1723:         /*0000*/                   IMAD.MOV.U32 R1, RZ, RZ, c[0x0][0x28] ;
1725:         /*0010*/                   S2R R4, SR_CTAID.X ;
1727:         /*0020*/                   UMOV UR9, 0x4a00 ;
1729:         /*0030*/                   ULDC.64 UR6, c[0x0][0x188] ;
1731:         /*0040*/                   UIMAD.WIDE.U32 UR4, UR9, UR6, URZ ;
1733:         /*0050*/                   UIMAD UR9, UR9, UR7, URZ ;
1735:         /*0060*/                   UIADD3 UR9, UR5, UR9, URZ ;
1737:         /*0070*/                   MOV R0, UR9 ;
1739:         /*0080*/                   ISETP.LT.U32.AND P0, PT, R4, UR4, PT ;
1741:         /*0090*/                   ISETP.GT.U32.AND.EX P0, PT, R0, RZ, PT, P0 ;
1743:         /*00a0*/              @!P0 EXIT ;
1745:         /*00b0*/                   S2R R10, SR_TID.X ;
1747:         /*00c0*/                   IMAD.MOV.U32 R5, RZ, RZ, RZ ;
1749:         /*00d0*/                   ULDC.64 UR10, c[0x0][0x178] ;
1751:         /*00e0*/                   IADD3 R7, P0, RZ, c[0x0][0x180], RZ ;
1753:         /*00f0*/                   ULOP3.LUT UR7, UR10, 0xfffffffe, URZ, 0xc0, !UPT ;
1755:         /*0100*/                   IMAD.SHL.U32 R2, R4.reuse, 0x2, RZ ;
1757:         /*0110*/                   ULOP3.LUT UR8, UR11, 0x1fffffff, URZ, 0xc0, !UPT ;
1759:         /*0120*/                   SHF.L.U64.HI R3, R4, 0x1, R5 ;
1761:         /*0130*/                   ULDC UR6, c[0x0][0xc] ;
1763:         /*0140*/                   LOP3.LUT R7, R7, 0xfffffffe, RZ, 0xc0, !PT ;
1765:         /*0150*/                   ULDC.64 UR10, c[0x0][0x160] ;
1767:         /*0160*/                   IADD3.X R11, RZ, c[0x0][0x184], RZ, P0, !PT ;
1769:         /*0170*/                   UIADD3 UR7, UP0, UR7, UR10, URZ ;
1771:         /*0180*/                   IADD3 R6, P0, P1, R2, c[0x0][0x170], R7 ;
1773:         /*0190*/                   USHF.L.U32 UR12, UR6, 0x1, URZ ;
1775:         /*01a0*/                   LOP3.LUT R11, R11, 0x1fffffff, RZ, 0xc0, !PT ;
1777:         /*01b0*/                   UIADD3.X UR8, UR8, UR11, URZ, UP0, !UPT ;
1779:         /*01c0*/                   USHF.R.U32.HI UR6, URZ, 0x1f, UR6 ;
1781:         /*01d0*/                   IADD3.X R7, R3, c[0x0][0x174], R11, P0, P1 ;
1783:         /*01e0*/                   ULDC.64 UR10, c[0x0][0x118] ;
1785:         /*01f0*/                   SHF.L.U32 R13, R10.reuse, 0x2, RZ ;
1787:         /*0200*/                   LOP3.LUT P1, R11, R10.reuse, 0x1f, RZ, 0xc0, !PT ;
1789:         /*0210*/                   ISETP.NE.U32.AND P2, PT, R10, RZ, PT ;
1791:         /*0220*/                   IMAD.WIDE.U32 R8, R13, 0x9400, R2 ;
1793:         /*0230*/                   MOV R3, UR8 ;
1795:         /*0240*/                   SHF.R.U32.HI R10, RZ, 0x5, R10 ;
1797:         /*0250*/                   IMAD.U32 R2, RZ, RZ, UR7 ;
1799:         /*0260*/                   IADD3 R8, P0, R8, c[0x0][0x168], RZ ;
1801:         /*0270*/                   ISETP.GT.U32.AND P3, PT, R11, 0x1b, PT ;
1803:         /*0280*/                   IMAD.WIDE.U32 R2, R13, 0x2, R2 ;
1805:         /*0290*/                   IADD3 R8, P4, R8, 0x12800, RZ ;
1807:         /*02a0*/                   SHF.L.U32 R11, R11, 0x2, RZ ;
1809:         /*02b0*/                   IMAD.SHL.U32 R10, R10, 0x4, RZ ;
1811:         /*02c0*/                   IADD3.X R9, RZ, c[0x0][0x16c], R9, P4, P0 ;
1813:         /*02d0*/                   IMAD.WIDE.U32 R12, R5, 0xdd67c8b, RZ ;
1815:         /*02e0*/                   BAR.SYNC.DEFER_BLOCKING 0x0 ;
1817:         /*02f0*/                   IMAD.WIDE.U32 R14, P0, R4, -0x2298375a, R12 ;
1819:         /*0300*/                   IMAD.WIDE.U32 R12, R4, 0xdd67c8b, RZ ;
1821:         /*0310*/                   MOV R16, R15 ;
1823:         /*0320*/                   IMAD.X R17, RZ, RZ, RZ, P0 ;
1825:         /*0330*/                   IADD3 RZ, P0, R13, R14, RZ ;
1827:         /*0340*/                   IMAD.WIDE.U32.X R12, R5, -0x2298375a, R16, P0 ;
1829:         /*0350*/                   SHF.R.U32.HI R0, RZ, 0xe, R13.reuse ;
1831:         /*0360*/                   SHF.R.U64 R17, R12, 0xe, R13 ;
1833:         /*0370*/                   IMAD R19, R0, 0x1c00, RZ ;
1835:         /*0380*/                   IMAD.WIDE.U32 R14, R17, 0x1c00, R2 ;
1837:         /*0390*/                   IMAD.WIDE.U32 R12, R17, -0x9400, R8 ;
1839:         /*03a0*/                   IMAD R0, R0, -0x9400, RZ ;
1841:         /*03b0*/                   IMAD.IADD R15, R15, 0x1, R19 ;
1843:         /*03c0*/                   IADD3 R13, R13, -R17, R0 ;
1845:         /*03d0*/                   LDG.E.U16.CONSTANT R0, [R14.64+0x4] ;
1847:         /*03e0*/                   LDG.E.U16.CONSTANT R17, [R14.64+0x6] ;
1849:         /*03f0*/                   LDG.E.U16.CONSTANT R22, [R12.64+0x9400] ;
1851:         /*0400*/                   LDG.E.U16.CONSTANT R16, [R14.64] ;
1853:         /*0410*/                   LDG.E.U16.CONSTANT R19, [R14.64+0x2] ;
1855:         /*0420*/                   LDG.E.U16.CONSTANT R21, [R12.64] ;
1857:         /*0430*/                   LDG.E.U16.CONSTANT R20, [R12.64+-0x9400] ;
1859:         /*0440*/                   LDG.E.U16.CONSTANT R18, [R12.64+-0x12800] ;
1861:         /*0450*/                   IADD3 R4, P0, R4, c[0x0][0xc], RZ ;
1863:         /*0460*/                   IADD3 R8, P5, R8, UR12, RZ ;
1865:         /*0470*/                   IMAD.X R5, RZ, RZ, R5, P0 ;
1867:         /*0480*/                   ISETP.GE.U32.AND P0, PT, R4, UR4, PT ;
1869:         /*0490*/                   IADD3.X R9, R9, UR6, RZ, P5, !PT ;
1871:         /*04a0*/                   ISETP.GE.AND.EX P0, PT, R5, UR9, PT, P0 ;
1873:         /*04b0*/                   LEA R17, R17, R0, 0x10 ;
1875:         /*04c0*/                   HADD2.F32 R23, -RZ, R22.H0_H0 ;
1877:         /*04d0*/                   HADD2.F32 R0, -RZ, R17.reuse.H1_H1 ;
1879:         /*04e0*/                   HADD2.F32 R17, -RZ, R17.H0_H0 ;
1881:         /*04f0*/                   FFMA R24, R0, R23, RZ ;
1883:         /*0500*/                   IMAD R16, R19, 0x10000, R16 ;
1885:         /*0510*/                   HADD2.F32 R22, -RZ, R21.H0_H0 ;
1887:         /*0520*/                   HADD2.F32 R0, -RZ, R16.H1_H1 ;
1889:         /*0530*/                   HADD2.F32 R15, -RZ, R20.H0_H0 ;
1891:         /*0540*/                   FFMA R17, R17, R22, R24 ;
1893:         /*0550*/                   HADD2.F32 R16, -RZ, R16.H0_H0 ;
1895:         /*0560*/                   HADD2.F32 R13, -RZ, R18.H0_H0 ;
1897:         /*0570*/                   FFMA R15, R0, R15, R17 ;
1899:         /*0580*/                   FFMA R13, R16, R13, R15 ;
1901:         /*0590*/                   FADD R13, RZ, R13 ;
1903:         /*05a0*/                   SHFL.BFLY PT, R0, R13, 0x1, 0x1f ;
1905:         /*05b0*/                   FADD R12, R13, R0 ;
1907:         /*05c0*/                   SHFL.BFLY PT, R15, R12, 0x2, 0x1f ;
1909:         /*05d0*/                   FADD R15, R12, R15 ;
1911:         /*05e0*/                   SHFL.BFLY PT, R0, R15, 0x4, 0x1f ;
1913:         /*05f0*/                   FADD R14, R15, R0 ;
1915:         /*0600*/                   MOV R0, RZ ;
1917:         /*0610*/                   SHFL.BFLY PT, R17, R14, 0x8, 0x1f ;
1919:         /*0620*/                   FADD R16, R14, R17 ;
1921:         /*0630*/                   SHFL.BFLY PT, R17, R16, 0x10, 0x1f ;
1923:         /*0640*/              @!P1 FADD R17, R16, R17 ;
1925:         /*0650*/              @!P1 STS [R10], R17 ;
1927:         /*0660*/                   BAR.SYNC.DEFER_BLOCKING 0x0 ;
1929:         /*0670*/              @!P3 LDS R0, [R11] ;
1931:         /*0680*/                   FADD R0, RZ, R0 ;
1933:         /*0690*/                   SHFL.BFLY PT, R13, R0, 0x1, 0x1f ;
1935:         /*06a0*/                   FADD R13, R0, R13 ;
1937:         /*06b0*/                   IADD3 R0, P4, R6, UR12, RZ ;
1939:         /*06c0*/                   SHFL.BFLY PT, R12, R13, 0x2, 0x1f ;
1941:         /*06d0*/                   FADD R12, R13, R12 ;
1943:         /*06e0*/                   SHFL.BFLY PT, R15, R12, 0x4, 0x1f ;
1945:         /*06f0*/                   FADD R15, R12, R15 ;
1947:         /*0700*/                   IADD3.X R12, R7, UR6, RZ, P4, !PT ;
1949:         /*0710*/                   SHFL.BFLY PT, R14, R15, 0x8, 0x1f ;
1951:         /*0720*/                   FADD R16, R15, R14 ;
1953:         /*0730*/                   SHFL.BFLY PT, R17, R16, 0x10, 0x1f ;
1955:         /*0740*/              @!P2 FADD R14, R16, R17 ;
1957:         /*0750*/              @!P2 F2FP.F16.F32.PACK_AB R14, RZ, R14 ;
1959:         /*0760*/              @!P2 STG.E.U16 [R6.64], R14 ;
1961:         /*0770*/               @P0 EXIT ;
1963:         /*0780*/                   MOV R6, R0 ;
1965:         /*0790*/                   IMAD.MOV.U32 R7, RZ, RZ, R12 ;
1967:         /*07a0*/                   BRA 0x2d0 ;
1969:         /*07b0*/                   BRA 0x7b0;
1971:         /*07c0*/                   NOP;
1973:         /*07d0*/                   NOP;
1975:         /*07e0*/                   NOP;
1977:         /*07f0*/                   NOP;
1979:         /*0800*/                   NOP;
1981:         /*0810*/                   NOP;
1983:         /*0820*/                   NOP;
1985:         /*0830*/                   NOP;
1987:         /*0840*/                   NOP;
1989:         /*0850*/                   NOP;
1991:         /*0860*/                   NOP;
1993:         /*0870*/                   NOP;
