7:         /*0000*/                   MOV R1, c[0x0][0x28] ;
9:         /*0010*/                   S2R R5, SR_CTAID.X ;
11:         /*0020*/                   UMOV UR9, 0x25200 ;
13:         /*0030*/                   ULDC.64 UR6, c[0x0][0x180] ;
15:         /*0040*/                   UIMAD.WIDE.U32 UR4, UR9, UR6, URZ ;
17:         /*0050*/                   UIMAD UR9, UR9, UR7, URZ ;
19:         /*0060*/                   UIADD3 UR9, UR5, UR9, URZ ;
21:         /*0070*/                   IMAD.U32 R0, RZ, RZ, UR9 ;
23:         /*0080*/                   ISETP.LT.U32.AND P0, PT, R5, UR4, PT ;
25:         /*0090*/                   ISETP.GT.U32.AND.EX P0, PT, R0, RZ, PT, P0 ;
27:         /*00a0*/              @!P0 EXIT ;
29:         /*00b0*/                   S2R R12, SR_TID.X ;
31:         /*00c0*/                   ULDC.64 UR10, c[0x0][0x178] ;
33:         /*00d0*/                   MOV R6, RZ ;
35:         /*00e0*/                   ULOP3.LUT UR7, UR10, 0xfffffffe, URZ, 0xc0, !UPT ;
37:         /*00f0*/                   IMAD.SHL.U32 R10, R5.reuse, 0x2, RZ ;
39:         /*0100*/                   ULOP3.LUT UR8, UR11, 0x1fffffff, URZ, 0xc0, !UPT ;
41:         /*0110*/                   SHF.L.U64.HI R11, R5, 0x1, R6 ;
43:         /*0120*/                   ULDC UR6, c[0x0][0xc] ;
45:         /*0130*/                   ULDC.64 UR10, c[0x0][0x160] ;
47:         /*0140*/                   UIADD3 UR7, UP0, UR7, UR10, URZ ;
49:         /*0150*/                   USHF.L.U32 UR12, UR6, 0x1, URZ ;
51:         /*0160*/                   UIADD3.X UR8, UR8, UR11, URZ, UP0, !UPT ;
53:         /*0170*/                   IMAD.U32 R2, RZ, RZ, UR7 ;
55:         /*0180*/                   USHF.R.U32.HI UR6, URZ, 0x1f, UR6 ;
57:         /*0190*/                   ULDC.64 UR10, c[0x0][0x118] ;
59:         /*01a0*/                   MOV R3, UR8 ;
61:         /*01b0*/                   SHF.L.U32 R7, R12, 0x2, RZ ;
63:         /*01c0*/                   ISETP.NE.U32.AND P2, PT, R12.reuse, RZ, PT ;
65:         /*01d0*/                   LOP3.LUT P3, R0, R12, 0x1f, RZ, 0xc0, !PT ;
67:         /*01e0*/                   IMAD.WIDE.U32 R8, R7.reuse, 0x4a400, R10 ;
69:         /*01f0*/                   SHF.R.U32.HI R12, RZ, 0x5, R12 ;
71:         /*0200*/                   ISETP.GT.U32.AND P1, PT, R0, 0x1b, PT ;
73:         /*0210*/                   IMAD.WIDE.U32 R2, R7, 0x2, R2 ;
75:         /*0220*/                   IADD3 R4, P4, R8, c[0x0][0x168], RZ ;
77:         /*0230*/                   IADD3 R8, P0, R10, c[0x0][0x170], RZ ;
79:         /*0240*/                   IADD3 R4, P5, R4, 0x94800, RZ ;
81:         /*0250*/                   SHF.L.U32 R10, R0, 0x2, RZ ;
83:         /*0260*/                   IADD3.X R7, RZ, c[0x0][0x16c], R9, P5, P4 ;
85:         /*0270*/                   IMAD.SHL.U32 R9, R12, 0x4, RZ ;
87:         /*0280*/                   IADD3.X R11, R11, c[0x0][0x174], RZ, P0, !PT ;
89:         /*0290*/                   IMAD.WIDE.U32 R12, R6, -0x3806e547, RZ ;
91:         /*02a0*/                   BAR.SYNC.DEFER_BLOCKING 0x0 ;
93:         /*02b0*/                   IMAD.WIDE.U32 R14, P0, R5, -0x23570ea8, R12 ;
95:         /*02c0*/                   IMAD.WIDE.U32 R12, R5, -0x3806e547, RZ ;
97:         /*02d0*/                   MOV R16, R15 ;
99:         /*02e0*/                   IMAD.X R17, RZ, RZ, RZ, P0 ;
101:         /*02f0*/                   IADD3 RZ, P0, R13, R14, RZ ;
103:         /*0300*/                   IMAD.WIDE.U32.X R12, R6, -0x23570ea8, R16, P0 ;
105:         /*0310*/                   SHF.R.U32.HI R0, RZ, 0x11, R13.reuse ;
107:         /*0320*/                   SHF.R.U64 R17, R12, 0x11, R13 ;
109:         /*0330*/                   IMAD.MOV.U32 R12, RZ, RZ, R4 ;
111:         /*0340*/                   MOV R13, R7 ;
113:         /*0350*/                   IMAD R19, R0, 0x1c00, RZ ;
115:         /*0360*/                   IMAD.WIDE.U32 R14, R17, 0x1c00, R2 ;
117:         /*0370*/                   IMAD.WIDE.U32 R12, R17, -0x4a400, R12 ;
119:         /*0380*/                   IMAD R0, R0, -0x4a400, RZ ;
121:         /*0390*/                   IMAD.IADD R15, R15, 0x1, R19 ;
123:         /*03a0*/                   IADD3 R13, R13, -R17, R0 ;
125:         /*03b0*/                   LDG.E.U16.CONSTANT R0, [R14.64+0x4] ;
127:         /*03c0*/                   LDG.E.U16.CONSTANT R17, [R14.64+0x6] ;
129:         /*03d0*/                   LDG.E.U16.CONSTANT R22, [R12.64+0x4a400] ;
131:         /*03e0*/                   LDG.E.U16.CONSTANT R16, [R14.64] ;
133:         /*03f0*/                   LDG.E.U16.CONSTANT R19, [R14.64+0x2] ;
135:         /*0400*/                   LDG.E.U16.CONSTANT R21, [R12.64] ;
137:         /*0410*/                   LDG.E.U16.CONSTANT R20, [R12.64+-0x4a400] ;
139:         /*0420*/                   LDG.E.U16.CONSTANT R18, [R12.64+-0x94800] ;
141:         /*0430*/                   IADD3 R5, P0, R5, c[0x0][0xc], RZ ;
143:         /*0440*/                   IADD3 R4, P5, R4, UR12, RZ ;
145:         /*0450*/                   IADD3.X R6, RZ, R6, RZ, P0, !PT ;
147:         /*0460*/                   ISETP.GE.U32.AND P0, PT, R5, UR4, PT ;
149:         /*0470*/                   IADD3.X R7, R7, UR6, RZ, P5, !PT ;
151:         /*0480*/                   ISETP.GE.AND.EX P0, PT, R6, UR9, PT, P0 ;
153:         /*0490*/                   LEA R17, R17, R0, 0x10 ;
155:         /*04a0*/                   HADD2.F32 R23, -RZ, R22.H0_H0 ;
157:         /*04b0*/                   HADD2.F32 R0, -RZ, R17.reuse.H1_H1 ;
159:         /*04c0*/                   HADD2.F32 R17, -RZ, R17.H0_H0 ;
161:         /*04d0*/                   FFMA R24, R0, R23, RZ ;
163:         /*04e0*/                   IMAD R16, R19, 0x10000, R16 ;
165:         /*04f0*/                   HADD2.F32 R22, -RZ, R21.H0_H0 ;
167:         /*0500*/                   HADD2.F32 R0, -RZ, R16.H1_H1 ;
169:         /*0510*/                   HADD2.F32 R15, -RZ, R20.H0_H0 ;
171:         /*0520*/                   FFMA R17, R17, R22, R24 ;
173:         /*0530*/                   HADD2.F32 R16, -RZ, R16.H0_H0 ;
175:         /*0540*/                   HADD2.F32 R13, -RZ, R18.H0_H0 ;
177:         /*0550*/                   FFMA R15, R0, R15, R17 ;
179:         /*0560*/                   FFMA R13, R16, R13, R15 ;
181:         /*0570*/                   FADD R13, RZ, R13 ;
183:         /*0580*/                   SHFL.BFLY PT, R0, R13, 0x1, 0x1f ;
185:         /*0590*/                   FADD R12, R13, R0 ;
187:         /*05a0*/                   SHFL.BFLY PT, R15, R12, 0x2, 0x1f ;
189:         /*05b0*/                   FADD R15, R12, R15 ;
191:         /*05c0*/                   SHFL.BFLY PT, R0, R15, 0x4, 0x1f ;
193:         /*05d0*/                   FADD R14, R15, R0 ;
195:         /*05e0*/                   MOV R0, RZ ;
197:         /*05f0*/                   SHFL.BFLY PT, R17, R14, 0x8, 0x1f ;
199:         /*0600*/                   FADD R17, R14, R17 ;
201:         /*0610*/                   SHFL.BFLY PT, R16, R17, 0x10, 0x1f ;
203:         /*0620*/              @!P3 FADD R16, R17, R16 ;
205:         /*0630*/              @!P3 STS [R9], R16 ;
207:         /*0640*/                   BAR.SYNC.DEFER_BLOCKING 0x0 ;
209:         /*0650*/              @!P1 LDS R0, [R10] ;
211:         /*0660*/                   FADD R0, RZ, R0 ;
213:         /*0670*/                   SHFL.BFLY PT, R13, R0, 0x1, 0x1f ;
215:         /*0680*/                   FADD R13, R0, R13 ;
217:         /*0690*/                   IADD3 R0, P4, R8, UR12, RZ ;
219:         /*06a0*/                   SHFL.BFLY PT, R12, R13, 0x2, 0x1f ;
221:         /*06b0*/                   FADD R12, R13, R12 ;
223:         /*06c0*/              @!P2 MOV R13, R11 ;
225:         /*06d0*/                   IADD3.X R11, R11, UR6, RZ, P4, !PT ;
227:         /*06e0*/                   SHFL.BFLY PT, R15, R12, 0x4, 0x1f ;
229:         /*06f0*/                   FADD R15, R12, R15 ;
231:         /*0700*/              @!P2 IMAD.MOV.U32 R12, RZ, RZ, R8 ;
233:         /*0710*/                   SHFL.BFLY PT, R14, R15, 0x8, 0x1f ;
235:         /*0720*/                   FADD R16, R15, R14 ;
237:         /*0730*/                   SHFL.BFLY PT, R17, R16, 0x10, 0x1f ;
239:         /*0740*/              @!P2 FADD R14, R16, R17 ;
241:         /*0750*/              @!P2 F2FP.F16.F32.PACK_AB R14, RZ, R14 ;
243:         /*0760*/              @!P2 STG.E.U16 [R12.64], R14 ;
245:         /*0770*/               @P0 EXIT ;
247:         /*0780*/                   MOV R8, R0 ;
249:         /*0790*/                   BRA 0x290 ;
251:         /*07a0*/                   BRA 0x7a0;
253:         /*07b0*/                   NOP;
255:         /*07c0*/                   NOP;
257:         /*07d0*/                   NOP;
259:         /*07e0*/                   NOP;
261:         /*07f0*/                   NOP;
263:         /*0800*/                   NOP;
265:         /*0810*/                   NOP;
267:         /*0820*/                   NOP;
269:         /*0830*/                   NOP;
271:         /*0840*/                   NOP;
273:         /*0850*/                   NOP;
275:         /*0860*/                   NOP;
277:         /*0870*/                   NOP;
