2000: /*0000*/ IMAD.MOV.U32 R1, RZ, RZ, c[0x0][0x28] ; 2002: /*0010*/ S2R R10, SR_CTAID.X ; 2004: /*0020*/ UMOV UR6, 0xe00 ; 2006: /*0030*/ ULDC.64 UR8, c[0x0][0x180] ; 2008: /*0040*/ UIMAD.WIDE.U32 UR4, UR6, UR8, URZ ; 2010: /*0050*/ UIMAD UR6, UR6, UR9, URZ ; 2012: /*0060*/ UIADD3 UR6, UR5, UR6, URZ ; 2014: /*0070*/ IMAD.U32 R0, RZ, RZ, UR6 ; 2016: /*0080*/ ISETP.LT.U32.AND P0, PT, R10, UR4, PT ; 2018: /*0090*/ ISETP.GT.U32.AND.EX P0, PT, R0, RZ, PT, P0 ; 2020: /*00a0*/ @!P0 EXIT ; 2022: /*00b0*/ S2R R11, SR_TID.X ; 2024: /*00c0*/ MOV R7, RZ ; 2026: /*00d0*/ IMAD.SHL.U32 R16, R10.reuse, 0x4, RZ ; 2028: /*00e0*/ MOV R4, c[0x0][0xc] ; 2030: /*00f0*/ ULDC.64 UR8, c[0x0][0x118] ; 2032: /*0100*/ SHF.L.U64.HI R17, R10, 0x2, R7 ; 2034: /*0110*/ IADD3 R14, P4, R16, c[0x0][0x178], RZ ; 2036: /*0120*/ IMAD.SHL.U32 R5, R4, 0x4, RZ ; 2038: /*0130*/ SHF.R.U32.HI R6, RZ, 0x1e, R4 ; 2040: /*0140*/ SHF.L.U32 R15, R11, 0x2, RZ ; 2042: /*0150*/ LOP3.LUT P2, R12, R11, 0x1f, RZ, 0xc0, !PT ; 2044: /*0160*/ IMAD.WIDE.U32 R2, R15, 0x3800, R16 ; 2046: /*0170*/ ISETP.GT.U32.AND P3, PT, R12.reuse, 0x1b, PT ; 2048: /*0180*/ IADD3.X R17, R17, c[0x0][0x17c], RZ, P4, !PT ; 2050: /*0190*/ IMAD.SHL.U32 R12, R12, 0x4, RZ ; 2052: /*01a0*/ IADD3 R8, P0, R2, c[0x0][0x170], RZ ; 2054: /*01b0*/ IMAD.MOV.U32 R2, RZ, RZ, 0x4 ; 2056: /*01c0*/ IADD3 R8, P1, R8, 0xa800, RZ ; 2058: /*01d0*/ IADD3.X R9, RZ, c[0x0][0x174], R3, P1, P0 ; 2060: /*01e0*/ IMAD.WIDE.U32 R2, R15, R2, c[0x0][0x168] ; 2062: /*01f0*/ ISETP.NE.U32.AND P1, PT, R11, RZ, PT ; 2064: /*0200*/ SHF.R.U32.HI R11, RZ, 0x5, R11 ; 2066: /*0210*/ LEA R13, P0, R10.reuse, c[0x0][0x160], 0x1 ; 2068: /*0220*/ SHF.L.U32 R11, R11, 0x2, RZ ; 2070: /*0230*/ LEA.HI.X R15, R10, c[0x0][0x164], R7, 0x1, P0 ; 2072: /*0240*/ SHF.R.U32.HI R25, RZ, 0x9, R7.reuse ; 2074: /*0250*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; 2076: /*0260*/ SHF.R.U64 R23, R10, 0x9, R7 ; 2078: /*0270*/ IMAD.WIDE.U32 R18, R25, -0x6db6db6d, RZ ; 2080: /*0280*/ IMAD.WIDE.U32 R20, P0, R23, 0x24924924, R18 ; 2082: /*0290*/ IMAD.WIDE.U32 R18, R23, -0x6db6db6d, RZ ; 2084: /*02a0*/ IADD3.X R23, RZ, RZ, RZ, P0, !PT ; 2086: /*02b0*/ IMAD.MOV.U32 R22, RZ, RZ, R21 ; 2088: /*02c0*/ IADD3 RZ, P0, R19, R20, RZ ; 2090: /*02d0*/ IMAD.MOV.U32 R21, RZ, RZ, R7 ; 2092: /*02e0*/ MOV R20, R10 ; 2094: /*02f0*/ IMAD.WIDE.U32.X R18, R25, 0x24924924, R22, P0 ; 2096: /*0300*/ IMAD R23, R19, -0xe00, RZ ; 2098: /*0310*/ IMAD.WIDE.U32 R20, R18, -0xe00, R20 ; 2100: /*0320*/ IMAD R19, R19, -0x3800, RZ ; 2102: /*0330*/ IADD3 R0, R21, -R18, R23 ; 2104: /*0340*/ IMAD.WIDE.U32 R22, R18, -0x3800, R8 ; 2106: /*0350*/ IADD3 R21, P0, R10, -R20, RZ ; 2108: /*0360*/ IADD3.X R0, R7, ~R0, RZ, P0, !PT ; 2110: /*0370*/ LEA R20, P0, R21, R2, 0x2 ; 2112: /*0380*/ IADD3 R19, R23, -R18, R19 ; 2114: /*0390*/ IMAD.MOV.U32 R18, RZ, RZ, R22 ; 2116: /*03a0*/ LEA.HI.X R21, R21, R3, R0, 0x2, P0 ; 2118: /*03b0*/ LDG.E.CONSTANT R0, [R18.64] ; 2120: /*03c0*/ LDG.E.CONSTANT R29, [R20.64+0xc] ; 2122: /*03d0*/ LDG.E.CONSTANT R23, [R18.64+-0x3800] ; 2124: /*03e0*/ LDG.E.CONSTANT R24, [R20.64+0x8] ; 2126: /*03f0*/ LDG.E.CONSTANT R22, [R18.64+-0x7000] ; 2128: /*0400*/ LDG.E.CONSTANT R27, [R20.64+0x4] ; 2130: /*0410*/ LDG.E.CONSTANT R16, [R18.64+-0xa800] ; 2132: /*0420*/ LDG.E.CONSTANT R25, [R20.64] ; 2134: /*0430*/ @!P1 MOV R18, R13 ; 2136: /*0440*/ @!P1 IMAD.MOV.U32 R19, RZ, RZ, R15 ; 2138: /*0450*/ FFMA R0, R0, R29, RZ ; 2140: /*0460*/ FFMA R23, R23, R24, R0 ; 2142: /*0470*/ @!P1 LDG.E.U16.CONSTANT R24, [R18.64] ; 2144: /*0480*/ FFMA R23, R22, R27, R23 ; 2146: /*0490*/ FFMA R16, R16, R25, R23 ; 2148: /*04a0*/ FADD R16, RZ, R16 ; 2150: /*04b0*/ SHFL.BFLY PT, R23, R16, 0x1, 0x1f ; 2152: /*04c0*/ FADD R23, R16, R23 ; 2154: /*04d0*/ SHFL.BFLY PT, R0, R23, 0x2, 0x1f ; 2156: /*04e0*/ FADD R20, R23, R0 ; 2158: /*04f0*/ SHFL.BFLY PT, R21, R20, 0x4, 0x1f ; 2160: /*0500*/ FADD R21, R20, R21 ; 2162: /*0510*/ SHFL.BFLY PT, R0, R21, 0x8, 0x1f ; 2164: /*0520*/ FADD R25, R21, R0 ; 2166: /*0530*/ SHFL.BFLY PT, R22, R25, 0x10, 0x1f ; 2168: /*0540*/ @!P2 FADD R22, R25, R22 ; 2170: /*0550*/ @!P2 STS [R11], R22 ; 2172: /*0560*/ MOV R0, RZ ; 2174: /*0570*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; 2176: /*0580*/ @!P3 LDS R0, [R12] ; 2178: /*0590*/ FADD R0, RZ, R0 ; 2180: /*05a0*/ SHFL.BFLY PT, R19, R0, 0x1, 0x1f ; 2182: /*05b0*/ FADD R19, R0, R19 ; 2184: /*05c0*/ SHFL.BFLY PT, R16, R19, 0x2, 0x1f ; 2186: /*05d0*/ FADD R16, R19, R16 ; 2188: /*05e0*/ SHFL.BFLY PT, R21, R16, 0x4, 0x1f ; 2190: /*05f0*/ FADD R21, R16, R21 ; 2192: /*0600*/ SHFL.BFLY PT, R18, R21, 0x8, 0x1f ; 2194: /*0610*/ IADD3 R10, P0, R10, c[0x0][0xc], RZ ; 2196: /*0620*/ FADD R18, R21, R18 ; 2198: /*0630*/ SHFL.BFLY PT, R23, R18, 0x10, 0x1f ; 2200: /*0640*/ IMAD.X R20, RZ, RZ, R7, P0 ; 2202: /*0650*/ ISETP.GE.U32.AND P0, PT, R10, UR4, PT ; 2204: /*0660*/ @!P1 HADD2.F32 R7, -RZ, R24.H0_H0 ; 2206: /*0670*/ ISETP.GE.AND.EX P0, PT, R20, UR6, PT, P0 ; 2208: /*0680*/ @!P1 MOV R16, R14 ; 2210: /*0690*/ LEA R13, P4, R4, R13, 0x1 ; 2212: /*06a0*/ IADD3 R8, P6, R5, R8, RZ ; 2214: /*06b0*/ @!P1 FADD R0, R18, R23 ; 2216: /*06c0*/ @!P1 FADD R7, R0, R7 ; 2218: /*06d0*/ IADD3 R0, P5, R5, R14, RZ ; 2220: /*06e0*/ @!P1 STG.E [R16.64], R7 ; 2222: /*06f0*/ LEA.HI.X R15, R4, R15, RZ, 0x1, P4 ; 2224: /*0700*/ IMAD.X R18, R6.reuse, 0x1, R17, P5 ; 2226: /*0710*/ IADD3.X R9, R6, R9, RZ, P6, !PT ; 2228: /*0720*/ @P0 EXIT ; 2230: /*0730*/ IMAD.MOV.U32 R7, RZ, RZ, R20 ; 2232: /*0740*/ MOV R14, R0 ; 2234: /*0750*/ IMAD.MOV.U32 R17, RZ, RZ, R18 ; 2236: /*0760*/ BRA 0x240 ; 2238: /*0770*/ BRA 0x770; 2240: /*0780*/ NOP; 2242: /*0790*/ NOP; 2244: /*07a0*/ NOP; 2246: /*07b0*/ NOP; 2248: /*07c0*/ NOP; 2250: /*07d0*/ NOP; 2252: /*07e0*/ NOP; 2254: /*07f0*/ NOP;