mirror of
https://github.com/brannondorsey/naive-hashcat.git
synced 2025-08-20 09:18:24 +02:00
8422 lines
206 KiB
Plaintext
8422 lines
206 KiB
Plaintext
//
|
||
// Generated by NVIDIA NVVM Compiler
|
||
//
|
||
// Compiler Build ID: CL-22053397
|
||
// Driver 375.66
|
||
// Based on LLVM 3.4svn
|
||
//
|
||
|
||
.version 5.0
|
||
.target sm_61, texmode_independent
|
||
.address_size 64
|
||
|
||
// .globl gpu_memset
|
||
|
||
.entry gpu_memset(
|
||
.param .u64 .ptr .global .align 16 gpu_memset_param_0,
|
||
.param .u32 gpu_memset_param_1,
|
||
.param .u32 gpu_memset_param_2
|
||
)
|
||
{
|
||
.reg .pred %p<2>;
|
||
.reg .b32 %r<9>;
|
||
.reg .b64 %rd<4>;
|
||
|
||
|
||
ld.param.u64 %rd1, [gpu_memset_param_0];
|
||
ld.param.u32 %r2, [gpu_memset_param_1];
|
||
ld.param.u32 %r3, [gpu_memset_param_2];
|
||
mov.b32 %r4, %envreg3;
|
||
mov.u32 %r5, %ctaid.x;
|
||
mov.u32 %r6, %ntid.x;
|
||
mad.lo.s32 %r7, %r5, %r6, %r4;
|
||
mov.u32 %r8, %tid.x;
|
||
add.s32 %r1, %r7, %r8;
|
||
setp.ge.u32 %p1, %r1, %r3;
|
||
@%p1 bra BB0_2;
|
||
|
||
mul.wide.u32 %rd2, %r1, 16;
|
||
add.s64 %rd3, %rd1, %rd2;
|
||
st.global.v4.u32 [%rd3], {%r2, %r2, %r2, %r2};
|
||
|
||
BB0_2:
|
||
ret;
|
||
}
|
||
|
||
// .globl m00500_init
|
||
.entry m00500_init(
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_0,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_1,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_2,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_3,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_4,
|
||
.param .u64 .ptr .global .align 1 m00500_init_param_5,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_6,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_7,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_8,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_9,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_10,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_11,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_12,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_13,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_14,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_15,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_16,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_17,
|
||
.param .u64 .ptr .global .align 1 m00500_init_param_18,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_19,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_20,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_21,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_22,
|
||
.param .u64 .ptr .global .align 4 m00500_init_param_23,
|
||
.param .u32 m00500_init_param_24,
|
||
.param .u32 m00500_init_param_25,
|
||
.param .u32 m00500_init_param_26,
|
||
.param .u32 m00500_init_param_27,
|
||
.param .u32 m00500_init_param_28,
|
||
.param .u32 m00500_init_param_29,
|
||
.param .u32 m00500_init_param_30,
|
||
.param .u32 m00500_init_param_31,
|
||
.param .u32 m00500_init_param_32,
|
||
.param .u32 m00500_init_param_33,
|
||
.param .u32 m00500_init_param_34
|
||
)
|
||
{
|
||
.reg .pred %p<214>;
|
||
.reg .b32 %r<3588>;
|
||
.reg .b64 %rd<13>;
|
||
|
||
|
||
ld.param.u64 %rd2, [m00500_init_param_0];
|
||
ld.param.u64 %rd4, [m00500_init_param_17];
|
||
ld.param.u32 %r245, [m00500_init_param_27];
|
||
ld.param.u32 %r246, [m00500_init_param_34];
|
||
mov.b32 %r247, %envreg3;
|
||
mov.u32 %r248, %ctaid.x;
|
||
mov.u32 %r249, %ntid.x;
|
||
mad.lo.s32 %r250, %r248, %r249, %r247;
|
||
mov.u32 %r251, %tid.x;
|
||
add.s32 %r1, %r250, %r251;
|
||
setp.ge.u32 %p1, %r1, %r246;
|
||
@%p1 bra BB1_312;
|
||
|
||
mul.wide.u32 %rd5, %r1, 80;
|
||
add.s64 %rd6, %rd2, %rd5;
|
||
ld.global.u32 %r2, [%rd6];
|
||
ld.global.u32 %r3, [%rd6+4];
|
||
ld.global.u32 %r4, [%rd6+8];
|
||
ld.global.u32 %r5, [%rd6+12];
|
||
mul.wide.u32 %rd7, %r245, 180;
|
||
add.s64 %rd8, %rd4, %rd7;
|
||
ld.global.u32 %r6, [%rd8];
|
||
ld.global.u32 %r7, [%rd8+4];
|
||
ld.global.u32 %r8, [%rd8+128];
|
||
ld.global.u32 %r1649, [%rd6+64];
|
||
and.b32 %r275, %r1649, 3;
|
||
mov.u32 %r276, 4;
|
||
sub.s32 %r277, %r276, %r275;
|
||
shl.b32 %r278, %r277, 2;
|
||
mov.u32 %r279, 1985229328;
|
||
shr.u32 %r280, %r279, %r278;
|
||
and.b32 %r263, %r280, 65535;
|
||
mov.u32 %r273, 0;
|
||
// inline asm
|
||
prmt.b32 %r252, %r273, %r6, %r263;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r256, %r6, %r7, %r263;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r260, %r7, %r273, %r263;
|
||
// inline asm
|
||
shr.u32 %r274, %r1649, 2;
|
||
mov.u32 %r272, %r273;
|
||
mov.u32 %r3577, %r273;
|
||
mov.u32 %r3578, %r273;
|
||
mov.u32 %r3579, %r273;
|
||
mov.u32 %r268, %r273;
|
||
mov.u32 %r3581, %r273;
|
||
mov.u32 %r3582, %r273;
|
||
mov.u32 %r3583, %r273;
|
||
mov.u32 %r3584, %r273;
|
||
setp.gt.s32 %p2, %r274, 5;
|
||
@%p2 bra BB1_11;
|
||
|
||
setp.gt.s32 %p10, %r274, 2;
|
||
@%p10 bra BB1_7;
|
||
|
||
setp.eq.s32 %p14, %r274, 0;
|
||
@%p14 bra BB1_27;
|
||
|
||
setp.eq.s32 %p15, %r274, 1;
|
||
@%p15 bra BB1_26;
|
||
bra.uni BB1_5;
|
||
|
||
BB1_26:
|
||
or.b32 %r15, %r252, %r3;
|
||
mov.u32 %r363, 0;
|
||
mov.u32 %r362, %r363;
|
||
mov.u32 %r3577, %r363;
|
||
mov.u32 %r3578, %r363;
|
||
mov.u32 %r3579, %r363;
|
||
mov.u32 %r3580, %r363;
|
||
mov.u32 %r3581, %r363;
|
||
mov.u32 %r3582, %r363;
|
||
mov.u32 %r3583, %r363;
|
||
mov.u32 %r3584, %r363;
|
||
mov.u32 %r1689, %r256;
|
||
mov.u32 %r1700, %r260;
|
||
mov.u32 %r1807, %r2;
|
||
mov.u32 %r3505, %r1807;
|
||
mov.u32 %r3575, %r363;
|
||
mov.u32 %r3576, %r362;
|
||
mov.u32 %r3585, %r15;
|
||
mov.u32 %r3586, %r1689;
|
||
mov.u32 %r3587, %r1700;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_11:
|
||
setp.gt.s32 %p3, %r274, 8;
|
||
@%p3 bra BB1_16;
|
||
|
||
setp.eq.s32 %p7, %r274, 6;
|
||
@%p7 bra BB1_23;
|
||
|
||
setp.eq.s32 %p8, %r274, 7;
|
||
@%p8 bra BB1_22;
|
||
bra.uni BB1_14;
|
||
|
||
BB1_22:
|
||
mov.u32 %r315, 0;
|
||
mov.u32 %r314, %r315;
|
||
mov.u32 %r3579, %r315;
|
||
mov.u32 %r312, %r315;
|
||
mov.u32 %r3581, %r315;
|
||
mov.u32 %r3582, %r315;
|
||
mov.u32 %r3583, %r315;
|
||
mov.u32 %r1658, %r5;
|
||
mov.u32 %r3577, %r256;
|
||
mov.u32 %r3578, %r260;
|
||
mov.u32 %r1719, %r4;
|
||
mov.u32 %r1759, %r3;
|
||
mov.u32 %r1801, %r2;
|
||
mov.u32 %r3584, %r252;
|
||
mov.u32 %r3505, %r1801;
|
||
mov.u32 %r3575, %r315;
|
||
mov.u32 %r3576, %r314;
|
||
mov.u32 %r3580, %r312;
|
||
mov.u32 %r3585, %r1759;
|
||
mov.u32 %r3586, %r1719;
|
||
mov.u32 %r3587, %r1658;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_7:
|
||
setp.eq.s32 %p11, %r274, 3;
|
||
@%p11 bra BB1_25;
|
||
|
||
setp.eq.s32 %p12, %r274, 4;
|
||
@%p12 bra BB1_24;
|
||
bra.uni BB1_9;
|
||
|
||
BB1_24:
|
||
mov.u32 %r336, 0;
|
||
mov.u32 %r335, %r336;
|
||
mov.u32 %r3577, %r336;
|
||
mov.u32 %r3578, %r336;
|
||
mov.u32 %r3579, %r336;
|
||
mov.u32 %r331, %r336;
|
||
mov.u32 %r3584, %r336;
|
||
mov.u32 %r1661, %r5;
|
||
mov.u32 %r3582, %r256;
|
||
mov.u32 %r3583, %r260;
|
||
mov.u32 %r1722, %r4;
|
||
mov.u32 %r1762, %r3;
|
||
mov.u32 %r1804, %r2;
|
||
mov.u32 %r3581, %r252;
|
||
mov.u32 %r3505, %r1804;
|
||
mov.u32 %r3575, %r336;
|
||
mov.u32 %r3576, %r335;
|
||
mov.u32 %r3580, %r331;
|
||
mov.u32 %r3585, %r1762;
|
||
mov.u32 %r3586, %r1722;
|
||
mov.u32 %r3587, %r1661;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_16:
|
||
setp.eq.s32 %p4, %r274, 9;
|
||
@%p4 bra BB1_21;
|
||
|
||
setp.eq.s32 %p5, %r274, 10;
|
||
@%p5 bra BB1_20;
|
||
bra.uni BB1_18;
|
||
|
||
BB1_20:
|
||
mov.u32 %r294, 0;
|
||
mov.u32 %r3577, %r294;
|
||
mov.u32 %r3578, %r294;
|
||
mov.u32 %r3581, %r294;
|
||
mov.u32 %r3582, %r294;
|
||
mov.u32 %r3583, %r294;
|
||
mov.u32 %r3584, %r294;
|
||
mov.u32 %r1655, %r5;
|
||
mov.u32 %r1680, %r256;
|
||
mov.u32 %r1691, %r260;
|
||
mov.u32 %r1716, %r4;
|
||
mov.u32 %r1756, %r3;
|
||
mov.u32 %r1798, %r2;
|
||
mov.u32 %r3579, %r252;
|
||
mov.u32 %r3505, %r1798;
|
||
mov.u32 %r3575, %r1691;
|
||
mov.u32 %r3576, %r294;
|
||
mov.u32 %r3580, %r1680;
|
||
mov.u32 %r3585, %r1756;
|
||
mov.u32 %r3586, %r1716;
|
||
mov.u32 %r3587, %r1655;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_27:
|
||
or.b32 %r16, %r252, %r2;
|
||
mov.u32 %r373, 0;
|
||
mov.u32 %r372, %r373;
|
||
mov.u32 %r3577, %r373;
|
||
mov.u32 %r3578, %r373;
|
||
mov.u32 %r3579, %r373;
|
||
mov.u32 %r3580, %r373;
|
||
mov.u32 %r3581, %r373;
|
||
mov.u32 %r3582, %r373;
|
||
mov.u32 %r3583, %r373;
|
||
mov.u32 %r3584, %r373;
|
||
mov.u32 %r1662, %r5;
|
||
mov.u32 %r1690, %r256;
|
||
mov.u32 %r1701, %r260;
|
||
mov.u32 %r3505, %r16;
|
||
mov.u32 %r3575, %r373;
|
||
mov.u32 %r3576, %r372;
|
||
mov.u32 %r3585, %r1690;
|
||
mov.u32 %r3586, %r1701;
|
||
mov.u32 %r3587, %r1662;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_5:
|
||
setp.eq.s32 %p16, %r274, 2;
|
||
mov.u32 %r1653, %r5;
|
||
mov.u32 %r1714, %r4;
|
||
mov.u32 %r1754, %r3;
|
||
mov.u32 %r1796, %r2;
|
||
mov.u32 %r1898, %r268;
|
||
mov.u32 %r1949, %r272;
|
||
mov.u32 %r1965, %r273;
|
||
mov.u32 %r3505, %r1796;
|
||
mov.u32 %r3575, %r1965;
|
||
mov.u32 %r3576, %r1949;
|
||
mov.u32 %r3580, %r1898;
|
||
mov.u32 %r3585, %r1754;
|
||
mov.u32 %r3586, %r1714;
|
||
mov.u32 %r3587, %r1653;
|
||
@%p16 bra BB1_6;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_6:
|
||
or.b32 %r14, %r252, %r4;
|
||
mov.u32 %r353, 0;
|
||
mov.u32 %r352, %r353;
|
||
mov.u32 %r3577, %r353;
|
||
mov.u32 %r3578, %r353;
|
||
mov.u32 %r3579, %r353;
|
||
mov.u32 %r3580, %r353;
|
||
mov.u32 %r3582, %r353;
|
||
mov.u32 %r3583, %r353;
|
||
mov.u32 %r3584, %r353;
|
||
mov.u32 %r1688, %r256;
|
||
mov.u32 %r3581, %r260;
|
||
mov.u32 %r1764, %r3;
|
||
mov.u32 %r1806, %r2;
|
||
mov.u32 %r3505, %r1806;
|
||
mov.u32 %r3575, %r353;
|
||
mov.u32 %r3576, %r352;
|
||
mov.u32 %r3585, %r1764;
|
||
mov.u32 %r3586, %r14;
|
||
mov.u32 %r3587, %r1688;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_23:
|
||
mov.u32 %r322, 0;
|
||
mov.u32 %r321, %r322;
|
||
mov.u32 %r3578, %r322;
|
||
mov.u32 %r3579, %r322;
|
||
mov.u32 %r318, %r322;
|
||
mov.u32 %r3581, %r322;
|
||
mov.u32 %r3582, %r322;
|
||
mov.u32 %r1659, %r5;
|
||
mov.u32 %r3584, %r256;
|
||
mov.u32 %r3577, %r260;
|
||
mov.u32 %r1720, %r4;
|
||
mov.u32 %r1760, %r3;
|
||
mov.u32 %r1802, %r2;
|
||
mov.u32 %r3583, %r252;
|
||
mov.u32 %r3505, %r1802;
|
||
mov.u32 %r3575, %r322;
|
||
mov.u32 %r3576, %r321;
|
||
mov.u32 %r3580, %r318;
|
||
mov.u32 %r3585, %r1760;
|
||
mov.u32 %r3586, %r1720;
|
||
mov.u32 %r3587, %r1659;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_14:
|
||
setp.eq.s32 %p9, %r274, 8;
|
||
mov.u32 %r1651, %r5;
|
||
mov.u32 %r1712, %r4;
|
||
mov.u32 %r1752, %r3;
|
||
mov.u32 %r1794, %r2;
|
||
mov.u32 %r1896, %r268;
|
||
mov.u32 %r1947, %r272;
|
||
mov.u32 %r1963, %r273;
|
||
mov.u32 %r3505, %r1794;
|
||
mov.u32 %r3575, %r1963;
|
||
mov.u32 %r3576, %r1947;
|
||
mov.u32 %r3580, %r1896;
|
||
mov.u32 %r3585, %r1752;
|
||
mov.u32 %r3586, %r1712;
|
||
mov.u32 %r3587, %r1651;
|
||
@%p9 bra BB1_15;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_15:
|
||
mov.u32 %r308, 0;
|
||
mov.u32 %r307, %r308;
|
||
mov.u32 %r306, %r308;
|
||
mov.u32 %r3581, %r308;
|
||
mov.u32 %r3582, %r308;
|
||
mov.u32 %r3583, %r308;
|
||
mov.u32 %r3584, %r308;
|
||
mov.u32 %r1657, %r5;
|
||
mov.u32 %r3578, %r256;
|
||
mov.u32 %r3579, %r260;
|
||
mov.u32 %r1718, %r4;
|
||
mov.u32 %r1758, %r3;
|
||
mov.u32 %r1800, %r2;
|
||
mov.u32 %r3577, %r252;
|
||
mov.u32 %r3505, %r1800;
|
||
mov.u32 %r3575, %r308;
|
||
mov.u32 %r3576, %r307;
|
||
mov.u32 %r3580, %r306;
|
||
mov.u32 %r3585, %r1758;
|
||
mov.u32 %r3586, %r1718;
|
||
mov.u32 %r3587, %r1657;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_25:
|
||
or.b32 %r13, %r252, %r5;
|
||
mov.u32 %r344, 0;
|
||
mov.u32 %r343, %r344;
|
||
mov.u32 %r3577, %r344;
|
||
mov.u32 %r3578, %r344;
|
||
mov.u32 %r3579, %r344;
|
||
mov.u32 %r3580, %r344;
|
||
mov.u32 %r3583, %r344;
|
||
mov.u32 %r3584, %r344;
|
||
mov.u32 %r3581, %r256;
|
||
mov.u32 %r3582, %r260;
|
||
mov.u32 %r1723, %r4;
|
||
mov.u32 %r1763, %r3;
|
||
mov.u32 %r1805, %r2;
|
||
mov.u32 %r3505, %r1805;
|
||
mov.u32 %r3575, %r344;
|
||
mov.u32 %r3576, %r343;
|
||
mov.u32 %r3585, %r1763;
|
||
mov.u32 %r3586, %r1723;
|
||
mov.u32 %r3587, %r13;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_9:
|
||
setp.eq.s32 %p13, %r274, 5;
|
||
mov.u32 %r1652, %r5;
|
||
mov.u32 %r1713, %r4;
|
||
mov.u32 %r1753, %r3;
|
||
mov.u32 %r1795, %r2;
|
||
mov.u32 %r1897, %r268;
|
||
mov.u32 %r1948, %r272;
|
||
mov.u32 %r1964, %r273;
|
||
mov.u32 %r3505, %r1795;
|
||
mov.u32 %r3575, %r1964;
|
||
mov.u32 %r3576, %r1948;
|
||
mov.u32 %r3580, %r1897;
|
||
mov.u32 %r3585, %r1753;
|
||
mov.u32 %r3586, %r1713;
|
||
mov.u32 %r3587, %r1652;
|
||
@%p13 bra BB1_10;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_10:
|
||
mov.u32 %r329, 0;
|
||
mov.u32 %r328, %r329;
|
||
mov.u32 %r3577, %r329;
|
||
mov.u32 %r3578, %r329;
|
||
mov.u32 %r3579, %r329;
|
||
mov.u32 %r324, %r329;
|
||
mov.u32 %r3581, %r329;
|
||
mov.u32 %r1660, %r5;
|
||
mov.u32 %r3583, %r256;
|
||
mov.u32 %r3584, %r260;
|
||
mov.u32 %r1721, %r4;
|
||
mov.u32 %r1761, %r3;
|
||
mov.u32 %r1803, %r2;
|
||
mov.u32 %r3582, %r252;
|
||
mov.u32 %r3505, %r1803;
|
||
mov.u32 %r3575, %r329;
|
||
mov.u32 %r3576, %r328;
|
||
mov.u32 %r3580, %r324;
|
||
mov.u32 %r3585, %r1761;
|
||
mov.u32 %r3586, %r1721;
|
||
mov.u32 %r3587, %r1660;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_21:
|
||
mov.u32 %r301, 0;
|
||
mov.u32 %r300, %r301;
|
||
mov.u32 %r3577, %r301;
|
||
mov.u32 %r3581, %r301;
|
||
mov.u32 %r3582, %r301;
|
||
mov.u32 %r3583, %r301;
|
||
mov.u32 %r3584, %r301;
|
||
mov.u32 %r1656, %r5;
|
||
mov.u32 %r3579, %r256;
|
||
mov.u32 %r1692, %r260;
|
||
mov.u32 %r1717, %r4;
|
||
mov.u32 %r1757, %r3;
|
||
mov.u32 %r1799, %r2;
|
||
mov.u32 %r3578, %r252;
|
||
mov.u32 %r3505, %r1799;
|
||
mov.u32 %r3575, %r301;
|
||
mov.u32 %r3576, %r300;
|
||
mov.u32 %r3580, %r1692;
|
||
mov.u32 %r3585, %r1757;
|
||
mov.u32 %r3586, %r1717;
|
||
mov.u32 %r3587, %r1656;
|
||
bra.uni BB1_28;
|
||
|
||
BB1_18:
|
||
setp.ne.s32 %p6, %r274, 11;
|
||
mov.u32 %r3505, %r2;
|
||
mov.u32 %r3575, %r273;
|
||
mov.u32 %r3576, %r272;
|
||
mov.u32 %r3580, %r268;
|
||
mov.u32 %r3585, %r3;
|
||
mov.u32 %r3586, %r4;
|
||
mov.u32 %r3587, %r5;
|
||
@%p6 bra BB1_28;
|
||
|
||
mov.u32 %r3577, 0;
|
||
mov.u32 %r3578, %r3577;
|
||
mov.u32 %r3579, %r3577;
|
||
mov.u32 %r3581, %r3577;
|
||
mov.u32 %r3582, %r3577;
|
||
mov.u32 %r3583, %r3577;
|
||
mov.u32 %r3584, %r3577;
|
||
mov.u32 %r1654, %r5;
|
||
mov.u32 %r1715, %r4;
|
||
mov.u32 %r1755, %r3;
|
||
mov.u32 %r1797, %r2;
|
||
mov.u32 %r3505, %r1797;
|
||
mov.u32 %r3575, %r256;
|
||
mov.u32 %r3576, %r260;
|
||
mov.u32 %r3580, %r252;
|
||
mov.u32 %r3585, %r1755;
|
||
mov.u32 %r3586, %r1715;
|
||
mov.u32 %r3587, %r1654;
|
||
|
||
BB1_28:
|
||
mov.u32 %r1702, %r3587;
|
||
mov.u32 %r1741, %r3586;
|
||
mov.u32 %r1782, %r3585;
|
||
mov.u32 %r1966, %r3575;
|
||
mov.u32 %r1828, %r3505;
|
||
mov.u32 %r30, %r1966;
|
||
mov.u32 %r29, %r3576;
|
||
mov.u32 %r28, %r3577;
|
||
mov.u32 %r27, %r3578;
|
||
mov.u32 %r26, %r3579;
|
||
mov.u32 %r25, %r3580;
|
||
mov.u32 %r24, %r3581;
|
||
mov.u32 %r23, %r3582;
|
||
mov.u32 %r22, %r3583;
|
||
mov.u32 %r21, %r3584;
|
||
mov.u32 %r20, %r1828;
|
||
mov.u32 %r19, %r1782;
|
||
mov.u32 %r18, %r1741;
|
||
mov.u32 %r17, %r1702;
|
||
add.s32 %r31, %r8, %r1649;
|
||
and.b32 %r395, %r31, 3;
|
||
sub.s32 %r397, %r276, %r395;
|
||
shl.b32 %r398, %r397, 2;
|
||
shr.u32 %r400, %r279, %r398;
|
||
and.b32 %r393, %r400, 65535;
|
||
// inline asm
|
||
prmt.b32 %r374, %r273, %r2, %r393;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r378, %r2, %r3, %r393;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r382, %r3, %r4, %r393;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r386, %r4, %r5, %r393;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r390, %r5, %r273, %r393;
|
||
// inline asm
|
||
shr.u32 %r394, %r31, 2;
|
||
setp.gt.s32 %p17, %r394, 4;
|
||
@%p17 bra BB1_37;
|
||
|
||
setp.gt.s32 %p24, %r394, 1;
|
||
@%p24 bra BB1_33;
|
||
|
||
setp.eq.s32 %p28, %r394, 0;
|
||
@%p28 bra BB1_50;
|
||
bra.uni BB1_31;
|
||
|
||
BB1_50:
|
||
or.b32 %r3504, %r374, %r20;
|
||
mov.u32 %r3571, %r21;
|
||
mov.u32 %r3570, %r22;
|
||
mov.u32 %r3569, %r23;
|
||
mov.u32 %r1909, %r25;
|
||
mov.u32 %r1922, %r26;
|
||
mov.u32 %r3565, %r27;
|
||
mov.u32 %r3564, %r28;
|
||
mov.u32 %r1962, %r29;
|
||
mov.u32 %r1977, %r30;
|
||
mov.u32 %r3572, %r378;
|
||
mov.u32 %r3573, %r382;
|
||
mov.u32 %r3574, %r386;
|
||
mov.u32 %r3568, %r390;
|
||
mov.u32 %r3562, %r1977;
|
||
mov.u32 %r3563, %r1962;
|
||
mov.u32 %r3566, %r1922;
|
||
mov.u32 %r3567, %r1909;
|
||
bra.uni BB1_51;
|
||
|
||
BB1_37:
|
||
setp.gt.s32 %p18, %r394, 6;
|
||
@%p18 bra BB1_41;
|
||
|
||
setp.eq.s32 %p22, %r394, 5;
|
||
@%p22 bra BB1_47;
|
||
bra.uni BB1_39;
|
||
|
||
BB1_47:
|
||
or.b32 %r3569, %r374, %r23;
|
||
mov.u32 %r3574, %r17;
|
||
mov.u32 %r3573, %r18;
|
||
mov.u32 %r3572, %r19;
|
||
mov.u32 %r3504, %r20;
|
||
mov.u32 %r3568, %r24;
|
||
mov.u32 %r1904, %r25;
|
||
mov.u32 %r1917, %r26;
|
||
mov.u32 %r1957, %r29;
|
||
mov.u32 %r1972, %r30;
|
||
mov.u32 %r3570, %r378;
|
||
mov.u32 %r3571, %r382;
|
||
mov.u32 %r3564, %r386;
|
||
mov.u32 %r3565, %r390;
|
||
mov.u32 %r3562, %r1972;
|
||
mov.u32 %r3563, %r1957;
|
||
mov.u32 %r3566, %r1917;
|
||
mov.u32 %r3567, %r1904;
|
||
bra.uni BB1_51;
|
||
|
||
BB1_33:
|
||
setp.eq.s32 %p25, %r394, 2;
|
||
@%p25 bra BB1_49;
|
||
|
||
setp.eq.s32 %p26, %r394, 3;
|
||
@%p26 bra BB1_48;
|
||
bra.uni BB1_35;
|
||
|
||
BB1_48:
|
||
or.b32 %r3574, %r374, %r17;
|
||
mov.u32 %r3573, %r18;
|
||
mov.u32 %r3572, %r19;
|
||
mov.u32 %r3504, %r20;
|
||
mov.u32 %r1906, %r25;
|
||
mov.u32 %r1919, %r26;
|
||
mov.u32 %r3565, %r27;
|
||
mov.u32 %r3564, %r28;
|
||
mov.u32 %r1959, %r29;
|
||
mov.u32 %r1974, %r30;
|
||
mov.u32 %r3568, %r378;
|
||
mov.u32 %r3569, %r382;
|
||
mov.u32 %r3570, %r386;
|
||
mov.u32 %r3571, %r390;
|
||
mov.u32 %r3562, %r1974;
|
||
mov.u32 %r3563, %r1959;
|
||
mov.u32 %r3566, %r1919;
|
||
mov.u32 %r3567, %r1906;
|
||
bra.uni BB1_51;
|
||
|
||
BB1_41:
|
||
setp.eq.s32 %p19, %r394, 7;
|
||
@%p19 bra BB1_46;
|
||
|
||
setp.eq.s32 %p20, %r394, 8;
|
||
@%p20 bra BB1_45;
|
||
bra.uni BB1_43;
|
||
|
||
BB1_45:
|
||
or.b32 %r3564, %r374, %r28;
|
||
mov.u32 %r3574, %r17;
|
||
mov.u32 %r3573, %r18;
|
||
mov.u32 %r3572, %r19;
|
||
mov.u32 %r3504, %r20;
|
||
mov.u32 %r3571, %r21;
|
||
mov.u32 %r3570, %r22;
|
||
mov.u32 %r3569, %r23;
|
||
mov.u32 %r3568, %r24;
|
||
mov.u32 %r1954, %r29;
|
||
mov.u32 %r3565, %r378;
|
||
mov.u32 %r3566, %r382;
|
||
mov.u32 %r3567, %r386;
|
||
mov.u32 %r3562, %r390;
|
||
mov.u32 %r3563, %r1954;
|
||
bra.uni BB1_51;
|
||
|
||
BB1_31:
|
||
setp.eq.s32 %p29, %r394, 1;
|
||
mov.u32 %r3574, %r17;
|
||
mov.u32 %r3573, %r18;
|
||
mov.u32 %r1785, %r19;
|
||
mov.u32 %r1831, %r20;
|
||
mov.u32 %r1854, %r21;
|
||
mov.u32 %r1866, %r22;
|
||
mov.u32 %r3569, %r23;
|
||
mov.u32 %r3568, %r24;
|
||
mov.u32 %r1902, %r25;
|
||
mov.u32 %r1916, %r26;
|
||
mov.u32 %r1929, %r27;
|
||
mov.u32 %r1941, %r28;
|
||
mov.u32 %r1953, %r29;
|
||
mov.u32 %r1969, %r30;
|
||
mov.u32 %r3504, %r1831;
|
||
mov.u32 %r3562, %r1969;
|
||
mov.u32 %r3563, %r1953;
|
||
mov.u32 %r3564, %r1941;
|
||
mov.u32 %r3565, %r1929;
|
||
mov.u32 %r3566, %r1916;
|
||
mov.u32 %r3567, %r1902;
|
||
mov.u32 %r3570, %r1866;
|
||
mov.u32 %r3571, %r1854;
|
||
mov.u32 %r3572, %r1785;
|
||
@%p29 bra BB1_32;
|
||
bra.uni BB1_51;
|
||
|
||
BB1_32:
|
||
or.b32 %r3572, %r374, %r19;
|
||
mov.u32 %r3504, %r20;
|
||
mov.u32 %r3571, %r21;
|
||
mov.u32 %r3570, %r22;
|
||
mov.u32 %r1908, %r25;
|
||
mov.u32 %r1921, %r26;
|
||
mov.u32 %r3565, %r27;
|
||
mov.u32 %r3564, %r28;
|
||
mov.u32 %r1961, %r29;
|
||
mov.u32 %r1976, %r30;
|
||
mov.u32 %r3573, %r378;
|
||
mov.u32 %r3574, %r382;
|
||
mov.u32 %r3568, %r386;
|
||
mov.u32 %r3569, %r390;
|
||
mov.u32 %r3562, %r1976;
|
||
mov.u32 %r3563, %r1961;
|
||
mov.u32 %r3566, %r1921;
|
||
mov.u32 %r3567, %r1908;
|
||
bra.uni BB1_51;
|
||
|
||
BB1_39:
|
||
setp.eq.s32 %p23, %r394, 6;
|
||
mov.u32 %r1703, %r17;
|
||
mov.u32 %r1742, %r18;
|
||
mov.u32 %r1783, %r19;
|
||
mov.u32 %r1829, %r20;
|
||
mov.u32 %r3571, %r21;
|
||
mov.u32 %r1864, %r22;
|
||
mov.u32 %r1876, %r23;
|
||
mov.u32 %r1888, %r24;
|
||
mov.u32 %r1900, %r25;
|
||
mov.u32 %r1914, %r26;
|
||
mov.u32 %r3565, %r27;
|
||
mov.u32 %r3564, %r28;
|
||
mov.u32 %r1951, %r29;
|
||
mov.u32 %r1967, %r30;
|
||
mov.u32 %r3504, %r1829;
|
||
mov.u32 %r3562, %r1967;
|
||
mov.u32 %r3563, %r1951;
|
||
mov.u32 %r3566, %r1914;
|
||
mov.u32 %r3567, %r1900;
|
||
mov.u32 %r3568, %r1888;
|
||
mov.u32 %r3569, %r1876;
|
||
mov.u32 %r3570, %r1864;
|
||
mov.u32 %r3572, %r1783;
|
||
mov.u32 %r3573, %r1742;
|
||
mov.u32 %r3574, %r1703;
|
||
@%p23 bra BB1_40;
|
||
bra.uni BB1_51;
|
||
|
||
BB1_40:
|
||
or.b32 %r3570, %r374, %r22;
|
||
mov.u32 %r3574, %r17;
|
||
mov.u32 %r3573, %r18;
|
||
mov.u32 %r3572, %r19;
|
||
mov.u32 %r3504, %r20;
|
||
mov.u32 %r3569, %r23;
|
||
mov.u32 %r3568, %r24;
|
||
mov.u32 %r1903, %r25;
|
||
mov.u32 %r1956, %r29;
|
||
mov.u32 %r1971, %r30;
|
||
mov.u32 %r3571, %r378;
|
||
mov.u32 %r3564, %r382;
|
||
mov.u32 %r3565, %r386;
|
||
mov.u32 %r3566, %r390;
|
||
mov.u32 %r3562, %r1971;
|
||
mov.u32 %r3563, %r1956;
|
||
mov.u32 %r3567, %r1903;
|
||
bra.uni BB1_51;
|
||
|
||
BB1_49:
|
||
or.b32 %r3573, %r374, %r18;
|
||
mov.u32 %r3572, %r19;
|
||
mov.u32 %r3504, %r20;
|
||
mov.u32 %r3571, %r21;
|
||
mov.u32 %r1907, %r25;
|
||
mov.u32 %r1920, %r26;
|
||
mov.u32 %r3565, %r27;
|
||
mov.u32 %r3564, %r28;
|
||
mov.u32 %r1960, %r29;
|
||
mov.u32 %r1975, %r30;
|
||
mov.u32 %r3574, %r378;
|
||
mov.u32 %r3568, %r382;
|
||
mov.u32 %r3569, %r386;
|
||
mov.u32 %r3570, %r390;
|
||
mov.u32 %r3562, %r1975;
|
||
mov.u32 %r3563, %r1960;
|
||
mov.u32 %r3566, %r1920;
|
||
mov.u32 %r3567, %r1907;
|
||
bra.uni BB1_51;
|
||
|
||
BB1_35:
|
||
setp.eq.s32 %p27, %r394, 4;
|
||
mov.u32 %r1704, %r17;
|
||
mov.u32 %r1743, %r18;
|
||
mov.u32 %r1784, %r19;
|
||
mov.u32 %r1830, %r20;
|
||
mov.u32 %r3571, %r21;
|
||
mov.u32 %r3570, %r22;
|
||
mov.u32 %r3569, %r23;
|
||
mov.u32 %r1889, %r24;
|
||
mov.u32 %r1901, %r25;
|
||
mov.u32 %r1915, %r26;
|
||
mov.u32 %r1928, %r27;
|
||
mov.u32 %r3564, %r28;
|
||
mov.u32 %r1952, %r29;
|
||
mov.u32 %r1968, %r30;
|
||
mov.u32 %r3504, %r1830;
|
||
mov.u32 %r3562, %r1968;
|
||
mov.u32 %r3563, %r1952;
|
||
mov.u32 %r3565, %r1928;
|
||
mov.u32 %r3566, %r1915;
|
||
mov.u32 %r3567, %r1901;
|
||
mov.u32 %r3568, %r1889;
|
||
mov.u32 %r3572, %r1784;
|
||
mov.u32 %r3573, %r1743;
|
||
mov.u32 %r3574, %r1704;
|
||
@%p27 bra BB1_36;
|
||
bra.uni BB1_51;
|
||
|
||
BB1_36:
|
||
or.b32 %r3568, %r374, %r24;
|
||
mov.u32 %r3574, %r17;
|
||
mov.u32 %r3573, %r18;
|
||
mov.u32 %r3572, %r19;
|
||
mov.u32 %r3504, %r20;
|
||
mov.u32 %r1905, %r25;
|
||
mov.u32 %r1918, %r26;
|
||
mov.u32 %r3565, %r27;
|
||
mov.u32 %r1958, %r29;
|
||
mov.u32 %r1973, %r30;
|
||
mov.u32 %r3569, %r378;
|
||
mov.u32 %r3570, %r382;
|
||
mov.u32 %r3571, %r386;
|
||
mov.u32 %r3564, %r390;
|
||
mov.u32 %r3562, %r1973;
|
||
mov.u32 %r3563, %r1958;
|
||
mov.u32 %r3566, %r1918;
|
||
mov.u32 %r3567, %r1905;
|
||
bra.uni BB1_51;
|
||
|
||
BB1_46:
|
||
or.b32 %r3571, %r374, %r21;
|
||
mov.u32 %r3574, %r17;
|
||
mov.u32 %r3573, %r18;
|
||
mov.u32 %r3572, %r19;
|
||
mov.u32 %r3504, %r20;
|
||
mov.u32 %r3570, %r22;
|
||
mov.u32 %r3569, %r23;
|
||
mov.u32 %r3568, %r24;
|
||
mov.u32 %r1955, %r29;
|
||
mov.u32 %r1970, %r30;
|
||
mov.u32 %r3564, %r378;
|
||
mov.u32 %r3565, %r382;
|
||
mov.u32 %r3566, %r386;
|
||
mov.u32 %r3567, %r390;
|
||
mov.u32 %r3562, %r1970;
|
||
mov.u32 %r3563, %r1955;
|
||
bra.uni BB1_51;
|
||
|
||
BB1_43:
|
||
setp.ne.s32 %p21, %r394, 9;
|
||
mov.u32 %r3504, %r20;
|
||
mov.u32 %r3562, %r30;
|
||
mov.u32 %r3563, %r29;
|
||
mov.u32 %r3564, %r28;
|
||
mov.u32 %r3565, %r27;
|
||
mov.u32 %r3566, %r26;
|
||
mov.u32 %r3567, %r25;
|
||
mov.u32 %r3568, %r24;
|
||
mov.u32 %r3569, %r23;
|
||
mov.u32 %r3570, %r22;
|
||
mov.u32 %r3571, %r21;
|
||
mov.u32 %r3572, %r19;
|
||
mov.u32 %r3573, %r18;
|
||
mov.u32 %r3574, %r17;
|
||
@%p21 bra BB1_51;
|
||
|
||
or.b32 %r3565, %r374, %r27;
|
||
mov.u32 %r3574, %r17;
|
||
mov.u32 %r3573, %r18;
|
||
mov.u32 %r3572, %r19;
|
||
mov.u32 %r3504, %r20;
|
||
mov.u32 %r3571, %r21;
|
||
mov.u32 %r3570, %r22;
|
||
mov.u32 %r3569, %r23;
|
||
mov.u32 %r3568, %r24;
|
||
mov.u32 %r3564, %r28;
|
||
mov.u32 %r3562, %r386;
|
||
mov.u32 %r3563, %r390;
|
||
mov.u32 %r3566, %r378;
|
||
mov.u32 %r3567, %r382;
|
||
|
||
BB1_51:
|
||
add.s32 %r402, %r31, %r1649;
|
||
shl.b32 %r403, %r402, 3;
|
||
and.b32 %r404, %r403, 24;
|
||
mov.u32 %r405, 128;
|
||
shl.b32 %r406, %r405, %r404;
|
||
setp.lt.u32 %p30, %r402, 4;
|
||
selp.b32 %r407, %r406, 0, %p30;
|
||
or.b32 %r408, %r3504, %r407;
|
||
and.b32 %r409, %r402, -4;
|
||
setp.eq.s32 %p31, %r409, 4;
|
||
selp.b32 %r410, %r406, 0, %p31;
|
||
or.b32 %r411, %r3572, %r410;
|
||
setp.eq.s32 %p32, %r409, 8;
|
||
selp.b32 %r412, %r406, 0, %p32;
|
||
or.b32 %r413, %r3573, %r412;
|
||
setp.eq.s32 %p33, %r409, 12;
|
||
selp.b32 %r414, %r406, 0, %p33;
|
||
or.b32 %r415, %r3574, %r414;
|
||
setp.eq.s32 %p34, %r409, 16;
|
||
selp.b32 %r416, %r406, 0, %p34;
|
||
or.b32 %r417, %r3568, %r416;
|
||
setp.eq.s32 %p35, %r409, 20;
|
||
selp.b32 %r418, %r406, 0, %p35;
|
||
or.b32 %r419, %r3569, %r418;
|
||
setp.eq.s32 %p36, %r409, 24;
|
||
selp.b32 %r420, %r406, 0, %p36;
|
||
or.b32 %r421, %r3570, %r420;
|
||
setp.eq.s32 %p37, %r409, 28;
|
||
selp.b32 %r422, %r406, 0, %p37;
|
||
or.b32 %r423, %r3571, %r422;
|
||
setp.eq.s32 %p38, %r409, 32;
|
||
selp.b32 %r424, %r406, 0, %p38;
|
||
or.b32 %r425, %r3564, %r424;
|
||
setp.eq.s32 %p39, %r409, 36;
|
||
selp.b32 %r426, %r406, 0, %p39;
|
||
or.b32 %r427, %r3565, %r426;
|
||
setp.eq.s32 %p40, %r409, 40;
|
||
selp.b32 %r428, %r406, 0, %p40;
|
||
or.b32 %r429, %r3566, %r428;
|
||
setp.eq.s32 %p41, %r409, 44;
|
||
selp.b32 %r430, %r406, 0, %p41;
|
||
or.b32 %r431, %r3567, %r430;
|
||
setp.eq.s32 %p42, %r409, 48;
|
||
selp.b32 %r432, %r406, 0, %p42;
|
||
or.b32 %r433, %r3562, %r432;
|
||
setp.eq.s32 %p43, %r409, 52;
|
||
selp.b32 %r434, %r406, 0, %p43;
|
||
or.b32 %r435, %r3563, %r434;
|
||
add.s32 %r436, %r408, -680876937;
|
||
shf.l.wrap.b32 %r437, %r436, %r436, 7;
|
||
add.s32 %r438, %r437, -271733879;
|
||
and.b32 %r439, %r438, 2004318071;
|
||
xor.b32 %r440, %r439, -1732584194;
|
||
add.s32 %r441, %r411, %r440;
|
||
add.s32 %r442, %r441, -117830708;
|
||
shf.l.wrap.b32 %r443, %r442, %r442, 12;
|
||
add.s32 %r444, %r443, %r438;
|
||
xor.b32 %r445, %r438, -271733879;
|
||
and.b32 %r446, %r444, %r445;
|
||
xor.b32 %r447, %r446, -271733879;
|
||
add.s32 %r448, %r413, %r447;
|
||
add.s32 %r449, %r448, -1126478375;
|
||
shf.l.wrap.b32 %r450, %r449, %r449, 17;
|
||
add.s32 %r451, %r450, %r444;
|
||
xor.b32 %r452, %r444, %r438;
|
||
and.b32 %r453, %r451, %r452;
|
||
xor.b32 %r454, %r453, %r438;
|
||
add.s32 %r455, %r415, %r454;
|
||
add.s32 %r456, %r455, -1316259209;
|
||
shf.l.wrap.b32 %r457, %r456, %r456, 22;
|
||
add.s32 %r458, %r457, %r451;
|
||
xor.b32 %r459, %r451, %r444;
|
||
and.b32 %r460, %r458, %r459;
|
||
xor.b32 %r461, %r460, %r444;
|
||
add.s32 %r462, %r417, %r437;
|
||
add.s32 %r463, %r462, %r461;
|
||
add.s32 %r464, %r463, -448152776;
|
||
shf.l.wrap.b32 %r465, %r464, %r464, 7;
|
||
add.s32 %r466, %r465, %r458;
|
||
xor.b32 %r467, %r458, %r451;
|
||
and.b32 %r468, %r466, %r467;
|
||
xor.b32 %r469, %r468, %r451;
|
||
add.s32 %r470, %r419, %r444;
|
||
add.s32 %r471, %r470, %r469;
|
||
add.s32 %r472, %r471, 1200080426;
|
||
shf.l.wrap.b32 %r473, %r472, %r472, 12;
|
||
add.s32 %r474, %r473, %r466;
|
||
xor.b32 %r475, %r466, %r458;
|
||
and.b32 %r476, %r474, %r475;
|
||
xor.b32 %r477, %r476, %r458;
|
||
add.s32 %r478, %r421, %r451;
|
||
add.s32 %r479, %r478, %r477;
|
||
add.s32 %r480, %r479, -1473231341;
|
||
shf.l.wrap.b32 %r481, %r480, %r480, 17;
|
||
add.s32 %r482, %r481, %r474;
|
||
xor.b32 %r483, %r474, %r466;
|
||
and.b32 %r484, %r482, %r483;
|
||
xor.b32 %r485, %r484, %r466;
|
||
add.s32 %r486, %r423, %r458;
|
||
add.s32 %r487, %r486, %r485;
|
||
add.s32 %r488, %r487, -45705983;
|
||
shf.l.wrap.b32 %r489, %r488, %r488, 22;
|
||
add.s32 %r490, %r489, %r482;
|
||
xor.b32 %r491, %r482, %r474;
|
||
and.b32 %r492, %r490, %r491;
|
||
xor.b32 %r493, %r492, %r474;
|
||
add.s32 %r494, %r425, %r466;
|
||
add.s32 %r495, %r494, %r493;
|
||
add.s32 %r496, %r495, 1770035416;
|
||
shf.l.wrap.b32 %r497, %r496, %r496, 7;
|
||
add.s32 %r498, %r497, %r490;
|
||
xor.b32 %r499, %r490, %r482;
|
||
and.b32 %r500, %r498, %r499;
|
||
xor.b32 %r501, %r500, %r482;
|
||
add.s32 %r502, %r427, %r474;
|
||
add.s32 %r503, %r502, %r501;
|
||
add.s32 %r504, %r503, -1958414417;
|
||
shf.l.wrap.b32 %r505, %r504, %r504, 12;
|
||
add.s32 %r506, %r505, %r498;
|
||
xor.b32 %r507, %r498, %r490;
|
||
and.b32 %r508, %r506, %r507;
|
||
xor.b32 %r509, %r508, %r490;
|
||
add.s32 %r510, %r429, %r482;
|
||
add.s32 %r511, %r510, %r509;
|
||
add.s32 %r512, %r511, -42063;
|
||
shf.l.wrap.b32 %r513, %r512, %r512, 17;
|
||
add.s32 %r514, %r513, %r506;
|
||
xor.b32 %r515, %r506, %r498;
|
||
and.b32 %r516, %r514, %r515;
|
||
xor.b32 %r517, %r516, %r498;
|
||
add.s32 %r518, %r431, %r490;
|
||
add.s32 %r519, %r518, %r517;
|
||
add.s32 %r520, %r519, -1990404162;
|
||
shf.l.wrap.b32 %r521, %r520, %r520, 22;
|
||
add.s32 %r522, %r521, %r514;
|
||
xor.b32 %r523, %r514, %r506;
|
||
and.b32 %r524, %r522, %r523;
|
||
xor.b32 %r525, %r524, %r506;
|
||
add.s32 %r526, %r433, %r498;
|
||
add.s32 %r527, %r526, %r525;
|
||
add.s32 %r528, %r527, 1804603682;
|
||
shf.l.wrap.b32 %r529, %r528, %r528, 7;
|
||
add.s32 %r530, %r529, %r522;
|
||
xor.b32 %r531, %r522, %r514;
|
||
and.b32 %r532, %r530, %r531;
|
||
xor.b32 %r533, %r532, %r514;
|
||
add.s32 %r534, %r435, %r506;
|
||
add.s32 %r535, %r534, %r533;
|
||
add.s32 %r536, %r535, -40341101;
|
||
shf.l.wrap.b32 %r537, %r536, %r536, 12;
|
||
add.s32 %r538, %r537, %r530;
|
||
xor.b32 %r539, %r530, %r522;
|
||
and.b32 %r540, %r538, %r539;
|
||
xor.b32 %r541, %r540, %r522;
|
||
add.s32 %r542, %r403, %r514;
|
||
add.s32 %r543, %r542, %r541;
|
||
add.s32 %r544, %r543, -1502002290;
|
||
shf.l.wrap.b32 %r545, %r544, %r544, 17;
|
||
add.s32 %r546, %r545, %r538;
|
||
xor.b32 %r547, %r538, %r530;
|
||
and.b32 %r548, %r546, %r547;
|
||
xor.b32 %r549, %r548, %r530;
|
||
add.s32 %r550, %r522, %r549;
|
||
add.s32 %r551, %r550, 1236535329;
|
||
shf.l.wrap.b32 %r552, %r551, %r551, 22;
|
||
add.s32 %r553, %r552, %r546;
|
||
xor.b32 %r554, %r553, %r546;
|
||
and.b32 %r555, %r554, %r538;
|
||
xor.b32 %r556, %r555, %r546;
|
||
add.s32 %r557, %r411, %r530;
|
||
add.s32 %r558, %r557, %r556;
|
||
add.s32 %r559, %r558, -165796510;
|
||
shf.l.wrap.b32 %r560, %r559, %r559, 5;
|
||
add.s32 %r561, %r560, %r553;
|
||
xor.b32 %r562, %r561, %r553;
|
||
and.b32 %r563, %r562, %r546;
|
||
xor.b32 %r564, %r563, %r553;
|
||
add.s32 %r565, %r421, %r538;
|
||
add.s32 %r566, %r565, %r564;
|
||
add.s32 %r567, %r566, -1069501632;
|
||
shf.l.wrap.b32 %r568, %r567, %r567, 9;
|
||
add.s32 %r569, %r568, %r561;
|
||
xor.b32 %r570, %r569, %r561;
|
||
and.b32 %r571, %r570, %r553;
|
||
xor.b32 %r572, %r571, %r561;
|
||
add.s32 %r573, %r431, %r546;
|
||
add.s32 %r574, %r573, %r572;
|
||
add.s32 %r575, %r574, 643717713;
|
||
shf.l.wrap.b32 %r576, %r575, %r575, 14;
|
||
add.s32 %r577, %r576, %r569;
|
||
xor.b32 %r578, %r577, %r569;
|
||
and.b32 %r579, %r578, %r561;
|
||
xor.b32 %r580, %r579, %r569;
|
||
add.s32 %r581, %r408, %r553;
|
||
add.s32 %r582, %r581, %r580;
|
||
add.s32 %r583, %r582, -373897302;
|
||
shf.l.wrap.b32 %r584, %r583, %r583, 20;
|
||
add.s32 %r585, %r584, %r577;
|
||
xor.b32 %r586, %r585, %r577;
|
||
and.b32 %r587, %r586, %r569;
|
||
xor.b32 %r588, %r587, %r577;
|
||
add.s32 %r589, %r419, %r561;
|
||
add.s32 %r590, %r589, %r588;
|
||
add.s32 %r591, %r590, -701558691;
|
||
shf.l.wrap.b32 %r592, %r591, %r591, 5;
|
||
add.s32 %r593, %r592, %r585;
|
||
xor.b32 %r594, %r593, %r585;
|
||
and.b32 %r595, %r594, %r577;
|
||
xor.b32 %r596, %r595, %r585;
|
||
add.s32 %r597, %r429, %r569;
|
||
add.s32 %r598, %r597, %r596;
|
||
add.s32 %r599, %r598, 38016083;
|
||
shf.l.wrap.b32 %r600, %r599, %r599, 9;
|
||
add.s32 %r601, %r600, %r593;
|
||
xor.b32 %r602, %r601, %r593;
|
||
and.b32 %r603, %r602, %r585;
|
||
xor.b32 %r604, %r603, %r593;
|
||
add.s32 %r605, %r577, %r604;
|
||
add.s32 %r606, %r605, -660478335;
|
||
shf.l.wrap.b32 %r607, %r606, %r606, 14;
|
||
add.s32 %r608, %r607, %r601;
|
||
xor.b32 %r609, %r608, %r601;
|
||
and.b32 %r610, %r609, %r593;
|
||
xor.b32 %r611, %r610, %r601;
|
||
add.s32 %r612, %r417, %r585;
|
||
add.s32 %r613, %r612, %r611;
|
||
add.s32 %r614, %r613, -405537848;
|
||
shf.l.wrap.b32 %r615, %r614, %r614, 20;
|
||
add.s32 %r616, %r615, %r608;
|
||
xor.b32 %r617, %r616, %r608;
|
||
and.b32 %r618, %r617, %r601;
|
||
xor.b32 %r619, %r618, %r608;
|
||
add.s32 %r620, %r427, %r593;
|
||
add.s32 %r621, %r620, %r619;
|
||
add.s32 %r622, %r621, 568446438;
|
||
shf.l.wrap.b32 %r623, %r622, %r622, 5;
|
||
add.s32 %r624, %r623, %r616;
|
||
xor.b32 %r625, %r624, %r616;
|
||
and.b32 %r626, %r625, %r608;
|
||
xor.b32 %r627, %r626, %r616;
|
||
add.s32 %r628, %r403, %r601;
|
||
add.s32 %r629, %r628, %r627;
|
||
add.s32 %r630, %r629, -1019803690;
|
||
shf.l.wrap.b32 %r631, %r630, %r630, 9;
|
||
add.s32 %r632, %r631, %r624;
|
||
xor.b32 %r633, %r632, %r624;
|
||
and.b32 %r634, %r633, %r616;
|
||
xor.b32 %r635, %r634, %r624;
|
||
add.s32 %r636, %r415, %r608;
|
||
add.s32 %r637, %r636, %r635;
|
||
add.s32 %r638, %r637, -187363961;
|
||
shf.l.wrap.b32 %r639, %r638, %r638, 14;
|
||
add.s32 %r640, %r639, %r632;
|
||
xor.b32 %r641, %r640, %r632;
|
||
and.b32 %r642, %r641, %r624;
|
||
xor.b32 %r643, %r642, %r632;
|
||
add.s32 %r644, %r425, %r616;
|
||
add.s32 %r645, %r644, %r643;
|
||
add.s32 %r646, %r645, 1163531501;
|
||
shf.l.wrap.b32 %r647, %r646, %r646, 20;
|
||
add.s32 %r648, %r647, %r640;
|
||
xor.b32 %r649, %r648, %r640;
|
||
and.b32 %r650, %r649, %r632;
|
||
xor.b32 %r651, %r650, %r640;
|
||
add.s32 %r652, %r435, %r624;
|
||
add.s32 %r653, %r652, %r651;
|
||
add.s32 %r654, %r653, -1444681467;
|
||
shf.l.wrap.b32 %r655, %r654, %r654, 5;
|
||
add.s32 %r656, %r655, %r648;
|
||
xor.b32 %r657, %r656, %r648;
|
||
and.b32 %r658, %r657, %r640;
|
||
xor.b32 %r659, %r658, %r648;
|
||
add.s32 %r660, %r413, %r632;
|
||
add.s32 %r661, %r660, %r659;
|
||
add.s32 %r662, %r661, -51403784;
|
||
shf.l.wrap.b32 %r663, %r662, %r662, 9;
|
||
add.s32 %r664, %r663, %r656;
|
||
xor.b32 %r665, %r664, %r656;
|
||
and.b32 %r666, %r665, %r648;
|
||
xor.b32 %r667, %r666, %r656;
|
||
add.s32 %r668, %r423, %r640;
|
||
add.s32 %r669, %r668, %r667;
|
||
add.s32 %r670, %r669, 1735328473;
|
||
shf.l.wrap.b32 %r671, %r670, %r670, 14;
|
||
add.s32 %r672, %r671, %r664;
|
||
xor.b32 %r673, %r672, %r664;
|
||
and.b32 %r674, %r673, %r656;
|
||
xor.b32 %r675, %r674, %r664;
|
||
add.s32 %r676, %r433, %r648;
|
||
add.s32 %r677, %r676, %r675;
|
||
add.s32 %r678, %r677, -1926607734;
|
||
shf.l.wrap.b32 %r679, %r678, %r678, 20;
|
||
add.s32 %r680, %r679, %r672;
|
||
xor.b32 %r681, %r673, %r680;
|
||
add.s32 %r682, %r419, %r656;
|
||
add.s32 %r683, %r682, %r681;
|
||
add.s32 %r684, %r683, -378558;
|
||
shf.l.wrap.b32 %r685, %r684, %r684, 4;
|
||
add.s32 %r686, %r685, %r680;
|
||
xor.b32 %r687, %r680, %r672;
|
||
xor.b32 %r688, %r687, %r686;
|
||
add.s32 %r689, %r425, %r664;
|
||
add.s32 %r690, %r689, %r688;
|
||
add.s32 %r691, %r690, -2022574463;
|
||
shf.l.wrap.b32 %r692, %r691, %r691, 11;
|
||
add.s32 %r693, %r692, %r686;
|
||
xor.b32 %r694, %r686, %r680;
|
||
xor.b32 %r695, %r694, %r693;
|
||
add.s32 %r696, %r431, %r672;
|
||
add.s32 %r697, %r696, %r695;
|
||
add.s32 %r698, %r697, 1839030562;
|
||
shf.l.wrap.b32 %r699, %r698, %r698, 16;
|
||
add.s32 %r700, %r699, %r693;
|
||
xor.b32 %r701, %r693, %r686;
|
||
xor.b32 %r702, %r701, %r700;
|
||
add.s32 %r703, %r403, %r680;
|
||
add.s32 %r704, %r703, %r702;
|
||
add.s32 %r705, %r704, -35309556;
|
||
shf.l.wrap.b32 %r706, %r705, %r705, 23;
|
||
add.s32 %r707, %r706, %r700;
|
||
xor.b32 %r708, %r700, %r693;
|
||
xor.b32 %r709, %r708, %r707;
|
||
add.s32 %r710, %r411, %r686;
|
||
add.s32 %r711, %r710, %r709;
|
||
add.s32 %r712, %r711, -1530992060;
|
||
shf.l.wrap.b32 %r713, %r712, %r712, 4;
|
||
add.s32 %r714, %r713, %r707;
|
||
xor.b32 %r715, %r707, %r700;
|
||
xor.b32 %r716, %r715, %r714;
|
||
add.s32 %r717, %r417, %r693;
|
||
add.s32 %r718, %r717, %r716;
|
||
add.s32 %r719, %r718, 1272893353;
|
||
shf.l.wrap.b32 %r720, %r719, %r719, 11;
|
||
add.s32 %r721, %r720, %r714;
|
||
xor.b32 %r722, %r714, %r707;
|
||
xor.b32 %r723, %r722, %r721;
|
||
add.s32 %r724, %r423, %r700;
|
||
add.s32 %r725, %r724, %r723;
|
||
add.s32 %r726, %r725, -155497632;
|
||
shf.l.wrap.b32 %r727, %r726, %r726, 16;
|
||
add.s32 %r728, %r727, %r721;
|
||
xor.b32 %r729, %r721, %r714;
|
||
xor.b32 %r730, %r729, %r728;
|
||
add.s32 %r731, %r429, %r707;
|
||
add.s32 %r732, %r731, %r730;
|
||
add.s32 %r733, %r732, -1094730640;
|
||
shf.l.wrap.b32 %r734, %r733, %r733, 23;
|
||
add.s32 %r735, %r734, %r728;
|
||
xor.b32 %r736, %r728, %r721;
|
||
xor.b32 %r737, %r736, %r735;
|
||
add.s32 %r738, %r435, %r714;
|
||
add.s32 %r739, %r738, %r737;
|
||
add.s32 %r740, %r739, 681279174;
|
||
shf.l.wrap.b32 %r741, %r740, %r740, 4;
|
||
add.s32 %r742, %r741, %r735;
|
||
xor.b32 %r743, %r735, %r728;
|
||
xor.b32 %r744, %r743, %r742;
|
||
add.s32 %r745, %r408, %r721;
|
||
add.s32 %r746, %r745, %r744;
|
||
add.s32 %r747, %r746, -358537222;
|
||
shf.l.wrap.b32 %r748, %r747, %r747, 11;
|
||
add.s32 %r749, %r748, %r742;
|
||
xor.b32 %r750, %r742, %r735;
|
||
xor.b32 %r751, %r750, %r749;
|
||
add.s32 %r752, %r415, %r728;
|
||
add.s32 %r753, %r752, %r751;
|
||
add.s32 %r754, %r753, -722521979;
|
||
shf.l.wrap.b32 %r755, %r754, %r754, 16;
|
||
add.s32 %r756, %r755, %r749;
|
||
xor.b32 %r757, %r749, %r742;
|
||
xor.b32 %r758, %r757, %r756;
|
||
add.s32 %r759, %r421, %r735;
|
||
add.s32 %r760, %r759, %r758;
|
||
add.s32 %r761, %r760, 76029189;
|
||
shf.l.wrap.b32 %r762, %r761, %r761, 23;
|
||
add.s32 %r763, %r762, %r756;
|
||
xor.b32 %r764, %r756, %r749;
|
||
xor.b32 %r765, %r764, %r763;
|
||
add.s32 %r766, %r427, %r742;
|
||
add.s32 %r767, %r766, %r765;
|
||
add.s32 %r768, %r767, -640364487;
|
||
shf.l.wrap.b32 %r769, %r768, %r768, 4;
|
||
add.s32 %r770, %r769, %r763;
|
||
xor.b32 %r771, %r763, %r756;
|
||
xor.b32 %r772, %r771, %r770;
|
||
add.s32 %r773, %r433, %r749;
|
||
add.s32 %r774, %r773, %r772;
|
||
add.s32 %r775, %r774, -421815835;
|
||
shf.l.wrap.b32 %r776, %r775, %r775, 11;
|
||
add.s32 %r777, %r776, %r770;
|
||
xor.b32 %r778, %r770, %r763;
|
||
xor.b32 %r779, %r778, %r777;
|
||
add.s32 %r780, %r756, %r779;
|
||
add.s32 %r781, %r780, 530742520;
|
||
shf.l.wrap.b32 %r782, %r781, %r781, 16;
|
||
add.s32 %r783, %r782, %r777;
|
||
xor.b32 %r784, %r777, %r770;
|
||
xor.b32 %r785, %r784, %r783;
|
||
add.s32 %r786, %r413, %r763;
|
||
add.s32 %r787, %r786, %r785;
|
||
add.s32 %r788, %r787, -995338651;
|
||
shf.l.wrap.b32 %r789, %r788, %r788, 23;
|
||
add.s32 %r790, %r789, %r783;
|
||
not.b32 %r791, %r777;
|
||
or.b32 %r792, %r790, %r791;
|
||
xor.b32 %r793, %r792, %r783;
|
||
add.s32 %r794, %r408, %r770;
|
||
add.s32 %r795, %r794, %r793;
|
||
add.s32 %r796, %r795, -198630844;
|
||
shf.l.wrap.b32 %r797, %r796, %r796, 6;
|
||
add.s32 %r798, %r797, %r790;
|
||
not.b32 %r799, %r783;
|
||
or.b32 %r800, %r798, %r799;
|
||
xor.b32 %r801, %r800, %r790;
|
||
add.s32 %r802, %r423, %r777;
|
||
add.s32 %r803, %r802, %r801;
|
||
add.s32 %r804, %r803, 1126891415;
|
||
shf.l.wrap.b32 %r805, %r804, %r804, 10;
|
||
add.s32 %r806, %r805, %r798;
|
||
not.b32 %r807, %r790;
|
||
or.b32 %r808, %r806, %r807;
|
||
xor.b32 %r809, %r808, %r798;
|
||
add.s32 %r810, %r403, %r783;
|
||
add.s32 %r811, %r810, %r809;
|
||
add.s32 %r812, %r811, -1416354905;
|
||
shf.l.wrap.b32 %r813, %r812, %r812, 15;
|
||
add.s32 %r814, %r813, %r806;
|
||
not.b32 %r815, %r798;
|
||
or.b32 %r816, %r814, %r815;
|
||
xor.b32 %r817, %r816, %r806;
|
||
add.s32 %r818, %r419, %r790;
|
||
add.s32 %r819, %r818, %r817;
|
||
add.s32 %r820, %r819, -57434055;
|
||
shf.l.wrap.b32 %r821, %r820, %r820, 21;
|
||
add.s32 %r822, %r821, %r814;
|
||
not.b32 %r823, %r806;
|
||
or.b32 %r824, %r822, %r823;
|
||
xor.b32 %r825, %r824, %r814;
|
||
add.s32 %r826, %r433, %r798;
|
||
add.s32 %r827, %r826, %r825;
|
||
add.s32 %r828, %r827, 1700485571;
|
||
shf.l.wrap.b32 %r829, %r828, %r828, 6;
|
||
add.s32 %r830, %r829, %r822;
|
||
not.b32 %r831, %r814;
|
||
or.b32 %r832, %r830, %r831;
|
||
xor.b32 %r833, %r832, %r822;
|
||
add.s32 %r834, %r415, %r806;
|
||
add.s32 %r835, %r834, %r833;
|
||
add.s32 %r836, %r835, -1894986606;
|
||
shf.l.wrap.b32 %r837, %r836, %r836, 10;
|
||
add.s32 %r838, %r837, %r830;
|
||
not.b32 %r839, %r822;
|
||
or.b32 %r840, %r838, %r839;
|
||
xor.b32 %r841, %r840, %r830;
|
||
add.s32 %r842, %r429, %r814;
|
||
add.s32 %r843, %r842, %r841;
|
||
add.s32 %r844, %r843, -1051523;
|
||
shf.l.wrap.b32 %r845, %r844, %r844, 15;
|
||
add.s32 %r846, %r845, %r838;
|
||
not.b32 %r847, %r830;
|
||
or.b32 %r848, %r846, %r847;
|
||
xor.b32 %r849, %r848, %r838;
|
||
add.s32 %r850, %r411, %r822;
|
||
add.s32 %r851, %r850, %r849;
|
||
add.s32 %r852, %r851, -2054922799;
|
||
shf.l.wrap.b32 %r853, %r852, %r852, 21;
|
||
add.s32 %r854, %r853, %r846;
|
||
not.b32 %r855, %r838;
|
||
or.b32 %r856, %r854, %r855;
|
||
xor.b32 %r857, %r856, %r846;
|
||
add.s32 %r858, %r425, %r830;
|
||
add.s32 %r859, %r858, %r857;
|
||
add.s32 %r860, %r859, 1873313359;
|
||
shf.l.wrap.b32 %r861, %r860, %r860, 6;
|
||
add.s32 %r862, %r861, %r854;
|
||
not.b32 %r863, %r846;
|
||
or.b32 %r864, %r862, %r863;
|
||
xor.b32 %r865, %r864, %r854;
|
||
add.s32 %r866, %r838, %r865;
|
||
add.s32 %r867, %r866, -30611744;
|
||
shf.l.wrap.b32 %r868, %r867, %r867, 10;
|
||
add.s32 %r869, %r868, %r862;
|
||
not.b32 %r870, %r854;
|
||
or.b32 %r871, %r869, %r870;
|
||
xor.b32 %r872, %r871, %r862;
|
||
add.s32 %r873, %r421, %r846;
|
||
add.s32 %r874, %r873, %r872;
|
||
add.s32 %r875, %r874, -1560198380;
|
||
shf.l.wrap.b32 %r876, %r875, %r875, 15;
|
||
add.s32 %r877, %r876, %r869;
|
||
not.b32 %r878, %r862;
|
||
or.b32 %r879, %r877, %r878;
|
||
xor.b32 %r880, %r879, %r869;
|
||
add.s32 %r881, %r435, %r854;
|
||
add.s32 %r882, %r881, %r880;
|
||
add.s32 %r883, %r882, 1309151649;
|
||
shf.l.wrap.b32 %r884, %r883, %r883, 21;
|
||
add.s32 %r885, %r884, %r877;
|
||
not.b32 %r886, %r869;
|
||
or.b32 %r887, %r885, %r886;
|
||
xor.b32 %r888, %r887, %r877;
|
||
add.s32 %r889, %r417, %r862;
|
||
add.s32 %r890, %r889, %r888;
|
||
add.s32 %r891, %r890, -145523070;
|
||
shf.l.wrap.b32 %r892, %r891, %r891, 6;
|
||
add.s32 %r893, %r892, %r885;
|
||
not.b32 %r894, %r877;
|
||
or.b32 %r895, %r893, %r894;
|
||
xor.b32 %r896, %r895, %r885;
|
||
add.s32 %r897, %r431, %r869;
|
||
add.s32 %r898, %r897, %r896;
|
||
add.s32 %r899, %r898, -1120210379;
|
||
shf.l.wrap.b32 %r900, %r899, %r899, 10;
|
||
add.s32 %r901, %r900, %r893;
|
||
not.b32 %r902, %r885;
|
||
or.b32 %r903, %r901, %r902;
|
||
xor.b32 %r904, %r903, %r893;
|
||
add.s32 %r905, %r413, %r877;
|
||
add.s32 %r906, %r905, %r904;
|
||
add.s32 %r907, %r906, 718787259;
|
||
shf.l.wrap.b32 %r908, %r907, %r907, 15;
|
||
add.s32 %r909, %r908, %r901;
|
||
not.b32 %r910, %r893;
|
||
or.b32 %r911, %r909, %r910;
|
||
xor.b32 %r912, %r911, %r901;
|
||
add.s32 %r913, %r427, %r885;
|
||
add.s32 %r914, %r913, %r912;
|
||
add.s32 %r915, %r914, -343485551;
|
||
shf.l.wrap.b32 %r916, %r915, %r915, 21;
|
||
add.s32 %r1648, %r893, 1732584193;
|
||
add.s32 %r917, %r909, %r916;
|
||
add.s32 %r1647, %r917, -271733879;
|
||
add.s32 %r1646, %r909, -1732584194;
|
||
add.s32 %r1645, %r901, 271733878;
|
||
mov.u32 %r401, 0;
|
||
setp.gt.s32 %p44, %r1649, 7;
|
||
@%p44 bra BB1_67;
|
||
|
||
setp.gt.s32 %p56, %r1649, 3;
|
||
@%p56 bra BB1_60;
|
||
|
||
setp.gt.s32 %p62, %r1649, 1;
|
||
@%p62 bra BB1_57;
|
||
|
||
setp.eq.s32 %p65, %r1649, 0;
|
||
@%p65 bra BB1_89;
|
||
bra.uni BB1_55;
|
||
|
||
BB1_89:
|
||
mov.u32 %r3503, 2371876;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r2090, %r401;
|
||
mov.u32 %r3552, %r2090;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_67:
|
||
setp.gt.s32 %p45, %r1649, 11;
|
||
@%p45 bra BB1_75;
|
||
|
||
setp.gt.s32 %p51, %r1649, 9;
|
||
@%p51 bra BB1_72;
|
||
|
||
setp.eq.s32 %p54, %r1649, 8;
|
||
@%p54 bra BB1_85;
|
||
bra.uni BB1_70;
|
||
|
||
BB1_85:
|
||
mov.u32 %r3560, 2371876;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2080, %r401;
|
||
mov.u32 %r3552, %r2080;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_60:
|
||
setp.gt.s32 %p57, %r1649, 5;
|
||
@%p57 bra BB1_64;
|
||
|
||
setp.eq.s32 %p60, %r1649, 4;
|
||
@%p60 bra BB1_87;
|
||
bra.uni BB1_62;
|
||
|
||
BB1_87:
|
||
mov.u32 %r3559, 2371876;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2087, %r401;
|
||
mov.u32 %r3552, %r2087;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_75:
|
||
setp.gt.s32 %p46, %r1649, 13;
|
||
@%p46 bra BB1_79;
|
||
|
||
setp.eq.s32 %p49, %r1649, 12;
|
||
@%p49 bra BB1_83;
|
||
bra.uni BB1_77;
|
||
|
||
BB1_83:
|
||
mov.u32 %r3561, 2371876;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2077, %r401;
|
||
mov.u32 %r3552, %r2077;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_57:
|
||
setp.eq.s32 %p63, %r1649, 2;
|
||
@%p63 bra BB1_88;
|
||
bra.uni BB1_58;
|
||
|
||
BB1_88:
|
||
or.b32 %r3503, %r2, 824442880;
|
||
mov.u32 %r3559, 36;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r2089, %r401;
|
||
mov.u32 %r3552, %r2089;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_72:
|
||
setp.eq.s32 %p52, %r1649, 10;
|
||
@%p52 bra BB1_84;
|
||
bra.uni BB1_73;
|
||
|
||
BB1_84:
|
||
or.b32 %r3560, %r4, 824442880;
|
||
mov.u32 %r3561, 36;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2079, %r401;
|
||
mov.u32 %r3552, %r2079;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_64:
|
||
setp.eq.s32 %p58, %r1649, 6;
|
||
@%p58 bra BB1_86;
|
||
bra.uni BB1_65;
|
||
|
||
BB1_86:
|
||
or.b32 %r3559, %r3, 824442880;
|
||
mov.u32 %r3560, 36;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2086, %r401;
|
||
mov.u32 %r3552, %r2086;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_79:
|
||
setp.eq.s32 %p47, %r1649, 14;
|
||
@%p47 bra BB1_82;
|
||
bra.uni BB1_80;
|
||
|
||
BB1_82:
|
||
or.b32 %r3561, %r5, 824442880;
|
||
mov.u32 %r919, 36;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r3552, %r919;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_55:
|
||
setp.eq.s32 %p66, %r1649, 1;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2075, %r401;
|
||
mov.u32 %r3552, %r2075;
|
||
@%p66 bra BB1_56;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_56:
|
||
or.b32 %r3503, %r2, 607200256;
|
||
mov.u32 %r944, 0;
|
||
mov.u32 %r3559, %r944;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3552, %r944;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_70:
|
||
setp.eq.s32 %p55, %r1649, 9;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2071, %r401;
|
||
mov.u32 %r3552, %r2071;
|
||
@%p55 bra BB1_71;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_71:
|
||
or.b32 %r3560, %r4, 607200256;
|
||
mov.u32 %r928, 0;
|
||
mov.u32 %r3561, %r928;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r3552, %r928;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_62:
|
||
setp.eq.s32 %p61, %r1649, 5;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2073, %r401;
|
||
mov.u32 %r3552, %r2073;
|
||
@%p61 bra BB1_63;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_63:
|
||
or.b32 %r3559, %r3, 607200256;
|
||
mov.u32 %r936, 0;
|
||
mov.u32 %r3560, %r936;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r3552, %r936;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_77:
|
||
setp.eq.s32 %p50, %r1649, 13;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2069, %r401;
|
||
mov.u32 %r3552, %r2069;
|
||
@%p50 bra BB1_78;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_78:
|
||
or.b32 %r3561, %r5, 607200256;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2076, %r401;
|
||
mov.u32 %r3552, %r2076;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_58:
|
||
setp.eq.s32 %p64, %r1649, 3;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2074, %r401;
|
||
mov.u32 %r3552, %r2074;
|
||
@%p64 bra BB1_59;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_59:
|
||
or.b32 %r3503, %r2, 603979776;
|
||
mov.u32 %r3559, 9265;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r2088, %r401;
|
||
mov.u32 %r3552, %r2088;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_73:
|
||
setp.eq.s32 %p53, %r1649, 11;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2070, %r401;
|
||
mov.u32 %r3552, %r2070;
|
||
@%p53 bra BB1_74;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_74:
|
||
or.b32 %r3560, %r4, 603979776;
|
||
mov.u32 %r3561, 9265;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2078, %r401;
|
||
mov.u32 %r3552, %r2078;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_65:
|
||
setp.eq.s32 %p59, %r1649, 7;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2072, %r401;
|
||
mov.u32 %r3552, %r2072;
|
||
@%p59 bra BB1_66;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_66:
|
||
or.b32 %r3559, %r3, 603979776;
|
||
mov.u32 %r3560, 9265;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r2085, %r401;
|
||
mov.u32 %r3552, %r2085;
|
||
bra.uni BB1_90;
|
||
|
||
BB1_80:
|
||
setp.ne.s32 %p48, %r1649, 15;
|
||
mov.u32 %r3561, %r5;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r3552, %r401;
|
||
@%p48 bra BB1_90;
|
||
|
||
or.b32 %r3561, %r5, 603979776;
|
||
mov.u32 %r918, 9265;
|
||
mov.u32 %r3560, %r4;
|
||
mov.u32 %r3559, %r3;
|
||
mov.u32 %r3503, %r2;
|
||
mov.u32 %r3552, %r918;
|
||
|
||
BB1_90:
|
||
mov.u32 %r2091, %r3552;
|
||
mov.u32 %r81, %r2091;
|
||
mov.u32 %r80, %r3503;
|
||
mov.u32 %r79, %r3559;
|
||
mov.u32 %r78, %r3560;
|
||
mov.u32 %r77, %r3561;
|
||
mov.u32 %r1635, 1985229328;
|
||
mov.u32 %r1634, 4;
|
||
add.s32 %r82, %r1649, 3;
|
||
and.b32 %r969, %r82, 3;
|
||
sub.s32 %r971, %r1634, %r969;
|
||
shl.b32 %r972, %r971, 2;
|
||
shr.u32 %r974, %r1635, %r972;
|
||
and.b32 %r958, %r974, 65535;
|
||
// inline asm
|
||
prmt.b32 %r947, %r401, %r6, %r958;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r951, %r6, %r7, %r958;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r955, %r7, %r401, %r958;
|
||
// inline asm
|
||
shr.u32 %r968, %r82, 2;
|
||
mov.u32 %r966, %r401;
|
||
mov.u32 %r3547, %r401;
|
||
mov.u32 %r3548, %r401;
|
||
mov.u32 %r3549, %r401;
|
||
mov.u32 %r962, %r401;
|
||
mov.u32 %r3553, %r401;
|
||
mov.u32 %r3554, %r401;
|
||
mov.u32 %r3555, %r401;
|
||
setp.gt.s32 %p67, %r968, 5;
|
||
@%p67 bra BB1_100;
|
||
|
||
setp.gt.s32 %p75, %r968, 2;
|
||
@%p75 bra BB1_96;
|
||
|
||
setp.eq.s32 %p79, %r968, 0;
|
||
@%p79 bra BB1_116;
|
||
|
||
setp.eq.s32 %p80, %r968, 1;
|
||
@%p80 bra BB1_115;
|
||
bra.uni BB1_94;
|
||
|
||
BB1_115:
|
||
or.b32 %r3556, %r947, %r79;
|
||
mov.u32 %r1049, 0;
|
||
mov.u32 %r1048, %r1049;
|
||
mov.u32 %r3547, %r1049;
|
||
mov.u32 %r3548, %r1049;
|
||
mov.u32 %r3549, %r1049;
|
||
mov.u32 %r3550, %r1049;
|
||
mov.u32 %r3553, %r1049;
|
||
mov.u32 %r3554, %r1049;
|
||
mov.u32 %r3555, %r1049;
|
||
mov.u32 %r3502, %r80;
|
||
mov.u32 %r3551, %r81;
|
||
mov.u32 %r3557, %r951;
|
||
mov.u32 %r3558, %r955;
|
||
mov.u32 %r3545, %r1049;
|
||
mov.u32 %r3546, %r1048;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_100:
|
||
setp.gt.s32 %p68, %r968, 8;
|
||
@%p68 bra BB1_105;
|
||
|
||
setp.eq.s32 %p72, %r968, 6;
|
||
@%p72 bra BB1_112;
|
||
|
||
setp.eq.s32 %p73, %r968, 7;
|
||
@%p73 bra BB1_111;
|
||
bra.uni BB1_103;
|
||
|
||
BB1_111:
|
||
mov.u32 %r1004, 0;
|
||
mov.u32 %r1003, %r1004;
|
||
mov.u32 %r3549, %r1004;
|
||
mov.u32 %r1001, %r1004;
|
||
mov.u32 %r3553, %r1004;
|
||
mov.u32 %r3554, %r1004;
|
||
mov.u32 %r3558, %r77;
|
||
mov.u32 %r3557, %r78;
|
||
mov.u32 %r3556, %r79;
|
||
mov.u32 %r3502, %r80;
|
||
mov.u32 %r3551, %r81;
|
||
mov.u32 %r3547, %r951;
|
||
mov.u32 %r3548, %r955;
|
||
mov.u32 %r3555, %r947;
|
||
mov.u32 %r3545, %r1004;
|
||
mov.u32 %r3546, %r1003;
|
||
mov.u32 %r3550, %r1001;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_96:
|
||
setp.eq.s32 %p76, %r968, 3;
|
||
@%p76 bra BB1_114;
|
||
|
||
setp.eq.s32 %p77, %r968, 4;
|
||
@%p77 bra BB1_113;
|
||
bra.uni BB1_98;
|
||
|
||
BB1_113:
|
||
or.b32 %r3551, %r947, %r81;
|
||
mov.u32 %r1023, 0;
|
||
mov.u32 %r1022, %r1023;
|
||
mov.u32 %r3547, %r1023;
|
||
mov.u32 %r3548, %r1023;
|
||
mov.u32 %r3549, %r1023;
|
||
mov.u32 %r3550, %r1023;
|
||
mov.u32 %r3555, %r1023;
|
||
mov.u32 %r3558, %r77;
|
||
mov.u32 %r3557, %r78;
|
||
mov.u32 %r3556, %r79;
|
||
mov.u32 %r3502, %r80;
|
||
mov.u32 %r3553, %r951;
|
||
mov.u32 %r3554, %r955;
|
||
mov.u32 %r3545, %r1023;
|
||
mov.u32 %r3546, %r1022;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_105:
|
||
setp.eq.s32 %p69, %r968, 9;
|
||
@%p69 bra BB1_110;
|
||
|
||
setp.eq.s32 %p70, %r968, 10;
|
||
@%p70 bra BB1_109;
|
||
bra.uni BB1_107;
|
||
|
||
BB1_109:
|
||
mov.u32 %r986, 0;
|
||
mov.u32 %r3547, %r986;
|
||
mov.u32 %r3548, %r986;
|
||
mov.u32 %r3553, %r986;
|
||
mov.u32 %r3554, %r986;
|
||
mov.u32 %r3555, %r986;
|
||
mov.u32 %r3558, %r77;
|
||
mov.u32 %r3557, %r78;
|
||
mov.u32 %r3556, %r79;
|
||
mov.u32 %r3502, %r80;
|
||
mov.u32 %r3551, %r81;
|
||
mov.u32 %r2104, %r951;
|
||
mov.u32 %r3545, %r955;
|
||
mov.u32 %r3549, %r947;
|
||
mov.u32 %r3546, %r986;
|
||
mov.u32 %r3550, %r2104;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_116:
|
||
or.b32 %r3502, %r947, %r80;
|
||
mov.u32 %r1058, 0;
|
||
mov.u32 %r1057, %r1058;
|
||
mov.u32 %r3547, %r1058;
|
||
mov.u32 %r3548, %r1058;
|
||
mov.u32 %r3549, %r1058;
|
||
mov.u32 %r3550, %r1058;
|
||
mov.u32 %r3553, %r1058;
|
||
mov.u32 %r3554, %r1058;
|
||
mov.u32 %r3555, %r1058;
|
||
mov.u32 %r3558, %r77;
|
||
mov.u32 %r3551, %r81;
|
||
mov.u32 %r3556, %r951;
|
||
mov.u32 %r3557, %r955;
|
||
mov.u32 %r3545, %r1058;
|
||
mov.u32 %r3546, %r1057;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_94:
|
||
setp.eq.s32 %p81, %r968, 2;
|
||
mov.u32 %r3558, %r77;
|
||
mov.u32 %r2030, %r78;
|
||
mov.u32 %r2043, %r79;
|
||
mov.u32 %r2057, %r80;
|
||
mov.u32 %r2084, %r401;
|
||
mov.u32 %r3551, %r81;
|
||
mov.u32 %r2225, %r962;
|
||
mov.u32 %r2276, %r966;
|
||
mov.u32 %r3502, %r2057;
|
||
mov.u32 %r3545, %r2084;
|
||
mov.u32 %r3546, %r2276;
|
||
mov.u32 %r3550, %r2225;
|
||
mov.u32 %r3556, %r2043;
|
||
mov.u32 %r3557, %r2030;
|
||
@%p81 bra BB1_95;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_95:
|
||
or.b32 %r3557, %r947, %r78;
|
||
mov.u32 %r1040, 0;
|
||
mov.u32 %r1039, %r1040;
|
||
mov.u32 %r3547, %r1040;
|
||
mov.u32 %r3548, %r1040;
|
||
mov.u32 %r3549, %r1040;
|
||
mov.u32 %r3550, %r1040;
|
||
mov.u32 %r3553, %r1040;
|
||
mov.u32 %r3554, %r1040;
|
||
mov.u32 %r3555, %r1040;
|
||
mov.u32 %r3556, %r79;
|
||
mov.u32 %r3502, %r80;
|
||
mov.u32 %r3558, %r951;
|
||
mov.u32 %r3551, %r955;
|
||
mov.u32 %r3545, %r1040;
|
||
mov.u32 %r3546, %r1039;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_112:
|
||
mov.u32 %r1010, 0;
|
||
mov.u32 %r1009, %r1010;
|
||
mov.u32 %r3548, %r1010;
|
||
mov.u32 %r3549, %r1010;
|
||
mov.u32 %r1006, %r1010;
|
||
mov.u32 %r3553, %r1010;
|
||
mov.u32 %r3558, %r77;
|
||
mov.u32 %r3557, %r78;
|
||
mov.u32 %r3556, %r79;
|
||
mov.u32 %r3502, %r80;
|
||
mov.u32 %r3551, %r81;
|
||
mov.u32 %r3555, %r951;
|
||
mov.u32 %r3547, %r955;
|
||
mov.u32 %r3554, %r947;
|
||
mov.u32 %r3545, %r1010;
|
||
mov.u32 %r3546, %r1009;
|
||
mov.u32 %r3550, %r1006;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_103:
|
||
setp.eq.s32 %p74, %r968, 8;
|
||
mov.u32 %r2015, %r77;
|
||
mov.u32 %r2028, %r78;
|
||
mov.u32 %r2041, %r79;
|
||
mov.u32 %r2055, %r80;
|
||
mov.u32 %r2082, %r401;
|
||
mov.u32 %r2092, %r81;
|
||
mov.u32 %r2223, %r962;
|
||
mov.u32 %r2274, %r966;
|
||
mov.u32 %r3502, %r2055;
|
||
mov.u32 %r3545, %r2082;
|
||
mov.u32 %r3546, %r2274;
|
||
mov.u32 %r3550, %r2223;
|
||
mov.u32 %r3551, %r2092;
|
||
mov.u32 %r3556, %r2041;
|
||
mov.u32 %r3557, %r2028;
|
||
mov.u32 %r3558, %r2015;
|
||
@%p74 bra BB1_104;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_104:
|
||
mov.u32 %r998, 0;
|
||
mov.u32 %r997, %r998;
|
||
mov.u32 %r996, %r998;
|
||
mov.u32 %r3553, %r998;
|
||
mov.u32 %r3554, %r998;
|
||
mov.u32 %r3555, %r998;
|
||
mov.u32 %r3558, %r77;
|
||
mov.u32 %r3557, %r78;
|
||
mov.u32 %r3556, %r79;
|
||
mov.u32 %r3502, %r80;
|
||
mov.u32 %r3551, %r81;
|
||
mov.u32 %r3548, %r951;
|
||
mov.u32 %r3549, %r955;
|
||
mov.u32 %r3547, %r947;
|
||
mov.u32 %r3545, %r998;
|
||
mov.u32 %r3546, %r997;
|
||
mov.u32 %r3550, %r996;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_114:
|
||
or.b32 %r3558, %r947, %r77;
|
||
mov.u32 %r1031, 0;
|
||
mov.u32 %r1030, %r1031;
|
||
mov.u32 %r3547, %r1031;
|
||
mov.u32 %r3548, %r1031;
|
||
mov.u32 %r3549, %r1031;
|
||
mov.u32 %r3550, %r1031;
|
||
mov.u32 %r3554, %r1031;
|
||
mov.u32 %r3555, %r1031;
|
||
mov.u32 %r3557, %r78;
|
||
mov.u32 %r3556, %r79;
|
||
mov.u32 %r3502, %r80;
|
||
mov.u32 %r3551, %r951;
|
||
mov.u32 %r3553, %r955;
|
||
mov.u32 %r3545, %r1031;
|
||
mov.u32 %r3546, %r1030;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_98:
|
||
setp.eq.s32 %p78, %r968, 5;
|
||
mov.u32 %r2016, %r77;
|
||
mov.u32 %r2029, %r78;
|
||
mov.u32 %r2042, %r79;
|
||
mov.u32 %r2056, %r80;
|
||
mov.u32 %r2083, %r401;
|
||
mov.u32 %r2093, %r81;
|
||
mov.u32 %r2224, %r962;
|
||
mov.u32 %r2275, %r966;
|
||
mov.u32 %r3502, %r2056;
|
||
mov.u32 %r3545, %r2083;
|
||
mov.u32 %r3546, %r2275;
|
||
mov.u32 %r3550, %r2224;
|
||
mov.u32 %r3551, %r2093;
|
||
mov.u32 %r3556, %r2042;
|
||
mov.u32 %r3557, %r2029;
|
||
mov.u32 %r3558, %r2016;
|
||
@%p78 bra BB1_99;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_99:
|
||
mov.u32 %r1016, 0;
|
||
mov.u32 %r1015, %r1016;
|
||
mov.u32 %r3547, %r1016;
|
||
mov.u32 %r3548, %r1016;
|
||
mov.u32 %r3549, %r1016;
|
||
mov.u32 %r1011, %r1016;
|
||
mov.u32 %r3558, %r77;
|
||
mov.u32 %r3557, %r78;
|
||
mov.u32 %r3556, %r79;
|
||
mov.u32 %r3502, %r80;
|
||
mov.u32 %r3551, %r81;
|
||
mov.u32 %r3554, %r951;
|
||
mov.u32 %r3555, %r955;
|
||
mov.u32 %r3553, %r947;
|
||
mov.u32 %r3545, %r1016;
|
||
mov.u32 %r3546, %r1015;
|
||
mov.u32 %r3550, %r1011;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_110:
|
||
mov.u32 %r992, 0;
|
||
mov.u32 %r991, %r992;
|
||
mov.u32 %r3547, %r992;
|
||
mov.u32 %r3553, %r992;
|
||
mov.u32 %r3554, %r992;
|
||
mov.u32 %r3555, %r992;
|
||
mov.u32 %r3558, %r77;
|
||
mov.u32 %r3557, %r78;
|
||
mov.u32 %r3556, %r79;
|
||
mov.u32 %r3502, %r80;
|
||
mov.u32 %r3551, %r81;
|
||
mov.u32 %r3549, %r951;
|
||
mov.u32 %r2116, %r955;
|
||
mov.u32 %r3548, %r947;
|
||
mov.u32 %r3545, %r992;
|
||
mov.u32 %r3546, %r991;
|
||
mov.u32 %r3550, %r2116;
|
||
bra.uni BB1_117;
|
||
|
||
BB1_107:
|
||
setp.ne.s32 %p71, %r968, 11;
|
||
mov.u32 %r2081, %r401;
|
||
mov.u32 %r3502, %r80;
|
||
mov.u32 %r3545, %r2081;
|
||
mov.u32 %r3546, %r966;
|
||
mov.u32 %r3550, %r962;
|
||
mov.u32 %r3551, %r81;
|
||
mov.u32 %r3556, %r79;
|
||
mov.u32 %r3557, %r78;
|
||
mov.u32 %r3558, %r77;
|
||
@%p71 bra BB1_117;
|
||
|
||
mov.u32 %r3547, 0;
|
||
mov.u32 %r3548, %r3547;
|
||
mov.u32 %r3549, %r3547;
|
||
mov.u32 %r3553, %r3547;
|
||
mov.u32 %r3554, %r3547;
|
||
mov.u32 %r3555, %r3547;
|
||
mov.u32 %r3558, %r77;
|
||
mov.u32 %r3557, %r78;
|
||
mov.u32 %r3556, %r79;
|
||
mov.u32 %r3502, %r80;
|
||
mov.u32 %r3551, %r81;
|
||
mov.u32 %r3545, %r951;
|
||
mov.u32 %r3546, %r955;
|
||
mov.u32 %r3550, %r947;
|
||
|
||
BB1_117:
|
||
mov.u32 %r104, %r3545;
|
||
mov.u32 %r103, %r3546;
|
||
mov.u32 %r102, %r3547;
|
||
mov.u32 %r101, %r3548;
|
||
mov.u32 %r100, %r3549;
|
||
mov.u32 %r99, %r3550;
|
||
mov.u32 %r98, %r3551;
|
||
mov.u32 %r97, %r3553;
|
||
mov.u32 %r96, %r3554;
|
||
mov.u32 %r95, %r3555;
|
||
mov.u32 %r94, %r3502;
|
||
mov.u32 %r93, %r3556;
|
||
mov.u32 %r92, %r3557;
|
||
mov.u32 %r91, %r3558;
|
||
add.s32 %r105, %r8, %r82;
|
||
@%p44 bra BB1_133;
|
||
|
||
setp.gt.s32 %p94, %r1649, 3;
|
||
@%p94 bra BB1_126;
|
||
|
||
setp.gt.s32 %p100, %r1649, 1;
|
||
@%p100 bra BB1_123;
|
||
|
||
setp.eq.s32 %p103, %r1649, 0;
|
||
@%p103 bra BB1_155;
|
||
bra.uni BB1_121;
|
||
|
||
BB1_155:
|
||
mov.u32 %r1648, 0;
|
||
mov.u32 %r1647, %r1648;
|
||
mov.u32 %r1646, %r1648;
|
||
mov.u32 %r1645, %r1648;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_133:
|
||
setp.gt.s32 %p83, %r1649, 11;
|
||
@%p83 bra BB1_141;
|
||
|
||
setp.gt.s32 %p89, %r1649, 9;
|
||
@%p89 bra BB1_138;
|
||
|
||
setp.eq.s32 %p92, %r1649, 8;
|
||
@%p92 bra BB1_152;
|
||
|
||
setp.eq.s32 %p93, %r1649, 9;
|
||
@%p93 bra BB1_137;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_137:
|
||
and.b32 %r1646, %r1646, 255;
|
||
mov.u32 %r1645, 0;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_126:
|
||
setp.gt.s32 %p95, %r1649, 5;
|
||
@%p95 bra BB1_130;
|
||
|
||
setp.eq.s32 %p98, %r1649, 4;
|
||
@%p98 bra BB1_154;
|
||
|
||
setp.eq.s32 %p99, %r1649, 5;
|
||
@%p99 bra BB1_129;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_129:
|
||
and.b32 %r1647, %r1647, 255;
|
||
bra.uni BB1_152;
|
||
|
||
BB1_141:
|
||
setp.gt.s32 %p84, %r1649, 13;
|
||
@%p84 bra BB1_145;
|
||
|
||
setp.eq.s32 %p87, %r1649, 12;
|
||
@%p87 bra BB1_149;
|
||
bra.uni BB1_143;
|
||
|
||
BB1_149:
|
||
mov.u32 %r1645, 0;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_123:
|
||
setp.eq.s32 %p101, %r1649, 2;
|
||
@%p101 bra BB1_153;
|
||
bra.uni BB1_124;
|
||
|
||
BB1_153:
|
||
and.b32 %r1648, %r1648, 65535;
|
||
bra.uni BB1_154;
|
||
|
||
BB1_138:
|
||
setp.eq.s32 %p90, %r1649, 10;
|
||
@%p90 bra BB1_150;
|
||
bra.uni BB1_139;
|
||
|
||
BB1_150:
|
||
and.b32 %r1646, %r1646, 65535;
|
||
mov.u32 %r1645, 0;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_130:
|
||
setp.eq.s32 %p96, %r1649, 6;
|
||
@%p96 bra BB1_151;
|
||
bra.uni BB1_131;
|
||
|
||
BB1_151:
|
||
and.b32 %r1647, %r1647, 65535;
|
||
bra.uni BB1_152;
|
||
|
||
BB1_145:
|
||
setp.eq.s32 %p85, %r1649, 14;
|
||
@%p85 bra BB1_148;
|
||
bra.uni BB1_146;
|
||
|
||
BB1_148:
|
||
and.b32 %r1645, %r1645, 65535;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_121:
|
||
setp.eq.s32 %p104, %r1649, 1;
|
||
@%p104 bra BB1_122;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_122:
|
||
and.b32 %r1648, %r1648, 255;
|
||
bra.uni BB1_154;
|
||
|
||
BB1_143:
|
||
setp.eq.s32 %p88, %r1649, 13;
|
||
@%p88 bra BB1_144;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_144:
|
||
and.b32 %r1645, %r1645, 255;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_124:
|
||
setp.eq.s32 %p102, %r1649, 3;
|
||
@%p102 bra BB1_125;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_125:
|
||
and.b32 %r1648, %r1648, 16777215;
|
||
|
||
BB1_154:
|
||
mov.u32 %r1647, 0;
|
||
mov.u32 %r1646, %r1647;
|
||
mov.u32 %r1645, %r1647;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_139:
|
||
setp.eq.s32 %p91, %r1649, 11;
|
||
@%p91 bra BB1_140;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_140:
|
||
and.b32 %r1646, %r1646, 16777215;
|
||
mov.u32 %r1645, 0;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_131:
|
||
setp.eq.s32 %p97, %r1649, 7;
|
||
@%p97 bra BB1_132;
|
||
bra.uni BB1_156;
|
||
|
||
BB1_132:
|
||
and.b32 %r1647, %r1647, 16777215;
|
||
|
||
BB1_152:
|
||
mov.u32 %r1646, 0;
|
||
mov.u32 %r1645, %r1646;
|
||
|
||
BB1_156:
|
||
mov.u32 %r1637, 1985229328;
|
||
mov.u32 %r1636, 4;
|
||
and.b32 %r1108, %r105, 3;
|
||
sub.s32 %r1110, %r1636, %r1108;
|
||
shl.b32 %r1111, %r1110, 2;
|
||
shr.u32 %r1113, %r1637, %r1111;
|
||
and.b32 %r1106, %r1113, 65535;
|
||
mov.u32 %r1105, 0;
|
||
// inline asm
|
||
prmt.b32 %r1087, %r1105, %r1648, %r1106;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r1091, %r1648, %r1647, %r1106;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r1095, %r1647, %r1646, %r1106;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r1099, %r1646, %r1645, %r1106;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r1103, %r1645, %r1105, %r1106;
|
||
// inline asm
|
||
shr.u32 %r1107, %r105, 2;
|
||
setp.gt.s32 %p105, %r1107, 4;
|
||
@%p105 bra BB1_165;
|
||
|
||
setp.gt.s32 %p112, %r1107, 1;
|
||
@%p112 bra BB1_161;
|
||
|
||
setp.eq.s32 %p116, %r1107, 0;
|
||
@%p116 bra BB1_178;
|
||
bra.uni BB1_159;
|
||
|
||
BB1_178:
|
||
or.b32 %r3501, %r1087, %r94;
|
||
mov.u32 %r3541, %r95;
|
||
mov.u32 %r3540, %r96;
|
||
mov.u32 %r3539, %r97;
|
||
mov.u32 %r2236, %r99;
|
||
mov.u32 %r2249, %r100;
|
||
mov.u32 %r3535, %r101;
|
||
mov.u32 %r3534, %r102;
|
||
mov.u32 %r2289, %r103;
|
||
mov.u32 %r2301, %r104;
|
||
mov.u32 %r3542, %r1091;
|
||
mov.u32 %r3543, %r1095;
|
||
mov.u32 %r3544, %r1099;
|
||
mov.u32 %r3538, %r1103;
|
||
mov.u32 %r3532, %r2301;
|
||
mov.u32 %r3533, %r2289;
|
||
mov.u32 %r3536, %r2249;
|
||
mov.u32 %r3537, %r2236;
|
||
bra.uni BB1_179;
|
||
|
||
BB1_165:
|
||
setp.gt.s32 %p106, %r1107, 6;
|
||
@%p106 bra BB1_169;
|
||
|
||
setp.eq.s32 %p110, %r1107, 5;
|
||
@%p110 bra BB1_175;
|
||
bra.uni BB1_167;
|
||
|
||
BB1_175:
|
||
or.b32 %r3539, %r1087, %r97;
|
||
mov.u32 %r3544, %r91;
|
||
mov.u32 %r3543, %r92;
|
||
mov.u32 %r3542, %r93;
|
||
mov.u32 %r3501, %r94;
|
||
mov.u32 %r3538, %r98;
|
||
mov.u32 %r2231, %r99;
|
||
mov.u32 %r2244, %r100;
|
||
mov.u32 %r2284, %r103;
|
||
mov.u32 %r2296, %r104;
|
||
mov.u32 %r3540, %r1091;
|
||
mov.u32 %r3541, %r1095;
|
||
mov.u32 %r3534, %r1099;
|
||
mov.u32 %r3535, %r1103;
|
||
mov.u32 %r3532, %r2296;
|
||
mov.u32 %r3533, %r2284;
|
||
mov.u32 %r3536, %r2244;
|
||
mov.u32 %r3537, %r2231;
|
||
bra.uni BB1_179;
|
||
|
||
BB1_161:
|
||
setp.eq.s32 %p113, %r1107, 2;
|
||
@%p113 bra BB1_177;
|
||
|
||
setp.eq.s32 %p114, %r1107, 3;
|
||
@%p114 bra BB1_176;
|
||
bra.uni BB1_163;
|
||
|
||
BB1_176:
|
||
or.b32 %r3544, %r1087, %r91;
|
||
mov.u32 %r3543, %r92;
|
||
mov.u32 %r3542, %r93;
|
||
mov.u32 %r3501, %r94;
|
||
mov.u32 %r2233, %r99;
|
||
mov.u32 %r2246, %r100;
|
||
mov.u32 %r3535, %r101;
|
||
mov.u32 %r3534, %r102;
|
||
mov.u32 %r2286, %r103;
|
||
mov.u32 %r2298, %r104;
|
||
mov.u32 %r3538, %r1091;
|
||
mov.u32 %r3539, %r1095;
|
||
mov.u32 %r3540, %r1099;
|
||
mov.u32 %r3541, %r1103;
|
||
mov.u32 %r3532, %r2298;
|
||
mov.u32 %r3533, %r2286;
|
||
mov.u32 %r3536, %r2246;
|
||
mov.u32 %r3537, %r2233;
|
||
bra.uni BB1_179;
|
||
|
||
BB1_169:
|
||
setp.eq.s32 %p107, %r1107, 7;
|
||
@%p107 bra BB1_174;
|
||
|
||
setp.eq.s32 %p108, %r1107, 8;
|
||
@%p108 bra BB1_173;
|
||
bra.uni BB1_171;
|
||
|
||
BB1_173:
|
||
or.b32 %r3534, %r1087, %r102;
|
||
mov.u32 %r3544, %r91;
|
||
mov.u32 %r3543, %r92;
|
||
mov.u32 %r3542, %r93;
|
||
mov.u32 %r3501, %r94;
|
||
mov.u32 %r3541, %r95;
|
||
mov.u32 %r3540, %r96;
|
||
mov.u32 %r3539, %r97;
|
||
mov.u32 %r3538, %r98;
|
||
mov.u32 %r2281, %r103;
|
||
mov.u32 %r3535, %r1091;
|
||
mov.u32 %r3536, %r1095;
|
||
mov.u32 %r3537, %r1099;
|
||
mov.u32 %r3532, %r1103;
|
||
mov.u32 %r3533, %r2281;
|
||
bra.uni BB1_179;
|
||
|
||
BB1_159:
|
||
setp.eq.s32 %p117, %r1107, 1;
|
||
mov.u32 %r3544, %r91;
|
||
mov.u32 %r3543, %r92;
|
||
mov.u32 %r2150, %r93;
|
||
mov.u32 %r2162, %r94;
|
||
mov.u32 %r2184, %r95;
|
||
mov.u32 %r2196, %r96;
|
||
mov.u32 %r3539, %r97;
|
||
mov.u32 %r3538, %r98;
|
||
mov.u32 %r2229, %r99;
|
||
mov.u32 %r2243, %r100;
|
||
mov.u32 %r2256, %r101;
|
||
mov.u32 %r2268, %r102;
|
||
mov.u32 %r2280, %r103;
|
||
mov.u32 %r2293, %r104;
|
||
mov.u32 %r3501, %r2162;
|
||
mov.u32 %r3532, %r2293;
|
||
mov.u32 %r3533, %r2280;
|
||
mov.u32 %r3534, %r2268;
|
||
mov.u32 %r3535, %r2256;
|
||
mov.u32 %r3536, %r2243;
|
||
mov.u32 %r3537, %r2229;
|
||
mov.u32 %r3540, %r2196;
|
||
mov.u32 %r3541, %r2184;
|
||
mov.u32 %r3542, %r2150;
|
||
@%p117 bra BB1_160;
|
||
bra.uni BB1_179;
|
||
|
||
BB1_160:
|
||
or.b32 %r3542, %r1087, %r93;
|
||
mov.u32 %r3501, %r94;
|
||
mov.u32 %r3541, %r95;
|
||
mov.u32 %r3540, %r96;
|
||
mov.u32 %r2235, %r99;
|
||
mov.u32 %r2248, %r100;
|
||
mov.u32 %r3535, %r101;
|
||
mov.u32 %r3534, %r102;
|
||
mov.u32 %r2288, %r103;
|
||
mov.u32 %r2300, %r104;
|
||
mov.u32 %r3543, %r1091;
|
||
mov.u32 %r3544, %r1095;
|
||
mov.u32 %r3538, %r1099;
|
||
mov.u32 %r3539, %r1103;
|
||
mov.u32 %r3532, %r2300;
|
||
mov.u32 %r3533, %r2288;
|
||
mov.u32 %r3536, %r2248;
|
||
mov.u32 %r3537, %r2235;
|
||
bra.uni BB1_179;
|
||
|
||
BB1_167:
|
||
setp.eq.s32 %p111, %r1107, 6;
|
||
mov.u32 %r2127, %r91;
|
||
mov.u32 %r2137, %r92;
|
||
mov.u32 %r2148, %r93;
|
||
mov.u32 %r2160, %r94;
|
||
mov.u32 %r3541, %r95;
|
||
mov.u32 %r2194, %r96;
|
||
mov.u32 %r2206, %r97;
|
||
mov.u32 %r2215, %r98;
|
||
mov.u32 %r2227, %r99;
|
||
mov.u32 %r2241, %r100;
|
||
mov.u32 %r3535, %r101;
|
||
mov.u32 %r3534, %r102;
|
||
mov.u32 %r2278, %r103;
|
||
mov.u32 %r2291, %r104;
|
||
mov.u32 %r3501, %r2160;
|
||
mov.u32 %r3532, %r2291;
|
||
mov.u32 %r3533, %r2278;
|
||
mov.u32 %r3536, %r2241;
|
||
mov.u32 %r3537, %r2227;
|
||
mov.u32 %r3538, %r2215;
|
||
mov.u32 %r3539, %r2206;
|
||
mov.u32 %r3540, %r2194;
|
||
mov.u32 %r3542, %r2148;
|
||
mov.u32 %r3543, %r2137;
|
||
mov.u32 %r3544, %r2127;
|
||
@%p111 bra BB1_168;
|
||
bra.uni BB1_179;
|
||
|
||
BB1_168:
|
||
or.b32 %r3540, %r1087, %r96;
|
||
mov.u32 %r3544, %r91;
|
||
mov.u32 %r3543, %r92;
|
||
mov.u32 %r3542, %r93;
|
||
mov.u32 %r3501, %r94;
|
||
mov.u32 %r3539, %r97;
|
||
mov.u32 %r3538, %r98;
|
||
mov.u32 %r2230, %r99;
|
||
mov.u32 %r2283, %r103;
|
||
mov.u32 %r2295, %r104;
|
||
mov.u32 %r3541, %r1091;
|
||
mov.u32 %r3534, %r1095;
|
||
mov.u32 %r3535, %r1099;
|
||
mov.u32 %r3536, %r1103;
|
||
mov.u32 %r3532, %r2295;
|
||
mov.u32 %r3533, %r2283;
|
||
mov.u32 %r3537, %r2230;
|
||
bra.uni BB1_179;
|
||
|
||
BB1_177:
|
||
or.b32 %r3543, %r1087, %r92;
|
||
mov.u32 %r3542, %r93;
|
||
mov.u32 %r3501, %r94;
|
||
mov.u32 %r3541, %r95;
|
||
mov.u32 %r2234, %r99;
|
||
mov.u32 %r2247, %r100;
|
||
mov.u32 %r3535, %r101;
|
||
mov.u32 %r3534, %r102;
|
||
mov.u32 %r2287, %r103;
|
||
mov.u32 %r2299, %r104;
|
||
mov.u32 %r3544, %r1091;
|
||
mov.u32 %r3538, %r1095;
|
||
mov.u32 %r3539, %r1099;
|
||
mov.u32 %r3540, %r1103;
|
||
mov.u32 %r3532, %r2299;
|
||
mov.u32 %r3533, %r2287;
|
||
mov.u32 %r3536, %r2247;
|
||
mov.u32 %r3537, %r2234;
|
||
bra.uni BB1_179;
|
||
|
||
BB1_163:
|
||
setp.eq.s32 %p115, %r1107, 4;
|
||
mov.u32 %r2128, %r91;
|
||
mov.u32 %r2138, %r92;
|
||
mov.u32 %r2149, %r93;
|
||
mov.u32 %r2161, %r94;
|
||
mov.u32 %r3541, %r95;
|
||
mov.u32 %r3540, %r96;
|
||
mov.u32 %r3539, %r97;
|
||
mov.u32 %r2216, %r98;
|
||
mov.u32 %r2228, %r99;
|
||
mov.u32 %r2242, %r100;
|
||
mov.u32 %r2255, %r101;
|
||
mov.u32 %r3534, %r102;
|
||
mov.u32 %r2279, %r103;
|
||
mov.u32 %r2292, %r104;
|
||
mov.u32 %r3501, %r2161;
|
||
mov.u32 %r3532, %r2292;
|
||
mov.u32 %r3533, %r2279;
|
||
mov.u32 %r3535, %r2255;
|
||
mov.u32 %r3536, %r2242;
|
||
mov.u32 %r3537, %r2228;
|
||
mov.u32 %r3538, %r2216;
|
||
mov.u32 %r3542, %r2149;
|
||
mov.u32 %r3543, %r2138;
|
||
mov.u32 %r3544, %r2128;
|
||
@%p115 bra BB1_164;
|
||
bra.uni BB1_179;
|
||
|
||
BB1_164:
|
||
or.b32 %r3538, %r1087, %r98;
|
||
mov.u32 %r3544, %r91;
|
||
mov.u32 %r3543, %r92;
|
||
mov.u32 %r3542, %r93;
|
||
mov.u32 %r3501, %r94;
|
||
mov.u32 %r2232, %r99;
|
||
mov.u32 %r2245, %r100;
|
||
mov.u32 %r3535, %r101;
|
||
mov.u32 %r2285, %r103;
|
||
mov.u32 %r2297, %r104;
|
||
mov.u32 %r3539, %r1091;
|
||
mov.u32 %r3540, %r1095;
|
||
mov.u32 %r3541, %r1099;
|
||
mov.u32 %r3534, %r1103;
|
||
mov.u32 %r3532, %r2297;
|
||
mov.u32 %r3533, %r2285;
|
||
mov.u32 %r3536, %r2245;
|
||
mov.u32 %r3537, %r2232;
|
||
bra.uni BB1_179;
|
||
|
||
BB1_174:
|
||
or.b32 %r3541, %r1087, %r95;
|
||
mov.u32 %r3544, %r91;
|
||
mov.u32 %r3543, %r92;
|
||
mov.u32 %r3542, %r93;
|
||
mov.u32 %r3501, %r94;
|
||
mov.u32 %r3540, %r96;
|
||
mov.u32 %r3539, %r97;
|
||
mov.u32 %r3538, %r98;
|
||
mov.u32 %r2282, %r103;
|
||
mov.u32 %r2294, %r104;
|
||
mov.u32 %r3534, %r1091;
|
||
mov.u32 %r3535, %r1095;
|
||
mov.u32 %r3536, %r1099;
|
||
mov.u32 %r3537, %r1103;
|
||
mov.u32 %r3532, %r2294;
|
||
mov.u32 %r3533, %r2282;
|
||
bra.uni BB1_179;
|
||
|
||
BB1_171:
|
||
setp.ne.s32 %p109, %r1107, 9;
|
||
mov.u32 %r3501, %r94;
|
||
mov.u32 %r3532, %r104;
|
||
mov.u32 %r3533, %r103;
|
||
mov.u32 %r3534, %r102;
|
||
mov.u32 %r3535, %r101;
|
||
mov.u32 %r3536, %r100;
|
||
mov.u32 %r3537, %r99;
|
||
mov.u32 %r3538, %r98;
|
||
mov.u32 %r3539, %r97;
|
||
mov.u32 %r3540, %r96;
|
||
mov.u32 %r3541, %r95;
|
||
mov.u32 %r3542, %r93;
|
||
mov.u32 %r3543, %r92;
|
||
mov.u32 %r3544, %r91;
|
||
@%p109 bra BB1_179;
|
||
|
||
or.b32 %r3535, %r1087, %r101;
|
||
mov.u32 %r3544, %r91;
|
||
mov.u32 %r3543, %r92;
|
||
mov.u32 %r3542, %r93;
|
||
mov.u32 %r3501, %r94;
|
||
mov.u32 %r3541, %r95;
|
||
mov.u32 %r3540, %r96;
|
||
mov.u32 %r3539, %r97;
|
||
mov.u32 %r3538, %r98;
|
||
mov.u32 %r3534, %r102;
|
||
mov.u32 %r3532, %r1099;
|
||
mov.u32 %r3533, %r1103;
|
||
mov.u32 %r3536, %r1091;
|
||
mov.u32 %r3537, %r1095;
|
||
|
||
BB1_179:
|
||
mov.u32 %r3498, %r3532;
|
||
mov.u32 %r3497, %r3533;
|
||
mov.u32 %r3496, %r3534;
|
||
mov.u32 %r3495, %r3535;
|
||
mov.u32 %r3494, %r3536;
|
||
mov.u32 %r3493, %r3537;
|
||
mov.u32 %r3492, %r3538;
|
||
mov.u32 %r3491, %r3539;
|
||
mov.u32 %r3490, %r3540;
|
||
mov.u32 %r3489, %r3541;
|
||
mov.u32 %r3488, %r3501;
|
||
mov.u32 %r3487, %r3542;
|
||
mov.u32 %r3486, %r3543;
|
||
mov.u32 %r3485, %r3544;
|
||
add.s32 %r1650, %r105, %r1649;
|
||
and.b32 %r152, %r3488, 255;
|
||
setp.eq.s32 %p118, %r1649, 0;
|
||
@%p118 bra BB1_311;
|
||
|
||
shl.b32 %r153, %r152, 8;
|
||
shl.b32 %r154, %r152, 16;
|
||
shl.b32 %r155, %r3488, 24;
|
||
mov.u32 %r3506, %r3485;
|
||
mov.u32 %r3508, %r3486;
|
||
mov.u32 %r3510, %r3487;
|
||
mov.u32 %r3499, %r3488;
|
||
mov.u32 %r3512, %r3489;
|
||
mov.u32 %r3514, %r3490;
|
||
mov.u32 %r3516, %r3491;
|
||
mov.u32 %r3518, %r3492;
|
||
mov.u32 %r3520, %r3493;
|
||
mov.u32 %r3522, %r3494;
|
||
mov.u32 %r3524, %r3495;
|
||
mov.u32 %r3526, %r3496;
|
||
mov.u32 %r3528, %r3497;
|
||
mov.u32 %r3530, %r3498;
|
||
|
||
BB1_181:
|
||
mov.u32 %r3406, %r3530;
|
||
mov.u32 %r3327, %r3528;
|
||
mov.u32 %r3248, %r3526;
|
||
mov.u32 %r3169, %r3524;
|
||
mov.u32 %r3090, %r3522;
|
||
mov.u32 %r3011, %r3520;
|
||
mov.u32 %r2932, %r3518;
|
||
mov.u32 %r2853, %r3516;
|
||
mov.u32 %r2774, %r3514;
|
||
mov.u32 %r2695, %r3512;
|
||
mov.u32 %r2616, %r3510;
|
||
mov.u32 %r2537, %r3508;
|
||
mov.u32 %r2458, %r3506;
|
||
mov.u32 %r2379, %r3499;
|
||
mov.u32 %r158, %r3406;
|
||
mov.u32 %r159, %r3327;
|
||
mov.u32 %r160, %r3248;
|
||
mov.u32 %r161, %r3169;
|
||
mov.u32 %r162, %r3090;
|
||
mov.u32 %r163, %r3011;
|
||
mov.u32 %r164, %r2932;
|
||
mov.u32 %r165, %r2853;
|
||
mov.u32 %r166, %r2774;
|
||
mov.u32 %r167, %r2695;
|
||
mov.u32 %r169, %r2616;
|
||
mov.u32 %r170, %r2537;
|
||
mov.u32 %r171, %r2458;
|
||
mov.u32 %r168, %r2379;
|
||
and.b32 %r1114, %r1649, 1;
|
||
setp.eq.b32 %p119, %r1114, 1;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3529, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
@%p119 bra BB1_310;
|
||
|
||
setp.gt.s32 %p120, %r1650, 27;
|
||
@%p120 bra BB1_234;
|
||
|
||
setp.gt.s32 %p160, %r1650, 13;
|
||
@%p160 bra BB1_209;
|
||
|
||
setp.gt.s32 %p180, %r1650, 6;
|
||
@%p180 bra BB1_197;
|
||
|
||
setp.gt.s32 %p190, %r1650, 2;
|
||
@%p190 bra BB1_190;
|
||
|
||
setp.eq.s32 %p196, %r1650, 0;
|
||
@%p196 bra BB1_309;
|
||
|
||
setp.eq.s32 %p197, %r1650, 1;
|
||
@%p197 bra BB1_308;
|
||
bra.uni BB1_188;
|
||
|
||
BB1_308:
|
||
or.b32 %r3500, %r168, %r153;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3402, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3402;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_234:
|
||
setp.gt.s32 %p121, %r1650, 41;
|
||
@%p121 bra BB1_260;
|
||
|
||
setp.gt.s32 %p141, %r1650, 34;
|
||
@%p141 bra BB1_248;
|
||
|
||
setp.gt.s32 %p151, %r1650, 30;
|
||
@%p151 bra BB1_241;
|
||
|
||
setp.eq.s32 %p157, %r1650, 28;
|
||
mov.u32 %r2371, %r152;
|
||
mov.u32 %r2431, %r168;
|
||
mov.u32 %r2510, %r171;
|
||
mov.u32 %r2589, %r170;
|
||
mov.u32 %r2668, %r169;
|
||
mov.u32 %r2826, %r166;
|
||
mov.u32 %r2905, %r165;
|
||
mov.u32 %r2984, %r164;
|
||
mov.u32 %r3059, %r163;
|
||
mov.u32 %r3138, %r162;
|
||
mov.u32 %r3217, %r161;
|
||
mov.u32 %r3296, %r160;
|
||
mov.u32 %r3375, %r159;
|
||
mov.u32 %r3454, %r158;
|
||
mov.u32 %r3500, %r2431;
|
||
mov.u32 %r3507, %r2510;
|
||
mov.u32 %r3509, %r2589;
|
||
mov.u32 %r3511, %r2668;
|
||
mov.u32 %r3513, %r2371;
|
||
mov.u32 %r3515, %r2826;
|
||
mov.u32 %r3517, %r2905;
|
||
mov.u32 %r3519, %r2984;
|
||
mov.u32 %r3521, %r3059;
|
||
mov.u32 %r3523, %r3138;
|
||
mov.u32 %r3525, %r3217;
|
||
mov.u32 %r3527, %r3296;
|
||
mov.u32 %r3529, %r3375;
|
||
mov.u32 %r3531, %r3454;
|
||
@%p157 bra BB1_310;
|
||
|
||
setp.eq.s32 %p158, %r1650, 29;
|
||
@%p158 bra BB1_296;
|
||
bra.uni BB1_239;
|
||
|
||
BB1_296:
|
||
or.b32 %r3513, %r167, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3374, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3374;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_209:
|
||
setp.gt.s32 %p161, %r1650, 20;
|
||
@%p161 bra BB1_222;
|
||
|
||
setp.gt.s32 %p171, %r1650, 16;
|
||
@%p171 bra BB1_215;
|
||
|
||
setp.eq.s32 %p177, %r1650, 14;
|
||
@%p177 bra BB1_303;
|
||
|
||
setp.eq.s32 %p178, %r1650, 15;
|
||
@%p178 bra BB1_302;
|
||
bra.uni BB1_213;
|
||
|
||
BB1_302:
|
||
or.b32 %r3507, %r171, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3388, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3388;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_260:
|
||
setp.gt.s32 %p122, %r1650, 48;
|
||
@%p122 bra BB1_273;
|
||
|
||
setp.gt.s32 %p132, %r1650, 44;
|
||
@%p132 bra BB1_266;
|
||
|
||
setp.eq.s32 %p138, %r1650, 42;
|
||
@%p138 bra BB1_291;
|
||
|
||
setp.eq.s32 %p139, %r1650, 43;
|
||
@%p139 bra BB1_290;
|
||
bra.uni BB1_264;
|
||
|
||
BB1_290:
|
||
or.b32 %r3523, %r162, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3360, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3360;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_197:
|
||
setp.gt.s32 %p181, %r1650, 9;
|
||
@%p181 bra BB1_202;
|
||
|
||
setp.eq.s32 %p187, %r1650, 7;
|
||
@%p187 bra BB1_305;
|
||
|
||
setp.eq.s32 %p188, %r1650, 8;
|
||
mov.u32 %r2376, %r152;
|
||
mov.u32 %r2451, %r168;
|
||
mov.u32 %r2526, %r171;
|
||
mov.u32 %r2688, %r169;
|
||
mov.u32 %r2763, %r167;
|
||
mov.u32 %r2842, %r166;
|
||
mov.u32 %r2921, %r165;
|
||
mov.u32 %r3000, %r164;
|
||
mov.u32 %r3079, %r163;
|
||
mov.u32 %r3158, %r162;
|
||
mov.u32 %r3237, %r161;
|
||
mov.u32 %r3316, %r160;
|
||
mov.u32 %r3395, %r159;
|
||
mov.u32 %r3474, %r158;
|
||
mov.u32 %r3500, %r2451;
|
||
mov.u32 %r3507, %r2526;
|
||
mov.u32 %r3509, %r2376;
|
||
mov.u32 %r3511, %r2688;
|
||
mov.u32 %r3513, %r2763;
|
||
mov.u32 %r3515, %r2842;
|
||
mov.u32 %r3517, %r2921;
|
||
mov.u32 %r3519, %r3000;
|
||
mov.u32 %r3521, %r3079;
|
||
mov.u32 %r3523, %r3158;
|
||
mov.u32 %r3525, %r3237;
|
||
mov.u32 %r3527, %r3316;
|
||
mov.u32 %r3529, %r3395;
|
||
mov.u32 %r3531, %r3474;
|
||
@%p188 bra BB1_310;
|
||
|
||
setp.eq.s32 %p189, %r1650, 9;
|
||
mov.u32 %r2400, %r168;
|
||
mov.u32 %r2479, %r171;
|
||
mov.u32 %r2558, %r170;
|
||
mov.u32 %r2637, %r169;
|
||
mov.u32 %r2716, %r167;
|
||
mov.u32 %r2795, %r166;
|
||
mov.u32 %r2874, %r165;
|
||
mov.u32 %r2953, %r164;
|
||
mov.u32 %r3032, %r163;
|
||
mov.u32 %r3111, %r162;
|
||
mov.u32 %r3190, %r161;
|
||
mov.u32 %r3269, %r160;
|
||
mov.u32 %r3348, %r159;
|
||
mov.u32 %r3427, %r158;
|
||
mov.u32 %r3500, %r2400;
|
||
mov.u32 %r3507, %r2479;
|
||
mov.u32 %r3509, %r2558;
|
||
mov.u32 %r3511, %r2637;
|
||
mov.u32 %r3513, %r2716;
|
||
mov.u32 %r3515, %r2795;
|
||
mov.u32 %r3517, %r2874;
|
||
mov.u32 %r3519, %r2953;
|
||
mov.u32 %r3521, %r3032;
|
||
mov.u32 %r3523, %r3111;
|
||
mov.u32 %r3525, %r3190;
|
||
mov.u32 %r3527, %r3269;
|
||
mov.u32 %r3529, %r3348;
|
||
mov.u32 %r3531, %r3427;
|
||
@%p189 bra BB1_201;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_201:
|
||
or.b32 %r3509, %r170, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3394, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3394;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_248:
|
||
setp.gt.s32 %p142, %r1650, 37;
|
||
@%p142 bra BB1_253;
|
||
|
||
setp.eq.s32 %p148, %r1650, 35;
|
||
@%p148 bra BB1_293;
|
||
|
||
setp.eq.s32 %p149, %r1650, 36;
|
||
mov.u32 %r2369, %r152;
|
||
mov.u32 %r2423, %r168;
|
||
mov.u32 %r2502, %r171;
|
||
mov.u32 %r2581, %r170;
|
||
mov.u32 %r2660, %r169;
|
||
mov.u32 %r2739, %r167;
|
||
mov.u32 %r2818, %r166;
|
||
mov.u32 %r2897, %r165;
|
||
mov.u32 %r2976, %r164;
|
||
mov.u32 %r3051, %r163;
|
||
mov.u32 %r3130, %r162;
|
||
mov.u32 %r3292, %r160;
|
||
mov.u32 %r3367, %r159;
|
||
mov.u32 %r3446, %r158;
|
||
mov.u32 %r3500, %r2423;
|
||
mov.u32 %r3507, %r2502;
|
||
mov.u32 %r3509, %r2581;
|
||
mov.u32 %r3511, %r2660;
|
||
mov.u32 %r3513, %r2739;
|
||
mov.u32 %r3515, %r2818;
|
||
mov.u32 %r3517, %r2897;
|
||
mov.u32 %r3519, %r2976;
|
||
mov.u32 %r3521, %r3051;
|
||
mov.u32 %r3523, %r3130;
|
||
mov.u32 %r3525, %r2369;
|
||
mov.u32 %r3527, %r3292;
|
||
mov.u32 %r3529, %r3367;
|
||
mov.u32 %r3531, %r3446;
|
||
@%p149 bra BB1_310;
|
||
|
||
setp.eq.s32 %p150, %r1650, 37;
|
||
mov.u32 %r2388, %r168;
|
||
mov.u32 %r2467, %r171;
|
||
mov.u32 %r2546, %r170;
|
||
mov.u32 %r2625, %r169;
|
||
mov.u32 %r2704, %r167;
|
||
mov.u32 %r2783, %r166;
|
||
mov.u32 %r2862, %r165;
|
||
mov.u32 %r2941, %r164;
|
||
mov.u32 %r3020, %r163;
|
||
mov.u32 %r3099, %r162;
|
||
mov.u32 %r3178, %r161;
|
||
mov.u32 %r3257, %r160;
|
||
mov.u32 %r3336, %r159;
|
||
mov.u32 %r3415, %r158;
|
||
mov.u32 %r3500, %r2388;
|
||
mov.u32 %r3507, %r2467;
|
||
mov.u32 %r3509, %r2546;
|
||
mov.u32 %r3511, %r2625;
|
||
mov.u32 %r3513, %r2704;
|
||
mov.u32 %r3515, %r2783;
|
||
mov.u32 %r3517, %r2862;
|
||
mov.u32 %r3519, %r2941;
|
||
mov.u32 %r3521, %r3020;
|
||
mov.u32 %r3523, %r3099;
|
||
mov.u32 %r3525, %r3178;
|
||
mov.u32 %r3527, %r3257;
|
||
mov.u32 %r3529, %r3336;
|
||
mov.u32 %r3531, %r3415;
|
||
@%p150 bra BB1_252;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_252:
|
||
or.b32 %r3525, %r161, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3366, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3366;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_222:
|
||
setp.gt.s32 %p162, %r1650, 23;
|
||
@%p162 bra BB1_227;
|
||
|
||
setp.eq.s32 %p168, %r1650, 21;
|
||
@%p168 bra BB1_299;
|
||
|
||
setp.eq.s32 %p169, %r1650, 22;
|
||
@%p169 bra BB1_298;
|
||
bra.uni BB1_225;
|
||
|
||
BB1_298:
|
||
or.b32 %r3517, %r165, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3381, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3381;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_273:
|
||
setp.gt.s32 %p123, %r1650, 51;
|
||
@%p123 bra BB1_278;
|
||
|
||
setp.eq.s32 %p129, %r1650, 49;
|
||
@%p129 bra BB1_287;
|
||
|
||
setp.eq.s32 %p130, %r1650, 50;
|
||
@%p130 bra BB1_286;
|
||
bra.uni BB1_276;
|
||
|
||
BB1_286:
|
||
or.b32 %r3531, %r158, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3353, %r159;
|
||
mov.u32 %r3529, %r3353;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_190:
|
||
setp.gt.s32 %p191, %r1650, 4;
|
||
@%p191 bra BB1_194;
|
||
|
||
setp.eq.s32 %p194, %r1650, 3;
|
||
@%p194 bra BB1_307;
|
||
bra.uni BB1_192;
|
||
|
||
BB1_307:
|
||
or.b32 %r3500, %r168, %r155;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3400, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3400;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_241:
|
||
setp.gt.s32 %p152, %r1650, 32;
|
||
@%p152 bra BB1_245;
|
||
|
||
setp.eq.s32 %p155, %r1650, 31;
|
||
@%p155 bra BB1_295;
|
||
bra.uni BB1_243;
|
||
|
||
BB1_295:
|
||
or.b32 %r3513, %r167, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3372, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3372;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_215:
|
||
setp.gt.s32 %p172, %r1650, 18;
|
||
@%p172 bra BB1_219;
|
||
|
||
setp.eq.s32 %p175, %r1650, 17;
|
||
@%p175 bra BB1_301;
|
||
bra.uni BB1_217;
|
||
|
||
BB1_301:
|
||
or.b32 %r3519, %r164, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3386, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3386;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_266:
|
||
setp.gt.s32 %p133, %r1650, 46;
|
||
@%p133 bra BB1_270;
|
||
|
||
setp.eq.s32 %p136, %r1650, 45;
|
||
@%p136 bra BB1_289;
|
||
bra.uni BB1_268;
|
||
|
||
BB1_289:
|
||
or.b32 %r3521, %r163, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3358, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3358;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_202:
|
||
setp.gt.s32 %p182, %r1650, 11;
|
||
@%p182 bra BB1_206;
|
||
|
||
setp.eq.s32 %p185, %r1650, 10;
|
||
@%p185 bra BB1_304;
|
||
bra.uni BB1_204;
|
||
|
||
BB1_304:
|
||
or.b32 %r3509, %r170, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3393, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3393;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_253:
|
||
setp.gt.s32 %p143, %r1650, 39;
|
||
@%p143 bra BB1_257;
|
||
|
||
setp.eq.s32 %p146, %r1650, 38;
|
||
@%p146 bra BB1_292;
|
||
bra.uni BB1_255;
|
||
|
||
BB1_292:
|
||
or.b32 %r3525, %r161, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3365, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3365;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_227:
|
||
setp.gt.s32 %p163, %r1650, 25;
|
||
@%p163 bra BB1_231;
|
||
|
||
setp.eq.s32 %p166, %r1650, 24;
|
||
mov.u32 %r2372, %r152;
|
||
mov.u32 %r2435, %r168;
|
||
mov.u32 %r2514, %r171;
|
||
mov.u32 %r2593, %r170;
|
||
mov.u32 %r2672, %r169;
|
||
mov.u32 %r2747, %r167;
|
||
mov.u32 %r2909, %r165;
|
||
mov.u32 %r2988, %r164;
|
||
mov.u32 %r3063, %r163;
|
||
mov.u32 %r3142, %r162;
|
||
mov.u32 %r3221, %r161;
|
||
mov.u32 %r3300, %r160;
|
||
mov.u32 %r3379, %r159;
|
||
mov.u32 %r3458, %r158;
|
||
mov.u32 %r3500, %r2435;
|
||
mov.u32 %r3507, %r2514;
|
||
mov.u32 %r3509, %r2593;
|
||
mov.u32 %r3511, %r2672;
|
||
mov.u32 %r3513, %r2747;
|
||
mov.u32 %r3515, %r2372;
|
||
mov.u32 %r3517, %r2909;
|
||
mov.u32 %r3519, %r2988;
|
||
mov.u32 %r3521, %r3063;
|
||
mov.u32 %r3523, %r3142;
|
||
mov.u32 %r3525, %r3221;
|
||
mov.u32 %r3527, %r3300;
|
||
mov.u32 %r3529, %r3379;
|
||
mov.u32 %r3531, %r3458;
|
||
@%p166 bra BB1_310;
|
||
|
||
setp.eq.s32 %p167, %r1650, 25;
|
||
mov.u32 %r2393, %r168;
|
||
mov.u32 %r2472, %r171;
|
||
mov.u32 %r2551, %r170;
|
||
mov.u32 %r2630, %r169;
|
||
mov.u32 %r2709, %r167;
|
||
mov.u32 %r2788, %r166;
|
||
mov.u32 %r2867, %r165;
|
||
mov.u32 %r2946, %r164;
|
||
mov.u32 %r3025, %r163;
|
||
mov.u32 %r3104, %r162;
|
||
mov.u32 %r3183, %r161;
|
||
mov.u32 %r3262, %r160;
|
||
mov.u32 %r3341, %r159;
|
||
mov.u32 %r3420, %r158;
|
||
mov.u32 %r3500, %r2393;
|
||
mov.u32 %r3507, %r2472;
|
||
mov.u32 %r3509, %r2551;
|
||
mov.u32 %r3511, %r2630;
|
||
mov.u32 %r3513, %r2709;
|
||
mov.u32 %r3515, %r2788;
|
||
mov.u32 %r3517, %r2867;
|
||
mov.u32 %r3519, %r2946;
|
||
mov.u32 %r3521, %r3025;
|
||
mov.u32 %r3523, %r3104;
|
||
mov.u32 %r3525, %r3183;
|
||
mov.u32 %r3527, %r3262;
|
||
mov.u32 %r3529, %r3341;
|
||
mov.u32 %r3531, %r3420;
|
||
@%p167 bra BB1_230;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_230:
|
||
or.b32 %r3515, %r166, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3378, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3378;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_278:
|
||
setp.gt.s32 %p124, %r1650, 53;
|
||
@%p124 bra BB1_282;
|
||
|
||
setp.eq.s32 %p127, %r1650, 52;
|
||
mov.u32 %r2407, %r168;
|
||
mov.u32 %r2486, %r171;
|
||
mov.u32 %r2565, %r170;
|
||
mov.u32 %r2644, %r169;
|
||
mov.u32 %r2723, %r167;
|
||
mov.u32 %r2802, %r166;
|
||
mov.u32 %r2881, %r165;
|
||
mov.u32 %r2960, %r164;
|
||
mov.u32 %r3039, %r163;
|
||
mov.u32 %r3118, %r162;
|
||
mov.u32 %r3197, %r161;
|
||
mov.u32 %r3276, %r160;
|
||
mov.u32 %r3434, %r158;
|
||
mov.u32 %r3500, %r2407;
|
||
mov.u32 %r3507, %r2486;
|
||
mov.u32 %r3509, %r2565;
|
||
mov.u32 %r3511, %r2644;
|
||
mov.u32 %r3513, %r2723;
|
||
mov.u32 %r3515, %r2802;
|
||
mov.u32 %r3517, %r2881;
|
||
mov.u32 %r3519, %r2960;
|
||
mov.u32 %r3521, %r3039;
|
||
mov.u32 %r3523, %r3118;
|
||
mov.u32 %r3525, %r3197;
|
||
mov.u32 %r3527, %r3276;
|
||
mov.u32 %r3529, %r152;
|
||
mov.u32 %r3531, %r3434;
|
||
@%p127 bra BB1_310;
|
||
|
||
setp.eq.s32 %p128, %r1650, 53;
|
||
mov.u32 %r2381, %r168;
|
||
mov.u32 %r2460, %r171;
|
||
mov.u32 %r2539, %r170;
|
||
mov.u32 %r2618, %r169;
|
||
mov.u32 %r2697, %r167;
|
||
mov.u32 %r2776, %r166;
|
||
mov.u32 %r2855, %r165;
|
||
mov.u32 %r2934, %r164;
|
||
mov.u32 %r3013, %r163;
|
||
mov.u32 %r3092, %r162;
|
||
mov.u32 %r3171, %r161;
|
||
mov.u32 %r3250, %r160;
|
||
mov.u32 %r3329, %r159;
|
||
mov.u32 %r3408, %r158;
|
||
mov.u32 %r3500, %r2381;
|
||
mov.u32 %r3507, %r2460;
|
||
mov.u32 %r3509, %r2539;
|
||
mov.u32 %r3511, %r2618;
|
||
mov.u32 %r3513, %r2697;
|
||
mov.u32 %r3515, %r2776;
|
||
mov.u32 %r3517, %r2855;
|
||
mov.u32 %r3519, %r2934;
|
||
mov.u32 %r3521, %r3013;
|
||
mov.u32 %r3523, %r3092;
|
||
mov.u32 %r3525, %r3171;
|
||
mov.u32 %r3527, %r3250;
|
||
mov.u32 %r3529, %r3329;
|
||
mov.u32 %r3531, %r3408;
|
||
@%p128 bra BB1_281;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_281:
|
||
or.b32 %r174, %r159, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r174;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_194:
|
||
setp.eq.s32 %p192, %r1650, 5;
|
||
@%p192 bra BB1_306;
|
||
bra.uni BB1_195;
|
||
|
||
BB1_306:
|
||
or.b32 %r3511, %r169, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3398, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3398;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_245:
|
||
setp.eq.s32 %p153, %r1650, 33;
|
||
@%p153 bra BB1_294;
|
||
bra.uni BB1_246;
|
||
|
||
BB1_294:
|
||
or.b32 %r3527, %r160, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3370, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3370;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_219:
|
||
setp.eq.s32 %p173, %r1650, 19;
|
||
@%p173 bra BB1_300;
|
||
bra.uni BB1_220;
|
||
|
||
BB1_300:
|
||
or.b32 %r3519, %r164, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3384, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3384;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_270:
|
||
setp.eq.s32 %p134, %r1650, 47;
|
||
@%p134 bra BB1_288;
|
||
bra.uni BB1_271;
|
||
|
||
BB1_288:
|
||
or.b32 %r3521, %r163, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3356, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3356;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_206:
|
||
setp.eq.s32 %p183, %r1650, 12;
|
||
mov.u32 %r2375, %r152;
|
||
mov.u32 %r2447, %r168;
|
||
mov.u32 %r2605, %r170;
|
||
mov.u32 %r2684, %r169;
|
||
mov.u32 %r2759, %r167;
|
||
mov.u32 %r2838, %r166;
|
||
mov.u32 %r2917, %r165;
|
||
mov.u32 %r2996, %r164;
|
||
mov.u32 %r3075, %r163;
|
||
mov.u32 %r3154, %r162;
|
||
mov.u32 %r3233, %r161;
|
||
mov.u32 %r3312, %r160;
|
||
mov.u32 %r3391, %r159;
|
||
mov.u32 %r3470, %r158;
|
||
mov.u32 %r3500, %r2447;
|
||
mov.u32 %r3507, %r2375;
|
||
mov.u32 %r3509, %r2605;
|
||
mov.u32 %r3511, %r2684;
|
||
mov.u32 %r3513, %r2759;
|
||
mov.u32 %r3515, %r2838;
|
||
mov.u32 %r3517, %r2917;
|
||
mov.u32 %r3519, %r2996;
|
||
mov.u32 %r3521, %r3075;
|
||
mov.u32 %r3523, %r3154;
|
||
mov.u32 %r3525, %r3233;
|
||
mov.u32 %r3527, %r3312;
|
||
mov.u32 %r3529, %r3391;
|
||
mov.u32 %r3531, %r3470;
|
||
@%p183 bra BB1_310;
|
||
|
||
setp.eq.s32 %p184, %r1650, 13;
|
||
mov.u32 %r2398, %r168;
|
||
mov.u32 %r2477, %r171;
|
||
mov.u32 %r2556, %r170;
|
||
mov.u32 %r2635, %r169;
|
||
mov.u32 %r2714, %r167;
|
||
mov.u32 %r2793, %r166;
|
||
mov.u32 %r2872, %r165;
|
||
mov.u32 %r2951, %r164;
|
||
mov.u32 %r3030, %r163;
|
||
mov.u32 %r3109, %r162;
|
||
mov.u32 %r3188, %r161;
|
||
mov.u32 %r3267, %r160;
|
||
mov.u32 %r3346, %r159;
|
||
mov.u32 %r3425, %r158;
|
||
mov.u32 %r3500, %r2398;
|
||
mov.u32 %r3507, %r2477;
|
||
mov.u32 %r3509, %r2556;
|
||
mov.u32 %r3511, %r2635;
|
||
mov.u32 %r3513, %r2714;
|
||
mov.u32 %r3515, %r2793;
|
||
mov.u32 %r3517, %r2872;
|
||
mov.u32 %r3519, %r2951;
|
||
mov.u32 %r3521, %r3030;
|
||
mov.u32 %r3523, %r3109;
|
||
mov.u32 %r3525, %r3188;
|
||
mov.u32 %r3527, %r3267;
|
||
mov.u32 %r3529, %r3346;
|
||
mov.u32 %r3531, %r3425;
|
||
@%p184 bra BB1_208;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_208:
|
||
or.b32 %r3507, %r171, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3390, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3390;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_257:
|
||
setp.eq.s32 %p144, %r1650, 40;
|
||
mov.u32 %r2368, %r152;
|
||
mov.u32 %r2419, %r168;
|
||
mov.u32 %r2498, %r171;
|
||
mov.u32 %r2577, %r170;
|
||
mov.u32 %r2656, %r169;
|
||
mov.u32 %r2735, %r167;
|
||
mov.u32 %r2814, %r166;
|
||
mov.u32 %r2893, %r165;
|
||
mov.u32 %r2972, %r164;
|
||
mov.u32 %r3047, %r163;
|
||
mov.u32 %r3209, %r161;
|
||
mov.u32 %r3288, %r160;
|
||
mov.u32 %r3363, %r159;
|
||
mov.u32 %r3442, %r158;
|
||
mov.u32 %r3500, %r2419;
|
||
mov.u32 %r3507, %r2498;
|
||
mov.u32 %r3509, %r2577;
|
||
mov.u32 %r3511, %r2656;
|
||
mov.u32 %r3513, %r2735;
|
||
mov.u32 %r3515, %r2814;
|
||
mov.u32 %r3517, %r2893;
|
||
mov.u32 %r3519, %r2972;
|
||
mov.u32 %r3521, %r3047;
|
||
mov.u32 %r3523, %r2368;
|
||
mov.u32 %r3525, %r3209;
|
||
mov.u32 %r3527, %r3288;
|
||
mov.u32 %r3529, %r3363;
|
||
mov.u32 %r3531, %r3442;
|
||
@%p144 bra BB1_310;
|
||
|
||
setp.eq.s32 %p145, %r1650, 41;
|
||
mov.u32 %r2386, %r168;
|
||
mov.u32 %r2465, %r171;
|
||
mov.u32 %r2544, %r170;
|
||
mov.u32 %r2623, %r169;
|
||
mov.u32 %r2702, %r167;
|
||
mov.u32 %r2781, %r166;
|
||
mov.u32 %r2860, %r165;
|
||
mov.u32 %r2939, %r164;
|
||
mov.u32 %r3018, %r163;
|
||
mov.u32 %r3097, %r162;
|
||
mov.u32 %r3176, %r161;
|
||
mov.u32 %r3255, %r160;
|
||
mov.u32 %r3334, %r159;
|
||
mov.u32 %r3413, %r158;
|
||
mov.u32 %r3500, %r2386;
|
||
mov.u32 %r3507, %r2465;
|
||
mov.u32 %r3509, %r2544;
|
||
mov.u32 %r3511, %r2623;
|
||
mov.u32 %r3513, %r2702;
|
||
mov.u32 %r3515, %r2781;
|
||
mov.u32 %r3517, %r2860;
|
||
mov.u32 %r3519, %r2939;
|
||
mov.u32 %r3521, %r3018;
|
||
mov.u32 %r3523, %r3097;
|
||
mov.u32 %r3525, %r3176;
|
||
mov.u32 %r3527, %r3255;
|
||
mov.u32 %r3529, %r3334;
|
||
mov.u32 %r3531, %r3413;
|
||
@%p145 bra BB1_259;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_259:
|
||
or.b32 %r3523, %r162, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3362, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3362;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_231:
|
||
setp.eq.s32 %p164, %r1650, 26;
|
||
@%p164 bra BB1_297;
|
||
bra.uni BB1_232;
|
||
|
||
BB1_297:
|
||
or.b32 %r3515, %r166, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3377, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3377;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_282:
|
||
setp.eq.s32 %p125, %r1650, 54;
|
||
@%p125 bra BB1_285;
|
||
bra.uni BB1_283;
|
||
|
||
BB1_285:
|
||
or.b32 %r173, %r159, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r173;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_309:
|
||
mov.u32 %r3500, %r152;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3403, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3403;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_188:
|
||
setp.eq.s32 %p198, %r1650, 2;
|
||
mov.u32 %r2403, %r168;
|
||
mov.u32 %r2482, %r171;
|
||
mov.u32 %r2561, %r170;
|
||
mov.u32 %r2640, %r169;
|
||
mov.u32 %r2719, %r167;
|
||
mov.u32 %r2798, %r166;
|
||
mov.u32 %r2877, %r165;
|
||
mov.u32 %r2956, %r164;
|
||
mov.u32 %r3035, %r163;
|
||
mov.u32 %r3114, %r162;
|
||
mov.u32 %r3193, %r161;
|
||
mov.u32 %r3272, %r160;
|
||
mov.u32 %r3351, %r159;
|
||
mov.u32 %r3430, %r158;
|
||
mov.u32 %r3500, %r2403;
|
||
mov.u32 %r3507, %r2482;
|
||
mov.u32 %r3509, %r2561;
|
||
mov.u32 %r3511, %r2640;
|
||
mov.u32 %r3513, %r2719;
|
||
mov.u32 %r3515, %r2798;
|
||
mov.u32 %r3517, %r2877;
|
||
mov.u32 %r3519, %r2956;
|
||
mov.u32 %r3521, %r3035;
|
||
mov.u32 %r3523, %r3114;
|
||
mov.u32 %r3525, %r3193;
|
||
mov.u32 %r3527, %r3272;
|
||
mov.u32 %r3529, %r3351;
|
||
mov.u32 %r3531, %r3430;
|
||
@%p198 bra BB1_189;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_189:
|
||
or.b32 %r3500, %r168, %r154;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3401, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3401;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_239:
|
||
setp.eq.s32 %p159, %r1650, 30;
|
||
mov.u32 %r2391, %r168;
|
||
mov.u32 %r2470, %r171;
|
||
mov.u32 %r2549, %r170;
|
||
mov.u32 %r2628, %r169;
|
||
mov.u32 %r2707, %r167;
|
||
mov.u32 %r2786, %r166;
|
||
mov.u32 %r2865, %r165;
|
||
mov.u32 %r2944, %r164;
|
||
mov.u32 %r3023, %r163;
|
||
mov.u32 %r3102, %r162;
|
||
mov.u32 %r3181, %r161;
|
||
mov.u32 %r3260, %r160;
|
||
mov.u32 %r3339, %r159;
|
||
mov.u32 %r3418, %r158;
|
||
mov.u32 %r3500, %r2391;
|
||
mov.u32 %r3507, %r2470;
|
||
mov.u32 %r3509, %r2549;
|
||
mov.u32 %r3511, %r2628;
|
||
mov.u32 %r3513, %r2707;
|
||
mov.u32 %r3515, %r2786;
|
||
mov.u32 %r3517, %r2865;
|
||
mov.u32 %r3519, %r2944;
|
||
mov.u32 %r3521, %r3023;
|
||
mov.u32 %r3523, %r3102;
|
||
mov.u32 %r3525, %r3181;
|
||
mov.u32 %r3527, %r3260;
|
||
mov.u32 %r3529, %r3339;
|
||
mov.u32 %r3531, %r3418;
|
||
@%p159 bra BB1_240;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_240:
|
||
or.b32 %r3513, %r167, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3373, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3373;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_303:
|
||
or.b32 %r3507, %r171, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3389, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3389;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_213:
|
||
setp.eq.s32 %p179, %r1650, 16;
|
||
mov.u32 %r2397, %r168;
|
||
mov.u32 %r2476, %r171;
|
||
mov.u32 %r2555, %r170;
|
||
mov.u32 %r2634, %r169;
|
||
mov.u32 %r2713, %r167;
|
||
mov.u32 %r2792, %r166;
|
||
mov.u32 %r2871, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3029, %r163;
|
||
mov.u32 %r3108, %r162;
|
||
mov.u32 %r3187, %r161;
|
||
mov.u32 %r3266, %r160;
|
||
mov.u32 %r3345, %r159;
|
||
mov.u32 %r3424, %r158;
|
||
mov.u32 %r3500, %r2397;
|
||
mov.u32 %r3507, %r2476;
|
||
mov.u32 %r3509, %r2555;
|
||
mov.u32 %r3511, %r2634;
|
||
mov.u32 %r3513, %r2713;
|
||
mov.u32 %r3515, %r2792;
|
||
mov.u32 %r3517, %r2871;
|
||
mov.u32 %r3521, %r3029;
|
||
mov.u32 %r3523, %r3108;
|
||
mov.u32 %r3525, %r3187;
|
||
mov.u32 %r3527, %r3266;
|
||
mov.u32 %r3529, %r3345;
|
||
mov.u32 %r3531, %r3424;
|
||
@%p179 bra BB1_214;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_214:
|
||
mov.u32 %r3519, %r152;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3387, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3387;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_291:
|
||
or.b32 %r3523, %r162, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3361, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3361;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_264:
|
||
setp.eq.s32 %p140, %r1650, 44;
|
||
mov.u32 %r2385, %r168;
|
||
mov.u32 %r2464, %r171;
|
||
mov.u32 %r2543, %r170;
|
||
mov.u32 %r2622, %r169;
|
||
mov.u32 %r2701, %r167;
|
||
mov.u32 %r2780, %r166;
|
||
mov.u32 %r2859, %r165;
|
||
mov.u32 %r2938, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3096, %r162;
|
||
mov.u32 %r3175, %r161;
|
||
mov.u32 %r3254, %r160;
|
||
mov.u32 %r3333, %r159;
|
||
mov.u32 %r3412, %r158;
|
||
mov.u32 %r3500, %r2385;
|
||
mov.u32 %r3507, %r2464;
|
||
mov.u32 %r3509, %r2543;
|
||
mov.u32 %r3511, %r2622;
|
||
mov.u32 %r3513, %r2701;
|
||
mov.u32 %r3515, %r2780;
|
||
mov.u32 %r3517, %r2859;
|
||
mov.u32 %r3519, %r2938;
|
||
mov.u32 %r3523, %r3096;
|
||
mov.u32 %r3525, %r3175;
|
||
mov.u32 %r3527, %r3254;
|
||
mov.u32 %r3529, %r3333;
|
||
mov.u32 %r3531, %r3412;
|
||
@%p140 bra BB1_265;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_265:
|
||
mov.u32 %r3521, %r152;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3359, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3359;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_305:
|
||
or.b32 %r3511, %r169, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3396, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3396;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_293:
|
||
or.b32 %r3527, %r160, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3368, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3368;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_299:
|
||
or.b32 %r3517, %r165, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3382, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3382;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_225:
|
||
setp.eq.s32 %p170, %r1650, 23;
|
||
mov.u32 %r2394, %r168;
|
||
mov.u32 %r2473, %r171;
|
||
mov.u32 %r2552, %r170;
|
||
mov.u32 %r2631, %r169;
|
||
mov.u32 %r2710, %r167;
|
||
mov.u32 %r2789, %r166;
|
||
mov.u32 %r2868, %r165;
|
||
mov.u32 %r2947, %r164;
|
||
mov.u32 %r3026, %r163;
|
||
mov.u32 %r3105, %r162;
|
||
mov.u32 %r3184, %r161;
|
||
mov.u32 %r3263, %r160;
|
||
mov.u32 %r3342, %r159;
|
||
mov.u32 %r3421, %r158;
|
||
mov.u32 %r3500, %r2394;
|
||
mov.u32 %r3507, %r2473;
|
||
mov.u32 %r3509, %r2552;
|
||
mov.u32 %r3511, %r2631;
|
||
mov.u32 %r3513, %r2710;
|
||
mov.u32 %r3515, %r2789;
|
||
mov.u32 %r3517, %r2868;
|
||
mov.u32 %r3519, %r2947;
|
||
mov.u32 %r3521, %r3026;
|
||
mov.u32 %r3523, %r3105;
|
||
mov.u32 %r3525, %r3184;
|
||
mov.u32 %r3527, %r3263;
|
||
mov.u32 %r3529, %r3342;
|
||
mov.u32 %r3531, %r3421;
|
||
@%p170 bra BB1_226;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_226:
|
||
or.b32 %r3517, %r165, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3380, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3380;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_287:
|
||
or.b32 %r3531, %r158, %r153;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3354, %r159;
|
||
mov.u32 %r3529, %r3354;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_276:
|
||
setp.eq.s32 %p131, %r1650, 51;
|
||
mov.u32 %r2382, %r168;
|
||
mov.u32 %r2461, %r171;
|
||
mov.u32 %r2540, %r170;
|
||
mov.u32 %r2619, %r169;
|
||
mov.u32 %r2698, %r167;
|
||
mov.u32 %r2777, %r166;
|
||
mov.u32 %r2856, %r165;
|
||
mov.u32 %r2935, %r164;
|
||
mov.u32 %r3014, %r163;
|
||
mov.u32 %r3093, %r162;
|
||
mov.u32 %r3172, %r161;
|
||
mov.u32 %r3251, %r160;
|
||
mov.u32 %r3330, %r159;
|
||
mov.u32 %r3409, %r158;
|
||
mov.u32 %r3500, %r2382;
|
||
mov.u32 %r3507, %r2461;
|
||
mov.u32 %r3509, %r2540;
|
||
mov.u32 %r3511, %r2619;
|
||
mov.u32 %r3513, %r2698;
|
||
mov.u32 %r3515, %r2777;
|
||
mov.u32 %r3517, %r2856;
|
||
mov.u32 %r3519, %r2935;
|
||
mov.u32 %r3521, %r3014;
|
||
mov.u32 %r3523, %r3093;
|
||
mov.u32 %r3525, %r3172;
|
||
mov.u32 %r3527, %r3251;
|
||
mov.u32 %r3529, %r3330;
|
||
mov.u32 %r3531, %r3409;
|
||
@%p131 bra BB1_277;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_277:
|
||
or.b32 %r3531, %r158, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3352, %r159;
|
||
mov.u32 %r3529, %r3352;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_192:
|
||
setp.eq.s32 %p195, %r1650, 4;
|
||
mov.u32 %r2402, %r168;
|
||
mov.u32 %r2481, %r171;
|
||
mov.u32 %r2560, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r2718, %r167;
|
||
mov.u32 %r2797, %r166;
|
||
mov.u32 %r2876, %r165;
|
||
mov.u32 %r2955, %r164;
|
||
mov.u32 %r3034, %r163;
|
||
mov.u32 %r3113, %r162;
|
||
mov.u32 %r3192, %r161;
|
||
mov.u32 %r3271, %r160;
|
||
mov.u32 %r3350, %r159;
|
||
mov.u32 %r3429, %r158;
|
||
mov.u32 %r3500, %r2402;
|
||
mov.u32 %r3507, %r2481;
|
||
mov.u32 %r3509, %r2560;
|
||
mov.u32 %r3513, %r2718;
|
||
mov.u32 %r3515, %r2797;
|
||
mov.u32 %r3517, %r2876;
|
||
mov.u32 %r3519, %r2955;
|
||
mov.u32 %r3521, %r3034;
|
||
mov.u32 %r3523, %r3113;
|
||
mov.u32 %r3525, %r3192;
|
||
mov.u32 %r3527, %r3271;
|
||
mov.u32 %r3529, %r3350;
|
||
mov.u32 %r3531, %r3429;
|
||
@%p195 bra BB1_193;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_193:
|
||
mov.u32 %r3511, %r152;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3399, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3399;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_243:
|
||
setp.eq.s32 %p156, %r1650, 32;
|
||
mov.u32 %r2390, %r168;
|
||
mov.u32 %r2469, %r171;
|
||
mov.u32 %r2548, %r170;
|
||
mov.u32 %r2627, %r169;
|
||
mov.u32 %r2706, %r167;
|
||
mov.u32 %r2785, %r166;
|
||
mov.u32 %r2864, %r165;
|
||
mov.u32 %r2943, %r164;
|
||
mov.u32 %r3022, %r163;
|
||
mov.u32 %r3101, %r162;
|
||
mov.u32 %r3180, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3338, %r159;
|
||
mov.u32 %r3417, %r158;
|
||
mov.u32 %r3500, %r2390;
|
||
mov.u32 %r3507, %r2469;
|
||
mov.u32 %r3509, %r2548;
|
||
mov.u32 %r3511, %r2627;
|
||
mov.u32 %r3513, %r2706;
|
||
mov.u32 %r3515, %r2785;
|
||
mov.u32 %r3517, %r2864;
|
||
mov.u32 %r3519, %r2943;
|
||
mov.u32 %r3521, %r3022;
|
||
mov.u32 %r3523, %r3101;
|
||
mov.u32 %r3525, %r3180;
|
||
mov.u32 %r3529, %r3338;
|
||
mov.u32 %r3531, %r3417;
|
||
@%p156 bra BB1_244;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_244:
|
||
mov.u32 %r3527, %r152;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3371, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3371;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_217:
|
||
setp.eq.s32 %p176, %r1650, 18;
|
||
mov.u32 %r2396, %r168;
|
||
mov.u32 %r2475, %r171;
|
||
mov.u32 %r2554, %r170;
|
||
mov.u32 %r2633, %r169;
|
||
mov.u32 %r2712, %r167;
|
||
mov.u32 %r2791, %r166;
|
||
mov.u32 %r2870, %r165;
|
||
mov.u32 %r2949, %r164;
|
||
mov.u32 %r3028, %r163;
|
||
mov.u32 %r3107, %r162;
|
||
mov.u32 %r3186, %r161;
|
||
mov.u32 %r3265, %r160;
|
||
mov.u32 %r3344, %r159;
|
||
mov.u32 %r3423, %r158;
|
||
mov.u32 %r3500, %r2396;
|
||
mov.u32 %r3507, %r2475;
|
||
mov.u32 %r3509, %r2554;
|
||
mov.u32 %r3511, %r2633;
|
||
mov.u32 %r3513, %r2712;
|
||
mov.u32 %r3515, %r2791;
|
||
mov.u32 %r3517, %r2870;
|
||
mov.u32 %r3519, %r2949;
|
||
mov.u32 %r3521, %r3028;
|
||
mov.u32 %r3523, %r3107;
|
||
mov.u32 %r3525, %r3186;
|
||
mov.u32 %r3527, %r3265;
|
||
mov.u32 %r3529, %r3344;
|
||
mov.u32 %r3531, %r3423;
|
||
@%p176 bra BB1_218;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_218:
|
||
or.b32 %r3519, %r164, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3385, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3385;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_268:
|
||
setp.eq.s32 %p137, %r1650, 46;
|
||
mov.u32 %r2384, %r168;
|
||
mov.u32 %r2463, %r171;
|
||
mov.u32 %r2542, %r170;
|
||
mov.u32 %r2621, %r169;
|
||
mov.u32 %r2700, %r167;
|
||
mov.u32 %r2779, %r166;
|
||
mov.u32 %r2858, %r165;
|
||
mov.u32 %r2937, %r164;
|
||
mov.u32 %r3016, %r163;
|
||
mov.u32 %r3095, %r162;
|
||
mov.u32 %r3174, %r161;
|
||
mov.u32 %r3253, %r160;
|
||
mov.u32 %r3332, %r159;
|
||
mov.u32 %r3411, %r158;
|
||
mov.u32 %r3500, %r2384;
|
||
mov.u32 %r3507, %r2463;
|
||
mov.u32 %r3509, %r2542;
|
||
mov.u32 %r3511, %r2621;
|
||
mov.u32 %r3513, %r2700;
|
||
mov.u32 %r3515, %r2779;
|
||
mov.u32 %r3517, %r2858;
|
||
mov.u32 %r3519, %r2937;
|
||
mov.u32 %r3521, %r3016;
|
||
mov.u32 %r3523, %r3095;
|
||
mov.u32 %r3525, %r3174;
|
||
mov.u32 %r3527, %r3253;
|
||
mov.u32 %r3529, %r3332;
|
||
mov.u32 %r3531, %r3411;
|
||
@%p137 bra BB1_269;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_269:
|
||
or.b32 %r3521, %r163, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3357, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3357;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_204:
|
||
setp.eq.s32 %p186, %r1650, 11;
|
||
mov.u32 %r2399, %r168;
|
||
mov.u32 %r2478, %r171;
|
||
mov.u32 %r2557, %r170;
|
||
mov.u32 %r2636, %r169;
|
||
mov.u32 %r2715, %r167;
|
||
mov.u32 %r2794, %r166;
|
||
mov.u32 %r2873, %r165;
|
||
mov.u32 %r2952, %r164;
|
||
mov.u32 %r3031, %r163;
|
||
mov.u32 %r3110, %r162;
|
||
mov.u32 %r3189, %r161;
|
||
mov.u32 %r3268, %r160;
|
||
mov.u32 %r3347, %r159;
|
||
mov.u32 %r3426, %r158;
|
||
mov.u32 %r3500, %r2399;
|
||
mov.u32 %r3507, %r2478;
|
||
mov.u32 %r3509, %r2557;
|
||
mov.u32 %r3511, %r2636;
|
||
mov.u32 %r3513, %r2715;
|
||
mov.u32 %r3515, %r2794;
|
||
mov.u32 %r3517, %r2873;
|
||
mov.u32 %r3519, %r2952;
|
||
mov.u32 %r3521, %r3031;
|
||
mov.u32 %r3523, %r3110;
|
||
mov.u32 %r3525, %r3189;
|
||
mov.u32 %r3527, %r3268;
|
||
mov.u32 %r3529, %r3347;
|
||
mov.u32 %r3531, %r3426;
|
||
@%p186 bra BB1_205;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_205:
|
||
or.b32 %r3509, %r170, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3392, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3392;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_255:
|
||
setp.eq.s32 %p147, %r1650, 39;
|
||
mov.u32 %r2387, %r168;
|
||
mov.u32 %r2466, %r171;
|
||
mov.u32 %r2545, %r170;
|
||
mov.u32 %r2624, %r169;
|
||
mov.u32 %r2703, %r167;
|
||
mov.u32 %r2782, %r166;
|
||
mov.u32 %r2861, %r165;
|
||
mov.u32 %r2940, %r164;
|
||
mov.u32 %r3019, %r163;
|
||
mov.u32 %r3098, %r162;
|
||
mov.u32 %r3177, %r161;
|
||
mov.u32 %r3256, %r160;
|
||
mov.u32 %r3335, %r159;
|
||
mov.u32 %r3414, %r158;
|
||
mov.u32 %r3500, %r2387;
|
||
mov.u32 %r3507, %r2466;
|
||
mov.u32 %r3509, %r2545;
|
||
mov.u32 %r3511, %r2624;
|
||
mov.u32 %r3513, %r2703;
|
||
mov.u32 %r3515, %r2782;
|
||
mov.u32 %r3517, %r2861;
|
||
mov.u32 %r3519, %r2940;
|
||
mov.u32 %r3521, %r3019;
|
||
mov.u32 %r3523, %r3098;
|
||
mov.u32 %r3525, %r3177;
|
||
mov.u32 %r3527, %r3256;
|
||
mov.u32 %r3529, %r3335;
|
||
mov.u32 %r3531, %r3414;
|
||
@%p147 bra BB1_256;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_256:
|
||
or.b32 %r3525, %r161, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3364, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3364;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_195:
|
||
setp.eq.s32 %p193, %r1650, 6;
|
||
mov.u32 %r2401, %r168;
|
||
mov.u32 %r2480, %r171;
|
||
mov.u32 %r2559, %r170;
|
||
mov.u32 %r2638, %r169;
|
||
mov.u32 %r2717, %r167;
|
||
mov.u32 %r2796, %r166;
|
||
mov.u32 %r2875, %r165;
|
||
mov.u32 %r2954, %r164;
|
||
mov.u32 %r3033, %r163;
|
||
mov.u32 %r3112, %r162;
|
||
mov.u32 %r3191, %r161;
|
||
mov.u32 %r3270, %r160;
|
||
mov.u32 %r3349, %r159;
|
||
mov.u32 %r3428, %r158;
|
||
mov.u32 %r3500, %r2401;
|
||
mov.u32 %r3507, %r2480;
|
||
mov.u32 %r3509, %r2559;
|
||
mov.u32 %r3511, %r2638;
|
||
mov.u32 %r3513, %r2717;
|
||
mov.u32 %r3515, %r2796;
|
||
mov.u32 %r3517, %r2875;
|
||
mov.u32 %r3519, %r2954;
|
||
mov.u32 %r3521, %r3033;
|
||
mov.u32 %r3523, %r3112;
|
||
mov.u32 %r3525, %r3191;
|
||
mov.u32 %r3527, %r3270;
|
||
mov.u32 %r3529, %r3349;
|
||
mov.u32 %r3531, %r3428;
|
||
@%p193 bra BB1_196;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_196:
|
||
or.b32 %r3511, %r169, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3397, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3397;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_246:
|
||
setp.eq.s32 %p154, %r1650, 34;
|
||
mov.u32 %r2389, %r168;
|
||
mov.u32 %r2468, %r171;
|
||
mov.u32 %r2547, %r170;
|
||
mov.u32 %r2626, %r169;
|
||
mov.u32 %r2705, %r167;
|
||
mov.u32 %r2784, %r166;
|
||
mov.u32 %r2863, %r165;
|
||
mov.u32 %r2942, %r164;
|
||
mov.u32 %r3021, %r163;
|
||
mov.u32 %r3100, %r162;
|
||
mov.u32 %r3179, %r161;
|
||
mov.u32 %r3258, %r160;
|
||
mov.u32 %r3337, %r159;
|
||
mov.u32 %r3416, %r158;
|
||
mov.u32 %r3500, %r2389;
|
||
mov.u32 %r3507, %r2468;
|
||
mov.u32 %r3509, %r2547;
|
||
mov.u32 %r3511, %r2626;
|
||
mov.u32 %r3513, %r2705;
|
||
mov.u32 %r3515, %r2784;
|
||
mov.u32 %r3517, %r2863;
|
||
mov.u32 %r3519, %r2942;
|
||
mov.u32 %r3521, %r3021;
|
||
mov.u32 %r3523, %r3100;
|
||
mov.u32 %r3525, %r3179;
|
||
mov.u32 %r3527, %r3258;
|
||
mov.u32 %r3529, %r3337;
|
||
mov.u32 %r3531, %r3416;
|
||
@%p154 bra BB1_247;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_247:
|
||
or.b32 %r3527, %r160, %r154;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3369, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3369;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_220:
|
||
setp.eq.s32 %p174, %r1650, 20;
|
||
mov.u32 %r2395, %r168;
|
||
mov.u32 %r2474, %r171;
|
||
mov.u32 %r2553, %r170;
|
||
mov.u32 %r2632, %r169;
|
||
mov.u32 %r2711, %r167;
|
||
mov.u32 %r2790, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r2948, %r164;
|
||
mov.u32 %r3027, %r163;
|
||
mov.u32 %r3106, %r162;
|
||
mov.u32 %r3185, %r161;
|
||
mov.u32 %r3264, %r160;
|
||
mov.u32 %r3343, %r159;
|
||
mov.u32 %r3422, %r158;
|
||
mov.u32 %r3500, %r2395;
|
||
mov.u32 %r3507, %r2474;
|
||
mov.u32 %r3509, %r2553;
|
||
mov.u32 %r3511, %r2632;
|
||
mov.u32 %r3513, %r2711;
|
||
mov.u32 %r3515, %r2790;
|
||
mov.u32 %r3519, %r2948;
|
||
mov.u32 %r3521, %r3027;
|
||
mov.u32 %r3523, %r3106;
|
||
mov.u32 %r3525, %r3185;
|
||
mov.u32 %r3527, %r3264;
|
||
mov.u32 %r3529, %r3343;
|
||
mov.u32 %r3531, %r3422;
|
||
@%p174 bra BB1_221;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_221:
|
||
mov.u32 %r3517, %r152;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3383, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3383;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_271:
|
||
setp.eq.s32 %p135, %r1650, 48;
|
||
mov.u32 %r2383, %r168;
|
||
mov.u32 %r2462, %r171;
|
||
mov.u32 %r2541, %r170;
|
||
mov.u32 %r2620, %r169;
|
||
mov.u32 %r2699, %r167;
|
||
mov.u32 %r2778, %r166;
|
||
mov.u32 %r2857, %r165;
|
||
mov.u32 %r2936, %r164;
|
||
mov.u32 %r3015, %r163;
|
||
mov.u32 %r3094, %r162;
|
||
mov.u32 %r3173, %r161;
|
||
mov.u32 %r3252, %r160;
|
||
mov.u32 %r3331, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3500, %r2383;
|
||
mov.u32 %r3507, %r2462;
|
||
mov.u32 %r3509, %r2541;
|
||
mov.u32 %r3511, %r2620;
|
||
mov.u32 %r3513, %r2699;
|
||
mov.u32 %r3515, %r2778;
|
||
mov.u32 %r3517, %r2857;
|
||
mov.u32 %r3519, %r2936;
|
||
mov.u32 %r3521, %r3015;
|
||
mov.u32 %r3523, %r3094;
|
||
mov.u32 %r3525, %r3173;
|
||
mov.u32 %r3527, %r3252;
|
||
mov.u32 %r3529, %r3331;
|
||
@%p135 bra BB1_272;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_272:
|
||
mov.u32 %r3531, %r152;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3355, %r159;
|
||
mov.u32 %r3529, %r3355;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_232:
|
||
setp.eq.s32 %p165, %r1650, 27;
|
||
mov.u32 %r2392, %r168;
|
||
mov.u32 %r2471, %r171;
|
||
mov.u32 %r2550, %r170;
|
||
mov.u32 %r2629, %r169;
|
||
mov.u32 %r2708, %r167;
|
||
mov.u32 %r2787, %r166;
|
||
mov.u32 %r2866, %r165;
|
||
mov.u32 %r2945, %r164;
|
||
mov.u32 %r3024, %r163;
|
||
mov.u32 %r3103, %r162;
|
||
mov.u32 %r3182, %r161;
|
||
mov.u32 %r3261, %r160;
|
||
mov.u32 %r3340, %r159;
|
||
mov.u32 %r3419, %r158;
|
||
mov.u32 %r3500, %r2392;
|
||
mov.u32 %r3507, %r2471;
|
||
mov.u32 %r3509, %r2550;
|
||
mov.u32 %r3511, %r2629;
|
||
mov.u32 %r3513, %r2708;
|
||
mov.u32 %r3515, %r2787;
|
||
mov.u32 %r3517, %r2866;
|
||
mov.u32 %r3519, %r2945;
|
||
mov.u32 %r3521, %r3024;
|
||
mov.u32 %r3523, %r3103;
|
||
mov.u32 %r3525, %r3182;
|
||
mov.u32 %r3527, %r3261;
|
||
mov.u32 %r3529, %r3340;
|
||
mov.u32 %r3531, %r3419;
|
||
@%p165 bra BB1_233;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_233:
|
||
or.b32 %r3515, %r166, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3376, %r159;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r3376;
|
||
bra.uni BB1_310;
|
||
|
||
BB1_283:
|
||
setp.ne.s32 %p126, %r1650, 55;
|
||
mov.u32 %r2380, %r168;
|
||
mov.u32 %r2459, %r171;
|
||
mov.u32 %r2538, %r170;
|
||
mov.u32 %r2617, %r169;
|
||
mov.u32 %r2696, %r167;
|
||
mov.u32 %r2775, %r166;
|
||
mov.u32 %r2854, %r165;
|
||
mov.u32 %r2933, %r164;
|
||
mov.u32 %r3012, %r163;
|
||
mov.u32 %r3091, %r162;
|
||
mov.u32 %r3170, %r161;
|
||
mov.u32 %r3249, %r160;
|
||
mov.u32 %r3328, %r159;
|
||
mov.u32 %r3407, %r158;
|
||
mov.u32 %r3500, %r2380;
|
||
mov.u32 %r3507, %r2459;
|
||
mov.u32 %r3509, %r2538;
|
||
mov.u32 %r3511, %r2617;
|
||
mov.u32 %r3513, %r2696;
|
||
mov.u32 %r3515, %r2775;
|
||
mov.u32 %r3517, %r2854;
|
||
mov.u32 %r3519, %r2933;
|
||
mov.u32 %r3521, %r3012;
|
||
mov.u32 %r3523, %r3091;
|
||
mov.u32 %r3525, %r3170;
|
||
mov.u32 %r3527, %r3249;
|
||
mov.u32 %r3529, %r3328;
|
||
mov.u32 %r3531, %r3407;
|
||
@%p126 bra BB1_310;
|
||
|
||
or.b32 %r172, %r159, %r155;
|
||
mov.u32 %r3500, %r168;
|
||
mov.u32 %r3507, %r171;
|
||
mov.u32 %r3509, %r170;
|
||
mov.u32 %r3511, %r169;
|
||
mov.u32 %r3513, %r167;
|
||
mov.u32 %r3515, %r166;
|
||
mov.u32 %r3517, %r165;
|
||
mov.u32 %r3519, %r164;
|
||
mov.u32 %r3521, %r163;
|
||
mov.u32 %r3523, %r162;
|
||
mov.u32 %r3525, %r161;
|
||
mov.u32 %r3527, %r160;
|
||
mov.u32 %r3531, %r158;
|
||
mov.u32 %r3529, %r172;
|
||
|
||
BB1_310:
|
||
mov.u32 %r3404, %r3529;
|
||
mov.u32 %r3530, %r3531;
|
||
mov.u32 %r3528, %r3404;
|
||
mov.u32 %r3526, %r3527;
|
||
mov.u32 %r3524, %r3525;
|
||
mov.u32 %r3522, %r3523;
|
||
mov.u32 %r3520, %r3521;
|
||
mov.u32 %r3518, %r3519;
|
||
mov.u32 %r3516, %r3517;
|
||
mov.u32 %r3514, %r3515;
|
||
mov.u32 %r3512, %r3513;
|
||
mov.u32 %r3510, %r3511;
|
||
mov.u32 %r3508, %r3509;
|
||
mov.u32 %r3506, %r3507;
|
||
mov.u32 %r3499, %r3500;
|
||
add.s32 %r1650, %r1650, 1;
|
||
shr.u32 %r1649, %r1649, 1;
|
||
setp.ne.s32 %p199, %r1649, 0;
|
||
mov.u32 %r3488, %r3499;
|
||
mov.u32 %r3485, %r3506;
|
||
mov.u32 %r3486, %r3508;
|
||
mov.u32 %r3487, %r3510;
|
||
mov.u32 %r3489, %r3512;
|
||
mov.u32 %r3490, %r3514;
|
||
mov.u32 %r3491, %r3516;
|
||
mov.u32 %r3492, %r3518;
|
||
mov.u32 %r3493, %r3520;
|
||
mov.u32 %r3494, %r3522;
|
||
mov.u32 %r3495, %r3524;
|
||
mov.u32 %r3496, %r3526;
|
||
mov.u32 %r3497, %r3528;
|
||
mov.u32 %r3498, %r3530;
|
||
@%p199 bra BB1_181;
|
||
|
||
BB1_311:
|
||
mov.b32 %r1644, %envreg3;
|
||
mov.u32 %r1643, %ntid.x;
|
||
mov.u32 %r1642, %ctaid.x;
|
||
mov.u32 %r1641, %tid.x;
|
||
mad.lo.s32 %r1640, %r1642, %r1643, %r1644;
|
||
add.s32 %r1639, %r1640, %r1641;
|
||
cvt.u64.u32 %rd12, %r1639;
|
||
ld.param.u64 %rd11, [m00500_init_param_4];
|
||
mov.u32 %r1638, 128;
|
||
shl.b32 %r1115, %r1650, 3;
|
||
and.b32 %r1116, %r1115, 24;
|
||
shl.b32 %r1118, %r1638, %r1116;
|
||
setp.lt.u32 %p200, %r1650, 4;
|
||
selp.b32 %r1119, %r1118, 0, %p200;
|
||
or.b32 %r1120, %r1119, %r3488;
|
||
and.b32 %r1121, %r1650, -4;
|
||
setp.eq.s32 %p201, %r1121, 4;
|
||
selp.b32 %r1122, %r1118, 0, %p201;
|
||
or.b32 %r1123, %r1122, %r3487;
|
||
setp.eq.s32 %p202, %r1121, 8;
|
||
selp.b32 %r1124, %r1118, 0, %p202;
|
||
or.b32 %r1125, %r1124, %r3486;
|
||
setp.eq.s32 %p203, %r1121, 12;
|
||
selp.b32 %r1126, %r1118, 0, %p203;
|
||
or.b32 %r1127, %r1126, %r3485;
|
||
setp.eq.s32 %p204, %r1121, 16;
|
||
selp.b32 %r1128, %r1118, 0, %p204;
|
||
or.b32 %r1129, %r1128, %r3492;
|
||
setp.eq.s32 %p205, %r1121, 20;
|
||
selp.b32 %r1130, %r1118, 0, %p205;
|
||
or.b32 %r1131, %r1130, %r3491;
|
||
setp.eq.s32 %p206, %r1121, 24;
|
||
selp.b32 %r1132, %r1118, 0, %p206;
|
||
or.b32 %r1133, %r1132, %r3490;
|
||
setp.eq.s32 %p207, %r1121, 28;
|
||
selp.b32 %r1134, %r1118, 0, %p207;
|
||
or.b32 %r1135, %r1134, %r3489;
|
||
setp.eq.s32 %p208, %r1121, 32;
|
||
selp.b32 %r1136, %r1118, 0, %p208;
|
||
or.b32 %r1137, %r1136, %r3496;
|
||
setp.eq.s32 %p209, %r1121, 36;
|
||
selp.b32 %r1138, %r1118, 0, %p209;
|
||
or.b32 %r1139, %r1138, %r3495;
|
||
setp.eq.s32 %p210, %r1121, 40;
|
||
selp.b32 %r1140, %r1118, 0, %p210;
|
||
or.b32 %r1141, %r1140, %r3494;
|
||
setp.eq.s32 %p211, %r1121, 44;
|
||
selp.b32 %r1142, %r1118, 0, %p211;
|
||
or.b32 %r1143, %r1142, %r3493;
|
||
setp.eq.s32 %p212, %r1121, 48;
|
||
selp.b32 %r1144, %r1118, 0, %p212;
|
||
or.b32 %r1145, %r1144, %r3498;
|
||
setp.eq.s32 %p213, %r1121, 52;
|
||
selp.b32 %r1146, %r1118, 0, %p213;
|
||
or.b32 %r1147, %r1146, %r3497;
|
||
add.s32 %r1148, %r1120, -680876937;
|
||
shf.l.wrap.b32 %r1149, %r1148, %r1148, 7;
|
||
add.s32 %r1150, %r1149, -271733879;
|
||
and.b32 %r1151, %r1150, 2004318071;
|
||
xor.b32 %r1152, %r1151, -1732584194;
|
||
add.s32 %r1153, %r1123, %r1152;
|
||
add.s32 %r1154, %r1153, -117830708;
|
||
shf.l.wrap.b32 %r1155, %r1154, %r1154, 12;
|
||
add.s32 %r1156, %r1155, %r1150;
|
||
xor.b32 %r1157, %r1150, -271733879;
|
||
and.b32 %r1158, %r1156, %r1157;
|
||
xor.b32 %r1159, %r1158, -271733879;
|
||
add.s32 %r1160, %r1125, %r1159;
|
||
add.s32 %r1161, %r1160, -1126478375;
|
||
shf.l.wrap.b32 %r1162, %r1161, %r1161, 17;
|
||
add.s32 %r1163, %r1162, %r1156;
|
||
xor.b32 %r1164, %r1156, %r1150;
|
||
and.b32 %r1165, %r1163, %r1164;
|
||
xor.b32 %r1166, %r1165, %r1150;
|
||
add.s32 %r1167, %r1127, %r1166;
|
||
add.s32 %r1168, %r1167, -1316259209;
|
||
shf.l.wrap.b32 %r1169, %r1168, %r1168, 22;
|
||
add.s32 %r1170, %r1169, %r1163;
|
||
xor.b32 %r1171, %r1163, %r1156;
|
||
and.b32 %r1172, %r1170, %r1171;
|
||
xor.b32 %r1173, %r1172, %r1156;
|
||
add.s32 %r1174, %r1129, %r1149;
|
||
add.s32 %r1175, %r1174, %r1173;
|
||
add.s32 %r1176, %r1175, -448152776;
|
||
shf.l.wrap.b32 %r1177, %r1176, %r1176, 7;
|
||
add.s32 %r1178, %r1177, %r1170;
|
||
xor.b32 %r1179, %r1170, %r1163;
|
||
and.b32 %r1180, %r1178, %r1179;
|
||
xor.b32 %r1181, %r1180, %r1163;
|
||
add.s32 %r1182, %r1131, %r1156;
|
||
add.s32 %r1183, %r1182, %r1181;
|
||
add.s32 %r1184, %r1183, 1200080426;
|
||
shf.l.wrap.b32 %r1185, %r1184, %r1184, 12;
|
||
add.s32 %r1186, %r1185, %r1178;
|
||
xor.b32 %r1187, %r1178, %r1170;
|
||
and.b32 %r1188, %r1186, %r1187;
|
||
xor.b32 %r1189, %r1188, %r1170;
|
||
add.s32 %r1190, %r1133, %r1163;
|
||
add.s32 %r1191, %r1190, %r1189;
|
||
add.s32 %r1192, %r1191, -1473231341;
|
||
shf.l.wrap.b32 %r1193, %r1192, %r1192, 17;
|
||
add.s32 %r1194, %r1193, %r1186;
|
||
xor.b32 %r1195, %r1186, %r1178;
|
||
and.b32 %r1196, %r1194, %r1195;
|
||
xor.b32 %r1197, %r1196, %r1178;
|
||
add.s32 %r1198, %r1135, %r1170;
|
||
add.s32 %r1199, %r1198, %r1197;
|
||
add.s32 %r1200, %r1199, -45705983;
|
||
shf.l.wrap.b32 %r1201, %r1200, %r1200, 22;
|
||
add.s32 %r1202, %r1201, %r1194;
|
||
xor.b32 %r1203, %r1194, %r1186;
|
||
and.b32 %r1204, %r1202, %r1203;
|
||
xor.b32 %r1205, %r1204, %r1186;
|
||
add.s32 %r1206, %r1137, %r1178;
|
||
add.s32 %r1207, %r1206, %r1205;
|
||
add.s32 %r1208, %r1207, 1770035416;
|
||
shf.l.wrap.b32 %r1209, %r1208, %r1208, 7;
|
||
add.s32 %r1210, %r1209, %r1202;
|
||
xor.b32 %r1211, %r1202, %r1194;
|
||
and.b32 %r1212, %r1210, %r1211;
|
||
xor.b32 %r1213, %r1212, %r1194;
|
||
add.s32 %r1214, %r1139, %r1186;
|
||
add.s32 %r1215, %r1214, %r1213;
|
||
add.s32 %r1216, %r1215, -1958414417;
|
||
shf.l.wrap.b32 %r1217, %r1216, %r1216, 12;
|
||
add.s32 %r1218, %r1217, %r1210;
|
||
xor.b32 %r1219, %r1210, %r1202;
|
||
and.b32 %r1220, %r1218, %r1219;
|
||
xor.b32 %r1221, %r1220, %r1202;
|
||
add.s32 %r1222, %r1141, %r1194;
|
||
add.s32 %r1223, %r1222, %r1221;
|
||
add.s32 %r1224, %r1223, -42063;
|
||
shf.l.wrap.b32 %r1225, %r1224, %r1224, 17;
|
||
add.s32 %r1226, %r1225, %r1218;
|
||
xor.b32 %r1227, %r1218, %r1210;
|
||
and.b32 %r1228, %r1226, %r1227;
|
||
xor.b32 %r1229, %r1228, %r1210;
|
||
add.s32 %r1230, %r1143, %r1202;
|
||
add.s32 %r1231, %r1230, %r1229;
|
||
add.s32 %r1232, %r1231, -1990404162;
|
||
shf.l.wrap.b32 %r1233, %r1232, %r1232, 22;
|
||
add.s32 %r1234, %r1233, %r1226;
|
||
xor.b32 %r1235, %r1226, %r1218;
|
||
and.b32 %r1236, %r1234, %r1235;
|
||
xor.b32 %r1237, %r1236, %r1218;
|
||
add.s32 %r1238, %r1145, %r1210;
|
||
add.s32 %r1239, %r1238, %r1237;
|
||
add.s32 %r1240, %r1239, 1804603682;
|
||
shf.l.wrap.b32 %r1241, %r1240, %r1240, 7;
|
||
add.s32 %r1242, %r1241, %r1234;
|
||
xor.b32 %r1243, %r1234, %r1226;
|
||
and.b32 %r1244, %r1242, %r1243;
|
||
xor.b32 %r1245, %r1244, %r1226;
|
||
add.s32 %r1246, %r1147, %r1218;
|
||
add.s32 %r1247, %r1246, %r1245;
|
||
add.s32 %r1248, %r1247, -40341101;
|
||
shf.l.wrap.b32 %r1249, %r1248, %r1248, 12;
|
||
add.s32 %r1250, %r1249, %r1242;
|
||
xor.b32 %r1251, %r1242, %r1234;
|
||
and.b32 %r1252, %r1250, %r1251;
|
||
xor.b32 %r1253, %r1252, %r1234;
|
||
add.s32 %r1254, %r1115, %r1226;
|
||
add.s32 %r1255, %r1254, %r1253;
|
||
add.s32 %r1256, %r1255, -1502002290;
|
||
shf.l.wrap.b32 %r1257, %r1256, %r1256, 17;
|
||
add.s32 %r1258, %r1257, %r1250;
|
||
xor.b32 %r1259, %r1250, %r1242;
|
||
and.b32 %r1260, %r1258, %r1259;
|
||
xor.b32 %r1261, %r1260, %r1242;
|
||
add.s32 %r1262, %r1234, %r1261;
|
||
add.s32 %r1263, %r1262, 1236535329;
|
||
shf.l.wrap.b32 %r1264, %r1263, %r1263, 22;
|
||
add.s32 %r1265, %r1264, %r1258;
|
||
xor.b32 %r1266, %r1265, %r1258;
|
||
and.b32 %r1267, %r1266, %r1250;
|
||
xor.b32 %r1268, %r1267, %r1258;
|
||
add.s32 %r1269, %r1123, %r1242;
|
||
add.s32 %r1270, %r1269, %r1268;
|
||
add.s32 %r1271, %r1270, -165796510;
|
||
shf.l.wrap.b32 %r1272, %r1271, %r1271, 5;
|
||
add.s32 %r1273, %r1272, %r1265;
|
||
xor.b32 %r1274, %r1273, %r1265;
|
||
and.b32 %r1275, %r1274, %r1258;
|
||
xor.b32 %r1276, %r1275, %r1265;
|
||
add.s32 %r1277, %r1133, %r1250;
|
||
add.s32 %r1278, %r1277, %r1276;
|
||
add.s32 %r1279, %r1278, -1069501632;
|
||
shf.l.wrap.b32 %r1280, %r1279, %r1279, 9;
|
||
add.s32 %r1281, %r1280, %r1273;
|
||
xor.b32 %r1282, %r1281, %r1273;
|
||
and.b32 %r1283, %r1282, %r1265;
|
||
xor.b32 %r1284, %r1283, %r1273;
|
||
add.s32 %r1285, %r1143, %r1258;
|
||
add.s32 %r1286, %r1285, %r1284;
|
||
add.s32 %r1287, %r1286, 643717713;
|
||
shf.l.wrap.b32 %r1288, %r1287, %r1287, 14;
|
||
add.s32 %r1289, %r1288, %r1281;
|
||
xor.b32 %r1290, %r1289, %r1281;
|
||
and.b32 %r1291, %r1290, %r1273;
|
||
xor.b32 %r1292, %r1291, %r1281;
|
||
add.s32 %r1293, %r1120, %r1265;
|
||
add.s32 %r1294, %r1293, %r1292;
|
||
add.s32 %r1295, %r1294, -373897302;
|
||
shf.l.wrap.b32 %r1296, %r1295, %r1295, 20;
|
||
add.s32 %r1297, %r1296, %r1289;
|
||
xor.b32 %r1298, %r1297, %r1289;
|
||
and.b32 %r1299, %r1298, %r1281;
|
||
xor.b32 %r1300, %r1299, %r1289;
|
||
add.s32 %r1301, %r1131, %r1273;
|
||
add.s32 %r1302, %r1301, %r1300;
|
||
add.s32 %r1303, %r1302, -701558691;
|
||
shf.l.wrap.b32 %r1304, %r1303, %r1303, 5;
|
||
add.s32 %r1305, %r1304, %r1297;
|
||
xor.b32 %r1306, %r1305, %r1297;
|
||
and.b32 %r1307, %r1306, %r1289;
|
||
xor.b32 %r1308, %r1307, %r1297;
|
||
add.s32 %r1309, %r1141, %r1281;
|
||
add.s32 %r1310, %r1309, %r1308;
|
||
add.s32 %r1311, %r1310, 38016083;
|
||
shf.l.wrap.b32 %r1312, %r1311, %r1311, 9;
|
||
add.s32 %r1313, %r1312, %r1305;
|
||
xor.b32 %r1314, %r1313, %r1305;
|
||
and.b32 %r1315, %r1314, %r1297;
|
||
xor.b32 %r1316, %r1315, %r1305;
|
||
add.s32 %r1317, %r1289, %r1316;
|
||
add.s32 %r1318, %r1317, -660478335;
|
||
shf.l.wrap.b32 %r1319, %r1318, %r1318, 14;
|
||
add.s32 %r1320, %r1319, %r1313;
|
||
xor.b32 %r1321, %r1320, %r1313;
|
||
and.b32 %r1322, %r1321, %r1305;
|
||
xor.b32 %r1323, %r1322, %r1313;
|
||
add.s32 %r1324, %r1129, %r1297;
|
||
add.s32 %r1325, %r1324, %r1323;
|
||
add.s32 %r1326, %r1325, -405537848;
|
||
shf.l.wrap.b32 %r1327, %r1326, %r1326, 20;
|
||
add.s32 %r1328, %r1327, %r1320;
|
||
xor.b32 %r1329, %r1328, %r1320;
|
||
and.b32 %r1330, %r1329, %r1313;
|
||
xor.b32 %r1331, %r1330, %r1320;
|
||
add.s32 %r1332, %r1139, %r1305;
|
||
add.s32 %r1333, %r1332, %r1331;
|
||
add.s32 %r1334, %r1333, 568446438;
|
||
shf.l.wrap.b32 %r1335, %r1334, %r1334, 5;
|
||
add.s32 %r1336, %r1335, %r1328;
|
||
xor.b32 %r1337, %r1336, %r1328;
|
||
and.b32 %r1338, %r1337, %r1320;
|
||
xor.b32 %r1339, %r1338, %r1328;
|
||
add.s32 %r1340, %r1115, %r1313;
|
||
add.s32 %r1341, %r1340, %r1339;
|
||
add.s32 %r1342, %r1341, -1019803690;
|
||
shf.l.wrap.b32 %r1343, %r1342, %r1342, 9;
|
||
add.s32 %r1344, %r1343, %r1336;
|
||
xor.b32 %r1345, %r1344, %r1336;
|
||
and.b32 %r1346, %r1345, %r1328;
|
||
xor.b32 %r1347, %r1346, %r1336;
|
||
add.s32 %r1348, %r1127, %r1320;
|
||
add.s32 %r1349, %r1348, %r1347;
|
||
add.s32 %r1350, %r1349, -187363961;
|
||
shf.l.wrap.b32 %r1351, %r1350, %r1350, 14;
|
||
add.s32 %r1352, %r1351, %r1344;
|
||
xor.b32 %r1353, %r1352, %r1344;
|
||
and.b32 %r1354, %r1353, %r1336;
|
||
xor.b32 %r1355, %r1354, %r1344;
|
||
add.s32 %r1356, %r1137, %r1328;
|
||
add.s32 %r1357, %r1356, %r1355;
|
||
add.s32 %r1358, %r1357, 1163531501;
|
||
shf.l.wrap.b32 %r1359, %r1358, %r1358, 20;
|
||
add.s32 %r1360, %r1359, %r1352;
|
||
xor.b32 %r1361, %r1360, %r1352;
|
||
and.b32 %r1362, %r1361, %r1344;
|
||
xor.b32 %r1363, %r1362, %r1352;
|
||
add.s32 %r1364, %r1147, %r1336;
|
||
add.s32 %r1365, %r1364, %r1363;
|
||
add.s32 %r1366, %r1365, -1444681467;
|
||
shf.l.wrap.b32 %r1367, %r1366, %r1366, 5;
|
||
add.s32 %r1368, %r1367, %r1360;
|
||
xor.b32 %r1369, %r1368, %r1360;
|
||
and.b32 %r1370, %r1369, %r1352;
|
||
xor.b32 %r1371, %r1370, %r1360;
|
||
add.s32 %r1372, %r1125, %r1344;
|
||
add.s32 %r1373, %r1372, %r1371;
|
||
add.s32 %r1374, %r1373, -51403784;
|
||
shf.l.wrap.b32 %r1375, %r1374, %r1374, 9;
|
||
add.s32 %r1376, %r1375, %r1368;
|
||
xor.b32 %r1377, %r1376, %r1368;
|
||
and.b32 %r1378, %r1377, %r1360;
|
||
xor.b32 %r1379, %r1378, %r1368;
|
||
add.s32 %r1380, %r1135, %r1352;
|
||
add.s32 %r1381, %r1380, %r1379;
|
||
add.s32 %r1382, %r1381, 1735328473;
|
||
shf.l.wrap.b32 %r1383, %r1382, %r1382, 14;
|
||
add.s32 %r1384, %r1383, %r1376;
|
||
xor.b32 %r1385, %r1384, %r1376;
|
||
and.b32 %r1386, %r1385, %r1368;
|
||
xor.b32 %r1387, %r1386, %r1376;
|
||
add.s32 %r1388, %r1145, %r1360;
|
||
add.s32 %r1389, %r1388, %r1387;
|
||
add.s32 %r1390, %r1389, -1926607734;
|
||
shf.l.wrap.b32 %r1391, %r1390, %r1390, 20;
|
||
add.s32 %r1392, %r1391, %r1384;
|
||
xor.b32 %r1393, %r1385, %r1392;
|
||
add.s32 %r1394, %r1131, %r1368;
|
||
add.s32 %r1395, %r1394, %r1393;
|
||
add.s32 %r1396, %r1395, -378558;
|
||
shf.l.wrap.b32 %r1397, %r1396, %r1396, 4;
|
||
add.s32 %r1398, %r1397, %r1392;
|
||
xor.b32 %r1399, %r1392, %r1384;
|
||
xor.b32 %r1400, %r1399, %r1398;
|
||
add.s32 %r1401, %r1137, %r1376;
|
||
add.s32 %r1402, %r1401, %r1400;
|
||
add.s32 %r1403, %r1402, -2022574463;
|
||
shf.l.wrap.b32 %r1404, %r1403, %r1403, 11;
|
||
add.s32 %r1405, %r1404, %r1398;
|
||
xor.b32 %r1406, %r1398, %r1392;
|
||
xor.b32 %r1407, %r1406, %r1405;
|
||
add.s32 %r1408, %r1143, %r1384;
|
||
add.s32 %r1409, %r1408, %r1407;
|
||
add.s32 %r1410, %r1409, 1839030562;
|
||
shf.l.wrap.b32 %r1411, %r1410, %r1410, 16;
|
||
add.s32 %r1412, %r1411, %r1405;
|
||
xor.b32 %r1413, %r1405, %r1398;
|
||
xor.b32 %r1414, %r1413, %r1412;
|
||
add.s32 %r1415, %r1115, %r1392;
|
||
add.s32 %r1416, %r1415, %r1414;
|
||
add.s32 %r1417, %r1416, -35309556;
|
||
shf.l.wrap.b32 %r1418, %r1417, %r1417, 23;
|
||
add.s32 %r1419, %r1418, %r1412;
|
||
xor.b32 %r1420, %r1412, %r1405;
|
||
xor.b32 %r1421, %r1420, %r1419;
|
||
add.s32 %r1422, %r1123, %r1398;
|
||
add.s32 %r1423, %r1422, %r1421;
|
||
add.s32 %r1424, %r1423, -1530992060;
|
||
shf.l.wrap.b32 %r1425, %r1424, %r1424, 4;
|
||
add.s32 %r1426, %r1425, %r1419;
|
||
xor.b32 %r1427, %r1419, %r1412;
|
||
xor.b32 %r1428, %r1427, %r1426;
|
||
add.s32 %r1429, %r1129, %r1405;
|
||
add.s32 %r1430, %r1429, %r1428;
|
||
add.s32 %r1431, %r1430, 1272893353;
|
||
shf.l.wrap.b32 %r1432, %r1431, %r1431, 11;
|
||
add.s32 %r1433, %r1432, %r1426;
|
||
xor.b32 %r1434, %r1426, %r1419;
|
||
xor.b32 %r1435, %r1434, %r1433;
|
||
add.s32 %r1436, %r1135, %r1412;
|
||
add.s32 %r1437, %r1436, %r1435;
|
||
add.s32 %r1438, %r1437, -155497632;
|
||
shf.l.wrap.b32 %r1439, %r1438, %r1438, 16;
|
||
add.s32 %r1440, %r1439, %r1433;
|
||
xor.b32 %r1441, %r1433, %r1426;
|
||
xor.b32 %r1442, %r1441, %r1440;
|
||
add.s32 %r1443, %r1141, %r1419;
|
||
add.s32 %r1444, %r1443, %r1442;
|
||
add.s32 %r1445, %r1444, -1094730640;
|
||
shf.l.wrap.b32 %r1446, %r1445, %r1445, 23;
|
||
add.s32 %r1447, %r1446, %r1440;
|
||
xor.b32 %r1448, %r1440, %r1433;
|
||
xor.b32 %r1449, %r1448, %r1447;
|
||
add.s32 %r1450, %r1147, %r1426;
|
||
add.s32 %r1451, %r1450, %r1449;
|
||
add.s32 %r1452, %r1451, 681279174;
|
||
shf.l.wrap.b32 %r1453, %r1452, %r1452, 4;
|
||
add.s32 %r1454, %r1453, %r1447;
|
||
xor.b32 %r1455, %r1447, %r1440;
|
||
xor.b32 %r1456, %r1455, %r1454;
|
||
add.s32 %r1457, %r1120, %r1433;
|
||
add.s32 %r1458, %r1457, %r1456;
|
||
add.s32 %r1459, %r1458, -358537222;
|
||
shf.l.wrap.b32 %r1460, %r1459, %r1459, 11;
|
||
add.s32 %r1461, %r1460, %r1454;
|
||
xor.b32 %r1462, %r1454, %r1447;
|
||
xor.b32 %r1463, %r1462, %r1461;
|
||
add.s32 %r1464, %r1127, %r1440;
|
||
add.s32 %r1465, %r1464, %r1463;
|
||
add.s32 %r1466, %r1465, -722521979;
|
||
shf.l.wrap.b32 %r1467, %r1466, %r1466, 16;
|
||
add.s32 %r1468, %r1467, %r1461;
|
||
xor.b32 %r1469, %r1461, %r1454;
|
||
xor.b32 %r1470, %r1469, %r1468;
|
||
add.s32 %r1471, %r1133, %r1447;
|
||
add.s32 %r1472, %r1471, %r1470;
|
||
add.s32 %r1473, %r1472, 76029189;
|
||
shf.l.wrap.b32 %r1474, %r1473, %r1473, 23;
|
||
add.s32 %r1475, %r1474, %r1468;
|
||
xor.b32 %r1476, %r1468, %r1461;
|
||
xor.b32 %r1477, %r1476, %r1475;
|
||
add.s32 %r1478, %r1139, %r1454;
|
||
add.s32 %r1479, %r1478, %r1477;
|
||
add.s32 %r1480, %r1479, -640364487;
|
||
shf.l.wrap.b32 %r1481, %r1480, %r1480, 4;
|
||
add.s32 %r1482, %r1481, %r1475;
|
||
xor.b32 %r1483, %r1475, %r1468;
|
||
xor.b32 %r1484, %r1483, %r1482;
|
||
add.s32 %r1485, %r1145, %r1461;
|
||
add.s32 %r1486, %r1485, %r1484;
|
||
add.s32 %r1487, %r1486, -421815835;
|
||
shf.l.wrap.b32 %r1488, %r1487, %r1487, 11;
|
||
add.s32 %r1489, %r1488, %r1482;
|
||
xor.b32 %r1490, %r1482, %r1475;
|
||
xor.b32 %r1491, %r1490, %r1489;
|
||
add.s32 %r1492, %r1468, %r1491;
|
||
add.s32 %r1493, %r1492, 530742520;
|
||
shf.l.wrap.b32 %r1494, %r1493, %r1493, 16;
|
||
add.s32 %r1495, %r1494, %r1489;
|
||
xor.b32 %r1496, %r1489, %r1482;
|
||
xor.b32 %r1497, %r1496, %r1495;
|
||
add.s32 %r1498, %r1125, %r1475;
|
||
add.s32 %r1499, %r1498, %r1497;
|
||
add.s32 %r1500, %r1499, -995338651;
|
||
shf.l.wrap.b32 %r1501, %r1500, %r1500, 23;
|
||
add.s32 %r1502, %r1501, %r1495;
|
||
not.b32 %r1503, %r1489;
|
||
or.b32 %r1504, %r1502, %r1503;
|
||
xor.b32 %r1505, %r1504, %r1495;
|
||
add.s32 %r1506, %r1120, %r1482;
|
||
add.s32 %r1507, %r1506, %r1505;
|
||
add.s32 %r1508, %r1507, -198630844;
|
||
shf.l.wrap.b32 %r1509, %r1508, %r1508, 6;
|
||
add.s32 %r1510, %r1509, %r1502;
|
||
not.b32 %r1511, %r1495;
|
||
or.b32 %r1512, %r1510, %r1511;
|
||
xor.b32 %r1513, %r1512, %r1502;
|
||
add.s32 %r1514, %r1135, %r1489;
|
||
add.s32 %r1515, %r1514, %r1513;
|
||
add.s32 %r1516, %r1515, 1126891415;
|
||
shf.l.wrap.b32 %r1517, %r1516, %r1516, 10;
|
||
add.s32 %r1518, %r1517, %r1510;
|
||
not.b32 %r1519, %r1502;
|
||
or.b32 %r1520, %r1518, %r1519;
|
||
xor.b32 %r1521, %r1520, %r1510;
|
||
add.s32 %r1522, %r1115, %r1495;
|
||
add.s32 %r1523, %r1522, %r1521;
|
||
add.s32 %r1524, %r1523, -1416354905;
|
||
shf.l.wrap.b32 %r1525, %r1524, %r1524, 15;
|
||
add.s32 %r1526, %r1525, %r1518;
|
||
not.b32 %r1527, %r1510;
|
||
or.b32 %r1528, %r1526, %r1527;
|
||
xor.b32 %r1529, %r1528, %r1518;
|
||
add.s32 %r1530, %r1131, %r1502;
|
||
add.s32 %r1531, %r1530, %r1529;
|
||
add.s32 %r1532, %r1531, -57434055;
|
||
shf.l.wrap.b32 %r1533, %r1532, %r1532, 21;
|
||
add.s32 %r1534, %r1533, %r1526;
|
||
not.b32 %r1535, %r1518;
|
||
or.b32 %r1536, %r1534, %r1535;
|
||
xor.b32 %r1537, %r1536, %r1526;
|
||
add.s32 %r1538, %r1145, %r1510;
|
||
add.s32 %r1539, %r1538, %r1537;
|
||
add.s32 %r1540, %r1539, 1700485571;
|
||
shf.l.wrap.b32 %r1541, %r1540, %r1540, 6;
|
||
add.s32 %r1542, %r1541, %r1534;
|
||
not.b32 %r1543, %r1526;
|
||
or.b32 %r1544, %r1542, %r1543;
|
||
xor.b32 %r1545, %r1544, %r1534;
|
||
add.s32 %r1546, %r1127, %r1518;
|
||
add.s32 %r1547, %r1546, %r1545;
|
||
add.s32 %r1548, %r1547, -1894986606;
|
||
shf.l.wrap.b32 %r1549, %r1548, %r1548, 10;
|
||
add.s32 %r1550, %r1549, %r1542;
|
||
not.b32 %r1551, %r1534;
|
||
or.b32 %r1552, %r1550, %r1551;
|
||
xor.b32 %r1553, %r1552, %r1542;
|
||
add.s32 %r1554, %r1141, %r1526;
|
||
add.s32 %r1555, %r1554, %r1553;
|
||
add.s32 %r1556, %r1555, -1051523;
|
||
shf.l.wrap.b32 %r1557, %r1556, %r1556, 15;
|
||
add.s32 %r1558, %r1557, %r1550;
|
||
not.b32 %r1559, %r1542;
|
||
or.b32 %r1560, %r1558, %r1559;
|
||
xor.b32 %r1561, %r1560, %r1550;
|
||
add.s32 %r1562, %r1123, %r1534;
|
||
add.s32 %r1563, %r1562, %r1561;
|
||
add.s32 %r1564, %r1563, -2054922799;
|
||
shf.l.wrap.b32 %r1565, %r1564, %r1564, 21;
|
||
add.s32 %r1566, %r1565, %r1558;
|
||
not.b32 %r1567, %r1550;
|
||
or.b32 %r1568, %r1566, %r1567;
|
||
xor.b32 %r1569, %r1568, %r1558;
|
||
add.s32 %r1570, %r1137, %r1542;
|
||
add.s32 %r1571, %r1570, %r1569;
|
||
add.s32 %r1572, %r1571, 1873313359;
|
||
shf.l.wrap.b32 %r1573, %r1572, %r1572, 6;
|
||
add.s32 %r1574, %r1573, %r1566;
|
||
not.b32 %r1575, %r1558;
|
||
or.b32 %r1576, %r1574, %r1575;
|
||
xor.b32 %r1577, %r1576, %r1566;
|
||
add.s32 %r1578, %r1550, %r1577;
|
||
add.s32 %r1579, %r1578, -30611744;
|
||
shf.l.wrap.b32 %r1580, %r1579, %r1579, 10;
|
||
add.s32 %r1581, %r1580, %r1574;
|
||
not.b32 %r1582, %r1566;
|
||
or.b32 %r1583, %r1581, %r1582;
|
||
xor.b32 %r1584, %r1583, %r1574;
|
||
add.s32 %r1585, %r1133, %r1558;
|
||
add.s32 %r1586, %r1585, %r1584;
|
||
add.s32 %r1587, %r1586, -1560198380;
|
||
shf.l.wrap.b32 %r1588, %r1587, %r1587, 15;
|
||
add.s32 %r1589, %r1588, %r1581;
|
||
not.b32 %r1590, %r1574;
|
||
or.b32 %r1591, %r1589, %r1590;
|
||
xor.b32 %r1592, %r1591, %r1581;
|
||
add.s32 %r1593, %r1147, %r1566;
|
||
add.s32 %r1594, %r1593, %r1592;
|
||
add.s32 %r1595, %r1594, 1309151649;
|
||
shf.l.wrap.b32 %r1596, %r1595, %r1595, 21;
|
||
add.s32 %r1597, %r1596, %r1589;
|
||
not.b32 %r1598, %r1581;
|
||
or.b32 %r1599, %r1597, %r1598;
|
||
xor.b32 %r1600, %r1599, %r1589;
|
||
add.s32 %r1601, %r1129, %r1574;
|
||
add.s32 %r1602, %r1601, %r1600;
|
||
add.s32 %r1603, %r1602, -145523070;
|
||
shf.l.wrap.b32 %r1604, %r1603, %r1603, 6;
|
||
add.s32 %r1605, %r1604, %r1597;
|
||
not.b32 %r1606, %r1589;
|
||
or.b32 %r1607, %r1605, %r1606;
|
||
xor.b32 %r1608, %r1607, %r1597;
|
||
add.s32 %r1609, %r1143, %r1581;
|
||
add.s32 %r1610, %r1609, %r1608;
|
||
add.s32 %r1611, %r1610, -1120210379;
|
||
shf.l.wrap.b32 %r1612, %r1611, %r1611, 10;
|
||
add.s32 %r1613, %r1612, %r1605;
|
||
not.b32 %r1614, %r1597;
|
||
or.b32 %r1615, %r1613, %r1614;
|
||
xor.b32 %r1616, %r1615, %r1605;
|
||
add.s32 %r1617, %r1125, %r1589;
|
||
add.s32 %r1618, %r1617, %r1616;
|
||
add.s32 %r1619, %r1618, 718787259;
|
||
shf.l.wrap.b32 %r1620, %r1619, %r1619, 15;
|
||
add.s32 %r1621, %r1620, %r1613;
|
||
not.b32 %r1622, %r1605;
|
||
or.b32 %r1623, %r1621, %r1622;
|
||
xor.b32 %r1624, %r1623, %r1613;
|
||
add.s32 %r1625, %r1139, %r1597;
|
||
add.s32 %r1626, %r1625, %r1624;
|
||
add.s32 %r1627, %r1626, -343485551;
|
||
shf.l.wrap.b32 %r1628, %r1627, %r1627, 21;
|
||
add.s32 %r1629, %r1605, 1732584193;
|
||
add.s32 %r1630, %r1621, %r1628;
|
||
add.s32 %r1631, %r1630, -271733879;
|
||
add.s32 %r1632, %r1621, -1732584194;
|
||
add.s32 %r1633, %r1613, 271733878;
|
||
shl.b64 %rd9, %rd12, 4;
|
||
add.s64 %rd10, %rd11, %rd9;
|
||
st.global.u32 [%rd10], %r1629;
|
||
st.global.u32 [%rd10+4], %r1631;
|
||
st.global.u32 [%rd10+8], %r1632;
|
||
st.global.u32 [%rd10+12], %r1633;
|
||
|
||
BB1_312:
|
||
ret;
|
||
|
||
BB1_146:
|
||
setp.ne.s32 %p86, %r1649, 15;
|
||
@%p86 bra BB1_156;
|
||
|
||
and.b32 %r1645, %r1645, 16777215;
|
||
bra.uni BB1_156;
|
||
}
|
||
|
||
// .globl m00500_loop
|
||
.entry m00500_loop(
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_0,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_1,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_2,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_3,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_4,
|
||
.param .u64 .ptr .global .align 1 m00500_loop_param_5,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_6,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_7,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_8,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_9,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_10,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_11,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_12,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_13,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_14,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_15,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_16,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_17,
|
||
.param .u64 .ptr .global .align 1 m00500_loop_param_18,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_19,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_20,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_21,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_22,
|
||
.param .u64 .ptr .global .align 4 m00500_loop_param_23,
|
||
.param .u32 m00500_loop_param_24,
|
||
.param .u32 m00500_loop_param_25,
|
||
.param .u32 m00500_loop_param_26,
|
||
.param .u32 m00500_loop_param_27,
|
||
.param .u32 m00500_loop_param_28,
|
||
.param .u32 m00500_loop_param_29,
|
||
.param .u32 m00500_loop_param_30,
|
||
.param .u32 m00500_loop_param_31,
|
||
.param .u32 m00500_loop_param_32,
|
||
.param .u32 m00500_loop_param_33,
|
||
.param .u32 m00500_loop_param_34
|
||
)
|
||
{
|
||
.reg .pred %p<83>;
|
||
.reg .b32 %r<2039>;
|
||
.reg .b64 %rd<14>;
|
||
|
||
|
||
ld.param.u64 %rd3, [m00500_loop_param_0];
|
||
ld.param.u64 %rd4, [m00500_loop_param_4];
|
||
ld.param.u64 %rd5, [m00500_loop_param_17];
|
||
ld.param.u32 %r216, [m00500_loop_param_27];
|
||
ld.param.u32 %r1039, [m00500_loop_param_28];
|
||
ld.param.u32 %r218, [m00500_loop_param_29];
|
||
ld.param.u32 %r219, [m00500_loop_param_34];
|
||
mov.b32 %r220, %envreg3;
|
||
mov.u32 %r221, %ctaid.x;
|
||
mov.u32 %r222, %ntid.x;
|
||
mad.lo.s32 %r223, %r221, %r222, %r220;
|
||
mov.u32 %r224, %tid.x;
|
||
add.s32 %r1, %r223, %r224;
|
||
setp.ge.u32 %p1, %r1, %r219;
|
||
@%p1 bra BB2_133;
|
||
|
||
mul.wide.u32 %rd6, %r1, 80;
|
||
add.s64 %rd7, %rd3, %rd6;
|
||
ld.global.u32 %r2, [%rd7+64];
|
||
ld.global.u32 %r3, [%rd7];
|
||
ld.global.u32 %r4, [%rd7+4];
|
||
ld.global.u32 %r5, [%rd7+8];
|
||
ld.global.u32 %r6, [%rd7+12];
|
||
mul.wide.u32 %rd8, %r216, 180;
|
||
add.s64 %rd1, %rd5, %rd8;
|
||
ld.global.u32 %r7, [%rd1];
|
||
ld.global.u32 %r8, [%rd1+4];
|
||
mul.wide.u32 %rd9, %r1, 16;
|
||
add.s64 %rd2, %rd4, %rd9;
|
||
ld.global.u32 %r1149, [%rd2];
|
||
ld.global.u32 %r1944, [%rd2+4];
|
||
ld.global.u32 %r1943, [%rd2+8];
|
||
ld.global.u32 %r1942, [%rd2+12];
|
||
setp.eq.s32 %p2, %r218, 0;
|
||
@%p2 bra BB2_132;
|
||
|
||
and.b32 %r226, %r2, 3;
|
||
shl.b32 %r227, %r226, 3;
|
||
mov.u32 %r228, 128;
|
||
shl.b32 %r229, %r228, %r227;
|
||
setp.lt.u32 %p3, %r2, 4;
|
||
selp.b32 %r230, %r229, 0, %p3;
|
||
or.b32 %r13, %r230, %r3;
|
||
and.b32 %r231, %r2, -4;
|
||
setp.eq.s32 %p4, %r231, 4;
|
||
selp.b32 %r232, %r229, 0, %p4;
|
||
or.b32 %r14, %r232, %r4;
|
||
setp.eq.s32 %p5, %r231, 8;
|
||
selp.b32 %r233, %r229, 0, %p5;
|
||
or.b32 %r15, %r233, %r5;
|
||
setp.gt.u32 %p6, %r2, 11;
|
||
selp.b32 %r234, %r229, 0, %p6;
|
||
or.b32 %r16, %r234, %r6;
|
||
ld.global.u32 %r235, [%rd1+128];
|
||
add.s32 %r17, %r235, 16;
|
||
and.b32 %r236, %r17, 3;
|
||
mov.u32 %r237, 4;
|
||
sub.s32 %r238, %r237, %r236;
|
||
shl.b32 %r239, %r238, 2;
|
||
mov.u32 %r240, 1985229328;
|
||
shr.u32 %r241, %r240, %r239;
|
||
and.b32 %r19, %r241, 65535;
|
||
shr.u32 %r20, %r17, 2;
|
||
add.s32 %r21, %r17, %r2;
|
||
sub.s32 %r242, %r237, %r226;
|
||
shl.b32 %r243, %r242, 2;
|
||
shr.u32 %r244, %r240, %r243;
|
||
and.b32 %r22, %r244, 65535;
|
||
shr.u32 %r23, %r2, 2;
|
||
add.s32 %r24, %r235, %r2;
|
||
mov.u32 %r1040, 0;
|
||
// inline asm
|
||
prmt.b32 %r437, %r3, %r4, %r19;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r441, %r4, %r5, %r19;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r445, %r5, %r6, %r19;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r268, %r7, %r8, %r22;
|
||
// inline asm
|
||
mov.u32 %r1153, %r1149;
|
||
mov.u32 %r1982, %r1944;
|
||
mov.u32 %r1984, %r1943;
|
||
mov.u32 %r1986, %r1942;
|
||
|
||
BB2_3:
|
||
mov.u32 %r1191, %r1986;
|
||
mov.u32 %r1176, %r1984;
|
||
mov.u32 %r1160, %r1982;
|
||
mov.u32 %r30, %r1191;
|
||
mov.u32 %r29, %r1176;
|
||
mov.u32 %r28, %r1160;
|
||
mov.u32 %r1076, %r1153;
|
||
mov.u32 %r27, %r1076;
|
||
and.b32 %r245, %r1039, 1;
|
||
setp.eq.b32 %p7, %r245, 1;
|
||
mul.wide.u32 %rd10, %r1039, -1431655765;
|
||
shr.u64 %rd11, %rd10, 33;
|
||
cvt.u32.u64 %r246, %rd11;
|
||
mul.lo.s32 %r247, %r246, 3;
|
||
sub.s32 %r31, %r1039, %r247;
|
||
mul.wide.u32 %rd12, %r1039, 613566757;
|
||
shr.u64 %rd13, %rd12, 32;
|
||
cvt.u32.u64 %r248, %rd13;
|
||
sub.s32 %r249, %r1039, %r248;
|
||
shr.u32 %r250, %r249, 1;
|
||
add.s32 %r251, %r250, %r248;
|
||
shr.u32 %r252, %r251, 2;
|
||
mul.lo.s32 %r253, %r252, 7;
|
||
sub.s32 %r32, %r1039, %r253;
|
||
@!%p7 bra BB2_81;
|
||
bra.uni BB2_4;
|
||
|
||
BB2_81:
|
||
setp.eq.s32 %p51, %r31, 0;
|
||
setp.eq.s32 %p52, %r32, 0;
|
||
or.pred %p53, %p51, %p52;
|
||
@%p53 bra BB2_105;
|
||
bra.uni BB2_82;
|
||
|
||
BB2_105:
|
||
add.s32 %r1037, %r235, 16;
|
||
setp.ne.s32 %p67, %r31, 0;
|
||
mov.u32 %r518, 0;
|
||
mov.u32 %r1972, %r518;
|
||
mov.u32 %r1973, %r518;
|
||
mov.u32 %r1974, %r518;
|
||
mov.u32 %r514, %r518;
|
||
mov.u32 %r513, %r518;
|
||
mov.u32 %r1979, %r518;
|
||
mov.u32 %r1980, %r518;
|
||
mov.u32 %r1072, %r1037;
|
||
mov.u32 %r1089, %r27;
|
||
mov.u32 %r1152, %r1089;
|
||
mov.u32 %r1172, %r28;
|
||
mov.u32 %r1187, %r29;
|
||
mov.u32 %r1201, %r30;
|
||
mov.u32 %r1783, %r8;
|
||
mov.u32 %r1801, %r7;
|
||
mov.u32 %r1971, %r518;
|
||
mov.u32 %r1975, %r514;
|
||
mov.u32 %r1976, %r513;
|
||
mov.u32 %r1977, %r1801;
|
||
mov.u32 %r1978, %r1783;
|
||
mov.u32 %r1981, %r1172;
|
||
mov.u32 %r1983, %r1187;
|
||
mov.u32 %r1985, %r1201;
|
||
@%p67 bra BB2_107;
|
||
|
||
add.s32 %r1038, %r2, 16;
|
||
setp.ne.s32 %p68, %r32, 0;
|
||
selp.b32 %r1980, %r6, 0, %p68;
|
||
selp.b32 %r1979, %r5, 0, %p68;
|
||
selp.b32 %r143, %r4, 0, %p68;
|
||
selp.b32 %r144, %r3, 0, %p68;
|
||
selp.b32 %r145, %r1038, 16, %p68;
|
||
mov.u32 %r524, 0;
|
||
mov.u32 %r1972, %r524;
|
||
mov.u32 %r1973, %r524;
|
||
mov.u32 %r1974, %r524;
|
||
mov.u32 %r520, %r524;
|
||
mov.u32 %r519, %r524;
|
||
mov.u32 %r1072, %r145;
|
||
mov.u32 %r1152, %r27;
|
||
mov.u32 %r1981, %r28;
|
||
mov.u32 %r1983, %r29;
|
||
mov.u32 %r1985, %r30;
|
||
mov.u32 %r1971, %r524;
|
||
mov.u32 %r1975, %r520;
|
||
mov.u32 %r1976, %r519;
|
||
mov.u32 %r1977, %r144;
|
||
mov.u32 %r1978, %r143;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_4:
|
||
mov.u32 %r263, 0;
|
||
mov.u32 %r2027, %r263;
|
||
mov.u32 %r2028, %r263;
|
||
mov.u32 %r2029, %r263;
|
||
mov.u32 %r2030, %r263;
|
||
mov.u32 %r2031, %r263;
|
||
mov.u32 %r2032, %r263;
|
||
mov.u32 %r2033, %r263;
|
||
mov.u32 %r2034, %r263;
|
||
mov.u32 %r2035, %r263;
|
||
setp.eq.s32 %p8, %r31, 0;
|
||
mov.u32 %r1058, %r2;
|
||
mov.u32 %r1157, %r3;
|
||
mov.u32 %r2026, %r263;
|
||
mov.u32 %r2036, %r4;
|
||
mov.u32 %r2037, %r5;
|
||
mov.u32 %r2038, %r6;
|
||
@%p8 bra BB2_32;
|
||
|
||
mov.u32 %r285, 0;
|
||
// inline asm
|
||
prmt.b32 %r264, %r285, %r7, %r22;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r272, %r8, %r285, %r22;
|
||
// inline asm
|
||
mov.u32 %r284, %r285;
|
||
mov.u32 %r2028, %r285;
|
||
mov.u32 %r2029, %r285;
|
||
mov.u32 %r2030, %r285;
|
||
mov.u32 %r280, %r285;
|
||
mov.u32 %r2032, %r285;
|
||
mov.u32 %r2033, %r285;
|
||
mov.u32 %r2034, %r285;
|
||
mov.u32 %r2035, %r285;
|
||
setp.gt.s32 %p9, %r23, 5;
|
||
@%p9 bra BB2_15;
|
||
|
||
setp.gt.s32 %p17, %r23, 2;
|
||
@%p17 bra BB2_11;
|
||
|
||
setp.eq.s32 %p21, %r23, 0;
|
||
@%p21 bra BB2_31;
|
||
|
||
setp.eq.s32 %p22, %r23, 1;
|
||
@%p22 bra BB2_30;
|
||
bra.uni BB2_9;
|
||
|
||
BB2_30:
|
||
or.b32 %r38, %r264, %r4;
|
||
mov.u32 %r368, 0;
|
||
mov.u32 %r367, %r368;
|
||
mov.u32 %r2028, %r368;
|
||
mov.u32 %r2029, %r368;
|
||
mov.u32 %r2030, %r368;
|
||
mov.u32 %r2031, %r368;
|
||
mov.u32 %r2032, %r368;
|
||
mov.u32 %r2033, %r368;
|
||
mov.u32 %r2034, %r368;
|
||
mov.u32 %r2035, %r368;
|
||
mov.u32 %r1054, %r24;
|
||
mov.u32 %r1058, %r1054;
|
||
mov.u32 %r1105, %r3;
|
||
mov.u32 %r1157, %r1105;
|
||
mov.u32 %r1225, %r268;
|
||
mov.u32 %r1236, %r272;
|
||
mov.u32 %r2026, %r368;
|
||
mov.u32 %r2027, %r367;
|
||
mov.u32 %r2036, %r38;
|
||
mov.u32 %r2037, %r1225;
|
||
mov.u32 %r2038, %r1236;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_82:
|
||
mov.u32 %r460, 0;
|
||
// inline asm
|
||
prmt.b32 %r433, %r460, %r3, %r19;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r449, %r6, %r460, %r19;
|
||
// inline asm
|
||
mov.u32 %r459, %r460;
|
||
mov.u32 %r1973, %r460;
|
||
mov.u32 %r457, %r460;
|
||
mov.u32 %r456, %r460;
|
||
mov.u32 %r455, %r460;
|
||
mov.u32 %r1979, %r460;
|
||
mov.u32 %r1980, %r460;
|
||
setp.gt.s32 %p54, %r20, 4;
|
||
@%p54 bra BB2_91;
|
||
|
||
setp.gt.s32 %p61, %r20, 1;
|
||
@%p61 bra BB2_87;
|
||
|
||
setp.eq.s32 %p65, %r20, 0;
|
||
@%p65 bra BB2_104;
|
||
bra.uni BB2_85;
|
||
|
||
BB2_104:
|
||
or.b32 %r1152, %r433, %r27;
|
||
mov.u32 %r510, 0;
|
||
mov.u32 %r509, %r510;
|
||
mov.u32 %r1973, %r510;
|
||
mov.u32 %r1974, %r510;
|
||
mov.u32 %r506, %r510;
|
||
mov.u32 %r505, %r510;
|
||
mov.u32 %r1979, %r510;
|
||
mov.u32 %r1980, %r510;
|
||
mov.u32 %r1071, %r21;
|
||
mov.u32 %r1072, %r1071;
|
||
mov.u32 %r1981, %r437;
|
||
mov.u32 %r1983, %r441;
|
||
mov.u32 %r1985, %r445;
|
||
mov.u32 %r1753, %r449;
|
||
mov.u32 %r1782, %r8;
|
||
mov.u32 %r1971, %r510;
|
||
mov.u32 %r1972, %r509;
|
||
mov.u32 %r1975, %r506;
|
||
mov.u32 %r1976, %r505;
|
||
mov.u32 %r1977, %r1753;
|
||
mov.u32 %r1978, %r1782;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_15:
|
||
setp.gt.s32 %p10, %r23, 8;
|
||
@%p10 bra BB2_20;
|
||
|
||
setp.eq.s32 %p14, %r23, 6;
|
||
@%p14 bra BB2_27;
|
||
|
||
setp.eq.s32 %p15, %r23, 7;
|
||
@%p15 bra BB2_26;
|
||
bra.uni BB2_18;
|
||
|
||
BB2_26:
|
||
mov.u32 %r320, 0;
|
||
mov.u32 %r319, %r320;
|
||
mov.u32 %r2030, %r320;
|
||
mov.u32 %r317, %r320;
|
||
mov.u32 %r2032, %r320;
|
||
mov.u32 %r2033, %r320;
|
||
mov.u32 %r2034, %r320;
|
||
mov.u32 %r1048, %r24;
|
||
mov.u32 %r1058, %r1048;
|
||
mov.u32 %r1099, %r3;
|
||
mov.u32 %r1157, %r1099;
|
||
mov.u32 %r1211, %r6;
|
||
mov.u32 %r2028, %r268;
|
||
mov.u32 %r2029, %r272;
|
||
mov.u32 %r1257, %r5;
|
||
mov.u32 %r1282, %r4;
|
||
mov.u32 %r2035, %r264;
|
||
mov.u32 %r2026, %r320;
|
||
mov.u32 %r2027, %r319;
|
||
mov.u32 %r2031, %r317;
|
||
mov.u32 %r2036, %r1282;
|
||
mov.u32 %r2037, %r1257;
|
||
mov.u32 %r2038, %r1211;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_91:
|
||
setp.gt.s32 %p55, %r20, 6;
|
||
@%p55 bra BB2_95;
|
||
|
||
setp.eq.s32 %p59, %r20, 5;
|
||
@%p59 bra BB2_101;
|
||
bra.uni BB2_93;
|
||
|
||
BB2_101:
|
||
or.b32 %r135, %r433, %r8;
|
||
mov.u32 %r476, 0;
|
||
mov.u32 %r475, %r476;
|
||
mov.u32 %r474, %r476;
|
||
mov.u32 %r473, %r476;
|
||
mov.u32 %r1066, %r21;
|
||
mov.u32 %r1072, %r1066;
|
||
mov.u32 %r1152, %r27;
|
||
mov.u32 %r1981, %r28;
|
||
mov.u32 %r1983, %r29;
|
||
mov.u32 %r1985, %r30;
|
||
mov.u32 %r1979, %r437;
|
||
mov.u32 %r1980, %r441;
|
||
mov.u32 %r1973, %r445;
|
||
mov.u32 %r1974, %r449;
|
||
mov.u32 %r1800, %r7;
|
||
mov.u32 %r1971, %r476;
|
||
mov.u32 %r1972, %r475;
|
||
mov.u32 %r1975, %r474;
|
||
mov.u32 %r1976, %r473;
|
||
mov.u32 %r1977, %r1800;
|
||
mov.u32 %r1978, %r135;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_11:
|
||
setp.eq.s32 %p18, %r23, 3;
|
||
@%p18 bra BB2_29;
|
||
|
||
setp.eq.s32 %p19, %r23, 4;
|
||
@%p19 bra BB2_28;
|
||
bra.uni BB2_13;
|
||
|
||
BB2_28:
|
||
mov.u32 %r341, 0;
|
||
mov.u32 %r340, %r341;
|
||
mov.u32 %r2028, %r341;
|
||
mov.u32 %r2029, %r341;
|
||
mov.u32 %r2030, %r341;
|
||
mov.u32 %r336, %r341;
|
||
mov.u32 %r2035, %r341;
|
||
mov.u32 %r1051, %r24;
|
||
mov.u32 %r1058, %r1051;
|
||
mov.u32 %r1102, %r3;
|
||
mov.u32 %r1157, %r1102;
|
||
mov.u32 %r1214, %r6;
|
||
mov.u32 %r2033, %r268;
|
||
mov.u32 %r2034, %r272;
|
||
mov.u32 %r1260, %r5;
|
||
mov.u32 %r1285, %r4;
|
||
mov.u32 %r2032, %r264;
|
||
mov.u32 %r2026, %r341;
|
||
mov.u32 %r2027, %r340;
|
||
mov.u32 %r2031, %r336;
|
||
mov.u32 %r2036, %r1285;
|
||
mov.u32 %r2037, %r1260;
|
||
mov.u32 %r2038, %r1214;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_87:
|
||
setp.eq.s32 %p62, %r20, 2;
|
||
@%p62 bra BB2_103;
|
||
|
||
setp.eq.s32 %p63, %r20, 3;
|
||
@%p63 bra BB2_102;
|
||
bra.uni BB2_89;
|
||
|
||
BB2_102:
|
||
or.b32 %r1985, %r433, %r30;
|
||
mov.u32 %r487, 0;
|
||
mov.u32 %r486, %r487;
|
||
mov.u32 %r1973, %r487;
|
||
mov.u32 %r1974, %r487;
|
||
mov.u32 %r483, %r487;
|
||
mov.u32 %r482, %r487;
|
||
mov.u32 %r1068, %r21;
|
||
mov.u32 %r1072, %r1068;
|
||
mov.u32 %r1152, %r27;
|
||
mov.u32 %r1981, %r28;
|
||
mov.u32 %r1983, %r29;
|
||
mov.u32 %r1684, %r437;
|
||
mov.u32 %r1693, %r441;
|
||
mov.u32 %r1979, %r445;
|
||
mov.u32 %r1980, %r449;
|
||
mov.u32 %r1971, %r487;
|
||
mov.u32 %r1972, %r486;
|
||
mov.u32 %r1975, %r483;
|
||
mov.u32 %r1976, %r482;
|
||
mov.u32 %r1977, %r1684;
|
||
mov.u32 %r1978, %r1693;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_20:
|
||
setp.eq.s32 %p11, %r23, 9;
|
||
@%p11 bra BB2_25;
|
||
|
||
setp.eq.s32 %p12, %r23, 10;
|
||
@%p12 bra BB2_24;
|
||
bra.uni BB2_22;
|
||
|
||
BB2_24:
|
||
mov.u32 %r299, 0;
|
||
mov.u32 %r2028, %r299;
|
||
mov.u32 %r2029, %r299;
|
||
mov.u32 %r2032, %r299;
|
||
mov.u32 %r2033, %r299;
|
||
mov.u32 %r2034, %r299;
|
||
mov.u32 %r2035, %r299;
|
||
mov.u32 %r1045, %r24;
|
||
mov.u32 %r1058, %r1045;
|
||
mov.u32 %r1096, %r3;
|
||
mov.u32 %r1157, %r1096;
|
||
mov.u32 %r1208, %r6;
|
||
mov.u32 %r1216, %r268;
|
||
mov.u32 %r1227, %r272;
|
||
mov.u32 %r1254, %r5;
|
||
mov.u32 %r1279, %r4;
|
||
mov.u32 %r2030, %r264;
|
||
mov.u32 %r2026, %r1227;
|
||
mov.u32 %r2027, %r299;
|
||
mov.u32 %r2031, %r1216;
|
||
mov.u32 %r2036, %r1279;
|
||
mov.u32 %r2037, %r1254;
|
||
mov.u32 %r2038, %r1208;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_95:
|
||
setp.eq.s32 %p56, %r20, 7;
|
||
@%p56 bra BB2_100;
|
||
|
||
setp.eq.s32 %p57, %r20, 8;
|
||
@%p57 bra BB2_99;
|
||
bra.uni BB2_97;
|
||
|
||
BB2_99:
|
||
mov.u32 %r466, 0;
|
||
mov.u32 %r1979, %r466;
|
||
mov.u32 %r1980, %r466;
|
||
mov.u32 %r1063, %r21;
|
||
mov.u32 %r1072, %r1063;
|
||
mov.u32 %r1152, %r27;
|
||
mov.u32 %r1981, %r28;
|
||
mov.u32 %r1983, %r29;
|
||
mov.u32 %r1985, %r30;
|
||
mov.u32 %r1679, %r437;
|
||
mov.u32 %r1688, %r441;
|
||
mov.u32 %r1697, %r445;
|
||
mov.u32 %r1973, %r433;
|
||
mov.u32 %r1745, %r449;
|
||
mov.u32 %r1779, %r8;
|
||
mov.u32 %r1797, %r7;
|
||
mov.u32 %r1971, %r1745;
|
||
mov.u32 %r1972, %r466;
|
||
mov.u32 %r1974, %r1679;
|
||
mov.u32 %r1975, %r1688;
|
||
mov.u32 %r1976, %r1697;
|
||
mov.u32 %r1977, %r1797;
|
||
mov.u32 %r1978, %r1779;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_85:
|
||
setp.eq.s32 %p66, %r20, 1;
|
||
mov.u32 %r1061, %r21;
|
||
mov.u32 %r1072, %r1061;
|
||
mov.u32 %r1079, %r27;
|
||
mov.u32 %r1152, %r1079;
|
||
mov.u32 %r1163, %r28;
|
||
mov.u32 %r1983, %r29;
|
||
mov.u32 %r1985, %r30;
|
||
mov.u32 %r1777, %r8;
|
||
mov.u32 %r1795, %r7;
|
||
mov.u32 %r1813, %r455;
|
||
mov.u32 %r1827, %r456;
|
||
mov.u32 %r1840, %r457;
|
||
mov.u32 %r1864, %r459;
|
||
mov.u32 %r1880, %r460;
|
||
mov.u32 %r1971, %r1880;
|
||
mov.u32 %r1972, %r1864;
|
||
mov.u32 %r1974, %r1840;
|
||
mov.u32 %r1975, %r1827;
|
||
mov.u32 %r1976, %r1813;
|
||
mov.u32 %r1977, %r1795;
|
||
mov.u32 %r1978, %r1777;
|
||
mov.u32 %r1981, %r1163;
|
||
@%p66 bra BB2_86;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_86:
|
||
or.b32 %r1981, %r433, %r28;
|
||
mov.u32 %r502, 0;
|
||
mov.u32 %r501, %r502;
|
||
mov.u32 %r1973, %r502;
|
||
mov.u32 %r1974, %r502;
|
||
mov.u32 %r498, %r502;
|
||
mov.u32 %r497, %r502;
|
||
mov.u32 %r1979, %r502;
|
||
mov.u32 %r1980, %r502;
|
||
mov.u32 %r1070, %r21;
|
||
mov.u32 %r1072, %r1070;
|
||
mov.u32 %r1152, %r27;
|
||
mov.u32 %r1983, %r437;
|
||
mov.u32 %r1985, %r441;
|
||
mov.u32 %r1704, %r445;
|
||
mov.u32 %r1752, %r449;
|
||
mov.u32 %r1971, %r502;
|
||
mov.u32 %r1972, %r501;
|
||
mov.u32 %r1975, %r498;
|
||
mov.u32 %r1976, %r497;
|
||
mov.u32 %r1977, %r1704;
|
||
mov.u32 %r1978, %r1752;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_93:
|
||
setp.eq.s32 %p60, %r20, 6;
|
||
mov.u32 %r1059, %r21;
|
||
mov.u32 %r1072, %r1059;
|
||
mov.u32 %r1077, %r27;
|
||
mov.u32 %r1152, %r1077;
|
||
mov.u32 %r1161, %r28;
|
||
mov.u32 %r1177, %r29;
|
||
mov.u32 %r1192, %r30;
|
||
mov.u32 %r1775, %r8;
|
||
mov.u32 %r1793, %r7;
|
||
mov.u32 %r1811, %r455;
|
||
mov.u32 %r1825, %r456;
|
||
mov.u32 %r1838, %r457;
|
||
mov.u32 %r1862, %r459;
|
||
mov.u32 %r1878, %r460;
|
||
mov.u32 %r1971, %r1878;
|
||
mov.u32 %r1972, %r1862;
|
||
mov.u32 %r1974, %r1838;
|
||
mov.u32 %r1975, %r1825;
|
||
mov.u32 %r1976, %r1811;
|
||
mov.u32 %r1977, %r1793;
|
||
mov.u32 %r1978, %r1775;
|
||
mov.u32 %r1981, %r1161;
|
||
mov.u32 %r1983, %r1177;
|
||
mov.u32 %r1985, %r1192;
|
||
@%p60 bra BB2_94;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_94:
|
||
mov.u32 %r472, 0;
|
||
mov.u32 %r471, %r472;
|
||
mov.u32 %r470, %r472;
|
||
mov.u32 %r1065, %r21;
|
||
mov.u32 %r1072, %r1065;
|
||
mov.u32 %r1152, %r27;
|
||
mov.u32 %r1981, %r28;
|
||
mov.u32 %r1983, %r29;
|
||
mov.u32 %r1985, %r30;
|
||
mov.u32 %r1980, %r437;
|
||
mov.u32 %r1973, %r441;
|
||
mov.u32 %r1699, %r445;
|
||
mov.u32 %r1979, %r433;
|
||
mov.u32 %r1747, %r449;
|
||
mov.u32 %r1781, %r8;
|
||
mov.u32 %r1799, %r7;
|
||
mov.u32 %r1971, %r472;
|
||
mov.u32 %r1972, %r471;
|
||
mov.u32 %r1974, %r1699;
|
||
mov.u32 %r1975, %r1747;
|
||
mov.u32 %r1976, %r470;
|
||
mov.u32 %r1977, %r1799;
|
||
mov.u32 %r1978, %r1781;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_31:
|
||
or.b32 %r39, %r264, %r3;
|
||
mov.u32 %r378, 0;
|
||
mov.u32 %r377, %r378;
|
||
mov.u32 %r2028, %r378;
|
||
mov.u32 %r2029, %r378;
|
||
mov.u32 %r2030, %r378;
|
||
mov.u32 %r2031, %r378;
|
||
mov.u32 %r2032, %r378;
|
||
mov.u32 %r2033, %r378;
|
||
mov.u32 %r2034, %r378;
|
||
mov.u32 %r2035, %r378;
|
||
mov.u32 %r1055, %r24;
|
||
mov.u32 %r1058, %r1055;
|
||
mov.u32 %r1157, %r39;
|
||
mov.u32 %r1215, %r6;
|
||
mov.u32 %r1226, %r268;
|
||
mov.u32 %r1237, %r272;
|
||
mov.u32 %r2026, %r378;
|
||
mov.u32 %r2027, %r377;
|
||
mov.u32 %r2036, %r1226;
|
||
mov.u32 %r2037, %r1237;
|
||
mov.u32 %r2038, %r1215;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_9:
|
||
setp.eq.s32 %p23, %r23, 2;
|
||
mov.u32 %r1043, %r24;
|
||
mov.u32 %r1058, %r1043;
|
||
mov.u32 %r1094, %r3;
|
||
mov.u32 %r1157, %r1094;
|
||
mov.u32 %r1206, %r6;
|
||
mov.u32 %r1252, %r5;
|
||
mov.u32 %r1277, %r4;
|
||
mov.u32 %r1362, %r280;
|
||
mov.u32 %r1417, %r284;
|
||
mov.u32 %r1434, %r285;
|
||
mov.u32 %r2026, %r1434;
|
||
mov.u32 %r2027, %r1417;
|
||
mov.u32 %r2031, %r1362;
|
||
mov.u32 %r2036, %r1277;
|
||
mov.u32 %r2037, %r1252;
|
||
mov.u32 %r2038, %r1206;
|
||
@%p23 bra BB2_10;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_10:
|
||
or.b32 %r37, %r264, %r5;
|
||
mov.u32 %r358, 0;
|
||
mov.u32 %r357, %r358;
|
||
mov.u32 %r2028, %r358;
|
||
mov.u32 %r2029, %r358;
|
||
mov.u32 %r2030, %r358;
|
||
mov.u32 %r2031, %r358;
|
||
mov.u32 %r2033, %r358;
|
||
mov.u32 %r2034, %r358;
|
||
mov.u32 %r2035, %r358;
|
||
mov.u32 %r1053, %r24;
|
||
mov.u32 %r1058, %r1053;
|
||
mov.u32 %r1104, %r3;
|
||
mov.u32 %r1157, %r1104;
|
||
mov.u32 %r1224, %r268;
|
||
mov.u32 %r2032, %r272;
|
||
mov.u32 %r1287, %r4;
|
||
mov.u32 %r2026, %r358;
|
||
mov.u32 %r2027, %r357;
|
||
mov.u32 %r2036, %r1287;
|
||
mov.u32 %r2037, %r37;
|
||
mov.u32 %r2038, %r1224;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_27:
|
||
mov.u32 %r327, 0;
|
||
mov.u32 %r326, %r327;
|
||
mov.u32 %r2029, %r327;
|
||
mov.u32 %r2030, %r327;
|
||
mov.u32 %r323, %r327;
|
||
mov.u32 %r2032, %r327;
|
||
mov.u32 %r2033, %r327;
|
||
mov.u32 %r1049, %r24;
|
||
mov.u32 %r1058, %r1049;
|
||
mov.u32 %r1100, %r3;
|
||
mov.u32 %r1157, %r1100;
|
||
mov.u32 %r1212, %r6;
|
||
mov.u32 %r2035, %r268;
|
||
mov.u32 %r2028, %r272;
|
||
mov.u32 %r1258, %r5;
|
||
mov.u32 %r1283, %r4;
|
||
mov.u32 %r2034, %r264;
|
||
mov.u32 %r2026, %r327;
|
||
mov.u32 %r2027, %r326;
|
||
mov.u32 %r2031, %r323;
|
||
mov.u32 %r2036, %r1283;
|
||
mov.u32 %r2037, %r1258;
|
||
mov.u32 %r2038, %r1212;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_18:
|
||
setp.eq.s32 %p16, %r23, 8;
|
||
mov.u32 %r1041, %r24;
|
||
mov.u32 %r1058, %r1041;
|
||
mov.u32 %r1092, %r3;
|
||
mov.u32 %r1157, %r1092;
|
||
mov.u32 %r1204, %r6;
|
||
mov.u32 %r1250, %r5;
|
||
mov.u32 %r1275, %r4;
|
||
mov.u32 %r1360, %r280;
|
||
mov.u32 %r1415, %r284;
|
||
mov.u32 %r1432, %r285;
|
||
mov.u32 %r2026, %r1432;
|
||
mov.u32 %r2027, %r1415;
|
||
mov.u32 %r2031, %r1360;
|
||
mov.u32 %r2036, %r1275;
|
||
mov.u32 %r2037, %r1250;
|
||
mov.u32 %r2038, %r1204;
|
||
@%p16 bra BB2_19;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_19:
|
||
mov.u32 %r313, 0;
|
||
mov.u32 %r312, %r313;
|
||
mov.u32 %r311, %r313;
|
||
mov.u32 %r2032, %r313;
|
||
mov.u32 %r2033, %r313;
|
||
mov.u32 %r2034, %r313;
|
||
mov.u32 %r2035, %r313;
|
||
mov.u32 %r1047, %r24;
|
||
mov.u32 %r1058, %r1047;
|
||
mov.u32 %r1098, %r3;
|
||
mov.u32 %r1157, %r1098;
|
||
mov.u32 %r1210, %r6;
|
||
mov.u32 %r2029, %r268;
|
||
mov.u32 %r2030, %r272;
|
||
mov.u32 %r1256, %r5;
|
||
mov.u32 %r1281, %r4;
|
||
mov.u32 %r2028, %r264;
|
||
mov.u32 %r2026, %r313;
|
||
mov.u32 %r2027, %r312;
|
||
mov.u32 %r2031, %r311;
|
||
mov.u32 %r2036, %r1281;
|
||
mov.u32 %r2037, %r1256;
|
||
mov.u32 %r2038, %r1210;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_29:
|
||
or.b32 %r36, %r264, %r6;
|
||
mov.u32 %r349, 0;
|
||
mov.u32 %r348, %r349;
|
||
mov.u32 %r2028, %r349;
|
||
mov.u32 %r2029, %r349;
|
||
mov.u32 %r2030, %r349;
|
||
mov.u32 %r2031, %r349;
|
||
mov.u32 %r2034, %r349;
|
||
mov.u32 %r2035, %r349;
|
||
mov.u32 %r1052, %r24;
|
||
mov.u32 %r1058, %r1052;
|
||
mov.u32 %r1103, %r3;
|
||
mov.u32 %r1157, %r1103;
|
||
mov.u32 %r2032, %r268;
|
||
mov.u32 %r2033, %r272;
|
||
mov.u32 %r1261, %r5;
|
||
mov.u32 %r1286, %r4;
|
||
mov.u32 %r2026, %r349;
|
||
mov.u32 %r2027, %r348;
|
||
mov.u32 %r2036, %r1286;
|
||
mov.u32 %r2037, %r1261;
|
||
mov.u32 %r2038, %r36;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_13:
|
||
setp.eq.s32 %p20, %r23, 5;
|
||
mov.u32 %r1042, %r24;
|
||
mov.u32 %r1058, %r1042;
|
||
mov.u32 %r1093, %r3;
|
||
mov.u32 %r1157, %r1093;
|
||
mov.u32 %r1205, %r6;
|
||
mov.u32 %r1251, %r5;
|
||
mov.u32 %r1276, %r4;
|
||
mov.u32 %r1361, %r280;
|
||
mov.u32 %r1416, %r284;
|
||
mov.u32 %r1433, %r285;
|
||
mov.u32 %r2026, %r1433;
|
||
mov.u32 %r2027, %r1416;
|
||
mov.u32 %r2031, %r1361;
|
||
mov.u32 %r2036, %r1276;
|
||
mov.u32 %r2037, %r1251;
|
||
mov.u32 %r2038, %r1205;
|
||
@%p20 bra BB2_14;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_14:
|
||
mov.u32 %r334, 0;
|
||
mov.u32 %r333, %r334;
|
||
mov.u32 %r2028, %r334;
|
||
mov.u32 %r2029, %r334;
|
||
mov.u32 %r2030, %r334;
|
||
mov.u32 %r329, %r334;
|
||
mov.u32 %r2032, %r334;
|
||
mov.u32 %r1050, %r24;
|
||
mov.u32 %r1058, %r1050;
|
||
mov.u32 %r1101, %r3;
|
||
mov.u32 %r1157, %r1101;
|
||
mov.u32 %r1213, %r6;
|
||
mov.u32 %r2034, %r268;
|
||
mov.u32 %r2035, %r272;
|
||
mov.u32 %r1259, %r5;
|
||
mov.u32 %r1284, %r4;
|
||
mov.u32 %r2033, %r264;
|
||
mov.u32 %r2026, %r334;
|
||
mov.u32 %r2027, %r333;
|
||
mov.u32 %r2031, %r329;
|
||
mov.u32 %r2036, %r1284;
|
||
mov.u32 %r2037, %r1259;
|
||
mov.u32 %r2038, %r1213;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_103:
|
||
or.b32 %r1983, %r433, %r29;
|
||
mov.u32 %r494, 0;
|
||
mov.u32 %r493, %r494;
|
||
mov.u32 %r1973, %r494;
|
||
mov.u32 %r1974, %r494;
|
||
mov.u32 %r490, %r494;
|
||
mov.u32 %r489, %r494;
|
||
mov.u32 %r1980, %r494;
|
||
mov.u32 %r1069, %r21;
|
||
mov.u32 %r1072, %r1069;
|
||
mov.u32 %r1152, %r27;
|
||
mov.u32 %r1981, %r28;
|
||
mov.u32 %r1985, %r437;
|
||
mov.u32 %r1694, %r441;
|
||
mov.u32 %r1703, %r445;
|
||
mov.u32 %r1979, %r449;
|
||
mov.u32 %r1971, %r494;
|
||
mov.u32 %r1972, %r493;
|
||
mov.u32 %r1975, %r490;
|
||
mov.u32 %r1976, %r489;
|
||
mov.u32 %r1977, %r1694;
|
||
mov.u32 %r1978, %r1703;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_89:
|
||
setp.eq.s32 %p64, %r20, 4;
|
||
mov.u32 %r1060, %r21;
|
||
mov.u32 %r1072, %r1060;
|
||
mov.u32 %r1078, %r27;
|
||
mov.u32 %r1152, %r1078;
|
||
mov.u32 %r1162, %r28;
|
||
mov.u32 %r1178, %r29;
|
||
mov.u32 %r1193, %r30;
|
||
mov.u32 %r1776, %r8;
|
||
mov.u32 %r1794, %r7;
|
||
mov.u32 %r1812, %r455;
|
||
mov.u32 %r1826, %r456;
|
||
mov.u32 %r1839, %r457;
|
||
mov.u32 %r1863, %r459;
|
||
mov.u32 %r1879, %r460;
|
||
mov.u32 %r1971, %r1879;
|
||
mov.u32 %r1972, %r1863;
|
||
mov.u32 %r1974, %r1839;
|
||
mov.u32 %r1975, %r1826;
|
||
mov.u32 %r1976, %r1812;
|
||
mov.u32 %r1977, %r1794;
|
||
mov.u32 %r1978, %r1776;
|
||
mov.u32 %r1981, %r1162;
|
||
mov.u32 %r1983, %r1178;
|
||
mov.u32 %r1985, %r1193;
|
||
@%p64 bra BB2_90;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_90:
|
||
or.b32 %r136, %r433, %r7;
|
||
mov.u32 %r481, 0;
|
||
mov.u32 %r480, %r481;
|
||
mov.u32 %r1974, %r481;
|
||
mov.u32 %r478, %r481;
|
||
mov.u32 %r477, %r481;
|
||
mov.u32 %r1067, %r21;
|
||
mov.u32 %r1072, %r1067;
|
||
mov.u32 %r1152, %r27;
|
||
mov.u32 %r1981, %r28;
|
||
mov.u32 %r1983, %r29;
|
||
mov.u32 %r1985, %r30;
|
||
mov.u32 %r1683, %r437;
|
||
mov.u32 %r1979, %r441;
|
||
mov.u32 %r1980, %r445;
|
||
mov.u32 %r1973, %r449;
|
||
mov.u32 %r1971, %r481;
|
||
mov.u32 %r1972, %r480;
|
||
mov.u32 %r1975, %r478;
|
||
mov.u32 %r1976, %r477;
|
||
mov.u32 %r1977, %r136;
|
||
mov.u32 %r1978, %r1683;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_25:
|
||
mov.u32 %r306, 0;
|
||
mov.u32 %r305, %r306;
|
||
mov.u32 %r2028, %r306;
|
||
mov.u32 %r2032, %r306;
|
||
mov.u32 %r2033, %r306;
|
||
mov.u32 %r2034, %r306;
|
||
mov.u32 %r2035, %r306;
|
||
mov.u32 %r1046, %r24;
|
||
mov.u32 %r1058, %r1046;
|
||
mov.u32 %r1097, %r3;
|
||
mov.u32 %r1157, %r1097;
|
||
mov.u32 %r1209, %r6;
|
||
mov.u32 %r2030, %r268;
|
||
mov.u32 %r1228, %r272;
|
||
mov.u32 %r1255, %r5;
|
||
mov.u32 %r1280, %r4;
|
||
mov.u32 %r2029, %r264;
|
||
mov.u32 %r2026, %r306;
|
||
mov.u32 %r2027, %r305;
|
||
mov.u32 %r2031, %r1228;
|
||
mov.u32 %r2036, %r1280;
|
||
mov.u32 %r2037, %r1255;
|
||
mov.u32 %r2038, %r1209;
|
||
bra.uni BB2_32;
|
||
|
||
BB2_22:
|
||
setp.ne.s32 %p13, %r23, 11;
|
||
mov.u32 %r1058, %r24;
|
||
mov.u32 %r1091, %r3;
|
||
mov.u32 %r1157, %r1091;
|
||
mov.u32 %r1203, %r6;
|
||
mov.u32 %r1249, %r5;
|
||
mov.u32 %r1274, %r4;
|
||
mov.u32 %r2026, %r285;
|
||
mov.u32 %r2027, %r284;
|
||
mov.u32 %r2031, %r280;
|
||
mov.u32 %r2036, %r1274;
|
||
mov.u32 %r2037, %r1249;
|
||
mov.u32 %r2038, %r1203;
|
||
@%p13 bra BB2_32;
|
||
|
||
mov.u32 %r2028, 0;
|
||
mov.u32 %r2029, %r2028;
|
||
mov.u32 %r2030, %r2028;
|
||
mov.u32 %r2032, %r2028;
|
||
mov.u32 %r2033, %r2028;
|
||
mov.u32 %r2034, %r2028;
|
||
mov.u32 %r2035, %r2028;
|
||
mov.u32 %r1044, %r24;
|
||
mov.u32 %r1058, %r1044;
|
||
mov.u32 %r1095, %r3;
|
||
mov.u32 %r1157, %r1095;
|
||
mov.u32 %r1207, %r6;
|
||
mov.u32 %r1253, %r5;
|
||
mov.u32 %r1278, %r4;
|
||
mov.u32 %r2026, %r268;
|
||
mov.u32 %r2027, %r272;
|
||
mov.u32 %r2031, %r264;
|
||
mov.u32 %r2036, %r1278;
|
||
mov.u32 %r2037, %r1253;
|
||
mov.u32 %r2038, %r1207;
|
||
|
||
BB2_32:
|
||
mov.u32 %r1238, %r2038;
|
||
mov.u32 %r1262, %r2037;
|
||
mov.u32 %r1288, %r2036;
|
||
mov.u32 %r1435, %r2026;
|
||
mov.u32 %r2000, %r1435;
|
||
mov.u32 %r2001, %r2027;
|
||
mov.u32 %r2002, %r2028;
|
||
mov.u32 %r2003, %r2029;
|
||
mov.u32 %r2004, %r2030;
|
||
mov.u32 %r2005, %r2031;
|
||
mov.u32 %r2006, %r2032;
|
||
mov.u32 %r2007, %r2033;
|
||
mov.u32 %r2008, %r2034;
|
||
mov.u32 %r2009, %r2035;
|
||
mov.u32 %r2010, %r1288;
|
||
mov.u32 %r2011, %r1262;
|
||
mov.u32 %r2012, %r1238;
|
||
mov.u32 %r1106, %r1157;
|
||
mov.u32 %r1155, %r1106;
|
||
mov.u32 %r1056, %r1058;
|
||
mov.u32 %r1057, %r1056;
|
||
setp.eq.s32 %p24, %r32, 0;
|
||
@%p24 bra BB2_57;
|
||
|
||
and.b32 %r400, %r1057, 3;
|
||
sub.s32 %r402, %r237, %r400;
|
||
shl.b32 %r403, %r402, 2;
|
||
shr.u32 %r405, %r240, %r403;
|
||
and.b32 %r398, %r405, 65535;
|
||
mov.u32 %r397, 0;
|
||
// inline asm
|
||
prmt.b32 %r379, %r397, %r3, %r398;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r383, %r3, %r4, %r398;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r387, %r4, %r5, %r398;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r391, %r5, %r6, %r398;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r395, %r6, %r397, %r398;
|
||
// inline asm
|
||
shr.u32 %r399, %r1057, 2;
|
||
setp.gt.s32 %p25, %r399, 4;
|
||
@%p25 bra BB2_42;
|
||
|
||
setp.gt.s32 %p32, %r399, 1;
|
||
@%p32 bra BB2_38;
|
||
|
||
setp.eq.s32 %p36, %r399, 0;
|
||
@%p36 bra BB2_55;
|
||
bra.uni BB2_36;
|
||
|
||
BB2_55:
|
||
or.b32 %r1156, %r379, %r1155;
|
||
mov.u32 %r2022, %r2009;
|
||
mov.u32 %r2021, %r2008;
|
||
mov.u32 %r2020, %r2007;
|
||
mov.u32 %r1374, %r2005;
|
||
mov.u32 %r1388, %r2004;
|
||
mov.u32 %r2016, %r2003;
|
||
mov.u32 %r2015, %r2002;
|
||
mov.u32 %r1431, %r2001;
|
||
mov.u32 %r1447, %r2000;
|
||
mov.u32 %r2023, %r383;
|
||
mov.u32 %r2024, %r387;
|
||
mov.u32 %r2025, %r391;
|
||
mov.u32 %r2019, %r395;
|
||
mov.u32 %r2013, %r1447;
|
||
mov.u32 %r2014, %r1431;
|
||
mov.u32 %r2017, %r1388;
|
||
mov.u32 %r2018, %r1374;
|
||
bra.uni BB2_56;
|
||
|
||
BB2_42:
|
||
setp.gt.s32 %p26, %r399, 6;
|
||
@%p26 bra BB2_46;
|
||
|
||
setp.eq.s32 %p30, %r399, 5;
|
||
@%p30 bra BB2_52;
|
||
bra.uni BB2_44;
|
||
|
||
BB2_52:
|
||
or.b32 %r2020, %r379, %r2007;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2025, %r2012;
|
||
mov.u32 %r2024, %r2011;
|
||
mov.u32 %r2023, %r2010;
|
||
mov.u32 %r2019, %r2006;
|
||
mov.u32 %r1369, %r2005;
|
||
mov.u32 %r1383, %r2004;
|
||
mov.u32 %r1426, %r2001;
|
||
mov.u32 %r1442, %r2000;
|
||
mov.u32 %r2021, %r383;
|
||
mov.u32 %r2022, %r387;
|
||
mov.u32 %r2015, %r391;
|
||
mov.u32 %r2016, %r395;
|
||
mov.u32 %r2013, %r1442;
|
||
mov.u32 %r2014, %r1426;
|
||
mov.u32 %r2017, %r1383;
|
||
mov.u32 %r2018, %r1369;
|
||
bra.uni BB2_56;
|
||
|
||
BB2_38:
|
||
setp.eq.s32 %p33, %r399, 2;
|
||
@%p33 bra BB2_54;
|
||
|
||
setp.eq.s32 %p34, %r399, 3;
|
||
@%p34 bra BB2_53;
|
||
bra.uni BB2_40;
|
||
|
||
BB2_53:
|
||
or.b32 %r2025, %r379, %r2012;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2024, %r2011;
|
||
mov.u32 %r2023, %r2010;
|
||
mov.u32 %r1371, %r2005;
|
||
mov.u32 %r1385, %r2004;
|
||
mov.u32 %r2016, %r2003;
|
||
mov.u32 %r2015, %r2002;
|
||
mov.u32 %r1428, %r2001;
|
||
mov.u32 %r1444, %r2000;
|
||
mov.u32 %r2019, %r383;
|
||
mov.u32 %r2020, %r387;
|
||
mov.u32 %r2021, %r391;
|
||
mov.u32 %r2022, %r395;
|
||
mov.u32 %r2013, %r1444;
|
||
mov.u32 %r2014, %r1428;
|
||
mov.u32 %r2017, %r1385;
|
||
mov.u32 %r2018, %r1371;
|
||
bra.uni BB2_56;
|
||
|
||
BB2_46:
|
||
setp.eq.s32 %p27, %r399, 7;
|
||
@%p27 bra BB2_51;
|
||
|
||
setp.eq.s32 %p28, %r399, 8;
|
||
@%p28 bra BB2_50;
|
||
bra.uni BB2_48;
|
||
|
||
BB2_50:
|
||
or.b32 %r2015, %r379, %r2002;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2025, %r2012;
|
||
mov.u32 %r2024, %r2011;
|
||
mov.u32 %r2023, %r2010;
|
||
mov.u32 %r2022, %r2009;
|
||
mov.u32 %r2021, %r2008;
|
||
mov.u32 %r2020, %r2007;
|
||
mov.u32 %r2019, %r2006;
|
||
mov.u32 %r1423, %r2001;
|
||
mov.u32 %r2016, %r383;
|
||
mov.u32 %r2017, %r387;
|
||
mov.u32 %r2018, %r391;
|
||
mov.u32 %r2013, %r395;
|
||
mov.u32 %r2014, %r1423;
|
||
bra.uni BB2_56;
|
||
|
||
BB2_36:
|
||
setp.eq.s32 %p37, %r399, 1;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2025, %r2012;
|
||
mov.u32 %r2024, %r2011;
|
||
mov.u32 %r2023, %r2010;
|
||
mov.u32 %r2022, %r2009;
|
||
mov.u32 %r2021, %r2008;
|
||
mov.u32 %r2020, %r2007;
|
||
mov.u32 %r2019, %r2006;
|
||
mov.u32 %r1367, %r2005;
|
||
mov.u32 %r1382, %r2004;
|
||
mov.u32 %r2016, %r2003;
|
||
mov.u32 %r2015, %r2002;
|
||
mov.u32 %r1422, %r2001;
|
||
mov.u32 %r1439, %r2000;
|
||
mov.u32 %r2013, %r1439;
|
||
mov.u32 %r2014, %r1422;
|
||
mov.u32 %r2017, %r1382;
|
||
mov.u32 %r2018, %r1367;
|
||
@%p37 bra BB2_37;
|
||
bra.uni BB2_56;
|
||
|
||
BB2_37:
|
||
or.b32 %r2023, %r379, %r2010;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2022, %r2009;
|
||
mov.u32 %r2021, %r2008;
|
||
mov.u32 %r1373, %r2005;
|
||
mov.u32 %r1387, %r2004;
|
||
mov.u32 %r2016, %r2003;
|
||
mov.u32 %r2015, %r2002;
|
||
mov.u32 %r1430, %r2001;
|
||
mov.u32 %r1446, %r2000;
|
||
mov.u32 %r2024, %r383;
|
||
mov.u32 %r2025, %r387;
|
||
mov.u32 %r2019, %r391;
|
||
mov.u32 %r2020, %r395;
|
||
mov.u32 %r2013, %r1446;
|
||
mov.u32 %r2014, %r1430;
|
||
mov.u32 %r2017, %r1387;
|
||
mov.u32 %r2018, %r1373;
|
||
bra.uni BB2_56;
|
||
|
||
BB2_44:
|
||
setp.eq.s32 %p31, %r399, 6;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2025, %r2012;
|
||
mov.u32 %r2024, %r2011;
|
||
mov.u32 %r2023, %r2010;
|
||
mov.u32 %r2022, %r2009;
|
||
mov.u32 %r2021, %r2008;
|
||
mov.u32 %r2020, %r2007;
|
||
mov.u32 %r2019, %r2006;
|
||
mov.u32 %r1365, %r2005;
|
||
mov.u32 %r1380, %r2004;
|
||
mov.u32 %r2016, %r2003;
|
||
mov.u32 %r2015, %r2002;
|
||
mov.u32 %r1420, %r2001;
|
||
mov.u32 %r1437, %r2000;
|
||
mov.u32 %r2013, %r1437;
|
||
mov.u32 %r2014, %r1420;
|
||
mov.u32 %r2017, %r1380;
|
||
mov.u32 %r2018, %r1365;
|
||
@%p31 bra BB2_45;
|
||
bra.uni BB2_56;
|
||
|
||
BB2_45:
|
||
or.b32 %r2021, %r379, %r2008;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2025, %r2012;
|
||
mov.u32 %r2024, %r2011;
|
||
mov.u32 %r2023, %r2010;
|
||
mov.u32 %r2020, %r2007;
|
||
mov.u32 %r2019, %r2006;
|
||
mov.u32 %r1368, %r2005;
|
||
mov.u32 %r1425, %r2001;
|
||
mov.u32 %r1441, %r2000;
|
||
mov.u32 %r2022, %r383;
|
||
mov.u32 %r2015, %r387;
|
||
mov.u32 %r2016, %r391;
|
||
mov.u32 %r2017, %r395;
|
||
mov.u32 %r2013, %r1441;
|
||
mov.u32 %r2014, %r1425;
|
||
mov.u32 %r2018, %r1368;
|
||
bra.uni BB2_56;
|
||
|
||
BB2_100:
|
||
mov.u32 %r469, 0;
|
||
mov.u32 %r468, %r469;
|
||
mov.u32 %r1979, %r469;
|
||
mov.u32 %r1064, %r21;
|
||
mov.u32 %r1072, %r1064;
|
||
mov.u32 %r1152, %r27;
|
||
mov.u32 %r1981, %r28;
|
||
mov.u32 %r1983, %r29;
|
||
mov.u32 %r1985, %r30;
|
||
mov.u32 %r1973, %r437;
|
||
mov.u32 %r1689, %r441;
|
||
mov.u32 %r1698, %r445;
|
||
mov.u32 %r1980, %r433;
|
||
mov.u32 %r1746, %r449;
|
||
mov.u32 %r1780, %r8;
|
||
mov.u32 %r1798, %r7;
|
||
mov.u32 %r1971, %r469;
|
||
mov.u32 %r1972, %r468;
|
||
mov.u32 %r1974, %r1689;
|
||
mov.u32 %r1975, %r1698;
|
||
mov.u32 %r1976, %r1746;
|
||
mov.u32 %r1977, %r1798;
|
||
mov.u32 %r1978, %r1780;
|
||
bra.uni BB2_107;
|
||
|
||
BB2_97:
|
||
setp.ne.s32 %p58, %r20, 9;
|
||
mov.u32 %r1072, %r21;
|
||
mov.u32 %r1152, %r27;
|
||
mov.u32 %r1971, %r460;
|
||
mov.u32 %r1972, %r459;
|
||
mov.u32 %r1974, %r457;
|
||
mov.u32 %r1975, %r456;
|
||
mov.u32 %r1976, %r455;
|
||
mov.u32 %r1977, %r7;
|
||
mov.u32 %r1978, %r8;
|
||
mov.u32 %r1981, %r28;
|
||
mov.u32 %r1983, %r29;
|
||
mov.u32 %r1985, %r30;
|
||
@%p58 bra BB2_107;
|
||
|
||
mov.u32 %r1973, 0;
|
||
mov.u32 %r1979, %r1973;
|
||
mov.u32 %r1980, %r1973;
|
||
mov.u32 %r1062, %r21;
|
||
mov.u32 %r1072, %r1062;
|
||
mov.u32 %r1152, %r27;
|
||
mov.u32 %r1981, %r28;
|
||
mov.u32 %r1983, %r29;
|
||
mov.u32 %r1985, %r30;
|
||
mov.u32 %r1778, %r8;
|
||
mov.u32 %r1796, %r7;
|
||
mov.u32 %r1971, %r445;
|
||
mov.u32 %r1972, %r449;
|
||
mov.u32 %r1974, %r433;
|
||
mov.u32 %r1975, %r437;
|
||
mov.u32 %r1976, %r441;
|
||
mov.u32 %r1977, %r1796;
|
||
mov.u32 %r1978, %r1778;
|
||
|
||
BB2_107:
|
||
mov.u32 %r1784, %r1978;
|
||
mov.u32 %r1802, %r1977;
|
||
mov.u32 %r1814, %r1976;
|
||
mov.u32 %r1828, %r1975;
|
||
mov.u32 %r1881, %r1971;
|
||
mov.u32 %r159, %r1881;
|
||
mov.u32 %r158, %r1972;
|
||
mov.u32 %r157, %r1973;
|
||
mov.u32 %r156, %r1974;
|
||
mov.u32 %r155, %r1828;
|
||
mov.u32 %r154, %r1814;
|
||
mov.u32 %r153, %r1802;
|
||
mov.u32 %r152, %r1784;
|
||
mov.u32 %r151, %r1979;
|
||
mov.u32 %r150, %r1980;
|
||
mov.u32 %r148, %r1981;
|
||
mov.u32 %r147, %r1983;
|
||
mov.u32 %r146, %r1985;
|
||
mov.u32 %r149, %r1152;
|
||
mov.u32 %r160, %r1072;
|
||
and.b32 %r546, %r160, 3;
|
||
sub.s32 %r548, %r237, %r546;
|
||
shl.b32 %r549, %r548, 2;
|
||
shr.u32 %r551, %r240, %r549;
|
||
and.b32 %r544, %r551, 65535;
|
||
mov.u32 %r543, 0;
|
||
// inline asm
|
||
prmt.b32 %r525, %r543, %r13, %r544;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r529, %r13, %r14, %r544;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r533, %r14, %r15, %r544;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r537, %r15, %r16, %r544;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r541, %r16, %r543, %r544;
|
||
// inline asm
|
||
shr.u32 %r545, %r160, 2;
|
||
setp.gt.s32 %p69, %r545, 4;
|
||
@%p69 bra BB2_116;
|
||
|
||
setp.gt.s32 %p76, %r545, 1;
|
||
@%p76 bra BB2_112;
|
||
|
||
setp.eq.s32 %p80, %r545, 0;
|
||
@%p80 bra BB2_129;
|
||
bra.uni BB2_110;
|
||
|
||
BB2_129:
|
||
or.b32 %r1151, %r525, %r149;
|
||
mov.u32 %r1967, %r150;
|
||
mov.u32 %r1966, %r151;
|
||
mov.u32 %r1965, %r152;
|
||
mov.u32 %r1824, %r154;
|
||
mov.u32 %r1837, %r155;
|
||
mov.u32 %r1961, %r156;
|
||
mov.u32 %r1960, %r157;
|
||
mov.u32 %r1877, %r158;
|
||
mov.u32 %r1892, %r159;
|
||
mov.u32 %r1968, %r529;
|
||
mov.u32 %r1969, %r533;
|
||
mov.u32 %r1970, %r537;
|
||
mov.u32 %r1964, %r541;
|
||
mov.u32 %r1958, %r1892;
|
||
mov.u32 %r1959, %r1877;
|
||
mov.u32 %r1962, %r1837;
|
||
mov.u32 %r1963, %r1824;
|
||
bra.uni BB2_130;
|
||
|
||
BB2_116:
|
||
setp.gt.s32 %p70, %r545, 6;
|
||
@%p70 bra BB2_120;
|
||
|
||
setp.eq.s32 %p74, %r545, 5;
|
||
@%p74 bra BB2_126;
|
||
bra.uni BB2_118;
|
||
|
||
BB2_126:
|
||
or.b32 %r1965, %r525, %r152;
|
||
mov.u32 %r1151, %r149;
|
||
mov.u32 %r1970, %r146;
|
||
mov.u32 %r1969, %r147;
|
||
mov.u32 %r1968, %r148;
|
||
mov.u32 %r1964, %r153;
|
||
mov.u32 %r1819, %r154;
|
||
mov.u32 %r1832, %r155;
|
||
mov.u32 %r1872, %r158;
|
||
mov.u32 %r1887, %r159;
|
||
mov.u32 %r1966, %r529;
|
||
mov.u32 %r1967, %r533;
|
||
mov.u32 %r1960, %r537;
|
||
mov.u32 %r1961, %r541;
|
||
mov.u32 %r1958, %r1887;
|
||
mov.u32 %r1959, %r1872;
|
||
mov.u32 %r1962, %r1832;
|
||
mov.u32 %r1963, %r1819;
|
||
bra.uni BB2_130;
|
||
|
||
BB2_112:
|
||
setp.eq.s32 %p77, %r545, 2;
|
||
@%p77 bra BB2_128;
|
||
|
||
setp.eq.s32 %p78, %r545, 3;
|
||
@%p78 bra BB2_127;
|
||
bra.uni BB2_114;
|
||
|
||
BB2_127:
|
||
or.b32 %r1970, %r525, %r146;
|
||
mov.u32 %r1151, %r149;
|
||
mov.u32 %r1969, %r147;
|
||
mov.u32 %r1968, %r148;
|
||
mov.u32 %r1821, %r154;
|
||
mov.u32 %r1834, %r155;
|
||
mov.u32 %r1961, %r156;
|
||
mov.u32 %r1960, %r157;
|
||
mov.u32 %r1874, %r158;
|
||
mov.u32 %r1889, %r159;
|
||
mov.u32 %r1964, %r529;
|
||
mov.u32 %r1965, %r533;
|
||
mov.u32 %r1966, %r537;
|
||
mov.u32 %r1967, %r541;
|
||
mov.u32 %r1958, %r1889;
|
||
mov.u32 %r1959, %r1874;
|
||
mov.u32 %r1962, %r1834;
|
||
mov.u32 %r1963, %r1821;
|
||
bra.uni BB2_130;
|
||
|
||
BB2_120:
|
||
setp.eq.s32 %p71, %r545, 7;
|
||
@%p71 bra BB2_125;
|
||
|
||
setp.eq.s32 %p72, %r545, 8;
|
||
@%p72 bra BB2_124;
|
||
bra.uni BB2_122;
|
||
|
||
BB2_124:
|
||
or.b32 %r1960, %r525, %r157;
|
||
mov.u32 %r1151, %r149;
|
||
mov.u32 %r1970, %r146;
|
||
mov.u32 %r1969, %r147;
|
||
mov.u32 %r1968, %r148;
|
||
mov.u32 %r1967, %r150;
|
||
mov.u32 %r1966, %r151;
|
||
mov.u32 %r1965, %r152;
|
||
mov.u32 %r1964, %r153;
|
||
mov.u32 %r1869, %r158;
|
||
mov.u32 %r1961, %r529;
|
||
mov.u32 %r1962, %r533;
|
||
mov.u32 %r1963, %r537;
|
||
mov.u32 %r1958, %r541;
|
||
mov.u32 %r1959, %r1869;
|
||
bra.uni BB2_130;
|
||
|
||
BB2_110:
|
||
setp.eq.s32 %p81, %r545, 1;
|
||
mov.u32 %r1138, %r149;
|
||
mov.u32 %r1151, %r1138;
|
||
mov.u32 %r1970, %r146;
|
||
mov.u32 %r1969, %r147;
|
||
mov.u32 %r1730, %r148;
|
||
mov.u32 %r1757, %r150;
|
||
mov.u32 %r1769, %r151;
|
||
mov.u32 %r1965, %r152;
|
||
mov.u32 %r1964, %r153;
|
||
mov.u32 %r1817, %r154;
|
||
mov.u32 %r1831, %r155;
|
||
mov.u32 %r1844, %r156;
|
||
mov.u32 %r1856, %r157;
|
||
mov.u32 %r1868, %r158;
|
||
mov.u32 %r1884, %r159;
|
||
mov.u32 %r1958, %r1884;
|
||
mov.u32 %r1959, %r1868;
|
||
mov.u32 %r1960, %r1856;
|
||
mov.u32 %r1961, %r1844;
|
||
mov.u32 %r1962, %r1831;
|
||
mov.u32 %r1963, %r1817;
|
||
mov.u32 %r1966, %r1769;
|
||
mov.u32 %r1967, %r1757;
|
||
mov.u32 %r1968, %r1730;
|
||
@%p81 bra BB2_111;
|
||
bra.uni BB2_130;
|
||
|
||
BB2_111:
|
||
or.b32 %r1968, %r525, %r148;
|
||
mov.u32 %r1151, %r149;
|
||
mov.u32 %r1967, %r150;
|
||
mov.u32 %r1966, %r151;
|
||
mov.u32 %r1823, %r154;
|
||
mov.u32 %r1836, %r155;
|
||
mov.u32 %r1961, %r156;
|
||
mov.u32 %r1960, %r157;
|
||
mov.u32 %r1876, %r158;
|
||
mov.u32 %r1891, %r159;
|
||
mov.u32 %r1969, %r529;
|
||
mov.u32 %r1970, %r533;
|
||
mov.u32 %r1964, %r537;
|
||
mov.u32 %r1965, %r541;
|
||
mov.u32 %r1958, %r1891;
|
||
mov.u32 %r1959, %r1876;
|
||
mov.u32 %r1962, %r1836;
|
||
mov.u32 %r1963, %r1823;
|
||
bra.uni BB2_130;
|
||
|
||
BB2_118:
|
||
setp.eq.s32 %p75, %r545, 6;
|
||
mov.u32 %r1136, %r149;
|
||
mov.u32 %r1151, %r1136;
|
||
mov.u32 %r1707, %r146;
|
||
mov.u32 %r1717, %r147;
|
||
mov.u32 %r1728, %r148;
|
||
mov.u32 %r1967, %r150;
|
||
mov.u32 %r1767, %r151;
|
||
mov.u32 %r1785, %r152;
|
||
mov.u32 %r1803, %r153;
|
||
mov.u32 %r1815, %r154;
|
||
mov.u32 %r1829, %r155;
|
||
mov.u32 %r1961, %r156;
|
||
mov.u32 %r1960, %r157;
|
||
mov.u32 %r1866, %r158;
|
||
mov.u32 %r1882, %r159;
|
||
mov.u32 %r1958, %r1882;
|
||
mov.u32 %r1959, %r1866;
|
||
mov.u32 %r1962, %r1829;
|
||
mov.u32 %r1963, %r1815;
|
||
mov.u32 %r1964, %r1803;
|
||
mov.u32 %r1965, %r1785;
|
||
mov.u32 %r1966, %r1767;
|
||
mov.u32 %r1968, %r1728;
|
||
mov.u32 %r1969, %r1717;
|
||
mov.u32 %r1970, %r1707;
|
||
@%p75 bra BB2_119;
|
||
bra.uni BB2_130;
|
||
|
||
BB2_119:
|
||
or.b32 %r1966, %r525, %r151;
|
||
mov.u32 %r1151, %r149;
|
||
mov.u32 %r1970, %r146;
|
||
mov.u32 %r1969, %r147;
|
||
mov.u32 %r1968, %r148;
|
||
mov.u32 %r1965, %r152;
|
||
mov.u32 %r1964, %r153;
|
||
mov.u32 %r1818, %r154;
|
||
mov.u32 %r1871, %r158;
|
||
mov.u32 %r1886, %r159;
|
||
mov.u32 %r1967, %r529;
|
||
mov.u32 %r1960, %r533;
|
||
mov.u32 %r1961, %r537;
|
||
mov.u32 %r1962, %r541;
|
||
mov.u32 %r1958, %r1886;
|
||
mov.u32 %r1959, %r1871;
|
||
mov.u32 %r1963, %r1818;
|
||
bra.uni BB2_130;
|
||
|
||
BB2_128:
|
||
or.b32 %r1969, %r525, %r147;
|
||
mov.u32 %r1151, %r149;
|
||
mov.u32 %r1968, %r148;
|
||
mov.u32 %r1967, %r150;
|
||
mov.u32 %r1822, %r154;
|
||
mov.u32 %r1835, %r155;
|
||
mov.u32 %r1961, %r156;
|
||
mov.u32 %r1960, %r157;
|
||
mov.u32 %r1875, %r158;
|
||
mov.u32 %r1890, %r159;
|
||
mov.u32 %r1970, %r529;
|
||
mov.u32 %r1964, %r533;
|
||
mov.u32 %r1965, %r537;
|
||
mov.u32 %r1966, %r541;
|
||
mov.u32 %r1958, %r1890;
|
||
mov.u32 %r1959, %r1875;
|
||
mov.u32 %r1962, %r1835;
|
||
mov.u32 %r1963, %r1822;
|
||
bra.uni BB2_130;
|
||
|
||
BB2_114:
|
||
setp.eq.s32 %p79, %r545, 4;
|
||
mov.u32 %r1137, %r149;
|
||
mov.u32 %r1151, %r1137;
|
||
mov.u32 %r1708, %r146;
|
||
mov.u32 %r1718, %r147;
|
||
mov.u32 %r1729, %r148;
|
||
mov.u32 %r1967, %r150;
|
||
mov.u32 %r1966, %r151;
|
||
mov.u32 %r1965, %r152;
|
||
mov.u32 %r1804, %r153;
|
||
mov.u32 %r1816, %r154;
|
||
mov.u32 %r1830, %r155;
|
||
mov.u32 %r1843, %r156;
|
||
mov.u32 %r1960, %r157;
|
||
mov.u32 %r1867, %r158;
|
||
mov.u32 %r1883, %r159;
|
||
mov.u32 %r1958, %r1883;
|
||
mov.u32 %r1959, %r1867;
|
||
mov.u32 %r1961, %r1843;
|
||
mov.u32 %r1962, %r1830;
|
||
mov.u32 %r1963, %r1816;
|
||
mov.u32 %r1964, %r1804;
|
||
mov.u32 %r1968, %r1729;
|
||
mov.u32 %r1969, %r1718;
|
||
mov.u32 %r1970, %r1708;
|
||
@%p79 bra BB2_115;
|
||
bra.uni BB2_130;
|
||
|
||
BB2_115:
|
||
or.b32 %r1964, %r525, %r153;
|
||
mov.u32 %r1151, %r149;
|
||
mov.u32 %r1970, %r146;
|
||
mov.u32 %r1969, %r147;
|
||
mov.u32 %r1968, %r148;
|
||
mov.u32 %r1820, %r154;
|
||
mov.u32 %r1833, %r155;
|
||
mov.u32 %r1961, %r156;
|
||
mov.u32 %r1873, %r158;
|
||
mov.u32 %r1888, %r159;
|
||
mov.u32 %r1965, %r529;
|
||
mov.u32 %r1966, %r533;
|
||
mov.u32 %r1967, %r537;
|
||
mov.u32 %r1960, %r541;
|
||
mov.u32 %r1958, %r1888;
|
||
mov.u32 %r1959, %r1873;
|
||
mov.u32 %r1962, %r1833;
|
||
mov.u32 %r1963, %r1820;
|
||
bra.uni BB2_130;
|
||
|
||
BB2_125:
|
||
or.b32 %r1967, %r525, %r150;
|
||
mov.u32 %r1151, %r149;
|
||
mov.u32 %r1970, %r146;
|
||
mov.u32 %r1969, %r147;
|
||
mov.u32 %r1968, %r148;
|
||
mov.u32 %r1966, %r151;
|
||
mov.u32 %r1965, %r152;
|
||
mov.u32 %r1964, %r153;
|
||
mov.u32 %r1870, %r158;
|
||
mov.u32 %r1885, %r159;
|
||
mov.u32 %r1960, %r529;
|
||
mov.u32 %r1961, %r533;
|
||
mov.u32 %r1962, %r537;
|
||
mov.u32 %r1963, %r541;
|
||
mov.u32 %r1958, %r1885;
|
||
mov.u32 %r1959, %r1870;
|
||
bra.uni BB2_130;
|
||
|
||
BB2_122:
|
||
setp.ne.s32 %p73, %r545, 9;
|
||
mov.u32 %r1151, %r149;
|
||
mov.u32 %r1958, %r159;
|
||
mov.u32 %r1959, %r158;
|
||
mov.u32 %r1960, %r157;
|
||
mov.u32 %r1961, %r156;
|
||
mov.u32 %r1962, %r155;
|
||
mov.u32 %r1963, %r154;
|
||
mov.u32 %r1964, %r153;
|
||
mov.u32 %r1965, %r152;
|
||
mov.u32 %r1966, %r151;
|
||
mov.u32 %r1967, %r150;
|
||
mov.u32 %r1968, %r148;
|
||
mov.u32 %r1969, %r147;
|
||
mov.u32 %r1970, %r146;
|
||
@%p73 bra BB2_130;
|
||
|
||
or.b32 %r1961, %r525, %r156;
|
||
mov.u32 %r1151, %r149;
|
||
mov.u32 %r1970, %r146;
|
||
mov.u32 %r1969, %r147;
|
||
mov.u32 %r1968, %r148;
|
||
mov.u32 %r1967, %r150;
|
||
mov.u32 %r1966, %r151;
|
||
mov.u32 %r1965, %r152;
|
||
mov.u32 %r1964, %r153;
|
||
mov.u32 %r1960, %r157;
|
||
mov.u32 %r1958, %r537;
|
||
mov.u32 %r1959, %r541;
|
||
mov.u32 %r1962, %r529;
|
||
mov.u32 %r1963, %r533;
|
||
|
||
BB2_130:
|
||
mov.u32 %r1945, %r1958;
|
||
mov.u32 %r1946, %r1959;
|
||
mov.u32 %r1947, %r1960;
|
||
mov.u32 %r1948, %r1961;
|
||
mov.u32 %r1949, %r1962;
|
||
mov.u32 %r1950, %r1963;
|
||
mov.u32 %r1951, %r1964;
|
||
mov.u32 %r1952, %r1965;
|
||
mov.u32 %r1953, %r1966;
|
||
mov.u32 %r1954, %r1967;
|
||
mov.u32 %r1955, %r1968;
|
||
mov.u32 %r1956, %r1969;
|
||
mov.u32 %r1957, %r1970;
|
||
mov.u32 %r1150, %r1151;
|
||
add.s32 %r1073, %r160, %r2;
|
||
bra.uni BB2_131;
|
||
|
||
BB2_54:
|
||
or.b32 %r2024, %r379, %r2011;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2023, %r2010;
|
||
mov.u32 %r2022, %r2009;
|
||
mov.u32 %r1372, %r2005;
|
||
mov.u32 %r1386, %r2004;
|
||
mov.u32 %r2016, %r2003;
|
||
mov.u32 %r2015, %r2002;
|
||
mov.u32 %r1429, %r2001;
|
||
mov.u32 %r1445, %r2000;
|
||
mov.u32 %r2025, %r383;
|
||
mov.u32 %r2019, %r387;
|
||
mov.u32 %r2020, %r391;
|
||
mov.u32 %r2021, %r395;
|
||
mov.u32 %r2013, %r1445;
|
||
mov.u32 %r2014, %r1429;
|
||
mov.u32 %r2017, %r1386;
|
||
mov.u32 %r2018, %r1372;
|
||
bra.uni BB2_56;
|
||
|
||
BB2_40:
|
||
setp.eq.s32 %p35, %r399, 4;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2025, %r2012;
|
||
mov.u32 %r2024, %r2011;
|
||
mov.u32 %r2023, %r2010;
|
||
mov.u32 %r2022, %r2009;
|
||
mov.u32 %r2021, %r2008;
|
||
mov.u32 %r2020, %r2007;
|
||
mov.u32 %r2019, %r2006;
|
||
mov.u32 %r1366, %r2005;
|
||
mov.u32 %r1381, %r2004;
|
||
mov.u32 %r2016, %r2003;
|
||
mov.u32 %r2015, %r2002;
|
||
mov.u32 %r1421, %r2001;
|
||
mov.u32 %r1438, %r2000;
|
||
mov.u32 %r2013, %r1438;
|
||
mov.u32 %r2014, %r1421;
|
||
mov.u32 %r2017, %r1381;
|
||
mov.u32 %r2018, %r1366;
|
||
@%p35 bra BB2_41;
|
||
bra.uni BB2_56;
|
||
|
||
BB2_41:
|
||
or.b32 %r2019, %r379, %r2006;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2025, %r2012;
|
||
mov.u32 %r2024, %r2011;
|
||
mov.u32 %r2023, %r2010;
|
||
mov.u32 %r1370, %r2005;
|
||
mov.u32 %r1384, %r2004;
|
||
mov.u32 %r2016, %r2003;
|
||
mov.u32 %r1427, %r2001;
|
||
mov.u32 %r1443, %r2000;
|
||
mov.u32 %r2020, %r383;
|
||
mov.u32 %r2021, %r387;
|
||
mov.u32 %r2022, %r391;
|
||
mov.u32 %r2015, %r395;
|
||
mov.u32 %r2013, %r1443;
|
||
mov.u32 %r2014, %r1427;
|
||
mov.u32 %r2017, %r1384;
|
||
mov.u32 %r2018, %r1370;
|
||
bra.uni BB2_56;
|
||
|
||
BB2_51:
|
||
or.b32 %r2022, %r379, %r2009;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2025, %r2012;
|
||
mov.u32 %r2024, %r2011;
|
||
mov.u32 %r2023, %r2010;
|
||
mov.u32 %r2021, %r2008;
|
||
mov.u32 %r2020, %r2007;
|
||
mov.u32 %r2019, %r2006;
|
||
mov.u32 %r1424, %r2001;
|
||
mov.u32 %r1440, %r2000;
|
||
mov.u32 %r2015, %r383;
|
||
mov.u32 %r2016, %r387;
|
||
mov.u32 %r2017, %r391;
|
||
mov.u32 %r2018, %r395;
|
||
mov.u32 %r2013, %r1440;
|
||
mov.u32 %r2014, %r1424;
|
||
bra.uni BB2_56;
|
||
|
||
BB2_48:
|
||
setp.ne.s32 %p29, %r399, 9;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2025, %r2012;
|
||
mov.u32 %r2024, %r2011;
|
||
mov.u32 %r2023, %r2010;
|
||
mov.u32 %r2022, %r2009;
|
||
mov.u32 %r2021, %r2008;
|
||
mov.u32 %r2020, %r2007;
|
||
mov.u32 %r2019, %r2006;
|
||
mov.u32 %r1364, %r2005;
|
||
mov.u32 %r1379, %r2004;
|
||
mov.u32 %r2016, %r2003;
|
||
mov.u32 %r2015, %r2002;
|
||
mov.u32 %r1419, %r2001;
|
||
mov.u32 %r1436, %r2000;
|
||
mov.u32 %r2013, %r1436;
|
||
mov.u32 %r2014, %r1419;
|
||
mov.u32 %r2017, %r1379;
|
||
mov.u32 %r2018, %r1364;
|
||
@%p29 bra BB2_56;
|
||
|
||
or.b32 %r2016, %r379, %r2003;
|
||
mov.u32 %r1156, %r1155;
|
||
mov.u32 %r2025, %r2012;
|
||
mov.u32 %r2024, %r2011;
|
||
mov.u32 %r2023, %r2010;
|
||
mov.u32 %r2022, %r2009;
|
||
mov.u32 %r2021, %r2008;
|
||
mov.u32 %r2020, %r2007;
|
||
mov.u32 %r2019, %r2006;
|
||
mov.u32 %r2015, %r2002;
|
||
mov.u32 %r2013, %r391;
|
||
mov.u32 %r2014, %r395;
|
||
mov.u32 %r2017, %r383;
|
||
mov.u32 %r2018, %r387;
|
||
|
||
BB2_56:
|
||
mov.u32 %r2000, %r2013;
|
||
mov.u32 %r2001, %r2014;
|
||
mov.u32 %r2002, %r2015;
|
||
mov.u32 %r2003, %r2016;
|
||
mov.u32 %r2004, %r2017;
|
||
mov.u32 %r2005, %r2018;
|
||
mov.u32 %r2006, %r2019;
|
||
mov.u32 %r2007, %r2020;
|
||
mov.u32 %r2008, %r2021;
|
||
mov.u32 %r2009, %r2022;
|
||
mov.u32 %r2010, %r2023;
|
||
mov.u32 %r2011, %r2024;
|
||
mov.u32 %r2012, %r2025;
|
||
mov.u32 %r1155, %r1156;
|
||
add.s32 %r1057, %r1057, %r2;
|
||
|
||
BB2_57:
|
||
mov.u32 %r98, %r2000;
|
||
mov.u32 %r97, %r2001;
|
||
mov.u32 %r96, %r2002;
|
||
mov.u32 %r95, %r2003;
|
||
mov.u32 %r94, %r2004;
|
||
mov.u32 %r93, %r2005;
|
||
mov.u32 %r92, %r2006;
|
||
mov.u32 %r91, %r2007;
|
||
mov.u32 %r90, %r2008;
|
||
mov.u32 %r89, %r2009;
|
||
mov.u32 %r87, %r2010;
|
||
mov.u32 %r86, %r2011;
|
||
mov.u32 %r85, %r2012;
|
||
mov.u32 %r88, %r1155;
|
||
mov.u32 %r1035, 128;
|
||
and.b32 %r427, %r1057, 3;
|
||
sub.s32 %r429, %r237, %r427;
|
||
shl.b32 %r430, %r429, 2;
|
||
shr.u32 %r432, %r240, %r430;
|
||
and.b32 %r425, %r432, 65535;
|
||
mov.u32 %r407, 0;
|
||
// inline asm
|
||
prmt.b32 %r406, %r407, %r27, %r425;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r410, %r27, %r28, %r425;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r414, %r28, %r29, %r425;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r418, %r29, %r30, %r425;
|
||
// inline asm
|
||
// inline asm
|
||
prmt.b32 %r422, %r30, %r1035, %r425;
|
||
// inline asm
|
||
shr.u32 %r426, %r1057, 2;
|
||
setp.gt.s32 %p38, %r426, 4;
|
||
@%p38 bra BB2_66;
|
||
|
||
setp.gt.s32 %p45, %r426, 1;
|
||
@%p45 bra BB2_62;
|
||
|
||
setp.eq.s32 %p49, %r426, 0;
|
||
@%p49 bra BB2_79;
|
||
bra.uni BB2_60;
|
||
|
||
BB2_79:
|
||
or.b32 %r1154, %r406, %r88;
|
||
mov.u32 %r1996, %r89;
|
||
mov.u32 %r1995, %r90;
|
||
mov.u32 %r1994, %r91;
|
||
mov.u32 %r1576, %r93;
|
||
mov.u32 %r1586, %r94;
|
||
mov.u32 %r1990, %r95;
|
||
mov.u32 %r1989, %r96;
|
||
mov.u32 %r1617, %r97;
|
||
mov.u32 %r1629, %r98;
|
||
mov.u32 %r1997, %r410;
|
||
mov.u32 %r1998, %r414;
|
||
mov.u32 %r1999, %r418;
|
||
mov.u32 %r1993, %r422;
|
||
mov.u32 %r1987, %r1629;
|
||
mov.u32 %r1988, %r1617;
|
||
mov.u32 %r1991, %r1586;
|
||
mov.u32 %r1992, %r1576;
|
||
bra.uni BB2_80;
|
||
|
||
BB2_66:
|
||
setp.gt.s32 %p39, %r426, 6;
|
||
@%p39 bra BB2_70;
|
||
|
||
setp.eq.s32 %p43, %r426, 5;
|
||
@%p43 bra BB2_76;
|
||
bra.uni BB2_68;
|
||
|
||
BB2_76:
|
||
or.b32 %r1994, %r406, %r91;
|
||
mov.u32 %r1154, %r88;
|
||
mov.u32 %r1999, %r85;
|
||
mov.u32 %r1998, %r86;
|
||
mov.u32 %r1997, %r87;
|
||
mov.u32 %r1993, %r92;
|
||
mov.u32 %r1571, %r93;
|
||
mov.u32 %r1581, %r94;
|
||
mov.u32 %r1612, %r97;
|
||
mov.u32 %r1624, %r98;
|
||
mov.u32 %r1995, %r410;
|
||
mov.u32 %r1996, %r414;
|
||
mov.u32 %r1989, %r418;
|
||
mov.u32 %r1990, %r422;
|
||
mov.u32 %r1987, %r1624;
|
||
mov.u32 %r1988, %r1612;
|
||
mov.u32 %r1991, %r1581;
|
||
mov.u32 %r1992, %r1571;
|
||
bra.uni BB2_80;
|
||
|
||
BB2_62:
|
||
setp.eq.s32 %p46, %r426, 2;
|
||
@%p46 bra BB2_78;
|
||
|
||
setp.eq.s32 %p47, %r426, 3;
|
||
@%p47 bra BB2_77;
|
||
bra.uni BB2_64;
|
||
|
||
BB2_77:
|
||
or.b32 %r1999, %r406, %r85;
|
||
mov.u32 %r1154, %r88;
|
||
mov.u32 %r1998, %r86;
|
||
mov.u32 %r1997, %r87;
|
||
mov.u32 %r1573, %r93;
|
||
mov.u32 %r1583, %r94;
|
||
mov.u32 %r1990, %r95;
|
||
mov.u32 %r1989, %r96;
|
||
mov.u32 %r1614, %r97;
|
||
mov.u32 %r1626, %r98;
|
||
mov.u32 %r1993, %r410;
|
||
mov.u32 %r1994, %r414;
|
||
mov.u32 %r1995, %r418;
|
||
mov.u32 %r1996, %r422;
|
||
mov.u32 %r1987, %r1626;
|
||
mov.u32 %r1988, %r1614;
|
||
mov.u32 %r1991, %r1583;
|
||
mov.u32 %r1992, %r1573;
|
||
bra.uni BB2_80;
|
||
|
||
BB2_70:
|
||
setp.eq.s32 %p40, %r426, 7;
|
||
@%p40 bra BB2_75;
|
||
|
||
setp.eq.s32 %p41, %r426, 8;
|
||
@%p41 bra BB2_74;
|
||
bra.uni BB2_72;
|
||
|
||
BB2_74:
|
||
or.b32 %r1989, %r406, %r96;
|
||
mov.u32 %r1154, %r88;
|
||
mov.u32 %r1999, %r85;
|
||
mov.u32 %r1998, %r86;
|
||
mov.u32 %r1997, %r87;
|
||
mov.u32 %r1996, %r89;
|
||
mov.u32 %r1995, %r90;
|
||
mov.u32 %r1994, %r91;
|
||
mov.u32 %r1993, %r92;
|
||
mov.u32 %r1609, %r97;
|
||
mov.u32 %r1990, %r410;
|
||
mov.u32 %r1991, %r414;
|
||
mov.u32 %r1992, %r418;
|
||
mov.u32 %r1987, %r422;
|
||
mov.u32 %r1988, %r1609;
|
||
bra.uni BB2_80;
|
||
|
||
BB2_60:
|
||
setp.eq.s32 %p50, %r426, 1;
|
||
mov.u32 %r1124, %r88;
|
||
mov.u32 %r1154, %r1124;
|
||
mov.u32 %r1999, %r85;
|
||
mov.u32 %r1998, %r86;
|
||
mov.u32 %r1521, %r87;
|
||
mov.u32 %r1533, %r89;
|
||
mov.u32 %r1542, %r90;
|
||
mov.u32 %r1994, %r91;
|
||
mov.u32 %r1993, %r92;
|
||
mov.u32 %r1569, %r93;
|
||
mov.u32 %r1580, %r94;
|
||
mov.u32 %r1590, %r95;
|
||
mov.u32 %r1599, %r96;
|
||
mov.u32 %r1608, %r97;
|
||
mov.u32 %r1621, %r98;
|
||
mov.u32 %r1987, %r1621;
|
||
mov.u32 %r1988, %r1608;
|
||
mov.u32 %r1989, %r1599;
|
||
mov.u32 %r1990, %r1590;
|
||
mov.u32 %r1991, %r1580;
|
||
mov.u32 %r1992, %r1569;
|
||
mov.u32 %r1995, %r1542;
|
||
mov.u32 %r1996, %r1533;
|
||
mov.u32 %r1997, %r1521;
|
||
@%p50 bra BB2_61;
|
||
bra.uni BB2_80;
|
||
|
||
BB2_61:
|
||
or.b32 %r1997, %r406, %r87;
|
||
mov.u32 %r1154, %r88;
|
||
mov.u32 %r1996, %r89;
|
||
mov.u32 %r1995, %r90;
|
||
mov.u32 %r1575, %r93;
|
||
mov.u32 %r1585, %r94;
|
||
mov.u32 %r1990, %r95;
|
||
mov.u32 %r1989, %r96;
|
||
mov.u32 %r1616, %r97;
|
||
mov.u32 %r1628, %r98;
|
||
mov.u32 %r1998, %r410;
|
||
mov.u32 %r1999, %r414;
|
||
mov.u32 %r1993, %r418;
|
||
mov.u32 %r1994, %r422;
|
||
mov.u32 %r1987, %r1628;
|
||
mov.u32 %r1988, %r1616;
|
||
mov.u32 %r1991, %r1585;
|
||
mov.u32 %r1992, %r1575;
|
||
bra.uni BB2_80;
|
||
|
||
BB2_68:
|
||
setp.eq.s32 %p44, %r426, 6;
|
||
mov.u32 %r1122, %r88;
|
||
mov.u32 %r1154, %r1122;
|
||
mov.u32 %r1498, %r85;
|
||
mov.u32 %r1508, %r86;
|
||
mov.u32 %r1519, %r87;
|
||
mov.u32 %r1996, %r89;
|
||
mov.u32 %r1540, %r90;
|
||
mov.u32 %r1549, %r91;
|
||
mov.u32 %r1558, %r92;
|
||
mov.u32 %r1567, %r93;
|
||
mov.u32 %r1578, %r94;
|
||
mov.u32 %r1990, %r95;
|
||
mov.u32 %r1989, %r96;
|
||
mov.u32 %r1606, %r97;
|
||
mov.u32 %r1619, %r98;
|
||
mov.u32 %r1987, %r1619;
|
||
mov.u32 %r1988, %r1606;
|
||
mov.u32 %r1991, %r1578;
|
||
mov.u32 %r1992, %r1567;
|
||
mov.u32 %r1993, %r1558;
|
||
mov.u32 %r1994, %r1549;
|
||
mov.u32 %r1995, %r1540;
|
||
mov.u32 %r1997, %r1519;
|
||
mov.u32 %r1998, %r1508;
|
||
mov.u32 %r1999, %r1498;
|
||
@%p44 bra BB2_69;
|
||
bra.uni BB2_80;
|
||
|
||
BB2_69:
|
||
or.b32 %r1995, %r406, %r90;
|
||
mov.u32 %r1154, %r88;
|
||
mov.u32 %r1999, %r85;
|
||
mov.u32 %r1998, %r86;
|
||
mov.u32 %r1997, %r87;
|
||
mov.u32 %r1994, %r91;
|
||
mov.u32 %r1993, %r92;
|
||
mov.u32 %r1570, %r93;
|
||
mov.u32 %r1611, %r97;
|
||
mov.u32 %r1623, %r98;
|
||
mov.u32 %r1996, %r410;
|
||
mov.u32 %r1989, %r414;
|
||
mov.u32 %r1990, %r418;
|
||
mov.u32 %r1991, %r422;
|
||
mov.u32 %r1987, %r1623;
|
||
mov.u32 %r1988, %r1611;
|
||
mov.u32 %r1992, %r1570;
|
||
bra.uni BB2_80;
|
||
|
||
BB2_78:
|
||
or.b32 %r1998, %r406, %r86;
|
||
mov.u32 %r1154, %r88;
|
||
mov.u32 %r1997, %r87;
|
||
mov.u32 %r1996, %r89;
|
||
mov.u32 %r1574, %r93;
|
||
mov.u32 %r1584, %r94;
|
||
mov.u32 %r1990, %r95;
|
||
mov.u32 %r1989, %r96;
|
||
mov.u32 %r1615, %r97;
|
||
mov.u32 %r1627, %r98;
|
||
mov.u32 %r1999, %r410;
|
||
mov.u32 %r1993, %r414;
|
||
mov.u32 %r1994, %r418;
|
||
mov.u32 %r1995, %r422;
|
||
mov.u32 %r1987, %r1627;
|
||
mov.u32 %r1988, %r1615;
|
||
mov.u32 %r1991, %r1584;
|
||
mov.u32 %r1992, %r1574;
|
||
bra.uni BB2_80;
|
||
|
||
BB2_64:
|
||
setp.eq.s32 %p48, %r426, 4;
|
||
mov.u32 %r1123, %r88;
|
||
mov.u32 %r1154, %r1123;
|
||
mov.u32 %r1499, %r85;
|
||
mov.u32 %r1509, %r86;
|
||
mov.u32 %r1520, %r87;
|
||
mov.u32 %r1996, %r89;
|
||
mov.u32 %r1995, %r90;
|
||
mov.u32 %r1994, %r91;
|
||
mov.u32 %r1559, %r92;
|
||
mov.u32 %r1568, %r93;
|
||
mov.u32 %r1579, %r94;
|
||
mov.u32 %r1589, %r95;
|
||
mov.u32 %r1989, %r96;
|
||
mov.u32 %r1607, %r97;
|
||
mov.u32 %r1620, %r98;
|
||
mov.u32 %r1987, %r1620;
|
||
mov.u32 %r1988, %r1607;
|
||
mov.u32 %r1990, %r1589;
|
||
mov.u32 %r1991, %r1579;
|
||
mov.u32 %r1992, %r1568;
|
||
mov.u32 %r1993, %r1559;
|
||
mov.u32 %r1997, %r1520;
|
||
mov.u32 %r1998, %r1509;
|
||
mov.u32 %r1999, %r1499;
|
||
@%p48 bra BB2_65;
|
||
bra.uni BB2_80;
|
||
|
||
BB2_65:
|
||
or.b32 %r1993, %r406, %r92;
|
||
mov.u32 %r1154, %r88;
|
||
mov.u32 %r1999, %r85;
|
||
mov.u32 %r1998, %r86;
|
||
mov.u32 %r1997, %r87;
|
||
mov.u32 %r1572, %r93;
|
||
mov.u32 %r1582, %r94;
|
||
mov.u32 %r1990, %r95;
|
||
mov.u32 %r1613, %r97;
|
||
mov.u32 %r1625, %r98;
|
||
mov.u32 %r1994, %r410;
|
||
mov.u32 %r1995, %r414;
|
||
mov.u32 %r1996, %r418;
|
||
mov.u32 %r1989, %r422;
|
||
mov.u32 %r1987, %r1625;
|
||
mov.u32 %r1988, %r1613;
|
||
mov.u32 %r1991, %r1582;
|
||
mov.u32 %r1992, %r1572;
|
||
bra.uni BB2_80;
|
||
|
||
BB2_75:
|
||
or.b32 %r1996, %r406, %r89;
|
||
mov.u32 %r1154, %r88;
|
||
mov.u32 %r1999, %r85;
|
||
mov.u32 %r1998, %r86;
|
||
mov.u32 %r1997, %r87;
|
||
mov.u32 %r1995, %r90;
|
||
mov.u32 %r1994, %r91;
|
||
mov.u32 %r1993, %r92;
|
||
mov.u32 %r1610, %r97;
|
||
mov.u32 %r1622, %r98;
|
||
mov.u32 %r1989, %r410;
|
||
mov.u32 %r1990, %r414;
|
||
mov.u32 %r1991, %r418;
|
||
mov.u32 %r1992, %r422;
|
||
mov.u32 %r1987, %r1622;
|
||
mov.u32 %r1988, %r1610;
|
||
bra.uni BB2_80;
|
||
|
||
BB2_72:
|
||
setp.ne.s32 %p42, %r426, 9;
|
||
mov.u32 %r1154, %r88;
|
||
mov.u32 %r1987, %r98;
|
||
mov.u32 %r1988, %r97;
|
||
mov.u32 %r1989, %r96;
|
||
mov.u32 %r1990, %r95;
|
||
mov.u32 %r1991, %r94;
|
||
mov.u32 %r1992, %r93;
|
||
mov.u32 %r1993, %r92;
|
||
mov.u32 %r1994, %r91;
|
||
mov.u32 %r1995, %r90;
|
||
mov.u32 %r1996, %r89;
|
||
mov.u32 %r1997, %r87;
|
||
mov.u32 %r1998, %r86;
|
||
mov.u32 %r1999, %r85;
|
||
@%p42 bra BB2_80;
|
||
|
||
or.b32 %r1990, %r406, %r95;
|
||
mov.u32 %r1154, %r88;
|
||
mov.u32 %r1999, %r85;
|
||
mov.u32 %r1998, %r86;
|
||
mov.u32 %r1997, %r87;
|
||
mov.u32 %r1996, %r89;
|
||
mov.u32 %r1995, %r90;
|
||
mov.u32 %r1994, %r91;
|
||
mov.u32 %r1993, %r92;
|
||
mov.u32 %r1989, %r96;
|
||
mov.u32 %r1987, %r418;
|
||
mov.u32 %r1988, %r422;
|
||
mov.u32 %r1991, %r410;
|
||
mov.u32 %r1992, %r414;
|
||
|
||
BB2_80:
|
||
mov.u32 %r1945, %r1987;
|
||
mov.u32 %r1946, %r1988;
|
||
mov.u32 %r1947, %r1989;
|
||
mov.u32 %r1948, %r1990;
|
||
mov.u32 %r1949, %r1991;
|
||
mov.u32 %r1950, %r1992;
|
||
mov.u32 %r1951, %r1993;
|
||
mov.u32 %r1952, %r1994;
|
||
mov.u32 %r1953, %r1995;
|
||
mov.u32 %r1954, %r1996;
|
||
mov.u32 %r1955, %r1997;
|
||
mov.u32 %r1956, %r1998;
|
||
mov.u32 %r1957, %r1999;
|
||
mov.u32 %r1150, %r1154;
|
||
add.s32 %r1073, %r1057, 16;
|
||
|
||
BB2_131:
|
||
ld.param.u32 %r1036, [m00500_loop_param_29];
|
||
add.s32 %r552, %r1150, -680876937;
|
||
shf.l.wrap.b32 %r553, %r552, %r552, 7;
|
||
add.s32 %r554, %r553, -271733879;
|
||
and.b32 %r555, %r554, 2004318071;
|
||
xor.b32 %r556, %r555, -1732584194;
|
||
add.s32 %r557, %r1955, %r556;
|
||
add.s32 %r558, %r557, -117830708;
|
||
shf.l.wrap.b32 %r559, %r558, %r558, 12;
|
||
add.s32 %r560, %r559, %r554;
|
||
xor.b32 %r561, %r554, -271733879;
|
||
and.b32 %r562, %r560, %r561;
|
||
xor.b32 %r563, %r562, -271733879;
|
||
add.s32 %r564, %r1956, %r563;
|
||
add.s32 %r565, %r564, -1126478375;
|
||
shf.l.wrap.b32 %r566, %r565, %r565, 17;
|
||
add.s32 %r567, %r566, %r560;
|
||
xor.b32 %r568, %r560, %r554;
|
||
and.b32 %r569, %r567, %r568;
|
||
xor.b32 %r570, %r569, %r554;
|
||
add.s32 %r571, %r1957, %r570;
|
||
add.s32 %r572, %r571, -1316259209;
|
||
shf.l.wrap.b32 %r573, %r572, %r572, 22;
|
||
add.s32 %r574, %r573, %r567;
|
||
xor.b32 %r575, %r567, %r560;
|
||
and.b32 %r576, %r574, %r575;
|
||
xor.b32 %r577, %r576, %r560;
|
||
add.s32 %r578, %r1951, %r553;
|
||
add.s32 %r579, %r578, %r577;
|
||
add.s32 %r580, %r579, -448152776;
|
||
shf.l.wrap.b32 %r581, %r580, %r580, 7;
|
||
add.s32 %r582, %r581, %r574;
|
||
xor.b32 %r583, %r574, %r567;
|
||
and.b32 %r584, %r582, %r583;
|
||
xor.b32 %r585, %r584, %r567;
|
||
add.s32 %r586, %r1952, %r560;
|
||
add.s32 %r587, %r586, %r585;
|
||
add.s32 %r588, %r587, 1200080426;
|
||
shf.l.wrap.b32 %r589, %r588, %r588, 12;
|
||
add.s32 %r590, %r589, %r582;
|
||
xor.b32 %r591, %r582, %r574;
|
||
and.b32 %r592, %r590, %r591;
|
||
xor.b32 %r593, %r592, %r574;
|
||
add.s32 %r594, %r1953, %r567;
|
||
add.s32 %r595, %r594, %r593;
|
||
add.s32 %r596, %r595, -1473231341;
|
||
shf.l.wrap.b32 %r597, %r596, %r596, 17;
|
||
add.s32 %r598, %r597, %r590;
|
||
xor.b32 %r599, %r590, %r582;
|
||
and.b32 %r600, %r598, %r599;
|
||
xor.b32 %r601, %r600, %r582;
|
||
add.s32 %r602, %r1954, %r574;
|
||
add.s32 %r603, %r602, %r601;
|
||
add.s32 %r604, %r603, -45705983;
|
||
shf.l.wrap.b32 %r605, %r604, %r604, 22;
|
||
add.s32 %r606, %r605, %r598;
|
||
xor.b32 %r607, %r598, %r590;
|
||
and.b32 %r608, %r606, %r607;
|
||
xor.b32 %r609, %r608, %r590;
|
||
add.s32 %r610, %r1947, %r582;
|
||
add.s32 %r611, %r610, %r609;
|
||
add.s32 %r612, %r611, 1770035416;
|
||
shf.l.wrap.b32 %r613, %r612, %r612, 7;
|
||
add.s32 %r614, %r613, %r606;
|
||
xor.b32 %r615, %r606, %r598;
|
||
and.b32 %r616, %r614, %r615;
|
||
xor.b32 %r617, %r616, %r598;
|
||
add.s32 %r618, %r1948, %r590;
|
||
add.s32 %r619, %r618, %r617;
|
||
add.s32 %r620, %r619, -1958414417;
|
||
shf.l.wrap.b32 %r621, %r620, %r620, 12;
|
||
add.s32 %r622, %r621, %r614;
|
||
xor.b32 %r623, %r614, %r606;
|
||
and.b32 %r624, %r622, %r623;
|
||
xor.b32 %r625, %r624, %r606;
|
||
add.s32 %r626, %r1949, %r598;
|
||
add.s32 %r627, %r626, %r625;
|
||
add.s32 %r628, %r627, -42063;
|
||
shf.l.wrap.b32 %r629, %r628, %r628, 17;
|
||
add.s32 %r630, %r629, %r622;
|
||
xor.b32 %r631, %r622, %r614;
|
||
and.b32 %r632, %r630, %r631;
|
||
xor.b32 %r633, %r632, %r614;
|
||
add.s32 %r634, %r1950, %r606;
|
||
add.s32 %r635, %r634, %r633;
|
||
add.s32 %r636, %r635, -1990404162;
|
||
shf.l.wrap.b32 %r637, %r636, %r636, 22;
|
||
add.s32 %r638, %r637, %r630;
|
||
xor.b32 %r639, %r630, %r622;
|
||
and.b32 %r640, %r638, %r639;
|
||
xor.b32 %r641, %r640, %r622;
|
||
add.s32 %r642, %r1945, %r614;
|
||
add.s32 %r643, %r642, %r641;
|
||
add.s32 %r644, %r643, 1804603682;
|
||
shf.l.wrap.b32 %r645, %r644, %r644, 7;
|
||
add.s32 %r646, %r645, %r638;
|
||
xor.b32 %r647, %r638, %r630;
|
||
and.b32 %r648, %r646, %r647;
|
||
xor.b32 %r649, %r648, %r630;
|
||
add.s32 %r650, %r1946, %r622;
|
||
add.s32 %r651, %r650, %r649;
|
||
add.s32 %r652, %r651, -40341101;
|
||
shf.l.wrap.b32 %r653, %r652, %r652, 12;
|
||
add.s32 %r654, %r653, %r646;
|
||
xor.b32 %r655, %r646, %r638;
|
||
and.b32 %r656, %r654, %r655;
|
||
xor.b32 %r657, %r656, %r638;
|
||
shl.b32 %r658, %r1073, 3;
|
||
add.s32 %r659, %r658, %r630;
|
||
add.s32 %r660, %r659, %r657;
|
||
add.s32 %r661, %r660, -1502002290;
|
||
shf.l.wrap.b32 %r662, %r661, %r661, 17;
|
||
add.s32 %r663, %r662, %r654;
|
||
xor.b32 %r664, %r654, %r646;
|
||
and.b32 %r665, %r663, %r664;
|
||
xor.b32 %r666, %r665, %r646;
|
||
add.s32 %r667, %r638, %r666;
|
||
add.s32 %r668, %r667, 1236535329;
|
||
shf.l.wrap.b32 %r669, %r668, %r668, 22;
|
||
add.s32 %r670, %r669, %r663;
|
||
xor.b32 %r671, %r670, %r663;
|
||
and.b32 %r672, %r671, %r654;
|
||
xor.b32 %r673, %r672, %r663;
|
||
add.s32 %r674, %r1955, %r646;
|
||
add.s32 %r675, %r674, %r673;
|
||
add.s32 %r676, %r675, -165796510;
|
||
shf.l.wrap.b32 %r677, %r676, %r676, 5;
|
||
add.s32 %r678, %r677, %r670;
|
||
xor.b32 %r679, %r678, %r670;
|
||
and.b32 %r680, %r679, %r663;
|
||
xor.b32 %r681, %r680, %r670;
|
||
add.s32 %r682, %r1953, %r654;
|
||
add.s32 %r683, %r682, %r681;
|
||
add.s32 %r684, %r683, -1069501632;
|
||
shf.l.wrap.b32 %r685, %r684, %r684, 9;
|
||
add.s32 %r686, %r685, %r678;
|
||
xor.b32 %r687, %r686, %r678;
|
||
and.b32 %r688, %r687, %r670;
|
||
xor.b32 %r689, %r688, %r678;
|
||
add.s32 %r690, %r1950, %r663;
|
||
add.s32 %r691, %r690, %r689;
|
||
add.s32 %r692, %r691, 643717713;
|
||
shf.l.wrap.b32 %r693, %r692, %r692, 14;
|
||
add.s32 %r694, %r693, %r686;
|
||
xor.b32 %r695, %r694, %r686;
|
||
and.b32 %r696, %r695, %r678;
|
||
xor.b32 %r697, %r696, %r686;
|
||
add.s32 %r698, %r1150, %r670;
|
||
add.s32 %r699, %r698, %r697;
|
||
add.s32 %r700, %r699, -373897302;
|
||
shf.l.wrap.b32 %r701, %r700, %r700, 20;
|
||
add.s32 %r702, %r701, %r694;
|
||
xor.b32 %r703, %r702, %r694;
|
||
and.b32 %r704, %r703, %r686;
|
||
xor.b32 %r705, %r704, %r694;
|
||
add.s32 %r706, %r1952, %r678;
|
||
add.s32 %r707, %r706, %r705;
|
||
add.s32 %r708, %r707, -701558691;
|
||
shf.l.wrap.b32 %r709, %r708, %r708, 5;
|
||
add.s32 %r710, %r709, %r702;
|
||
xor.b32 %r711, %r710, %r702;
|
||
and.b32 %r712, %r711, %r694;
|
||
xor.b32 %r713, %r712, %r702;
|
||
add.s32 %r714, %r1949, %r686;
|
||
add.s32 %r715, %r714, %r713;
|
||
add.s32 %r716, %r715, 38016083;
|
||
shf.l.wrap.b32 %r717, %r716, %r716, 9;
|
||
add.s32 %r718, %r717, %r710;
|
||
xor.b32 %r719, %r718, %r710;
|
||
and.b32 %r720, %r719, %r702;
|
||
xor.b32 %r721, %r720, %r710;
|
||
add.s32 %r722, %r694, %r721;
|
||
add.s32 %r723, %r722, -660478335;
|
||
shf.l.wrap.b32 %r724, %r723, %r723, 14;
|
||
add.s32 %r725, %r724, %r718;
|
||
xor.b32 %r726, %r725, %r718;
|
||
and.b32 %r727, %r726, %r710;
|
||
xor.b32 %r728, %r727, %r718;
|
||
add.s32 %r729, %r1951, %r702;
|
||
add.s32 %r730, %r729, %r728;
|
||
add.s32 %r731, %r730, -405537848;
|
||
shf.l.wrap.b32 %r732, %r731, %r731, 20;
|
||
add.s32 %r733, %r732, %r725;
|
||
xor.b32 %r734, %r733, %r725;
|
||
and.b32 %r735, %r734, %r718;
|
||
xor.b32 %r736, %r735, %r725;
|
||
add.s32 %r737, %r1948, %r710;
|
||
add.s32 %r738, %r737, %r736;
|
||
add.s32 %r739, %r738, 568446438;
|
||
shf.l.wrap.b32 %r740, %r739, %r739, 5;
|
||
add.s32 %r741, %r740, %r733;
|
||
xor.b32 %r742, %r741, %r733;
|
||
and.b32 %r743, %r742, %r725;
|
||
xor.b32 %r744, %r743, %r733;
|
||
add.s32 %r745, %r658, %r718;
|
||
add.s32 %r746, %r745, %r744;
|
||
add.s32 %r747, %r746, -1019803690;
|
||
shf.l.wrap.b32 %r748, %r747, %r747, 9;
|
||
add.s32 %r749, %r748, %r741;
|
||
xor.b32 %r750, %r749, %r741;
|
||
and.b32 %r751, %r750, %r733;
|
||
xor.b32 %r752, %r751, %r741;
|
||
add.s32 %r753, %r1957, %r725;
|
||
add.s32 %r754, %r753, %r752;
|
||
add.s32 %r755, %r754, -187363961;
|
||
shf.l.wrap.b32 %r756, %r755, %r755, 14;
|
||
add.s32 %r757, %r756, %r749;
|
||
xor.b32 %r758, %r757, %r749;
|
||
and.b32 %r759, %r758, %r741;
|
||
xor.b32 %r760, %r759, %r749;
|
||
add.s32 %r761, %r1947, %r733;
|
||
add.s32 %r762, %r761, %r760;
|
||
add.s32 %r763, %r762, 1163531501;
|
||
shf.l.wrap.b32 %r764, %r763, %r763, 20;
|
||
add.s32 %r765, %r764, %r757;
|
||
xor.b32 %r766, %r765, %r757;
|
||
and.b32 %r767, %r766, %r749;
|
||
xor.b32 %r768, %r767, %r757;
|
||
add.s32 %r769, %r1946, %r741;
|
||
add.s32 %r770, %r769, %r768;
|
||
add.s32 %r771, %r770, -1444681467;
|
||
shf.l.wrap.b32 %r772, %r771, %r771, 5;
|
||
add.s32 %r773, %r772, %r765;
|
||
xor.b32 %r774, %r773, %r765;
|
||
and.b32 %r775, %r774, %r757;
|
||
xor.b32 %r776, %r775, %r765;
|
||
add.s32 %r777, %r1956, %r749;
|
||
add.s32 %r778, %r777, %r776;
|
||
add.s32 %r779, %r778, -51403784;
|
||
shf.l.wrap.b32 %r780, %r779, %r779, 9;
|
||
add.s32 %r781, %r780, %r773;
|
||
xor.b32 %r782, %r781, %r773;
|
||
and.b32 %r783, %r782, %r765;
|
||
xor.b32 %r784, %r783, %r773;
|
||
add.s32 %r785, %r1954, %r757;
|
||
add.s32 %r786, %r785, %r784;
|
||
add.s32 %r787, %r786, 1735328473;
|
||
shf.l.wrap.b32 %r788, %r787, %r787, 14;
|
||
add.s32 %r789, %r788, %r781;
|
||
xor.b32 %r790, %r789, %r781;
|
||
and.b32 %r791, %r790, %r773;
|
||
xor.b32 %r792, %r791, %r781;
|
||
add.s32 %r793, %r1945, %r765;
|
||
add.s32 %r794, %r793, %r792;
|
||
add.s32 %r795, %r794, -1926607734;
|
||
shf.l.wrap.b32 %r796, %r795, %r795, 20;
|
||
add.s32 %r797, %r796, %r789;
|
||
xor.b32 %r798, %r790, %r797;
|
||
add.s32 %r799, %r1952, %r773;
|
||
add.s32 %r800, %r799, %r798;
|
||
add.s32 %r801, %r800, -378558;
|
||
shf.l.wrap.b32 %r802, %r801, %r801, 4;
|
||
add.s32 %r803, %r802, %r797;
|
||
xor.b32 %r804, %r797, %r789;
|
||
xor.b32 %r805, %r804, %r803;
|
||
add.s32 %r806, %r1947, %r781;
|
||
add.s32 %r807, %r806, %r805;
|
||
add.s32 %r808, %r807, -2022574463;
|
||
shf.l.wrap.b32 %r809, %r808, %r808, 11;
|
||
add.s32 %r810, %r809, %r803;
|
||
xor.b32 %r811, %r803, %r797;
|
||
xor.b32 %r812, %r811, %r810;
|
||
add.s32 %r813, %r1950, %r789;
|
||
add.s32 %r814, %r813, %r812;
|
||
add.s32 %r815, %r814, 1839030562;
|
||
shf.l.wrap.b32 %r816, %r815, %r815, 16;
|
||
add.s32 %r817, %r816, %r810;
|
||
xor.b32 %r818, %r810, %r803;
|
||
xor.b32 %r819, %r818, %r817;
|
||
add.s32 %r820, %r658, %r797;
|
||
add.s32 %r821, %r820, %r819;
|
||
add.s32 %r822, %r821, -35309556;
|
||
shf.l.wrap.b32 %r823, %r822, %r822, 23;
|
||
add.s32 %r824, %r823, %r817;
|
||
xor.b32 %r825, %r817, %r810;
|
||
xor.b32 %r826, %r825, %r824;
|
||
add.s32 %r827, %r1955, %r803;
|
||
add.s32 %r828, %r827, %r826;
|
||
add.s32 %r829, %r828, -1530992060;
|
||
shf.l.wrap.b32 %r830, %r829, %r829, 4;
|
||
add.s32 %r831, %r830, %r824;
|
||
xor.b32 %r832, %r824, %r817;
|
||
xor.b32 %r833, %r832, %r831;
|
||
add.s32 %r834, %r1951, %r810;
|
||
add.s32 %r835, %r834, %r833;
|
||
add.s32 %r836, %r835, 1272893353;
|
||
shf.l.wrap.b32 %r837, %r836, %r836, 11;
|
||
add.s32 %r838, %r837, %r831;
|
||
xor.b32 %r839, %r831, %r824;
|
||
xor.b32 %r840, %r839, %r838;
|
||
add.s32 %r841, %r1954, %r817;
|
||
add.s32 %r842, %r841, %r840;
|
||
add.s32 %r843, %r842, -155497632;
|
||
shf.l.wrap.b32 %r844, %r843, %r843, 16;
|
||
add.s32 %r845, %r844, %r838;
|
||
xor.b32 %r846, %r838, %r831;
|
||
xor.b32 %r847, %r846, %r845;
|
||
add.s32 %r848, %r1949, %r824;
|
||
add.s32 %r849, %r848, %r847;
|
||
add.s32 %r850, %r849, -1094730640;
|
||
shf.l.wrap.b32 %r851, %r850, %r850, 23;
|
||
add.s32 %r852, %r851, %r845;
|
||
xor.b32 %r853, %r845, %r838;
|
||
xor.b32 %r854, %r853, %r852;
|
||
add.s32 %r855, %r1946, %r831;
|
||
add.s32 %r856, %r855, %r854;
|
||
add.s32 %r857, %r856, 681279174;
|
||
shf.l.wrap.b32 %r858, %r857, %r857, 4;
|
||
add.s32 %r859, %r858, %r852;
|
||
xor.b32 %r860, %r852, %r845;
|
||
xor.b32 %r861, %r860, %r859;
|
||
add.s32 %r862, %r1150, %r838;
|
||
add.s32 %r863, %r862, %r861;
|
||
add.s32 %r864, %r863, -358537222;
|
||
shf.l.wrap.b32 %r865, %r864, %r864, 11;
|
||
add.s32 %r866, %r865, %r859;
|
||
xor.b32 %r867, %r859, %r852;
|
||
xor.b32 %r868, %r867, %r866;
|
||
add.s32 %r869, %r1957, %r845;
|
||
add.s32 %r870, %r869, %r868;
|
||
add.s32 %r871, %r870, -722521979;
|
||
shf.l.wrap.b32 %r872, %r871, %r871, 16;
|
||
add.s32 %r873, %r872, %r866;
|
||
xor.b32 %r874, %r866, %r859;
|
||
xor.b32 %r875, %r874, %r873;
|
||
add.s32 %r876, %r1953, %r852;
|
||
add.s32 %r877, %r876, %r875;
|
||
add.s32 %r878, %r877, 76029189;
|
||
shf.l.wrap.b32 %r879, %r878, %r878, 23;
|
||
add.s32 %r880, %r879, %r873;
|
||
xor.b32 %r881, %r873, %r866;
|
||
xor.b32 %r882, %r881, %r880;
|
||
add.s32 %r883, %r1948, %r859;
|
||
add.s32 %r884, %r883, %r882;
|
||
add.s32 %r885, %r884, -640364487;
|
||
shf.l.wrap.b32 %r886, %r885, %r885, 4;
|
||
add.s32 %r887, %r886, %r880;
|
||
xor.b32 %r888, %r880, %r873;
|
||
xor.b32 %r889, %r888, %r887;
|
||
add.s32 %r890, %r1945, %r866;
|
||
add.s32 %r891, %r890, %r889;
|
||
add.s32 %r892, %r891, -421815835;
|
||
shf.l.wrap.b32 %r893, %r892, %r892, 11;
|
||
add.s32 %r894, %r893, %r887;
|
||
xor.b32 %r895, %r887, %r880;
|
||
xor.b32 %r896, %r895, %r894;
|
||
add.s32 %r897, %r873, %r896;
|
||
add.s32 %r898, %r897, 530742520;
|
||
shf.l.wrap.b32 %r899, %r898, %r898, 16;
|
||
add.s32 %r900, %r899, %r894;
|
||
xor.b32 %r901, %r894, %r887;
|
||
xor.b32 %r902, %r901, %r900;
|
||
add.s32 %r903, %r1956, %r880;
|
||
add.s32 %r904, %r903, %r902;
|
||
add.s32 %r905, %r904, -995338651;
|
||
shf.l.wrap.b32 %r906, %r905, %r905, 23;
|
||
add.s32 %r907, %r906, %r900;
|
||
not.b32 %r908, %r894;
|
||
or.b32 %r909, %r907, %r908;
|
||
xor.b32 %r910, %r909, %r900;
|
||
add.s32 %r911, %r1150, %r887;
|
||
add.s32 %r912, %r911, %r910;
|
||
add.s32 %r913, %r912, -198630844;
|
||
shf.l.wrap.b32 %r914, %r913, %r913, 6;
|
||
add.s32 %r915, %r914, %r907;
|
||
not.b32 %r916, %r900;
|
||
or.b32 %r917, %r915, %r916;
|
||
xor.b32 %r918, %r917, %r907;
|
||
add.s32 %r919, %r1954, %r894;
|
||
add.s32 %r920, %r919, %r918;
|
||
add.s32 %r921, %r920, 1126891415;
|
||
shf.l.wrap.b32 %r922, %r921, %r921, 10;
|
||
add.s32 %r923, %r922, %r915;
|
||
not.b32 %r924, %r907;
|
||
or.b32 %r925, %r923, %r924;
|
||
xor.b32 %r926, %r925, %r915;
|
||
add.s32 %r927, %r658, %r900;
|
||
add.s32 %r928, %r927, %r926;
|
||
add.s32 %r929, %r928, -1416354905;
|
||
shf.l.wrap.b32 %r930, %r929, %r929, 15;
|
||
add.s32 %r931, %r930, %r923;
|
||
not.b32 %r932, %r915;
|
||
or.b32 %r933, %r931, %r932;
|
||
xor.b32 %r934, %r933, %r923;
|
||
add.s32 %r935, %r1952, %r907;
|
||
add.s32 %r936, %r935, %r934;
|
||
add.s32 %r937, %r936, -57434055;
|
||
shf.l.wrap.b32 %r938, %r937, %r937, 21;
|
||
add.s32 %r939, %r938, %r931;
|
||
not.b32 %r940, %r923;
|
||
or.b32 %r941, %r939, %r940;
|
||
xor.b32 %r942, %r941, %r931;
|
||
add.s32 %r943, %r1945, %r915;
|
||
add.s32 %r944, %r943, %r942;
|
||
add.s32 %r945, %r944, 1700485571;
|
||
shf.l.wrap.b32 %r946, %r945, %r945, 6;
|
||
add.s32 %r947, %r946, %r939;
|
||
not.b32 %r948, %r931;
|
||
or.b32 %r949, %r947, %r948;
|
||
xor.b32 %r950, %r949, %r939;
|
||
add.s32 %r951, %r1957, %r923;
|
||
add.s32 %r952, %r951, %r950;
|
||
add.s32 %r953, %r952, -1894986606;
|
||
shf.l.wrap.b32 %r954, %r953, %r953, 10;
|
||
add.s32 %r955, %r954, %r947;
|
||
not.b32 %r956, %r939;
|
||
or.b32 %r957, %r955, %r956;
|
||
xor.b32 %r958, %r957, %r947;
|
||
add.s32 %r959, %r1949, %r931;
|
||
add.s32 %r960, %r959, %r958;
|
||
add.s32 %r961, %r960, -1051523;
|
||
shf.l.wrap.b32 %r962, %r961, %r961, 15;
|
||
add.s32 %r963, %r962, %r955;
|
||
not.b32 %r964, %r947;
|
||
or.b32 %r965, %r963, %r964;
|
||
xor.b32 %r966, %r965, %r955;
|
||
add.s32 %r967, %r1955, %r939;
|
||
add.s32 %r968, %r967, %r966;
|
||
add.s32 %r969, %r968, -2054922799;
|
||
shf.l.wrap.b32 %r970, %r969, %r969, 21;
|
||
add.s32 %r971, %r970, %r963;
|
||
not.b32 %r972, %r955;
|
||
or.b32 %r973, %r971, %r972;
|
||
xor.b32 %r974, %r973, %r963;
|
||
add.s32 %r975, %r1947, %r947;
|
||
add.s32 %r976, %r975, %r974;
|
||
add.s32 %r977, %r976, 1873313359;
|
||
shf.l.wrap.b32 %r978, %r977, %r977, 6;
|
||
add.s32 %r979, %r978, %r971;
|
||
not.b32 %r980, %r963;
|
||
or.b32 %r981, %r979, %r980;
|
||
xor.b32 %r982, %r981, %r971;
|
||
add.s32 %r983, %r955, %r982;
|
||
add.s32 %r984, %r983, -30611744;
|
||
shf.l.wrap.b32 %r985, %r984, %r984, 10;
|
||
add.s32 %r986, %r985, %r979;
|
||
not.b32 %r987, %r971;
|
||
or.b32 %r988, %r986, %r987;
|
||
xor.b32 %r989, %r988, %r979;
|
||
add.s32 %r990, %r1953, %r963;
|
||
add.s32 %r991, %r990, %r989;
|
||
add.s32 %r992, %r991, -1560198380;
|
||
shf.l.wrap.b32 %r993, %r992, %r992, 15;
|
||
add.s32 %r994, %r993, %r986;
|
||
not.b32 %r995, %r979;
|
||
or.b32 %r996, %r994, %r995;
|
||
xor.b32 %r997, %r996, %r986;
|
||
add.s32 %r998, %r1946, %r971;
|
||
add.s32 %r999, %r998, %r997;
|
||
add.s32 %r1000, %r999, 1309151649;
|
||
shf.l.wrap.b32 %r1001, %r1000, %r1000, 21;
|
||
add.s32 %r1002, %r1001, %r994;
|
||
not.b32 %r1003, %r986;
|
||
or.b32 %r1004, %r1002, %r1003;
|
||
xor.b32 %r1005, %r1004, %r994;
|
||
add.s32 %r1006, %r1951, %r979;
|
||
add.s32 %r1007, %r1006, %r1005;
|
||
add.s32 %r1008, %r1007, -145523070;
|
||
shf.l.wrap.b32 %r1009, %r1008, %r1008, 6;
|
||
add.s32 %r1010, %r1009, %r1002;
|
||
not.b32 %r1011, %r994;
|
||
or.b32 %r1012, %r1010, %r1011;
|
||
xor.b32 %r1013, %r1012, %r1002;
|
||
add.s32 %r1014, %r1950, %r986;
|
||
add.s32 %r1015, %r1014, %r1013;
|
||
add.s32 %r1016, %r1015, -1120210379;
|
||
shf.l.wrap.b32 %r1017, %r1016, %r1016, 10;
|
||
add.s32 %r1018, %r1017, %r1010;
|
||
not.b32 %r1019, %r1002;
|
||
or.b32 %r1020, %r1018, %r1019;
|
||
xor.b32 %r1021, %r1020, %r1010;
|
||
add.s32 %r1022, %r1956, %r994;
|
||
add.s32 %r1023, %r1022, %r1021;
|
||
add.s32 %r1024, %r1023, 718787259;
|
||
shf.l.wrap.b32 %r1025, %r1024, %r1024, 15;
|
||
add.s32 %r1026, %r1025, %r1018;
|
||
not.b32 %r1027, %r1010;
|
||
or.b32 %r1028, %r1026, %r1027;
|
||
xor.b32 %r1029, %r1028, %r1018;
|
||
add.s32 %r1030, %r1948, %r1002;
|
||
add.s32 %r1031, %r1030, %r1029;
|
||
add.s32 %r1032, %r1031, -343485551;
|
||
shf.l.wrap.b32 %r1033, %r1032, %r1032, 21;
|
||
add.s32 %r1986, %r1018, 271733878;
|
||
add.s32 %r1984, %r1026, -1732584194;
|
||
add.s32 %r1034, %r1026, %r1033;
|
||
add.s32 %r1982, %r1034, -271733879;
|
||
add.s32 %r1153, %r1010, 1732584193;
|
||
add.s32 %r1039, %r1039, 1;
|
||
add.s32 %r1040, %r1040, 1;
|
||
setp.lt.u32 %p82, %r1040, %r1036;
|
||
mov.u32 %r1149, %r1153;
|
||
mov.u32 %r1944, %r1982;
|
||
mov.u32 %r1943, %r1984;
|
||
mov.u32 %r1942, %r1986;
|
||
@%p82 bra BB2_3;
|
||
|
||
BB2_132:
|
||
st.global.u32 [%rd2], %r1149;
|
||
st.global.u32 [%rd2+4], %r1944;
|
||
st.global.u32 [%rd2+8], %r1943;
|
||
st.global.u32 [%rd2+12], %r1942;
|
||
|
||
BB2_133:
|
||
ret;
|
||
}
|
||
|
||
// .globl m00500_comp
|
||
.entry m00500_comp(
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_0,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_1,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_2,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_3,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_4,
|
||
.param .u64 .ptr .global .align 1 m00500_comp_param_5,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_6,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_7,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_8,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_9,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_10,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_11,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_12,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_13,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_14,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_15,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_16,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_17,
|
||
.param .u64 .ptr .global .align 1 m00500_comp_param_18,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_19,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_20,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_21,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_22,
|
||
.param .u64 .ptr .global .align 4 m00500_comp_param_23,
|
||
.param .u32 m00500_comp_param_24,
|
||
.param .u32 m00500_comp_param_25,
|
||
.param .u32 m00500_comp_param_26,
|
||
.param .u32 m00500_comp_param_27,
|
||
.param .u32 m00500_comp_param_28,
|
||
.param .u32 m00500_comp_param_29,
|
||
.param .u32 m00500_comp_param_30,
|
||
.param .u32 m00500_comp_param_31,
|
||
.param .u32 m00500_comp_param_32,
|
||
.param .u32 m00500_comp_param_33,
|
||
.param .u32 m00500_comp_param_34
|
||
)
|
||
{
|
||
.reg .pred %p<25>;
|
||
.reg .b32 %r<106>;
|
||
.reg .b64 %rd<41>;
|
||
|
||
|
||
ld.param.u64 %rd2, [m00500_comp_param_4];
|
||
ld.param.u64 %rd3, [m00500_comp_param_6];
|
||
ld.param.u64 %rd4, [m00500_comp_param_7];
|
||
ld.param.u64 %rd5, [m00500_comp_param_8];
|
||
ld.param.u64 %rd6, [m00500_comp_param_9];
|
||
ld.param.u64 %rd7, [m00500_comp_param_10];
|
||
ld.param.u64 %rd8, [m00500_comp_param_11];
|
||
ld.param.u64 %rd9, [m00500_comp_param_12];
|
||
ld.param.u64 %rd10, [m00500_comp_param_13];
|
||
ld.param.u64 %rd11, [m00500_comp_param_14];
|
||
ld.param.u64 %rd12, [m00500_comp_param_15];
|
||
ld.param.u64 %rd13, [m00500_comp_param_16];
|
||
ld.param.u64 %rd14, [m00500_comp_param_19];
|
||
ld.param.u32 %r27, [m00500_comp_param_24];
|
||
ld.param.u32 %r28, [m00500_comp_param_25];
|
||
ld.param.u32 %r29, [m00500_comp_param_26];
|
||
ld.param.u32 %r30, [m00500_comp_param_27];
|
||
ld.param.u32 %r31, [m00500_comp_param_31];
|
||
ld.param.u32 %r32, [m00500_comp_param_32];
|
||
ld.param.u32 %r33, [m00500_comp_param_34];
|
||
mov.b32 %r34, %envreg3;
|
||
mov.u32 %r35, %ctaid.x;
|
||
mov.u32 %r36, %ntid.x;
|
||
mad.lo.s32 %r37, %r35, %r36, %r34;
|
||
mov.u32 %r38, %tid.x;
|
||
add.s32 %r1, %r37, %r38;
|
||
setp.ge.u32 %p1, %r1, %r33;
|
||
@%p1 bra BB3_26;
|
||
|
||
mul.wide.u32 %rd15, %r1, 16;
|
||
add.s64 %rd16, %rd2, %rd15;
|
||
ld.global.u32 %r2, [%rd16+4];
|
||
ld.global.u32 %r3, [%rd16+8];
|
||
ld.global.u32 %r4, [%rd16+12];
|
||
and.b32 %r5, %r28, 31;
|
||
ld.global.u32 %r6, [%rd16];
|
||
shr.u32 %r39, %r6, %r5;
|
||
and.b32 %r40, %r39, %r27;
|
||
mul.wide.u32 %rd17, %r40, 4;
|
||
add.s64 %rd18, %rd3, %rd17;
|
||
and.b32 %r41, %r6, 31;
|
||
mov.u32 %r42, 1;
|
||
shl.b32 %r7, %r42, %r41;
|
||
ld.global.u32 %r43, [%rd18];
|
||
and.b32 %r44, %r43, %r7;
|
||
setp.eq.s32 %p2, %r44, 0;
|
||
@%p2 bra BB3_26;
|
||
|
||
shr.u32 %r45, %r2, %r5;
|
||
and.b32 %r46, %r45, %r27;
|
||
mul.wide.u32 %rd19, %r46, 4;
|
||
add.s64 %rd20, %rd4, %rd19;
|
||
and.b32 %r47, %r2, 31;
|
||
shl.b32 %r8, %r42, %r47;
|
||
ld.global.u32 %r49, [%rd20];
|
||
and.b32 %r50, %r49, %r8;
|
||
setp.eq.s32 %p3, %r50, 0;
|
||
@%p3 bra BB3_26;
|
||
|
||
shr.u32 %r51, %r3, %r5;
|
||
and.b32 %r52, %r51, %r27;
|
||
mul.wide.u32 %rd21, %r52, 4;
|
||
add.s64 %rd22, %rd5, %rd21;
|
||
and.b32 %r53, %r3, 31;
|
||
shl.b32 %r9, %r42, %r53;
|
||
ld.global.u32 %r55, [%rd22];
|
||
and.b32 %r56, %r55, %r9;
|
||
setp.eq.s32 %p4, %r56, 0;
|
||
@%p4 bra BB3_26;
|
||
|
||
shr.u32 %r57, %r4, %r5;
|
||
and.b32 %r58, %r57, %r27;
|
||
mul.wide.u32 %rd23, %r58, 4;
|
||
add.s64 %rd24, %rd6, %rd23;
|
||
and.b32 %r59, %r4, 31;
|
||
shl.b32 %r10, %r42, %r59;
|
||
ld.global.u32 %r61, [%rd24];
|
||
and.b32 %r62, %r61, %r10;
|
||
setp.eq.s32 %p5, %r62, 0;
|
||
@%p5 bra BB3_26;
|
||
|
||
and.b32 %r11, %r29, 31;
|
||
shr.u32 %r63, %r6, %r11;
|
||
and.b32 %r64, %r63, %r27;
|
||
mul.wide.u32 %rd25, %r64, 4;
|
||
add.s64 %rd26, %rd7, %rd25;
|
||
ld.global.u32 %r65, [%rd26];
|
||
and.b32 %r66, %r65, %r7;
|
||
setp.eq.s32 %p6, %r66, 0;
|
||
@%p6 bra BB3_26;
|
||
|
||
shr.u32 %r67, %r2, %r11;
|
||
and.b32 %r68, %r67, %r27;
|
||
mul.wide.u32 %rd27, %r68, 4;
|
||
add.s64 %rd28, %rd8, %rd27;
|
||
ld.global.u32 %r69, [%rd28];
|
||
and.b32 %r70, %r69, %r8;
|
||
setp.eq.s32 %p7, %r70, 0;
|
||
@%p7 bra BB3_26;
|
||
|
||
shr.u32 %r71, %r3, %r11;
|
||
and.b32 %r72, %r71, %r27;
|
||
mul.wide.u32 %rd29, %r72, 4;
|
||
add.s64 %rd30, %rd9, %rd29;
|
||
ld.global.u32 %r73, [%rd30];
|
||
and.b32 %r74, %r73, %r9;
|
||
setp.eq.s32 %p8, %r74, 0;
|
||
@%p8 bra BB3_26;
|
||
|
||
shr.u32 %r75, %r4, %r11;
|
||
and.b32 %r76, %r75, %r27;
|
||
mul.wide.u32 %rd31, %r76, 4;
|
||
add.s64 %rd32, %rd10, %rd31;
|
||
ld.global.u32 %r77, [%rd32];
|
||
and.b32 %r78, %r77, %r10;
|
||
setp.eq.s32 %p9, %r78, 0;
|
||
@%p9 bra BB3_26;
|
||
|
||
setp.eq.s32 %p10, %r31, 0;
|
||
mov.u32 %r97, 0;
|
||
mov.u32 %r79, -1;
|
||
mov.u32 %r105, %r79;
|
||
@%p10 bra BB3_21;
|
||
|
||
mov.u32 %r96, %r31;
|
||
|
||
BB3_11:
|
||
mov.u32 %r12, %r96;
|
||
shr.u32 %r14, %r12, 1;
|
||
add.s32 %r15, %r14, %r97;
|
||
cvt.u64.u32 %rd33, %r15;
|
||
cvt.u64.u32 %rd34, %r32;
|
||
add.s64 %rd35, %rd33, %rd34;
|
||
shl.b64 %rd36, %rd35, 4;
|
||
add.s64 %rd1, %rd12, %rd36;
|
||
ld.global.u32 %r16, [%rd1+12];
|
||
setp.gt.u32 %p11, %r4, %r16;
|
||
mov.u32 %r103, %r42;
|
||
@%p11 bra BB3_19;
|
||
|
||
setp.lt.u32 %p12, %r4, %r16;
|
||
mov.u32 %r82, -1;
|
||
mov.u32 %r103, %r82;
|
||
@%p12 bra BB3_19;
|
||
|
||
ld.global.u32 %r17, [%rd1+8];
|
||
setp.gt.u32 %p13, %r3, %r17;
|
||
mov.u32 %r98, %r42;
|
||
mov.u32 %r103, %r98;
|
||
@%p13 bra BB3_19;
|
||
|
||
setp.lt.u32 %p14, %r3, %r17;
|
||
mov.u32 %r101, %r82;
|
||
mov.u32 %r103, %r101;
|
||
@%p14 bra BB3_19;
|
||
|
||
ld.global.u32 %r18, [%rd1+4];
|
||
setp.gt.u32 %p15, %r2, %r18;
|
||
mov.u32 %r99, %r42;
|
||
mov.u32 %r103, %r99;
|
||
@%p15 bra BB3_19;
|
||
|
||
setp.lt.u32 %p16, %r2, %r18;
|
||
mov.u32 %r102, %r82;
|
||
mov.u32 %r103, %r102;
|
||
@%p16 bra BB3_19;
|
||
|
||
ld.global.u32 %r19, [%rd1];
|
||
setp.gt.u32 %p17, %r6, %r19;
|
||
mov.u32 %r100, %r42;
|
||
mov.u32 %r103, %r100;
|
||
@%p17 bra BB3_19;
|
||
|
||
setp.lt.u32 %p18, %r6, %r19;
|
||
selp.b32 %r20, -1, 0, %p18;
|
||
mov.u32 %r103, %r20;
|
||
|
||
BB3_19:
|
||
mov.u32 %r21, %r103;
|
||
add.s32 %r88, %r14, 1;
|
||
setp.gt.s32 %p19, %r21, 0;
|
||
selp.b32 %r89, %r88, 0, %p19;
|
||
add.s32 %r97, %r89, %r97;
|
||
selp.b32 %r90, -1, 0, %p19;
|
||
add.s32 %r91, %r90, %r12;
|
||
shr.u32 %r23, %r91, 1;
|
||
setp.eq.s32 %p20, %r21, 0;
|
||
mov.u32 %r105, %r15;
|
||
@%p20 bra BB3_21;
|
||
|
||
setp.ne.s32 %p21, %r23, 0;
|
||
mov.u32 %r96, %r23;
|
||
mov.u32 %r104, %r79;
|
||
mov.u32 %r105, %r104;
|
||
@%p21 bra BB3_11;
|
||
|
||
BB3_21:
|
||
setp.eq.s32 %p22, %r105, -1;
|
||
@%p22 bra BB3_26;
|
||
|
||
add.s32 %r25, %r105, %r32;
|
||
mul.wide.u32 %rd37, %r25, 4;
|
||
add.s64 %rd38, %rd13, %rd37;
|
||
atom.global.add.u32 %r93, [%rd38], 1;
|
||
setp.ne.s32 %p23, %r93, 0;
|
||
@%p23 bra BB3_26;
|
||
|
||
atom.global.add.u32 %r26, [%rd14], 1;
|
||
setp.lt.u32 %p24, %r26, %r31;
|
||
@%p24 bra BB3_25;
|
||
bra.uni BB3_24;
|
||
|
||
BB3_25:
|
||
mul.wide.u32 %rd39, %r26, 20;
|
||
add.s64 %rd40, %rd11, %rd39;
|
||
st.global.u32 [%rd40], %r30;
|
||
st.global.u32 [%rd40+4], %r105;
|
||
st.global.u32 [%rd40+8], %r25;
|
||
st.global.u32 [%rd40+12], %r1;
|
||
mov.u32 %r95, 0;
|
||
st.global.u32 [%rd40+16], %r95;
|
||
bra.uni BB3_26;
|
||
|
||
BB3_24:
|
||
atom.global.add.u32 %r94, [%rd14], -1;
|
||
|
||
BB3_26:
|
||
ret;
|
||
}
|
||
|
||
|
||
|