284:         /*0000*/                   IMAD.MOV.U32 R1, RZ, RZ, c[0x0][0x28] ;
286:         /*0010*/                   S2R R6, SR_CTAID.X ;
288:         /*0020*/                   UMOV UR8, 0xe00 ;
290:         /*0030*/                   ULDC.64 UR4, c[0x0][0x180] ;
292:         /*0040*/                   UIMAD.WIDE.U32 UR6, UR8, UR4, URZ ;
294:         /*0050*/                   UIMAD UR8, UR8, UR5, URZ ;
296:         /*0060*/                   UIADD3 UR8, UR7, UR8, URZ ;
298:         /*0070*/                   IMAD.U32 R0, RZ, RZ, UR8 ;
300:         /*0080*/                   ISETP.LT.U32.AND P0, PT, R6, UR6, PT ;
302:         /*0090*/                   ISETP.GT.U32.AND.EX P0, PT, R0, RZ, PT, P0 ;
304:         /*00a0*/              @!P0 EXIT ;
306:         /*00b0*/                   S2R R8, SR_TID.X ;
308:         /*00c0*/                   MOV R7, RZ ;
310:         /*00d0*/                   IMAD.SHL.U32 R10, R6.reuse, 0x4, RZ ;
312:         /*00e0*/                   ULDC UR9, c[0x0][0xc] ;
314:         /*00f0*/                   IMAD.MOV.U32 R4, RZ, RZ, 0x4 ;
316:         /*0100*/                   SHF.L.U64.HI R11, R6, 0x2, R7 ;
318:         /*0110*/                   UMOV UR4, URZ ;
320:         /*0120*/                   IADD3 R23, P4, R10, c[0x0][0x178], RZ ;
322:         /*0130*/                   UMOV UR5, URZ ;
324:         /*0140*/                   ULDC.64 UR10, c[0x0][0x118] ;
326:         /*0150*/                   SHF.L.U32 R5, R8, 0x2, RZ ;
328:         /*0160*/                   ISETP.NE.U32.AND P2, PT, R8, RZ, PT ;
330:         /*0170*/                   IMAD.WIDE.U32 R2, R5, 0x3800, R10 ;
332:         /*0180*/                   IMAD.WIDE.U32 R4, R5, R4, c[0x0][0x160] ;
334:         /*0190*/                   IADD3 R2, P0, R2, c[0x0][0x168], RZ ;
336:         /*01a0*/                   IADD3 R2, P1, R2, 0xa800, RZ ;
338:         /*01b0*/                   IADD3.X R3, RZ, c[0x0][0x16c], R3, P1, P0 ;
340:         /*01c0*/                   LOP3.LUT P1, R9, R8, 0x1f, RZ, 0xc0, !PT ;
342:         /*01d0*/                   SHF.R.U32.HI R8, RZ, 0x5, R8 ;
344:         /*01e0*/                   IADD3 R22, P0, R10, c[0x0][0x170], RZ ;
346:         /*01f0*/                   ISETP.GT.U32.AND P3, PT, R9.reuse, 0x1b, PT ;
348:         /*0200*/                   IMAD.SHL.U32 R9, R9, 0x4, RZ ;
350:         /*0210*/                   SHF.L.U32 R8, R8, 0x2, RZ ;
352:         /*0220*/                   IADD3.X R24, R11.reuse, c[0x0][0x174], RZ, P0, !PT ;
354:         /*0230*/                   IADD3.X R10, R11, c[0x0][0x17c], RZ, P4, !PT ;
356:         /*0240*/                   SHF.R.U32.HI R19, RZ, 0x9, R7.reuse ;
358:         /*0250*/                   BAR.SYNC.DEFER_BLOCKING 0x0 ;
360:         /*0260*/                   SHF.R.U64 R11, R6, 0x9, R7 ;
362:         /*0270*/                   IMAD.WIDE.U32 R12, R19, -0x6db6db6d, RZ ;
364:         /*0280*/                   IMAD.WIDE.U32 R14, P0, R11, 0x24924924, R12 ;
366:         /*0290*/                   IMAD.WIDE.U32 R12, R11, -0x6db6db6d, RZ ;
368:         /*02a0*/                   IADD3.X R17, RZ, RZ, RZ, P0, !PT ;
370:         /*02b0*/                   IMAD.MOV.U32 R16, RZ, RZ, R15 ;
372:         /*02c0*/                   IADD3 RZ, P0, R13, R14, RZ ;
374:         /*02d0*/                   IMAD.WIDE.U32.X R12, R19, 0x24924924, R16, P0 ;
376:         /*02e0*/                   IMAD R11, R13, -0xe00, RZ ;
378:         /*02f0*/                   IMAD.WIDE.U32 R16, R12, -0xe00, R6 ;
380:         /*0300*/                   IMAD.WIDE.U32 R14, R12, -0x3800, R2 ;
382:         /*0310*/                   IADD3 R0, R17, -R12, R11 ;
384:         /*0320*/                   IADD3 R17, P4, R6, -R16, RZ ;
386:         /*0330*/                   IMAD R13, R13, -0x3800, RZ ;
388:         /*0340*/                   IADD3 R14, P0, R14, UR4, RZ ;
390:         /*0350*/                   IADD3.X R0, R7, ~R0, RZ, P4, !PT ;
392:         /*0360*/                   IADD3 R13, R15, -R12, R13 ;
394:         /*0370*/                   LEA R12, P4, R17, R4, 0x2 ;
396:         /*0380*/                   IADD3.X R15, R13, UR5, RZ, P0, !PT ;
398:         /*0390*/                   LEA.HI.X R13, R17, R5, R0, 0x2, P4 ;
400:         /*03a0*/                   LDG.E.CONSTANT R0, [R14.64] ;
402:         /*03b0*/                   LDG.E.CONSTANT R21, [R12.64+0xc] ;
404:         /*03c0*/                   LDG.E.CONSTANT R17, [R14.64+-0x3800] ;
406:         /*03d0*/                   LDG.E.CONSTANT R20, [R12.64+0x8] ;
408:         /*03e0*/                   LDG.E.CONSTANT R16, [R14.64+-0x7000] ;
410:         /*03f0*/                   LDG.E.CONSTANT R19, [R12.64+0x4] ;
412:         /*0400*/                   LDG.E.CONSTANT R11, [R14.64+-0xa800] ;
414:         /*0410*/                   LDG.E.CONSTANT R18, [R12.64] ;
416:         /*0420*/                   FFMA R0, R0, R21, RZ ;
418:         /*0430*/                   FFMA R17, R17, R20, R0 ;
420:         /*0440*/                   FFMA R16, R16, R19, R17 ;
422:         /*0450*/                   FFMA R11, R11, R18, R16 ;
424:         /*0460*/              @!P2 IADD3 R16, P0, R22, UR4, RZ ;
426:         /*0470*/                   FADD R11, RZ, R11 ;
428:         /*0480*/              @!P2 IADD3.X R17, R24, UR5, RZ, P0, !PT ;
430:         /*0490*/                   SHFL.BFLY PT, R0, R11, 0x1, 0x1f ;
432:         /*04a0*/              @!P2 LDG.E.CONSTANT R17, [R16.64] ;
434:         /*04b0*/                   FADD R18, R11, R0 ;
436:         /*04c0*/                   SHFL.BFLY PT, R19, R18, 0x2, 0x1f ;
438:         /*04d0*/                   FADD R19, R18, R19 ;
440:         /*04e0*/                   SHFL.BFLY PT, R0, R19, 0x4, 0x1f ;
442:         /*04f0*/                   FADD R12, R19, R0 ;
444:         /*0500*/                   SHFL.BFLY PT, R13, R12, 0x8, 0x1f ;
446:         /*0510*/                   FADD R13, R12, R13 ;
448:         /*0520*/                   SHFL.BFLY PT, R0, R13, 0x10, 0x1f ;
450:         /*0530*/              @!P1 FADD R21, R13, R0 ;
452:         /*0540*/              @!P1 STS [R8], R21 ;
454:         /*0550*/                   IMAD.MOV.U32 R0, RZ, RZ, RZ ;
456:         /*0560*/                   BAR.SYNC.DEFER_BLOCKING 0x0 ;
458:         /*0570*/              @!P3 LDS R0, [R9] ;
460:         /*0580*/                   FADD R0, RZ, R0 ;
462:         /*0590*/                   SHFL.BFLY PT, R11, R0, 0x1, 0x1f ;
464:         /*05a0*/                   FADD R11, R0, R11 ;
466:         /*05b0*/                   SHFL.BFLY PT, R12, R11, 0x2, 0x1f ;
468:         /*05c0*/                   FADD R14, R11, R12 ;
470:         /*05d0*/                   SHFL.BFLY PT, R13, R14, 0x4, 0x1f ;
472:         /*05e0*/                   FADD R15, R14, R13 ;
474:         /*05f0*/                   SHFL.BFLY PT, R12, R15, 0x8, 0x1f ;
476:         /*0600*/                   FADD R18, R15, R12 ;
478:         /*0610*/                   SHFL.BFLY PT, R13, R18, 0x10, 0x1f ;
480:         /*0620*/              @!P2 IADD3 R12, P0, R23, UR4, RZ ;
482:         /*0630*/                   IADD3 R6, P4, R6, c[0x0][0xc], RZ ;
484:         /*0640*/                   IADD3.X R7, RZ, R7, RZ, P4, !PT ;
486:         /*0650*/                   ULEA UR4, UP0, UR9, UR4, 0x2 ;
488:         /*0660*/              @!P2 FADD R16, R18, R13 ;
490:         /*0670*/              @!P2 IADD3.X R13, R10, UR5, RZ, P0, !PT ;
492:         /*0680*/                   ISETP.GE.U32.AND P0, PT, R6, UR6, PT ;
494:         /*0690*/                   ISETP.GE.AND.EX P0, PT, R7, UR8, PT, P0 ;
496:         /*06a0*/              @!P2 FADD R17, R16, R17 ;
498:         /*06b0*/                   ULEA.HI.X UR5, UR9, UR5, URZ, 0x2, UP0 ;
500:         /*06c0*/              @!P2 STG.E [R12.64], R17 ;
502:         /*06d0*/               @P0 EXIT ;
504:         /*06e0*/                   BRA 0x240 ;
506:         /*06f0*/                   BRA 0x6f0;
508:         /*0700*/                   NOP;
510:         /*0710*/                   NOP;
512:         /*0720*/                   NOP;
514:         /*0730*/                   NOP;
516:         /*0740*/                   NOP;
518:         /*0750*/                   NOP;
520:         /*0760*/                   NOP;
522:         /*0770*/                   NOP;
