Fatbin elf code: ================ arch = sm_90 code version = [1,7] host = linux compile_size = 64bit identifier = /workspace/bench.cu code for sm_90 Fatbin elf code: ================ arch = sm_90 code version = [1,7] host = linux compile_size = 64bit identifier = /workspace/bench.cu code for sm_90 Function : _Z12ragged_queueILb0EEvPKjPVjPjPKiP5StampiS4_ .headerflags @"EF_CUDA_TEXMODE_UNIFIED EF_CUDA_64BIT_ADDRESS EF_CUDA_SM90 EF_CUDA_VIRTUAL_SM(EF_CUDA_SM90)" /*0000*/ LDC R1, c[0x0][0x28] ; /* 0x00000a00ff017b82 */ /* 0x000e220000000800 */ /*0010*/ S2R R0, SR_TID.X ; /* 0x0000000000007919 */ /* 0x000e620000002100 */ /*0020*/ ULDC.64 UR6, c[0x0][0x208] ; /* 0x0000820000067ab9 */ /* 0x000fca0000000a00 */ /*0030*/ ISETP.NE.AND P0, PT, R0, RZ, PT ; /* 0x000000ff0000720c */ /* 0x002fe20003f05270 */ /*0040*/ BSSY B0, 0xe0 ; /* 0x0000009000007945 */ /* 0x000fd80003800000 */ /*0050*/ @P0 BRA 0xd0 ; /* 0x00000000001c0947 */ /* 0x000fea0003800000 */ /*0060*/ LDC.64 R2, c[0x0][0x240] ; /* 0x00009000ff027b82 */ /* 0x000e620000000a00 */ /*0070*/ IMAD.MOV.U32 R5, RZ, RZ, 0x1 ; /* 0x00000001ff057424 */ /* 0x000fca00078e00ff */ /*0080*/ ATOMG.E.ADD.STRONG.GPU PT, R2, desc[UR6][R2.64], R5 ; /* 0x00000005020279a8 */ /* 0x002ea400081ee1c6 */ /*0090*/ S2UR UR5, SR_CgaCtaId ; /* 0x00000000000579c3 */ /* 0x000e620000008800 */ /*00a0*/ UMOV UR4, 0x400 ; /* 0x0000040000047882 */ /* 0x000fe40000000000 */ /*00b0*/ ULEA UR4, UR5, UR4, 0x18 ; /* 0x0000000405047291 */ /* 0x002fd2000f8ec03f */ /*00c0*/ STS [UR4], R2 ; /* 0x00000002ff007988 */ /* 0x0043e60008000804 */ /*00d0*/ BSYNC B0 ; /* 0x0000000000007941 */ /* 0x000fea0003800000 */ /*00e0*/ S2UR UR5, SR_CgaCtaId ; /* 0x00000000000579c3 */ /* 0x000eb00000008800 */ /*00f0*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fec0000010000 */ /*0100*/ UMOV UR4, 0x400 ; /* 0x0000040000047882 */ /* 0x000fc40000000000 */ /*0110*/ ULEA UR5, UR5, UR4, 0x18 ; /* 0x0000000405057291 */ /* 0x004fc6000f8ec03f */ /*0120*/ ULDC UR4, c[0x0][0x238] ; /* 0x00008e0000047ab9 */ /* 0x000fcc0000000800 */ /*0130*/ LDS R7, [UR5] ; /* 0x00000005ff077984 */ /* 0x000ea40008000800 */ /*0140*/ ISETP.GE.U32.AND P0, PT, R7, UR4, PT ; /* 0x0000000407007c0c */ /* 0x004fda000bf06070 */ /*0150*/ @P0 EXIT ; /* 0x000000000000094d */ /* 0x000fea0003800000 */ /*0160*/ LDC.64 R4, c[0x0][0x228] ; /* 0x00008a00ff047b82 */ /* 0x000eb00000000a00 */ /*0170*/ LDC.64 R2, c[0x0][0x210] ; /* 0x00008400ff027b82 */ /* 0x002e620000000a00 */ /*0180*/ IMAD.WIDE R4, R7, 0x4, R4 ; /* 0x0000000407047825 */ /* 0x004fcc00078e0204 */ /*0190*/ LDG.E R4, desc[UR6][R4.64] ; /* 0x0000000604047981 */ /* 0x000ea2000c1e1900 */ /*01a0*/ IMAD R7, R7, 0x100, R0 ; /* 0x0000010007077824 */ /* 0x000fc800078e0200 */ /*01b0*/ IMAD.WIDE R2, R7, 0x4, R2 ; /* 0x0000000407027825 */ /* 0x002fca00078e0202 */ /*01c0*/ LDG.E R9, desc[UR6][R2.64] ; /* 0x0000000602097981 */ /* 0x000362000c1e1900 */ /*01d0*/ ISETP.GE.AND P0, PT, R4, 0x1, PT ; /* 0x000000010400780c */ /* 0x004fda0003f06270 */ /*01e0*/ @!P0 BRA 0x250 ; /* 0x0000000000188947 */ /* 0x002fea0003800000 */ /*01f0*/ IMAD.MOV.U32 R2, RZ, RZ, 0x19660d ; /* 0x0019660dff027424 */ /* 0x000fe200078e00ff */ /*0200*/ UMOV UR4, URZ ; /* 0x0000003f00047c82 */ /* 0x000fcc0008000000 */ /*0210*/ UIADD3 UR4, UR4, 0x1, URZ ; /* 0x0000000104047890 */ /* 0x000fe2000fffe03f */ /*0220*/ IMAD R9, R9, R2, 0x3c6ef35f ; /* 0x3c6ef35f09097424 */ /* 0x020fca00078e0202 */ /*0230*/ ISETP.LE.AND P0, PT, R4, UR4, PT ; /* 0x0000000404007c0c */ /* 0x000fda000bf03270 */ /*0240*/ @!P0 BRA 0x210 ; /* 0xfffffffc00f08947 */ /* 0x000fea000383ffff */ /*0250*/ ULDC.64 UR8, c[0x0][0x218] ; /* 0x0000860000087ab9 */ /* 0x000fe20000000a00 */ /*0260*/ SHF.R.S32.HI R2, RZ, 0x1f, R7 ; /* 0x0000001fff027819 */ /* 0x000fe40000011407 */ /*0270*/ LEA R4, P0, R7, UR8, 0x2 ; /* 0x0000000807047c11 */ /* 0x000fc8000f8010ff */ /*0280*/ LEA.HI.X R5, R7, UR9, R2, 0x2, P0 ; /* 0x0000000907057c11 */ /* 0x000fe400080f1402 */ /*0290*/ LDC.64 R2, c[0x0][0x218] ; /* 0x00008600ff027b82 */ /* 0x000e660000000a00 */ /*02a0*/ STG.E.STRONG.SYS desc[UR6][R4.64], R9 ; /* 0x0000000904007986 */ /* 0x020fea000c115906 */ /*02b0*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000ff00000010000 */ /*02c0*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fec0000010000 */ /*02d0*/ LDS R7, [UR5] ; /* 0x00000005ff077984 */ /* 0x000ea40008000800 */ /*02e0*/ LEA R11, R7, R0, 0x8 ; /* 0x00000000070b7211 */ /* 0x004fca00078e40ff */ /*02f0*/ IMAD.WIDE R2, R11, 0x4, R2 ; /* 0x000000040b027825 */ /* 0x002fca00078e0202 */ /*0300*/ LDG.E.STRONG.SYS R7, desc[UR6][R2.64] ; /* 0x0000000602077981 */ /* 0x000362000c1f5900 */ /*0310*/ IMAD.MOV.U32 R6, RZ, RZ, RZ ; /* 0x000000ffff067224 */ /* 0x000fe200078e00ff */ /*0320*/ SHF.R.S32.HI R8, RZ, 0x1f, R11 ; /* 0x0000001fff087819 */ /* 0x000fe2000001140b */ /*0330*/ IMAD.MOV.U32 R10, RZ, RZ, 0x1 ; /* 0x00000001ff0a7424 */ /* 0x000fce00078e00ff */ /*0340*/ IADD3 R6, R6, 0x1, RZ ; /* 0x0000000106067810 */ /* 0x000fe20007ffe0ff */ /*0350*/ IMAD R7, R7, 0x15a4e35, R10 ; /* 0x015a4e3507077824 */ /* 0x020fc600078e020a */ /*0360*/ ISETP.NE.AND P0, PT, R6, 0x200, PT ; /* 0x000002000600780c */ /* 0x000fda0003f05270 */ /*0370*/ @P0 BRA 0x340 ; /* 0xfffffffc00f00947 */ /* 0x000fea000383ffff */ /*0380*/ ULDC.64 UR4, c[0x0][0x220] ; /* 0x0000880000047ab9 */ /* 0x000fe40000000a00 */ /*0390*/ LEA R2, P0, R11, UR4, 0x2 ; /* 0x000000040b027c11 */ /* 0x002fc8000f8010ff */ /*03a0*/ LEA.HI.X R3, R11, UR5, R8, 0x2, P0 ; /* 0x000000050b037c11 */ /* 0x000fca00080f1408 */ /*03b0*/ STG.E desc[UR6][R2.64], R7 ; /* 0x0000000702007986 */ /* 0x0005e2000c101906 */ /*03c0*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000ff00000010000 */ /*03d0*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fec0000010000 */ /*03e0*/ BRA 0x30 ; /* 0xfffffffc00107947 */ /* 0x004fea000383ffff */ /*03f0*/ BRA 0x3f0; /* 0xfffffffc00fc7947 */ /* 0x000fc0000383ffff */ /*0400*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0410*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0420*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0430*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0440*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0450*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0460*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0470*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ .......... Function : _Z12ragged_fusedILb0EEvPKjPVjPjPKiP5Stamp .headerflags @"EF_CUDA_TEXMODE_UNIFIED EF_CUDA_64BIT_ADDRESS EF_CUDA_SM90 EF_CUDA_VIRTUAL_SM(EF_CUDA_SM90)" /*0000*/ LDC R1, c[0x0][0x28] ; /* 0x00000a00ff017b82 */ /* 0x000e220000000800 */ /*0010*/ S2R R7, SR_CTAID.X ; /* 0x0000000000077919 */ /* 0x000e6e0000002500 */ /*0020*/ LDC.64 R4, c[0x0][0x228] ; /* 0x00008a00ff047b82 */ /* 0x000e620000000a00 */ /*0030*/ ULDC.64 UR6, c[0x0][0x208] ; /* 0x0000820000067ab9 */ /* 0x000fe20000000a00 */ /*0040*/ ULDC.64 UR4, c[0x0][0x218] ; /* 0x0000860000047ab9 */ /* 0x000fe20000000a00 */ /*0050*/ S2R R0, SR_TID.X ; /* 0x0000000000007919 */ /* 0x000eaa0000002100 */ /*0060*/ LDC.64 R2, c[0x0][0x210] ; /* 0x00008400ff027b82 */ /* 0x000ee20000000a00 */ /*0070*/ IMAD.WIDE R4, R7, 0x4, R4 ; /* 0x0000000407047825 */ /* 0x002fcc00078e0204 */ /*0080*/ LDG.E R4, desc[UR6][R4.64] ; /* 0x0000000604047981 */ /* 0x000f22000c1e1900 */ /*0090*/ IMAD R7, R7, 0x100, R0 ; /* 0x0000010007077824 */ /* 0x004fc800078e0200 */ /*00a0*/ IMAD.SHL.U32 R8, R7.reuse, 0x4, RZ ; /* 0x0000000407087824 */ /* 0x040fe200078e00ff */ /*00b0*/ SHF.R.S32.HI R0, RZ, 0x1f, R7 ; /* 0x0000001fff007819 */ /* 0x000fe20000011407 */ /*00c0*/ IMAD.WIDE R2, R7, 0x4, R2 ; /* 0x0000000407027825 */ /* 0x008fc600078e0202 */ /*00d0*/ SHF.L.U64.HI R9, R7, 0x2, R0 ; /* 0x0000000207097819 */ /* 0x000fe40000010200 */ /*00e0*/ IADD3 R6, P1, R8, UR4, RZ ; /* 0x0000000408067c10 */ /* 0x000fe2000ff3e0ff */ /*00f0*/ LDG.E R3, desc[UR6][R2.64] ; /* 0x0000000602037981 */ /* 0x000366000c1e1900 */ /*0100*/ IADD3.X R7, R9, UR5, RZ, P1, !PT ; /* 0x0000000509077c10 */ /* 0x000fe40008ffe4ff */ /*0110*/ ISETP.GE.AND P0, PT, R4, 0x1, PT ; /* 0x000000010400780c */ /* 0x010fda0003f06270 */ /*0120*/ @!P0 BRA 0x190 ; /* 0x0000000000188947 */ /* 0x003fea0003800000 */ /*0130*/ IMAD.MOV.U32 R0, RZ, RZ, 0x19660d ; /* 0x0019660dff007424 */ /* 0x000fe200078e00ff */ /*0140*/ UMOV UR4, URZ ; /* 0x0000003f00047c82 */ /* 0x000fcc0008000000 */ /*0150*/ UIADD3 UR4, UR4, 0x1, URZ ; /* 0x0000000104047890 */ /* 0x000fe2000fffe03f */ /*0160*/ IMAD R3, R3, R0, 0x3c6ef35f ; /* 0x3c6ef35f03037424 */ /* 0x020fca00078e0200 */ /*0170*/ ISETP.LE.AND P0, PT, R4, UR4, PT ; /* 0x0000000404007c0c */ /* 0x000fda000bf03270 */ /*0180*/ @!P0 BRA 0x150 ; /* 0xfffffffc00f08947 */ /* 0x000fea000383ffff */ /*0190*/ STG.E.STRONG.SYS desc[UR6][R6.64], R3 ; /* 0x0000000306007986 */ /* 0x0201e2000c115906 */ /*01a0*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000ff00000010000 */ /*01b0*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fec0000010000 */ /*01c0*/ LDG.E.STRONG.SYS R5, desc[UR6][R6.64] ; /* 0x0000000606057981 */ /* 0x000162000c1f5900 */ /*01d0*/ HFMA2.MMA R0, -RZ, RZ, 0, 0 ; /* 0x00000000ff007435 */ /* 0x000fe200000001ff */ /*01e0*/ IMAD.MOV.U32 R2, RZ, RZ, 0x1 ; /* 0x00000001ff027424 */ /* 0x000fd400078e00ff */ /*01f0*/ IADD3 R0, R0, 0x1, RZ ; /* 0x0000000100007810 */ /* 0x000fe20007ffe0ff */ /*0200*/ IMAD R5, R5, 0x15a4e35, R2 ; /* 0x015a4e3505057824 */ /* 0x020fc600078e0202 */ /*0210*/ ISETP.NE.AND P0, PT, R0, 0x200, PT ; /* 0x000002000000780c */ /* 0x000fda0003f05270 */ /*0220*/ @P0 BRA 0x1f0 ; /* 0xfffffffc00f00947 */ /* 0x000fea000383ffff */ /*0230*/ ULDC.64 UR4, c[0x0][0x220] ; /* 0x0000880000047ab9 */ /* 0x000fe40000000a00 */ /*0240*/ IADD3 R2, P0, R8, UR4, RZ ; /* 0x0000000408027c10 */ /* 0x000fc8000ff1e0ff */ /*0250*/ IADD3.X R3, R9, UR5, RZ, P0, !PT ; /* 0x0000000509037c10 */ /* 0x001fca00087fe4ff */ /*0260*/ STG.E desc[UR6][R2.64], R5 ; /* 0x0000000502007986 */ /* 0x000fe2000c101906 */ /*0270*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fec0000010000 */ /*0280*/ EXIT ; /* 0x000000000000794d */ /* 0x000fea0003800000 */ /*0290*/ BRA 0x290; /* 0xfffffffc00fc7947 */ /* 0x000fc0000383ffff */ /*02a0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*02b0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*02c0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*02d0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*02e0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*02f0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0300*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0310*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0320*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0330*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0340*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0350*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0360*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0370*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ .......... Function : _Z8ragged_bILb0EEvPVKjPjP5Stamp .headerflags @"EF_CUDA_TEXMODE_UNIFIED EF_CUDA_64BIT_ADDRESS EF_CUDA_SM90 EF_CUDA_VIRTUAL_SM(EF_CUDA_SM90)" /*0000*/ LDC R1, c[0x0][0x28] ; /* 0x00000a00ff017b82 */ /* 0x000e220000000800 */ /*0010*/ S2R R5, SR_CTAID.X ; /* 0x0000000000057919 */ /* 0x000e6e0000002500 */ /*0020*/ LDC.64 R2, c[0x0][0x210] ; /* 0x00008400ff027b82 */ /* 0x000ea20000000a00 */ /*0030*/ ULDC.64 UR4, c[0x0][0x208] ; /* 0x0000820000047ab9 */ /* 0x000fe20000000a00 */ /*0040*/ S2R R0, SR_TID.X ; /* 0x0000000000007919 */ /* 0x000e640000002100 */ /*0050*/ IMAD R5, R5, 0x100, R0 ; /* 0x0000010005057824 */ /* 0x002fc800078e0200 */ /*0060*/ IMAD.WIDE R2, R5, 0x4, R2 ; /* 0x0000000405027825 */ /* 0x004fca00078e0202 */ /*0070*/ LDG.E.STRONG.SYS R7, desc[UR4][R2.64] ; /* 0x0000000402077981 */ /* 0x000362000c1f5900 */ /*0080*/ HFMA2.MMA R6, -RZ, RZ, 0, 5.9604644775390625e-08 ; /* 0x00000001ff067435 */ /* 0x000fe200000001ff */ /*0090*/ IMAD.MOV.U32 R0, RZ, RZ, RZ ; /* 0x000000ffff007224 */ /* 0x000fe200078e00ff */ /*00a0*/ SHF.R.S32.HI R4, RZ, 0x1f, R5 ; /* 0x0000001fff047819 */ /* 0x001fd20000011405 */ /*00b0*/ VIADD R0, R0, 0x1 ; /* 0x0000000100007836 */ /* 0x000fe40000000000 */ /*00c0*/ IMAD R7, R7, 0x15a4e35, R6 ; /* 0x015a4e3507077824 */ /* 0x020fc600078e0206 */ /*00d0*/ ISETP.NE.AND P0, PT, R0, 0x200, PT ; /* 0x000002000000780c */ /* 0x000fda0003f05270 */ /*00e0*/ @P0 BRA 0xb0 ; /* 0xfffffffc00f00947 */ /* 0x000fea000383ffff */ /*00f0*/ ULDC.64 UR6, c[0x0][0x218] ; /* 0x0000860000067ab9 */ /* 0x000fe40000000a00 */ /*0100*/ LEA R2, P0, R5, UR6, 0x2 ; /* 0x0000000605027c11 */ /* 0x002fc8000f8010ff */ /*0110*/ LEA.HI.X R3, R5, UR7, R4, 0x2, P0 ; /* 0x0000000705037c11 */ /* 0x000fca00080f1404 */ /*0120*/ STG.E desc[UR4][R2.64], R7 ; /* 0x0000000702007986 */ /* 0x000fe2000c101904 */ /*0130*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fec0000010000 */ /*0140*/ EXIT ; /* 0x000000000000794d */ /* 0x000fea0003800000 */ /*0150*/ BRA 0x150; /* 0xfffffffc00fc7947 */ /* 0x000fc0000383ffff */ /*0160*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0170*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0180*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0190*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01a0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01b0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01c0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01d0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01e0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01f0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ .......... Function : _Z8ragged_aILb0EEvPKjPVjPKiP5Stamp .headerflags @"EF_CUDA_TEXMODE_UNIFIED EF_CUDA_64BIT_ADDRESS EF_CUDA_SM90 EF_CUDA_VIRTUAL_SM(EF_CUDA_SM90)" /*0000*/ LDC R1, c[0x0][0x28] ; /* 0x00000a00ff017b82 */ /* 0x000e220000000800 */ /*0010*/ S2R R7, SR_CTAID.X ; /* 0x0000000000077919 */ /* 0x000e6e0000002500 */ /*0020*/ LDC.64 R4, c[0x0][0x220] ; /* 0x00008800ff047b82 */ /* 0x000e620000000a00 */ /*0030*/ ULDC.64 UR6, c[0x0][0x208] ; /* 0x0000820000067ab9 */ /* 0x000fe20000000a00 */ /*0040*/ S2R R0, SR_TID.X ; /* 0x0000000000007919 */ /* 0x000eac0000002100 */ /*0050*/ LDC.64 R2, c[0x0][0x210] ; /* 0x00008400ff027b82 */ /* 0x000ee20000000a00 */ /*0060*/ IMAD.WIDE R4, R7, 0x4, R4 ; /* 0x0000000407047825 */ /* 0x002fcc00078e0204 */ /*0070*/ LDG.E R4, desc[UR6][R4.64] ; /* 0x0000000604047981 */ /* 0x000f22000c1e1900 */ /*0080*/ IMAD R7, R7, 0x100, R0 ; /* 0x0000010007077824 */ /* 0x004fc800078e0200 */ /*0090*/ IMAD.WIDE R2, R7, 0x4, R2 ; /* 0x0000000407027825 */ /* 0x008fca00078e0202 */ /*00a0*/ LDG.E R9, desc[UR6][R2.64] ; /* 0x0000000602097981 */ /* 0x000362000c1e1900 */ /*00b0*/ ISETP.GE.AND P0, PT, R4, 0x1, PT ; /* 0x000000010400780c */ /* 0x010fda0003f06270 */ /*00c0*/ @!P0 BRA 0x130 ; /* 0x0000000000188947 */ /* 0x003fea0003800000 */ /*00d0*/ IMAD.MOV.U32 R0, RZ, RZ, 0x19660d ; /* 0x0019660dff007424 */ /* 0x000fe200078e00ff */ /*00e0*/ UMOV UR4, URZ ; /* 0x0000003f00047c82 */ /* 0x000fcc0008000000 */ /*00f0*/ UIADD3 UR4, UR4, 0x1, URZ ; /* 0x0000000104047890 */ /* 0x000fe2000fffe03f */ /*0100*/ IMAD R9, R9, R0, 0x3c6ef35f ; /* 0x3c6ef35f09097424 */ /* 0x020fca00078e0200 */ /*0110*/ ISETP.LE.AND P0, PT, R4, UR4, PT ; /* 0x0000000404007c0c */ /* 0x000fda000bf03270 */ /*0120*/ @!P0 BRA 0xf0 ; /* 0xfffffffc00f08947 */ /* 0x000fea000383ffff */ /*0130*/ ULDC.64 UR4, c[0x0][0x218] ; /* 0x0000860000047ab9 */ /* 0x000fe20000000a00 */ /*0140*/ SHF.R.S32.HI R0, RZ, 0x1f, R7 ; /* 0x0000001fff007819 */ /* 0x000fe40000011407 */ /*0150*/ LEA R2, P0, R7, UR4, 0x2 ; /* 0x0000000407027c11 */ /* 0x000fc8000f8010ff */ /*0160*/ LEA.HI.X R3, R7, UR5, R0, 0x2, P0 ; /* 0x0000000507037c11 */ /* 0x000fca00080f1400 */ /*0170*/ STG.E.STRONG.SYS desc[UR6][R2.64], R9 ; /* 0x0000000902007986 */ /* 0x020fe2000c115906 */ /*0180*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fec0000010000 */ /*0190*/ EXIT ; /* 0x000000000000794d */ /* 0x000fea0003800000 */ /*01a0*/ BRA 0x1a0; /* 0xfffffffc00fc7947 */ /* 0x000fc0000383ffff */ /*01b0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01c0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01d0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01e0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01f0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0200*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0210*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0220*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0230*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0240*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0250*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0260*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0270*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ .......... Function : _Z12ragged_queueILb1EEvPKjPVjPjPKiP5StampiS4_ .headerflags @"EF_CUDA_TEXMODE_UNIFIED EF_CUDA_64BIT_ADDRESS EF_CUDA_SM90 EF_CUDA_VIRTUAL_SM(EF_CUDA_SM90)" /*0000*/ LDC R1, c[0x0][0x28] ; /* 0x00000a00ff017b82 */ /* 0x000e220000000800 */ /*0010*/ S2R R0, SR_TID.X ; /* 0x0000000000007919 */ /* 0x000e620000002100 */ /*0020*/ ULDC.64 UR6, c[0x0][0x208] ; /* 0x0000820000067ab9 */ /* 0x000fe20000000a00 */ /*0030*/ S2R R3, SR_CTAID.X ; /* 0x0000000000037919 */ /* 0x000ea80000002500 */ /*0040*/ ISETP.NE.AND P0, PT, R0, RZ, PT ; /* 0x000000ff0000720c */ /* 0x002fe20003f05270 */ /*0050*/ BSSY B0, 0xf0 ; /* 0x0000009000007945 */ /* 0x000fd80003800000 */ /*0060*/ @P0 BRA 0xe0 ; /* 0x00000000001c0947 */ /* 0x008fea0003800000 */ /*0070*/ LDC.64 R4, c[0x0][0x240] ; /* 0x00009000ff047b82 */ /* 0x000e620000000a00 */ /*0080*/ IMAD.MOV.U32 R7, RZ, RZ, 0x1 ; /* 0x00000001ff077424 */ /* 0x000fca00078e00ff */ /*0090*/ ATOMG.E.ADD.STRONG.GPU PT, R4, desc[UR6][R4.64], R7 ; /* 0x00000007040479a8 */ /* 0x002ee400081ee1c6 */ /*00a0*/ S2UR UR5, SR_CgaCtaId ; /* 0x00000000000579c3 */ /* 0x000e620000008800 */ /*00b0*/ UMOV UR4, 0x400 ; /* 0x0000040000047882 */ /* 0x000fe40000000000 */ /*00c0*/ ULEA UR4, UR5, UR4, 0x18 ; /* 0x0000000405047291 */ /* 0x002fd2000f8ec03f */ /*00d0*/ STS [UR4], R4 ; /* 0x00000004ff007988 */ /* 0x0083e60008000804 */ /*00e0*/ BSYNC B0 ; /* 0x0000000000007941 */ /* 0x000fea0003800000 */ /*00f0*/ S2UR UR5, SR_CgaCtaId ; /* 0x00000000000579c3 */ /* 0x000ef00000008800 */ /*0100*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fec0000010000 */ /*0110*/ UMOV UR4, 0x400 ; /* 0x0000040000047882 */ /* 0x000fc40000000000 */ /*0120*/ ULEA UR5, UR5, UR4, 0x18 ; /* 0x0000000405057291 */ /* 0x008fc6000f8ec03f */ /*0130*/ ULDC UR4, c[0x0][0x238] ; /* 0x00008e0000047ab9 */ /* 0x000fcc0000000800 */ /*0140*/ LDS R15, [UR5] ; /* 0x00000005ff0f7984 */ /* 0x000ee40008000800 */ /*0150*/ ISETP.GE.U32.AND P1, PT, R15, UR4, PT ; /* 0x000000040f007c0c */ /* 0x008fda000bf26070 */ /*0160*/ @P1 EXIT ; /* 0x000000000000194d */ /* 0x000fea0003800000 */ /*0170*/ LDC.64 R4, c[0x0][0x230] ; /* 0x00008c00ff047b82 */ /* 0x002e620000000a00 */ /*0180*/ SHF.R.S32.HI R2, RZ, 0x1f, R15 ; /* 0x0000001fff027819 */ /* 0x000fe4000001140f */ /*0190*/ ISETP.NE.AND P1, PT, R0, RZ, PT ; /* 0x000000ff0000720c */ /* 0x000fc60003f25270 */ /*01a0*/ IMAD R9, R2, 0x30, RZ ; /* 0x0000003002097824 */ /* 0x000fe400078e02ff */ /*01b0*/ IMAD.WIDE.U32 R6, R15, 0x30, R4 ; /* 0x000000300f067825 */ /* 0x002fc800078e0004 */ /*01c0*/ IMAD.IADD R7, R7, 0x1, R9 ; /* 0x0000000107077824 */ /* 0x000fc800078e0209 */ /*01d0*/ @!P1 CS2R R12, SR_GLOBALTIMERLO ; /* 0x00000000000c9805 */ /* 0x000fe20000015200 */ /*01e0*/ @!P1 S2R R2, SR_VIRTUALSMID ; /* 0x0000000000029919 */ /* 0x000e620000004300 */ /*01f0*/ LDC.64 R10, c[0x0][0x228] ; /* 0x00008a00ff0a7b82 */ /* 0x000ee20000000a00 */ /*0200*/ ULDC.64 UR8, c[0x0][0x218] ; /* 0x0000860000087ab9 */ /* 0x000fe40000000a00 */ /*0210*/ @!P1 STG.E.64 desc[UR6][R6.64], R12 ; /* 0x0000000c06009986 */ /* 0x0009ea000c101b06 */ /*0220*/ LDC.64 R8, c[0x0][0x210] ; /* 0x00008400ff087b82 */ /* 0x000ea20000000a00 */ /*0230*/ IMAD.WIDE R10, R15.reuse, 0x4, R10 ; /* 0x000000040f0a7825 */ /* 0x048fe200078e020a */ /*0240*/ @!P1 STG.E.64 desc[UR6][R6.64+0x20], R2 ; /* 0x0000200206009986 */ /* 0x0069ea000c101b06 */ /*0250*/ LDG.E R10, desc[UR6][R10.64] ; /* 0x000000060a0a7981 */ /* 0x000ea2000c1e1900 */ /*0260*/ IMAD R15, R15, 0x100, R0 ; /* 0x000001000f0f7824 */ /* 0x000fc800078e0200 */ /*0270*/ IMAD.WIDE R8, R15, 0x4, R8 ; /* 0x000000040f087825 */ /* 0x000fcc00078e0208 */ /*0280*/ LDG.E R9, desc[UR6][R8.64] ; /* 0x0000000608097981 */ /* 0x000962000c1e1900 */ /*0290*/ SHF.R.S32.HI R16, RZ, 0x1f, R15 ; /* 0x0000001fff107819 */ /* 0x000fe4000001140f */ /*02a0*/ LEA R14, P2, R15, UR8, 0x2 ; /* 0x000000080f0e7c11 */ /* 0x000fc8000f8410ff */ /*02b0*/ LEA.HI.X R15, R15, UR9, R16, 0x2, P2 ; /* 0x000000090f0f7c11 */ /* 0x000fe400090f1410 */ /*02c0*/ ISETP.GE.AND P1, PT, R10, 0x1, PT ; /* 0x000000010a00780c */ /* 0x004fda0003f26270 */ /*02d0*/ @!P1 BRA 0x340 ; /* 0x0000000000189947 */ /* 0x010fea0003800000 */ /*02e0*/ IMAD.MOV.U32 R2, RZ, RZ, 0x19660d ; /* 0x0019660dff027424 */ /* 0x000fe200078e00ff */ /*02f0*/ UMOV UR4, URZ ; /* 0x0000003f00047c82 */ /* 0x000fcc0008000000 */ /*0300*/ UIADD3 UR4, UR4, 0x1, URZ ; /* 0x0000000104047890 */ /* 0x000fe2000fffe03f */ /*0310*/ IMAD R9, R9, R2, 0x3c6ef35f ; /* 0x3c6ef35f09097424 */ /* 0x020fca00078e0202 */ /*0320*/ ISETP.LE.AND P1, PT, R10, UR4, PT ; /* 0x000000040a007c0c */ /* 0x000fda000bf23270 */ /*0330*/ @!P1 BRA 0x300 ; /* 0xfffffffc00f09947 */ /* 0x000fea000383ffff */ /*0340*/ STG.E.STRONG.SYS desc[UR6][R14.64], R9 ; /* 0x000000090e007986 */ /* 0x0203e2000c115906 */ /*0350*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fe20000010000 */ /*0360*/ @!P0 CS2R R8, SR_GLOBALTIMERLO ; /* 0x0000000000088805 */ /* 0x002fca0000015200 */ /*0370*/ @!P0 STG.E.64 desc[UR6][R6.64+0x8], R8 ; /* 0x0000080806008986 */ /* 0x000fe4000c101b06 */ /*0380*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fec0000010000 */ /*0390*/ LDS R11, [UR5] ; /* 0x00000005ff0b7984 */ /* 0x000e640008000800 */ /*03a0*/ IMAD.WIDE R4, R11.reuse, 0x30, R4 ; /* 0x000000300b047825 */ /* 0x042fe200078e0204 */ /*03b0*/ @!P0 CS2R R8, SR_GLOBALTIMERLO ; /* 0x0000000000088805 */ /* 0x000fe20000015200 */ /*03c0*/ @!P0 S2R R14, SR_VIRTUALSMID ; /* 0x00000000000e8919 */ /* 0x000e620000004300 */ /*03d0*/ LDC.64 R6, c[0x0][0x218] ; /* 0x00008600ff067b82 */ /* 0x000ea20000000a00 */ /*03e0*/ IMAD R11, R11, 0x100, R0 ; /* 0x000001000b0b7824 */ /* 0x000fc400078e0200 */ /*03f0*/ IMAD.MOV.U32 R15, RZ, RZ, R3 ; /* 0x000000ffff0f7224 */ /* 0x000fe200078e0003 */ /*0400*/ @!P0 STG.E.64 desc[UR6][R4.64+0x10], R8 ; /* 0x0000100804008986 */ /* 0x0007e2000c101b06 */ /*0410*/ IMAD.WIDE R6, R11, 0x4, R6 ; /* 0x000000040b067825 */ /* 0x004fc600078e0206 */ /*0420*/ @!P0 STG.E.64 desc[UR6][R4.64+0x28], R14 ; /* 0x0000280e04008986 */ /* 0x0027e8000c101b06 */ /*0430*/ LDG.E.STRONG.SYS R13, desc[UR6][R6.64] ; /* 0x00000006060d7981 */ /* 0x000762000c1f5900 */ /*0440*/ IMAD.MOV.U32 R2, RZ, RZ, RZ ; /* 0x000000ffff027224 */ /* 0x000fe200078e00ff */ /*0450*/ SHF.R.S32.HI R10, RZ, 0x1f, R11 ; /* 0x0000001fff0a7819 */ /* 0x000fe2000001140b */ /*0460*/ IMAD.MOV.U32 R12, RZ, RZ, 0x1 ; /* 0x00000001ff0c7424 */ /* 0x000fce00078e00ff */ /*0470*/ VIADD R2, R2, 0x1 ; /* 0x0000000102027836 */ /* 0x000fe40000000000 */ /*0480*/ IMAD R13, R13, 0x15a4e35, R12 ; /* 0x015a4e350d0d7824 */ /* 0x020fc600078e020c */ /*0490*/ ISETP.NE.AND P1, PT, R2, 0x200, PT ; /* 0x000002000200780c */ /* 0x000fda0003f25270 */ /*04a0*/ @P1 BRA 0x470 ; /* 0xfffffffc00f01947 */ /* 0x000fea000383ffff */ /*04b0*/ ULDC.64 UR4, c[0x0][0x220] ; /* 0x0000880000047ab9 */ /* 0x000fe40000000a00 */ /*04c0*/ LEA R6, P1, R11, UR4, 0x2 ; /* 0x000000040b067c11 */ /* 0x008fc8000f8210ff */ /*04d0*/ LEA.HI.X R7, R11, UR5, R10, 0x2, P1 ; /* 0x000000050b077c11 */ /* 0x000fca00088f140a */ /*04e0*/ STG.E desc[UR6][R6.64], R13 ; /* 0x0000000d06007986 */ /* 0x0003e2000c101906 */ /*04f0*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fe20000010000 */ /*0500*/ @!P0 CS2R R6, SR_GLOBALTIMERLO ; /* 0x0000000000068805 */ /* 0x002fca0000015200 */ /*0510*/ @!P0 STG.E.64 desc[UR6][R4.64+0x18], R6 ; /* 0x0000180604008986 */ /* 0x0007e4000c101b06 */ /*0520*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fec0000010000 */ /*0530*/ BRA 0x40 ; /* 0xfffffff800c07947 */ /* 0x000fea000383ffff */ /*0540*/ BRA 0x540; /* 0xfffffffc00fc7947 */ /* 0x000fc0000383ffff */ /*0550*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0560*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0570*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0580*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0590*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*05a0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*05b0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*05c0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*05d0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*05e0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*05f0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ .......... Function : _Z12ragged_fusedILb1EEvPKjPVjPjPKiP5Stamp .headerflags @"EF_CUDA_TEXMODE_UNIFIED EF_CUDA_64BIT_ADDRESS EF_CUDA_SM90 EF_CUDA_VIRTUAL_SM(EF_CUDA_SM90)" /*0000*/ LDC R1, c[0x0][0x28] ; /* 0x00000a00ff017b82 */ /* 0x000e220000000800 */ /*0010*/ S2R R0, SR_TID.X ; /* 0x0000000000007919 */ /* 0x000e6e0000002100 */ /*0020*/ LDC.64 R2, c[0x0][0x230] ; /* 0x00008c00ff027b82 */ /* 0x000ea20000000a00 */ /*0030*/ ULDC.64 UR6, c[0x0][0x208] ; /* 0x0000820000067ab9 */ /* 0x000fe20000000a00 */ /*0040*/ S2R R5, SR_CTAID.X ; /* 0x0000000000057919 */ /* 0x000ea20000002500 */ /*0050*/ ISETP.NE.AND P0, PT, R0, RZ, PT ; /* 0x000000ff0000720c */ /* 0x002fe20003f05270 */ /*0060*/ IMAD.WIDE R2, R5, 0x30, R2 ; /* 0x0000003005027825 */ /* 0x004fd800078e0202 */ /*0070*/ @!P0 CS2R R10, SR_GLOBALTIMERLO ; /* 0x00000000000a8805 */ /* 0x000fe20000015200 */ /*0080*/ @!P0 S2R R4, SR_VIRTUALSMID ; /* 0x0000000000048919 */ /* 0x000e620000004300 */ /*0090*/ LDC.64 R8, c[0x0][0x228] ; /* 0x00008a00ff087b82 */ /* 0x000ea60000000a00 */ /*00a0*/ @!P0 STG.E.64 desc[UR6][R2.64], R10 ; /* 0x0000000a02008986 */ /* 0x0007ea000c101b06 */ /*00b0*/ LDC.64 R6, c[0x0][0x210] ; /* 0x00008400ff067b82 */ /* 0x000f220000000a00 */ /*00c0*/ IMAD R13, R5.reuse, 0x100, R0 ; /* 0x00000100050d7824 */ /* 0x040fe200078e0200 */ /*00d0*/ ULDC.64 UR4, c[0x0][0x218] ; /* 0x0000860000047ab9 */ /* 0x000fe20000000a00 */ /*00e0*/ IMAD.WIDE R8, R5, 0x4, R8 ; /* 0x0000000405087825 */ /* 0x004fe200078e0208 */ /*00f0*/ @!P0 STG.E.64 desc[UR6][R2.64+0x20], R4 ; /* 0x0000200402008986 */ /* 0x0023ea000c101b06 */ /*0100*/ LDG.E R8, desc[UR6][R8.64] ; /* 0x0000000608087981 */ /* 0x000ea2000c1e1900 */ /*0110*/ IMAD.WIDE R6, R13, 0x4, R6 ; /* 0x000000040d067825 */ /* 0x010fcc00078e0206 */ /*0120*/ LDG.E R7, desc[UR6][R6.64] ; /* 0x0000000606077981 */ /* 0x000362000c1e1900 */ /*0130*/ IMAD.SHL.U32 R14, R13, 0x4, RZ ; /* 0x000000040d0e7824 */ /* 0x000fe200078e00ff */ /*0140*/ SHF.R.S32.HI R12, RZ, 0x1f, R13 ; /* 0x0000001fff0c7819 */ /* 0x000fc8000001140d */ /*0150*/ SHF.L.U64.HI R12, R13, 0x2, R12 ; /* 0x000000020d0c7819 */ /* 0x000fe4000001020c */ /*0160*/ IADD3 R10, P1, R14, UR4, RZ ; /* 0x000000040e0a7c10 */ /* 0x008fc8000ff3e0ff */ /*0170*/ IADD3.X R11, R12, UR5, RZ, P1, !PT ; /* 0x000000050c0b7c10 */ /* 0x000fe40008ffe4ff */ /*0180*/ ISETP.GE.AND P0, PT, R8, 0x1, PT ; /* 0x000000010800780c */ /* 0x004fda0003f06270 */ /*0190*/ @!P0 BRA 0x200 ; /* 0x0000000000188947 */ /* 0x003fea0003800000 */ /*01a0*/ IMAD.MOV.U32 R4, RZ, RZ, 0x19660d ; /* 0x0019660dff047424 */ /* 0x000fe200078e00ff */ /*01b0*/ UMOV UR4, URZ ; /* 0x0000003f00047c82 */ /* 0x000fcc0008000000 */ /*01c0*/ UIADD3 UR4, UR4, 0x1, URZ ; /* 0x0000000104047890 */ /* 0x000fe2000fffe03f */ /*01d0*/ IMAD R7, R7, R4, 0x3c6ef35f ; /* 0x3c6ef35f07077424 */ /* 0x020fca00078e0204 */ /*01e0*/ ISETP.LE.AND P0, PT, R8, UR4, PT ; /* 0x0000000408007c0c */ /* 0x000fda000bf03270 */ /*01f0*/ @!P0 BRA 0x1c0 ; /* 0xfffffffc00f08947 */ /* 0x000fea000383ffff */ /*0200*/ STG.E.STRONG.SYS desc[UR6][R10.64], R7 ; /* 0x000000070a007986 */ /* 0x0201e2000c115906 */ /*0210*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fe20000010000 */ /*0220*/ ISETP.NE.AND P0, PT, R0, RZ, PT ; /* 0x000000ff0000720c */ /* 0x000fda0003f05270 */ /*0230*/ @!P0 CS2R R6, SR_GLOBALTIMERLO ; /* 0x0000000000068805 */ /* 0x001fca0000015200 */ /*0240*/ @!P0 STG.E.64 desc[UR6][R2.64+0x8], R6 ; /* 0x0000080602008986 */ /* 0x0001e2000c101b06 */ /*0250*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fe20000010000 */ /*0260*/ @!P0 CS2R R6, SR_GLOBALTIMERLO ; /* 0x0000000000068805 */ /* 0x001fca0000015200 */ /*0270*/ @!P0 S2R R4, SR_VIRTUALSMID ; /* 0x0000000000048919 */ /* 0x000e220000004300 */ /*0280*/ @!P0 STG.E.64 desc[UR6][R2.64+0x10], R6 ; /* 0x0000100602008986 */ /* 0x0003e8000c101b06 */ /*0290*/ @!P0 STG.E.64 desc[UR6][R2.64+0x28], R4 ; /* 0x0000280402008986 */ /* 0x0013e8000c101b06 */ /*02a0*/ LDG.E.STRONG.SYS R11, desc[UR6][R10.64] ; /* 0x000000060a0b7981 */ /* 0x000362000c1f5900 */ /*02b0*/ IMAD.MOV.U32 R0, RZ, RZ, RZ ; /* 0x000000ffff007224 */ /* 0x000fc400078e00ff */ /*02c0*/ IMAD.MOV.U32 R8, RZ, RZ, 0x1 ; /* 0x00000001ff087424 */ /* 0x000fce00078e00ff */ /*02d0*/ VIADD R0, R0, 0x1 ; /* 0x0000000100007836 */ /* 0x000fe40000000000 */ /*02e0*/ IMAD R11, R11, 0x15a4e35, R8 ; /* 0x015a4e350b0b7824 */ /* 0x020fc600078e0208 */ /*02f0*/ ISETP.NE.AND P1, PT, R0, 0x200, PT ; /* 0x000002000000780c */ /* 0x000fda0003f25270 */ /*0300*/ @P1 BRA 0x2d0 ; /* 0xfffffffc00f01947 */ /* 0x000fea000383ffff */ /*0310*/ ULDC.64 UR4, c[0x0][0x220] ; /* 0x0000880000047ab9 */ /* 0x000fe40000000a00 */ /*0320*/ IADD3 R4, P1, R14, UR4, RZ ; /* 0x000000040e047c10 */ /* 0x002fc8000ff3e0ff */ /*0330*/ IADD3.X R5, R12, UR5, RZ, P1, !PT ; /* 0x000000050c057c10 */ /* 0x000fca0008ffe4ff */ /*0340*/ STG.E desc[UR6][R4.64], R11 ; /* 0x0000000b04007986 */ /* 0x0001e2000c101906 */ /*0350*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fec0000010000 */ /*0360*/ @P0 EXIT ; /* 0x000000000000094d */ /* 0x000fea0003800000 */ /*0370*/ CS2R R4, SR_GLOBALTIMERLO ; /* 0x0000000000047805 */ /* 0x001fca0000015200 */ /*0380*/ STG.E.64 desc[UR6][R2.64+0x18], R4 ; /* 0x0000180402007986 */ /* 0x000fe2000c101b06 */ /*0390*/ EXIT ; /* 0x000000000000794d */ /* 0x000fea0003800000 */ /*03a0*/ BRA 0x3a0; /* 0xfffffffc00fc7947 */ /* 0x000fc0000383ffff */ /*03b0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*03c0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*03d0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*03e0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*03f0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0400*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0410*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0420*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0430*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0440*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0450*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0460*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0470*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ .......... Function : _Z8ragged_bILb1EEvPVKjPjP5Stamp .headerflags @"EF_CUDA_TEXMODE_UNIFIED EF_CUDA_64BIT_ADDRESS EF_CUDA_SM90 EF_CUDA_VIRTUAL_SM(EF_CUDA_SM90)" /*0000*/ LDC R1, c[0x0][0x28] ; /* 0x00000a00ff017b82 */ /* 0x000e220000000800 */ /*0010*/ S2R R0, SR_TID.X ; /* 0x0000000000007919 */ /* 0x000e6e0000002100 */ /*0020*/ LDC.64 R2, c[0x0][0x220] ; /* 0x00008800ff027b82 */ /* 0x000ea20000000a00 */ /*0030*/ ULDC.64 UR4, c[0x0][0x208] ; /* 0x0000820000047ab9 */ /* 0x000fe20000000a00 */ /*0040*/ S2R R13, SR_CTAID.X ; /* 0x00000000000d7919 */ /* 0x000ea20000002500 */ /*0050*/ ISETP.NE.AND P0, PT, R0, RZ, PT ; /* 0x000000ff0000720c */ /* 0x002fe20003f05270 */ /*0060*/ IMAD.WIDE R2, R13, 0x30, R2 ; /* 0x000000300d027825 */ /* 0x004fd800078e0202 */ /*0070*/ @!P0 CS2R R6, SR_GLOBALTIMERLO ; /* 0x0000000000068805 */ /* 0x000fe20000015200 */ /*0080*/ @!P0 S2R R12, SR_VIRTUALSMID ; /* 0x00000000000c8919 */ /* 0x000e620000004300 */ /*0090*/ LDC.64 R4, c[0x0][0x210] ; /* 0x00008400ff047b82 */ /* 0x000ea20000000a00 */ /*00a0*/ IMAD R9, R13, 0x100, R0 ; /* 0x000001000d097824 */ /* 0x000fe400078e0200 */ /*00b0*/ @!P0 STG.E.64 desc[UR4][R2.64+0x10], R6 ; /* 0x0000100602008986 */ /* 0x0007e4000c101b04 */ /*00c0*/ IMAD.WIDE R4, R9, 0x4, R4 ; /* 0x0000000409047825 */ /* 0x004fe400078e0204 */ /*00d0*/ @!P0 STG.E.64 desc[UR4][R2.64+0x28], R12 ; /* 0x0000280c02008986 */ /* 0x0027e8000c101b04 */ /*00e0*/ LDG.E.STRONG.SYS R11, desc[UR4][R4.64] ; /* 0x00000004040b7981 */ /* 0x000762000c1f5900 */ /*00f0*/ IMAD.MOV.U32 R0, RZ, RZ, RZ ; /* 0x000000ffff007224 */ /* 0x000fe200078e00ff */ /*0100*/ SHF.R.S32.HI R8, RZ, 0x1f, R9 ; /* 0x0000001fff087819 */ /* 0x000fe20000011409 */ /*0110*/ IMAD.MOV.U32 R10, RZ, RZ, 0x1 ; /* 0x00000001ff0a7424 */ /* 0x001fce00078e00ff */ /*0120*/ VIADD R0, R0, 0x1 ; /* 0x0000000100007836 */ /* 0x000fe40000000000 */ /*0130*/ IMAD R11, R11, 0x15a4e35, R10 ; /* 0x015a4e350b0b7824 */ /* 0x020fc600078e020a */ /*0140*/ ISETP.NE.AND P1, PT, R0, 0x200, PT ; /* 0x000002000000780c */ /* 0x000fda0003f25270 */ /*0150*/ @P1 BRA 0x120 ; /* 0xfffffffc00f01947 */ /* 0x000fea000383ffff */ /*0160*/ ULDC.64 UR6, c[0x0][0x218] ; /* 0x0000860000067ab9 */ /* 0x000fe40000000a00 */ /*0170*/ LEA R4, P1, R9, UR6, 0x2 ; /* 0x0000000609047c11 */ /* 0x008fc8000f8210ff */ /*0180*/ LEA.HI.X R5, R9, UR7, R8, 0x2, P1 ; /* 0x0000000709057c11 */ /* 0x000fca00088f1408 */ /*0190*/ STG.E desc[UR4][R4.64], R11 ; /* 0x0000000b04007986 */ /* 0x0001e2000c101904 */ /*01a0*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fec0000010000 */ /*01b0*/ @P0 EXIT ; /* 0x000000000000094d */ /* 0x000fea0003800000 */ /*01c0*/ CS2R R4, SR_GLOBALTIMERLO ; /* 0x0000000000047805 */ /* 0x001fca0000015200 */ /*01d0*/ STG.E.64 desc[UR4][R2.64+0x18], R4 ; /* 0x0000180402007986 */ /* 0x000fe2000c101b04 */ /*01e0*/ EXIT ; /* 0x000000000000794d */ /* 0x000fea0003800000 */ /*01f0*/ BRA 0x1f0; /* 0xfffffffc00fc7947 */ /* 0x000fc0000383ffff */ /*0200*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0210*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0220*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0230*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0240*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0250*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0260*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0270*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ .......... Function : _Z8ragged_aILb1EEvPKjPVjPKiP5Stamp .headerflags @"EF_CUDA_TEXMODE_UNIFIED EF_CUDA_64BIT_ADDRESS EF_CUDA_SM90 EF_CUDA_VIRTUAL_SM(EF_CUDA_SM90)" /*0000*/ LDC R1, c[0x0][0x28] ; /* 0x00000a00ff017b82 */ /* 0x000e220000000800 */ /*0010*/ S2R R0, SR_TID.X ; /* 0x0000000000007919 */ /* 0x000e6e0000002100 */ /*0020*/ LDC.64 R2, c[0x0][0x228] ; /* 0x00008a00ff027b82 */ /* 0x000ea20000000a00 */ /*0030*/ ULDC.64 UR6, c[0x0][0x208] ; /* 0x0000820000067ab9 */ /* 0x000fe20000000a00 */ /*0040*/ S2R R13, SR_CTAID.X ; /* 0x00000000000d7919 */ /* 0x000ea20000002500 */ /*0050*/ ISETP.NE.AND P0, PT, R0, RZ, PT ; /* 0x000000ff0000720c */ /* 0x002fe20003f05270 */ /*0060*/ IMAD.WIDE R2, R13, 0x30, R2 ; /* 0x000000300d027825 */ /* 0x004fd800078e0202 */ /*0070*/ @!P0 CS2R R8, SR_GLOBALTIMERLO ; /* 0x0000000000088805 */ /* 0x000fe20000015200 */ /*0080*/ @!P0 S2R R12, SR_VIRTUALSMID ; /* 0x00000000000c8919 */ /* 0x000e620000004300 */ /*0090*/ LDC.64 R6, c[0x0][0x220] ; /* 0x00008800ff067b82 */ /* 0x000ea60000000a00 */ /*00a0*/ @!P0 STG.E.64 desc[UR6][R2.64], R8 ; /* 0x0000000802008986 */ /* 0x0007ea000c101b06 */ /*00b0*/ LDC.64 R4, c[0x0][0x210] ; /* 0x00008400ff047b82 */ /* 0x000f220000000a00 */ /*00c0*/ IMAD R11, R13.reuse, 0x100, R0 ; /* 0x000001000d0b7824 */ /* 0x040fe200078e0200 */ /*00d0*/ ULDC.64 UR4, c[0x0][0x218] ; /* 0x0000860000047ab9 */ /* 0x000fe20000000a00 */ /*00e0*/ IMAD.WIDE R6, R13, 0x4, R6 ; /* 0x000000040d067825 */ /* 0x004fe200078e0206 */ /*00f0*/ @!P0 STG.E.64 desc[UR6][R2.64+0x20], R12 ; /* 0x0000200c02008986 */ /* 0x0027ea000c101b06 */ /*0100*/ LDG.E R6, desc[UR6][R6.64] ; /* 0x0000000606067981 */ /* 0x000ea2000c1e1900 */ /*0110*/ IMAD.WIDE R4, R11, 0x4, R4 ; /* 0x000000040b047825 */ /* 0x010fcc00078e0204 */ /*0120*/ LDG.E R5, desc[UR6][R4.64] ; /* 0x0000000604057981 */ /* 0x000762000c1e1900 */ /*0130*/ SHF.R.S32.HI R14, RZ, 0x1f, R11 ; /* 0x0000001fff0e7819 */ /* 0x000fe4000001140b */ /*0140*/ LEA R10, P1, R11, UR4, 0x2 ; /* 0x000000040b0a7c11 */ /* 0x000fc8000f8210ff */ /*0150*/ LEA.HI.X R11, R11, UR5, R14, 0x2, P1 ; /* 0x000000050b0b7c11 */ /* 0x000fe400088f140e */ /*0160*/ ISETP.GE.AND P0, PT, R6, 0x1, PT ; /* 0x000000010600780c */ /* 0x004fda0003f06270 */ /*0170*/ @!P0 BRA 0x1e0 ; /* 0x0000000000188947 */ /* 0x009fea0003800000 */ /*0180*/ IMAD.MOV.U32 R4, RZ, RZ, 0x19660d ; /* 0x0019660dff047424 */ /* 0x000fe200078e00ff */ /*0190*/ UMOV UR4, URZ ; /* 0x0000003f00047c82 */ /* 0x000fcc0008000000 */ /*01a0*/ UIADD3 UR4, UR4, 0x1, URZ ; /* 0x0000000104047890 */ /* 0x000fe2000fffe03f */ /*01b0*/ IMAD R5, R5, R4, 0x3c6ef35f ; /* 0x3c6ef35f05057424 */ /* 0x020fca00078e0204 */ /*01c0*/ ISETP.LE.AND P0, PT, R6, UR4, PT ; /* 0x0000000406007c0c */ /* 0x000fda000bf03270 */ /*01d0*/ @!P0 BRA 0x1a0 ; /* 0xfffffffc00f08947 */ /* 0x000fea000383ffff */ /*01e0*/ STG.E.STRONG.SYS desc[UR6][R10.64], R5 ; /* 0x000000050a007986 */ /* 0x0201e2000c115906 */ /*01f0*/ BAR.SYNC.DEFER_BLOCKING 0x0 ; /* 0x0000000000007b1d */ /* 0x000fe20000010000 */ /*0200*/ ISETP.NE.AND P0, PT, R0, RZ, PT ; /* 0x000000ff0000720c */ /* 0x000fda0003f05270 */ /*0210*/ @P0 EXIT ; /* 0x000000000000094d */ /* 0x001fea0003800000 */ /*0220*/ CS2R R4, SR_GLOBALTIMERLO ; /* 0x0000000000047805 */ /* 0x000fca0000015200 */ /*0230*/ STG.E.64 desc[UR6][R2.64+0x8], R4 ; /* 0x0000080402007986 */ /* 0x000fe2000c101b06 */ /*0240*/ EXIT ; /* 0x000000000000794d */ /* 0x000fea0003800000 */ /*0250*/ BRA 0x250; /* 0xfffffffc00fc7947 */ /* 0x000fc0000383ffff */ /*0260*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0270*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0280*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0290*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*02a0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*02b0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*02c0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*02d0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*02e0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*02f0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ .......... Function : _Z15fused_pointwisePKfPfi .headerflags @"EF_CUDA_TEXMODE_UNIFIED EF_CUDA_64BIT_ADDRESS EF_CUDA_SM90 EF_CUDA_VIRTUAL_SM(EF_CUDA_SM90)" /*0000*/ LDC R1, c[0x0][0x28] ; /* 0x00000a00ff017b82 */ /* 0x000e220000000800 */ /*0010*/ S2R R0, SR_TID.X ; /* 0x0000000000007919 */ /* 0x000e6e0000002100 */ /*0020*/ S2UR UR4, SR_CTAID.X ; /* 0x00000000000479c3 */ /* 0x000e700000002500 */ /*0030*/ LDC R7, c[0x0][RZ] ; /* 0x00000000ff077b82 */ /* 0x000e640000000800 */ /*0040*/ IMAD R7, R7, UR4, R0 ; /* 0x0000000407077c24 */ /* 0x002fe2000f8e0200 */ /*0050*/ ULDC UR4, c[0x0][0x220] ; /* 0x0000880000047ab9 */ /* 0x000fc80000000800 */ /*0060*/ ISETP.GE.AND P0, PT, R7, UR4, PT ; /* 0x0000000407007c0c */ /* 0x000fda000bf06270 */ /*0070*/ @P0 EXIT ; /* 0x000000000000094d */ /* 0x001fea0003800000 */ /*0080*/ LDC.64 R2, c[0x0][0x210] ; /* 0x00008400ff027b82 */ /* 0x000e220000000a00 */ /*0090*/ ULDC.64 UR4, c[0x0][0x208] ; /* 0x0000820000047ab9 */ /* 0x000fce0000000a00 */ /*00a0*/ LDC.64 R4, c[0x0][0x218] ; /* 0x00008600ff047b82 */ /* 0x000e620000000a00 */ /*00b0*/ IMAD.WIDE R2, R7, 0x4, R2 ; /* 0x0000000407027825 */ /* 0x001fcc00078e0202 */ /*00c0*/ LDG.E R2, desc[UR4][R2.64] ; /* 0x0000000402027981 */ /* 0x000ea2000c1e1900 */ /*00d0*/ IMAD.WIDE R4, R7, 0x4, R4 ; /* 0x0000000407047825 */ /* 0x002fc800078e0204 */ /*00e0*/ FADD R0, R2, 1 ; /* 0x3f80000002007421 */ /* 0x004fc80000000000 */ /*00f0*/ FMUL R0, R0, R0 ; /* 0x0000000000007220 */ /* 0x000fc80000400000 */ /*0100*/ FMUL R7, R0, 0.5 ; /* 0x3f00000000077820 */ /* 0x000fca0000400000 */ /*0110*/ STG.E desc[UR4][R4.64], R7 ; /* 0x0000000704007986 */ /* 0x000fe2000c101904 */ /*0120*/ EXIT ; /* 0x000000000000794d */ /* 0x000fea0003800000 */ /*0130*/ BRA 0x130; /* 0xfffffffc00fc7947 */ /* 0x000fc0000383ffff */ /*0140*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0150*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0160*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0170*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0180*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0190*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01a0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01b0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01c0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01d0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01e0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01f0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ .......... Function : _Z10half_valuePKfPfi .headerflags @"EF_CUDA_TEXMODE_UNIFIED EF_CUDA_64BIT_ADDRESS EF_CUDA_SM90 EF_CUDA_VIRTUAL_SM(EF_CUDA_SM90)" /*0000*/ LDC R1, c[0x0][0x28] ; /* 0x00000a00ff017b82 */ /* 0x000e220000000800 */ /*0010*/ S2R R0, SR_TID.X ; /* 0x0000000000007919 */ /* 0x000e6e0000002100 */ /*0020*/ S2UR UR4, SR_CTAID.X ; /* 0x00000000000479c3 */ /* 0x000e700000002500 */ /*0030*/ LDC R7, c[0x0][RZ] ; /* 0x00000000ff077b82 */ /* 0x000e640000000800 */ /*0040*/ IMAD R7, R7, UR4, R0 ; /* 0x0000000407077c24 */ /* 0x002fe2000f8e0200 */ /*0050*/ ULDC UR4, c[0x0][0x220] ; /* 0x0000880000047ab9 */ /* 0x000fc80000000800 */ /*0060*/ ISETP.GE.AND P0, PT, R7, UR4, PT ; /* 0x0000000407007c0c */ /* 0x000fda000bf06270 */ /*0070*/ @P0 EXIT ; /* 0x000000000000094d */ /* 0x001fea0003800000 */ /*0080*/ LDC.64 R2, c[0x0][0x210] ; /* 0x00008400ff027b82 */ /* 0x000e220000000a00 */ /*0090*/ ULDC.64 UR4, c[0x0][0x208] ; /* 0x0000820000047ab9 */ /* 0x000fce0000000a00 */ /*00a0*/ LDC.64 R4, c[0x0][0x218] ; /* 0x00008600ff047b82 */ /* 0x000e620000000a00 */ /*00b0*/ IMAD.WIDE R2, R7, 0x4, R2 ; /* 0x0000000407027825 */ /* 0x001fcc00078e0202 */ /*00c0*/ LDG.E R2, desc[UR4][R2.64] ; /* 0x0000000402027981 */ /* 0x000ea2000c1e1900 */ /*00d0*/ IMAD.WIDE R4, R7, 0x4, R4 ; /* 0x0000000407047825 */ /* 0x002fc800078e0204 */ /*00e0*/ FMUL R7, R2, 0.5 ; /* 0x3f00000002077820 */ /* 0x004fca0000400000 */ /*00f0*/ STG.E desc[UR4][R4.64], R7 ; /* 0x0000000704007986 */ /* 0x000fe2000c101904 */ /*0100*/ EXIT ; /* 0x000000000000794d */ /* 0x000fea0003800000 */ /*0110*/ BRA 0x110; /* 0xfffffffc00fc7947 */ /* 0x000fc0000383ffff */ /*0120*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0130*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0140*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0150*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0160*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0170*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0180*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0190*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01a0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01b0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01c0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01d0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01e0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01f0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ .......... Function : _Z6squarePKfPfi .headerflags @"EF_CUDA_TEXMODE_UNIFIED EF_CUDA_64BIT_ADDRESS EF_CUDA_SM90 EF_CUDA_VIRTUAL_SM(EF_CUDA_SM90)" /*0000*/ LDC R1, c[0x0][0x28] ; /* 0x00000a00ff017b82 */ /* 0x000e220000000800 */ /*0010*/ S2R R0, SR_TID.X ; /* 0x0000000000007919 */ /* 0x000e6e0000002100 */ /*0020*/ S2UR UR4, SR_CTAID.X ; /* 0x00000000000479c3 */ /* 0x000e700000002500 */ /*0030*/ LDC R7, c[0x0][RZ] ; /* 0x00000000ff077b82 */ /* 0x000e640000000800 */ /*0040*/ IMAD R7, R7, UR4, R0 ; /* 0x0000000407077c24 */ /* 0x002fe2000f8e0200 */ /*0050*/ ULDC UR4, c[0x0][0x220] ; /* 0x0000880000047ab9 */ /* 0x000fc80000000800 */ /*0060*/ ISETP.GE.AND P0, PT, R7, UR4, PT ; /* 0x0000000407007c0c */ /* 0x000fda000bf06270 */ /*0070*/ @P0 EXIT ; /* 0x000000000000094d */ /* 0x001fea0003800000 */ /*0080*/ LDC.64 R2, c[0x0][0x210] ; /* 0x00008400ff027b82 */ /* 0x000e220000000a00 */ /*0090*/ ULDC.64 UR4, c[0x0][0x208] ; /* 0x0000820000047ab9 */ /* 0x000fce0000000a00 */ /*00a0*/ LDC.64 R4, c[0x0][0x218] ; /* 0x00008600ff047b82 */ /* 0x000e620000000a00 */ /*00b0*/ IMAD.WIDE R2, R7, 0x4, R2 ; /* 0x0000000407027825 */ /* 0x001fcc00078e0202 */ /*00c0*/ LDG.E R2, desc[UR4][R2.64] ; /* 0x0000000402027981 */ /* 0x000ea2000c1e1900 */ /*00d0*/ IMAD.WIDE R4, R7, 0x4, R4 ; /* 0x0000000407047825 */ /* 0x002fc800078e0204 */ /*00e0*/ FMUL R7, R2, R2 ; /* 0x0000000202077220 */ /* 0x004fca0000400000 */ /*00f0*/ STG.E desc[UR4][R4.64], R7 ; /* 0x0000000704007986 */ /* 0x000fe2000c101904 */ /*0100*/ EXIT ; /* 0x000000000000794d */ /* 0x000fea0003800000 */ /*0110*/ BRA 0x110; /* 0xfffffffc00fc7947 */ /* 0x000fc0000383ffff */ /*0120*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0130*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0140*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0150*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0160*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0170*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0180*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0190*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01a0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01b0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01c0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01d0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01e0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01f0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ .......... Function : _Z7add_onePKfPfi .headerflags @"EF_CUDA_TEXMODE_UNIFIED EF_CUDA_64BIT_ADDRESS EF_CUDA_SM90 EF_CUDA_VIRTUAL_SM(EF_CUDA_SM90)" /*0000*/ LDC R1, c[0x0][0x28] ; /* 0x00000a00ff017b82 */ /* 0x000e220000000800 */ /*0010*/ S2R R0, SR_TID.X ; /* 0x0000000000007919 */ /* 0x000e6e0000002100 */ /*0020*/ S2UR UR4, SR_CTAID.X ; /* 0x00000000000479c3 */ /* 0x000e700000002500 */ /*0030*/ LDC R7, c[0x0][RZ] ; /* 0x00000000ff077b82 */ /* 0x000e640000000800 */ /*0040*/ IMAD R7, R7, UR4, R0 ; /* 0x0000000407077c24 */ /* 0x002fe2000f8e0200 */ /*0050*/ ULDC UR4, c[0x0][0x220] ; /* 0x0000880000047ab9 */ /* 0x000fc80000000800 */ /*0060*/ ISETP.GE.AND P0, PT, R7, UR4, PT ; /* 0x0000000407007c0c */ /* 0x000fda000bf06270 */ /*0070*/ @P0 EXIT ; /* 0x000000000000094d */ /* 0x001fea0003800000 */ /*0080*/ LDC.64 R2, c[0x0][0x210] ; /* 0x00008400ff027b82 */ /* 0x000e220000000a00 */ /*0090*/ ULDC.64 UR4, c[0x0][0x208] ; /* 0x0000820000047ab9 */ /* 0x000fce0000000a00 */ /*00a0*/ LDC.64 R4, c[0x0][0x218] ; /* 0x00008600ff047b82 */ /* 0x000e620000000a00 */ /*00b0*/ IMAD.WIDE R2, R7, 0x4, R2 ; /* 0x0000000407027825 */ /* 0x001fcc00078e0202 */ /*00c0*/ LDG.E R2, desc[UR4][R2.64] ; /* 0x0000000402027981 */ /* 0x000ea2000c1e1900 */ /*00d0*/ IMAD.WIDE R4, R7, 0x4, R4 ; /* 0x0000000407047825 */ /* 0x002fc800078e0204 */ /*00e0*/ FADD R7, R2, 1 ; /* 0x3f80000002077421 */ /* 0x004fca0000000000 */ /*00f0*/ STG.E desc[UR4][R4.64], R7 ; /* 0x0000000704007986 */ /* 0x000fe2000c101904 */ /*0100*/ EXIT ; /* 0x000000000000794d */ /* 0x000fea0003800000 */ /*0110*/ BRA 0x110; /* 0xfffffffc00fc7947 */ /* 0x000fc0000383ffff */ /*0120*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0130*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0140*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0150*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0160*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0170*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0180*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*0190*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01a0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01b0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01c0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01d0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01e0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ /*01f0*/ NOP; /* 0x0000000000007918 */ /* 0x000fc00000000000 */ .......... Fatbin ptx code: ================ arch = sm_90 code version = [8,7] host = linux compile_size = 64bit compressed identifier = /workspace/bench.cu ptxasOptions = -v --fmad false --generate-line-info