s[44:47] dmask:0x3 ; F0900300 016D1314 s_nop 0 ; BF800000 image_sample_l v[7:8], v[7:10], s[52:59], s[44:47] dmask:0x3 ; F0900300 016D0707 v_mov_b32_e32 v31, v9 ; 7E3E0309 s_waitcnt vmcnt(2) ; BF8C0F72 v_mad_f32 v27, -v54, v17, v17 ; D1C1001B 24462336 v_mad_f32 v28, -v54, v18, v18 ; D1C1001C 244A2536 v_cmp_ngt_f32_e64 s[12:13], |v12|, s34 ; D04B010C 0000450C image_sample_l v[17:18], v[29:32], s[52:59], s[44:47] dmask:0x3 ; F0900300 016D111D s_waitcnt vmcnt(1) ; BF8C0F71 v_mad_f32 v7, -v54, v7, v7 ; D1C10007 241E0F36 v_cndmask_b32_e64 v9, v14, 1.0, s[10:11] ; D1000009 0029E50E v_mov_b32_e32 v12, 0x3cf5c28f ; 7E1802FF 3CF5C28F v_mad_f32 v8, -v54, v8, v8 ; D1C10008 24221136 s_waitcnt vmcnt(0) ; BF8C0F70 v_mac_f32_e32 v7, v54, v17 ; 2C0E2336 v_cndmask_b32_e64 v17, v14, 1.0, s[28:29] ; D1000011 0071E50E v_cndmask_b32_e64 v9, v9, v12, s[12:13] ; D1000009 00321909 v_mac_f32_e32 v8, v54, v18 ; 2C102536 v_mad_f32 v18, -v54, v9, v9 ; D1C10012 24261336 v_cndmask_b32_e32 v17, v12, v17, vcc ; 0022230C v_mov_b32_e32 v9, 0x3ca3d70a ; 7E1202FF 3CA3D70A v_mac_f32_e32 v18, v54, v17 ; 2C242336 v_cndmask_b32_e64 v17, v9, 0, s[10:11] ; D1000011 00290109 v_cndmask_b32_e64 v17, v17, v13, s[12:13] ; D1000011 00321B11 v_mac_f32_e32 v27, v54, v19 ; 2C362736 v_cndmask_b32_e64 v19, v9, 0, s[28:29] ; D1000013 00710109 v_cndmask_b32_e32 v19, v13, v19, vcc ; 0026270D v_mad_f32 v17, -v54, v17, v17 ; D1C10011 24462336 v_mac_f32_e32 v17, v54, v19 ; 2C222736 v_mov_b32_e32 v19, 0x3c449ba6 ; 7E2602FF 3C449BA6 v_cndmask_b32_e64 v29, v19, 0, vcc ; D100001D 01A90113 v_cmp_ngt_f32_e64 vcc, |v15|, s34 ; D04B016A 0000450F v_cndmask_b32_e64 v15, v14, 1.0, s[6:7] ; D100000F 0019E50E v_cndmask_b32_e64 v14, v14, 1.0, s[8:9] ; D100000E 0021E50E v_cndmask_b32_e32 v14, v14, v12, vcc ; 001C190E v_cndmask_b32_e64 v12, v15, v12, s[0:1] ; D100000C 0002190F v_cndmask_b32_e64 v15, v9, 0, s[6:7] ; D100000F 00190109 v_cndmask_b32_e64 v9, v9, 0, s[8:9] ; D1000009 00210109 v_mad_f32 v12, -v54, v12, v12 ; D1C1000C 24321936 v_cndmask_b32_e32 v9, v9, v13, vcc ; 00121B09 v_cndmask_b32_e64 v13, v15, v13, s[0:1] ; D100000D 00021B0F v_mad_f32 v13, -v54, v13, v13 ; D1C1000D 24361B36 v_mac_f32_e32 v12, v54, v14 ; 2C181D36 v_cndmask_b32_e64 v14, 0, v19, s[0:1] ; D100000E 00022680 v_mac_f32_e32 v13, v54, v9 ; 2C1A1336 v_cndmask_b32_e32 v9, 0, v19, vcc ; 00122680 v_mad_f32 v14, -v54, v14, v14 ; D1C1000E 243A1D36 v_mac_f32_e32 v14, v54, v9 ; 2C1C1336 v_mul_f32_e32 v9, s33, v4 ; 0A120821 v_floor_f32_e32 v9, v9 ; 7E123F09 v_mad_f32 v15, s33, v4, -v9 ; D1C1000F 84260821 v_mad_f32 v9, -v15, v55, v55 ; D1C10009 24DE6F0F v_mac_f32_e32 v9, v15, v24 ; 2C12310F v_mac_f32_e32 v28, v54, v20 ; 2C382936 v_cndmask_b32_e64 v20, 0, v19, s[12:13] ; D1000014 00322680 v_mad_f32 v19, -v15, v56, v56 ; D1C10013 24E2710F v_mac_f32_e32 v19, v15, v33 ; 2C26430F v_mul_f32_e32 v9, 0x3e4ccccd, v9 ; 0A1212FF 3E4CCCCD s_load_dwordx8 s[8:15], s[4:5], 0x0 ; C00E0202 00000000 v_mac_f32_e32 v9, 0x3f333333, v19 ; 2C1226FF 3F333333 v_mad_f32 v19, -v15, v57, v57 ; D1C10013 24E6730F v_mac_f32_e32 v19, v15, v40 ; 2C26510F v_mac_f32_e32 v9, 0x3dcccccd, v19 ; 2C1226FF 3DCCCCCD v_mad_f32 v19, -v15, v28, v28 ; D1C10013 2472390F v_sub_f32_e32 v28, 1.0, v15 ; 04381EF2 v_mac_f32_e32 v19, v15, v8 ; 2C26110F v_mad_f32 v8, -v27, v28, v9 ; D1C10008 2426391B v_mad_f32 v27, -v7, v15, v8 ; D1C1001B 24221F07 s_waitcnt lgkmcnt(0) ; BF8C007F image_sample v[7:9], v[4:5], s[8:15], s[40:43] dmask:0x7 ; F0800700 01420704 v_mad_f32 v4, -v15, v18, v18 ; D1C10004 244A250F s_waitcnt vmcnt(0) ; BF8C0F70 v_mad_f32 v5, v28, v55, v7 ; D1C10005 041E6F1C v_mad_f32 v7, v28, v56, v8 ; D1C10007 0422711C v_mac_f32_e32 v4, v15, v12 ; 2C08190F v_mad_f32 v8, -v15, v17, v17 ; D1C10008 2446230F v_mac_f32_e32 v8, v15, v13 ; 2C101B0F v_mad_f32 v4, v0, v4, -v4 ; D1C10004 84120900 v_mac_f32_e32 v4, v8, v6 ; 2C080D08 v_mad_f32 v8, -v10, v10, 1.0 ; D1C10008 23CA150A v_mad_f32 v8, -v6, v6, v8 ; D1C10008 24220D06 v_cmp_eq_f32_e32 vcc, 0, v8 ; 7C841080 v_rsq_f32_e64 v8, |v8| ; D1640108 00000108 v_mad_f32 v20, -v54, v20, v20 ; D1C10014 24522936 v_mac_f32_e32 v20, v54, v29 ; 2C283B36 v_mac_f32_e32 v5, v15, v24 ; 2C0A310F v_cndmask_b32_e32 v8, v8, v11, vcc ; 00101708 v_cmp_eq_f32_e32 vcc, 0, v8 ; 7C841080 v_rcp_f32_e32 v8, v8 ; 7E104508 v_mad_f32 v24, -v15, v20, v20 ; D1C10018 2452290F v_mac_f32_e32 v24, v15, v14 ; 2C301D0F v_mad_f32 v21, -v54, v34, v34 ; D1C10015 248A4536 v_cndmask_b32_e32 v8, v8, v11, vcc ; 00101708 v_mac_f32_e32 v4, v24, v8 ; 2C081118 v_mad_f32 v22, -v54, v35, v35 ; D1C10016 248E4736 v_mad_f32 v18, v18, v28, -v4 ; D1C10012 84123912 v_mac_f32_e32 v18, v15, v12 ; 2C24190F v_mad_f32 v12, v17, v28, -v4 ; D1C1000C 84123911 v_mad_f32 v4, v20, v28, -v4 ; D1C10004 84123914 v_mac_f32_e32 v22, v54, v38 ; 2C2C4D36 v_mac_f32_e32 v21, v54, v37 ; 2C2A4B36 v_mad_f32 v26, -v54, v36, v36 ; D1C1001A 24924936 v_mad_f32 v17, -v15, v22, v22 ; D1C10011 245A2D0F v_mac_f32_e32 v4, v15, v14 ; 2C081D0F v_mad_f32 v14, -v15, v21, v21 ; D1C1000E 24562B0F v_mac_f32_e32 v14, v15, v25 ; 2C1C330F v_mac_f32_e32 v17, v15, v32 ; 2C22410F v_mac_f32_e32 v26, v54, v39 ; 2C344F36 v_mac_f32_e32 v9, v28, v57 ; 2C12731C v_mac_f32_e32 v14, v28, v21 ; 2C1C2B1C v_mad_f32 v20, v28, v22, v17 ; D1C10014 04462D1C v_mac_f32_e32 v12, v15, v13 ; 2C181B0F v_mad_f32 v13, -v15, v26, v26 ; D1C1000D 246A350F v_mac_f32_e32 v7, v15, v33 ; 2C0E430F v_mac_f32_e32 v9, v15, v40 ; 2C12510F v_mac_f32_e32 v13, v15, v23 ; 2C1A2F0F v_mac_f32_e32 v14, v15, v25 ; 2C1C330F v_mac_f32_e32 v20, v15, v32 ; 2C28410F v_mul_f32_e32 v15, v27, v19 ; 0A1E271B v_mul_f32_e32 v19, v18, v18 ; 0A262512 v_mac_f32_e32 v19, v12, v12 ; 2C26190C v_mac_f32_e32 v19, v4, v4 ; 2C260904 v_sqrt_f32_e32 v21, v19 ; 7E2A4F13 v_rsq_f32_e32 v19, v19 ; 7E264913 v_cmp_neq_f32_e32 vcc, 0, v21 ; 7C9A2A80 s_buffer_load_dword s6, s[36:39], 0xe0 ; C0220192 000000E0 s_buffer_load_dword s7, s[36:39], 0xe4 ; C02201D2 000000E4 v_cndmask_b32_e32 v19, v11, v19, vcc ; 0026270B v_mul_f32_e32 v12, v19, v12 ; 0A181913 v_mul_f32_e32 v18, v19, v18 ; 0A242513 v_mul_f32_e32 v19, v19, v4 ; 0A260913 v_mad_f32 v4, 2.0, v13, -1.0 ; D1C10004 03CE1AF4 v_mad_f32 v4, v14, v4, -v4 ; D1C10004 8412090E v_mad_f32 v13, 2.0, v17, -1.0 ; D1C1000D 03CE22F4 v_mad_f32 v14, -v4, v4, 1.0 ; D1C1000E 23CA0904 v_mad_f32 v14, -v13, v13, v14 ; D1C1000E 243A1B0D v_cmp_eq_f32_e32 vcc, 0, v14 ; 7C841C80 v_rsq_f32_e64 v14, |v14| ; D164010E 0000010E v_mov_b32_e32 v17, 0x7ee66665 ; 7E2202FF 7EE66665 v_mad_f32 v1, v19, v1, v19 ; D1C10001 044E0313 v_mad_f32 v1, v12, v8, -v1 ; D1C10001 8406110C v_cndmask_b32_e32 v14, v14, v11, vcc ; 001C170E v_cmp_eq_f32_e32 vcc, 0, v14 ; 7C841C80 v_rcp_f32_e32 v14, v14 ; 7E1C450E v_mad_f32 v0, v0, v12, -v12 ; D1C10000 84321900 v_mad_f32 v0, v18, v6, -v0 ; D1C10000 84020D12 s_buffer_load_dword s8, s[36:39], 0xe8 ; C0220212 000000E8 v_mul_f32_e32 v14, 0x3ee66666, v14 ; 0A1C1CFF 3EE66666 v_cndmask_b32_e32 v14, v14, v17, vcc ; 001C230E v_mul_f32_e32 v17, v4, v4 ; 0A220904 v_mac_f32_e32 v17, v13, v13 ; 2C221B0D v_mac_f32_e32 v17, v14, v14 ; 2C221D0E v_sqrt_f32_e32 v13, v17 ; 7E1A4F11 v_cmp_neq_f32_e32 vcc, 0, v13 ; 7C9A1A80 v_rsq_f32_e32 v13, v17 ; 7E1A4911 s_buffer_load_dword s9, s[36:39], 0xd8 ; C0220252 000000D8 s_load_dwordx4 s[0:3], s[4:5], 0x170 ; C00A0002 00000170 v_cndmask_b32_e32 v13, v11, v13, vcc ; 001A1B0B v_mul_f32_e32 v17, v13, v4 ; 0A22090D v_mad_f32 v4, v20, v13, -v13 ; D1C10004 84361B14 v_mul_f32_e32 v13, v13, v14 ; 0A1A1D0D v_mul_f32_e32 v14, v18, v8 ; 0A1C1112 v_mul_f32_e32 v20, v4, v1 ; 0A280304 v_mad_f32 v14, v19, v10, -v14 ; D1C1000E 843A1513 v_mul_f32_e32 v14, v4, v14 ; 0A1C1D04 v_mac_f32_e32 v20, v18, v17 ; 2C282312 v_mul_f32_e32 v21, v4, v0 ; 0A2A0104 v_mac_f32_e32 v14, v12, v17 ; 2C1C230C v_mac_f32_e32 v20, v10, v13 ; 2C281B0A v_mac_f32_e32 v21, v19, v17 ; 2C2A2313 v_mac_f32_e32 v14, v6, v13 ; 2C1C1B06 v_mul_f32_e32 v6, s31, v20 ; 0A0C281F v_add_f32_e32 v1, -0.5, v7 ; 02020EF1 s_waitcnt lgkmcnt(0) ; BF8C007F v_mul_f32_e32 v7, s6, v20 ; 0A0E2806 v_mac_f32_e32 v21, v8, v13 ; 2C2A1B08 v_mac_f32_e32 v6, s32, v14 ; 2C0C1C20 v_mac_f32_e32 v7, s7, v14 ; 2C0E1C07 v_mul_f32_e32 v8, s20, v20 ; 0A102814 v_mac_f32_e32 v6, s30, v21 ; 2C0C2A1E v_mac_f32_e32 v8, s18, v14 ; 2C101C12 v_mac_f32_e32 v7, s8, v21 ; 2C0E2A08 v_mul_f32_e32 v10, v6, v6 ; 0A140D06 v_mac_f32_e32 v8, s9, v21 ; 2C102A09 v_mac_f32_e32 v10, v7, v7 ; 2C140F07 v_mac_f32_e32 v10, v8, v8 ; 2C141108 v_rsq_f32_e32 v12, v10 ; 7E18490A v_add_f32_e32 v0, -0.5, v5 ; 02000AF1 v_add_f32_e32 v5, -0.5, v9 ; 020A12F1 v_sqrt_f32_e32 v9, v10 ; 7E124F0A v_cmp_neq_f32_e32 vcc, 0, v9 ; 7C9A1280 v_cndmask_b32_e32 v9, v11, v12, vcc ; 0012190B v_mul_f32_e32 v6, v9, v6 ; 0A0C0D09 v_mul_f32_e32 v7, v9, v7 ; 0A0E0F09 v_mul_f32_e32 v8, v9, v8 ; 0A101109 v_max3_f32 v13, |v8|, |v6|, |v7| ; D1D3070D 041E0D08 v_rcp_f32_e32 v14, v13 ; 7E1C450D v_bfrev_b32_e32 v9, -2 ; 7E1258C2 v_sub_f32_e64 v17, |v7|, v13 ; D1020111 00021B07 v_cmp_eq_f32_e32 vcc, 0, v13 ; 7C841A80 v_add_f32_e64 v4, v15, 0 clamp ; D1018004 0001010F v_sub_f32_e64 v15, |v8|, v13 ; D102010F 00021B08 v_cndmask_b32_e32 v13, v14, v11, vcc ; 001A170E v_and_b32_e32 v10, v9, v6 ; 26140D09 v_and_b32_e32 v12, v9, v7 ; 26180F09 v_cmp_ngt_f32_e32 vcc, 0, v17 ; 7C962280 v_and_b32_e32 v9, v9, v8 ; 26121109 v_mul_f32_e32 v14, v13, v6 ; 0A1C0D0D v_cndmask_b32_e32 v6, v12, v10, vcc ; 000C150C v_cmp_gt_f32_e32 vcc, 0, v15 ; 7C881E80 v_mul_f32_e32 v18, v13, v7 ; 0A240F0D v_cndmask_b32_e32 v7, v10, v6, vcc ; 000E0D0A v_cndmask_b32_e32 v9, v12, v9, vcc ; 0012130C v_subrev_f32_e32 v6, v9, v7 ; 060C0F09 v_cmp_gt_f32_e32 vcc, 0, v6 ; 7C880C80 v_cndmask_b32_e32 v6, v7, v9, vcc ; 000C1307 v_rcp_f32_e32 v10, v6 ; 7E144506 v_cndmask_b32_e32 v7, v9, v7, vcc ; 000E0F09 v_cmp_eq_f32_e32 vcc, 0, v6 ; 7C840C80 v_mul_f32_e32 v8, v13, v8 ; 0A10110D v_cndmask_b32_e32 v9, v10, v11, vcc ; 0012170A s_mov_b32 vcc_hi, m0 ; BEEB007C s_add_u32 m0, s89, 0x100 ; 807CFF59 00000100 s_buffer_load_dword s4, s[84:87], m0 ; C020012A 0000007C s_add_u32 m0, s89, 0x200 ; 807CFF59 00000200 s_waitcnt lgkmcnt(0) ; BF8C007F s_buffer_load_dword s5, s[84:87], m0 ; C020016A 0000007C v_mul_f32_e32 v7, v9, v7 ; 0A0E0F09 s_waitcnt lgkmcnt(0) ; BF8C007F s_and_b64 exec, exec, s[4:5] ; 86FE047E s_add_u32 m0, s89, 0x300 ; 807CFF59 00000300 s_buffer_load_dword s4, s[84:87], m0 ; C020012A 0000007C s_add_u32 m0, s89, 0x400 ; 807CFF59 00000400 s_waitcnt lgkmcnt(0) ; BF8C007F s_buffer_load_dword s5, s[84:87], m0 ; C020016A 0000007C s_add_u32 m0, s89, 0x500 ; 807CFF59 00000500 s_waitcnt lgkmcnt(0) ; BF8C007F s_buffer_load_dword s6, s[84:87], m0 ; C02001AA 0000007C s_add_u32 m0, s89, 0x600 ; 807CFF59 00000600 s_waitcnt lgkmcnt(0) ; BF8C007F s_buffer_load_dword s7, s[84:87], m0 ; C02001EA 0000007C s_add_u32 m0, s89, 0x700 ; 807CFF59 00000700 s_waitcnt lgkmcnt(0) ; BF8C007F s_buffer_load_dword s8, s[84:87], m0 ; C020022A 0000007C s_add_u32 m0, s89, 0x800 ; 807CFF59 00000800 s_waitcnt lgkmcnt(0) ; BF8C007F s_buffer_load_dword s9, s[84:87], m0 ; C020026A 0000007C s_add_u32 m0, s89, 0x900 ; 807CFF59 00000900 s_waitcnt lgkmcnt(0) ; BF8C007F s_buffer_load_dword s10, s[84:87], m0 ; C02002AA 0000007C s_add_u32 m0, s89, 0xa00 ; 807CFF59 00000A00 s_waitcnt lgkmcnt(0) ; BF8C007F s_buffer_load_dword s11, s[84:87], m0 ; C02002EA 0000007C s_mov_b32 m0, s19 ; BEFC0013 s_mov_b32 vcc_hi, m0 ; BEEB007C v_mov_b32_e32 v11, 0x3ecccccd ; 7E1602FF 3ECCCCCD v_mov_b32_e32 v32, v16 ; 7E400310 s_waitcnt lgkmcnt(0) ; BF8C007F image_sample v6, v[6:7], s[4:11], s[0:3] dmask:0x1 ; F0800100 00010606 v_interp_p1_f32 v12, v2, 2, 0, [m0] ; D4300202 v_interp_p2_f32 v12, [v12], v3, 2, 0, [m0] ; D4310203 s_mov_b32 m0, s89 ; BEFC0059 s_buffer_load_dword s10, s[84:87], m0 ; C02002AA 0000007C s_waitcnt vmcnt(0) ; BF8C0F70 v_mul_f32_e32 v9, v6, v18 ; 0A122506 v_mul_f32_e32 v7, v6, v14 ; 0A0E1D06 v_mul_f32_e32 v6, v6, v8 ; 0A0C1106 v_mad_f32 v8, v7, 0.5, 0.5 ; D1C10008 03C1E107 v_mad_f32 v10, v6, 0.5, 0.5 ; D1C1000A 03C1E106 v_mov_b32_e32 v2, v5 ; 7E040305 v_mad_f32 v9, v9, 0.5, 0.5 ; D1C10009 03C1E109 v_mov_b32_e32 v3, 0 ; 7E060280 v_mov_b32_e32 v7, 0 ; 7E0E0280 v_mov_b32_e32 v5, v4 ; 7E0A0304 v_mov_b32_e32 v6, v4 ; 7E0C0304 v_mov_b32_e32 v13, v12 ; 7E1A030C v_mov_b32_e32 v14, v12 ; 7E1C030C v_mov_b32_e32 v15, v12 ; 7E1E030C s_waitcnt lgkmcnt(0) ; BF8C007F s_dcache_wb ; C0840000 00000000 Shader epilog disassembly: v_cvt_pkrtz_f16_f32_e64 v0, v0, v1 ; D2960000 00020300 v_cvt_pkrtz_f16_f32_e64 v1, v2, v3 ; D2960001 00020702 v_cvt_pkrtz_f16_f32_e64 v2, v4, v5 ; D2960002 00020B04 v_cvt_pkrtz_f16_f32_e64 v3, v6, v7 ; D2960003 00020F06 exp 15, 0, 1, 0, 0, v0, v1, v0, v0 ; C400040F 00000100 v_cvt_pkrtz_f16_f32_e64 v4, v8, v9 ; D2960004 00021308 v_cvt_pkrtz_f16_f32_e64 v5, v10, v11 ; D2960005 0002170A exp 15, 1, 1, 0, 0, v2, v3, v0, v0 ; C400041F 00000302 v_cvt_pkrtz_f16_f32_e64 v6, v12, v13 ; D2960006 00021B0C v_cvt_pkrtz_f16_f32_e64 v7, v14, v15 ; D2960007 00021F0E exp 15, 2, 1, 0, 0, v4, v5, v0, v0 ; C400042F 00000504 v_cvt_pkrtz_f16_f32_e64 v8, v16, v17 ; D2960008 00022310 v_cvt_pkrtz_f16_f32_e64 v9, v18, v19 ; D2960009 00022712 exp 15, 3, 1, 0, 0, v6, v7, v0, v0 ; C400043F 00000706 v_cvt_pkrtz_f16_f32_e64 v10, v20, v21 ; D296000A 00022B14 v_cvt_pkrtz_f16_f32_e64 v11, v22, v23 ; D296000B 00022F16 exp 15, 4, 1, 0, 0, v8, v9, v0, v0 ; C400044F 00000908 v_cvt_pkrtz_f16_f32_e64 v12, v24, v25 ; D296000C 00023318 v_cvt_pkrtz_f16_f32_e64 v13, v26, v27 ; D296000D 0002371A exp 15, 5, 1, 0, 0, v10, v11, v0, v0 ; C400045F 00000B0A v_cvt_pkrtz_f16_f32_e64 v14, v28, v29 ; D296000E 00023B1C v_cvt_pkrtz_f16_f32_e64 v15, v30, v31 ; D296000F 00023F1E exp 15, 6, 1, 0, 0, v12, v13, v0, v0 ; C400046F 00000D0C exp 15, 7, 1, 1, 1, v14, v15, v0, v0 ; C4001C7F 00000F0E s_endpgm ; BF810000 *** SHADER CONFIG *** SPI_PS_INPUT_ADDR = 0xd077 SPI_PS_INPUT_ENA = 0x0002 *** SHADER STATS *** SGPRS: 96 VGPRS: 60 Spilled SGPRs: 11 Spilled VGPRs: 0 Code Size: 5028 bytes LDS: 0 blocks Scratch: 3072 bytes per wave Max Waves: 4 ********************