Files
ispc/examples/stencil/1.s
2013-11-08 14:17:26 +01:00

176 lines
17 KiB
ArmAsm

code for sm_35
Function : stencil_step_task
.headerflags @"EF_CUDA_SM35 EF_CUDA_PTX_SM(EF_CUDA_SM35)"
/* 0x0880a010a0a01000 */
/*0008*/ MOV R1, c[0x0][0x44]; /* 0x64c03c00089c0006 */
/*0010*/ S2R R10, SR_CTAID.X; /* 0x86400000129c002a */
/*0018*/ MOV R12, c[0x0][0x160]; /* 0x64c03c002c1c0032 */
/*0020*/ IADD R0, R10, c[0x0][0x150]; /* 0x608000002a1c2802 */
/*0028*/ IADD R11, R0, 0x1; /* 0xc0800000009c002d */
/*0030*/ MOV R13, c[0x0][0x164]; /* 0x64c03c002c9c0036 */
/*0038*/ ISETP.GE.AND P0, PT, R0, R11, PT; /* 0xdb681c00059c001e */
/* 0x08a0a1ac118d8d8c */
/*0048*/ LD.E.64 R8, [R12]; /* 0xc5800000001c3020 */
/*0050*/ LD.E.64 R6, [R12+0x8]; /* 0xc5800000041c3018 */
/*0058*/ LD.E.64 R4, [R12+0x10]; /* 0xc5800000081c3010 */
/*0060*/ LD.E.64 R2, [R12+0x18]; /* 0xc58000000c1c3008 */
/*0068*/ @P0 EXIT ; /* 0x180000000000003c */
/*0070*/ MOV R11, c[0x0][0x158]; /* 0x64c03c002b1c002e */
/*0078*/ IMUL R41, R11, c[0x0][0x154]; /* 0x61c018002a9c2ca6 */
/* 0x08b0a000a010a010 */
/*0088*/ IADD R11, R10, c[0x0][0x150]; /* 0x608000002a1c282e */
/*0090*/ SHF.L R40, RZ, 0x1, R41; /* 0xb7c0a400009ffca1 */
/*0098*/ I2I.S32.S32 R10, -R40; /* 0xe6010000141ce82a */
/*00a0*/ IADD R49, R11, 0x1; /* 0xc0800000009c2cc5 */
/*00a8*/ SHF.L R28, RZ, 0x3, R10; /* 0xb7c02800019ffc71 */
/*00b0*/ MOV R10, c[0x0][0x148]; /* 0x64c03c00291c002a */
/*00b8*/ ISETP.GE.AND P0, PT, R10, c[0x0][0x14c], PT; /* 0x5b681c00299c281e */
/* 0x0880acb0a00010ac */
/*00c8*/ @P0 BRA 0x4f0; /* 0x120000021000003c */
/*00d0*/ MOV R29, c[0x0][0x148]; /* 0x64c03c00291c0076 */
/*00d8*/ IMUL R42, R0, R41; /* 0xe1c01800149c00aa */
/*00e0*/ MOV R10, c[0x0][0x140]; /* 0x64c03c00281c002a */
/*00e8*/ ISETP.GE.AND P0, PT, R10, c[0x0][0x144], PT; /* 0x5b681c00289c281e */
/*00f0*/ @P0 BRA 0x4d8; /* 0x12000001f000003c */
/*00f8*/ MOV R10, c[0x0][0x154]; /* 0x64c03c002a9c002a */
/* 0x0880888010a0109c */
/*0108*/ IMAD R44, R29, c[0x0][0x154], R42; /* 0x5108a8002a9c74b2 */
/*0110*/ SHF.L R11, RZ, 0x1, R10; /* 0xb7c02800009ffc2d */
/*0118*/ MOV R39, c[0x0][0x140]; /* 0x64c03c00281c009e */
/*0120*/ IMAD R34, R10, -0x2, R44; /* 0xa908b3ffff1c2889 */
/*0128*/ IADD R43, R44, R11; /* 0xe0800000059cb0ae */
/*0130*/ I2I.S32.S32 R10, -R11; /* 0xe6010000059ce82a */
/*0138*/ IMAD R36, R41, -0x2, R44; /* 0xa908b3ffff1ca491 */
/* 0x08a0001084108480 */
/*0148*/ IADD R32, R44, c[0x0][0x154]; /* 0x608000002a9cb082 */
/*0150*/ IADD R33, R44, R41; /* 0xe0800000149cb086 */
/*0158*/ IADD R35, R44, R40; /* 0xe0800000141cb08e */
/*0160*/ IMAD R38, R41, 0x3, R44; /* 0xa108b000019ca499 */
/*0168*/ SHF.L R47, RZ, 0x3, R10; /* 0xb7c02800019ffcbd */
/*0170*/ IADD R37, R43, c[0x0][0x154]; /* 0x608000002a9cac96 */
/*0178*/ S2R R10, SR_TID.X; /* 0x86400000109c002a */
/* 0x08a0b0a010908c10 */
/*0188*/ MOV32I R48, 0x8; /* 0x74000000041fc0c2 */
/*0190*/ IADD R45, R10, R39; /* 0xe0800000139c28b6 */
/*0198*/ BFE R30, R47, 0x11f; /* 0xc00800008f9cbc79 */
/*01a0*/ IADD R46, R45, R44; /* 0xe0800000161cb4ba */
/*01a8*/ IADD R14, R32, R45; /* 0xe0800000169c803a */
/*01b0*/ IMAD R10.CC, R46, R48, c[0x0][0x170]; /* 0x910cc0002e1cb82a */
/*01b8*/ IMAD.HI.X R11, R46, R48, c[0x0][0x174]; /* 0x9318c0002e9cb82e */
/* 0x0881cc118c118c10 */
/*01c8*/ IADD R27, R37, R45; /* 0xe0800000169c946e */
/*01d0*/ LD.E.64 R12, [R10+-0x8]; /* 0xc5fffffffc1c2830 */
/*01d8*/ BFE R50, R28, 0x11f; /* 0xc00800008f9c70c9 */
/*01e0*/ LD.E.64 R24, [R10+0x8]; /* 0xc5800000041c2860 */
/*01e8*/ ISETP.GE.AND P0, PT, R45, c[0x0][0x144], PT; /* 0x5b681c00289cb41e */
/*01f0*/ LD.E.64 R18, [R10+-0x18]; /* 0xc5fffffff41c2848 */
/*01f8*/ DADD R20, R24, R12; /* 0xe3800000061c6052 */
/* 0x098c10a011ac8188 */
/*0208*/ LD.E.64 R22, [R10+0x18]; /* 0xc58000000c1c2858 */
/*0210*/ IMAD R16.CC, R14, R48, c[0x0][0x170]; /* 0x910cc0002e1c3842 */
/*0218*/ LD.E.64 R12, [R10+-0x10]; /* 0xc5fffffff81c2830 */
/*0220*/ IMAD.HI.X R17, R14, R48, c[0x0][0x174]; /* 0x9318c0002e9c3846 */
/*0228*/ IADD R25, R43, R45; /* 0xe0800000169cac66 */
/*0230*/ LD.E.64 R14, [R16]; /* 0xc5800000001c4038 */
/*0238*/ DADD R22, R22, R18; /* 0xe3800000091c585a */
/* 0x0994808c848cb180 */
/*0248*/ LD.E.64 R18, [R10+0x10]; /* 0xc5800000081c2848 */
/*0250*/ IMAD R26.CC, R27, R48, c[0x0][0x170]; /* 0x910cc0002e1c6c6a */
/*0258*/ IMAD.HI.X R27, R27, R48, c[0x0][0x174]; /* 0x9318c0002e9c6c6e */
/*0260*/ IMAD R24.CC, R25, R48, c[0x0][0x170]; /* 0x910cc0002e1c6462 */
/*0268*/ DADD R14, R20, R14; /* 0xe3800000071c503a */
/*0270*/ DADD R20, R18, R12; /* 0xe3800000061c4852 */
/*0278*/ LD.E.64 R12, [R26]; /* 0xc5800000001c6830 */
/* 0x08b080118010c080 */
/*0288*/ IMAD.HI.X R25, R25, R48, c[0x0][0x174]; /* 0x9318c0002e9c6466 */
/*0290*/ IADD R16.CC, R16, R47; /* 0xe0840000179c4042 */
/*0298*/ LD.E.64 R18, [R24]; /* 0xc5800000001c6048 */
/*02a0*/ DADD R12, R22, R12; /* 0xe3800000061c5832 */
/*02a8*/ IADD.X R17, R17, R30; /* 0xe08040000f1c4446 */
/*02b0*/ IADD R31, R34, R45; /* 0xe0800000169c887e */
/*02b8*/ IADD R22.CC, R16, R47; /* 0xe0840000179c405a */
/* 0x089980818880a010 */
/*02c8*/ IADD.X R23, R17, R30; /* 0xe08040000f1c445e */
/*02d0*/ IMAD R26.CC, R31, R48, c[0x0][0x170]; /* 0x910cc0002e1c7c6a */
/*02d8*/ DADD R20, R20, R18; /* 0xe3800000091c5052 */
/*02e0*/ LD.E.64 R18, [R16]; /* 0xc5800000001c4048 */
/*02e8*/ IMAD.HI.X R27, R31, R48, c[0x0][0x174]; /* 0x9318c0002e9c7c6e */
/*02f0*/ LD.E.64 R24, [R22]; /* 0xc5800000001c5860 */
/*02f8*/ IADD R51, R33, R45; /* 0xe0800000169c84ce */
/* 0x088880ac818c11b8 */
/*0308*/ LD.E.64 R30, [R26]; /* 0xc5800000001c6878 */
/*0310*/ LD.E.64 R26, [R10]; /* 0xc5800000001c2868 */
/*0318*/ DADD R14, R14, R18; /* 0xe3800000091c383a */
/*0320*/ IMAD R18.CC, R51, R48, c[0x0][0x170]; /* 0x910cc0002e1ccc4a */
/*0328*/ IADD R17, R35, R45; /* 0xe0800000169c8c46 */
/*0330*/ IMAD.HI.X R19, R51, R48, c[0x0][0x174]; /* 0x9318c0002e9ccc4e */
/*0338*/ DADD R22, R20, R30; /* 0xe38000000f1c505a */
/* 0x098c10a0999c1090 */
/*0348*/ IMAD R16.CC, R17, R48, c[0x0][0x170]; /* 0x910cc0002e1c4442 */
/*0350*/ LD.E.64 R20, [R18]; /* 0xc5800000001c4850 */
/*0358*/ DADD R12, R12, R24; /* 0xe38000000c1c3032 */
/*0360*/ IMAD.HI.X R17, R17, R48, c[0x0][0x174]; /* 0x9318c0002e9c4446 */
/*0368*/ IADD R18.CC, R18, R28; /* 0xe08400000e1c484a */
/*0370*/ LD.E.64 R24, [R16]; /* 0xc5800000001c4060 */
/*0378*/ DADD R20, R14, R20; /* 0xe38000000a1c3852 */
/* 0x088080b4a18010cc */
/*0388*/ IADD.X R19, R19, R50; /* 0xe0804000191c4c4e */
/*0390*/ LD.E.64 R14, [R18]; /* 0xc5800000001c4838 */
/*0398*/ DADD R22, R22, R24; /* 0xe38000000c1c585a */
/*03a0*/ IADD R25, R36, R45; /* 0xe0800000169c9066 */
/*03a8*/ IMAD R16.CC, R25, R48, c[0x0][0x170]; /* 0x910cc0002e1c6442 */
/*03b0*/ DADD R20, R20, R14; /* 0xe3800000071c5052 */
/*03b8*/ IADD R15, R38, R45; /* 0xe0800000169c983e */
/* 0x09a010b081ac809c */
/*03c8*/ IMAD.HI.X R17, R25, R48, c[0x0][0x174]; /* 0x9318c0002e9c6446 */
/*03d0*/ IMAD R14.CC, R15, R48, c[0x0][0x170]; /* 0x910cc0002e1c3c3a */
/*03d8*/ LD.E.64 R24, [R16]; /* 0xc5800000001c4060 */
/*03e0*/ IMAD.HI.X R15, R15, R48, c[0x0][0x174]; /* 0x9318c0002e9c3c3e */
/*03e8*/ IADD R18.CC, R18, R28; /* 0xe08400000e1c484a */
/*03f0*/ LD.E.64 R30, [R14]; /* 0xc5800000001c3878 */
/*03f8*/ IADD.X R19, R19, R50; /* 0xe0804000191c4c4e */
/* 0x08a480a480b58010 */
/*0408*/ LD.E.64 R50, [R18]; /* 0xc5800000001c48c8 */
/*0410*/ DMUL R20, R6, R20; /* 0xe40000000a1c1852 */
/*0418*/ DADD R22, R22, R24; /* 0xe38000000c1c585a */
/*0420*/ DADD R12, R12, R30; /* 0xe38000000f1c3032 */
/*0428*/ DFMA R24, R8, R26, R20; /* 0xdb8050000d1c2062 */
/*0430*/ DFMA R16, R4, R22, R24; /* 0xdb8060000b1c1042 */
/*0438*/ DADD R12, R12, R50; /* 0xe3800000191c3032 */
/* 0x08908cb0a010ac80 */
/*0448*/ DFMA R10, R2, R12, R16; /* 0xdb804000061c082a */
/*0450*/ @P0 BRA.U 0x4b8; /* 0x120000003000023c */
/*0458*/ @!P0 MOV32I R17, 0x8; /* 0x740000000423c046 */
/*0460*/ @!P0 DADD R18, R26, R26; /* 0xe38000000d20684a */
/*0468*/ @!P0 IMAD R14.CC, R46, R17, c[0x0][0x178]; /* 0x910c44002f20b83a */
/*0470*/ @!P0 IMAD.HI.X R15, R46, R17, c[0x0][0x17c]; /* 0x931844002fa0b83e */
/*0478*/ @!P0 IMAD R16.CC, R46, R17, c[0x0][0x168]; /* 0x910c44002d20b842 */
/* 0x08a180a5dc10bd9c */
/*0488*/ @!P0 LD.E.64 R12, [R14]; /* 0xc580000000203830 */
/*0490*/ @!P0 IMAD.HI.X R17, R46, R17, c[0x0][0x16c]; /* 0x931844002da0b846 */
/*0498*/ @!P0 LD.E.64 R20, [R16]; /* 0xc580000000204050 */
/*04a0*/ @!P0 DADD R22, R18, -R12; /* 0xe38100000620485a */
/*04a8*/ @!P0 DFMA R10, R20, R10, R22; /* 0xdb8058000520502a */
/*04b0*/ @!P0 ST.E.64 [R14], R10; /* 0xe580000000203828 */
/*04b8*/ IADD R39, R39, 0x20; /* 0xc0800000101c9c9d */
/* 0x08b0a0b8b0a0b8b0 */
/*04c8*/ ISETP.LT.AND P0, PT, R39, c[0x0][0x144], PT; /* 0x5b181c00289c9c1e */
/*04d0*/ @P0 BRA 0x178; /* 0x12007ffe5000003c */
/*04d8*/ IADD R29, R29, 0x1; /* 0xc0800000009c7475 */
/*04e0*/ ISETP.LT.AND P0, PT, R29, c[0x0][0x14c], PT; /* 0x5b181c00299c741e */
/*04e8*/ @P0 BRA 0xe0; /* 0x12007ffdf800003c */
/*04f0*/ IADD R0, R0, 0x1; /* 0xc0800000009c0001 */
/*04f8*/ ISETP.LT.AND P0, PT, R0, R49, PT; /* 0xdb181c00189c001e */
/* 0x0800000000b810b8 */
/*0508*/ @P0 BRA 0xb0; /* 0x12007ffdd000003c */
/*0510*/ MOV RZ, RZ; /* 0xe4c03c007f9c03fe */
/*0518*/ EXIT ; /* 0x18000000001c003c */
/*0520*/ BRA 0x520; /* 0x12007ffffc1c003c */
/*0528*/ NOP; /* 0x85800000001c3c02 */
/*0530*/ NOP; /* 0x85800000001c3c02 */
/*0538*/ NOP; /* 0x85800000001c3c02 */
..................................