13190:         /*0000*/                   IMAD.MOV.U32 R1, RZ, RZ, c[0x0][0x28] ;
13192:         /*0010*/                   S2R R4, SR_CTAID.X ;
13194:         /*0020*/                   UMOV UR7, 0xe00 ;
13196:         /*0030*/                   ULDC.64 UR8, c[0x0][0x190] ;
13198:         /*0040*/                   UIMAD.WIDE.U32 UR4, UR7, UR8, URZ ;
13200:         /*0050*/                   UIMAD UR7, UR7, UR9, URZ ;
13202:         /*0060*/                   UIADD3 UR7, UR5, UR7, URZ ;
13204:         /*0070*/                   IMAD.U32 R0, RZ, RZ, UR7 ;
13206:         /*0080*/                   ISETP.LT.U32.AND P0, PT, R4, UR4, PT ;
13208:         /*0090*/                   ISETP.GT.U32.AND.EX P0, PT, R0, RZ, PT, P0 ;
13210:         /*00a0*/              @!P0 EXIT ;
13212:         /*00b0*/                   S2R R11, SR_TID.X ;
13214:         /*00c0*/                   IMAD.MOV.U32 R5, RZ, RZ, RZ ;
13216:         /*00d0*/                   SHF.L.U32 R12, R4.reuse, 0x1, RZ ;
13218:         /*00e0*/                   IMAD.MOV.U32 R17, RZ, RZ, 0x1fffffff ;
13220:         /*00f0*/                   MOV R2, c[0x0][0x180] ;
13222:         /*0100*/                   ULDC UR6, c[0x0][0xc] ;
13224:         /*0110*/                   SHF.L.U64.HI R13, R4, 0x1, R5 ;
13226:         /*0120*/                   USHF.L.U32 UR10, UR6, 0x1, URZ ;
13228:         /*0130*/                   IADD3 R2, P0, P1, R2, c[0x0][0x160], RZ ;
13230:         /*0140*/                   USHF.R.U32.HI UR6, URZ, 0x1f, UR6 ;
13232:         /*0150*/                   LOP3.LUT R0, R17.reuse, c[0x0][0x184], RZ, 0xc0, !PT ;
13234:         /*0160*/                   ULDC.64 UR8, c[0x0][0x118] ;
13236:         /*0170*/                   LOP3.LUT R17, R17, c[0x0][0x18c], RZ, 0xc0, !PT ;
13238:         /*0180*/                   IADD3 R10, P3, P5, RZ, c[0x0][0x188], R12 ;
13240:         /*0190*/                   IADD3.X R3, RZ, c[0x0][0x164], R0, P0, P1 ;
13242:         /*01a0*/                   IADD3.X R17, RZ, R17, R13, P3, P5 ;
13244:         /*01b0*/                   IADD3 R10, P5, R10, c[0x0][0x178], RZ ;
13246:         /*01c0*/                   IMAD.SHL.U32 R9, R11.reuse, 0x4, RZ ;
13248:         /*01d0*/                   ISETP.NE.U32.AND P2, PT, R11, RZ, PT ;
13250:         /*01e0*/                   SHF.R.U32.HI R0, RZ, 0x5, R11 ;
13252:         /*01f0*/                   IMAD.WIDE.U32 R6, R9, 0x1c00, R12 ;
13254:         /*0200*/                   LOP3.LUT P1, R11, R11, 0x1f, RZ, 0xc0, !PT ;
13256:         /*0210*/                   IADD3.X R17, R17, c[0x0][0x17c], RZ, P5, !PT ;
13258:         /*0220*/                   IMAD.WIDE.U32 R2, R9, 0x2, R2 ;
13260:         /*0230*/                   IADD3 R8, P4, R6, c[0x0][0x168], RZ ;
13262:         /*0240*/                   IADD3 R6, P0, R12, c[0x0][0x170], RZ ;
13264:         /*0250*/                   IADD3 R8, P6, R8, 0x5400, RZ ;
13266:         /*0260*/                   ISETP.GT.U32.AND P3, PT, R11.reuse, 0x1b, PT ;
13268:         /*0270*/                   IADD3.X R9, RZ, c[0x0][0x16c], R7, P6, P4 ;
13270:         /*0280*/                   IMAD.SHL.U32 R7, R0, 0x4, RZ ;
13272:         /*0290*/                   SHF.L.U32 R11, R11, 0x2, RZ ;
13274:         /*02a0*/                   IADD3.X R15, R13, c[0x0][0x174], RZ, P0, !PT ;
13276:         /*02b0*/                   SHF.R.U32.HI R23, RZ, 0x9, R5.reuse ;
13278:         /*02c0*/                   BAR.SYNC.DEFER_BLOCKING 0x0 ;
13280:         /*02d0*/                   SHF.R.U64 R21, R4, 0x9, R5 ;
13282:         /*02e0*/                   IMAD.WIDE.U32 R12, R23, -0x6db6db6d, RZ ;
13284:         /*02f0*/                   IMAD.WIDE.U32 R18, P0, R21, 0x24924924, R12 ;
13286:         /*0300*/                   IMAD.WIDE.U32 R12, R21, -0x6db6db6d, RZ ;
13288:         /*0310*/                   IMAD.X R21, RZ, RZ, RZ, P0 ;
13290:         /*0320*/                   IADD3 RZ, P0, R13, R18, RZ ;
13292:         /*0330*/                   IMAD.MOV.U32 R20, RZ, RZ, R19 ;
13294:         /*0340*/                   IMAD.WIDE.U32.X R12, R23, 0x24924924, R20, P0 ;
13296:         /*0350*/                   IMAD R21, R13, -0xe00, RZ ;
13298:         /*0360*/                   IMAD.WIDE.U32 R18, R12, -0xe00, R4 ;
13300:         /*0370*/                   IMAD R13, R13, -0x1c00, RZ ;
13302:         /*0380*/                   IADD3 R0, R19, -R12, R21 ;
13304:         /*0390*/                   IADD3 R21, P0, R4, -R18, RZ ;
13306:         /*03a0*/                   IMAD.WIDE.U32 R18, R12, -0x1c00, R8 ;
13308:         /*03b0*/                   IADD3.X R0, R5, ~R0, RZ, P0, !PT ;
13310:         /*03c0*/                   IMAD.IADD R27, R13, 0x1, -R12 ;
13312:         /*03d0*/                   LEA R20, P0, R21, R2, 0x1 ;
13314:         /*03e0*/                   IMAD.IADD R19, R19, 0x1, R27 ;
13316:         /*03f0*/                   LEA.HI.X R21, R21, R3, R0, 0x1, P0 ;
13318:         /*0400*/                   LDG.E.U16.CONSTANT R25, [R18.64] ;
13320:         /*0410*/                   LDG.E.U16.CONSTANT R0, [R20.64+0x4] ;
13322:         /*0420*/                   LDG.E.U16.CONSTANT R13, [R20.64+0x6] ;
13324:         /*0430*/                   LDG.E.U16.CONSTANT R14, [R20.64] ;
13326:         /*0440*/                   LDG.E.U16.CONSTANT R23, [R20.64+0x2] ;
13328:         /*0450*/                   LDG.E.U16.CONSTANT R24, [R18.64+-0x1c00] ;
13330:         /*0460*/                   LDG.E.U16.CONSTANT R22, [R18.64+-0x3800] ;
13332:         /*0470*/                   LDG.E.U16.CONSTANT R16, [R18.64+-0x5400] ;
13334:         /*0480*/                   HADD2.F32 R25, -RZ, R25.H0_H0 ;
13336:         /*0490*/                   LEA R13, R13, R0, 0x10 ;
13338:         /*04a0*/                   HADD2.F32 R0, -RZ, R13.reuse.H1_H1 ;
13340:         /*04b0*/                   HADD2.F32 R13, -RZ, R13.H0_H0 ;
13342:         /*04c0*/                   IMAD R14, R23, 0x10000, R14 ;
13344:         /*04d0*/                   FFMA R26, R0, R25, RZ ;
13346:         /*04e0*/                   HADD2.F32 R24, -RZ, R24.H0_H0 ;
13348:         /*04f0*/                   HADD2.F32 R0, -RZ, R14.H1_H1 ;
13350:         /*0500*/                   HADD2.F32 R21, -RZ, R22.H0_H0 ;
13352:         /*0510*/                   FFMA R23, R13, R24, R26 ;
13354:         /*0520*/                   HADD2.F32 R14, -RZ, R14.H0_H0 ;
13356:         /*0530*/                   HADD2.F32 R13, -RZ, R16.H0_H0 ;
13358:         /*0540*/                   FFMA R21, R0, R21, R23 ;
13360:         /*0550*/                   FFMA R13, R14, R13, R21 ;
13362:         /*0560*/                   FADD R0, RZ, R13 ;
13364:         /*0570*/                   SHFL.BFLY PT, R13, R0, 0x1, 0x1f ;
13366:         /*0580*/                   FADD R16, R0, R13 ;
13368:         /*0590*/                   SHFL.BFLY PT, R19, R16, 0x2, 0x1f ;
13370:         /*05a0*/              @!P2 IMAD.MOV.U32 R14, RZ, RZ, R6 ;
13372:         /*05b0*/              @!P2 IMAD.WIDE.U32 R12, R12, -0x1c00, R14 ;
13374:         /*05c0*/              @!P2 IADD3 R13, R27, R13, RZ ;
13376:         /*05d0*/              @!P2 LDG.E.U16.CONSTANT R23, [R12.64] ;
13378:         /*05e0*/                   FADD R19, R16, R19 ;
13380:         /*05f0*/                   SHFL.BFLY PT, R14, R19, 0x4, 0x1f ;
13382:         /*0600*/                   FADD R14, R19, R14 ;
13384:         /*0610*/                   SHFL.BFLY PT, R21, R14, 0x8, 0x1f ;
13386:         /*0620*/                   FADD R21, R14, R21 ;
13388:         /*0630*/                   SHFL.BFLY PT, R16, R21, 0x10, 0x1f ;
13390:         /*0640*/              @!P1 FADD R16, R21, R16 ;
13392:         /*0650*/              @!P1 STS [R7], R16 ;
13394:         /*0660*/                   IMAD.MOV.U32 R0, RZ, RZ, RZ ;
13396:         /*0670*/                   BAR.SYNC.DEFER_BLOCKING 0x0 ;
13398:         /*0680*/              @!P3 LDS R0, [R11] ;
13400:         /*0690*/                   FADD R0, RZ, R0 ;
13402:         /*06a0*/                   SHFL.BFLY PT, R13, R0, 0x1, 0x1f ;
13404:         /*06b0*/                   FADD R13, R0, R13 ;
13406:         /*06c0*/                   SHFL.BFLY PT, R12, R13, 0x2, 0x1f ;
13408:         /*06d0*/                   FADD R12, R13, R12 ;
13410:         /*06e0*/                   SHFL.BFLY PT, R19, R12, 0x4, 0x1f ;
13412:         /*06f0*/                   FADD R19, R12, R19 ;
13414:         /*0700*/                   SHFL.BFLY PT, R14, R19, 0x8, 0x1f ;
13416:         /*0710*/                   IADD3 R4, P0, R4, c[0x0][0xc], RZ ;
13418:         /*0720*/                   FADD R16, R19, R14 ;
13420:         /*0730*/                   SHFL.BFLY PT, R21, R16, 0x10, 0x1f ;
13422:         /*0740*/                   IMAD.X R5, RZ, RZ, R5, P0 ;
13424:         /*0750*/                   ISETP.GE.U32.AND P0, PT, R4, UR4, PT ;
13426:         /*0760*/                   ISETP.GE.AND.EX P0, PT, R5, UR7, PT, P0 ;
13428:         /*0770*/                   IADD3 R6, P5, R6, UR10, RZ ;
13430:         /*0780*/                   IADD3 R0, P4, R10, UR10, RZ ;
13432:         /*0790*/                   IADD3 R8, P6, R8, UR10, RZ ;
13434:         /*07a0*/              @!P2 FADD R14, R16, R21 ;
13436:         /*07b0*/              @!P2 F2FP.F16.F32.PACK_AB R14, RZ, R14 ;
13438:         /*07c0*/              @!P2 MOV R16, R10 ;
13440:         /*07d0*/              @!P2 HADD2 R14, R14.H0_H0, R23.H0_H0 ;
13442:         /*07e0*/                   IADD3.X R12, R17, UR6, RZ, P4, !PT ;
13444:         /*07f0*/                   IADD3.X R15, R15, UR6, RZ, P5, !PT ;
13446:         /*0800*/              @!P2 STG.E.U16 [R16.64], R14 ;
13448:         /*0810*/                   IADD3.X R9, R9, UR6, RZ, P6, !PT ;
13450:         /*0820*/               @P0 EXIT ;
13452:         /*0830*/                   IMAD.MOV.U32 R10, RZ, RZ, R0 ;
13454:         /*0840*/                   MOV R17, R12 ;
13456:         /*0850*/                   BRA 0x2b0 ;
13458:         /*0860*/                   BRA 0x860;
13460:         /*0870*/                   NOP;
13462:         /*0880*/                   NOP;
13464:         /*0890*/                   NOP;
13466:         /*08a0*/                   NOP;
13468:         /*08b0*/                   NOP;
13470:         /*08c0*/                   NOP;
13472:         /*08d0*/                   NOP;
13474:         /*08e0*/                   NOP;
13476:         /*08f0*/                   NOP;
