set.le.s32 $p0|$o127, $r38, c1[0x0008]
@$p0.ne bra.label label5
shl.u32 $r4, s[0x0020], 0x00000003
mov.b64 $r0, g[$r45]
add.u32 $r45, $r4, $r45
mov.b64 $r2, g[$r45]
add.u32 $r45, $r4, $r45
mov.b32 $ofs2, 0x00000480
mov.b32 $ofs3, 0x00000080
mov.b32 $ofs4, 0x00000480
lds.b32 $r6, s[$ofs4+0x0064]
mad.rn.f32 $r47, $r6, $r3, -$r47
mad.rn.f32 $r46, $r6, $r2, $r46
lds.b32 $r7, s[0x0060]
lds.b32 $r6, s[$ofs3+0x0024]
mad.rn.f32 $r46, $r7, $r1, $r46
mad.rn.f32 $r47, $r6, $r2, -$r47
mul.rn.f32 $r5, s[$ofs2+0x0020], $r1
mad.rn.f32 $r46, $r6, $r3, $r46
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r46, s[$ofs2+0x0020], $r0, $r46
add.rn.f32 $r47, $r47, $r5
lds.b32 $r6, s[$ofs4+0x0068]
mad.rn.f32 $r44, $r6, $r3, -$r44
mad.rn.f32 $r43, $r6, $r2, $r43
lds.b32 $r7, s[0x0064]
lds.b32 $r6, s[$ofs3+0x0028]
mad.rn.f32 $r43, $r7, $r1, $r43
mad.rn.f32 $r44, $r6, $r2, -$r44
mul.rn.f32 $r5, s[$ofs2+0x0024], $r1
mad.rn.f32 $r43, $r6, $r3, $r43
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r43, s[$ofs2+0x0024], $r0, $r43
add.rn.f32 $r44, $r44, $r5
lds.b32 $r6, s[$ofs4+0x006c]
mad.rn.f32 $r42, $r6, $r3, -$r42
mad.rn.f32 $r41, $r6, $r2, $r41
lds.b32 $r7, s[0x0068]
lds.b32 $r6, s[$ofs3+0x002c]
mad.rn.f32 $r41, $r7, $r1, $r41
mad.rn.f32 $r42, $r6, $r2, -$r42
mul.rn.f32 $r5, s[$ofs2+0x0028], $r1
mad.rn.f32 $r41, $r6, $r3, $r41
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r41, s[$ofs2+0x0028], $r0, $r41
add.rn.f32 $r42, $r42, $r5
lds.b32 $r6, s[$ofs4+0x0070]
mad.rn.f32 $r36, $r6, $r3, -$r36
mad.rn.f32 $r35, $r6, $r2, $r35
lds.b32 $r7, s[0x006c]
lds.b32 $r6, s[$ofs3+0x0030]
mad.rn.f32 $r35, $r7, $r1, $r35
mad.rn.f32 $r36, $r6, $r2, -$r36
mul.rn.f32 $r5, s[$ofs2+0x002c], $r1
mad.rn.f32 $r35, $r6, $r3, $r35
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r35, s[$ofs2+0x002c], $r0, $r35
add.rn.f32 $r36, $r36, $r5
lds.b32 $r6, s[$ofs4+0x0074]
mad.rn.f32 $r34, $r6, $r3, -$r34
mad.rn.f32 $r33, $r6, $r2, $r33
lds.b32 $r7, s[0x0070]
lds.b32 $r6, s[$ofs3+0x0034]
mad.rn.f32 $r33, $r7, $r1, $r33
mad.rn.f32 $r34, $r6, $r2, -$r34
mul.rn.f32 $r5, s[$ofs2+0x0030], $r1
mad.rn.f32 $r33, $r6, $r3, $r33
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r33, s[$ofs2+0x0030], $r0, $r33
add.rn.f32 $r34, $r34, $r5
lds.b32 $r6, s[$ofs4+0x0078]
mad.rn.f32 $r32, $r6, $r3, -$r32
mad.rn.f32 $r31, $r6, $r2, $r31
lds.b32 $r7, s[0x0074]
lds.b32 $r6, s[$ofs3+0x0038]
mad.rn.f32 $r31, $r7, $r1, $r31
mad.rn.f32 $r32, $r6, $r2, -$r32
mul.rn.f32 $r5, s[$ofs2+0x0034], $r1
mad.rn.f32 $r31, $r6, $r3, $r31
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r31, s[$ofs2+0x0034], $r0, $r31
add.rn.f32 $r32, $r32, $r5
lds.b32 $r6, s[$ofs4+0x007c]
mad.rn.f32 $r30, $r6, $r3, -$r30
mad.rn.f32 $r29, $r6, $r2, $r29
lds.b32 $r7, s[0x0078]
lds.b32 $r6, s[$ofs3+0x003c]
mad.rn.f32 $r29, $r7, $r1, $r29
mad.rn.f32 $r30, $r6, $r2, -$r30
mul.rn.f32 $r5, s[$ofs2+0x0038], $r1
mad.rn.f32 $r29, $r6, $r3, $r29
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r29, s[$ofs2+0x0038], $r0, $r29
add.rn.f32 $r30, $r30, $r5
mov.b32 $ofs4, 0x00000500
lds.b32 $r6, s[$ofs4+0x0000]
mad.rn.f32 $r27, $r6, $r3, -$r27
mad.rn.f32 $r26, $r6, $r2, $r26
lds.b32 $r7, s[0x007c]
lds.b32 $r6, s[$ofs3+0x0040]
mad.rn.f32 $r26, $r7, $r1, $r26
mad.rn.f32 $r27, $r6, $r2, -$r27
mul.rn.f32 $r5, s[$ofs2+0x003c], $r1
mad.rn.f32 $r26, $r6, $r3, $r26
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r26, s[$ofs2+0x003c], $r0, $r26
add.rn.f32 $r27, $r27, $r5
mov.b32 $ofs1, 0x00000080
lds.b32 $r6, s[$ofs4+0x0004]
mad.rn.f32 $r25, $r6, $r3, -$r25
mad.rn.f32 $r24, $r6, $r2, $r24
lds.b32 $r7, s[$ofs1+0x0000]
lds.b32 $r6, s[$ofs3+0x0044]
mad.rn.f32 $r24, $r7, $r1, $r24
mad.rn.f32 $r25, $r6, $r2, -$r25
mul.rn.f32 $r5, s[$ofs2+0x0040], $r1
mad.rn.f32 $r24, $r6, $r3, $r24
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r24, s[$ofs2+0x0040], $r0, $r24
add.rn.f32 $r25, $r25, $r5
lds.b32 $r6, s[$ofs4+0x0008]
mad.rn.f32 $r23, $r6, $r3, -$r23
mad.rn.f32 $r22, $r6, $r2, $r22
lds.b32 $r7, s[$ofs1+0x0004]
lds.b32 $r6, s[$ofs3+0x0048]
mad.rn.f32 $r22, $r7, $r1, $r22
mad.rn.f32 $r23, $r6, $r2, -$r23
mul.rn.f32 $r5, s[$ofs2+0x0044], $r1
mad.rn.f32 $r22, $r6, $r3, $r22
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r22, s[$ofs2+0x0044], $r0, $r22
add.rn.f32 $r23, $r23, $r5
lds.b32 $r6, s[$ofs4+0x000c]
mad.rn.f32 $r10, $r6, $r3, -$r10
mad.rn.f32 $r11, $r6, $r2, $r11
lds.b32 $r7, s[$ofs1+0x0008]
lds.b32 $r6, s[$ofs3+0x004c]
mad.rn.f32 $r11, $r7, $r1, $r11
mad.rn.f32 $r10, $r6, $r2, -$r10
mul.rn.f32 $r5, s[$ofs2+0x0048], $r1
mad.rn.f32 $r11, $r6, $r3, $r11
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r11, s[$ofs2+0x0048], $r0, $r11
add.rn.f32 $r10, $r10, $r5
lds.b32 $r6, s[$ofs4+0x0010]
mad.rn.f32 $r12, $r6, $r3, -$r12
mad.rn.f32 $r13, $r6, $r2, $r13
lds.b32 $r7, s[$ofs1+0x000c]
lds.b32 $r6, s[$ofs3+0x0050]
mad.rn.f32 $r13, $r7, $r1, $r13
mad.rn.f32 $r12, $r6, $r2, -$r12
mul.rn.f32 $r5, s[$ofs2+0x004c], $r1
mad.rn.f32 $r13, $r6, $r3, $r13
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r13, s[$ofs2+0x004c], $r0, $r13
add.rn.f32 $r12, $r12, $r5
lds.b32 $r6, s[$ofs4+0x0014]
mad.rn.f32 $r14, $r6, $r3, -$r14
mad.rn.f32 $r15, $r6, $r2, $r15
lds.b32 $r7, s[$ofs1+0x0010]
lds.b32 $r6, s[$ofs3+0x0054]
mad.rn.f32 $r15, $r7, $r1, $r15
mad.rn.f32 $r14, $r6, $r2, -$r14
mul.rn.f32 $r5, s[$ofs2+0x0050], $r1
mad.rn.f32 $r15, $r6, $r3, $r15
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r15, s[$ofs2+0x0050], $r0, $r15
add.rn.f32 $r14, $r14, $r5
lds.b32 $r6, s[$ofs4+0x0018]
mad.rn.f32 $r16, $r6, $r3, -$r16
mad.rn.f32 $r17, $r6, $r2, $r17
lds.b32 $r7, s[$ofs1+0x0014]
lds.b32 $r6, s[$ofs3+0x0058]
mad.rn.f32 $r17, $r7, $r1, $r17
mad.rn.f32 $r16, $r6, $r2, -$r16
mul.rn.f32 $r5, s[$ofs2+0x0054], $r1
mad.rn.f32 $r17, $r6, $r3, $r17
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r17, s[$ofs2+0x0054], $r0, $r17
add.rn.f32 $r16, $r16, $r5
lds.b32 $r6, s[$ofs4+0x001c]
mad.rn.f32 $r18, $r6, $r3, -$r18
mad.rn.f32 $r19, $r6, $r2, $r19
lds.b32 $r7, s[$ofs1+0x0018]
lds.b32 $r6, s[$ofs3+0x005c]
mad.rn.f32 $r19, $r7, $r1, $r19
mad.rn.f32 $r18, $r6, $r2, -$r18
mul.rn.f32 $r5, s[$ofs2+0x0058], $r1
mad.rn.f32 $r19, $r6, $r3, $r19
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r19, s[$ofs2+0x0058], $r0, $r19
add.rn.f32 $r18, $r18, $r5
lds.b32 $r6, s[$ofs4+0x0020]
mad.rn.f32 $r20, $r6, $r3, -$r20
mad.rn.f32 $r21, $r6, $r2, $r21
lds.b32 $r7, s[$ofs1+0x001c]
lds.b32 $r6, s[$ofs3+0x0060]
mad.rn.f32 $r21, $r7, $r1, $r21
mad.rn.f32 $r20, $r6, $r2, -$r20
mul.rn.f32 $r5, s[$ofs2+0x005c], $r1
mad.rn.f32 $r21, $r6, $r3, $r21
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r21, s[$ofs2+0x005c], $r0, $r21
add.rn.f32 $r20, $r20, $r5
mov.b64 $r0, g[$r45]
add.u32 $r45, $r4, $r45
mov.b64 $r2, g[$r45]
add.u32 $r45, $r4, $r45
mov.b32 $ofs2, 0x00000500
mov.b32 $ofs3, 0x00000100
lds.b32 $r6, s[$ofs4+0x006c]
mad.rn.f32 $r47, $r6, $r3, -$r47
mad.rn.f32 $r46, $r6, $r2, $r46
lds.b32 $r7, s[$ofs1+0x0068]
lds.b32 $r6, s[$ofs3+0x002c]
mad.rn.f32 $r46, $r7, $r1, $r46
mad.rn.f32 $r47, $r6, $r2, -$r47
mul.rn.f32 $r5, s[$ofs2+0x0028], $r1
mad.rn.f32 $r46, $r6, $r3, $r46
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r46, s[$ofs2+0x0028], $r0, $r46
add.rn.f32 $r47, $r47, $r5
lds.b32 $r6, s[$ofs4+0x0070]
mad.rn.f32 $r44, $r6, $r3, -$r44
mad.rn.f32 $r43, $r6, $r2, $r43
lds.b32 $r7, s[$ofs1+0x006c]
lds.b32 $r6, s[$ofs3+0x0030]
mad.rn.f32 $r43, $r7, $r1, $r43
mad.rn.f32 $r44, $r6, $r2, -$r44
mul.rn.f32 $r5, s[$ofs2+0x002c], $r1
mad.rn.f32 $r43, $r6, $r3, $r43
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r43, s[$ofs2+0x002c], $r0, $r43
add.rn.f32 $r44, $r44, $r5
lds.b32 $r6, s[$ofs4+0x0074]
mad.rn.f32 $r42, $r6, $r3, -$r42
mad.rn.f32 $r41, $r6, $r2, $r41
lds.b32 $r7, s[$ofs1+0x0070]
lds.b32 $r6, s[$ofs3+0x0034]
mad.rn.f32 $r41, $r7, $r1, $r41
mad.rn.f32 $r42, $r6, $r2, -$r42
mul.rn.f32 $r5, s[$ofs2+0x0030], $r1
mad.rn.f32 $r41, $r6, $r3, $r41
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r41, s[$ofs2+0x0030], $r0, $r41
add.rn.f32 $r42, $r42, $r5
lds.b32 $r6, s[$ofs4+0x0078]
mad.rn.f32 $r36, $r6, $r3, -$r36
mad.rn.f32 $r35, $r6, $r2, $r35
lds.b32 $r7, s[$ofs1+0x0074]
lds.b32 $r6, s[$ofs3+0x0038]
mad.rn.f32 $r35, $r7, $r1, $r35
mad.rn.f32 $r36, $r6, $r2, -$r36
mul.rn.f32 $r5, s[$ofs2+0x0034], $r1
mad.rn.f32 $r35, $r6, $r3, $r35
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r35, s[$ofs2+0x0034], $r0, $r35
add.rn.f32 $r36, $r36, $r5
lds.b32 $r6, s[$ofs4+0x007c]
mad.rn.f32 $r34, $r6, $r3, -$r34
mad.rn.f32 $r33, $r6, $r2, $r33
lds.b32 $r7, s[$ofs1+0x0078]
lds.b32 $r6, s[$ofs3+0x003c]
mad.rn.f32 $r33, $r7, $r1, $r33
mad.rn.f32 $r34, $r6, $r2, -$r34
mul.rn.f32 $r5, s[$ofs2+0x0038], $r1
mad.rn.f32 $r33, $r6, $r3, $r33
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r33, s[$ofs2+0x0038], $r0, $r33
add.rn.f32 $r34, $r34, $r5
mov.b32 $ofs4, 0x00000580
lds.b32 $r6, s[$ofs4+0x0000]
mad.rn.f32 $r32, $r6, $r3, -$r32
mad.rn.f32 $r31, $r6, $r2, $r31
lds.b32 $r7, s[$ofs1+0x007c]
lds.b32 $r6, s[$ofs3+0x0040]
mad.rn.f32 $r31, $r7, $r1, $r31
mad.rn.f32 $r32, $r6, $r2, -$r32
mul.rn.f32 $r5, s[$ofs2+0x003c], $r1
mad.rn.f32 $r31, $r6, $r3, $r31
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r31, s[$ofs2+0x003c], $r0, $r31
add.rn.f32 $r32, $r32, $r5
mov.b32 $ofs1, 0x00000100
lds.b32 $r6, s[$ofs4+0x0004]
mad.rn.f32 $r30, $r6, $r3, -$r30
mad.rn.f32 $r29, $r6, $r2, $r29
lds.b32 $r7, s[$ofs1+0x0000]
lds.b32 $r6, s[$ofs3+0x0044]
mad.rn.f32 $r29, $r7, $r1, $r29
mad.rn.f32 $r30, $r6, $r2, -$r30
mul.rn.f32 $r5, s[$ofs2+0x0040], $r1
mad.rn.f32 $r29, $r6, $r3, $r29
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r29, s[$ofs2+0x0040], $r0, $r29
add.rn.f32 $r30, $r30, $r5
lds.b32 $r6, s[$ofs4+0x0008]
mad.rn.f32 $r27, $r6, $r3, -$r27
mad.rn.f32 $r26, $r6, $r2, $r26
lds.b32 $r7, s[$ofs1+0x0004]
lds.b32 $r6, s[$ofs3+0x0048]
mad.rn.f32 $r26, $r7, $r1, $r26
mad.rn.f32 $r27, $r6, $r2, -$r27
mul.rn.f32 $r5, s[$ofs2+0x0044], $r1
mad.rn.f32 $r26, $r6, $r3, $r26
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r26, s[$ofs2+0x0044], $r0, $r26
add.rn.f32 $r27, $r27, $r5
lds.b32 $r6, s[$ofs4+0x000c]
mad.rn.f32 $r25, $r6, $r3, -$r25
mad.rn.f32 $r24, $r6, $r2, $r24
lds.b32 $r7, s[$ofs1+0x0008]
lds.b32 $r6, s[$ofs3+0x004c]
mad.rn.f32 $r24, $r7, $r1, $r24
mad.rn.f32 $r25, $r6, $r2, -$r25
mul.rn.f32 $r5, s[$ofs2+0x0048], $r1
mad.rn.f32 $r24, $r6, $r3, $r24
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r24, s[$ofs2+0x0048], $r0, $r24
add.rn.f32 $r25, $r25, $r5
lds.b32 $r6, s[$ofs4+0x0010]
mad.rn.f32 $r23, $r6, $r3, -$r23
mad.rn.f32 $r22, $r6, $r2, $r22
lds.b32 $r7, s[$ofs1+0x000c]
lds.b32 $r6, s[$ofs3+0x0050]
mad.rn.f32 $r22, $r7, $r1, $r22
mad.rn.f32 $r23, $r6, $r2, -$r23
mul.rn.f32 $r5, s[$ofs2+0x004c], $r1
mad.rn.f32 $r22, $r6, $r3, $r22
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r22, s[$ofs2+0x004c], $r0, $r22
add.rn.f32 $r23, $r23, $r5
lds.b32 $r6, s[$ofs4+0x0014]
mad.rn.f32 $r10, $r6, $r3, -$r10
mad.rn.f32 $r11, $r6, $r2, $r11
lds.b32 $r7, s[$ofs1+0x0010]
lds.b32 $r6, s[$ofs3+0x0054]
mad.rn.f32 $r11, $r7, $r1, $r11
mad.rn.f32 $r10, $r6, $r2, -$r10
mul.rn.f32 $r5, s[$ofs2+0x0050], $r1
mad.rn.f32 $r11, $r6, $r3, $r11
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r11, s[$ofs2+0x0050], $r0, $r11
add.rn.f32 $r10, $r10, $r5
lds.b32 $r6, s[$ofs4+0x0018]
mad.rn.f32 $r12, $r6, $r3, -$r12
mad.rn.f32 $r13, $r6, $r2, $r13
lds.b32 $r7, s[$ofs1+0x0014]
lds.b32 $r6, s[$ofs3+0x0058]
mad.rn.f32 $r13, $r7, $r1, $r13
mad.rn.f32 $r12, $r6, $r2, -$r12
mul.rn.f32 $r5, s[$ofs2+0x0054], $r1
mad.rn.f32 $r13, $r6, $r3, $r13
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r13, s[$ofs2+0x0054], $r0, $r13
add.rn.f32 $r12, $r12, $r5
lds.b32 $r6, s[$ofs4+0x001c]
mad.rn.f32 $r14, $r6, $r3, -$r14
mad.rn.f32 $r15, $r6, $r2, $r15
lds.b32 $r7, s[$ofs1+0x0018]
lds.b32 $r6, s[$ofs3+0x005c]
mad.rn.f32 $r15, $r7, $r1, $r15
mad.rn.f32 $r14, $r6, $r2, -$r14
mul.rn.f32 $r5, s[$ofs2+0x0058], $r1
mad.rn.f32 $r15, $r6, $r3, $r15
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r15, s[$ofs2+0x0058], $r0, $r15
add.rn.f32 $r14, $r14, $r5
lds.b32 $r6, s[$ofs4+0x0020]
mad.rn.f32 $r16, $r6, $r3, -$r16
mad.rn.f32 $r17, $r6, $r2, $r17
lds.b32 $r7, s[$ofs1+0x001c]
lds.b32 $r6, s[$ofs3+0x0060]
mad.rn.f32 $r17, $r7, $r1, $r17
mad.rn.f32 $r16, $r6, $r2, -$r16
mul.rn.f32 $r5, s[$ofs2+0x005c], $r1
mad.rn.f32 $r17, $r6, $r3, $r17
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r17, s[$ofs2+0x005c], $r0, $r17
add.rn.f32 $r16, $r16, $r5
lds.b32 $r6, s[$ofs4+0x0024]
mad.rn.f32 $r18, $r6, $r3, -$r18
mad.rn.f32 $r19, $r6, $r2, $r19
lds.b32 $r7, s[$ofs1+0x0020]
lds.b32 $r6, s[$ofs3+0x0064]
mad.rn.f32 $r19, $r7, $r1, $r19
mad.rn.f32 $r18, $r6, $r2, -$r18
mul.rn.f32 $r5, s[$ofs2+0x0060], $r1
mad.rn.f32 $r19, $r6, $r3, $r19
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r19, s[$ofs2+0x0060], $r0, $r19
add.rn.f32 $r18, $r18, $r5
lds.b32 $r6, s[$ofs4+0x0028]
mad.rn.f32 $r20, $r6, $r3, -$r20
mad.rn.f32 $r21, $r6, $r2, $r21
lds.b32 $r7, s[$ofs1+0x0024]
lds.b32 $r6, s[$ofs3+0x0068]
mad.rn.f32 $r21, $r7, $r1, $r21
mad.rn.f32 $r20, $r6, $r2, -$r20
mul.rn.f32 $r5, s[$ofs2+0x0064], $r1
mad.rn.f32 $r21, $r6, $r3, $r21
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r21, s[$ofs2+0x0064], $r0, $r21
add.rn.f32 $r20, $r20, $r5
mov.b64 $r0, g[$r45]
add.u32 $r45, $r4, $r45
mov.b64 $r2, g[$r45]
add.u32 $r45, $r4, $r45
mov.b32 $ofs2, 0x00000580
mov.b32 $ofs3, 0x00000180
lds.b32 $r6, s[$ofs4+0x0074]
mad.rn.f32 $r47, $r6, $r3, -$r47
mad.rn.f32 $r46, $r6, $r2, $r46
lds.b32 $r7, s[$ofs1+0x0070]
lds.b32 $r6, s[$ofs3+0x0034]
mad.rn.f32 $r46, $r7, $r1, $r46
mad.rn.f32 $r47, $r6, $r2, -$r47
mul.rn.f32 $r5, s[$ofs2+0x0030], $r1
mad.rn.f32 $r46, $r6, $r3, $r46
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r46, s[$ofs2+0x0030], $r0, $r46
add.rn.f32 $r47, $r47, $r5
lds.b32 $r6, s[$ofs4+0x0078]
mad.rn.f32 $r44, $r6, $r3, -$r44
mad.rn.f32 $r43, $r6, $r2, $r43
lds.b32 $r7, s[$ofs1+0x0074]
lds.b32 $r6, s[$ofs3+0x0038]
mad.rn.f32 $r43, $r7, $r1, $r43
mad.rn.f32 $r44, $r6, $r2, -$r44
mul.rn.f32 $r5, s[$ofs2+0x0034], $r1
mad.rn.f32 $r43, $r6, $r3, $r43
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r43, s[$ofs2+0x0034], $r0, $r43
add.rn.f32 $r44, $r44, $r5
lds.b32 $r6, s[$ofs4+0x007c]
mad.rn.f32 $r42, $r6, $r3, -$r42
mad.rn.f32 $r41, $r6, $r2, $r41
lds.b32 $r7, s[$ofs1+0x0078]
lds.b32 $r6, s[$ofs3+0x003c]
mad.rn.f32 $r41, $r7, $r1, $r41
mad.rn.f32 $r42, $r6, $r2, -$r42
mul.rn.f32 $r5, s[$ofs2+0x0038], $r1
mad.rn.f32 $r41, $r6, $r3, $r41
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r41, s[$ofs2+0x0038], $r0, $r41
add.rn.f32 $r42, $r42, $r5
mov.b32 $ofs4, 0x00000600
lds.b32 $r6, s[$ofs4+0x0000]
mad.rn.f32 $r36, $r6, $r3, -$r36
mad.rn.f32 $r35, $r6, $r2, $r35
lds.b32 $r7, s[$ofs1+0x007c]
lds.b32 $r6, s[$ofs3+0x0040]
mad.rn.f32 $r35, $r7, $r1, $r35
mad.rn.f32 $r36, $r6, $r2, -$r36
mul.rn.f32 $r5, s[$ofs2+0x003c], $r1
mad.rn.f32 $r35, $r6, $r3, $r35
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r35, s[$ofs2+0x003c], $r0, $r35
add.rn.f32 $r36, $r36, $r5
mov.b32 $ofs1, 0x00000180
lds.b32 $r6, s[$ofs4+0x0004]
mad.rn.f32 $r34, $r6, $r3, -$r34
mad.rn.f32 $r33, $r6, $r2, $r33
lds.b32 $r7, s[$ofs1+0x0000]
lds.b32 $r6, s[$ofs3+0x0044]
mad.rn.f32 $r33, $r7, $r1, $r33
mad.rn.f32 $r34, $r6, $r2, -$r34
mul.rn.f32 $r5, s[$ofs2+0x0040], $r1
mad.rn.f32 $r33, $r6, $r3, $r33
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r33, s[$ofs2+0x0040], $r0, $r33
add.rn.f32 $r34, $r34, $r5
lds.b32 $r6, s[$ofs4+0x0008]
mad.rn.f32 $r32, $r6, $r3, -$r32
mad.rn.f32 $r31, $r6, $r2, $r31
lds.b32 $r7, s[$ofs1+0x0004]
lds.b32 $r6, s[$ofs3+0x0048]
mad.rn.f32 $r31, $r7, $r1, $r31
mad.rn.f32 $r32, $r6, $r2, -$r32
mul.rn.f32 $r5, s[$ofs2+0x0044], $r1
mad.rn.f32 $r31, $r6, $r3, $r31
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r31, s[$ofs2+0x0044], $r0, $r31
add.rn.f32 $r32, $r32, $r5
lds.b32 $r6, s[$ofs4+0x000c]
mad.rn.f32 $r30, $r6, $r3, -$r30
mad.rn.f32 $r29, $r6, $r2, $r29
lds.b32 $r7, s[$ofs1+0x0008]
lds.b32 $r6, s[$ofs3+0x004c]
mad.rn.f32 $r29, $r7, $r1, $r29
mad.rn.f32 $r30, $r6, $r2, -$r30
mul.rn.f32 $r5, s[$ofs2+0x0048], $r1
mad.rn.f32 $r29, $r6, $r3, $r29
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r29, s[$ofs2+0x0048], $r0, $r29
add.rn.f32 $r30, $r30, $r5
lds.b32 $r6, s[$ofs4+0x0010]
mad.rn.f32 $r27, $r6, $r3, -$r27
mad.rn.f32 $r26, $r6, $r2, $r26
lds.b32 $r7, s[$ofs1+0x000c]
lds.b32 $r6, s[$ofs3+0x0050]
mad.rn.f32 $r26, $r7, $r1, $r26
mad.rn.f32 $r27, $r6, $r2, -$r27
mul.rn.f32 $r5, s[$ofs2+0x004c], $r1
mad.rn.f32 $r26, $r6, $r3, $r26
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r26, s[$ofs2+0x004c], $r0, $r26
add.rn.f32 $r27, $r27, $r5
lds.b32 $r6, s[$ofs4+0x0014]
mad.rn.f32 $r25, $r6, $r3, -$r25
mad.rn.f32 $r24, $r6, $r2, $r24
lds.b32 $r7, s[$ofs1+0x0010]
lds.b32 $r6, s[$ofs3+0x0054]
mad.rn.f32 $r24, $r7, $r1, $r24
mad.rn.f32 $r25, $r6, $r2, -$r25
mul.rn.f32 $r5, s[$ofs2+0x0050], $r1
mad.rn.f32 $r24, $r6, $r3, $r24
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r24, s[$ofs2+0x0050], $r0, $r24
add.rn.f32 $r25, $r25, $r5
lds.b32 $r6, s[$ofs4+0x0018]
mad.rn.f32 $r23, $r6, $r3, -$r23
mad.rn.f32 $r22, $r6, $r2, $r22
lds.b32 $r7, s[$ofs1+0x0014]
lds.b32 $r6, s[$ofs3+0x0058]
mad.rn.f32 $r22, $r7, $r1, $r22
mad.rn.f32 $r23, $r6, $r2, -$r23
mul.rn.f32 $r5, s[$ofs2+0x0054], $r1
mad.rn.f32 $r22, $r6, $r3, $r22
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r22, s[$ofs2+0x0054], $r0, $r22
add.rn.f32 $r23, $r23, $r5
lds.b32 $r6, s[$ofs4+0x001c]
mad.rn.f32 $r10, $r6, $r3, -$r10
mad.rn.f32 $r11, $r6, $r2, $r11
lds.b32 $r7, s[$ofs1+0x0018]
lds.b32 $r6, s[$ofs3+0x005c]
mad.rn.f32 $r11, $r7, $r1, $r11
mad.rn.f32 $r10, $r6, $r2, -$r10
mul.rn.f32 $r5, s[$ofs2+0x0058], $r1
mad.rn.f32 $r11, $r6, $r3, $r11
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r11, s[$ofs2+0x0058], $r0, $r11
add.rn.f32 $r10, $r10, $r5
lds.b32 $r6, s[$ofs4+0x0020]
mad.rn.f32 $r12, $r6, $r3, -$r12
mad.rn.f32 $r13, $r6, $r2, $r13
lds.b32 $r7, s[$ofs1+0x001c]
lds.b32 $r6, s[$ofs3+0x0060]
mad.rn.f32 $r13, $r7, $r1, $r13
mad.rn.f32 $r12, $r6, $r2, -$r12
mul.rn.f32 $r5, s[$ofs2+0x005c], $r1
mad.rn.f32 $r13, $r6, $r3, $r13
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r13, s[$ofs2+0x005c], $r0, $r13
add.rn.f32 $r12, $r12, $r5
lds.b32 $r6, s[$ofs4+0x0024]
mad.rn.f32 $r14, $r6, $r3, -$r14
mad.rn.f32 $r15, $r6, $r2, $r15
lds.b32 $r7, s[$ofs1+0x0020]
lds.b32 $r6, s[$ofs3+0x0064]
mad.rn.f32 $r15, $r7, $r1, $r15
mad.rn.f32 $r14, $r6, $r2, -$r14
mul.rn.f32 $r5, s[$ofs2+0x0060], $r1
mad.rn.f32 $r15, $r6, $r3, $r15
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r15, s[$ofs2+0x0060], $r0, $r15
add.rn.f32 $r14, $r14, $r5
lds.b32 $r6, s[$ofs4+0x0028]
mad.rn.f32 $r16, $r6, $r3, -$r16
mad.rn.f32 $r17, $r6, $r2, $r17
lds.b32 $r7, s[$ofs1+0x0024]
lds.b32 $r6, s[$ofs3+0x0068]
mad.rn.f32 $r17, $r7, $r1, $r17
mad.rn.f32 $r16, $r6, $r2, -$r16
mul.rn.f32 $r5, s[$ofs2+0x0064], $r1
mad.rn.f32 $r17, $r6, $r3, $r17
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r17, s[$ofs2+0x0064], $r0, $r17
add.rn.f32 $r16, $r16, $r5
lds.b32 $r6, s[$ofs4+0x002c]
mad.rn.f32 $r18, $r6, $r3, -$r18
mad.rn.f32 $r19, $r6, $r2, $r19
lds.b32 $r7, s[$ofs1+0x0028]
lds.b32 $r6, s[$ofs3+0x006c]
mad.rn.f32 $r19, $r7, $r1, $r19
mad.rn.f32 $r18, $r6, $r2, -$r18
mul.rn.f32 $r5, s[$ofs2+0x0068], $r1
mad.rn.f32 $r19, $r6, $r3, $r19
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r19, s[$ofs2+0x0068], $r0, $r19
add.rn.f32 $r18, $r18, $r5
lds.b32 $r6, s[$ofs4+0x0030]
mad.rn.f32 $r20, $r6, $r3, -$r20
mad.rn.f32 $r21, $r6, $r2, $r21
lds.b32 $r7, s[$ofs1+0x002c]
lds.b32 $r6, s[$ofs3+0x0070]
mad.rn.f32 $r21, $r7, $r1, $r21
mad.rn.f32 $r20, $r6, $r2, -$r20
mul.rn.f32 $r5, s[$ofs2+0x006c], $r1
mad.rn.f32 $r21, $r6, $r3, $r21
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r21, s[$ofs2+0x006c], $r0, $r21
add.rn.f32 $r20, $r20, $r5
mov.b64 $r0, g[$r45]
add.u32 $r45, $r4, $r45
mov.b64 $r2, g[$r45]
add.u32 $r45, $r4, $r45
mov.b32 $ofs2, 0x00000600
mov.b32 $ofs3, 0x00000200
lds.b32 $r6, s[$ofs4+0x007c]
mad.rn.f32 $r47, $r6, $r3, -$r47
mad.rn.f32 $r46, $r6, $r2, $r46
lds.b32 $r7, s[$ofs1+0x0078]
lds.b32 $r6, s[$ofs3+0x003c]
mad.rn.f32 $r46, $r7, $r1, $r46
mad.rn.f32 $r47, $r6, $r2, -$r47
mul.rn.f32 $r5, s[$ofs2+0x0038], $r1
mad.rn.f32 $r46, $r6, $r3, $r46
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r46, s[$ofs2+0x0038], $r0, $r46
add.rn.f32 $r47, $r47, $r5
mov.b32 $ofs4, 0x00000680
lds.b32 $r6, s[$ofs4+0x0000]
mad.rn.f32 $r44, $r6, $r3, -$r44
mad.rn.f32 $r43, $r6, $r2, $r43
lds.b32 $r7, s[$ofs1+0x007c]
lds.b32 $r6, s[$ofs3+0x0040]
mad.rn.f32 $r43, $r7, $r1, $r43
mad.rn.f32 $r44, $r6, $r2, -$r44
mul.rn.f32 $r5, s[$ofs2+0x003c], $r1
mad.rn.f32 $r43, $r6, $r3, $r43
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r43, s[$ofs2+0x003c], $r0, $r43
add.rn.f32 $r44, $r44, $r5
mov.b32 $ofs1, 0x00000200
lds.b32 $r6, s[$ofs4+0x0004]
mad.rn.f32 $r42, $r6, $r3, -$r42
mad.rn.f32 $r41, $r6, $r2, $r41
lds.b32 $r7, s[$ofs1+0x0000]
lds.b32 $r6, s[$ofs3+0x0044]
mad.rn.f32 $r41, $r7, $r1, $r41
mad.rn.f32 $r42, $r6, $r2, -$r42
mul.rn.f32 $r5, s[$ofs2+0x0040], $r1
mad.rn.f32 $r41, $r6, $r3, $r41
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r41, s[$ofs2+0x0040], $r0, $r41
add.rn.f32 $r42, $r42, $r5
lds.b32 $r6, s[$ofs4+0x0008]
mad.rn.f32 $r36, $r6, $r3, -$r36
mad.rn.f32 $r35, $r6, $r2, $r35
lds.b32 $r7, s[$ofs1+0x0004]
lds.b32 $r6, s[$ofs3+0x0048]
mad.rn.f32 $r35, $r7, $r1, $r35
mad.rn.f32 $r36, $r6, $r2, -$r36
mul.rn.f32 $r5, s[$ofs2+0x0044], $r1
mad.rn.f32 $r35, $r6, $r3, $r35
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r35, s[$ofs2+0x0044], $r0, $r35
add.rn.f32 $r36, $r36, $r5
lds.b32 $r6, s[$ofs4+0x000c]
mad.rn.f32 $r34, $r6, $r3, -$r34
mad.rn.f32 $r33, $r6, $r2, $r33
lds.b32 $r7, s[$ofs1+0x0008]
lds.b32 $r6, s[$ofs3+0x004c]
mad.rn.f32 $r33, $r7, $r1, $r33
mad.rn.f32 $r34, $r6, $r2, -$r34
mul.rn.f32 $r5, s[$ofs2+0x0048], $r1
mad.rn.f32 $r33, $r6, $r3, $r33
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r33, s[$ofs2+0x0048], $r0, $r33
add.rn.f32 $r34, $r34, $r5
lds.b32 $r6, s[$ofs4+0x0010]
mad.rn.f32 $r32, $r6, $r3, -$r32
mad.rn.f32 $r31, $r6, $r2, $r31
lds.b32 $r7, s[$ofs1+0x000c]
lds.b32 $r6, s[$ofs3+0x0050]
mad.rn.f32 $r31, $r7, $r1, $r31
mad.rn.f32 $r32, $r6, $r2, -$r32
mul.rn.f32 $r5, s[$ofs2+0x004c], $r1
mad.rn.f32 $r31, $r6, $r3, $r31
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r31, s[$ofs2+0x004c], $r0, $r31
add.rn.f32 $r32, $r32, $r5
lds.b32 $r6, s[$ofs4+0x0014]
mad.rn.f32 $r30, $r6, $r3, -$r30
mad.rn.f32 $r29, $r6, $r2, $r29
lds.b32 $r7, s[$ofs1+0x0010]
lds.b32 $r6, s[$ofs3+0x0054]
mad.rn.f32 $r29, $r7, $r1, $r29
mad.rn.f32 $r30, $r6, $r2, -$r30
mul.rn.f32 $r5, s[$ofs2+0x0050], $r1
mad.rn.f32 $r29, $r6, $r3, $r29
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r29, s[$ofs2+0x0050], $r0, $r29
add.rn.f32 $r30, $r30, $r5
lds.b32 $r6, s[$ofs4+0x0018]
mad.rn.f32 $r27, $r6, $r3, -$r27
mad.rn.f32 $r26, $r6, $r2, $r26
lds.b32 $r7, s[$ofs1+0x0014]
lds.b32 $r6, s[$ofs3+0x0058]
mad.rn.f32 $r26, $r7, $r1, $r26
mad.rn.f32 $r27, $r6, $r2, -$r27
mul.rn.f32 $r5, s[$ofs2+0x0054], $r1
mad.rn.f32 $r26, $r6, $r3, $r26
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r26, s[$ofs2+0x0054], $r0, $r26
add.rn.f32 $r27, $r27, $r5
lds.b32 $r6, s[$ofs4+0x001c]
mad.rn.f32 $r25, $r6, $r3, -$r25
mad.rn.f32 $r24, $r6, $r2, $r24
lds.b32 $r7, s[$ofs1+0x0018]
lds.b32 $r6, s[$ofs3+0x005c]
mad.rn.f32 $r24, $r7, $r1, $r24
mad.rn.f32 $r25, $r6, $r2, -$r25
mul.rn.f32 $r5, s[$ofs2+0x0058], $r1
mad.rn.f32 $r24, $r6, $r3, $r24
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r24, s[$ofs2+0x0058], $r0, $r24
add.rn.f32 $r25, $r25, $r5
lds.b32 $r6, s[$ofs4+0x0020]
mad.rn.f32 $r23, $r6, $r3, -$r23
mad.rn.f32 $r22, $r6, $r2, $r22
lds.b32 $r7, s[$ofs1+0x001c]
lds.b32 $r6, s[$ofs3+0x0060]
mad.rn.f32 $r22, $r7, $r1, $r22
mad.rn.f32 $r23, $r6, $r2, -$r23
mul.rn.f32 $r5, s[$ofs2+0x005c], $r1
mad.rn.f32 $r22, $r6, $r3, $r22
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r22, s[$ofs2+0x005c], $r0, $r22
add.rn.f32 $r23, $r23, $r5
lds.b32 $r6, s[$ofs4+0x0024]
mad.rn.f32 $r10, $r6, $r3, -$r10
mad.rn.f32 $r11, $r6, $r2, $r11
lds.b32 $r7, s[$ofs1+0x0020]
lds.b32 $r6, s[$ofs3+0x0064]
mad.rn.f32 $r11, $r7, $r1, $r11
mad.rn.f32 $r10, $r6, $r2, -$r10
mul.rn.f32 $r5, s[$ofs2+0x0060], $r1
mad.rn.f32 $r11, $r6, $r3, $r11
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r11, s[$ofs2+0x0060], $r0, $r11
add.rn.f32 $r10, $r10, $r5
lds.b32 $r6, s[$ofs4+0x0028]
mad.rn.f32 $r12, $r6, $r3, -$r12
mad.rn.f32 $r13, $r6, $r2, $r13
lds.b32 $r7, s[$ofs1+0x0024]
lds.b32 $r6, s[$ofs3+0x0068]
mad.rn.f32 $r13, $r7, $r1, $r13
mad.rn.f32 $r12, $r6, $r2, -$r12
mul.rn.f32 $r5, s[$ofs2+0x0064], $r1
mad.rn.f32 $r13, $r6, $r3, $r13
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r13, s[$ofs2+0x0064], $r0, $r13
add.rn.f32 $r12, $r12, $r5
lds.b32 $r6, s[$ofs4+0x002c]
mad.rn.f32 $r14, $r6, $r3, -$r14
mad.rn.f32 $r15, $r6, $r2, $r15
lds.b32 $r7, s[$ofs1+0x0028]
lds.b32 $r6, s[$ofs3+0x006c]
mad.rn.f32 $r15, $r7, $r1, $r15
mad.rn.f32 $r14, $r6, $r2, -$r14
mul.rn.f32 $r5, s[$ofs2+0x0068], $r1
mad.rn.f32 $r15, $r6, $r3, $r15
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r15, s[$ofs2+0x0068], $r0, $r15
add.rn.f32 $r14, $r14, $r5
lds.b32 $r6, s[$ofs4+0x0030]
mad.rn.f32 $r16, $r6, $r3, -$r16
mad.rn.f32 $r17, $r6, $r2, $r17
lds.b32 $r7, s[$ofs1+0x002c]
lds.b32 $r6, s[$ofs3+0x0070]
mad.rn.f32 $r17, $r7, $r1, $r17
mad.rn.f32 $r16, $r6, $r2, -$r16
mul.rn.f32 $r5, s[$ofs2+0x006c], $r1
mad.rn.f32 $r17, $r6, $r3, $r17
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r17, s[$ofs2+0x006c], $r0, $r17
add.rn.f32 $r16, $r16, $r5
lds.b32 $r6, s[$ofs4+0x0034]
mad.rn.f32 $r18, $r6, $r3, -$r18
mad.rn.f32 $r19, $r6, $r2, $r19
lds.b32 $r7, s[$ofs1+0x0030]
lds.b32 $r6, s[$ofs3+0x0074]
mad.rn.f32 $r19, $r7, $r1, $r19
mad.rn.f32 $r18, $r6, $r2, -$r18
mul.rn.f32 $r5, s[$ofs2+0x0070], $r1
mad.rn.f32 $r19, $r6, $r3, $r19
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r19, s[$ofs2+0x0070], $r0, $r19
add.rn.f32 $r18, $r18, $r5
lds.b32 $r6, s[$ofs4+0x0038]
mad.rn.f32 $r20, $r6, $r3, -$r20
mad.rn.f32 $r21, $r6, $r2, $r21
lds.b32 $r7, s[$ofs1+0x0034]
lds.b32 $r6, s[$ofs3+0x0078]
mad.rn.f32 $r21, $r7, $r1, $r21
mad.rn.f32 $r20, $r6, $r2, -$r20
mul.rn.f32 $r5, s[$ofs2+0x0074], $r1
mad.rn.f32 $r21, $r6, $r3, $r21
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r21, s[$ofs2+0x0074], $r0, $r21
add.rn.f32 $r20, $r20, $r5
mov.b64 $r0, g[$r45]
add.u32 $r45, $r4, $r45
mov.b64 $r2, g[$r45]
add.u32 $r45, $r4, $r45
mov.b32 $ofs1, 0x00000280
mov.b32 $ofs2, 0x00000680
mov.b32 $ofs3, 0x00000280
mov.b32 $ofs4, 0x00000700
lds.b32 $r6, s[$ofs4+0x0004]
mad.rn.f32 $r47, $r6, $r3, -$r47
mad.rn.f32 $r46, $r6, $r2, $r46
lds.b32 $r7, s[$ofs1+0x0000]
lds.b32 $r6, s[$ofs3+0x0044]
mad.rn.f32 $r46, $r7, $r1, $r46
mad.rn.f32 $r47, $r6, $r2, -$r47
mul.rn.f32 $r5, s[$ofs2+0x0040], $r1
mad.rn.f32 $r46, $r6, $r3, $r46
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r46, s[$ofs2+0x0040], $r0, $r46
add.rn.f32 $r47, $r47, $r5
lds.b32 $r6, s[$ofs4+0x0008]
mad.rn.f32 $r44, $r6, $r3, -$r44
mad.rn.f32 $r43, $r6, $r2, $r43
lds.b32 $r7, s[$ofs1+0x0004]
lds.b32 $r6, s[$ofs3+0x0048]
mad.rn.f32 $r43, $r7, $r1, $r43
mad.rn.f32 $r44, $r6, $r2, -$r44
mul.rn.f32 $r5, s[$ofs2+0x0044], $r1
mad.rn.f32 $r43, $r6, $r3, $r43
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r43, s[$ofs2+0x0044], $r0, $r43
add.rn.f32 $r44, $r44, $r5
lds.b32 $r6, s[$ofs4+0x000c]
mad.rn.f32 $r42, $r6, $r3, -$r42
mad.rn.f32 $r41, $r6, $r2, $r41
lds.b32 $r7, s[$ofs1+0x0008]
lds.b32 $r6, s[$ofs3+0x004c]
mad.rn.f32 $r41, $r7, $r1, $r41
mad.rn.f32 $r42, $r6, $r2, -$r42
mul.rn.f32 $r5, s[$ofs2+0x0048], $r1
mad.rn.f32 $r41, $r6, $r3, $r41
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r41, s[$ofs2+0x0048], $r0, $r41
add.rn.f32 $r42, $r42, $r5
lds.b32 $r6, s[$ofs4+0x0010]
mad.rn.f32 $r36, $r6, $r3, -$r36
mad.rn.f32 $r35, $r6, $r2, $r35
lds.b32 $r7, s[$ofs1+0x000c]
lds.b32 $r6, s[$ofs3+0x0050]
mad.rn.f32 $r35, $r7, $r1, $r35
mad.rn.f32 $r36, $r6, $r2, -$r36
mul.rn.f32 $r5, s[$ofs2+0x004c], $r1
mad.rn.f32 $r35, $r6, $r3, $r35
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r35, s[$ofs2+0x004c], $r0, $r35
add.rn.f32 $r36, $r36, $r5
lds.b32 $r6, s[$ofs4+0x0014]
mad.rn.f32 $r34, $r6, $r3, -$r34
mad.rn.f32 $r33, $r6, $r2, $r33
lds.b32 $r7, s[$ofs1+0x0010]
lds.b32 $r6, s[$ofs3+0x0054]
mad.rn.f32 $r33, $r7, $r1, $r33
mad.rn.f32 $r34, $r6, $r2, -$r34
mul.rn.f32 $r5, s[$ofs2+0x0050], $r1
mad.rn.f32 $r33, $r6, $r3, $r33
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r33, s[$ofs2+0x0050], $r0, $r33
add.rn.f32 $r34, $r34, $r5
lds.b32 $r6, s[$ofs4+0x0018]
mad.rn.f32 $r32, $r6, $r3, -$r32
mad.rn.f32 $r31, $r6, $r2, $r31
lds.b32 $r7, s[$ofs1+0x0014]
lds.b32 $r6, s[$ofs3+0x0058]
mad.rn.f32 $r31, $r7, $r1, $r31
mad.rn.f32 $r32, $r6, $r2, -$r32
mul.rn.f32 $r5, s[$ofs2+0x0054], $r1
mad.rn.f32 $r31, $r6, $r3, $r31
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r31, s[$ofs2+0x0054], $r0, $r31
add.rn.f32 $r32, $r32, $r5
lds.b32 $r6, s[$ofs4+0x001c]
mad.rn.f32 $r30, $r6, $r3, -$r30
mad.rn.f32 $r29, $r6, $r2, $r29
lds.b32 $r7, s[$ofs1+0x0018]
lds.b32 $r6, s[$ofs3+0x005c]
mad.rn.f32 $r29, $r7, $r1, $r29
mad.rn.f32 $r30, $r6, $r2, -$r30
mul.rn.f32 $r5, s[$ofs2+0x0058], $r1
mad.rn.f32 $r29, $r6, $r3, $r29
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r29, s[$ofs2+0x0058], $r0, $r29
add.rn.f32 $r30, $r30, $r5
lds.b32 $r6, s[$ofs4+0x0020]
mad.rn.f32 $r27, $r6, $r3, -$r27
mad.rn.f32 $r26, $r6, $r2, $r26
lds.b32 $r7, s[$ofs1+0x001c]
lds.b32 $r6, s[$ofs3+0x0060]
mad.rn.f32 $r26, $r7, $r1, $r26
mad.rn.f32 $r27, $r6, $r2, -$r27
mul.rn.f32 $r5, s[$ofs2+0x005c], $r1
mad.rn.f32 $r26, $r6, $r3, $r26
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r26, s[$ofs2+0x005c], $r0, $r26
add.rn.f32 $r27, $r27, $r5
lds.b32 $r6, s[$ofs4+0x0024]
mad.rn.f32 $r25, $r6, $r3, -$r25
mad.rn.f32 $r24, $r6, $r2, $r24
lds.b32 $r7, s[$ofs1+0x0020]
lds.b32 $r6, s[$ofs3+0x0064]
mad.rn.f32 $r24, $r7, $r1, $r24
mad.rn.f32 $r25, $r6, $r2, -$r25
mul.rn.f32 $r5, s[$ofs2+0x0060], $r1
mad.rn.f32 $r24, $r6, $r3, $r24
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r24, s[$ofs2+0x0060], $r0, $r24
add.rn.f32 $r25, $r25, $r5
lds.b32 $r6, s[$ofs4+0x0028]
mad.rn.f32 $r23, $r6, $r3, -$r23
mad.rn.f32 $r22, $r6, $r2, $r22
lds.b32 $r7, s[$ofs1+0x0024]
lds.b32 $r6, s[$ofs3+0x0068]
mad.rn.f32 $r22, $r7, $r1, $r22
mad.rn.f32 $r23, $r6, $r2, -$r23
mul.rn.f32 $r5, s[$ofs2+0x0064], $r1
mad.rn.f32 $r22, $r6, $r3, $r22
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r22, s[$ofs2+0x0064], $r0, $r22
add.rn.f32 $r23, $r23, $r5
lds.b32 $r6, s[$ofs4+0x002c]
mad.rn.f32 $r10, $r6, $r3, -$r10
mad.rn.f32 $r11, $r6, $r2, $r11
lds.b32 $r7, s[$ofs1+0x0028]
lds.b32 $r6, s[$ofs3+0x006c]
mad.rn.f32 $r11, $r7, $r1, $r11
mad.rn.f32 $r10, $r6, $r2, -$r10
mul.rn.f32 $r5, s[$ofs2+0x0068], $r1
mad.rn.f32 $r11, $r6, $r3, $r11
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r11, s[$ofs2+0x0068], $r0, $r11
add.rn.f32 $r10, $r10, $r5
lds.b32 $r6, s[$ofs4+0x0030]
mad.rn.f32 $r12, $r6, $r3, -$r12
mad.rn.f32 $r13, $r6, $r2, $r13
lds.b32 $r7, s[$ofs1+0x002c]
lds.b32 $r6, s[$ofs3+0x0070]
mad.rn.f32 $r13, $r7, $r1, $r13
mad.rn.f32 $r12, $r6, $r2, -$r12
mul.rn.f32 $r5, s[$ofs2+0x006c], $r1
mad.rn.f32 $r13, $r6, $r3, $r13
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r13, s[$ofs2+0x006c], $r0, $r13
add.rn.f32 $r12, $r12, $r5
lds.b32 $r6, s[$ofs4+0x0034]
mad.rn.f32 $r14, $r6, $r3, -$r14
mad.rn.f32 $r15, $r6, $r2, $r15
lds.b32 $r7, s[$ofs1+0x0030]
lds.b32 $r6, s[$ofs3+0x0074]
mad.rn.f32 $r15, $r7, $r1, $r15
mad.rn.f32 $r14, $r6, $r2, -$r14
mul.rn.f32 $r5, s[$ofs2+0x0070], $r1
mad.rn.f32 $r15, $r6, $r3, $r15
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r15, s[$ofs2+0x0070], $r0, $r15
add.rn.f32 $r14, $r14, $r5
lds.b32 $r6, s[$ofs4+0x0038]
mad.rn.f32 $r16, $r6, $r3, -$r16
mad.rn.f32 $r17, $r6, $r2, $r17
lds.b32 $r7, s[$ofs1+0x0034]
lds.b32 $r6, s[$ofs3+0x0078]
mad.rn.f32 $r17, $r7, $r1, $r17
mad.rn.f32 $r16, $r6, $r2, -$r16
mul.rn.f32 $r5, s[$ofs2+0x0074], $r1
mad.rn.f32 $r17, $r6, $r3, $r17
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r17, s[$ofs2+0x0074], $r0, $r17
add.rn.f32 $r16, $r16, $r5
lds.b32 $r6, s[$ofs4+0x003c]
mad.rn.f32 $r18, $r6, $r3, -$r18
mad.rn.f32 $r19, $r6, $r2, $r19
lds.b32 $r7, s[$ofs1+0x0038]
lds.b32 $r6, s[$ofs3+0x007c]
mad.rn.f32 $r19, $r7, $r1, $r19
mad.rn.f32 $r18, $r6, $r2, -$r18
mul.rn.f32 $r5, s[$ofs2+0x0078], $r1
mad.rn.f32 $r19, $r6, $r3, $r19
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r19, s[$ofs2+0x0078], $r0, $r19
add.rn.f32 $r18, $r18, $r5
mov.b32 $ofs3, 0x00000300
lds.b32 $r6, s[$ofs4+0x0040]
mad.rn.f32 $r20, $r6, $r3, -$r20
mad.rn.f32 $r21, $r6, $r2, $r21
lds.b32 $r7, s[$ofs1+0x003c]
lds.b32 $r6, s[$ofs3+0x0000]
mad.rn.f32 $r21, $r7, $r1, $r21
mad.rn.f32 $r20, $r6, $r2, -$r20
mul.rn.f32 $r5, s[$ofs2+0x007c], $r1
mad.rn.f32 $r21, $r6, $r3, $r21
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r21, s[$ofs2+0x007c], $r0, $r21
add.rn.f32 $r20, $r20, $r5
mov.b64 $r0, g[$r45]
add.u32 $r45, $r4, $r45
mov.b64 $r2, g[$r45]
add.u32 $r45, $r4, $r45
mov.b32 $ofs1, 0x00000300
mov.b32 $ofs2, 0x00000700
mov.b32 $ofs4, 0x00000780
lds.b32 $r6, s[$ofs4+0x000c]
mad.rn.f32 $r47, $r6, $r3, -$r47
mad.rn.f32 $r46, $r6, $r2, $r46
lds.b32 $r7, s[$ofs1+0x0008]
lds.b32 $r6, s[$ofs3+0x004c]
mad.rn.f32 $r46, $r7, $r1, $r46
mad.rn.f32 $r47, $r6, $r2, -$r47
mul.rn.f32 $r5, s[$ofs2+0x0048], $r1
mad.rn.f32 $r46, $r6, $r3, $r46
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r46, s[$ofs2+0x0048], $r0, $r46
add.rn.f32 $r47, $r47, $r5
lds.b32 $r6, s[$ofs4+0x0010]
mad.rn.f32 $r44, $r6, $r3, -$r44
mad.rn.f32 $r43, $r6, $r2, $r43
lds.b32 $r7, s[$ofs1+0x000c]
lds.b32 $r6, s[$ofs3+0x0050]
mad.rn.f32 $r43, $r7, $r1, $r43
mad.rn.f32 $r44, $r6, $r2, -$r44
mul.rn.f32 $r5, s[$ofs2+0x004c], $r1
mad.rn.f32 $r43, $r6, $r3, $r43
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r43, s[$ofs2+0x004c], $r0, $r43
add.rn.f32 $r44, $r44, $r5
lds.b32 $r6, s[$ofs4+0x0014]
mad.rn.f32 $r42, $r6, $r3, -$r42
mad.rn.f32 $r41, $r6, $r2, $r41
lds.b32 $r7, s[$ofs1+0x0010]
lds.b32 $r6, s[$ofs3+0x0054]
mad.rn.f32 $r41, $r7, $r1, $r41
mad.rn.f32 $r42, $r6, $r2, -$r42
mul.rn.f32 $r5, s[$ofs2+0x0050], $r1
mad.rn.f32 $r41, $r6, $r3, $r41
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r41, s[$ofs2+0x0050], $r0, $r41
add.rn.f32 $r42, $r42, $r5
lds.b32 $r6, s[$ofs4+0x0018]
mad.rn.f32 $r36, $r6, $r3, -$r36
mad.rn.f32 $r35, $r6, $r2, $r35
lds.b32 $r7, s[$ofs1+0x0014]
lds.b32 $r6, s[$ofs3+0x0058]
mad.rn.f32 $r35, $r7, $r1, $r35
mad.rn.f32 $r36, $r6, $r2, -$r36
mul.rn.f32 $r5, s[$ofs2+0x0054], $r1
mad.rn.f32 $r35, $r6, $r3, $r35
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r35, s[$ofs2+0x0054], $r0, $r35
add.rn.f32 $r36, $r36, $r5
lds.b32 $r6, s[$ofs4+0x001c]
mad.rn.f32 $r34, $r6, $r3, -$r34
mad.rn.f32 $r33, $r6, $r2, $r33
lds.b32 $r7, s[$ofs1+0x0018]
lds.b32 $r6, s[$ofs3+0x005c]
mad.rn.f32 $r33, $r7, $r1, $r33
mad.rn.f32 $r34, $r6, $r2, -$r34
mul.rn.f32 $r5, s[$ofs2+0x0058], $r1
mad.rn.f32 $r33, $r6, $r3, $r33
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r33, s[$ofs2+0x0058], $r0, $r33
add.rn.f32 $r34, $r34, $r5
lds.b32 $r6, s[$ofs4+0x0020]
mad.rn.f32 $r32, $r6, $r3, -$r32
mad.rn.f32 $r31, $r6, $r2, $r31
lds.b32 $r7, s[$ofs1+0x001c]
lds.b32 $r6, s[$ofs3+0x0060]
mad.rn.f32 $r31, $r7, $r1, $r31
mad.rn.f32 $r32, $r6, $r2, -$r32
mul.rn.f32 $r5, s[$ofs2+0x005c], $r1
mad.rn.f32 $r31, $r6, $r3, $r31
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r31, s[$ofs2+0x005c], $r0, $r31
add.rn.f32 $r32, $r32, $r5
lds.b32 $r6, s[$ofs4+0x0024]
mad.rn.f32 $r30, $r6, $r3, -$r30
mad.rn.f32 $r29, $r6, $r2, $r29
lds.b32 $r7, s[$ofs1+0x0020]
lds.b32 $r6, s[$ofs3+0x0064]
mad.rn.f32 $r29, $r7, $r1, $r29
mad.rn.f32 $r30, $r6, $r2, -$r30
mul.rn.f32 $r5, s[$ofs2+0x0060], $r1
mad.rn.f32 $r29, $r6, $r3, $r29
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r29, s[$ofs2+0x0060], $r0, $r29
add.rn.f32 $r30, $r30, $r5
lds.b32 $r6, s[$ofs4+0x0028]
mad.rn.f32 $r27, $r6, $r3, -$r27
mad.rn.f32 $r26, $r6, $r2, $r26
lds.b32 $r7, s[$ofs1+0x0024]
lds.b32 $r6, s[$ofs3+0x0068]
mad.rn.f32 $r26, $r7, $r1, $r26
mad.rn.f32 $r27, $r6, $r2, -$r27
mul.rn.f32 $r5, s[$ofs2+0x0064], $r1
mad.rn.f32 $r26, $r6, $r3, $r26
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r26, s[$ofs2+0x0064], $r0, $r26
add.rn.f32 $r27, $r27, $r5
lds.b32 $r6, s[$ofs4+0x002c]
mad.rn.f32 $r25, $r6, $r3, -$r25
mad.rn.f32 $r24, $r6, $r2, $r24
lds.b32 $r7, s[$ofs1+0x0028]
lds.b32 $r6, s[$ofs3+0x006c]
mad.rn.f32 $r24, $r7, $r1, $r24
mad.rn.f32 $r25, $r6, $r2, -$r25
mul.rn.f32 $r5, s[$ofs2+0x0068], $r1
mad.rn.f32 $r24, $r6, $r3, $r24
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r24, s[$ofs2+0x0068], $r0, $r24
add.rn.f32 $r25, $r25, $r5
lds.b32 $r6, s[$ofs4+0x0030]
mad.rn.f32 $r23, $r6, $r3, -$r23
mad.rn.f32 $r22, $r6, $r2, $r22
lds.b32 $r7, s[$ofs1+0x002c]
lds.b32 $r6, s[$ofs3+0x0070]
mad.rn.f32 $r22, $r7, $r1, $r22
mad.rn.f32 $r23, $r6, $r2, -$r23
mul.rn.f32 $r5, s[$ofs2+0x006c], $r1
mad.rn.f32 $r22, $r6, $r3, $r22
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r22, s[$ofs2+0x006c], $r0, $r22
add.rn.f32 $r23, $r23, $r5
lds.b32 $r6, s[$ofs4+0x0034]
mad.rn.f32 $r10, $r6, $r3, -$r10
mad.rn.f32 $r11, $r6, $r2, $r11
lds.b32 $r7, s[$ofs1+0x0030]
lds.b32 $r6, s[$ofs3+0x0074]
mad.rn.f32 $r11, $r7, $r1, $r11
mad.rn.f32 $r10, $r6, $r2, -$r10
mul.rn.f32 $r5, s[$ofs2+0x0070], $r1
mad.rn.f32 $r11, $r6, $r3, $r11
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r11, s[$ofs2+0x0070], $r0, $r11
add.rn.f32 $r10, $r10, $r5
lds.b32 $r6, s[$ofs4+0x0038]
mad.rn.f32 $r12, $r6, $r3, -$r12
mad.rn.f32 $r13, $r6, $r2, $r13
lds.b32 $r7, s[$ofs1+0x0034]
lds.b32 $r6, s[$ofs3+0x0078]
mad.rn.f32 $r13, $r7, $r1, $r13
mad.rn.f32 $r12, $r6, $r2, -$r12
mul.rn.f32 $r5, s[$ofs2+0x0074], $r1
mad.rn.f32 $r13, $r6, $r3, $r13
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r13, s[$ofs2+0x0074], $r0, $r13
add.rn.f32 $r12, $r12, $r5
lds.b32 $r6, s[$ofs4+0x003c]
mad.rn.f32 $r14, $r6, $r3, -$r14
mad.rn.f32 $r15, $r6, $r2, $r15
lds.b32 $r7, s[$ofs1+0x0038]
lds.b32 $r6, s[$ofs3+0x007c]
mad.rn.f32 $r15, $r7, $r1, $r15
mad.rn.f32 $r14, $r6, $r2, -$r14
mul.rn.f32 $r5, s[$ofs2+0x0078], $r1
mad.rn.f32 $r15, $r6, $r3, $r15
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r15, s[$ofs2+0x0078], $r0, $r15
add.rn.f32 $r14, $r14, $r5
mov.b32 $ofs3, 0x00000380
lds.b32 $r6, s[$ofs4+0x0040]
mad.rn.f32 $r16, $r6, $r3, -$r16
mad.rn.f32 $r17, $r6, $r2, $r17
lds.b32 $r7, s[$ofs1+0x003c]
lds.b32 $r6, s[$ofs3+0x0000]
mad.rn.f32 $r17, $r7, $r1, $r17
mad.rn.f32 $r16, $r6, $r2, -$r16
mul.rn.f32 $r5, s[$ofs2+0x007c], $r1
mad.rn.f32 $r17, $r6, $r3, $r17
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r17, s[$ofs2+0x007c], $r0, $r17
add.rn.f32 $r16, $r16, $r5
mov.b32 $ofs2, 0x00000780
lds.b32 $r6, s[$ofs4+0x0044]
mad.rn.f32 $r18, $r6, $r3, -$r18
mad.rn.f32 $r19, $r6, $r2, $r19
lds.b32 $r7, s[$ofs1+0x0040]
lds.b32 $r6, s[$ofs3+0x0004]
mad.rn.f32 $r19, $r7, $r1, $r19
mad.rn.f32 $r18, $r6, $r2, -$r18
mul.rn.f32 $r5, s[$ofs2+0x0000], $r1
mad.rn.f32 $r19, $r6, $r3, $r19
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r19, s[$ofs2+0x0000], $r0, $r19
add.rn.f32 $r18, $r18, $r5
lds.b32 $r6, s[$ofs4+0x0048]
mad.rn.f32 $r20, $r6, $r3, -$r20
mad.rn.f32 $r21, $r6, $r2, $r21
lds.b32 $r7, s[$ofs1+0x0044]
lds.b32 $r6, s[$ofs3+0x0008]
mad.rn.f32 $r21, $r7, $r1, $r21
mad.rn.f32 $r20, $r6, $r2, -$r20
mul.rn.f32 $r5, s[$ofs2+0x0004], $r1
mad.rn.f32 $r21, $r6, $r3, $r21
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r21, s[$ofs2+0x0004], $r0, $r21
add.rn.f32 $r20, $r20, $r5
mov.b64 $r0, g[$r45]
add.u32 $r45, $r4, $r45
mov.b64 $r2, g[$r45]
add.u32 $r45, $r4, $r45
mov.b32 $ofs1, 0x00000380
mov.b32 $ofs4, 0x00000800
lds.b32 $r6, s[$ofs4+0x0014]
mad.rn.f32 $r47, $r6, $r3, -$r47
mad.rn.f32 $r46, $r6, $r2, $r46
lds.b32 $r7, s[$ofs1+0x0010]
lds.b32 $r6, s[$ofs3+0x0054]
mad.rn.f32 $r46, $r7, $r1, $r46
mad.rn.f32 $r47, $r6, $r2, -$r47
mul.rn.f32 $r5, s[$ofs2+0x0050], $r1
mad.rn.f32 $r46, $r6, $r3, $r46
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r46, s[$ofs2+0x0050], $r0, $r46
add.rn.f32 $r47, $r47, $r5
lds.b32 $r6, s[$ofs4+0x0018]
mad.rn.f32 $r44, $r6, $r3, -$r44
mad.rn.f32 $r43, $r6, $r2, $r43
lds.b32 $r7, s[$ofs1+0x0014]
lds.b32 $r6, s[$ofs3+0x0058]
mad.rn.f32 $r43, $r7, $r1, $r43
mad.rn.f32 $r44, $r6, $r2, -$r44
mul.rn.f32 $r5, s[$ofs2+0x0054], $r1
mad.rn.f32 $r43, $r6, $r3, $r43
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r43, s[$ofs2+0x0054], $r0, $r43
add.rn.f32 $r44, $r44, $r5
lds.b32 $r6, s[$ofs4+0x001c]
mad.rn.f32 $r42, $r6, $r3, -$r42
mad.rn.f32 $r41, $r6, $r2, $r41
lds.b32 $r7, s[$ofs1+0x0018]
lds.b32 $r6, s[$ofs3+0x005c]
mad.rn.f32 $r41, $r7, $r1, $r41
mad.rn.f32 $r42, $r6, $r2, -$r42
mul.rn.f32 $r5, s[$ofs2+0x0058], $r1
mad.rn.f32 $r41, $r6, $r3, $r41
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r41, s[$ofs2+0x0058], $r0, $r41
add.rn.f32 $r42, $r42, $r5
lds.b32 $r6, s[$ofs4+0x0020]
mad.rn.f32 $r36, $r6, $r3, -$r36
mad.rn.f32 $r35, $r6, $r2, $r35
lds.b32 $r7, s[$ofs1+0x001c]
lds.b32 $r6, s[$ofs3+0x0060]
mad.rn.f32 $r35, $r7, $r1, $r35
mad.rn.f32 $r36, $r6, $r2, -$r36
mul.rn.f32 $r5, s[$ofs2+0x005c], $r1
mad.rn.f32 $r35, $r6, $r3, $r35
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r35, s[$ofs2+0x005c], $r0, $r35
add.rn.f32 $r36, $r36, $r5
lds.b32 $r6, s[$ofs4+0x0024]
mad.rn.f32 $r34, $r6, $r3, -$r34
mad.rn.f32 $r33, $r6, $r2, $r33
lds.b32 $r7, s[$ofs1+0x0020]
lds.b32 $r6, s[$ofs3+0x0064]
mad.rn.f32 $r33, $r7, $r1, $r33
mad.rn.f32 $r34, $r6, $r2, -$r34
mul.rn.f32 $r5, s[$ofs2+0x0060], $r1
mad.rn.f32 $r33, $r6, $r3, $r33
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r33, s[$ofs2+0x0060], $r0, $r33
add.rn.f32 $r34, $r34, $r5
lds.b32 $r6, s[$ofs4+0x0028]
mad.rn.f32 $r32, $r6, $r3, -$r32
mad.rn.f32 $r31, $r6, $r2, $r31
lds.b32 $r7, s[$ofs1+0x0024]
lds.b32 $r6, s[$ofs3+0x0068]
mad.rn.f32 $r31, $r7, $r1, $r31
mad.rn.f32 $r32, $r6, $r2, -$r32
mul.rn.f32 $r5, s[$ofs2+0x0064], $r1
mad.rn.f32 $r31, $r6, $r3, $r31
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r31, s[$ofs2+0x0064], $r0, $r31
add.rn.f32 $r32, $r32, $r5
lds.b32 $r6, s[$ofs4+0x002c]
mad.rn.f32 $r30, $r6, $r3, -$r30
mad.rn.f32 $r29, $r6, $r2, $r29
lds.b32 $r7, s[$ofs1+0x0028]
lds.b32 $r6, s[$ofs3+0x006c]
mad.rn.f32 $r29, $r7, $r1, $r29
mad.rn.f32 $r30, $r6, $r2, -$r30
mul.rn.f32 $r5, s[$ofs2+0x0068], $r1
mad.rn.f32 $r29, $r6, $r3, $r29
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r29, s[$ofs2+0x0068], $r0, $r29
add.rn.f32 $r30, $r30, $r5
lds.b32 $r6, s[$ofs4+0x0030]
mad.rn.f32 $r27, $r6, $r3, -$r27
mad.rn.f32 $r26, $r6, $r2, $r26
lds.b32 $r7, s[$ofs1+0x002c]
lds.b32 $r6, s[$ofs3+0x0070]
mad.rn.f32 $r26, $r7, $r1, $r26
mad.rn.f32 $r27, $r6, $r2, -$r27
mul.rn.f32 $r5, s[$ofs2+0x006c], $r1
mad.rn.f32 $r26, $r6, $r3, $r26
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r26, s[$ofs2+0x006c], $r0, $r26
add.rn.f32 $r27, $r27, $r5
lds.b32 $r6, s[$ofs4+0x0034]
mad.rn.f32 $r25, $r6, $r3, -$r25
mad.rn.f32 $r24, $r6, $r2, $r24
lds.b32 $r7, s[$ofs1+0x0030]
lds.b32 $r6, s[$ofs3+0x0074]
mad.rn.f32 $r24, $r7, $r1, $r24
mad.rn.f32 $r25, $r6, $r2, -$r25
mul.rn.f32 $r5, s[$ofs2+0x0070], $r1
mad.rn.f32 $r24, $r6, $r3, $r24
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r24, s[$ofs2+0x0070], $r0, $r24
add.rn.f32 $r25, $r25, $r5
lds.b32 $r6, s[$ofs4+0x0038]
mad.rn.f32 $r23, $r6, $r3, -$r23
mad.rn.f32 $r22, $r6, $r2, $r22
lds.b32 $r7, s[$ofs1+0x0034]
lds.b32 $r6, s[$ofs3+0x0078]
mad.rn.f32 $r22, $r7, $r1, $r22
mad.rn.f32 $r23, $r6, $r2, -$r23
mul.rn.f32 $r5, s[$ofs2+0x0074], $r1
mad.rn.f32 $r22, $r6, $r3, $r22
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r22, s[$ofs2+0x0074], $r0, $r22
add.rn.f32 $r23, $r23, $r5
lds.b32 $r6, s[$ofs4+0x003c]
mad.rn.f32 $r10, $r6, $r3, -$r10
mad.rn.f32 $r11, $r6, $r2, $r11
lds.b32 $r7, s[$ofs1+0x0038]
lds.b32 $r6, s[$ofs3+0x007c]
mad.rn.f32 $r11, $r7, $r1, $r11
mad.rn.f32 $r10, $r6, $r2, -$r10
mul.rn.f32 $r5, s[$ofs2+0x0078], $r1
mad.rn.f32 $r11, $r6, $r3, $r11
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r11, s[$ofs2+0x0078], $r0, $r11
add.rn.f32 $r10, $r10, $r5
mov.b32 $ofs3, 0x00000400
lds.b32 $r6, s[$ofs4+0x0040]
mad.rn.f32 $r12, $r6, $r3, -$r12
mad.rn.f32 $r13, $r6, $r2, $r13
lds.b32 $r7, s[$ofs1+0x003c]
lds.b32 $r6, s[$ofs3+0x0000]
mad.rn.f32 $r13, $r7, $r1, $r13
mad.rn.f32 $r12, $r6, $r2, -$r12
mul.rn.f32 $r5, s[$ofs2+0x007c], $r1
mad.rn.f32 $r13, $r6, $r3, $r13
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r13, s[$ofs2+0x007c], $r0, $r13
add.rn.f32 $r12, $r12, $r5
mov.b32 $ofs2, 0x00000800
lds.b32 $r6, s[$ofs4+0x0044]
mad.rn.f32 $r14, $r6, $r3, -$r14
mad.rn.f32 $r15, $r6, $r2, $r15
lds.b32 $r7, s[$ofs1+0x0040]
lds.b32 $r6, s[$ofs3+0x0004]
mad.rn.f32 $r15, $r7, $r1, $r15
mad.rn.f32 $r14, $r6, $r2, -$r14
mul.rn.f32 $r5, s[$ofs2+0x0000], $r1
mad.rn.f32 $r15, $r6, $r3, $r15
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r15, s[$ofs2+0x0000], $r0, $r15
add.rn.f32 $r14, $r14, $r5
lds.b32 $r6, s[$ofs4+0x0048]
mad.rn.f32 $r16, $r6, $r3, -$r16
mad.rn.f32 $r17, $r6, $r2, $r17
lds.b32 $r7, s[$ofs1+0x0044]
lds.b32 $r6, s[$ofs3+0x0008]
mad.rn.f32 $r17, $r7, $r1, $r17
mad.rn.f32 $r16, $r6, $r2, -$r16
mul.rn.f32 $r5, s[$ofs2+0x0004], $r1
mad.rn.f32 $r17, $r6, $r3, $r17
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r17, s[$ofs2+0x0004], $r0, $r17
add.rn.f32 $r16, $r16, $r5
lds.b32 $r6, s[$ofs4+0x004c]
mad.rn.f32 $r18, $r6, $r3, -$r18
mad.rn.f32 $r19, $r6, $r2, $r19
lds.b32 $r7, s[$ofs1+0x0048]
lds.b32 $r6, s[$ofs3+0x000c]
mad.rn.f32 $r19, $r7, $r1, $r19
mad.rn.f32 $r18, $r6, $r2, -$r18
mul.rn.f32 $r5, s[$ofs2+0x0008], $r1
mad.rn.f32 $r19, $r6, $r3, $r19
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r19, s[$ofs2+0x0008], $r0, $r19
add.rn.f32 $r18, $r18, $r5
lds.b32 $r6, s[$ofs4+0x0050]
mad.rn.f32 $r20, $r6, $r3, -$r20
mad.rn.f32 $r21, $r6, $r2, $r21
lds.b32 $r7, s[$ofs1+0x004c]
lds.b32 $r6, s[$ofs3+0x0010]
mad.rn.f32 $r21, $r7, $r1, $r21
mad.rn.f32 $r20, $r6, $r2, -$r20
mul.rn.f32 $r5, s[$ofs2+0x000c], $r1
mad.rn.f32 $r21, $r6, $r3, $r21
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r21, s[$ofs2+0x000c], $r0, $r21
add.rn.f32 $r20, $r20, $r5
mov.b64 $r0, g[$r45]
add.u32 $r45, $r4, $r45
mov.b64 $r2, g[$r45]
add.u32 $r45, $r4, $r45
mov.b32 $ofs1, 0x00000400
mov.b32 $ofs4, 0x00000880
lds.b32 $r6, s[$ofs4+0x001c]
mad.rn.f32 $r47, $r6, $r3, -$r47
mad.rn.f32 $r46, $r6, $r2, $r46
lds.b32 $r7, s[$ofs1+0x0018]
lds.b32 $r6, s[$ofs3+0x005c]
mad.rn.f32 $r46, $r7, $r1, $r46
mad.rn.f32 $r47, $r6, $r2, -$r47
mul.rn.f32 $r5, s[$ofs2+0x0058], $r1
mad.rn.f32 $r46, $r6, $r3, $r46
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r46, s[$ofs2+0x0058], $r0, $r46
add.rn.f32 $r47, $r47, $r5
lds.b32 $r6, s[$ofs4+0x0020]
mad.rn.f32 $r44, $r6, $r3, -$r44
mad.rn.f32 $r43, $r6, $r2, $r43
lds.b32 $r7, s[$ofs1+0x001c]
lds.b32 $r6, s[$ofs3+0x0060]
mad.rn.f32 $r43, $r7, $r1, $r43
mad.rn.f32 $r44, $r6, $r2, -$r44
mul.rn.f32 $r5, s[$ofs2+0x005c], $r1
mad.rn.f32 $r43, $r6, $r3, $r43
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r43, s[$ofs2+0x005c], $r0, $r43
add.rn.f32 $r44, $r44, $r5
lds.b32 $r6, s[$ofs4+0x0024]
mad.rn.f32 $r42, $r6, $r3, -$r42
mad.rn.f32 $r41, $r6, $r2, $r41
lds.b32 $r7, s[$ofs1+0x0020]
lds.b32 $r6, s[$ofs3+0x0064]
mad.rn.f32 $r41, $r7, $r1, $r41
mad.rn.f32 $r42, $r6, $r2, -$r42
mul.rn.f32 $r5, s[$ofs2+0x0060], $r1
mad.rn.f32 $r41, $r6, $r3, $r41
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r41, s[$ofs2+0x0060], $r0, $r41
add.rn.f32 $r42, $r42, $r5
lds.b32 $r6, s[$ofs4+0x0028]
mad.rn.f32 $r36, $r6, $r3, -$r36
mad.rn.f32 $r35, $r6, $r2, $r35
lds.b32 $r7, s[$ofs1+0x0024]
lds.b32 $r6, s[$ofs3+0x0068]
mad.rn.f32 $r35, $r7, $r1, $r35
mad.rn.f32 $r36, $r6, $r2, -$r36
mul.rn.f32 $r5, s[$ofs2+0x0064], $r1
mad.rn.f32 $r35, $r6, $r3, $r35
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r35, s[$ofs2+0x0064], $r0, $r35
add.rn.f32 $r36, $r36, $r5
lds.b32 $r6, s[$ofs4+0x002c]
mad.rn.f32 $r34, $r6, $r3, -$r34
mad.rn.f32 $r33, $r6, $r2, $r33
lds.b32 $r7, s[$ofs1+0x0028]
lds.b32 $r6, s[$ofs3+0x006c]
mad.rn.f32 $r33, $r7, $r1, $r33
mad.rn.f32 $r34, $r6, $r2, -$r34
mul.rn.f32 $r5, s[$ofs2+0x0068], $r1
mad.rn.f32 $r33, $r6, $r3, $r33
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r33, s[$ofs2+0x0068], $r0, $r33
add.rn.f32 $r34, $r34, $r5
lds.b32 $r6, s[$ofs4+0x0030]
mad.rn.f32 $r32, $r6, $r3, -$r32
mad.rn.f32 $r31, $r6, $r2, $r31
lds.b32 $r7, s[$ofs1+0x002c]
lds.b32 $r6, s[$ofs3+0x0070]
mad.rn.f32 $r31, $r7, $r1, $r31
mad.rn.f32 $r32, $r6, $r2, -$r32
mul.rn.f32 $r5, s[$ofs2+0x006c], $r1
mad.rn.f32 $r31, $r6, $r3, $r31
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r31, s[$ofs2+0x006c], $r0, $r31
add.rn.f32 $r32, $r32, $r5
lds.b32 $r6, s[$ofs4+0x0034]
mad.rn.f32 $r30, $r6, $r3, -$r30
mad.rn.f32 $r29, $r6, $r2, $r29
lds.b32 $r7, s[$ofs1+0x0030]
lds.b32 $r6, s[$ofs3+0x0074]
mad.rn.f32 $r29, $r7, $r1, $r29
mad.rn.f32 $r30, $r6, $r2, -$r30
mul.rn.f32 $r5, s[$ofs2+0x0070], $r1
mad.rn.f32 $r29, $r6, $r3, $r29
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r29, s[$ofs2+0x0070], $r0, $r29
add.rn.f32 $r30, $r30, $r5
lds.b32 $r6, s[$ofs4+0x0038]
mad.rn.f32 $r27, $r6, $r3, -$r27
mad.rn.f32 $r26, $r6, $r2, $r26
lds.b32 $r7, s[$ofs1+0x0034]
lds.b32 $r6, s[$ofs3+0x0078]
mad.rn.f32 $r26, $r7, $r1, $r26
mad.rn.f32 $r27, $r6, $r2, -$r27
mul.rn.f32 $r5, s[$ofs2+0x0074], $r1
mad.rn.f32 $r26, $r6, $r3, $r26
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r26, s[$ofs2+0x0074], $r0, $r26
add.rn.f32 $r27, $r27, $r5
lds.b32 $r6, s[$ofs4+0x003c]
mad.rn.f32 $r25, $r6, $r3, -$r25
mad.rn.f32 $r24, $r6, $r2, $r24
lds.b32 $r7, s[$ofs1+0x0038]
lds.b32 $r6, s[$ofs3+0x007c]
mad.rn.f32 $r24, $r7, $r1, $r24
mad.rn.f32 $r25, $r6, $r2, -$r25
mul.rn.f32 $r5, s[$ofs2+0x0078], $r1
mad.rn.f32 $r24, $r6, $r3, $r24
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r24, s[$ofs2+0x0078], $r0, $r24
add.rn.f32 $r25, $r25, $r5
mov.b32 $ofs3, 0x00000480
lds.b32 $r6, s[$ofs4+0x0040]
mad.rn.f32 $r23, $r6, $r3, -$r23
mad.rn.f32 $r22, $r6, $r2, $r22
lds.b32 $r7, s[$ofs1+0x003c]
lds.b32 $r6, s[$ofs3+0x0000]
mad.rn.f32 $r22, $r7, $r1, $r22
mad.rn.f32 $r23, $r6, $r2, -$r23
mul.rn.f32 $r5, s[$ofs2+0x007c], $r1
mad.rn.f32 $r22, $r6, $r3, $r22
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r22, s[$ofs2+0x007c], $r0, $r22
add.rn.f32 $r23, $r23, $r5
mov.b32 $ofs2, 0x00000880
lds.b32 $r6, s[$ofs4+0x0044]
mad.rn.f32 $r10, $r6, $r3, -$r10
mad.rn.f32 $r11, $r6, $r2, $r11
lds.b32 $r7, s[$ofs1+0x0040]
lds.b32 $r6, s[$ofs3+0x0004]
mad.rn.f32 $r11, $r7, $r1, $r11
mad.rn.f32 $r10, $r6, $r2, -$r10
mul.rn.f32 $r5, s[$ofs2+0x0000], $r1
mad.rn.f32 $r11, $r6, $r3, $r11
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r11, s[$ofs2+0x0000], $r0, $r11
add.rn.f32 $r10, $r10, $r5
lds.b32 $r6, s[$ofs4+0x0048]
mad.rn.f32 $r12, $r6, $r3, -$r12
mad.rn.f32 $r13, $r6, $r2, $r13
lds.b32 $r7, s[$ofs1+0x0044]
lds.b32 $r6, s[$ofs3+0x0008]
mad.rn.f32 $r13, $r7, $r1, $r13
mad.rn.f32 $r12, $r6, $r2, -$r12
mul.rn.f32 $r5, s[$ofs2+0x0004], $r1
mad.rn.f32 $r13, $r6, $r3, $r13
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r13, s[$ofs2+0x0004], $r0, $r13
add.rn.f32 $r12, $r12, $r5
lds.b32 $r6, s[$ofs4+0x004c]
mad.rn.f32 $r14, $r6, $r3, -$r14
mad.rn.f32 $r15, $r6, $r2, $r15
lds.b32 $r7, s[$ofs1+0x0048]
lds.b32 $r6, s[$ofs3+0x000c]
mad.rn.f32 $r15, $r7, $r1, $r15
mad.rn.f32 $r14, $r6, $r2, -$r14
mul.rn.f32 $r5, s[$ofs2+0x0008], $r1
mad.rn.f32 $r15, $r6, $r3, $r15
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r15, s[$ofs2+0x0008], $r0, $r15
add.rn.f32 $r14, $r14, $r5
lds.b32 $r6, s[$ofs4+0x0050]
mad.rn.f32 $r16, $r6, $r3, -$r16
mad.rn.f32 $r17, $r6, $r2, $r17
lds.b32 $r7, s[$ofs1+0x004c]
lds.b32 $r6, s[$ofs3+0x0010]
mad.rn.f32 $r17, $r7, $r1, $r17
mad.rn.f32 $r16, $r6, $r2, -$r16
mul.rn.f32 $r5, s[$ofs2+0x000c], $r1
mad.rn.f32 $r17, $r6, $r3, $r17
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r17, s[$ofs2+0x000c], $r0, $r17
add.rn.f32 $r16, $r16, $r5
lds.b32 $r6, s[$ofs4+0x0054]
mad.rn.f32 $r18, $r6, $r3, -$r18
mad.rn.f32 $r19, $r6, $r2, $r19
lds.b32 $r7, s[$ofs1+0x0050]
lds.b32 $r6, s[$ofs3+0x0014]
mad.rn.f32 $r19, $r7, $r1, $r19
mad.rn.f32 $r18, $r6, $r2, -$r18
mul.rn.f32 $r5, s[$ofs2+0x0010], $r1
mad.rn.f32 $r19, $r6, $r3, $r19
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r19, s[$ofs2+0x0010], $r0, $r19
add.rn.f32 $r18, $r18, $r5
lds.b32 $r6, s[$ofs4+0x0058]
mad.rn.f32 $r20, $r6, $r3, -$r20
mad.rn.f32 $r21, $r6, $r2, $r21
lds.b32 $r7, s[$ofs1+0x0054]
lds.b32 $r6, s[$ofs3+0x0018]
mad.rn.f32 $r21, $r7, $r1, $r21
mad.rn.f32 $r20, $r6, $r2, -$r20
mul.rn.f32 $r5, s[$ofs2+0x0014], $r1
mad.rn.f32 $r21, $r6, $r3, $r21
mad.rn.f32 $r5, $r7, $r0, -$r5
mad.rn.f32 $r21, s[$ofs2+0x0014], $r0, $r21
add.rn.f32 $r20, $r20, $r5
bar.sync.u32 0x00000000
add.b32 $r38, $r38, 0xfffffff0
