// Disassembling matrix_simpletile (0)
000000: 10000005 0403c780 mov.b32 $r1, $r0
000008: d0800601 00400780 and.b16 $r0.lo, $r1.hi, c1[0x0000]
000010: a0000409 04000780 cvt.u32.u16 $r2, $r1.lo
000018: 40100005 00000003 mul24.lo.u32.u16.u16 $r1, $r0.lo, 0x0010
000020: 20000419 04004780 add.u32 $r6, $r2, $r1
000028: 40000005 00000007 mul24.lo.u32.u16.u16 $r1, $r0.lo, 0x0040
000030: 20078c0d 00000013 add.b32 $r3, $r6, 0x00000107
000038: 201c8205 00000043 add.b32 $r1, $r1, 0x0000041c
000040: 00020609 c0000780 movsh.b32 $ofs2, $r3, 0x00000002
000048: 00000205 c0000780 movsh.b32 $ofs1, $r1, 0x00000000
000050: a004d003 00000000 join.label label1
000058: 10004e05 0023c780 mov.b16 $r0.hi, s[0x000e]
000060: a0004c05 04200780 cvt.u32.u16 $r1, s[0x000c]
000068: 10008025 00000003 mov.b32 $r9, 0x00000000
000070: 4000830d 00000403 mul24.lo.s32.s16.s16 $r67, $r0.hi, 0x4000
000078: 30040215 c4100780 shl.u32 $r5, $r1, 0x00000004
000080: 40008101 00000043 mul24.lo.s32.s16.s16 $r64, $r0.lo, 0x0400
000088: 20008611 00000043 add.b32 $r4, $r3, 0x00000400
000090: 2003800c          add.half.b32 $r3, $r0, $r3
000094: 20048010          add.half.b32 $r4, $r0, $r4
000098: 20058014          add.half.b32 $r5, $r0, $r5
00009c: 2003840c          add.half.b32 $r3, $r2, $r3
0000a0: 20048410          add.half.b32 $r4, $r2, $r4
0000a4: 20058414          add.half.b32 $r5, $r2, $r5
0000a8: 3002060d c4100780 shl.u32 $r3, $r3, 0x00000002
0000b0: 30020811 c4100780 shl.u32 $r4, $r4, 0x00000002
0000b8: 30020a15 c4100780 shl.u32 $r5, $r5, 0x00000002
0000c0: 2103ea0c          add.half.b32 $r3, s[0x0014], $r3
0000c4: 2104ea10          add.half.b32 $r4, s[0x0014], $r4
0000c8: 2000cc15 04214780 add.u32 $r5, s[0x0018], $r5
0000d0: d00e0621 80c00780 label0: mov.u32 $r8, g[$r3]
0000d8: d00e0a1d 80c00780 mov.u32 $r7, g[$r5]
0000e0: 08000001 e4220780 mov.b32 s[$ofs2+0x0000], $r8
0000e8: 00020c0d c0000780 movsh.b32 $ofs3, $r6, 0x00000002
0000f0: 0c000e01 e421c780 mov.b32 s[$ofs3+0x001c], $r7
0000f8: 861ffe03 00000000 bar.sync.u32 0x00000000
000100: 0002040d c0000780 movsh.b32 $ofs3, $r2, 0x00000002
000108: dc073811 20000780 add.b32 $ofs4, $ofs3, 0x0000039c
000110: 1000c021 0423c784 mov.b32 $r8, s[$ofs4+0x0000]
000118: 1000e01d 0423c784 mov.b32 $r7, s[$ofs4+0x0040]
000120: dc013811 20000780 add.b32 $ofs4, $ofs3, 0x0000009c
000128: 1000c031 0423c784 mov.b32 $r12, s[$ofs4+0x0000]
000130: 1000e02d 0423c784 mov.b32 $r11, s[$ofs4+0x0040]
000138: 1c00ce29 0423c780 mov.b32 $r10, s[$ofs3+0x001c]
000140: e40ac029 00224780 mad.rn.f32 $r10, s[$ofs1+0x0000], $r10, $r9
000148: 1c00ee25 0423c780 mov.b32 $r9, s[$ofs3+0x005c]
000150: dc023811 20000780 add.b32 $ofs4, $ofs3, 0x0000011c
000158: e409c235 00228780 mad.rn.f32 $r13, s[$ofs1+0x0004], $r9, $r10
000160: 1000c029 0423c784 mov.b32 $r10, s[$ofs4+0x0000]
000168: 1000e025 0423c784 mov.b32 $r9, s[$ofs4+0x0040]
000170: e40cc431 00234780 mad.rn.f32 $r12, s[$ofs1+0x0008], $r12, $r13
000178: dc033811 20000780 add.b32 $ofs4, $ofs3, 0x0000019c
000180: e40bc635 00230780 mad.rn.f32 $r13, s[$ofs1+0x000c], $r11, $r12
000188: 1000c031 0423c784 mov.b32 $r12, s[$ofs4+0x0000]
000190: 1000e02d 0423c784 mov.b32 $r11, s[$ofs4+0x0040]
000198: e40ac829 00234780 mad.rn.f32 $r10, s[$ofs1+0x0010], $r10, $r13
0001a0: dc043811 20000780 add.b32 $ofs4, $ofs3, 0x0000021c
0001a8: e409ca35 00228780 mad.rn.f32 $r13, s[$ofs1+0x0014], $r9, $r10
0001b0: 1000c029 0423c784 mov.b32 $r10, s[$ofs4+0x0000]
0001b8: 1000e025 0423c784 mov.b32 $r9, s[$ofs4+0x0040]
0001c0: e40ccc31 00234780 mad.rn.f32 $r12, s[$ofs1+0x0018], $r12, $r13
0001c8: dc053811 20000780 add.b32 $ofs4, $ofs3, 0x0000029c
0001d0: e40bce35 00230780 mad.rn.f32 $r13, s[$ofs1+0x001c], $r11, $r12
0001d8: 1000c031 0423c784 mov.b32 $r12, s[$ofs4+0x0000]
0001e0: 1000e02d 0423c784 mov.b32 $r11, s[$ofs4+0x0040]
0001e8: e40ad029 00234780 mad.rn.f32 $r10, s[$ofs1+0x0020], $r10, $r13
0001f0: e409d225 00228780 mad.rn.f32 $r9, s[$ofs1+0x0024], $r9, $r10
0001f8: e40cd425 00224780 mad.rn.f32 $r9, s[$ofs1+0x0028], $r12, $r9
000200: dc06380d 20000780 add.b32 $ofs3, $ofs3, 0x0000031c
000208: e40bd629 00224780 mad.rn.f32 $r10, s[$ofs1+0x002c], $r11, $r9
000210: 1c00c025 0423c780 mov.b32 $r9, s[$ofs3+0x0000]
000218: e409d829 00228780 mad.rn.f32 $r10, s[$ofs1+0x0030], $r9, $r10
000220: 1c00e025 0423c780 mov.b32 $r9, s[$ofs3+0x0040]
000228: e409da25 00228780 mad.rn.f32 $r9, s[$ofs1+0x0034], $r9, $r10
000230: e408dc21 00224780 mad.rn.f32 $r8, s[$ofs1+0x0038], $r8, $r9
000238: e407de25 00220780 mad.rn.f32 $r9, s[$ofs1+0x003c], $r7, $r8
000240: 861ffe03 00000000 bar.sync.u32 0x00000000
000248: 2000860d 00000007 add.b32 $r3, $r3, 0x00000040
000250: 300407fd 6c0147c8 set.ne.s32 $p0|$o127, $r3, $r4
000258: 20008a15 00001003 add.b32 $r5, $r5, 0x00010000
000260: 1001a003 00000280 @$p0.ne bra.label label0
000268: 10004e19 0023c782 label1: mov.join.b16 $r3.lo, s[0x000e]
000270: 40008d0d 00000043 mul24.lo.s32.s16.s16 $r67, $r3.lo, 0x0400
000278: 20000205 0400c780 add.u32 $r1, $r1, $r3
000280: 30040205 c4100780 shl.u32 $r1, $r1, 0x00000004
000288: 20018000          add.half.b32 $r0, $r0, $r1
00028c: 20008400          add.half.b32 $r0, $r2, $r0
000290: 30020001 c4100780 shl.u32 $r0, $r0, 0x00000002
000298: 2000c801 04200780 add.u32 $r0, s[0x0010], $r0
0002a0: d00e0025 a0c00781 mov.end.u32 g[$r0], $r9
// segment: const (1:0000)
0000: 000003ff 
// Disassembling _Z22matrix_large_tile_basePfS_S_ (1)
000000: 10000005 0403c780 mov.b32 $r1, $r0
000008: d0820601 00400780 and.b16 $r0.lo, $r1.hi, c1[0x0004]
000010: a0000465 04000780 cvt.u32.u16 $r25, $r1.lo
000018: a0000049 04000780 cvt.u32.u16 $r18, $r0.lo
000020: 60803201 60448780 mad24.lo.u32 $r0, $r25, c1[0x0000], $r18
000028: 00020005 c0000780 movsh.b32 $ofs1, $r0, 0x00000002
000030: a0152003 00000000 join.label label1
000038: 10008001 00000003 mov.b32 $r0, 0x00000000
000040: 10008005 00000003 mov.b32 $r1, 0x00000000
000048: 10008009 00000003 mov.b32 $r2, 0x00000000
000050: 1000800d 00000003 mov.b32 $r3, 0x00000000
000058: 10008011 00000003 mov.b32 $r4, 0x00000000
000060: 10008015 00000003 mov.b32 $r5, 0x00000000
000068: 10008019 00000003 mov.b32 $r6, 0x00000000
000070: 1000801d 00000003 mov.b32 $r7, 0x00000000
000078: 10008021 00000003 mov.b32 $r8, 0x00000000
000080: 10008025 00000003 mov.b32 $r9, 0x00000000
000088: 10008029 00000003 mov.b32 $r10, 0x00000000
000090: 1000802d 00000003 mov.b32 $r11, 0x00000000
000098: 10008031 00000003 mov.b32 $r12, 0x00000000
0000a0: 10008035 00000003 mov.b32 $r13, 0x00000000
0000a8: 10008039 00000003 mov.b32 $r14, 0x00000000
0000b0: 10008041 00000003 mov.b32 $r16, 0x00000000
0000b8: 41002c5d 00000013 mul24.lo.u32.u16.u16 $r23, s[0x000c], 0x0100
0000c0: 41102e59 00000003 mul24.lo.u32.u16.u16 $r22, s[0x000e], 0x0010
0000c8: 2017b244          add.half.b32 $r17, $r25, $r23
0000cc: 2016a43c          add.half.b32 $r15, $r18, $r22
0000d0: 30042461 c4100780 shl.u32 $r24, $r18, 0x00000004
0000d8: 300a1e3d c4100780 shl.u32 $r15, $r15, 0x0000000a
0000e0: 2011b044          add.half.b32 $r17, $r24, $r17
0000e4: 200fb23c          add.half.b32 $r15, $r25, $r15
0000e8: 30022245 c4100780 shl.u32 $r17, $r17, 0x00000002
0000f0: 30021e3d c4100780 shl.u32 $r15, $r15, 0x00000002
0000f8: 2111ea54          add.half.b32 $r21, s[0x0014], $r17
0000fc: 210fe848          add.half.b32 $r18, s[0x0010], $r15
000100: 2000aa51 00001003 add.b32 $r20, $r21, 0x00010000
000108: 2000a44d 00000103 add.b32 $r19, $r18, 0x00001000
000110: d00e243d 80c00780 label0: mov.u32 $r15, g[$r18]
000118: 04000e01 e423c780 mov.b32 s[$ofs1+0x001c], $r15
000120: 861ffe03 00000000 bar.sync.u32 0x00000000
000128: d00e2a3d 80c00780 mov.u32 $r15, g[$r21]
000130: e00fce01 00200780 mad.rn.f32 $r0, s[0x001c], $r15, $r0
000138: e00fd005 00204780 mad.rn.f32 $r1, s[0x0020], $r15, $r1
000140: e00fd209 00208780 mad.rn.f32 $r2, s[0x0024], $r15, $r2
000148: e00fd40d 0020c780 mad.rn.f32 $r3, s[0x0028], $r15, $r3
000150: e00fd611 00210780 mad.rn.f32 $r4, s[0x002c], $r15, $r4
000158: e00fd815 00214780 mad.rn.f32 $r5, s[0x0030], $r15, $r5
000160: e00fda19 00218780 mad.rn.f32 $r6, s[0x0034], $r15, $r6
000168: e00fdc1d 0021c780 mad.rn.f32 $r7, s[0x0038], $r15, $r7
000170: e00fde21 00220780 mad.rn.f32 $r8, s[0x003c], $r15, $r8
000178: e00fe025 00224780 mad.rn.f32 $r9, s[0x0040], $r15, $r9
000180: e00fe229 00228780 mad.rn.f32 $r10, s[0x0044], $r15, $r10
000188: e00fe42d 0022c780 mad.rn.f32 $r11, s[0x0048], $r15, $r11
000190: e00fe631 00230780 mad.rn.f32 $r12, s[0x004c], $r15, $r12
000198: e00fe835 00234780 mad.rn.f32 $r13, s[0x0050], $r15, $r13
0001a0: e00fea39 00238780 mad.rn.f32 $r14, s[0x0054], $r15, $r14
0001a8: e00fec45 00240780 mad.rn.f32 $r17, s[0x0058], $r15, $r16
0001b0: d001000d 20000780 mov.b32 $ofs3, 0x00000080
0001b8: 2000aa3d 00000103 add.b32 $r15, $r21, 0x00001000
0001c0: d00e1e3d 80c00780 mov.u32 $r15, g[$r15]
0001c8: 2000aa41 00000203 add.b32 $r16, $r21, 0x00002000
0001d0: d00e2041 80c00780 mov.u32 $r16, g[$r16]
0001d8: e00ff001 00200780 mad.rn.f32 $r0, s[0x0060], $r15, $r0
0001e0: e00ff205 00204780 mad.rn.f32 $r1, s[0x0064], $r15, $r1
0001e8: e00ff409 00208780 mad.rn.f32 $r2, s[0x0068], $r15, $r2
0001f0: e00ff60d 0020c780 mad.rn.f32 $r3, s[0x006c], $r15, $r3
0001f8: e00ff811 00210780 mad.rn.f32 $r4, s[0x0070], $r15, $r4
000200: e00ffa15 00214780 mad.rn.f32 $r5, s[0x0074], $r15, $r5
000208: e00ffc19 00218780 mad.rn.f32 $r6, s[0x0078], $r15, $r6
000210: e00ffe1d 0021c780 mad.rn.f32 $r7, s[0x007c], $r15, $r7
000218: ec0fc021 00220780 mad.rn.f32 $r8, s[$ofs3+0x0000], $r15, $r8
000220: ec0fc225 00224780 mad.rn.f32 $r9, s[$ofs3+0x0004], $r15, $r9
000228: ec0fc429 00228780 mad.rn.f32 $r10, s[$ofs3+0x0008], $r15, $r10
000230: ec0fc62d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x000c], $r15, $r11
000238: ec0fc831 00230780 mad.rn.f32 $r12, s[$ofs3+0x0010], $r15, $r12
000240: ec0fca35 00234780 mad.rn.f32 $r13, s[$ofs3+0x0014], $r15, $r13
000248: ec0fcc39 00238780 mad.rn.f32 $r14, s[$ofs3+0x0018], $r15, $r14
000250: ec0fce3d 00244780 mad.rn.f32 $r15, s[$ofs3+0x001c], $r15, $r17
000258: ec10d201 00200780 mad.rn.f32 $r0, s[$ofs3+0x0024], $r16, $r0
000260: ec10d405 00204780 mad.rn.f32 $r1, s[$ofs3+0x0028], $r16, $r1
000268: ec10d609 00208780 mad.rn.f32 $r2, s[$ofs3+0x002c], $r16, $r2
000270: ec10d80d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0030], $r16, $r3
000278: ec10da11 00210780 mad.rn.f32 $r4, s[$ofs3+0x0034], $r16, $r4
000280: ec10dc15 00214780 mad.rn.f32 $r5, s[$ofs3+0x0038], $r16, $r5
000288: ec10de19 00218780 mad.rn.f32 $r6, s[$ofs3+0x003c], $r16, $r6
000290: ec10e01d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0040], $r16, $r7
000298: ec10e221 00220780 mad.rn.f32 $r8, s[$ofs3+0x0044], $r16, $r8
0002a0: ec10e425 00224780 mad.rn.f32 $r9, s[$ofs3+0x0048], $r16, $r9
0002a8: ec10e629 00228780 mad.rn.f32 $r10, s[$ofs3+0x004c], $r16, $r10
0002b0: ec10e82d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x0050], $r16, $r11
0002b8: ec10ea31 00230780 mad.rn.f32 $r12, s[$ofs3+0x0054], $r16, $r12
0002c0: ec10ec35 00234780 mad.rn.f32 $r13, s[$ofs3+0x0058], $r16, $r13
0002c8: ec10ee39 00238780 mad.rn.f32 $r14, s[$ofs3+0x005c], $r16, $r14
0002d0: ec10f041 0023c780 mad.rn.f32 $r16, s[$ofs3+0x0060], $r16, $r15
0002d8: 2000aa3d 00000303 add.b32 $r15, $r21, 0x00003000
0002e0: d00e1e3d 80c00780 mov.u32 $r15, g[$r15]
0002e8: d0020009 20000780 mov.b32 $ofs2, 0x00000100
0002f0: 2000aa45 00000403 add.b32 $r17, $r21, 0x00004000
0002f8: d00e2245 80c00780 mov.u32 $r17, g[$r17]
000300: ec0ff401 00200780 mad.rn.f32 $r0, s[$ofs3+0x0068], $r15, $r0
000308: ec0ff605 00204780 mad.rn.f32 $r1, s[$ofs3+0x006c], $r15, $r1
000310: ec0ff809 00208780 mad.rn.f32 $r2, s[$ofs3+0x0070], $r15, $r2
000318: ec0ffa0d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0074], $r15, $r3
000320: ec0ffc11 00210780 mad.rn.f32 $r4, s[$ofs3+0x0078], $r15, $r4
000328: ec0ffe15 00214780 mad.rn.f32 $r5, s[$ofs3+0x007c], $r15, $r5
000330: e80fc019 00218780 mad.rn.f32 $r6, s[$ofs2+0x0000], $r15, $r6
000338: e80fc21d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0004], $r15, $r7
000340: e80fc421 00220780 mad.rn.f32 $r8, s[$ofs2+0x0008], $r15, $r8
000348: e80fc625 00224780 mad.rn.f32 $r9, s[$ofs2+0x000c], $r15, $r9
000350: e80fc829 00228780 mad.rn.f32 $r10, s[$ofs2+0x0010], $r15, $r10
000358: e80fca2d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0014], $r15, $r11
000360: e80fcc31 00230780 mad.rn.f32 $r12, s[$ofs2+0x0018], $r15, $r12
000368: e80fce35 00234780 mad.rn.f32 $r13, s[$ofs2+0x001c], $r15, $r13
000370: e80fd039 00238780 mad.rn.f32 $r14, s[$ofs2+0x0020], $r15, $r14
000378: e80fd23d 00240780 mad.rn.f32 $r15, s[$ofs2+0x0024], $r15, $r16
000380: 2000aa41 00000503 add.b32 $r16, $r21, 0x00005000
000388: d00e2041 80c00780 mov.u32 $r16, g[$r16]
000390: e811d601 00200780 mad.rn.f32 $r0, s[$ofs2+0x002c], $r17, $r0
000398: e811d805 00204780 mad.rn.f32 $r1, s[$ofs2+0x0030], $r17, $r1
0003a0: e811da09 00208780 mad.rn.f32 $r2, s[$ofs2+0x0034], $r17, $r2
0003a8: e811dc0d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0038], $r17, $r3
0003b0: e811de11 00210780 mad.rn.f32 $r4, s[$ofs2+0x003c], $r17, $r4
0003b8: e811e015 00214780 mad.rn.f32 $r5, s[$ofs2+0x0040], $r17, $r5
0003c0: e811e219 00218780 mad.rn.f32 $r6, s[$ofs2+0x0044], $r17, $r6
0003c8: e811e41d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0048], $r17, $r7
0003d0: e811e621 00220780 mad.rn.f32 $r8, s[$ofs2+0x004c], $r17, $r8
0003d8: e811e825 00224780 mad.rn.f32 $r9, s[$ofs2+0x0050], $r17, $r9
0003e0: e811ea29 00228780 mad.rn.f32 $r10, s[$ofs2+0x0054], $r17, $r10
0003e8: e811ec2d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0058], $r17, $r11
0003f0: e811ee31 00230780 mad.rn.f32 $r12, s[$ofs2+0x005c], $r17, $r12
0003f8: e811f035 00234780 mad.rn.f32 $r13, s[$ofs2+0x0060], $r17, $r13
000400: e811f239 00238780 mad.rn.f32 $r14, s[$ofs2+0x0064], $r17, $r14
000408: e811f43d 0023c780 mad.rn.f32 $r15, s[$ofs2+0x0068], $r17, $r15
000410: d003000d 20000780 mov.b32 $ofs3, 0x00000180
000418: 2000aa45 00000603 add.b32 $r17, $r21, 0x00006000
000420: d00e2245 80c00780 mov.u32 $r17, g[$r17]
000428: e810f801 00200780 mad.rn.f32 $r0, s[$ofs2+0x0070], $r16, $r0
000430: e810fa05 00204780 mad.rn.f32 $r1, s[$ofs2+0x0074], $r16, $r1
000438: e810fc09 00208780 mad.rn.f32 $r2, s[$ofs2+0x0078], $r16, $r2
000440: e810fe0d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x007c], $r16, $r3
000448: ec10c011 00210780 mad.rn.f32 $r4, s[$ofs3+0x0000], $r16, $r4
000450: ec10c215 00214780 mad.rn.f32 $r5, s[$ofs3+0x0004], $r16, $r5
000458: ec10c419 00218780 mad.rn.f32 $r6, s[$ofs3+0x0008], $r16, $r6
000460: ec10c61d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x000c], $r16, $r7
000468: ec10c821 00220780 mad.rn.f32 $r8, s[$ofs3+0x0010], $r16, $r8
000470: ec10ca25 00224780 mad.rn.f32 $r9, s[$ofs3+0x0014], $r16, $r9
000478: ec10cc29 00228780 mad.rn.f32 $r10, s[$ofs3+0x0018], $r16, $r10
000480: ec10ce2d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x001c], $r16, $r11
000488: ec10d031 00230780 mad.rn.f32 $r12, s[$ofs3+0x0020], $r16, $r12
000490: ec10d235 00234780 mad.rn.f32 $r13, s[$ofs3+0x0024], $r16, $r13
000498: ec10d439 00238780 mad.rn.f32 $r14, s[$ofs3+0x0028], $r16, $r14
0004a0: ec10d63d 0023c780 mad.rn.f32 $r15, s[$ofs3+0x002c], $r16, $r15
0004a8: 2000aa41 00000703 add.b32 $r16, $r21, 0x00007000
0004b0: d00e2041 80c00780 mov.u32 $r16, g[$r16]
0004b8: ec11da01 00200780 mad.rn.f32 $r0, s[$ofs3+0x0034], $r17, $r0
0004c0: ec11dc05 00204780 mad.rn.f32 $r1, s[$ofs3+0x0038], $r17, $r1
0004c8: ec11de09 00208780 mad.rn.f32 $r2, s[$ofs3+0x003c], $r17, $r2
0004d0: ec11e00d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0040], $r17, $r3
0004d8: ec11e211 00210780 mad.rn.f32 $r4, s[$ofs3+0x0044], $r17, $r4
0004e0: ec11e415 00214780 mad.rn.f32 $r5, s[$ofs3+0x0048], $r17, $r5
0004e8: ec11e619 00218780 mad.rn.f32 $r6, s[$ofs3+0x004c], $r17, $r6
0004f0: ec11e81d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0050], $r17, $r7
0004f8: ec11ea21 00220780 mad.rn.f32 $r8, s[$ofs3+0x0054], $r17, $r8
000500: ec11ec25 00224780 mad.rn.f32 $r9, s[$ofs3+0x0058], $r17, $r9
000508: ec11ee29 00228780 mad.rn.f32 $r10, s[$ofs3+0x005c], $r17, $r10
000510: ec11f02d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x0060], $r17, $r11
000518: ec11f231 00230780 mad.rn.f32 $r12, s[$ofs3+0x0064], $r17, $r12
000520: ec11f435 00234780 mad.rn.f32 $r13, s[$ofs3+0x0068], $r17, $r13
000528: ec11f639 00238780 mad.rn.f32 $r14, s[$ofs3+0x006c], $r17, $r14
000530: ec11f83d 0023c780 mad.rn.f32 $r15, s[$ofs3+0x0070], $r17, $r15
000538: d0040009 20000780 mov.b32 $ofs2, 0x00000200
000540: ec10fc01 00200780 mad.rn.f32 $r0, s[$ofs3+0x0078], $r16, $r0
000548: ec10fe05 00204780 mad.rn.f32 $r1, s[$ofs3+0x007c], $r16, $r1
000550: e810c009 00208780 mad.rn.f32 $r2, s[$ofs2+0x0000], $r16, $r2
000558: e810c20d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0004], $r16, $r3
000560: e810c411 00210780 mad.rn.f32 $r4, s[$ofs2+0x0008], $r16, $r4
000568: e810c615 00214780 mad.rn.f32 $r5, s[$ofs2+0x000c], $r16, $r5
000570: e810c819 00218780 mad.rn.f32 $r6, s[$ofs2+0x0010], $r16, $r6
000578: e810ca1d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0014], $r16, $r7
000580: e810cc21 00220780 mad.rn.f32 $r8, s[$ofs2+0x0018], $r16, $r8
000588: e810ce25 00224780 mad.rn.f32 $r9, s[$ofs2+0x001c], $r16, $r9
000590: e810d029 00228780 mad.rn.f32 $r10, s[$ofs2+0x0020], $r16, $r10
000598: e810d22d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0024], $r16, $r11
0005a0: e810d431 00230780 mad.rn.f32 $r12, s[$ofs2+0x0028], $r16, $r12
0005a8: e810d635 00234780 mad.rn.f32 $r13, s[$ofs2+0x002c], $r16, $r13
0005b0: e810d839 00238780 mad.rn.f32 $r14, s[$ofs2+0x0030], $r16, $r14
0005b8: e810da45 0023c780 mad.rn.f32 $r17, s[$ofs2+0x0034], $r16, $r15
0005c0: d005000d 20000780 mov.b32 $ofs3, 0x00000280
0005c8: 2000a83d 0ffff803 add.b32 $r15, $r20, 0xffff8000
0005d0: d00e1e3d 80c00780 mov.u32 $r15, g[$r15]
0005d8: 2000a841 0ffff903 add.b32 $r16, $r20, 0xffff9000
0005e0: d00e2041 80c00780 mov.u32 $r16, g[$r16]
0005e8: e80fde01 00200780 mad.rn.f32 $r0, s[$ofs2+0x003c], $r15, $r0
0005f0: e80fe005 00204780 mad.rn.f32 $r1, s[$ofs2+0x0040], $r15, $r1
0005f8: e80fe209 00208780 mad.rn.f32 $r2, s[$ofs2+0x0044], $r15, $r2
000600: e80fe40d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0048], $r15, $r3
000608: e80fe611 00210780 mad.rn.f32 $r4, s[$ofs2+0x004c], $r15, $r4
000610: e80fe815 00214780 mad.rn.f32 $r5, s[$ofs2+0x0050], $r15, $r5
000618: e80fea19 00218780 mad.rn.f32 $r6, s[$ofs2+0x0054], $r15, $r6
000620: e80fec1d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0058], $r15, $r7
000628: e80fee21 00220780 mad.rn.f32 $r8, s[$ofs2+0x005c], $r15, $r8
000630: e80ff025 00224780 mad.rn.f32 $r9, s[$ofs2+0x0060], $r15, $r9
000638: e80ff229 00228780 mad.rn.f32 $r10, s[$ofs2+0x0064], $r15, $r10
000640: e80ff42d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0068], $r15, $r11
000648: e80ff631 00230780 mad.rn.f32 $r12, s[$ofs2+0x006c], $r15, $r12
000650: e80ff835 00234780 mad.rn.f32 $r13, s[$ofs2+0x0070], $r15, $r13
000658: e80ffa39 00238780 mad.rn.f32 $r14, s[$ofs2+0x0074], $r15, $r14
000660: e80ffc3d 00244780 mad.rn.f32 $r15, s[$ofs2+0x0078], $r15, $r17
000668: ec10c001 00200780 mad.rn.f32 $r0, s[$ofs3+0x0000], $r16, $r0
000670: ec10c205 00204780 mad.rn.f32 $r1, s[$ofs3+0x0004], $r16, $r1
000678: ec10c409 00208780 mad.rn.f32 $r2, s[$ofs3+0x0008], $r16, $r2
000680: ec10c60d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x000c], $r16, $r3
000688: ec10c811 00210780 mad.rn.f32 $r4, s[$ofs3+0x0010], $r16, $r4
000690: ec10ca15 00214780 mad.rn.f32 $r5, s[$ofs3+0x0014], $r16, $r5
000698: ec10cc19 00218780 mad.rn.f32 $r6, s[$ofs3+0x0018], $r16, $r6
0006a0: ec10ce1d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x001c], $r16, $r7
0006a8: ec10d021 00220780 mad.rn.f32 $r8, s[$ofs3+0x0020], $r16, $r8
0006b0: ec10d225 00224780 mad.rn.f32 $r9, s[$ofs3+0x0024], $r16, $r9
0006b8: ec10d429 00228780 mad.rn.f32 $r10, s[$ofs3+0x0028], $r16, $r10
0006c0: ec10d62d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x002c], $r16, $r11
0006c8: ec10d831 00230780 mad.rn.f32 $r12, s[$ofs3+0x0030], $r16, $r12
0006d0: ec10da35 00234780 mad.rn.f32 $r13, s[$ofs3+0x0034], $r16, $r13
0006d8: ec10dc39 00238780 mad.rn.f32 $r14, s[$ofs3+0x0038], $r16, $r14
0006e0: ec10de45 0023c780 mad.rn.f32 $r17, s[$ofs3+0x003c], $r16, $r15
0006e8: d0060009 20000780 mov.b32 $ofs2, 0x00000300
0006f0: 2000a83d 0ffffa03 add.b32 $r15, $r20, 0xffffa000
0006f8: d00e1e3d 80c00780 mov.u32 $r15, g[$r15]
000700: 2000a841 0ffffb03 add.b32 $r16, $r20, 0xffffb000
000708: d00e2041 80c00780 mov.u32 $r16, g[$r16]
000710: ec0fe201 00200780 mad.rn.f32 $r0, s[$ofs3+0x0044], $r15, $r0
000718: ec0fe405 00204780 mad.rn.f32 $r1, s[$ofs3+0x0048], $r15, $r1
000720: ec0fe609 00208780 mad.rn.f32 $r2, s[$ofs3+0x004c], $r15, $r2
000728: ec0fe80d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0050], $r15, $r3
000730: ec0fea11 00210780 mad.rn.f32 $r4, s[$ofs3+0x0054], $r15, $r4
000738: ec0fec15 00214780 mad.rn.f32 $r5, s[$ofs3+0x0058], $r15, $r5
000740: ec0fee19 00218780 mad.rn.f32 $r6, s[$ofs3+0x005c], $r15, $r6
000748: ec0ff01d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0060], $r15, $r7
000750: ec0ff221 00220780 mad.rn.f32 $r8, s[$ofs3+0x0064], $r15, $r8
000758: ec0ff425 00224780 mad.rn.f32 $r9, s[$ofs3+0x0068], $r15, $r9
000760: ec0ff629 00228780 mad.rn.f32 $r10, s[$ofs3+0x006c], $r15, $r10
000768: ec0ff82d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x0070], $r15, $r11
000770: ec0ffa31 00230780 mad.rn.f32 $r12, s[$ofs3+0x0074], $r15, $r12
000778: ec0ffc35 00234780 mad.rn.f32 $r13, s[$ofs3+0x0078], $r15, $r13
000780: ec0ffe39 00238780 mad.rn.f32 $r14, s[$ofs3+0x007c], $r15, $r14
000788: e80fc03d 00244780 mad.rn.f32 $r15, s[$ofs2+0x0000], $r15, $r17
000790: e810c401 00200780 mad.rn.f32 $r0, s[$ofs2+0x0008], $r16, $r0
000798: e810c605 00204780 mad.rn.f32 $r1, s[$ofs2+0x000c], $r16, $r1
0007a0: e810c809 00208780 mad.rn.f32 $r2, s[$ofs2+0x0010], $r16, $r2
0007a8: e810ca0d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0014], $r16, $r3
0007b0: e810cc11 00210780 mad.rn.f32 $r4, s[$ofs2+0x0018], $r16, $r4
0007b8: e810ce15 00214780 mad.rn.f32 $r5, s[$ofs2+0x001c], $r16, $r5
0007c0: e810d019 00218780 mad.rn.f32 $r6, s[$ofs2+0x0020], $r16, $r6
0007c8: e810d21d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0024], $r16, $r7
0007d0: e810d421 00220780 mad.rn.f32 $r8, s[$ofs2+0x0028], $r16, $r8
0007d8: e810d625 00224780 mad.rn.f32 $r9, s[$ofs2+0x002c], $r16, $r9
0007e0: e810d829 00228780 mad.rn.f32 $r10, s[$ofs2+0x0030], $r16, $r10
0007e8: e810da2d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0034], $r16, $r11
0007f0: e810dc31 00230780 mad.rn.f32 $r12, s[$ofs2+0x0038], $r16, $r12
0007f8: e810de35 00234780 mad.rn.f32 $r13, s[$ofs2+0x003c], $r16, $r13
000800: e810e039 00238780 mad.rn.f32 $r14, s[$ofs2+0x0040], $r16, $r14
000808: e810e245 0023c780 mad.rn.f32 $r17, s[$ofs2+0x0044], $r16, $r15
000810: d007000d 20000780 mov.b32 $ofs3, 0x00000380
000818: 2000a83d 0ffffc03 add.b32 $r15, $r20, 0xffffc000
000820: d00e1e3d 80c00780 mov.u32 $r15, g[$r15]
000828: 2000a841 0ffffd03 add.b32 $r16, $r20, 0xffffd000
000830: d00e2041 80c00780 mov.u32 $r16, g[$r16]
000838: e80fe601 00200780 mad.rn.f32 $r0, s[$ofs2+0x004c], $r15, $r0
000840: e80fe805 00204780 mad.rn.f32 $r1, s[$ofs2+0x0050], $r15, $r1
000848: e80fea09 00208780 mad.rn.f32 $r2, s[$ofs2+0x0054], $r15, $r2
000850: e80fec0d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0058], $r15, $r3
000858: e80fee11 00210780 mad.rn.f32 $r4, s[$ofs2+0x005c], $r15, $r4
000860: e80ff015 00214780 mad.rn.f32 $r5, s[$ofs2+0x0060], $r15, $r5
000868: e80ff219 00218780 mad.rn.f32 $r6, s[$ofs2+0x0064], $r15, $r6
000870: e80ff41d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0068], $r15, $r7
000878: e80ff621 00220780 mad.rn.f32 $r8, s[$ofs2+0x006c], $r15, $r8
000880: e80ff825 00224780 mad.rn.f32 $r9, s[$ofs2+0x0070], $r15, $r9
000888: e80ffa29 00228780 mad.rn.f32 $r10, s[$ofs2+0x0074], $r15, $r10
000890: e80ffc2d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0078], $r15, $r11
000898: e80ffe31 00230780 mad.rn.f32 $r12, s[$ofs2+0x007c], $r15, $r12
0008a0: ec0fc035 00234780 mad.rn.f32 $r13, s[$ofs3+0x0000], $r15, $r13
0008a8: ec0fc239 00238780 mad.rn.f32 $r14, s[$ofs3+0x0004], $r15, $r14
0008b0: ec0fc43d 00244780 mad.rn.f32 $r15, s[$ofs3+0x0008], $r15, $r17
0008b8: ec10c801 00200780 mad.rn.f32 $r0, s[$ofs3+0x0010], $r16, $r0
0008c0: ec10ca05 00204780 mad.rn.f32 $r1, s[$ofs3+0x0014], $r16, $r1
0008c8: ec10cc09 00208780 mad.rn.f32 $r2, s[$ofs3+0x0018], $r16, $r2
0008d0: ec10ce0d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x001c], $r16, $r3
0008d8: ec10d011 00210780 mad.rn.f32 $r4, s[$ofs3+0x0020], $r16, $r4
0008e0: ec10d215 00214780 mad.rn.f32 $r5, s[$ofs3+0x0024], $r16, $r5
0008e8: ec10d419 00218780 mad.rn.f32 $r6, s[$ofs3+0x0028], $r16, $r6
0008f0: ec10d61d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x002c], $r16, $r7
0008f8: ec10d821 00220780 mad.rn.f32 $r8, s[$ofs3+0x0030], $r16, $r8
000900: ec10da25 00224780 mad.rn.f32 $r9, s[$ofs3+0x0034], $r16, $r9
000908: ec10dc29 00228780 mad.rn.f32 $r10, s[$ofs3+0x0038], $r16, $r10
000910: ec10de2d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x003c], $r16, $r11
000918: ec10e031 00230780 mad.rn.f32 $r12, s[$ofs3+0x0040], $r16, $r12
000920: ec10e235 00234780 mad.rn.f32 $r13, s[$ofs3+0x0044], $r16, $r13
000928: ec10e439 00238780 mad.rn.f32 $r14, s[$ofs3+0x0048], $r16, $r14
000930: ec10e641 0023c780 mad.rn.f32 $r16, s[$ofs3+0x004c], $r16, $r15
000938: d0080009 20000780 mov.b32 $ofs2, 0x00000400
000940: 2000a83d 0ffffe03 add.b32 $r15, $r20, 0xffffe000
000948: d00e1e3d 80c00780 mov.u32 $r15, g[$r15]
000950: 2000a845 0fffff03 add.b32 $r17, $r20, 0xfffff000
000958: d00e2245 80c00780 mov.u32 $r17, g[$r17]
000960: ec0fea01 00200780 mad.rn.f32 $r0, s[$ofs3+0x0054], $r15, $r0
000968: ec0fec05 00204780 mad.rn.f32 $r1, s[$ofs3+0x0058], $r15, $r1
000970: ec0fee09 00208780 mad.rn.f32 $r2, s[$ofs3+0x005c], $r15, $r2
000978: ec0ff00d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0060], $r15, $r3
000980: ec0ff211 00210780 mad.rn.f32 $r4, s[$ofs3+0x0064], $r15, $r4
000988: ec0ff415 00214780 mad.rn.f32 $r5, s[$ofs3+0x0068], $r15, $r5
000990: ec0ff619 00218780 mad.rn.f32 $r6, s[$ofs3+0x006c], $r15, $r6
000998: ec0ff81d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0070], $r15, $r7
0009a0: ec0ffa21 00220780 mad.rn.f32 $r8, s[$ofs3+0x0074], $r15, $r8
0009a8: ec0ffc25 00224780 mad.rn.f32 $r9, s[$ofs3+0x0078], $r15, $r9
0009b0: ec0ffe29 00228780 mad.rn.f32 $r10, s[$ofs3+0x007c], $r15, $r10
0009b8: e80fc02d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0000], $r15, $r11
0009c0: e80fc231 00230780 mad.rn.f32 $r12, s[$ofs2+0x0004], $r15, $r12
0009c8: e80fc435 00234780 mad.rn.f32 $r13, s[$ofs2+0x0008], $r15, $r13
0009d0: e80fc639 00238780 mad.rn.f32 $r14, s[$ofs2+0x000c], $r15, $r14
0009d8: e80fc83d 00240780 mad.rn.f32 $r15, s[$ofs2+0x0010], $r15, $r16
0009e0: e811cc01 00200780 mad.rn.f32 $r0, s[$ofs2+0x0018], $r17, $r0
0009e8: e811ce05 00204780 mad.rn.f32 $r1, s[$ofs2+0x001c], $r17, $r1
0009f0: e811d009 00208780 mad.rn.f32 $r2, s[$ofs2+0x0020], $r17, $r2
0009f8: e811d20d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0024], $r17, $r3
000a00: e811d411 00210780 mad.rn.f32 $r4, s[$ofs2+0x0028], $r17, $r4
000a08: e811d615 00214780 mad.rn.f32 $r5, s[$ofs2+0x002c], $r17, $r5
000a10: e811d819 00218780 mad.rn.f32 $r6, s[$ofs2+0x0030], $r17, $r6
000a18: e811da1d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0034], $r17, $r7
000a20: e811dc21 00220780 mad.rn.f32 $r8, s[$ofs2+0x0038], $r17, $r8
000a28: e811de25 00224780 mad.rn.f32 $r9, s[$ofs2+0x003c], $r17, $r9
000a30: e811e029 00228780 mad.rn.f32 $r10, s[$ofs2+0x0040], $r17, $r10
000a38: e811e22d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0044], $r17, $r11
000a40: e811e431 00230780 mad.rn.f32 $r12, s[$ofs2+0x0048], $r17, $r12
000a48: e811e635 00234780 mad.rn.f32 $r13, s[$ofs2+0x004c], $r17, $r13
000a50: e811e839 00238780 mad.rn.f32 $r14, s[$ofs2+0x0050], $r17, $r14
000a58: e811ea41 0023c780 mad.rn.f32 $r16, s[$ofs2+0x0054], $r17, $r15
000a60: 861ffe03 00000000 bar.sync.u32 0x00000000
000a68: 2000a449 00000007 add.b32 $r18, $r18, 0x00000040
000a70: 301325fd 6c0147c8 set.ne.s32 $p0|$o127, $r18, $r19
000a78: 2000aa55 00001003 add.b32 $r21, $r21, 0x00010000
000a80: 2000a851 00001003 add.b32 $r20, $r20, 0x00010000
000a88: 10022003 00000280 @$p0.ne bra.label label0
000a90: 300a2c3d c4100782 label1: shl.join.u32 $r15, $r22, 0x0000000a
000a98: 20179e3c          add.half.b32 $r15, $r15, $r23
000a9c: 20199e3c          add.half.b32 $r15, $r15, $r25
000aa0: 2000303d 0403c780 add.u32 $r15, $r24, $r15
000aa8: 30021e3d c4100780 shl.u32 $r15, $r15, 0x00000002
000ab0: 2000cc3d 0423c780 add.u32 $r15, s[0x0018], $r15
000ab8: d00e1e01 a0c00780 mov.u32 g[$r15], $r0
000ac0: 20009e01 00000103 add.b32 $r0, $r15, 0x00001000
000ac8: d00e0005 a0c00780 mov.u32 g[$r0], $r1
000ad0: 20009e01 00000203 add.b32 $r0, $r15, 0x00002000
000ad8: d00e0009 a0c00780 mov.u32 g[$r0], $r2
000ae0: 20009e01 00000303 add.b32 $r0, $r15, 0x00003000
000ae8: d00e000d a0c00780 mov.u32 g[$r0], $r3
000af0: 20009e01 00000403 add.b32 $r0, $r15, 0x00004000
000af8: d00e0011 a0c00780 mov.u32 g[$r0], $r4
000b00: 20009e01 00000503 add.b32 $r0, $r15, 0x00005000
000b08: d00e0015 a0c00780 mov.u32 g[$r0], $r5
000b10: 20009e01 00000603 add.b32 $r0, $r15, 0x00006000
000b18: d00e0019 a0c00780 mov.u32 g[$r0], $r6
000b20: 20009e01 00000703 add.b32 $r0, $r15, 0x00007000
000b28: 20009e05 00001003 add.b32 $r1, $r15, 0x00010000
000b30: d00e001d a0c00780 mov.u32 g[$r0], $r7
000b38: 20008201 0ffff803 add.b32 $r0, $r1, 0xffff8000
000b40: d00e0021 a0c00780 mov.u32 g[$r0], $r8
000b48: 20008201 0ffff903 add.b32 $r0, $r1, 0xffff9000
000b50: d00e0025 a0c00780 mov.u32 g[$r0], $r9
000b58: 20008201 0ffffa03 add.b32 $r0, $r1, 0xffffa000
000b60: d00e0029 a0c00780 mov.u32 g[$r0], $r10
000b68: 20008201 0ffffb03 add.b32 $r0, $r1, 0xffffb000
000b70: d00e002d a0c00780 mov.u32 g[$r0], $r11
000b78: 20008201 0ffffc03 add.b32 $r0, $r1, 0xffffc000
000b80: d00e0031 a0c00780 mov.u32 g[$r0], $r12
000b88: 20008201 0ffffd03 add.b32 $r0, $r1, 0xffffd000
000b90: d00e0035 a0c00780 mov.u32 g[$r0], $r13
000b98: 20008201 0ffffe03 add.b32 $r0, $r1, 0xffffe000
000ba0: d00e0039 a0c00780 mov.u32 g[$r0], $r14
000ba8: 20008201 0fffff03 add.b32 $r0, $r1, 0xfffff000
000bb0: d00e0041 a0c00781 mov.end.u32 g[$r0], $r16
// segment: const (1:0000)
0000: 00000011 000003ff 
// Disassembling _Z27matrix_large_tile_optimizedPfS_S_ (2)
000000: d0800205 00400780 and.b16 $r0.hi, $r0.hi, c1[0x0000]
000008: a0000065 04000780 cvt.u32.u16 $r25, $r0.lo
000010: a0000245 04000780 cvt.u32.u16 $r17, $r0.hi
000018: 60813201 60444780 mad24.lo.u32 $r0, $r25, c1[0x0004], $r17
000020: 00020005 c0000780 movsh.b32 $ofs1, $r0, 0x00000002
000028: a015b003 00000000 join.label label1
000030: 41002c6d 00000013 mul24.lo.u32.u16.u16 $r27, s[0x000c], 0x0100
000038: 41102e71 00000003 mul24.lo.u32.u16.u16 $r28, s[0x000e], 0x0010
000040: 10008055 00000003 mov.b32 $r21, 0x00000000
000048: 10008039 00000003 mov.b32 $r14, 0x00000000
000050: 10008035 00000003 mov.b32 $r13, 0x00000000
000058: 10008031 00000003 mov.b32 $r12, 0x00000000
000060: 1000802d 00000003 mov.b32 $r11, 0x00000000
000068: 10008029 00000003 mov.b32 $r10, 0x00000000
000070: 10008025 00000003 mov.b32 $r9, 0x00000000
000078: 10008021 00000003 mov.b32 $r8, 0x00000000
000080: 1000801d 00000003 mov.b32 $r7, 0x00000000
000088: 10008019 00000003 mov.b32 $r6, 0x00000000
000090: 10008015 00000003 mov.b32 $r5, 0x00000000
000098: 10008011 00000003 mov.b32 $r4, 0x00000000
0000a0: 1000800d 00000003 mov.b32 $r3, 0x00000000
0000a8: 10008009 00000003 mov.b32 $r2, 0x00000000
0000b0: 10008005 00000003 mov.b32 $r1, 0x00000000
0000b8: 10008001 00000003 mov.b32 $r0, 0x00000000
0000c0: 201bb23c          add.half.b32 $r15, $r25, $r27
0000c4: 201ca240          add.half.b32 $r16, $r17, $r28
0000c8: 30042269 c4100780 shl.u32 $r26, $r17, 0x00000004
0000d0: 300a2041 c4100780 shl.u32 $r16, $r16, 0x0000000a
0000d8: 200fb43c          add.half.b32 $r15, $r26, $r15
0000dc: 2010b240          add.half.b32 $r16, $r25, $r16
0000e0: 30021e3d c4100780 shl.u32 $r15, $r15, 0x00000002
0000e8: 30022045 c4100780 shl.u32 $r17, $r16, 0x00000002
0000f0: 2000ca41 0423c780 add.u32 $r16, s[0x0014], $r15
0000f8: d00e203d 80c00780 mov.u32 $r15, g[$r16]
000100: 2000c861 04244780 add.u32 $r24, s[0x0010], $r17
000108: d00e3059 80c00780 mov.u32 $r22, g[$r24]
000110: 2000a045 00000103 add.b32 $r17, $r16, 0x00001000
000118: d00e224d 80c00780 mov.u32 $r19, g[$r17]
000120: 2000a045 00000203 add.b32 $r17, $r16, 0x00002000
000128: d00e2249 80c00780 mov.u32 $r18, g[$r17]
000130: 2000a045 00000303 add.b32 $r17, $r16, 0x00003000
000138: d00e2245 80c00780 mov.u32 $r17, g[$r17]
000140: 2000a051 00001003 add.b32 $r20, $r16, 0x00010000
000148: 2000b05d 00000007 add.b32 $r23, $r24, 0x00000040
000150: 2000b061 00000103 add.b32 $r24, $r24, 0x00001000
000158: 04000e01 e4258780 label0: mov.b32 s[$ofs1+0x001c], $r22
000160: 861ffe03 00000000 bar.sync.u32 0x00000000
000168: e00fd039 00238780 mad.rn.f32 $r14, s[0x0020], $r15, $r14
000170: e00fd235 00234780 mad.rn.f32 $r13, s[0x0024], $r15, $r13
000178: e00fd431 00230780 mad.rn.f32 $r12, s[0x0028], $r15, $r12
000180: e00fd62d 0022c780 mad.rn.f32 $r11, s[0x002c], $r15, $r11
000188: e00fd829 00228780 mad.rn.f32 $r10, s[0x0030], $r15, $r10
000190: e00fda25 00224780 mad.rn.f32 $r9, s[0x0034], $r15, $r9
000198: e00fdc21 00220780 mad.rn.f32 $r8, s[0x0038], $r15, $r8
0001a0: e00fde1d 0021c780 mad.rn.f32 $r7, s[0x003c], $r15, $r7
0001a8: e00fe019 00218780 mad.rn.f32 $r6, s[0x0040], $r15, $r6
0001b0: e00fe215 00214780 mad.rn.f32 $r5, s[0x0044], $r15, $r5
0001b8: e00fe411 00210780 mad.rn.f32 $r4, s[0x0048], $r15, $r4
0001c0: e00fe60d 0020c780 mad.rn.f32 $r3, s[0x004c], $r15, $r3
0001c8: e00fe809 00208780 mad.rn.f32 $r2, s[0x0050], $r15, $r2
0001d0: e00fea05 00204780 mad.rn.f32 $r1, s[0x0054], $r15, $r1
0001d8: e00fec01 00200780 mad.rn.f32 $r0, s[0x0058], $r15, $r0
0001e0: e00fce3d 00254780 mad.rn.f32 $r15, s[0x001c], $r15, $r21
0001e8: d001000d 20000780 mov.b32 $ofs3, 0x00000080
0001f0: e013f239 00238780 mad.rn.f32 $r14, s[0x0064], $r19, $r14
0001f8: e013f435 00234780 mad.rn.f32 $r13, s[0x0068], $r19, $r13
000200: e013f631 00230780 mad.rn.f32 $r12, s[0x006c], $r19, $r12
000208: e013f82d 0022c780 mad.rn.f32 $r11, s[0x0070], $r19, $r11
000210: e013fa29 00228780 mad.rn.f32 $r10, s[0x0074], $r19, $r10
000218: e013fc25 00224780 mad.rn.f32 $r9, s[0x0078], $r19, $r9
000220: e013fe21 00220780 mad.rn.f32 $r8, s[0x007c], $r19, $r8
000228: ec13c01d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0000], $r19, $r7
000230: ec13c219 00218780 mad.rn.f32 $r6, s[$ofs3+0x0004], $r19, $r6
000238: ec13c415 00214780 mad.rn.f32 $r5, s[$ofs3+0x0008], $r19, $r5
000240: ec13c611 00210780 mad.rn.f32 $r4, s[$ofs3+0x000c], $r19, $r4
000248: ec13c80d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0010], $r19, $r3
000250: ec13ca09 00208780 mad.rn.f32 $r2, s[$ofs3+0x0014], $r19, $r2
000258: ec13cc05 00204780 mad.rn.f32 $r1, s[$ofs3+0x0018], $r19, $r1
000260: ec13ce01 00200780 mad.rn.f32 $r0, s[$ofs3+0x001c], $r19, $r0
000268: e013f03d 0023c780 mad.rn.f32 $r15, s[0x0060], $r19, $r15
000270: ec12d439 00238780 mad.rn.f32 $r14, s[$ofs3+0x0028], $r18, $r14
000278: ec12d635 00234780 mad.rn.f32 $r13, s[$ofs3+0x002c], $r18, $r13
000280: ec12d831 00230780 mad.rn.f32 $r12, s[$ofs3+0x0030], $r18, $r12
000288: ec12da2d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x0034], $r18, $r11
000290: ec12dc29 00228780 mad.rn.f32 $r10, s[$ofs3+0x0038], $r18, $r10
000298: ec12de25 00224780 mad.rn.f32 $r9, s[$ofs3+0x003c], $r18, $r9
0002a0: ec12e021 00220780 mad.rn.f32 $r8, s[$ofs3+0x0040], $r18, $r8
0002a8: ec12e21d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0044], $r18, $r7
0002b0: ec12e419 00218780 mad.rn.f32 $r6, s[$ofs3+0x0048], $r18, $r6
0002b8: ec12e615 00214780 mad.rn.f32 $r5, s[$ofs3+0x004c], $r18, $r5
0002c0: ec12e811 00210780 mad.rn.f32 $r4, s[$ofs3+0x0050], $r18, $r4
0002c8: ec12ea0d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0054], $r18, $r3
0002d0: ec12ec09 00208780 mad.rn.f32 $r2, s[$ofs3+0x0058], $r18, $r2
0002d8: ec12ee05 00204780 mad.rn.f32 $r1, s[$ofs3+0x005c], $r18, $r1
0002e0: ec12f001 00200780 mad.rn.f32 $r0, s[$ofs3+0x0060], $r18, $r0
0002e8: ec12d249 0023c780 mad.rn.f32 $r18, s[$ofs3+0x0024], $r18, $r15
0002f0: 2000a03d 00000403 add.b32 $r15, $r16, 0x00004000
0002f8: d00e1e3d 80c00780 mov.u32 $r15, g[$r15]
000300: d0020009 20000780 mov.b32 $ofs2, 0x00000100
000308: ec11f639 00238780 mad.rn.f32 $r14, s[$ofs3+0x006c], $r17, $r14
000310: ec11f835 00234780 mad.rn.f32 $r13, s[$ofs3+0x0070], $r17, $r13
000318: ec11fa31 00230780 mad.rn.f32 $r12, s[$ofs3+0x0074], $r17, $r12
000320: ec11fc2d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x0078], $r17, $r11
000328: ec11fe29 00228780 mad.rn.f32 $r10, s[$ofs3+0x007c], $r17, $r10
000330: e811c025 00224780 mad.rn.f32 $r9, s[$ofs2+0x0000], $r17, $r9
000338: e811c221 00220780 mad.rn.f32 $r8, s[$ofs2+0x0004], $r17, $r8
000340: e811c41d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0008], $r17, $r7
000348: e811c619 00218780 mad.rn.f32 $r6, s[$ofs2+0x000c], $r17, $r6
000350: e811c815 00214780 mad.rn.f32 $r5, s[$ofs2+0x0010], $r17, $r5
000358: e811ca11 00210780 mad.rn.f32 $r4, s[$ofs2+0x0014], $r17, $r4
000360: e811cc0d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0018], $r17, $r3
000368: e811ce09 00208780 mad.rn.f32 $r2, s[$ofs2+0x001c], $r17, $r2
000370: e811d005 00204780 mad.rn.f32 $r1, s[$ofs2+0x0020], $r17, $r1
000378: e811d201 00200780 mad.rn.f32 $r0, s[$ofs2+0x0024], $r17, $r0
000380: ec11f449 00248780 mad.rn.f32 $r18, s[$ofs3+0x0068], $r17, $r18
000388: 2000a045 00000503 add.b32 $r17, $r16, 0x00005000
000390: d00e2245 80c00780 mov.u32 $r17, g[$r17]
000398: e80fd649 00248780 mad.rn.f32 $r18, s[$ofs2+0x002c], $r15, $r18
0003a0: e80fd839 00238780 mad.rn.f32 $r14, s[$ofs2+0x0030], $r15, $r14
0003a8: e80fda35 00234780 mad.rn.f32 $r13, s[$ofs2+0x0034], $r15, $r13
0003b0: e80fdc31 00230780 mad.rn.f32 $r12, s[$ofs2+0x0038], $r15, $r12
0003b8: e80fde2d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x003c], $r15, $r11
0003c0: e80fe029 00228780 mad.rn.f32 $r10, s[$ofs2+0x0040], $r15, $r10
0003c8: e80fe225 00224780 mad.rn.f32 $r9, s[$ofs2+0x0044], $r15, $r9
0003d0: e80fe421 00220780 mad.rn.f32 $r8, s[$ofs2+0x0048], $r15, $r8
0003d8: e80fe61d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x004c], $r15, $r7
0003e0: e80fe819 00218780 mad.rn.f32 $r6, s[$ofs2+0x0050], $r15, $r6
0003e8: e80fea15 00214780 mad.rn.f32 $r5, s[$ofs2+0x0054], $r15, $r5
0003f0: e80fec11 00210780 mad.rn.f32 $r4, s[$ofs2+0x0058], $r15, $r4
0003f8: e80fee0d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x005c], $r15, $r3
000400: e80ff009 00208780 mad.rn.f32 $r2, s[$ofs2+0x0060], $r15, $r2
000408: e80ff205 00204780 mad.rn.f32 $r1, s[$ofs2+0x0064], $r15, $r1
000410: e80ff43d 00200780 mad.rn.f32 $r15, s[$ofs2+0x0068], $r15, $r0
000418: d003000d 20000780 mov.b32 $ofs3, 0x00000180
000420: 2000a001 00000603 add.b32 $r0, $r16, 0x00006000
000428: d00e0001 80c00780 mov.u32 $r0, g[$r0]
000430: e811f849 00248780 mad.rn.f32 $r18, s[$ofs2+0x0070], $r17, $r18
000438: e811fa39 00238780 mad.rn.f32 $r14, s[$ofs2+0x0074], $r17, $r14
000440: e811fc35 00234780 mad.rn.f32 $r13, s[$ofs2+0x0078], $r17, $r13
000448: e811fe31 00230780 mad.rn.f32 $r12, s[$ofs2+0x007c], $r17, $r12
000450: ec11c02d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x0000], $r17, $r11
000458: ec11c229 00228780 mad.rn.f32 $r10, s[$ofs3+0x0004], $r17, $r10
000460: ec11c425 00224780 mad.rn.f32 $r9, s[$ofs3+0x0008], $r17, $r9
000468: ec11c621 00220780 mad.rn.f32 $r8, s[$ofs3+0x000c], $r17, $r8
000470: ec11c81d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0010], $r17, $r7
000478: ec11ca19 00218780 mad.rn.f32 $r6, s[$ofs3+0x0014], $r17, $r6
000480: ec11cc15 00214780 mad.rn.f32 $r5, s[$ofs3+0x0018], $r17, $r5
000488: ec11ce11 00210780 mad.rn.f32 $r4, s[$ofs3+0x001c], $r17, $r4
000490: ec11d00d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0020], $r17, $r3
000498: ec11d209 00208780 mad.rn.f32 $r2, s[$ofs3+0x0024], $r17, $r2
0004a0: ec11d405 00204780 mad.rn.f32 $r1, s[$ofs3+0x0028], $r17, $r1
0004a8: ec11d645 0023c780 mad.rn.f32 $r17, s[$ofs3+0x002c], $r17, $r15
0004b0: 2000a03d 00000703 add.b32 $r15, $r16, 0x00007000
0004b8: d00e1e3d 80c00780 mov.u32 $r15, g[$r15]
0004c0: ec00da49 00248780 mad.rn.f32 $r18, s[$ofs3+0x0034], $r0, $r18
0004c8: ec00dc39 00238780 mad.rn.f32 $r14, s[$ofs3+0x0038], $r0, $r14
0004d0: ec00de35 00234780 mad.rn.f32 $r13, s[$ofs3+0x003c], $r0, $r13
0004d8: ec00e031 00230780 mad.rn.f32 $r12, s[$ofs3+0x0040], $r0, $r12
0004e0: ec00e22d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x0044], $r0, $r11
0004e8: ec00e429 00228780 mad.rn.f32 $r10, s[$ofs3+0x0048], $r0, $r10
0004f0: ec00e625 00224780 mad.rn.f32 $r9, s[$ofs3+0x004c], $r0, $r9
0004f8: ec00e821 00220780 mad.rn.f32 $r8, s[$ofs3+0x0050], $r0, $r8
000500: ec00ea1d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0054], $r0, $r7
000508: ec00ec19 00218780 mad.rn.f32 $r6, s[$ofs3+0x0058], $r0, $r6
000510: ec00ee15 00214780 mad.rn.f32 $r5, s[$ofs3+0x005c], $r0, $r5
000518: ec00f011 00210780 mad.rn.f32 $r4, s[$ofs3+0x0060], $r0, $r4
000520: ec00f20d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0064], $r0, $r3
000528: ec00f409 00208780 mad.rn.f32 $r2, s[$ofs3+0x0068], $r0, $r2
000530: ec00f605 00204780 mad.rn.f32 $r1, s[$ofs3+0x006c], $r0, $r1
000538: ec00f801 00244780 mad.rn.f32 $r0, s[$ofs3+0x0070], $r0, $r17
000540: d0040009 20000780 mov.b32 $ofs2, 0x00000200
000548: ec0ffc45 00248780 mad.rn.f32 $r17, s[$ofs3+0x0078], $r15, $r18
000550: ec0ffe39 00238780 mad.rn.f32 $r14, s[$ofs3+0x007c], $r15, $r14
000558: e80fc035 00234780 mad.rn.f32 $r13, s[$ofs2+0x0000], $r15, $r13
000560: e80fc231 00230780 mad.rn.f32 $r12, s[$ofs2+0x0004], $r15, $r12
000568: e80fc42d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0008], $r15, $r11
000570: e80fc629 00228780 mad.rn.f32 $r10, s[$ofs2+0x000c], $r15, $r10
000578: e80fc825 00224780 mad.rn.f32 $r9, s[$ofs2+0x0010], $r15, $r9
000580: e80fca21 00220780 mad.rn.f32 $r8, s[$ofs2+0x0014], $r15, $r8
000588: e80fcc1d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0018], $r15, $r7
000590: e80fce19 00218780 mad.rn.f32 $r6, s[$ofs2+0x001c], $r15, $r6
000598: e80fd015 00214780 mad.rn.f32 $r5, s[$ofs2+0x0020], $r15, $r5
0005a0: e80fd211 00210780 mad.rn.f32 $r4, s[$ofs2+0x0024], $r15, $r4
0005a8: e80fd40d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0028], $r15, $r3
0005b0: e80fd609 00208780 mad.rn.f32 $r2, s[$ofs2+0x002c], $r15, $r2
0005b8: e80fd805 00204780 mad.rn.f32 $r1, s[$ofs2+0x0030], $r15, $r1
0005c0: e80fda3d 00200780 mad.rn.f32 $r15, s[$ofs2+0x0034], $r15, $r0
0005c8: d005000d 20000780 mov.b32 $ofs3, 0x00000280
0005d0: 2000a801 0ffff803 add.b32 $r0, $r20, 0xffff8000
0005d8: d00e0049 80c00780 mov.u32 $r18, g[$r0]
0005e0: 2000a801 0ffff903 add.b32 $r0, $r20, 0xffff9000
0005e8: d00e0001 80c00780 mov.u32 $r0, g[$r0]
0005f0: e812de45 00244780 mad.rn.f32 $r17, s[$ofs2+0x003c], $r18, $r17
0005f8: e812e039 00238780 mad.rn.f32 $r14, s[$ofs2+0x0040], $r18, $r14
000600: e812e235 00234780 mad.rn.f32 $r13, s[$ofs2+0x0044], $r18, $r13
000608: e812e431 00230780 mad.rn.f32 $r12, s[$ofs2+0x0048], $r18, $r12
000610: e812e62d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x004c], $r18, $r11
000618: e812e829 00228780 mad.rn.f32 $r10, s[$ofs2+0x0050], $r18, $r10
000620: e812ea25 00224780 mad.rn.f32 $r9, s[$ofs2+0x0054], $r18, $r9
000628: e812ec21 00220780 mad.rn.f32 $r8, s[$ofs2+0x0058], $r18, $r8
000630: e812ee1d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x005c], $r18, $r7
000638: e812f019 00218780 mad.rn.f32 $r6, s[$ofs2+0x0060], $r18, $r6
000640: e812f215 00214780 mad.rn.f32 $r5, s[$ofs2+0x0064], $r18, $r5
000648: e812f411 00210780 mad.rn.f32 $r4, s[$ofs2+0x0068], $r18, $r4
000650: e812f60d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x006c], $r18, $r3
000658: e812f809 00208780 mad.rn.f32 $r2, s[$ofs2+0x0070], $r18, $r2
000660: e812fa05 00204780 mad.rn.f32 $r1, s[$ofs2+0x0074], $r18, $r1
000668: e812fc3d 0023c780 mad.rn.f32 $r15, s[$ofs2+0x0078], $r18, $r15
000670: ec00c045 00244780 mad.rn.f32 $r17, s[$ofs3+0x0000], $r0, $r17
000678: ec00c239 00238780 mad.rn.f32 $r14, s[$ofs3+0x0004], $r0, $r14
000680: ec00c435 00234780 mad.rn.f32 $r13, s[$ofs3+0x0008], $r0, $r13
000688: ec00c631 00230780 mad.rn.f32 $r12, s[$ofs3+0x000c], $r0, $r12
000690: ec00c82d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x0010], $r0, $r11
000698: ec00ca29 00228780 mad.rn.f32 $r10, s[$ofs3+0x0014], $r0, $r10
0006a0: ec00cc25 00224780 mad.rn.f32 $r9, s[$ofs3+0x0018], $r0, $r9
0006a8: ec00ce21 00220780 mad.rn.f32 $r8, s[$ofs3+0x001c], $r0, $r8
0006b0: ec00d01d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0020], $r0, $r7
0006b8: ec00d219 00218780 mad.rn.f32 $r6, s[$ofs3+0x0024], $r0, $r6
0006c0: ec00d415 00214780 mad.rn.f32 $r5, s[$ofs3+0x0028], $r0, $r5
0006c8: ec00d611 00210780 mad.rn.f32 $r4, s[$ofs3+0x002c], $r0, $r4
0006d0: ec00d80d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0030], $r0, $r3
0006d8: ec00da09 00208780 mad.rn.f32 $r2, s[$ofs3+0x0034], $r0, $r2
0006e0: ec00dc05 00204780 mad.rn.f32 $r1, s[$ofs3+0x0038], $r0, $r1
0006e8: ec00de3d 0023c780 mad.rn.f32 $r15, s[$ofs3+0x003c], $r0, $r15
0006f0: d0060009 20000780 mov.b32 $ofs2, 0x00000300
0006f8: 2000a801 0ffffa03 add.b32 $r0, $r20, 0xffffa000
000700: d00e0049 80c00780 mov.u32 $r18, g[$r0]
000708: 2000a801 0ffffb03 add.b32 $r0, $r20, 0xffffb000
000710: d00e0001 80c00780 mov.u32 $r0, g[$r0]
000718: ec12e245 00244780 mad.rn.f32 $r17, s[$ofs3+0x0044], $r18, $r17
000720: ec12e439 00238780 mad.rn.f32 $r14, s[$ofs3+0x0048], $r18, $r14
000728: ec12e635 00234780 mad.rn.f32 $r13, s[$ofs3+0x004c], $r18, $r13
000730: ec12e831 00230780 mad.rn.f32 $r12, s[$ofs3+0x0050], $r18, $r12
000738: ec12ea2d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x0054], $r18, $r11
000740: ec12ec29 00228780 mad.rn.f32 $r10, s[$ofs3+0x0058], $r18, $r10
000748: ec12ee25 00224780 mad.rn.f32 $r9, s[$ofs3+0x005c], $r18, $r9
000750: ec12f021 00220780 mad.rn.f32 $r8, s[$ofs3+0x0060], $r18, $r8
000758: ec12f21d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0064], $r18, $r7
000760: ec12f419 00218780 mad.rn.f32 $r6, s[$ofs3+0x0068], $r18, $r6
000768: ec12f615 00214780 mad.rn.f32 $r5, s[$ofs3+0x006c], $r18, $r5
000770: ec12f811 00210780 mad.rn.f32 $r4, s[$ofs3+0x0070], $r18, $r4
000778: ec12fa0d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0074], $r18, $r3
000780: ec12fc09 00208780 mad.rn.f32 $r2, s[$ofs3+0x0078], $r18, $r2
000788: ec12fe05 00204780 mad.rn.f32 $r1, s[$ofs3+0x007c], $r18, $r1
000790: e812c03d 0023c780 mad.rn.f32 $r15, s[$ofs2+0x0000], $r18, $r15
000798: e800c445 00244780 mad.rn.f32 $r17, s[$ofs2+0x0008], $r0, $r17
0007a0: e800c639 00238780 mad.rn.f32 $r14, s[$ofs2+0x000c], $r0, $r14
0007a8: e800c835 00234780 mad.rn.f32 $r13, s[$ofs2+0x0010], $r0, $r13
0007b0: e800ca31 00230780 mad.rn.f32 $r12, s[$ofs2+0x0014], $r0, $r12
0007b8: e800cc2d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0018], $r0, $r11
0007c0: e800ce29 00228780 mad.rn.f32 $r10, s[$ofs2+0x001c], $r0, $r10
0007c8: e800d025 00224780 mad.rn.f32 $r9, s[$ofs2+0x0020], $r0, $r9
0007d0: e800d221 00220780 mad.rn.f32 $r8, s[$ofs2+0x0024], $r0, $r8
0007d8: e800d41d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0028], $r0, $r7
0007e0: e800d619 00218780 mad.rn.f32 $r6, s[$ofs2+0x002c], $r0, $r6
0007e8: e800d815 00214780 mad.rn.f32 $r5, s[$ofs2+0x0030], $r0, $r5
0007f0: e800da11 00210780 mad.rn.f32 $r4, s[$ofs2+0x0034], $r0, $r4
0007f8: e800dc0d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0038], $r0, $r3
000800: e800de09 00208780 mad.rn.f32 $r2, s[$ofs2+0x003c], $r0, $r2
000808: e800e005 00204780 mad.rn.f32 $r1, s[$ofs2+0x0040], $r0, $r1
000810: e800e201 0023c780 mad.rn.f32 $r0, s[$ofs2+0x0044], $r0, $r15
000818: d007000d 20000780 mov.b32 $ofs3, 0x00000380
000820: 2000a83d 0ffffc03 add.b32 $r15, $r20, 0xffffc000
000828: d00e1e3d 80c00780 mov.u32 $r15, g[$r15]
000830: 2000a849 0ffffd03 add.b32 $r18, $r20, 0xffffd000
000838: d00e2449 80c00780 mov.u32 $r18, g[$r18]
000840: e80fe645 00244780 mad.rn.f32 $r17, s[$ofs2+0x004c], $r15, $r17
000848: e80fe839 00238780 mad.rn.f32 $r14, s[$ofs2+0x0050], $r15, $r14
000850: e80fea35 00234780 mad.rn.f32 $r13, s[$ofs2+0x0054], $r15, $r13
000858: e80fec31 00230780 mad.rn.f32 $r12, s[$ofs2+0x0058], $r15, $r12
000860: e80fee2d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x005c], $r15, $r11
000868: e80ff029 00228780 mad.rn.f32 $r10, s[$ofs2+0x0060], $r15, $r10
000870: e80ff225 00224780 mad.rn.f32 $r9, s[$ofs2+0x0064], $r15, $r9
000878: e80ff421 00220780 mad.rn.f32 $r8, s[$ofs2+0x0068], $r15, $r8
000880: e80ff61d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x006c], $r15, $r7
000888: e80ff819 00218780 mad.rn.f32 $r6, s[$ofs2+0x0070], $r15, $r6
000890: e80ffa15 00214780 mad.rn.f32 $r5, s[$ofs2+0x0074], $r15, $r5
000898: e80ffc11 00210780 mad.rn.f32 $r4, s[$ofs2+0x0078], $r15, $r4
0008a0: e80ffe0d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x007c], $r15, $r3
0008a8: ec0fc009 00208780 mad.rn.f32 $r2, s[$ofs3+0x0000], $r15, $r2
0008b0: ec0fc205 00204780 mad.rn.f32 $r1, s[$ofs3+0x0004], $r15, $r1
0008b8: ec0fc401 00200780 mad.rn.f32 $r0, s[$ofs3+0x0008], $r15, $r0
0008c0: ec12c83d 00244780 mad.rn.f32 $r15, s[$ofs3+0x0010], $r18, $r17
0008c8: ec12ca39 00238780 mad.rn.f32 $r14, s[$ofs3+0x0014], $r18, $r14
0008d0: ec12cc35 00234780 mad.rn.f32 $r13, s[$ofs3+0x0018], $r18, $r13
0008d8: ec12ce31 00230780 mad.rn.f32 $r12, s[$ofs3+0x001c], $r18, $r12
0008e0: ec12d02d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x0020], $r18, $r11
0008e8: ec12d229 00228780 mad.rn.f32 $r10, s[$ofs3+0x0024], $r18, $r10
0008f0: ec12d425 00224780 mad.rn.f32 $r9, s[$ofs3+0x0028], $r18, $r9
0008f8: ec12d621 00220780 mad.rn.f32 $r8, s[$ofs3+0x002c], $r18, $r8
000900: ec12d81d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0030], $r18, $r7
000908: ec12da19 00218780 mad.rn.f32 $r6, s[$ofs3+0x0034], $r18, $r6
000910: ec12dc15 00214780 mad.rn.f32 $r5, s[$ofs3+0x0038], $r18, $r5
000918: ec12de11 00210780 mad.rn.f32 $r4, s[$ofs3+0x003c], $r18, $r4
000920: ec12e00d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0040], $r18, $r3
000928: ec12e209 00208780 mad.rn.f32 $r2, s[$ofs3+0x0044], $r18, $r2
000930: ec12e405 00204780 mad.rn.f32 $r1, s[$ofs3+0x0048], $r18, $r1
000938: ec12e645 00200780 mad.rn.f32 $r17, s[$ofs3+0x004c], $r18, $r0
000940: 2000a801 0ffffe03 add.b32 $r0, $r20, 0xffffe000
000948: d00e0049 80c00780 mov.u32 $r18, g[$r0]
000950: d0080009 20000780 mov.b32 $ofs2, 0x00000400
000958: 2000a801 0fffff03 add.b32 $r0, $r20, 0xfffff000
000960: d00e0001 80c00780 mov.u32 $r0, g[$r0]
000968: d00e2e59 80c00780 mov.u32 $r22, g[$r23]
000970: ec12ea3d 0023c780 mad.rn.f32 $r15, s[$ofs3+0x0054], $r18, $r15
000978: ec12ec39 00238780 mad.rn.f32 $r14, s[$ofs3+0x0058], $r18, $r14
000980: ec12ee35 00234780 mad.rn.f32 $r13, s[$ofs3+0x005c], $r18, $r13
000988: ec12f031 00230780 mad.rn.f32 $r12, s[$ofs3+0x0060], $r18, $r12
000990: ec12f22d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x0064], $r18, $r11
000998: ec12f429 00228780 mad.rn.f32 $r10, s[$ofs3+0x0068], $r18, $r10
0009a0: ec12f625 00224780 mad.rn.f32 $r9, s[$ofs3+0x006c], $r18, $r9
0009a8: ec12f821 00220780 mad.rn.f32 $r8, s[$ofs3+0x0070], $r18, $r8
0009b0: ec12fa1d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0074], $r18, $r7
0009b8: ec12fc19 00218780 mad.rn.f32 $r6, s[$ofs3+0x0078], $r18, $r6
0009c0: ec12fe15 00214780 mad.rn.f32 $r5, s[$ofs3+0x007c], $r18, $r5
0009c8: e812c011 00210780 mad.rn.f32 $r4, s[$ofs2+0x0000], $r18, $r4
0009d0: e812c20d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0004], $r18, $r3
0009d8: e812c409 00208780 mad.rn.f32 $r2, s[$ofs2+0x0008], $r18, $r2
0009e0: e812c605 00204780 mad.rn.f32 $r1, s[$ofs2+0x000c], $r18, $r1
0009e8: e812c845 00244780 mad.rn.f32 $r17, s[$ofs2+0x0010], $r18, $r17
0009f0: e800cc55 0023c780 mad.rn.f32 $r21, s[$ofs2+0x0018], $r0, $r15
0009f8: e800ce39 00238780 mad.rn.f32 $r14, s[$ofs2+0x001c], $r0, $r14
000a00: e800d035 00234780 mad.rn.f32 $r13, s[$ofs2+0x0020], $r0, $r13
000a08: e800d231 00230780 mad.rn.f32 $r12, s[$ofs2+0x0024], $r0, $r12
000a10: e800d42d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0028], $r0, $r11
000a18: e800d629 00228780 mad.rn.f32 $r10, s[$ofs2+0x002c], $r0, $r10
000a20: e800d825 00224780 mad.rn.f32 $r9, s[$ofs2+0x0030], $r0, $r9
000a28: e800da21 00220780 mad.rn.f32 $r8, s[$ofs2+0x0034], $r0, $r8
000a30: e800dc1d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0038], $r0, $r7
000a38: e800de19 00218780 mad.rn.f32 $r6, s[$ofs2+0x003c], $r0, $r6
000a40: e800e015 00214780 mad.rn.f32 $r5, s[$ofs2+0x0040], $r0, $r5
000a48: e800e211 00210780 mad.rn.f32 $r4, s[$ofs2+0x0044], $r0, $r4
000a50: e800e40d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0048], $r0, $r3
000a58: e800e609 00208780 mad.rn.f32 $r2, s[$ofs2+0x004c], $r0, $r2
000a60: e800e805 00204780 mad.rn.f32 $r1, s[$ofs2+0x0050], $r0, $r1
000a68: e800ea01 00244780 mad.rn.f32 $r0, s[$ofs2+0x0054], $r0, $r17
000a70: 2000ae5d 00000007 add.b32 $r23, $r23, 0x00000040
000a78: 2000a041 00001003 add.b32 $r16, $r16, 0x00010000
000a80: d00e203d 80c00780 mov.u32 $r15, g[$r16]
000a88: 2000a851 00001003 add.b32 $r20, $r20, 0x00010000
000a90: 2000a045 00000103 add.b32 $r17, $r16, 0x00001000
000a98: d00e224d 80c00780 mov.u32 $r19, g[$r17]
000aa0: 2000a045 00000203 add.b32 $r17, $r16, 0x00002000
000aa8: d00e2249 80c00780 mov.u32 $r18, g[$r17]
000ab0: 2000a045 00000303 add.b32 $r17, $r16, 0x00003000
000ab8: d00e2245 80c00780 mov.u32 $r17, g[$r17]
000ac0: 861ffe03 00000000 bar.sync.u32 0x00000000
000ac8: 301731fd 640107c8 set.gt.u32 $p0|$o127, $r24, $r23
000ad0: 1002b003 00000280 @$p0.ne bra.label label0
000ad8: 04000e01 e4258782 label1: mov.join.b32 s[$ofs1+0x001c], $r22
000ae0: 861ffe03 00000000 bar.sync.u32 0x00000000
000ae8: 300a3851 c4100780 shl.u32 $r20, $r28, 0x0000000a
000af0: 201ba850          add.half.b32 $r20, $r20, $r27
000af4: 2019a850          add.half.b32 $r20, $r20, $r25
000af8: 20003451 04050780 add.u32 $r20, $r26, $r20
000b00: e00fd039 00238780 mad.rn.f32 $r14, s[0x0020], $r15, $r14
000b08: e00fd235 00234780 mad.rn.f32 $r13, s[0x0024], $r15, $r13
000b10: e00fd431 00230780 mad.rn.f32 $r12, s[0x0028], $r15, $r12
000b18: e00fd62d 0022c780 mad.rn.f32 $r11, s[0x002c], $r15, $r11
000b20: e00fd829 00228780 mad.rn.f32 $r10, s[0x0030], $r15, $r10
000b28: e00fda25 00224780 mad.rn.f32 $r9, s[0x0034], $r15, $r9
000b30: e00fdc21 00220780 mad.rn.f32 $r8, s[0x0038], $r15, $r8
000b38: e00fde1d 0021c780 mad.rn.f32 $r7, s[0x003c], $r15, $r7
000b40: e00fe019 00218780 mad.rn.f32 $r6, s[0x0040], $r15, $r6
000b48: e00fe215 00214780 mad.rn.f32 $r5, s[0x0044], $r15, $r5
000b50: e00fe411 00210780 mad.rn.f32 $r4, s[0x0048], $r15, $r4
000b58: e00fe60d 0020c780 mad.rn.f32 $r3, s[0x004c], $r15, $r3
000b60: e00fe809 00208780 mad.rn.f32 $r2, s[0x0050], $r15, $r2
000b68: e00fea05 00204780 mad.rn.f32 $r1, s[0x0054], $r15, $r1
000b70: e00fce55 00254780 mad.rn.f32 $r21, s[0x001c], $r15, $r21
000b78: e00fec01 00200780 mad.rn.f32 $r0, s[0x0058], $r15, $r0
000b80: d0010009 20000780 mov.b32 $ofs2, 0x00000080
000b88: e013f239 00238780 mad.rn.f32 $r14, s[0x0064], $r19, $r14
000b90: e013f435 00234780 mad.rn.f32 $r13, s[0x0068], $r19, $r13
000b98: e013f631 00230780 mad.rn.f32 $r12, s[0x006c], $r19, $r12
000ba0: e013f82d 0022c780 mad.rn.f32 $r11, s[0x0070], $r19, $r11
000ba8: e013fa29 00228780 mad.rn.f32 $r10, s[0x0074], $r19, $r10
000bb0: e013fc25 00224780 mad.rn.f32 $r9, s[0x0078], $r19, $r9
000bb8: e013fe21 00220780 mad.rn.f32 $r8, s[0x007c], $r19, $r8
000bc0: e813c01d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0000], $r19, $r7
000bc8: e813c219 00218780 mad.rn.f32 $r6, s[$ofs2+0x0004], $r19, $r6
000bd0: e813c415 00214780 mad.rn.f32 $r5, s[$ofs2+0x0008], $r19, $r5
000bd8: e813c611 00210780 mad.rn.f32 $r4, s[$ofs2+0x000c], $r19, $r4
000be0: e813c80d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0010], $r19, $r3
000be8: e813ca09 00208780 mad.rn.f32 $r2, s[$ofs2+0x0014], $r19, $r2
000bf0: e813cc05 00204780 mad.rn.f32 $r1, s[$ofs2+0x0018], $r19, $r1
000bf8: e013f03d 00254780 mad.rn.f32 $r15, s[0x0060], $r19, $r21
000c00: e813ce4d 00200780 mad.rn.f32 $r19, s[$ofs2+0x001c], $r19, $r0
000c08: 2000a001 00000403 add.b32 $r0, $r16, 0x00004000
000c10: d00e0001 80c00780 mov.u32 $r0, g[$r0]
000c18: e812d439 00238780 mad.rn.f32 $r14, s[$ofs2+0x0028], $r18, $r14
000c20: e812d635 00234780 mad.rn.f32 $r13, s[$ofs2+0x002c], $r18, $r13
000c28: e812d831 00230780 mad.rn.f32 $r12, s[$ofs2+0x0030], $r18, $r12
000c30: e812da2d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0034], $r18, $r11
000c38: e812dc29 00228780 mad.rn.f32 $r10, s[$ofs2+0x0038], $r18, $r10
000c40: e812de25 00224780 mad.rn.f32 $r9, s[$ofs2+0x003c], $r18, $r9
000c48: e812e021 00220780 mad.rn.f32 $r8, s[$ofs2+0x0040], $r18, $r8
000c50: e812e21d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0044], $r18, $r7
000c58: e812e419 00218780 mad.rn.f32 $r6, s[$ofs2+0x0048], $r18, $r6
000c60: e812e615 00214780 mad.rn.f32 $r5, s[$ofs2+0x004c], $r18, $r5
000c68: e812e811 00210780 mad.rn.f32 $r4, s[$ofs2+0x0050], $r18, $r4
000c70: e812ea0d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0054], $r18, $r3
000c78: e812ec09 00208780 mad.rn.f32 $r2, s[$ofs2+0x0058], $r18, $r2
000c80: e812ee05 00204780 mad.rn.f32 $r1, s[$ofs2+0x005c], $r18, $r1
000c88: e812d23d 0023c780 mad.rn.f32 $r15, s[$ofs2+0x0024], $r18, $r15
000c90: e812f049 0024c780 mad.rn.f32 $r18, s[$ofs2+0x0060], $r18, $r19
000c98: d0020005 20000780 mov.b32 $ofs1, 0x00000100
000ca0: e811f639 00238780 mad.rn.f32 $r14, s[$ofs2+0x006c], $r17, $r14
000ca8: e811f835 00234780 mad.rn.f32 $r13, s[$ofs2+0x0070], $r17, $r13
000cb0: e811fa31 00230780 mad.rn.f32 $r12, s[$ofs2+0x0074], $r17, $r12
000cb8: e811fc2d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0078], $r17, $r11
000cc0: e811fe29 00228780 mad.rn.f32 $r10, s[$ofs2+0x007c], $r17, $r10
000cc8: e411c025 00224780 mad.rn.f32 $r9, s[$ofs1+0x0000], $r17, $r9
000cd0: e411c221 00220780 mad.rn.f32 $r8, s[$ofs1+0x0004], $r17, $r8
000cd8: e411c41d 0021c780 mad.rn.f32 $r7, s[$ofs1+0x0008], $r17, $r7
000ce0: e411c619 00218780 mad.rn.f32 $r6, s[$ofs1+0x000c], $r17, $r6
000ce8: e411c815 00214780 mad.rn.f32 $r5, s[$ofs1+0x0010], $r17, $r5
000cf0: e411ca11 00210780 mad.rn.f32 $r4, s[$ofs1+0x0014], $r17, $r4
000cf8: e411cc0d 0020c780 mad.rn.f32 $r3, s[$ofs1+0x0018], $r17, $r3
000d00: e411ce09 00208780 mad.rn.f32 $r2, s[$ofs1+0x001c], $r17, $r2
000d08: e411d005 00204780 mad.rn.f32 $r1, s[$ofs1+0x0020], $r17, $r1
000d10: e811f43d 0023c780 mad.rn.f32 $r15, s[$ofs2+0x0068], $r17, $r15
000d18: e411d245 00248780 mad.rn.f32 $r17, s[$ofs1+0x0024], $r17, $r18
000d20: e400d63d 0023c780 mad.rn.f32 $r15, s[$ofs1+0x002c], $r0, $r15
000d28: e400d839 00238780 mad.rn.f32 $r14, s[$ofs1+0x0030], $r0, $r14
000d30: e400da35 00234780 mad.rn.f32 $r13, s[$ofs1+0x0034], $r0, $r13
000d38: e400dc31 00230780 mad.rn.f32 $r12, s[$ofs1+0x0038], $r0, $r12
000d40: e400de2d 0022c780 mad.rn.f32 $r11, s[$ofs1+0x003c], $r0, $r11
000d48: e400e029 00228780 mad.rn.f32 $r10, s[$ofs1+0x0040], $r0, $r10
000d50: e400e225 00224780 mad.rn.f32 $r9, s[$ofs1+0x0044], $r0, $r9
000d58: e400e421 00220780 mad.rn.f32 $r8, s[$ofs1+0x0048], $r0, $r8
000d60: e400e61d 0021c780 mad.rn.f32 $r7, s[$ofs1+0x004c], $r0, $r7
000d68: e400e819 00218780 mad.rn.f32 $r6, s[$ofs1+0x0050], $r0, $r6
000d70: e400ea15 00214780 mad.rn.f32 $r5, s[$ofs1+0x0054], $r0, $r5
000d78: e400ec11 00210780 mad.rn.f32 $r4, s[$ofs1+0x0058], $r0, $r4
000d80: e400ee0d 0020c780 mad.rn.f32 $r3, s[$ofs1+0x005c], $r0, $r3
000d88: e400f009 00208780 mad.rn.f32 $r2, s[$ofs1+0x0060], $r0, $r2
000d90: e400f205 00204780 mad.rn.f32 $r1, s[$ofs1+0x0064], $r0, $r1
000d98: e400f445 00244780 mad.rn.f32 $r17, s[$ofs1+0x0068], $r0, $r17
000da0: d0030009 20000780 mov.b32 $ofs2, 0x00000180
000da8: 2000a001 00000503 add.b32 $r0, $r16, 0x00005000
000db0: d00e0049 80c00780 mov.u32 $r18, g[$r0]
000db8: 2000a001 00000603 add.b32 $r0, $r16, 0x00006000
000dc0: d00e0001 80c00780 mov.u32 $r0, g[$r0]
000dc8: e412f83d 0023c780 mad.rn.f32 $r15, s[$ofs1+0x0070], $r18, $r15
000dd0: e412fa39 00238780 mad.rn.f32 $r14, s[$ofs1+0x0074], $r18, $r14
000dd8: e412fc35 00234780 mad.rn.f32 $r13, s[$ofs1+0x0078], $r18, $r13
000de0: e412fe31 00230780 mad.rn.f32 $r12, s[$ofs1+0x007c], $r18, $r12
000de8: e812c02d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0000], $r18, $r11
000df0: e812c229 00228780 mad.rn.f32 $r10, s[$ofs2+0x0004], $r18, $r10
000df8: e812c425 00224780 mad.rn.f32 $r9, s[$ofs2+0x0008], $r18, $r9
000e00: e812c621 00220780 mad.rn.f32 $r8, s[$ofs2+0x000c], $r18, $r8
000e08: e812c81d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0010], $r18, $r7
000e10: e812ca19 00218780 mad.rn.f32 $r6, s[$ofs2+0x0014], $r18, $r6
000e18: e812cc15 00214780 mad.rn.f32 $r5, s[$ofs2+0x0018], $r18, $r5
000e20: e812ce11 00210780 mad.rn.f32 $r4, s[$ofs2+0x001c], $r18, $r4
000e28: e812d00d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0020], $r18, $r3
000e30: e812d209 00208780 mad.rn.f32 $r2, s[$ofs2+0x0024], $r18, $r2
000e38: e812d405 00204780 mad.rn.f32 $r1, s[$ofs2+0x0028], $r18, $r1
000e40: e812d645 00244780 mad.rn.f32 $r17, s[$ofs2+0x002c], $r18, $r17
000e48: e800da3d 0023c780 mad.rn.f32 $r15, s[$ofs2+0x0034], $r0, $r15
000e50: e800dc39 00238780 mad.rn.f32 $r14, s[$ofs2+0x0038], $r0, $r14
000e58: e800de35 00234780 mad.rn.f32 $r13, s[$ofs2+0x003c], $r0, $r13
000e60: e800e031 00230780 mad.rn.f32 $r12, s[$ofs2+0x0040], $r0, $r12
000e68: e800e22d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0044], $r0, $r11
000e70: e800e429 00228780 mad.rn.f32 $r10, s[$ofs2+0x0048], $r0, $r10
000e78: e800e625 00224780 mad.rn.f32 $r9, s[$ofs2+0x004c], $r0, $r9
000e80: e800e821 00220780 mad.rn.f32 $r8, s[$ofs2+0x0050], $r0, $r8
000e88: e800ea1d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0054], $r0, $r7
000e90: e800ec19 00218780 mad.rn.f32 $r6, s[$ofs2+0x0058], $r0, $r6
000e98: e800ee15 00214780 mad.rn.f32 $r5, s[$ofs2+0x005c], $r0, $r5
000ea0: e800f011 00210780 mad.rn.f32 $r4, s[$ofs2+0x0060], $r0, $r4
000ea8: e800f20d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0064], $r0, $r3
000eb0: e800f409 00208780 mad.rn.f32 $r2, s[$ofs2+0x0068], $r0, $r2
000eb8: e800f605 00204780 mad.rn.f32 $r1, s[$ofs2+0x006c], $r0, $r1
000ec0: e800f845 00244780 mad.rn.f32 $r17, s[$ofs2+0x0070], $r0, $r17
000ec8: d0040005 20000780 mov.b32 $ofs1, 0x00000200
000ed0: 2000a001 00001003 add.b32 $r0, $r16, 0x00010000
000ed8: 2000a041 00000703 add.b32 $r16, $r16, 0x00007000
000ee0: d00e2049 80c00780 mov.u32 $r18, g[$r16]
000ee8: 20008041 0ffff803 add.b32 $r16, $r0, 0xffff8000
000ef0: d00e2041 80c00780 mov.u32 $r16, g[$r16]
000ef8: e812fc3d 0023c780 mad.rn.f32 $r15, s[$ofs2+0x0078], $r18, $r15
000f00: e812fe39 00238780 mad.rn.f32 $r14, s[$ofs2+0x007c], $r18, $r14
000f08: e412c035 00234780 mad.rn.f32 $r13, s[$ofs1+0x0000], $r18, $r13
000f10: e412c231 00230780 mad.rn.f32 $r12, s[$ofs1+0x0004], $r18, $r12
000f18: e412c42d 0022c780 mad.rn.f32 $r11, s[$ofs1+0x0008], $r18, $r11
000f20: e412c629 00228780 mad.rn.f32 $r10, s[$ofs1+0x000c], $r18, $r10
000f28: e412c825 00224780 mad.rn.f32 $r9, s[$ofs1+0x0010], $r18, $r9
000f30: e412ca21 00220780 mad.rn.f32 $r8, s[$ofs1+0x0014], $r18, $r8
000f38: e412cc1d 0021c780 mad.rn.f32 $r7, s[$ofs1+0x0018], $r18, $r7
000f40: e412ce19 00218780 mad.rn.f32 $r6, s[$ofs1+0x001c], $r18, $r6
000f48: e412d015 00214780 mad.rn.f32 $r5, s[$ofs1+0x0020], $r18, $r5
000f50: e412d211 00210780 mad.rn.f32 $r4, s[$ofs1+0x0024], $r18, $r4
000f58: e412d40d 0020c780 mad.rn.f32 $r3, s[$ofs1+0x0028], $r18, $r3
000f60: e412d609 00208780 mad.rn.f32 $r2, s[$ofs1+0x002c], $r18, $r2
000f68: e412d805 00204780 mad.rn.f32 $r1, s[$ofs1+0x0030], $r18, $r1
000f70: e412da45 00244780 mad.rn.f32 $r17, s[$ofs1+0x0034], $r18, $r17
000f78: e410de3d 0023c780 mad.rn.f32 $r15, s[$ofs1+0x003c], $r16, $r15
000f80: e410e039 00238780 mad.rn.f32 $r14, s[$ofs1+0x0040], $r16, $r14
000f88: e410e235 00234780 mad.rn.f32 $r13, s[$ofs1+0x0044], $r16, $r13
000f90: e410e431 00230780 mad.rn.f32 $r12, s[$ofs1+0x0048], $r16, $r12
000f98: e410e62d 0022c780 mad.rn.f32 $r11, s[$ofs1+0x004c], $r16, $r11
000fa0: e410e829 00228780 mad.rn.f32 $r10, s[$ofs1+0x0050], $r16, $r10
000fa8: e410ea25 00224780 mad.rn.f32 $r9, s[$ofs1+0x0054], $r16, $r9
000fb0: e410ec21 00220780 mad.rn.f32 $r8, s[$ofs1+0x0058], $r16, $r8
000fb8: e410ee1d 0021c780 mad.rn.f32 $r7, s[$ofs1+0x005c], $r16, $r7
000fc0: e410f019 00218780 mad.rn.f32 $r6, s[$ofs1+0x0060], $r16, $r6
000fc8: e410f215 00214780 mad.rn.f32 $r5, s[$ofs1+0x0064], $r16, $r5
000fd0: e410f411 00210780 mad.rn.f32 $r4, s[$ofs1+0x0068], $r16, $r4
000fd8: e410f60d 0020c780 mad.rn.f32 $r3, s[$ofs1+0x006c], $r16, $r3
000fe0: e410f809 00208780 mad.rn.f32 $r2, s[$ofs1+0x0070], $r16, $r2
000fe8: e410fa05 00204780 mad.rn.f32 $r1, s[$ofs1+0x0074], $r16, $r1
000ff0: e410fc59 00244780 mad.rn.f32 $r22, s[$ofs1+0x0078], $r16, $r17
000ff8: d0050011 20000780 mov.b32 $ofs4, 0x00000280
001000: d006000d 20000780 mov.b32 $ofs3, 0x00000300
001008: d0070009 20000780 mov.b32 $ofs2, 0x00000380
001010: d0080005 20000780 mov.b32 $ofs1, 0x00000400
001018: 20008041 0ffff903 add.b32 $r16, $r0, 0xffff9000
001020: d00e205d 80c00780 mov.u32 $r23, g[$r16]
001028: 20008041 0ffffa03 add.b32 $r16, $r0, 0xffffa000
001030: d00e2055 80c00780 mov.u32 $r21, g[$r16]
001038: 20008041 0ffffb03 add.b32 $r16, $r0, 0xffffb000
001040: d00e204d 80c00780 mov.u32 $r19, g[$r16]
001048: 20008041 0ffffc03 add.b32 $r16, $r0, 0xffffc000
001050: d00e2049 80c00780 mov.u32 $r18, g[$r16]
001058: 20008041 0ffffd03 add.b32 $r16, $r0, 0xffffd000
001060: d00e2045 80c00780 mov.u32 $r17, g[$r16]
001068: 20008041 0ffffe03 add.b32 $r16, $r0, 0xffffe000
001070: d00e2041 80c00780 mov.u32 $r16, g[$r16]
001078: 20008001 0fffff03 add.b32 $r0, $r0, 0xfffff000
001080: d00e0001 80c00780 mov.u32 $r0, g[$r0]
001088: e017c03d 0023c784 mad.rn.f32 $r15, s[$ofs4+0x0000], $r23, $r15
001090: 30022851 c4100780 shl.u32 $r20, $r20, 0x00000002
001098: e015e23d 0023c784 mad.rn.f32 $r15, s[$ofs4+0x0044], $r21, $r15
0010a0: 2000cc51 04250780 add.u32 $r20, s[0x0018], $r20
0010a8: ec13c43d 0023c780 mad.rn.f32 $r15, s[$ofs3+0x0008], $r19, $r15
0010b0: e017c239 00238784 mad.rn.f32 $r14, s[$ofs4+0x0004], $r23, $r14
0010b8: ec12e63d 0023c780 mad.rn.f32 $r15, s[$ofs3+0x004c], $r18, $r15
0010c0: e015e439 00238784 mad.rn.f32 $r14, s[$ofs4+0x0048], $r21, $r14
0010c8: e811c83d 0023c780 mad.rn.f32 $r15, s[$ofs2+0x0010], $r17, $r15
0010d0: ec13c639 00238780 mad.rn.f32 $r14, s[$ofs3+0x000c], $r19, $r14
0010d8: e810ea3d 0023c780 mad.rn.f32 $r15, s[$ofs2+0x0054], $r16, $r15
0010e0: ec12e839 00238780 mad.rn.f32 $r14, s[$ofs3+0x0050], $r18, $r14
0010e8: e400cc3d 0023c780 mad.rn.f32 $r15, s[$ofs1+0x0018], $r0, $r15
0010f0: e811ca39 00238780 mad.rn.f32 $r14, s[$ofs2+0x0014], $r17, $r14
0010f8: d00e283d a0c00780 mov.u32 g[$r20], $r15
001100: e810ec39 00238780 mad.rn.f32 $r14, s[$ofs2+0x0058], $r16, $r14
001108: 2000a83d 00000103 add.b32 $r15, $r20, 0x00001000
001110: e400ce39 00238780 mad.rn.f32 $r14, s[$ofs1+0x001c], $r0, $r14
001118: e017c435 00234784 mad.rn.f32 $r13, s[$ofs4+0x0008], $r23, $r13
001120: d00e1e39 a0c00780 mov.u32 g[$r15], $r14
001128: e015e635 00234784 mad.rn.f32 $r13, s[$ofs4+0x004c], $r21, $r13
001130: 2000a839 00000203 add.b32 $r14, $r20, 0x00002000
001138: ec13c835 00234780 mad.rn.f32 $r13, s[$ofs3+0x0010], $r19, $r13
001140: e017c631 00230784 mad.rn.f32 $r12, s[$ofs4+0x000c], $r23, $r12
001148: ec12ea35 00234780 mad.rn.f32 $r13, s[$ofs3+0x0054], $r18, $r13
001150: e015e831 00230784 mad.rn.f32 $r12, s[$ofs4+0x0050], $r21, $r12
001158: e811cc35 00234780 mad.rn.f32 $r13, s[$ofs2+0x0018], $r17, $r13
001160: ec13ca31 00230780 mad.rn.f32 $r12, s[$ofs3+0x0014], $r19, $r12
001168: e810ee35 00234780 mad.rn.f32 $r13, s[$ofs2+0x005c], $r16, $r13
001170: ec12ec31 00230780 mad.rn.f32 $r12, s[$ofs3+0x0058], $r18, $r12
001178: e400d035 00234780 mad.rn.f32 $r13, s[$ofs1+0x0020], $r0, $r13
001180: e811ce31 00230780 mad.rn.f32 $r12, s[$ofs2+0x001c], $r17, $r12
001188: d00e1c35 a0c00780 mov.u32 g[$r14], $r13
001190: e810f031 00230780 mad.rn.f32 $r12, s[$ofs2+0x0060], $r16, $r12
001198: 2000a835 00000303 add.b32 $r13, $r20, 0x00003000
0011a0: e400d231 00230780 mad.rn.f32 $r12, s[$ofs1+0x0024], $r0, $r12
0011a8: e017c82d 0022c784 mad.rn.f32 $r11, s[$ofs4+0x0010], $r23, $r11
0011b0: d00e1a31 a0c00780 mov.u32 g[$r13], $r12
0011b8: e015ea2d 0022c784 mad.rn.f32 $r11, s[$ofs4+0x0054], $r21, $r11
0011c0: 2000a831 00000403 add.b32 $r12, $r20, 0x00004000
0011c8: ec13cc2d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x0018], $r19, $r11
0011d0: e017ca29 00228784 mad.rn.f32 $r10, s[$ofs4+0x0014], $r23, $r10
0011d8: ec12ee2d 0022c780 mad.rn.f32 $r11, s[$ofs3+0x005c], $r18, $r11
0011e0: e015ec29 00228784 mad.rn.f32 $r10, s[$ofs4+0x0058], $r21, $r10
0011e8: e811d02d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0020], $r17, $r11
0011f0: ec13ce29 00228780 mad.rn.f32 $r10, s[$ofs3+0x001c], $r19, $r10
0011f8: e810f22d 0022c780 mad.rn.f32 $r11, s[$ofs2+0x0064], $r16, $r11
001200: ec12f029 00228780 mad.rn.f32 $r10, s[$ofs3+0x0060], $r18, $r10
001208: e400d42d 0022c780 mad.rn.f32 $r11, s[$ofs1+0x0028], $r0, $r11
001210: e811d229 00228780 mad.rn.f32 $r10, s[$ofs2+0x0024], $r17, $r10
001218: d00e182d a0c00780 mov.u32 g[$r12], $r11
001220: e810f429 00228780 mad.rn.f32 $r10, s[$ofs2+0x0068], $r16, $r10
001228: 2000a82d 00000503 add.b32 $r11, $r20, 0x00005000
001230: e400d629 00228780 mad.rn.f32 $r10, s[$ofs1+0x002c], $r0, $r10
001238: e017cc25 00224784 mad.rn.f32 $r9, s[$ofs4+0x0018], $r23, $r9
001240: d00e1629 a0c00780 mov.u32 g[$r11], $r10
001248: e015ee25 00224784 mad.rn.f32 $r9, s[$ofs4+0x005c], $r21, $r9
001250: 2000a829 00000603 add.b32 $r10, $r20, 0x00006000
001258: ec13d025 00224780 mad.rn.f32 $r9, s[$ofs3+0x0020], $r19, $r9
001260: e017ce21 00220784 mad.rn.f32 $r8, s[$ofs4+0x001c], $r23, $r8
001268: ec12f225 00224780 mad.rn.f32 $r9, s[$ofs3+0x0064], $r18, $r9
001270: e015f021 00220784 mad.rn.f32 $r8, s[$ofs4+0x0060], $r21, $r8
001278: e811d425 00224780 mad.rn.f32 $r9, s[$ofs2+0x0028], $r17, $r9
001280: ec13d221 00220780 mad.rn.f32 $r8, s[$ofs3+0x0024], $r19, $r8
001288: e810f625 00224780 mad.rn.f32 $r9, s[$ofs2+0x006c], $r16, $r9
001290: ec12f421 00220780 mad.rn.f32 $r8, s[$ofs3+0x0068], $r18, $r8
001298: e400d825 00224780 mad.rn.f32 $r9, s[$ofs1+0x0030], $r0, $r9
0012a0: e811d621 00220780 mad.rn.f32 $r8, s[$ofs2+0x002c], $r17, $r8
0012a8: d00e1425 a0c00780 mov.u32 g[$r10], $r9
0012b0: e810f821 00220780 mad.rn.f32 $r8, s[$ofs2+0x0070], $r16, $r8
0012b8: 2000a825 00000703 add.b32 $r9, $r20, 0x00007000
0012c0: e400da21 00220780 mad.rn.f32 $r8, s[$ofs1+0x0034], $r0, $r8
0012c8: 2000a829 00001003 add.b32 $r10, $r20, 0x00010000
0012d0: d00e1221 a0c00780 mov.u32 g[$r9], $r8
0012d8: 20009421 0ffff803 add.b32 $r8, $r10, 0xffff8000
0012e0: e017d01d 0021c784 mad.rn.f32 $r7, s[$ofs4+0x0020], $r23, $r7
0012e8: e017d219 00218784 mad.rn.f32 $r6, s[$ofs4+0x0024], $r23, $r6
0012f0: e015f21d 0021c784 mad.rn.f32 $r7, s[$ofs4+0x0064], $r21, $r7
0012f8: e017d415 00214784 mad.rn.f32 $r5, s[$ofs4+0x0028], $r23, $r5
001300: ec13d41d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x0028], $r19, $r7
001308: e017d611 00210784 mad.rn.f32 $r4, s[$ofs4+0x002c], $r23, $r4
001310: ec12f61d 0021c780 mad.rn.f32 $r7, s[$ofs3+0x006c], $r18, $r7
001318: e017d80d 0020c784 mad.rn.f32 $r3, s[$ofs4+0x0030], $r23, $r3
001320: e811d81d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0030], $r17, $r7
001328: e017da09 00208784 mad.rn.f32 $r2, s[$ofs4+0x0034], $r23, $r2
001330: e810fa1d 0021c780 mad.rn.f32 $r7, s[$ofs2+0x0074], $r16, $r7
001338: e017dc05 00204784 mad.rn.f32 $r1, s[$ofs4+0x0038], $r23, $r1
001340: e400dc1d 0021c780 mad.rn.f32 $r7, s[$ofs1+0x0038], $r0, $r7
001348: e017de25 00258784 mad.rn.f32 $r9, s[$ofs4+0x003c], $r23, $r22
001350: d00e101d a0c00780 mov.u32 g[$r8], $r7
001358: 2000941d 0ffff903 add.b32 $r7, $r10, 0xffff9000
001360: e015f419 00218784 mad.rn.f32 $r6, s[$ofs4+0x0068], $r21, $r6
001368: e015f615 00214784 mad.rn.f32 $r5, s[$ofs4+0x006c], $r21, $r5
001370: ec13d619 00218780 mad.rn.f32 $r6, s[$ofs3+0x002c], $r19, $r6
001378: e015f811 00210784 mad.rn.f32 $r4, s[$ofs4+0x0070], $r21, $r4
001380: ec12f819 00218780 mad.rn.f32 $r6, s[$ofs3+0x0070], $r18, $r6
001388: e015fa0d 0020c784 mad.rn.f32 $r3, s[$ofs4+0x0074], $r21, $r3
001390: e811da19 00218780 mad.rn.f32 $r6, s[$ofs2+0x0034], $r17, $r6
001398: e015fc09 00208784 mad.rn.f32 $r2, s[$ofs4+0x0078], $r21, $r2
0013a0: e810fc19 00218780 mad.rn.f32 $r6, s[$ofs2+0x0078], $r16, $r6
0013a8: e015fe05 00204784 mad.rn.f32 $r1, s[$ofs4+0x007c], $r21, $r1
0013b0: e400de19 00218780 mad.rn.f32 $r6, s[$ofs1+0x003c], $r0, $r6
0013b8: ec15c021 00224780 mad.rn.f32 $r8, s[$ofs3+0x0000], $r21, $r9
0013c0: d00e0e19 a0c00780 mov.u32 g[$r7], $r6
0013c8: 20009419 0ffffa03 add.b32 $r6, $r10, 0xffffa000
0013d0: ec13d815 00214780 mad.rn.f32 $r5, s[$ofs3+0x0030], $r19, $r5
0013d8: ec13da11 00210780 mad.rn.f32 $r4, s[$ofs3+0x0034], $r19, $r4
0013e0: ec12fa15 00214780 mad.rn.f32 $r5, s[$ofs3+0x0074], $r18, $r5
0013e8: ec13dc0d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x0038], $r19, $r3
0013f0: e811dc15 00214780 mad.rn.f32 $r5, s[$ofs2+0x0038], $r17, $r5
0013f8: ec13de09 00208780 mad.rn.f32 $r2, s[$ofs3+0x003c], $r19, $r2
001400: e810fe15 00214780 mad.rn.f32 $r5, s[$ofs2+0x007c], $r16, $r5
001408: ec13e005 00204780 mad.rn.f32 $r1, s[$ofs3+0x0040], $r19, $r1
001410: e400e015 00214780 mad.rn.f32 $r5, s[$ofs1+0x0040], $r0, $r5
001418: ec13e21d 00220780 mad.rn.f32 $r7, s[$ofs3+0x0044], $r19, $r8
001420: d00e0c15 a0c00780 mov.u32 g[$r6], $r5
001428: 20009415 0ffffb03 add.b32 $r5, $r10, 0xffffb000
001430: ec12fc11 00210780 mad.rn.f32 $r4, s[$ofs3+0x0078], $r18, $r4
001438: ec12fe0d 0020c780 mad.rn.f32 $r3, s[$ofs3+0x007c], $r18, $r3
001440: e811de11 00210780 mad.rn.f32 $r4, s[$ofs2+0x003c], $r17, $r4
001448: e812c009 00208780 mad.rn.f32 $r2, s[$ofs2+0x0000], $r18, $r2
001450: e410c011 00210780 mad.rn.f32 $r4, s[$ofs1+0x0000], $r16, $r4
001458: e812c205 00204780 mad.rn.f32 $r1, s[$ofs2+0x0004], $r18, $r1
001460: e400e211 00210780 mad.rn.f32 $r4, s[$ofs1+0x0044], $r0, $r4
001468: e812c419 0021c780 mad.rn.f32 $r6, s[$ofs2+0x0008], $r18, $r7
001470: d00e0a11 a0c00780 mov.u32 g[$r5], $r4
001478: 20009411 0ffffc03 add.b32 $r4, $r10, 0xffffc000
001480: e811e00d 0020c780 mad.rn.f32 $r3, s[$ofs2+0x0040], $r17, $r3
001488: e811e209 00208780 mad.rn.f32 $r2, s[$ofs2+0x0044], $r17, $r2
001490: e410c20d 0020c780 mad.rn.f32 $r3, s[$ofs1+0x0004], $r16, $r3
001498: e811e405 00204780 mad.rn.f32 $r1, s[$ofs2+0x0048], $r17, $r1
0014a0: e400e40d 0020c780 mad.rn.f32 $r3, s[$ofs1+0x0048], $r0, $r3
0014a8: e811e615 00218780 mad.rn.f32 $r5, s[$ofs2+0x004c], $r17, $r6
0014b0: d00e080d a0c00780 mov.u32 g[$r4], $r3
0014b8: 2000940d 0ffffd03 add.b32 $r3, $r10, 0xffffd000
0014c0: e410c409 00208780 mad.rn.f32 $r2, s[$ofs1+0x0008], $r16, $r2
0014c8: e410c605 00204780 mad.rn.f32 $r1, s[$ofs1+0x000c], $r16, $r1
0014d0: e400e609 00208780 mad.rn.f32 $r2, s[$ofs1+0x004c], $r0, $r2
0014d8: e410c811 00214780 mad.rn.f32 $r4, s[$ofs1+0x0010], $r16, $r5
0014e0: d00e0609 a0c00780 mov.u32 g[$r3], $r2
0014e8: 20009409 0ffffe03 add.b32 $r2, $r10, 0xffffe000
0014f0: e400e805 00204780 mad.rn.f32 $r1, s[$ofs1+0x0050], $r0, $r1
0014f8: e400ea01 00210780 mad.rn.f32 $r0, s[$ofs1+0x0054], $r0, $r4
001500: d00e0405 a0c00780 mov.u32 g[$r2], $r1
001508: 20009405 0fffff03 add.b32 $r1, $r10, 0xfffff000
001510: d00e0201 a0c00781 mov.end.u32 g[$r1], $r0
// segment: const (1:0000)
0000: 000003ff 00000011