// Disassembling _Z11mem3_kernelPiS_ii .entry _Z11mem3_kernelPiS_ii { .lmem 0 .smem 48 .reg 10 .bar 1 set.gt.s32 $p1|$r1, s[0x0020], $r124 cvt.rn.s32 $p0|$o127, $r1 cvt.rn.u32.u16 $r0, $r0.lo @$p1.eq bra.label label2 join.label label1 cvt.rn.u32.u16 $r7, %ntid.y shl.u32 $r8, $r0, 0x00000002 mov.half.b32 $r1, $r0 add.half.b32 $r6, s[0x0020], $r0 shl.u32 $r5, $r7, 0x00000002 add.b32 $r4, $r8, 0x00000030 add.u32 $r8, s[0x0010], $r8 label0: mov.u32 $r9, g[$r8] movsh.b32 $ofs1, $r4, 0x00000000 add.u32 $r1, $r1, $r7 set.lt.s32 $p1|$o127, $r1, $r6 mov.b32 s[$ofs1+0x0000], $r9 add.half.b32 $r8, $r5, $r8 add.half.b32 $r4, $r4, $r5 @$p1.ne bra.label label0 label1: nop.join label2: bar.sync.u32 0x00000000 set.eq.s32 $p2|$r1, $r0, $r124 cvt.rn.s32 $p1|$o127, $r1 mov.b32 $r1, %clock @$p2.neu shl.u32 $r2, $r1, 0x00000001 bar.sync.u32 0x00000000 cvt.rn.abs.u32.s32 $r1, s[0x0024] mov.b32 $r4, 0xfc0fc0fd mul24.lo.u32.u16.u16 $r5, $r1.lo, $r4.hi mad24.lo.u32.u16.u16.u32 $p2|$r5, $r1.hi, $r4.lo, $r5 shl.u32 $r6, $r5, 0x00000010 shr.u32 $r5, $r5, 0x00000010 mad24.lo.u32.u16.u16.u32 $p3|$o127, $r1.lo, $r4.lo, $r6 @$p2.cf add.u32 $r5, $r5, c1[0x0000] mad24c1.lo.u32.u16.u16.u32 $r1, -$r1.hi, $r4.hi, -$r5 shr.s32 $r1, $r1, 0x00000007 set.lt.s32 $p2|$o127, s[0x0024], $r124 subr.b32 $r4, $r1, 0x00000000 @$p2.ne mov.b32 $r1, $r4 shl.u32 $r4, $r1, 0x00000007 shl.u32 $r1, $r1, 0x00000001 add.u32 $r4, $r4, $r1 set.le.s32 $p2|$o127, $r4, $r124 mov.b32 $r5, $r124 @$p2.ne bra.label label4 movsh.b32 $ofs1, $r0, 0x00000002 label3: mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 add.b32 $r5, $r5, 0x00000082 set.gt.s32 $p2|$o127, $r4, $r5 @$p2.ne bra.label label3 label4: set.le.s32 $p2|$o127, s[0x0024], $r5 @$p2.ne bra.label label6 movsh.b32 $ofs1, $r0, 0x00000002 label5: mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0430], $r1 mov.b32 s[$ofs1+0x0430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x0c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x1c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2830], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x2c30], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3030], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3430], $r1 mov.b32 $r1, s[$ofs1+0x0030] mov.b32 s[$ofs1+0x3830], $r1 bar.sync.u32 0x00000000 add.b32 $r5, $r5, 0x00000001 set.ne.s32 $p2|$o127, s[0x0024], $r5 @$p2.ne bra.label label5 label6: mov.b32 $r1, %clock @$p1.neu shl.u32 $r3, $r1, 0x00000001 bar.sync.u32 0x00000000 @$p0.eq bra.label label9 cvt.rn.u32.u16 $r7, %ntid.y shl.u32 $r1, $r0, 0x00000002 join.label label8 add.u32 $r6, s[0x0020], $r0 shl.u32 $r5, $r7, 0x00000002 add.b32 $r4, $r1, 0x00000030 add.u32 $r8, s[0x0018], $r1 label7: movsh.b32 $ofs1, $r4, 0x00000000 mov.half.b32 $r1, s[$ofs1+0x0000] add.half.b32 $r0, $r0, $r7 mov.u32 g[$r8], $r1 set.lt.s32 $p0|$o127, $r0, $r6 add.half.b32 $r4, $r4, $r5 add.half.b32 $r8, $r5, $r8 @$p0.ne bra.label label7 label8: nop.join label9: bar.sync.u32 0x00000000 @$p1.eq return sub.half.b32 $r1, $r3, $r66 mov.half.b32 $r0, s[0x0018] mov.end.u32 g[$r0], $r1 #.constseg 1:0x0000 const #{ #d.u32 0x00010000 // 0000 #} }