code for sm_35 Function : stencil_step_task .headerflags @"EF_CUDA_SM35 EF_CUDA_PTX_SM(EF_CUDA_SM35)" /* 0x0880acb0a0a0a000 */ /*0008*/ MOV R1, c[0x0][0x44]; /* 0x64c03c00089c0006 */ /*0010*/ S2R R10, SR_CTAID.X; /* 0x86400000129c002a */ /*0018*/ IADD R44, R10, c[0x0][0x150]; /* 0x608000002a1c28b2 */ /*0020*/ IADD R0, R44, 0x1; /* 0xc0800000009cb001 */ /*0028*/ ISETP.GE.AND P0, PT, R44, R0, PT; /* 0xdb681c00001cb01e */ /*0030*/ @P0 EXIT ; /* 0x180000000000003c */ /*0038*/ MOV R11, c[0x0][0x154]; /* 0x64c03c002a9c002e */ /* 0x0888108010a01080 */ /*0048*/ IADD R41, R10, c[0x0][0x150]; /* 0x608000002a1c28a6 */ /*0050*/ MOV R12, c[0x0][0x160]; /* 0x64c03c002c1c0032 */ /*0058*/ MOV R13, c[0x0][0x164]; /* 0x64c03c002c9c0036 */ /*0060*/ IMUL R35, R11, c[0x0][0x158]; /* 0x61c018002b1c2c8e */ /*0068*/ LD.E.64 R8, [R12]; /* 0xc5800000001c3020 */ /*0070*/ SHF.L R36, RZ, 0x1, R11; /* 0xb7c02c00009ffc91 */ /*0078*/ MOV R42, c[0x0][0x148]; /* 0x64c03c00291c00aa */ /* 0x088c80108c108c10 */ /*0088*/ LD.E.64 R6, [R12+0x8]; /* 0xc5800000041c3018 */ /*0090*/ IMUL R0, R11, 0x3; /* 0xc1c01800019c2c01 */ /*0098*/ LD.E.64 R4, [R12+0x10]; /* 0xc5800000081c3010 */ /*00a0*/ IMUL R18, R11, -0x3; /* 0xc9c01bfffe9c2c49 */ /*00a8*/ SHF.L R37, RZ, 0x1, R35; /* 0xb7c08c00009ffc95 */ /*00b0*/ LD.E.64 R2, [R12+0x18]; /* 0xc58000000c1c3008 */ /*00b8*/ IMUL R19, R35, 0x3; /* 0xc1c01800019c8c4d */ /* 0x0880acb0a0acb000 */ /*00c8*/ IMUL R20, R35, -0x3; /* 0xc9c01bfffe9c8c51 */ /*00d0*/ ISETP.GE.AND P0, PT, R42, c[0x0][0x14c], PT; /* 0x5b681c00299ca81e */ /*00d8*/ @P0 BRA 0x6d8; /* 0x12000002fc00003c */ /*00e0*/ MOV R10, c[0x0][0x140]; /* 0x64c03c00281c002a */ /*00e8*/ ISETP.LT.AND P0, PT, R10, c[0x0][0x144], PT; /* 0x5b181c00289c281e */ /*00f0*/ @!P0 BRA 0x6d8; /* 0x12000002f020003c */ /*00f8*/ IMUL R40, R44, R35; /* 0xe1c01800119cb0a2 */ /* 0x088880108c10a000 */ /*0108*/ MOV R21, c[0x0][0x148]; /* 0x64c03c00291c0056 */ /*0110*/ IMAD R39, R21, c[0x0][0x154], R40; /* 0x5108a0002a9c549e */ /*0118*/ MOV R34, c[0x0][0x140]; /* 0x64c03c00281c008a */ /*0120*/ IADD R29, R39, R37; /* 0xe0800000129c9c76 */ /*0128*/ IADD R22, R39, c[0x0][0x154]; /* 0x608000002a9c9c5a */ /*0130*/ ISUB R32, R39, R37; /* 0xe0880000129c9c82 */ /*0138*/ IADD R23, R39, R36; /* 0xe0800000121c9c5e */ /* 0x0880808080108c10 */ /*0148*/ ISUB R24, R39, c[0x0][0x154]; /* 0x608800002a9c9c62 */ /*0150*/ IADD R25, R39, R0; /* 0xe0800000001c9c66 */ /*0158*/ ISUB R26, R39, R36; /* 0xe0880000121c9c6a */ /*0160*/ IADD R27, R39, R35; /* 0xe0800000119c9c6e */ /*0168*/ IADD R28, R39, R18; /* 0xe0800000091c9c72 */ /*0170*/ ISUB R30, R39, R35; /* 0xe0880000119c9c7a */ /*0178*/ IADD R33, R39, R20; /* 0xe08000000a1c9c86 */ /* 0x08a0acb0a0a0a000 */ /*0188*/ IADD R31, R39, R19; /* 0xe0800000099c9c7e */ /*0190*/ S2R R10, SR_TID.X; /* 0x86400000109c002a */ /*0198*/ LOP.AND R11, R10, 0x1f; /* 0xc20000000f9c282d */ /*01a0*/ IADD R43, R11, R34; /* 0xe0800000111c2cae */ /*01a8*/ ISETP.GE.AND P0, PT, R43, c[0x0][0x144], PT; /* 0x5b681c00289cac1e */ /*01b0*/ @P0 BRA.U 0x6a0; /* 0x120000027400023c */ /*01b8*/ @!P0 IADD R10, R39, R43; /* 0xe080000015a09c2a */ /* 0x08a0108c109c80a0 */ /*01c8*/ @!P0 SHF.L R38, RZ, 0x3, R10; /* 0xb7c0280001a3fc99 */ /*01d0*/ @!P0 IADD R10, R38, -0x8; /* 0xc88003fffc209829 */ /*01d8*/ @!P0 IADD R11, R38, 0x8; /* 0xc08000000420982d */ /*01e0*/ @!P0 BFE R12, R10, 0x11f; /* 0xc00800008fa02831 */ /*01e8*/ @!P0 IADD R54.CC, R10, c[0x0][0x170]; /* 0x608400002e2028da */ /*01f0*/ @!P0 IADD R10, R38, -0x10; /* 0xc88003fff8209829 */ /*01f8*/ @!P0 BFE R13, R11, 0x11f; /* 0xc00800008fa02c35 */ /* 0x08808080a0108c10 */ /*0208*/ @!P0 IADD.X R55, R12, c[0x0][0x174]; /* 0x608040002ea030de */ /*0210*/ @!P0 IADD R46.CC, R11, c[0x0][0x170]; /* 0x608400002e202cba */ /*0218*/ @!P0 IADD R11, R38, 0x10; /* 0xc08000000820982d */ /*0220*/ @!P0 BFE R14, R10, 0x11f; /* 0xc00800008fa02839 */ /*0228*/ @!P0 IADD.X R47, R13, c[0x0][0x174]; /* 0x608040002ea034be */ /*0230*/ @!P0 IADD R48.CC, R10, c[0x0][0x170]; /* 0x608400002e2028c2 */ /*0238*/ @!P0 IADD R10, R22, R43; /* 0xe080000015a0582a */ /* 0x08ac108080909410 */ /*0248*/ @!P0 LD.E.64 R12, [R54]; /* 0xc58000000020d830 */ /*0250*/ @!P0 BFE R15, R11, 0x11f; /* 0xc00800008fa02c3d */ /*0258*/ @!P0 LD.E.64 R16, [R46]; /* 0xc58000000020b840 */ /*0260*/ @!P0 IADD.X R49, R14, c[0x0][0x174]; /* 0x608040002ea038c6 */ /*0268*/ @!P0 IADD R52.CC, R11, c[0x0][0x170]; /* 0x608400002e202cd2 */ /*0270*/ @!P0 SHF.L R50, RZ, 0x3, R10; /* 0xb7c0280001a3fcc9 */ /*0278*/ @!P0 IADD R14, R23, R43; /* 0xe080000015a05c3a */ /* 0x08908c108c108010 */ /*0288*/ @!P0 IADD.X R53, R15, c[0x0][0x174]; /* 0x608040002ea03cd6 */ /*0290*/ @!P0 BFE R51, R50, 0x11f; /* 0xc00800008fa0c8cd */ /*0298*/ @!P0 IADD R50.CC, R50, c[0x0][0x170]; /* 0x608400002e20c8ca */ /*02a0*/ @!P0 SHF.L R45, RZ, 0x3, R14; /* 0xb7c0380001a3fcb5 */ /*02a8*/ @!P0 LD.E.64 R10, [R48]; /* 0xc58000000020c028 */ /*02b0*/ @!P0 DADD R12, R12, R16; /* 0xe380000008203032 */ /*02b8*/ @!P0 LD.E.64 R14, [R52]; /* 0xc58000000020d038 */ /* 0x089c8010b0108c10 */ /*02c8*/ @!P0 IADD.X R51, R51, c[0x0][0x174]; /* 0x608040002ea0ccce */ /*02d0*/ @!P0 BFE R17, R45, 0x11f; /* 0xc00800008fa0b445 */ /*02d8*/ @!P0 IADD R16, R24, R43; /* 0xe080000015a06042 */ /*02e0*/ @!P0 IADD R46.CC, R45, c[0x0][0x170]; /* 0x608400002e20b4ba */ /*02e8*/ @!P0 SHF.L R45, RZ, 0x3, R16; /* 0xb7c0400001a3fcb5 */ /*02f0*/ @!P0 IADD.X R47, R17, c[0x0][0x174]; /* 0x608040002ea044be */ /*02f8*/ @!P0 LD.E.64 R16, [R50]; /* 0xc58000000020c840 */ /* 0x08848010a8108080 */ /*0308*/ @!P0 IADD R54.CC, R45, c[0x0][0x170]; /* 0x608400002e20b4da */ /*0310*/ @!P0 DADD R48, R10, R14; /* 0xe3800000072028c2 */ /*0318*/ @!P0 BFE R11, R45, 0x11f; /* 0xc00800008fa0b42d */ /*0320*/ @!P0 IADD R10, R26, R43; /* 0xe080000015a0682a */ /*0328*/ @!P0 IADD.X R55, R11, c[0x0][0x174]; /* 0x608040002ea02cde */ /*0330*/ @!P0 SHF.L R45, RZ, 0x3, R10; /* 0xb7c0280001a3fcb5 */ /*0338*/ @!P0 LD.E.64 R14, [R46]; /* 0xc58000000020b838 */ /* 0x0890988010801094 */ /*0348*/ @!P0 DADD R16, R12, R16; /* 0xe380000008203042 */ /*0350*/ @!P0 IADD R13, R27, R43; /* 0xe080000015a06c36 */ /*0358*/ @!P0 LD.E.64 R10, [R54]; /* 0xc58000000020d828 */ /*0360*/ @!P0 BFE R53, R45, 0x11f; /* 0xc00800008fa0b4d5 */ /*0368*/ @!P0 IADD R52.CC, R45, c[0x0][0x170]; /* 0x608400002e20b4d2 */ /*0370*/ @!P0 IADD R12, R29, R43; /* 0xe080000015a07432 */ /*0378*/ @!P0 SHF.L R13, RZ, 0x3, R13; /* 0xb7c0340001a3fc35 */ /* 0x0894801094108c10 */ /*0388*/ @!P0 IADD.X R53, R53, c[0x0][0x174]; /* 0x608040002ea0d4d6 */ /*0390*/ @!P0 SHF.L R45, RZ, 0x3, R12; /* 0xb7c0300001a3fcb5 */ /*0398*/ @!P0 BFE R46, R13, 0x11f; /* 0xc00800008fa034b9 */ /*03a0*/ @!P0 IADD R50.CC, R13, c[0x0][0x170]; /* 0x608400002e2034ca */ /*03a8*/ @!P0 LD.E.64 R12, [R52]; /* 0xc58000000020d030 */ /*03b0*/ @!P0 DADD R16, R16, R10; /* 0xe380000005204042 */ /*03b8*/ @!P0 BFE R10, R45, 0x11f; /* 0xc00800008fa0b429 */ /* 0x08a0108c109c8010 */ /*03c8*/ @!P0 IADD.X R51, R46, c[0x0][0x174]; /* 0x608040002ea0b8ce */ /*03d0*/ @!P0 IADD R54.CC, R45, c[0x0][0x170]; /* 0x608400002e20b4da */ /*03d8*/ @!P0 IADD R45, R30, R43; /* 0xe080000015a078b6 */ /*03e0*/ @!P0 LD.E.64 R46, [R50]; /* 0xc58000000020c8b8 */ /*03e8*/ @!P0 DADD R14, R48, R14; /* 0xe38000000720c03a */ /*03f0*/ @!P0 IADD.X R55, R10, c[0x0][0x174]; /* 0x608040002ea028de */ /*03f8*/ @!P0 SHF.L R48, RZ, 0x3, R45; /* 0xb7c0b40001a3fcc1 */ /* 0x088480a080108010 */ /*0408*/ @!P0 IADD R45, R32, R43; /* 0xe080000015a080b6 */ /*0410*/ @!P0 LD.E.64 R10, [R54]; /* 0xc58000000020d828 */ /*0418*/ @!P0 BFE R49, R48, 0x11f; /* 0xc00800008fa0c0c5 */ /*0420*/ @!P0 IADD R48.CC, R48, c[0x0][0x170]; /* 0x608400002e20c0c2 */ /*0428*/ @!P0 DADD R14, R14, R12; /* 0xe38000000620383a */ /*0430*/ @!P0 DADD R12, R16, R46; /* 0xe380000017204032 */ /*0438*/ @!P0 SHF.L R46, RZ, 0x3, R45; /* 0xb7c0b40001a3fcb9 */ /* 0x0880808010b08010 */ /*0448*/ @!P0 IADD.X R49, R49, c[0x0][0x174]; /* 0x608040002ea0c4c6 */ /*0450*/ @!P0 BFE R45, R38, 0x11f; /* 0xc00800008fa098b5 */ /*0458*/ @!P0 IADD R16.CC, R38, c[0x0][0x170]; /* 0x608400002e209842 */ /*0460*/ @!P0 IADD.X R17, R45, c[0x0][0x174]; /* 0x608040002ea0b446 */ /*0468*/ @!P0 LD.E.64 R50, [R48]; /* 0xc58000000020c0c8 */ /*0470*/ @!P0 DADD R14, R14, R10; /* 0xe38000000520383a */ /*0478*/ @!P0 BFE R10, R46, 0x11f; /* 0xc00800008fa0b829 */ /* 0x0880bc109c1080b0 */ /*0488*/ @!P0 IADD R46.CC, R46, c[0x0][0x170]; /* 0x608400002e20b8ba */ /*0490*/ @!P0 IADD.X R47, R10, c[0x0][0x174]; /* 0x608040002ea028be */ /*0498*/ @!P0 LD.E.64 R10, [R16]; /* 0xc580000000204028 */ /*04a0*/ @!P0 IADD R48, R38, -0x18; /* 0xc88003fff42098c1 */ /*04a8*/ @!P0 LD.E.64 R52, [R46]; /* 0xc58000000020b8d0 */ /*04b0*/ @!P0 DADD R12, R12, R50; /* 0xe380000019203032 */ /*04b8*/ @!P0 DMUL R50, R8, R10; /* 0xe4000000052020ca */ /* 0x08b08010b01080a0 */ /*04c8*/ @!P0 IADD R46, R38, 0x18; /* 0xc08000000c2098b9 */ /*04d0*/ @!P0 DFMA R16, R6, R12, R50; /* 0xdb80c80006201842 */ /*04d8*/ @!P0 BFE R13, R48, 0x11f; /* 0xc00800008fa0c035 */ /*04e0*/ @!P0 IADD R12.CC, R48, c[0x0][0x170]; /* 0x608400002e20c032 */ /*04e8*/ @!P0 BFE R47, R46, 0x11f; /* 0xc00800008fa0b8bd */ /*04f0*/ @!P0 IADD.X R13, R13, c[0x0][0x174]; /* 0x608040002ea03436 */ /*04f8*/ @!P0 IADD R46.CC, R46, c[0x0][0x170]; /* 0x608400002e20b8ba */ /* 0x08a0a080dc109c80 */ /*0508*/ @!P0 IADD.X R47, R47, c[0x0][0x174]; /* 0x608040002ea0bcbe */ /*0510*/ @!P0 LD.E.64 R48, [R12]; /* 0xc5800000002030c0 */ /*0518*/ @!P0 LD.E.64 R50, [R46]; /* 0xc58000000020b8c8 */ /*0520*/ @!P0 DADD R14, R14, R52; /* 0xe38000001a20383a */ /*0528*/ @!P0 DADD R12, R48, R50; /* 0xe38000001920c032 */ /*0530*/ @!P0 IADD R48, R25, R43; /* 0xe080000015a064c2 */ /*0538*/ @!P0 SHF.L R46, RZ, 0x3, R48; /* 0xb7c0c00001a3fcb9 */ /* 0x08a080dc10a0b010 */ /*0548*/ @!P0 BFE R47, R46, 0x11f; /* 0xc00800008fa0b8bd */ /*0550*/ @!P0 IADD R46.CC, R46, c[0x0][0x170]; /* 0x608400002e20b8ba */ /*0558*/ @!P0 IADD.X R47, R47, c[0x0][0x174]; /* 0x608040002ea0bcbe */ /*0560*/ @!P0 LD.E.64 R48, [R46]; /* 0xc58000000020b8c0 */ /*0568*/ @!P0 DADD R10, R10, R10; /* 0xe38000000520282a */ /*0570*/ @!P0 DADD R12, R12, R48; /* 0xe380000018203032 */ /*0578*/ @!P0 IADD R48, R28, R43; /* 0xe080000015a070c2 */ /* 0x08a080dca0b010a0 */ /*0588*/ @!P0 SHF.L R46, RZ, 0x3, R48; /* 0xb7c0c00001a3fcb9 */ /*0590*/ @!P0 BFE R47, R46, 0x11f; /* 0xc00800008fa0b8bd */ /*0598*/ @!P0 IADD R46.CC, R46, c[0x0][0x170]; /* 0x608400002e20b8ba */ /*05a0*/ @!P0 IADD.X R47, R47, c[0x0][0x174]; /* 0x608040002ea0bcbe */ /*05a8*/ @!P0 LD.E.64 R48, [R46]; /* 0xc58000000020b8c0 */ /*05b0*/ @!P0 DADD R12, R12, R48; /* 0xe380000018203032 */ /*05b8*/ @!P0 IADD R48, R31, R43; /* 0xe080000015a07cc2 */ /* 0x0880a010b010a010 */ /*05c8*/ @!P0 IADD R43, R33, R43; /* 0xe080000015a084ae */ /*05d0*/ @!P0 SHF.L R46, RZ, 0x3, R48; /* 0xb7c0c00001a3fcb9 */ /*05d8*/ @!P0 BFE R47, R46, 0x11f; /* 0xc00800008fa0b8bd */ /*05e0*/ @!P0 IADD R48.CC, R46, c[0x0][0x170]; /* 0x608400002e20b8c2 */ /*05e8*/ @!P0 IADD.X R49, R47, c[0x0][0x174]; /* 0x608040002ea0bcc6 */ /*05f0*/ @!P0 SHF.L R43, RZ, 0x3, R43; /* 0xb7c0ac0001a3fcad */ /*05f8*/ @!P0 LD.E.64 R46, [R48]; /* 0xc58000000020c0b8 */ /* 0x0880909c80a080d8 */ /*0608*/ @!P0 IADD R52.CC, R43, c[0x0][0x170]; /* 0x608400002e20acd2 */ /*0610*/ @!P0 DADD R46, R12, R46; /* 0xe3800000172030ba */ /*0618*/ @!P0 BFE R12, R43, 0x11f; /* 0xc00800008fa0ac31 */ /*0620*/ @!P0 IADD.X R53, R12, c[0x0][0x174]; /* 0x608040002ea030d6 */ /*0628*/ @!P0 IADD R12.CC, R38, c[0x0][0x178]; /* 0x608400002f209832 */ /*0630*/ @!P0 LD.E.64 R48, [R52]; /* 0xc58000000020d0c0 */ /*0638*/ @!P0 IADD.X R13, R45, c[0x0][0x17c]; /* 0x608040002fa0b436 */ /* 0x08cc8c10a48090b0 */ /*0648*/ @!P0 IADD R50.CC, R38, c[0x0][0x168]; /* 0x608400002d2098ca */ /*0650*/ @!P0 IADD.X R51, R45, c[0x0][0x16c]; /* 0x608040002da0b4ce */ /*0658*/ @!P0 DADD R46, R46, R48; /* 0xe38000001820b8ba */ /*0660*/ @!P0 DFMA R48, R4, R14, R16; /* 0xdb804000072010c2 */ /*0668*/ @!P0 LD.E.64 R16, [R12]; /* 0xc580000000203040 */ /*0670*/ @!P0 DFMA R48, R2, R46, R48; /* 0xdb80c000172008c2 */ /*0678*/ @!P0 LD.E.64 R14, [R50]; /* 0xc58000000020c838 */ /* 0x08a0b8b0a000a4a4 */ /*0688*/ @!P0 DADD R10, R10, -R16; /* 0xe38100000820282a */ /*0690*/ @!P0 DFMA R10, R48, R14, R10; /* 0xdb8028000720c02a */ /*0698*/ @!P0 ST.E.64 [R12], R10; /* 0xe580000000203028 */ /*06a0*/ IADD R34, R34, 0x20; /* 0xc0800000101c8889 */ /*06a8*/ ISETP.LT.AND P0, PT, R34, c[0x0][0x144], PT; /* 0x5b181c00289c881e */ /*06b0*/ @P0 BRA 0x190; /* 0x12007ffd6c00003c */ /*06b8*/ IADD R21, R21, 0x1; /* 0xc0800000009c5455 */ /* 0x08b810b8b010b8b0 */ /*06c8*/ ISETP.EQ.AND P0, PT, R21, c[0x0][0x14c], PT; /* 0x5b281c00299c541e */ /*06d0*/ @!P0 BRA 0x110; /* 0x12007ffd1c20003c */ /*06d8*/ ISETP.NE.AND P0, PT, R44, R41, PT; /* 0xdb581c00149cb01e */ /*06e0*/ IADD R44, R44, 0x1; /* 0xc0800000009cb0b1 */ /*06e8*/ @P0 BRA 0xd0; /* 0x12007ffcf000003c */ /*06f0*/ MOV RZ, RZ; /* 0xe4c03c007f9c03fe */ /*06f8*/ EXIT ; /* 0x18000000001c003c */ /*0700*/ BRA 0x700; /* 0x12007ffffc1c003c */ /*0708*/ NOP; /* 0x85800000001c3c02 */ /*0710*/ NOP; /* 0x85800000001c3c02 */ /*0718*/ NOP; /* 0x85800000001c3c02 */ /*0720*/ NOP; /* 0x85800000001c3c02 */ /*0728*/ NOP; /* 0x85800000001c3c02 */ /*0730*/ NOP; /* 0x85800000001c3c02 */ /*0738*/ NOP; /* 0x85800000001c3c02 */ ..................................