// // Generated by NVIDIA NVVM Compiler // // Compiler Build ID: CL-31968024 // Cuda compilation tools, release 12.0, V12.0.76 // Based on NVVM 7.0.1 // .version 8.0 .target sm_52 .address_size 64 // .globl __raygen__oxMain .const .align 16 .b8 params[1184]; .global .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162}; .visible .entry __raygen__oxMain() { .local .align 4 .b8 __local_depot0[28]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<209>; .reg .b16 %rs<109>; .reg .f32 %f<1674>; .reg .b32 %r<1546>; .reg .f64 %fd<5>; .reg .b64 %rd<127>; mov.u64 %SPL, __local_depot0; add.u64 %rd1, %SPL, 0; // begin inline asm call (%r62), _optix_get_launch_index_x, (); // end inline asm // begin inline asm call (%r63), _optix_get_launch_index_y, (); // end inline asm // begin inline asm call (%r65), _optix_get_launch_dimension_x, (); // end inline asm // begin inline asm call (%r66), _optix_get_launch_dimension_y, (); // end inline asm ld.const.u64 %rd24, [params+400]; cvta.to.global.u64 %rd25, %rd24; ld.const.u32 %r68, [params+392]; mad.lo.s32 %r69, %r68, %r63, %r62; mul.wide.u32 %rd26, %r69, 4; add.s64 %rd2, %rd25, %rd26; ld.global.v2.u8 {%rs15, %rs100}, [%rd2]; or.b16 %rs17, %rs15, %rs100; and.b16 %rs18, %rs17, 255; setp.eq.s16 %p8, %rs18, 0; @%p8 bra $L__BB0_2; ld.global.u8 %rs99, [%rd2+2]; bra.uni $L__BB0_3; $L__BB0_2: ld.global.u8 %rs99, [%rd2+2]; setp.eq.s16 %p9, %rs99, 0; mov.f32 %f1619, 0f00000000; mov.u16 %rs100, 0; mov.f32 %f1620, %f1619; mov.f32 %f1621, %f1619; @%p9 bra $L__BB0_4; $L__BB0_3: cvt.rn.f32.u16 %f227, %rs15; div.rn.f32 %f228, %f227, 0f437F0000; fma.rn.f32 %f229, %f228, 0f40000000, 0fBF800000; and.b16 %rs21, %rs100, 255; cvt.rn.f32.u16 %f230, %rs21; div.rn.f32 %f231, %f230, 0f437F0000; fma.rn.f32 %f232, %f231, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f233, %rs99; div.rn.f32 %f234, %f233, 0f437F0000; fma.rn.f32 %f235, %f234, 0f40000000, 0fBF800000; mul.f32 %f236, %f232, %f232; fma.rn.f32 %f237, %f229, %f229, %f236; fma.rn.f32 %f238, %f235, %f235, %f237; sqrt.rn.f32 %f239, %f238; rcp.rn.f32 %f240, %f239; mul.f32 %f1621, %f240, %f235; mul.f32 %f1620, %f240, %f232; mul.f32 %f1619, %f229, %f240; $L__BB0_4: ld.const.v2.u32 {%r70, %r71}, [params]; add.s32 %r5, %r70, %r62; add.s32 %r6, %r71, %r63; setp.eq.f32 %p10, %f1619, 0f00000000; setp.eq.f32 %p11, %f1620, 0f00000000; and.pred %p12, %p10, %p11; setp.eq.f32 %p13, %f1621, 0f00000000; and.pred %p14, %p13, %p12; @%p14 bra $L__BB0_160; bra.uni $L__BB0_5; $L__BB0_160: ld.const.u32 %r61, [params+104]; and.b32 %r1520, %r61, 1; setp.eq.b32 %p201, %r1520, 1; mov.pred %p202, 0; xor.pred %p203, %p201, %p202; not.pred %p204, %p203; @%p204 bra $L__BB0_162; ld.const.u64 %rd107, [params+144]; cvta.to.global.u64 %rd108, %rd107; ld.const.u32 %r1521, [params+136]; mad.lo.s32 %r1522, %r1521, %r6, %r5; mul.wide.u32 %rd109, %r1522, 4; add.s64 %rd110, %rd108, %rd109; mov.u16 %rs80, 0; st.global.v4.u8 [%rd110], {%rs80, %rs80, %rs80, %rs80}; $L__BB0_162: and.b32 %r1523, %r61, 8; setp.eq.s32 %p205, %r1523, 0; @%p205 bra $L__BB0_164; ld.const.u64 %rd111, [params+192]; cvta.to.global.u64 %rd112, %rd111; ld.const.u32 %r1524, [params+184]; mad.lo.s32 %r1525, %r1524, %r6, %r5; mov.f32 %f1608, 0f00000000; cvt.rzi.u32.f32 %r1526, %f1608; mul.wide.u32 %rd113, %r1525, 2; add.s64 %rd114, %rd112, %rd113; mov.u16 %rs81, 0; cvt.u16.u32 %rs82, %r1526; st.global.v2.u8 [%rd114], {%rs82, %rs81}; $L__BB0_164: and.b32 %r1527, %r61, 4; setp.eq.s32 %p206, %r1527, 0; @%p206 bra $L__BB0_168; ld.const.u32 %r1528, [params+108]; setp.eq.s32 %p207, %r1528, 0; ld.const.u64 %rd115, [params+224]; cvta.to.global.u64 %rd116, %rd115; ld.const.u32 %r1529, [params+216]; mad.lo.s32 %r1530, %r1529, %r6, %r5; mul.wide.u32 %rd117, %r1530, 8; add.s64 %rd22, %rd116, %rd117; @%p207 bra $L__BB0_167; ld.global.v4.u16 {%rs89, %rs90, %rs91, %rs92}, [%rd22]; // begin inline asm { cvt.f32.f16 %f1609, %rs89;} // end inline asm // begin inline asm { cvt.f32.f16 %f1610, %rs90;} // end inline asm // begin inline asm { cvt.f32.f16 %f1611, %rs91;} // end inline asm add.f32 %f1612, %f1609, 0f00000000; add.f32 %f1613, %f1610, 0f00000000; add.f32 %f1614, %f1611, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs88, %f1614;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs87, %f1613;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs86, %f1612;} // end inline asm mov.u16 %rs93, 0; st.global.v4.u16 [%rd22], {%rs86, %rs87, %rs88, %rs93}; bra.uni $L__BB0_168; $L__BB0_5: ld.const.v4.f32 {%f241, %f242, %f243, %f244}, [params+512]; neg.f32 %f13, %f242; neg.f32 %f14, %f243; mul.f32 %f245, %f1619, %f241; mul.f32 %f246, %f1620, %f13; sub.f32 %f247, %f246, %f245; mul.f32 %f248, %f1621, %f243; sub.f32 %f15, %f247, %f248; setp.gtu.f32 %p15, %f15, 0f00000000; ld.const.u32 %r7, [params+544]; setp.ne.s32 %p16, %r7, 0; or.pred %p17, %p16, %p15; @%p17 bra $L__BB0_7; ld.const.u32 %r8, [params+104]; and.b32 %r74, %r8, 32; setp.eq.s32 %p18, %r74, 0; @%p18 bra $L__BB0_150; bra.uni $L__BB0_7; $L__BB0_150: and.b32 %r1506, %r8, 1; setp.eq.b32 %p193, %r1506, 1; mov.pred %p194, 0; xor.pred %p195, %p193, %p194; not.pred %p196, %p195; @%p196 bra $L__BB0_152; ld.const.u64 %rd92, [params+144]; cvta.to.global.u64 %rd93, %rd92; ld.const.u32 %r1507, [params+136]; mad.lo.s32 %r1508, %r1507, %r6, %r5; mul.wide.u32 %rd94, %r1508, 4; add.s64 %rd95, %rd93, %rd94; mov.u16 %rs59, 0; mov.u16 %rs60, 255; st.global.v4.u8 [%rd95], {%rs59, %rs59, %rs59, %rs60}; $L__BB0_152: and.b32 %r1509, %r8, 8; setp.eq.s32 %p197, %r1509, 0; @%p197 bra $L__BB0_154; ld.const.u64 %rd96, [params+192]; cvta.to.global.u64 %rd97, %rd96; ld.const.u32 %r1510, [params+184]; mad.lo.s32 %r1511, %r1510, %r6, %r5; mov.f32 %f1596, 0f00000000; cvt.rzi.u32.f32 %r1512, %f1596; mul.wide.u32 %rd98, %r1511, 2; add.s64 %rd99, %rd97, %rd98; cvt.u16.u32 %rs61, %r1512; mov.u16 %rs62, 255; st.global.v2.u8 [%rd99], {%rs61, %rs62}; $L__BB0_154: and.b32 %r1513, %r8, 4; setp.eq.s32 %p198, %r1513, 0; @%p198 bra $L__BB0_158; ld.const.u32 %r1514, [params+108]; setp.eq.s32 %p199, %r1514, 0; ld.const.u64 %rd100, [params+224]; cvta.to.global.u64 %rd101, %rd100; ld.const.u32 %r1515, [params+216]; mad.lo.s32 %r1516, %r1515, %r6, %r5; mul.wide.u32 %rd102, %r1516, 8; add.s64 %rd21, %rd101, %rd102; @%p199 bra $L__BB0_157; ld.global.v4.u16 {%rs70, %rs71, %rs72, %rs73}, [%rd21]; // begin inline asm { cvt.f32.f16 %f1597, %rs70;} // end inline asm // begin inline asm { cvt.f32.f16 %f1598, %rs71;} // end inline asm // begin inline asm { cvt.f32.f16 %f1599, %rs72;} // end inline asm add.f32 %f1600, %f1597, 0f00000000; add.f32 %f1601, %f1598, 0f00000000; add.f32 %f1602, %f1599, 0f00000000; mov.f32 %f1603, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs69, %f1603;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs68, %f1602;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs67, %f1601;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs66, %f1600;} // end inline asm st.global.v4.u16 [%rd21], {%rs66, %rs67, %rs68, %rs69}; bra.uni $L__BB0_158; $L__BB0_7: ld.const.u64 %rd27, [params+432]; cvta.to.global.u64 %rd28, %rd27; ld.const.u32 %r75, [params+424]; mul.lo.s32 %r76, %r75, %r63; add.s32 %r77, %r76, %r62; mul.wide.u32 %rd29, %r77, 12; add.s64 %rd30, %rd28, %rd29; ld.global.f32 %f251, [%rd30]; mul.f32 %f252, %f251, 0f3456BF95; ld.global.f32 %f253, [%rd30+4]; mul.f32 %f254, %f253, 0f3456BF95; ld.global.f32 %f255, [%rd30+8]; mul.f32 %f256, %f255, 0f3456BF95; abs.f32 %f257, %f1619; div.rn.f32 %f258, %f252, %f257; abs.f32 %f259, %f1620; div.rn.f32 %f260, %f254, %f259; abs.f32 %f261, %f1621; div.rn.f32 %f262, %f256, %f261; abs.f32 %f263, %f258; abs.f32 %f264, %f260; abs.f32 %f265, %f262; mov.f32 %f266, 0f38D1B717; max.f32 %f267, %f263, %f266; max.f32 %f268, %f264, %f266; max.f32 %f269, %f265, %f266; fma.rn.f32 %f16, %f1619, %f267, %f251; fma.rn.f32 %f17, %f1620, %f268, %f253; fma.rn.f32 %f18, %f1621, %f269, %f255; setp.gt.f32 %p19, %f257, %f261; neg.f32 %f270, %f1620; selp.f32 %f271, %f270, 0f00000000, %p19; mov.f32 %f1632, 0f00000000; neg.f32 %f272, %f1621; selp.f32 %f273, %f1619, %f272, %p19; selp.f32 %f274, 0f00000000, %f1620, %p19; mul.f32 %f275, %f273, %f273; fma.rn.f32 %f276, %f271, %f271, %f275; fma.rn.f32 %f277, %f274, %f274, %f276; sqrt.rn.f32 %f278, %f277; rcp.rn.f32 %f279, %f278; mul.f32 %f19, %f271, %f279; mul.f32 %f20, %f273, %f279; mul.f32 %f21, %f274, %f279; mul.f32 %f280, %f1621, %f20; mul.f32 %f281, %f1620, %f21; sub.f32 %f22, %f280, %f281; mul.f32 %f282, %f1619, %f21; mul.f32 %f283, %f1621, %f19; sub.f32 %f23, %f282, %f283; mul.f32 %f284, %f1620, %f19; mul.f32 %f285, %f1619, %f20; sub.f32 %f24, %f284, %f285; add.s32 %r78, %r65, -1; add.s32 %r79, %r62, 1; min.u32 %r80, %r79, %r78; add.s32 %r81, %r76, %r80; mul.wide.u32 %rd31, %r81, 12; add.s64 %rd32, %rd28, %rd31; ld.global.f32 %f286, [%rd32]; setp.eq.f32 %p20, %f286, 0f00000000; ld.global.f32 %f287, [%rd32+4]; setp.eq.f32 %p21, %f287, 0f00000000; and.pred %p22, %p20, %p21; ld.global.f32 %f288, [%rd32+8]; setp.eq.f32 %p23, %f288, 0f00000000; and.pred %p24, %p22, %p23; selp.f32 %f289, %f251, %f286, %p24; selp.f32 %f290, %f253, %f287, %p24; selp.f32 %f291, %f255, %f288, %p24; add.s32 %r82, %r66, -1; add.s32 %r83, %r63, 1; min.u32 %r84, %r83, %r82; mad.lo.s32 %r85, %r75, %r84, %r62; mul.wide.u32 %rd33, %r85, 12; add.s64 %rd34, %rd28, %rd33; ld.global.f32 %f292, [%rd34]; setp.eq.f32 %p25, %f292, 0f00000000; ld.global.f32 %f293, [%rd34+4]; setp.eq.f32 %p26, %f293, 0f00000000; and.pred %p27, %p25, %p26; ld.global.f32 %f294, [%rd34+8]; setp.eq.f32 %p28, %f294, 0f00000000; and.pred %p29, %p27, %p28; selp.f32 %f295, %f251, %f292, %p29; selp.f32 %f296, %f253, %f293, %p29; selp.f32 %f297, %f255, %f294, %p29; sub.f32 %f298, %f289, %f251; sub.f32 %f299, %f290, %f253; sub.f32 %f300, %f291, %f255; sub.f32 %f301, %f295, %f251; sub.f32 %f302, %f296, %f253; sub.f32 %f303, %f297, %f255; abs.f32 %f304, %f298; abs.f32 %f305, %f301; max.f32 %f306, %f304, %f305; abs.f32 %f307, %f299; abs.f32 %f308, %f302; max.f32 %f309, %f307, %f308; abs.f32 %f310, %f300; abs.f32 %f311, %f303; max.f32 %f312, %f310, %f311; max.f32 %f313, %f306, %f309; max.f32 %f314, %f313, %f312; setp.eq.f32 %p30, %f298, 0f00000000; setp.eq.f32 %p31, %f299, 0f00000000; and.pred %p32, %p30, %p31; setp.eq.f32 %p33, %f300, 0f00000000; and.pred %p34, %p33, %p32; setp.eq.f32 %p35, %f301, 0f00000000; setp.eq.f32 %p36, %f302, 0f00000000; and.pred %p37, %p35, %p36; setp.eq.f32 %p38, %f303, 0f00000000; and.pred %p39, %p38, %p37; or.pred %p40, %p34, %p39; selp.f32 %f25, 0f00000000, %f314, %p40; abs.f32 %f315, %f243; abs.f32 %f316, %f241; setp.gt.f32 %p41, %f316, %f315; selp.f32 %f317, %f13, 0f00000000, %p41; selp.f32 %f318, %f241, %f14, %p41; selp.f32 %f319, 0f00000000, %f242, %p41; mul.f32 %f320, %f318, %f318; fma.rn.f32 %f321, %f317, %f317, %f320; fma.rn.f32 %f322, %f319, %f319, %f321; sqrt.rn.f32 %f323, %f322; rcp.rn.f32 %f324, %f323; mul.f32 %f26, %f317, %f324; mul.f32 %f27, %f318, %f324; mul.f32 %f28, %f319, %f324; ld.const.u64 %rd35, [params+128]; cvta.to.global.u64 %rd36, %rd35; ld.const.u32 %r86, [params+120]; mad.lo.s32 %r87, %r86, %r63, %r62; mul.wide.u32 %rd37, %r87, 4; add.s64 %rd3, %rd36, %rd37; ld.const.u32 %r9, [params+540]; setp.lt.s32 %p42, %r9, 1; mov.f32 %f1633, %f1632; @%p42 bra $L__BB0_54; cvt.rn.f32.s32 %f327, %r9; rcp.rn.f32 %f29, %f327; ld.global.u32 %r1538, [%rd3]; ld.const.f32 %f30, [params+536]; ld.const.u8 %rs23, [params+104]; and.b16 %rs24, %rs23, 32; setp.eq.s16 %p43, %rs24, 0; mov.u32 %r88, 0; selp.f32 %f31, 0f3F800000, 0f41200000, %p43; mul.f32 %f32, %f16, 0f3456BF95; mul.f32 %f33, %f17, 0f3456BF95; mul.f32 %f34, %f18, 0f3456BF95; ld.const.u64 %rd4, [params+96]; mul.f32 %f328, %f241, %f27; mul.f32 %f329, %f242, %f26; sub.f32 %f35, %f329, %f328; mul.f32 %f330, %f243, %f26; mul.f32 %f331, %f241, %f28; sub.f32 %f36, %f331, %f330; mul.f32 %f332, %f242, %f28; mul.f32 %f333, %f243, %f27; sub.f32 %f37, %f333, %f332; abs.f32 %f416, %f33; abs.f32 %f417, %f32; max.f32 %f418, %f417, %f416; abs.f32 %f419, %f34; max.f32 %f420, %f418, %f419; mov.u32 %r1535, %r88; $L__BB0_9: mov.u32 %r1537, %r88; $L__BB0_10: cvt.rn.f32.s32 %f1618, %r1535; mad.lo.s32 %r90, %r1538, 1664525, 1013904223; and.b32 %r91, %r90, 16777215; cvt.rn.f32.u32 %f334, %r91; fma.rn.f32 %f335, %f334, 0f33800000, %f1618; mul.f32 %f336, %f29, %f335; mad.lo.s32 %r15, %r90, 1664525, 1013904223; and.b32 %r92, %r15, 16777215; cvt.rn.f32.u32 %f337, %r92; cvt.rn.f32.s32 %f338, %r1537; fma.rn.f32 %f339, %f337, 0f33800000, %f338; mul.f32 %f340, %f29, %f339; sqrt.rn.f32 %f43, %f336; mul.f32 %f44, %f340, 0f40C90FDB; mul.f32 %f341, %f44, 0f3F22F983; cvt.rni.s32.f32 %r1545, %f341; cvt.rn.f32.s32 %f342, %r1545; mov.f32 %f343, 0fBFC90FDA; fma.rn.f32 %f344, %f342, %f343, %f44; mov.f32 %f345, 0fB3A22168; fma.rn.f32 %f346, %f342, %f345, %f344; mov.f32 %f347, 0fA7C234C5; fma.rn.f32 %f1629, %f342, %f347, %f346; abs.f32 %f46, %f44; setp.ltu.f32 %p44, %f46, 0f47CE4780; add.s64 %rd5, %rd1, 24; mov.u32 %r1542, %r1545; mov.f32 %f1626, %f1629; @%p44 bra $L__BB0_18; setp.eq.f32 %p45, %f46, 0f7F800000; @%p45 bra $L__BB0_17; bra.uni $L__BB0_12; $L__BB0_17: mov.f32 %f350, 0f00000000; mul.rn.f32 %f1626, %f44, %f350; mov.u32 %r1542, 0; bra.uni $L__BB0_18; $L__BB0_12: mov.b32 %r17, %f44; bfe.u32 %r94, %r17, 23, 8; add.s32 %r18, %r94, -128; shl.b32 %r95, %r17, 8; or.b32 %r19, %r95, -2147483648; shr.u32 %r20, %r18, 5; mov.u64 %rd124, 0; mov.u32 %r1539, 0; mov.u64 %rd123, __cudart_i2opi_f; mov.u64 %rd122, %rd1; $L__BB0_13: .pragma "nounroll"; ld.global.nc.u32 %r96, [%rd123]; mad.wide.u32 %rd40, %r96, %r19, %rd124; shr.u64 %rd124, %rd40, 32; st.local.u32 [%rd122], %rd40; add.s64 %rd123, %rd123, 4; add.s64 %rd122, %rd122, 4; add.s32 %r1539, %r1539, 1; setp.ne.s32 %p46, %r1539, 6; @%p46 bra $L__BB0_13; st.local.u32 [%rd5], %rd124; mov.u32 %r97, 4; sub.s32 %r23, %r97, %r20; mov.u32 %r98, 6; sub.s32 %r99, %r98, %r20; mul.wide.s32 %rd41, %r99, 4; add.s64 %rd42, %rd1, %rd41; ld.local.u32 %r1540, [%rd42]; ld.local.u32 %r1541, [%rd42+-4]; and.b32 %r26, %r18, 31; setp.eq.s32 %p47, %r26, 0; @%p47 bra $L__BB0_16; mov.u32 %r100, 32; sub.s32 %r101, %r100, %r26; shr.u32 %r102, %r1541, %r101; shl.b32 %r103, %r1540, %r26; add.s32 %r1540, %r102, %r103; mul.wide.s32 %rd43, %r23, 4; add.s64 %rd44, %rd1, %rd43; ld.local.u32 %r104, [%rd44]; shr.u32 %r105, %r104, %r101; shl.b32 %r106, %r1541, %r26; add.s32 %r1541, %r105, %r106; $L__BB0_16: and.b32 %r107, %r17, -2147483648; shr.u32 %r108, %r1541, 30; shl.b32 %r109, %r1540, 2; or.b32 %r110, %r108, %r109; shr.u32 %r111, %r110, 31; shr.u32 %r112, %r1540, 30; add.s32 %r113, %r111, %r112; neg.s32 %r114, %r113; setp.eq.s32 %p48, %r107, 0; selp.b32 %r1542, %r113, %r114, %p48; setp.ne.s32 %p49, %r111, 0; xor.b32 %r115, %r107, -2147483648; selp.b32 %r116, %r115, %r107, %p49; selp.b32 %r117, -1, 0, %p49; xor.b32 %r118, %r110, %r117; shl.b32 %r119, %r1541, 2; xor.b32 %r120, %r119, %r117; cvt.u64.u32 %rd45, %r118; cvt.u64.u32 %rd46, %r120; bfi.b64 %rd47, %rd45, %rd46, 32, 32; cvt.rn.f64.s64 %fd1, %rd47; mul.f64 %fd2, %fd1, 0d3BF921FB54442D19; cvt.rn.f32.f64 %f348, %fd2; setp.eq.s32 %p50, %r116, 0; neg.f32 %f349, %f348; selp.f32 %f1626, %f348, %f349, %p50; $L__BB0_18: add.s32 %r33, %r1542, 1; and.b32 %r34, %r33, 1; setp.eq.s32 %p51, %r34, 0; selp.f32 %f50, %f1626, 0f3F800000, %p51; mul.rn.f32 %f51, %f1626, %f1626; mov.f32 %f1627, 0fB94D4153; @%p51 bra $L__BB0_20; mov.f32 %f352, 0fBAB607ED; mov.f32 %f353, 0f37CBAC00; fma.rn.f32 %f1627, %f353, %f51, %f352; $L__BB0_20: selp.f32 %f354, 0f3C0885E4, 0f3D2AAABB, %p51; fma.rn.f32 %f355, %f1627, %f51, %f354; selp.f32 %f356, 0fBE2AAAA8, 0fBEFFFFFF, %p51; fma.rn.f32 %f357, %f355, %f51, %f356; mov.f32 %f358, 0f00000000; fma.rn.f32 %f359, %f51, %f50, %f358; fma.rn.f32 %f1628, %f357, %f359, %f50; and.b32 %r122, %r33, 2; setp.eq.s32 %p53, %r122, 0; @%p53 bra $L__BB0_22; mov.f32 %f361, 0fBF800000; fma.rn.f32 %f1628, %f1628, %f361, %f358; $L__BB0_22: @%p44 bra $L__BB0_30; setp.eq.f32 %p55, %f46, 0f7F800000; @%p55 bra $L__BB0_29; bra.uni $L__BB0_24; $L__BB0_29: mov.f32 %f364, 0f00000000; mul.rn.f32 %f1629, %f44, %f364; mov.u32 %r1545, 0; bra.uni $L__BB0_30; $L__BB0_24: mov.b32 %r35, %f44; bfe.u32 %r123, %r35, 23, 8; add.s32 %r36, %r123, -128; shl.b32 %r124, %r35, 8; or.b32 %r37, %r124, -2147483648; shr.u32 %r38, %r36, 5; mov.u64 %rd125, 0; mov.u64 %rd126, %rd125; $L__BB0_25: .pragma "nounroll"; shl.b64 %rd50, %rd125, 2; mov.u64 %rd51, __cudart_i2opi_f; add.s64 %rd52, %rd51, %rd50; ld.global.nc.u32 %r125, [%rd52]; mad.wide.u32 %rd53, %r125, %r37, %rd126; shr.u64 %rd126, %rd53, 32; add.s64 %rd54, %rd1, %rd50; st.local.u32 [%rd54], %rd53; cvt.u32.u64 %r126, %rd125; add.s32 %r127, %r126, 1; cvt.s64.s32 %rd125, %r127; setp.ne.s32 %p56, %r127, 6; @%p56 bra $L__BB0_25; st.local.u32 [%rd5], %rd126; mov.u32 %r128, 4; sub.s32 %r39, %r128, %r38; mov.u32 %r129, 6; sub.s32 %r130, %r129, %r38; mul.wide.s32 %rd55, %r130, 4; add.s64 %rd56, %rd1, %rd55; ld.local.u32 %r1543, [%rd56]; ld.local.u32 %r1544, [%rd56+-4]; and.b32 %r42, %r36, 31; setp.eq.s32 %p57, %r42, 0; @%p57 bra $L__BB0_28; mov.u32 %r131, 32; sub.s32 %r132, %r131, %r42; shr.u32 %r133, %r1544, %r132; shl.b32 %r134, %r1543, %r42; add.s32 %r1543, %r133, %r134; mul.wide.s32 %rd57, %r39, 4; add.s64 %rd58, %rd1, %rd57; ld.local.u32 %r135, [%rd58]; shr.u32 %r136, %r135, %r132; shl.b32 %r137, %r1544, %r42; add.s32 %r1544, %r136, %r137; $L__BB0_28: and.b32 %r138, %r35, -2147483648; shr.u32 %r139, %r1544, 30; shl.b32 %r140, %r1543, 2; or.b32 %r141, %r139, %r140; shr.u32 %r142, %r141, 31; shr.u32 %r143, %r1543, 30; add.s32 %r144, %r142, %r143; neg.s32 %r145, %r144; setp.eq.s32 %p58, %r138, 0; selp.b32 %r1545, %r144, %r145, %p58; setp.ne.s32 %p59, %r142, 0; xor.b32 %r146, %r138, -2147483648; selp.b32 %r147, %r146, %r138, %p59; selp.b32 %r148, -1, 0, %p59; xor.b32 %r149, %r141, %r148; shl.b32 %r150, %r1544, 2; xor.b32 %r151, %r150, %r148; cvt.u64.u32 %rd59, %r149; cvt.u64.u32 %rd60, %r151; bfi.b64 %rd61, %rd59, %rd60, 32, 32; cvt.rn.f64.s64 %fd3, %rd61; mul.f64 %fd4, %fd3, 0d3BF921FB54442D19; cvt.rn.f32.f64 %f362, %fd4; setp.eq.s32 %p60, %r147, 0; neg.f32 %f363, %f362; selp.f32 %f1629, %f362, %f363, %p60; $L__BB0_30: mul.f32 %f60, %f43, %f1628; and.b32 %r49, %r1545, 1; setp.eq.s32 %p61, %r49, 0; selp.f32 %f61, %f1629, 0f3F800000, %p61; mul.rn.f32 %f62, %f1629, %f1629; mov.f32 %f1630, 0fB94D4153; @%p61 bra $L__BB0_32; mov.f32 %f366, 0fBAB607ED; mov.f32 %f367, 0f37CBAC00; fma.rn.f32 %f1630, %f367, %f62, %f366; $L__BB0_32: selp.f32 %f368, 0f3C0885E4, 0f3D2AAABB, %p61; fma.rn.f32 %f369, %f1630, %f62, %f368; selp.f32 %f370, 0fBE2AAAA8, 0fBEFFFFFF, %p61; fma.rn.f32 %f371, %f369, %f62, %f370; mov.f32 %f372, 0f00000000; fma.rn.f32 %f373, %f62, %f61, %f372; fma.rn.f32 %f1631, %f371, %f373, %f61; and.b32 %r153, %r1545, 2; setp.eq.s32 %p63, %r153, 0; @%p63 bra $L__BB0_34; mov.f32 %f375, 0fBF800000; fma.rn.f32 %f1631, %f1631, %f375, %f372; $L__BB0_34: mul.f32 %f385, %f60, %f60; mov.f32 %f386, 0f3F800000; sub.f32 %f387, %f386, %f385; mul.f32 %f388, %f43, %f1631; mul.f32 %f389, %f388, %f388; sub.f32 %f390, %f387, %f389; max.f32 %f391, %f372, %f390; sqrt.rn.f32 %f392, %f391; mul.f32 %f393, %f26, %f388; mul.f32 %f394, %f27, %f388; mul.f32 %f395, %f28, %f388; fma.rn.f32 %f396, %f37, %f60, %f393; fma.rn.f32 %f397, %f36, %f60, %f394; fma.rn.f32 %f398, %f35, %f60, %f395; fma.rn.f32 %f399, %f241, %f392, %f396; fma.rn.f32 %f400, %f242, %f392, %f397; fma.rn.f32 %f401, %f243, %f392, %f398; add.f32 %f402, %f241, %f399; add.f32 %f403, %f242, %f400; add.f32 %f404, %f243, %f401; mul.f32 %f405, %f30, %f402; mul.f32 %f406, %f30, %f403; mul.f32 %f407, %f30, %f404; sub.f32 %f408, %f405, %f241; sub.f32 %f409, %f406, %f242; sub.f32 %f410, %f407, %f243; mul.f32 %f411, %f409, %f409; fma.rn.f32 %f412, %f408, %f408, %f411; fma.rn.f32 %f413, %f410, %f410, %f412; sqrt.rn.f32 %f414, %f413; rcp.rn.f32 %f415, %f414; mul.f32 %f68, %f415, %f408; mul.f32 %f69, %f415, %f409; mul.f32 %f70, %f415, %f410; mov.f32 %f421, 0f38D1B717; max.f32 %f422, %f420, %f421; mul.f32 %f382, %f31, %f422; mad.lo.s32 %r225, %r15, 1664525, 1013904223; and.b32 %r226, %r225, 16777215; cvt.rn.f32.u32 %f423, %r226; mul.f32 %f424, %f423, 0f33800000; fma.rn.f32 %f425, %f424, 0f40000000, 0fBF800000; mad.lo.s32 %r50, %r225, 1664525, 1013904223; and.b32 %r227, %r50, 16777215; cvt.rn.f32.u32 %f426, %r227; mul.f32 %f427, %f426, 0f33800000; fma.rn.f32 %f428, %f427, 0f40000000, 0fBF800000; mul.f32 %f429, %f19, %f428; mul.f32 %f430, %f20, %f428; mul.f32 %f431, %f21, %f428; fma.rn.f32 %f432, %f22, %f425, %f429; fma.rn.f32 %f433, %f23, %f425, %f430; fma.rn.f32 %f434, %f24, %f425, %f431; mul.f32 %f435, %f433, %f433; fma.rn.f32 %f436, %f432, %f432, %f435; fma.rn.f32 %f437, %f434, %f434, %f436; sqrt.rn.f32 %f438, %f437; rcp.rn.f32 %f439, %f438; mul.f32 %f440, %f432, %f439; mul.f32 %f441, %f433, %f439; mul.f32 %f442, %f434, %f439; mul.f32 %f443, %f1620, %f442; mul.f32 %f444, %f1621, %f441; sub.f32 %f445, %f443, %f444; mul.f32 %f446, %f1621, %f440; mul.f32 %f447, %f1619, %f442; sub.f32 %f448, %f446, %f447; mul.f32 %f449, %f1619, %f441; mul.f32 %f450, %f1620, %f440; sub.f32 %f451, %f449, %f450; mul.f32 %f452, %f440, 0f3E24A490; mul.f32 %f453, %f441, 0f3E24A490; mul.f32 %f454, %f442, 0f3E24A490; fma.rn.f32 %f455, %f25, %f452, %f16; fma.rn.f32 %f456, %f25, %f453, %f17; fma.rn.f32 %f457, %f25, %f454, %f18; mul.f32 %f458, %f445, 0fBECECEDE; mul.f32 %f459, %f448, 0fBECECEDE; mul.f32 %f460, %f451, 0fBECECEDE; fma.rn.f32 %f72, %f25, %f458, %f455; fma.rn.f32 %f73, %f25, %f459, %f456; fma.rn.f32 %f74, %f25, %f460, %f457; sub.f32 %f461, %f72, %f16; sub.f32 %f462, %f73, %f17; sub.f32 %f463, %f74, %f18; mul.f32 %f464, %f462, %f462; fma.rn.f32 %f465, %f461, %f461, %f464; fma.rn.f32 %f466, %f463, %f463, %f465; sqrt.rn.f32 %f383, %f466; rcp.rn.f32 %f467, %f383; mul.f32 %f379, %f467, %f461; mul.f32 %f380, %f467, %f462; mul.f32 %f381, %f467, %f463; mov.u32 %r190, 2; mov.u32 %r192, 1; mov.u32 %r193, 1065353216; mov.u32 %r224, 0; // begin inline asm call(%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164,%r165,%r166,%r167,%r168,%r169,%r170,%r171,%r172,%r173,%r174,%r175,%r176,%r177,%r178,%r179,%r180,%r181,%r182,%r183,%r184,%r185),_optix_trace_typed_32,(%r224,%rd4,%f16,%f17,%f18,%f379,%f380,%f381,%f382,%f383,%f372,%r192,%r224,%r192,%r190,%r192,%r192,%r193,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224); // end inline asm mov.b32 %f468, %r154; setp.neu.f32 %p64, %f468, 0f3F800000; mov.u16 %rs102, 0; @%p64 bra $L__BB0_36; mov.f32 %f476, 0f6C4ECB8F; mov.f32 %f477, 0f00000000; // begin inline asm call(%r228,%r229,%r230,%r231,%r232,%r233,%r234,%r235,%r236,%r237,%r238,%r239,%r240,%r241,%r242,%r243,%r244,%r245,%r246,%r247,%r248,%r249,%r250,%r251,%r252,%r253,%r254,%r255,%r256,%r257,%r258,%r259),_optix_trace_typed_32,(%r224,%rd4,%f72,%f73,%f74,%f68,%f69,%f70,%f382,%f476,%f477,%r192,%r224,%r192,%r190,%r192,%r192,%r193,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224); // end inline asm mov.b32 %f478, %r228; add.f32 %f1632, %f1632, %f478; add.f32 %f1633, %f1633, 0f3F800000; mov.u16 %rs102, 1; $L__BB0_36: mad.lo.s32 %r370, %r50, 1664525, 1013904223; and.b32 %r371, %r370, 16777215; cvt.rn.f32.u32 %f488, %r371; mul.f32 %f489, %f488, 0f33800000; fma.rn.f32 %f490, %f489, 0f40000000, 0fBF800000; mad.lo.s32 %r51, %r370, 1664525, 1013904223; and.b32 %r372, %r51, 16777215; cvt.rn.f32.u32 %f491, %r372; mul.f32 %f492, %f491, 0f33800000; fma.rn.f32 %f493, %f492, 0f40000000, 0fBF800000; mul.f32 %f494, %f19, %f493; mul.f32 %f495, %f20, %f493; mul.f32 %f496, %f21, %f493; fma.rn.f32 %f497, %f22, %f490, %f494; fma.rn.f32 %f498, %f23, %f490, %f495; fma.rn.f32 %f499, %f24, %f490, %f496; mul.f32 %f500, %f498, %f498; fma.rn.f32 %f501, %f497, %f497, %f500; fma.rn.f32 %f502, %f499, %f499, %f501; sqrt.rn.f32 %f503, %f502; rcp.rn.f32 %f504, %f503; mul.f32 %f505, %f497, %f504; mul.f32 %f506, %f498, %f504; mul.f32 %f507, %f499, %f504; mul.f32 %f508, %f1620, %f507; mul.f32 %f509, %f1621, %f506; sub.f32 %f510, %f508, %f509; mul.f32 %f511, %f1621, %f505; mul.f32 %f512, %f1619, %f507; sub.f32 %f513, %f511, %f512; mul.f32 %f514, %f1619, %f506; mul.f32 %f515, %f1620, %f505; sub.f32 %f516, %f514, %f515; mul.f32 %f517, %f505, 0fBE24A490; mul.f32 %f518, %f506, 0fBE24A490; mul.f32 %f519, %f507, 0fBE24A490; fma.rn.f32 %f520, %f25, %f517, %f16; fma.rn.f32 %f521, %f25, %f518, %f17; fma.rn.f32 %f522, %f25, %f519, %f18; mul.f32 %f523, %f510, 0f3ECACABC; mul.f32 %f524, %f513, 0f3ECACABC; mul.f32 %f525, %f516, 0f3ECACABC; fma.rn.f32 %f79, %f25, %f523, %f520; fma.rn.f32 %f80, %f25, %f524, %f521; fma.rn.f32 %f81, %f25, %f525, %f522; sub.f32 %f526, %f79, %f16; sub.f32 %f527, %f80, %f17; sub.f32 %f528, %f81, %f18; mul.f32 %f529, %f527, %f527; fma.rn.f32 %f530, %f526, %f526, %f529; fma.rn.f32 %f531, %f528, %f528, %f530; sqrt.rn.f32 %f486, %f531; rcp.rn.f32 %f532, %f486; mul.f32 %f482, %f532, %f526; mul.f32 %f483, %f532, %f527; mul.f32 %f484, %f532, %f528; mov.f32 %f487, 0f00000000; // begin inline asm call(%r299,%r300,%r301,%r302,%r303,%r304,%r305,%r306,%r307,%r308,%r309,%r310,%r311,%r312,%r313,%r314,%r315,%r316,%r317,%r318,%r319,%r320,%r321,%r322,%r323,%r324,%r325,%r326,%r327,%r328,%r329,%r330),_optix_trace_typed_32,(%r224,%rd4,%f16,%f17,%f18,%f482,%f483,%f484,%f382,%f486,%f487,%r192,%r224,%r192,%r190,%r192,%r192,%r193,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224,%r224); // end inline asm mov.b32 %f533, %r299; setp.neu.f32 %p65, %f533, 0f3F800000; @%p65 bra $L__BB0_38; mov.f32 %f541, 0f6C4ECB8F; mov.u32 %r409, 2; mov.u32 %r411, 1; mov.u32 %r412, 1065353216; mov.u32 %r443, 0; // begin inline asm call(%r373,%r374,%r375,%r376,%r377,%r378,%r379,%r380,%r381,%r382,%r383,%r384,%r385,%r386,%r387,%r388,%r389,%r390,%r391,%r392,%r393,%r394,%r395,%r396,%r397,%r398,%r399,%r400,%r401,%r402,%r403,%r404),_optix_trace_typed_32,(%r443,%rd4,%f79,%f80,%f81,%f68,%f69,%f70,%f382,%f541,%f487,%r411,%r443,%r411,%r409,%r411,%r411,%r412,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443,%r443); // end inline asm mov.b32 %f543, %r373; add.f32 %f1632, %f1632, %f543; add.f32 %f1633, %f1633, 0f3F800000; mov.u16 %rs102, 1; $L__BB0_38: mad.lo.s32 %r515, %r51, 1664525, 1013904223; and.b32 %r516, %r515, 16777215; cvt.rn.f32.u32 %f553, %r516; mul.f32 %f554, %f553, 0f33800000; fma.rn.f32 %f555, %f554, 0f40000000, 0fBF800000; mad.lo.s32 %r52, %r515, 1664525, 1013904223; and.b32 %r517, %r52, 16777215; cvt.rn.f32.u32 %f556, %r517; mul.f32 %f557, %f556, 0f33800000; fma.rn.f32 %f558, %f557, 0f40000000, 0fBF800000; mul.f32 %f559, %f19, %f558; mul.f32 %f560, %f20, %f558; mul.f32 %f561, %f21, %f558; fma.rn.f32 %f562, %f22, %f555, %f559; fma.rn.f32 %f563, %f23, %f555, %f560; fma.rn.f32 %f564, %f24, %f555, %f561; mul.f32 %f565, %f563, %f563; fma.rn.f32 %f566, %f562, %f562, %f565; fma.rn.f32 %f567, %f564, %f564, %f566; sqrt.rn.f32 %f568, %f567; rcp.rn.f32 %f569, %f568; mul.f32 %f570, %f562, %f569; mul.f32 %f571, %f563, %f569; mul.f32 %f572, %f564, %f569; mul.f32 %f573, %f1620, %f572; mul.f32 %f574, %f1621, %f571; sub.f32 %f575, %f573, %f574; mul.f32 %f576, %f1621, %f570; mul.f32 %f577, %f1619, %f572; sub.f32 %f578, %f576, %f577; mul.f32 %f579, %f1619, %f571; mul.f32 %f580, %f1620, %f570; sub.f32 %f581, %f579, %f580; mul.f32 %f582, %f570, 0f3F23A3A8; mul.f32 %f583, %f571, 0f3F23A3A8; mul.f32 %f584, %f572, 0f3F23A3A8; fma.rn.f32 %f585, %f25, %f582, %f16; fma.rn.f32 %f586, %f25, %f583, %f17; fma.rn.f32 %f587, %f25, %f584, %f18; mul.f32 %f588, %f575, 0f3E24A490; mul.f32 %f589, %f578, 0f3E24A490; mul.f32 %f590, %f581, 0f3E24A490; fma.rn.f32 %f86, %f25, %f588, %f585; fma.rn.f32 %f87, %f25, %f589, %f586; fma.rn.f32 %f88, %f25, %f590, %f587; sub.f32 %f591, %f86, %f16; sub.f32 %f592, %f87, %f17; sub.f32 %f593, %f88, %f18; mul.f32 %f594, %f592, %f592; fma.rn.f32 %f595, %f591, %f591, %f594; fma.rn.f32 %f596, %f593, %f593, %f595; sqrt.rn.f32 %f551, %f596; rcp.rn.f32 %f597, %f551; mul.f32 %f547, %f597, %f591; mul.f32 %f548, %f597, %f592; mul.f32 %f549, %f597, %f593; mov.u32 %r480, 2; mov.u32 %r482, 1; mov.u32 %r483, 1065353216; mov.u32 %r514, 0; // begin inline asm call(%r444,%r445,%r446,%r447,%r448,%r449,%r450,%r451,%r452,%r453,%r454,%r455,%r456,%r457,%r458,%r459,%r460,%r461,%r462,%r463,%r464,%r465,%r466,%r467,%r468,%r469,%r470,%r471,%r472,%r473,%r474,%r475),_optix_trace_typed_32,(%r514,%rd4,%f16,%f17,%f18,%f547,%f548,%f549,%f382,%f551,%f487,%r482,%r514,%r482,%r480,%r482,%r482,%r483,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514); // end inline asm mov.b32 %f598, %r444; setp.neu.f32 %p66, %f598, 0f3F800000; @%p66 bra $L__BB0_40; mov.f32 %f606, 0f6C4ECB8F; mov.f32 %f607, 0f00000000; // begin inline asm call(%r518,%r519,%r520,%r521,%r522,%r523,%r524,%r525,%r526,%r527,%r528,%r529,%r530,%r531,%r532,%r533,%r534,%r535,%r536,%r537,%r538,%r539,%r540,%r541,%r542,%r543,%r544,%r545,%r546,%r547,%r548,%r549),_optix_trace_typed_32,(%r514,%rd4,%f86,%f87,%f88,%f68,%f69,%f70,%f382,%f606,%f607,%r482,%r514,%r482,%r480,%r482,%r482,%r483,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514); // end inline asm mov.b32 %f608, %r518; add.f32 %f1632, %f1632, %f608; add.f32 %f1633, %f1633, 0f3F800000; mov.u16 %rs102, 1; $L__BB0_40: mad.lo.s32 %r660, %r52, 1664525, 1013904223; and.b32 %r661, %r660, 16777215; cvt.rn.f32.u32 %f618, %r661; mul.f32 %f619, %f618, 0f33800000; fma.rn.f32 %f620, %f619, 0f40000000, 0fBF800000; mad.lo.s32 %r53, %r660, 1664525, 1013904223; and.b32 %r662, %r53, 16777215; cvt.rn.f32.u32 %f621, %r662; mul.f32 %f622, %f621, 0f33800000; fma.rn.f32 %f623, %f622, 0f40000000, 0fBF800000; mul.f32 %f624, %f19, %f623; mul.f32 %f625, %f20, %f623; mul.f32 %f626, %f21, %f623; fma.rn.f32 %f627, %f22, %f620, %f624; fma.rn.f32 %f628, %f23, %f620, %f625; fma.rn.f32 %f629, %f24, %f620, %f626; mul.f32 %f630, %f628, %f628; fma.rn.f32 %f631, %f627, %f627, %f630; fma.rn.f32 %f632, %f629, %f629, %f631; sqrt.rn.f32 %f633, %f632; rcp.rn.f32 %f634, %f633; mul.f32 %f635, %f627, %f634; mul.f32 %f636, %f628, %f634; mul.f32 %f637, %f629, %f634; mul.f32 %f638, %f1620, %f637; mul.f32 %f639, %f1621, %f636; sub.f32 %f640, %f638, %f639; mul.f32 %f641, %f1621, %f635; mul.f32 %f642, %f1619, %f637; sub.f32 %f643, %f641, %f642; mul.f32 %f644, %f1619, %f636; mul.f32 %f645, %f1620, %f635; sub.f32 %f646, %f644, %f645; mul.f32 %f647, %f635, 0fBECECEDE; mul.f32 %f648, %f636, 0fBECECEDE; mul.f32 %f649, %f637, 0fBECECEDE; fma.rn.f32 %f650, %f25, %f647, %f16; fma.rn.f32 %f651, %f25, %f648, %f17; fma.rn.f32 %f652, %f25, %f649, %f18; mul.f32 %f653, %f640, 0fBF23A3A9; mul.f32 %f654, %f643, 0fBF23A3A9; mul.f32 %f655, %f646, 0fBF23A3A9; fma.rn.f32 %f93, %f25, %f653, %f650; fma.rn.f32 %f94, %f25, %f654, %f651; fma.rn.f32 %f95, %f25, %f655, %f652; sub.f32 %f656, %f93, %f16; sub.f32 %f657, %f94, %f17; sub.f32 %f658, %f95, %f18; mul.f32 %f659, %f657, %f657; fma.rn.f32 %f660, %f656, %f656, %f659; fma.rn.f32 %f661, %f658, %f658, %f660; sqrt.rn.f32 %f616, %f661; rcp.rn.f32 %f662, %f616; mul.f32 %f612, %f662, %f656; mul.f32 %f613, %f662, %f657; mul.f32 %f614, %f662, %f658; mov.f32 %f617, 0f00000000; // begin inline asm call(%r589,%r590,%r591,%r592,%r593,%r594,%r595,%r596,%r597,%r598,%r599,%r600,%r601,%r602,%r603,%r604,%r605,%r606,%r607,%r608,%r609,%r610,%r611,%r612,%r613,%r614,%r615,%r616,%r617,%r618,%r619,%r620),_optix_trace_typed_32,(%r514,%rd4,%f16,%f17,%f18,%f612,%f613,%f614,%f382,%f616,%f617,%r482,%r514,%r482,%r480,%r482,%r482,%r483,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514,%r514); // end inline asm mov.b32 %f663, %r589; setp.neu.f32 %p67, %f663, 0f3F800000; @%p67 bra $L__BB0_42; mov.f32 %f671, 0f6C4ECB8F; mov.u32 %r699, 2; mov.u32 %r701, 1; mov.u32 %r702, 1065353216; mov.u32 %r733, 0; // begin inline asm call(%r663,%r664,%r665,%r666,%r667,%r668,%r669,%r670,%r671,%r672,%r673,%r674,%r675,%r676,%r677,%r678,%r679,%r680,%r681,%r682,%r683,%r684,%r685,%r686,%r687,%r688,%r689,%r690,%r691,%r692,%r693,%r694),_optix_trace_typed_32,(%r733,%rd4,%f93,%f94,%f95,%f68,%f69,%f70,%f382,%f671,%f617,%r701,%r733,%r701,%r699,%r701,%r701,%r702,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733,%r733); // end inline asm mov.b32 %f673, %r663; add.f32 %f1632, %f1632, %f673; add.f32 %f1633, %f1633, 0f3F800000; mov.u16 %rs102, 1; $L__BB0_42: mad.lo.s32 %r805, %r53, 1664525, 1013904223; and.b32 %r806, %r805, 16777215; cvt.rn.f32.u32 %f683, %r806; mul.f32 %f684, %f683, 0f33800000; fma.rn.f32 %f685, %f684, 0f40000000, 0fBF800000; mad.lo.s32 %r54, %r805, 1664525, 1013904223; and.b32 %r807, %r54, 16777215; cvt.rn.f32.u32 %f686, %r807; mul.f32 %f687, %f686, 0f33800000; fma.rn.f32 %f688, %f687, 0f40000000, 0fBF800000; mul.f32 %f689, %f19, %f688; mul.f32 %f690, %f20, %f688; mul.f32 %f691, %f21, %f688; fma.rn.f32 %f692, %f22, %f685, %f689; fma.rn.f32 %f693, %f23, %f685, %f690; fma.rn.f32 %f694, %f24, %f685, %f691; mul.f32 %f695, %f693, %f693; fma.rn.f32 %f696, %f692, %f692, %f695; fma.rn.f32 %f697, %f694, %f694, %f696; sqrt.rn.f32 %f698, %f697; rcp.rn.f32 %f699, %f698; mul.f32 %f700, %f692, %f699; mul.f32 %f701, %f693, %f699; mul.f32 %f702, %f694, %f699; mul.f32 %f703, %f1620, %f702; mul.f32 %f704, %f1621, %f701; sub.f32 %f705, %f703, %f704; mul.f32 %f706, %f1621, %f700; mul.f32 %f707, %f1619, %f702; sub.f32 %f708, %f706, %f707; mul.f32 %f709, %f1619, %f701; mul.f32 %f710, %f1620, %f700; sub.f32 %f711, %f709, %f710; mul.f32 %f712, %f700, 0fBF23A3A9; mul.f32 %f713, %f701, 0fBF23A3A9; mul.f32 %f714, %f702, 0fBF23A3A9; fma.rn.f32 %f715, %f25, %f712, %f16; fma.rn.f32 %f716, %f25, %f713, %f17; fma.rn.f32 %f717, %f25, %f714, %f18; mul.f32 %f718, %f705, 0f3F23A3A8; mul.f32 %f719, %f708, 0f3F23A3A8; mul.f32 %f720, %f711, 0f3F23A3A8; fma.rn.f32 %f100, %f25, %f718, %f715; fma.rn.f32 %f101, %f25, %f719, %f716; fma.rn.f32 %f102, %f25, %f720, %f717; sub.f32 %f721, %f100, %f16; sub.f32 %f722, %f101, %f17; sub.f32 %f723, %f102, %f18; mul.f32 %f724, %f722, %f722; fma.rn.f32 %f725, %f721, %f721, %f724; fma.rn.f32 %f726, %f723, %f723, %f725; sqrt.rn.f32 %f681, %f726; rcp.rn.f32 %f727, %f681; mul.f32 %f677, %f727, %f721; mul.f32 %f678, %f727, %f722; mul.f32 %f679, %f727, %f723; mov.u32 %r770, 2; mov.u32 %r772, 1; mov.u32 %r773, 1065353216; mov.u32 %r804, 0; // begin inline asm call(%r734,%r735,%r736,%r737,%r738,%r739,%r740,%r741,%r742,%r743,%r744,%r745,%r746,%r747,%r748,%r749,%r750,%r751,%r752,%r753,%r754,%r755,%r756,%r757,%r758,%r759,%r760,%r761,%r762,%r763,%r764,%r765),_optix_trace_typed_32,(%r804,%rd4,%f16,%f17,%f18,%f677,%f678,%f679,%f382,%f681,%f617,%r772,%r804,%r772,%r770,%r772,%r772,%r773,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804); // end inline asm mov.b32 %f728, %r734; setp.neu.f32 %p68, %f728, 0f3F800000; @%p68 bra $L__BB0_44; mov.f32 %f736, 0f6C4ECB8F; mov.f32 %f737, 0f00000000; // begin inline asm call(%r808,%r809,%r810,%r811,%r812,%r813,%r814,%r815,%r816,%r817,%r818,%r819,%r820,%r821,%r822,%r823,%r824,%r825,%r826,%r827,%r828,%r829,%r830,%r831,%r832,%r833,%r834,%r835,%r836,%r837,%r838,%r839),_optix_trace_typed_32,(%r804,%rd4,%f100,%f101,%f102,%f68,%f69,%f70,%f382,%f736,%f737,%r772,%r804,%r772,%r770,%r772,%r772,%r773,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804); // end inline asm mov.b32 %f738, %r808; add.f32 %f1632, %f1632, %f738; add.f32 %f1633, %f1633, 0f3F800000; mov.u16 %rs102, 1; $L__BB0_44: mad.lo.s32 %r950, %r54, 1664525, 1013904223; and.b32 %r951, %r950, 16777215; cvt.rn.f32.u32 %f748, %r951; mul.f32 %f749, %f748, 0f33800000; fma.rn.f32 %f750, %f749, 0f40000000, 0fBF800000; mad.lo.s32 %r55, %r950, 1664525, 1013904223; and.b32 %r952, %r55, 16777215; cvt.rn.f32.u32 %f751, %r952; mul.f32 %f752, %f751, 0f33800000; fma.rn.f32 %f753, %f752, 0f40000000, 0fBF800000; mul.f32 %f754, %f19, %f753; mul.f32 %f755, %f20, %f753; mul.f32 %f756, %f21, %f753; fma.rn.f32 %f757, %f22, %f750, %f754; fma.rn.f32 %f758, %f23, %f750, %f755; fma.rn.f32 %f759, %f24, %f750, %f756; mul.f32 %f760, %f758, %f758; fma.rn.f32 %f761, %f757, %f757, %f760; fma.rn.f32 %f762, %f759, %f759, %f761; sqrt.rn.f32 %f763, %f762; rcp.rn.f32 %f764, %f763; mul.f32 %f765, %f757, %f764; mul.f32 %f766, %f758, %f764; mul.f32 %f767, %f759, %f764; mul.f32 %f768, %f1620, %f767; mul.f32 %f769, %f1621, %f766; sub.f32 %f770, %f768, %f769; mul.f32 %f771, %f1621, %f765; mul.f32 %f772, %f1619, %f767; sub.f32 %f773, %f771, %f772; mul.f32 %f774, %f1619, %f766; mul.f32 %f775, %f1620, %f765; sub.f32 %f776, %f774, %f775; mul.f32 %f777, %f765, 0fBF61E1E3; mul.f32 %f778, %f766, 0fBF61E1E3; mul.f32 %f779, %f767, 0fBF61E1E3; fma.rn.f32 %f780, %f25, %f777, %f16; fma.rn.f32 %f781, %f25, %f778, %f17; fma.rn.f32 %f782, %f25, %f779, %f18; mul.f32 %f783, %f770, 0fBE24A490; mul.f32 %f784, %f773, 0fBE24A490; mul.f32 %f785, %f776, 0fBE24A490; fma.rn.f32 %f107, %f25, %f783, %f780; fma.rn.f32 %f108, %f25, %f784, %f781; fma.rn.f32 %f109, %f25, %f785, %f782; sub.f32 %f786, %f107, %f16; sub.f32 %f787, %f108, %f17; sub.f32 %f788, %f109, %f18; mul.f32 %f789, %f787, %f787; fma.rn.f32 %f790, %f786, %f786, %f789; fma.rn.f32 %f791, %f788, %f788, %f790; sqrt.rn.f32 %f746, %f791; rcp.rn.f32 %f792, %f746; mul.f32 %f742, %f792, %f786; mul.f32 %f743, %f792, %f787; mul.f32 %f744, %f792, %f788; mov.f32 %f747, 0f00000000; // begin inline asm call(%r879,%r880,%r881,%r882,%r883,%r884,%r885,%r886,%r887,%r888,%r889,%r890,%r891,%r892,%r893,%r894,%r895,%r896,%r897,%r898,%r899,%r900,%r901,%r902,%r903,%r904,%r905,%r906,%r907,%r908,%r909,%r910),_optix_trace_typed_32,(%r804,%rd4,%f16,%f17,%f18,%f742,%f743,%f744,%f382,%f746,%f747,%r772,%r804,%r772,%r770,%r772,%r772,%r773,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804,%r804); // end inline asm mov.b32 %f793, %r879; setp.neu.f32 %p69, %f793, 0f3F800000; @%p69 bra $L__BB0_46; mov.f32 %f801, 0f6C4ECB8F; mov.u32 %r989, 2; mov.u32 %r991, 1; mov.u32 %r992, 1065353216; mov.u32 %r1023, 0; // begin inline asm call(%r953,%r954,%r955,%r956,%r957,%r958,%r959,%r960,%r961,%r962,%r963,%r964,%r965,%r966,%r967,%r968,%r969,%r970,%r971,%r972,%r973,%r974,%r975,%r976,%r977,%r978,%r979,%r980,%r981,%r982,%r983,%r984),_optix_trace_typed_32,(%r1023,%rd4,%f107,%f108,%f109,%f68,%f69,%f70,%f382,%f801,%f747,%r991,%r1023,%r991,%r989,%r991,%r991,%r992,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023,%r1023); // end inline asm mov.b32 %f803, %r953; add.f32 %f1632, %f1632, %f803; add.f32 %f1633, %f1633, 0f3F800000; mov.u16 %rs102, 1; $L__BB0_46: mad.lo.s32 %r1095, %r55, 1664525, 1013904223; and.b32 %r1096, %r1095, 16777215; cvt.rn.f32.u32 %f813, %r1096; mul.f32 %f814, %f813, 0f33800000; fma.rn.f32 %f815, %f814, 0f40000000, 0fBF800000; mad.lo.s32 %r56, %r1095, 1664525, 1013904223; and.b32 %r1097, %r56, 16777215; cvt.rn.f32.u32 %f816, %r1097; mul.f32 %f817, %f816, 0f33800000; fma.rn.f32 %f818, %f817, 0f40000000, 0fBF800000; mul.f32 %f819, %f19, %f818; mul.f32 %f820, %f20, %f818; mul.f32 %f821, %f21, %f818; fma.rn.f32 %f822, %f22, %f815, %f819; fma.rn.f32 %f823, %f23, %f815, %f820; fma.rn.f32 %f824, %f24, %f815, %f821; mul.f32 %f825, %f823, %f823; fma.rn.f32 %f826, %f822, %f822, %f825; fma.rn.f32 %f827, %f824, %f824, %f826; sqrt.rn.f32 %f828, %f827; rcp.rn.f32 %f829, %f828; mul.f32 %f830, %f822, %f829; mul.f32 %f831, %f823, %f829; mul.f32 %f832, %f824, %f829; mul.f32 %f833, %f1620, %f832; mul.f32 %f834, %f1621, %f831; sub.f32 %f835, %f833, %f834; mul.f32 %f836, %f1621, %f830; mul.f32 %f837, %f1619, %f832; sub.f32 %f838, %f836, %f837; mul.f32 %f839, %f1619, %f831; mul.f32 %f840, %f1620, %f830; sub.f32 %f841, %f839, %f840; mul.f32 %f842, %f830, 0f3ECACABC; mul.f32 %f843, %f831, 0f3ECACABC; mul.f32 %f844, %f832, 0f3ECACABC; fma.rn.f32 %f845, %f25, %f842, %f16; fma.rn.f32 %f846, %f25, %f843, %f17; fma.rn.f32 %f847, %f25, %f844, %f18; mul.f32 %f848, %f835, 0f3F61E1D2; mul.f32 %f849, %f838, 0f3F61E1D2; mul.f32 %f850, %f841, 0f3F61E1D2; fma.rn.f32 %f114, %f25, %f848, %f845; fma.rn.f32 %f115, %f25, %f849, %f846; fma.rn.f32 %f116, %f25, %f850, %f847; sub.f32 %f851, %f114, %f16; sub.f32 %f852, %f115, %f17; sub.f32 %f853, %f116, %f18; mul.f32 %f854, %f852, %f852; fma.rn.f32 %f855, %f851, %f851, %f854; fma.rn.f32 %f856, %f853, %f853, %f855; sqrt.rn.f32 %f811, %f856; rcp.rn.f32 %f857, %f811; mul.f32 %f807, %f857, %f851; mul.f32 %f808, %f857, %f852; mul.f32 %f809, %f857, %f853; mov.u32 %r1060, 2; mov.u32 %r1062, 1; mov.u32 %r1063, 1065353216; mov.u32 %r1094, 0; // begin inline asm call(%r1024,%r1025,%r1026,%r1027,%r1028,%r1029,%r1030,%r1031,%r1032,%r1033,%r1034,%r1035,%r1036,%r1037,%r1038,%r1039,%r1040,%r1041,%r1042,%r1043,%r1044,%r1045,%r1046,%r1047,%r1048,%r1049,%r1050,%r1051,%r1052,%r1053,%r1054,%r1055),_optix_trace_typed_32,(%r1094,%rd4,%f16,%f17,%f18,%f807,%f808,%f809,%f382,%f811,%f747,%r1062,%r1094,%r1062,%r1060,%r1062,%r1062,%r1063,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094); // end inline asm mov.b32 %f858, %r1024; setp.neu.f32 %p70, %f858, 0f3F800000; @%p70 bra $L__BB0_48; mov.f32 %f866, 0f6C4ECB8F; mov.f32 %f867, 0f00000000; // begin inline asm call(%r1098,%r1099,%r1100,%r1101,%r1102,%r1103,%r1104,%r1105,%r1106,%r1107,%r1108,%r1109,%r1110,%r1111,%r1112,%r1113,%r1114,%r1115,%r1116,%r1117,%r1118,%r1119,%r1120,%r1121,%r1122,%r1123,%r1124,%r1125,%r1126,%r1127,%r1128,%r1129),_optix_trace_typed_32,(%r1094,%rd4,%f114,%f115,%f116,%f68,%f69,%f70,%f382,%f866,%f867,%r1062,%r1094,%r1062,%r1060,%r1062,%r1062,%r1063,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094); // end inline asm mov.b32 %f868, %r1098; add.f32 %f1632, %f1632, %f868; add.f32 %f1633, %f1633, 0f3F800000; mov.u16 %rs102, 1; $L__BB0_48: mad.lo.s32 %r1240, %r56, 1664525, 1013904223; and.b32 %r1241, %r1240, 16777215; cvt.rn.f32.u32 %f878, %r1241; mul.f32 %f879, %f878, 0f33800000; fma.rn.f32 %f880, %f879, 0f40000000, 0fBF800000; mad.lo.s32 %r1538, %r1240, 1664525, 1013904223; and.b32 %r1242, %r1538, 16777215; cvt.rn.f32.u32 %f881, %r1242; mul.f32 %f882, %f881, 0f33800000; fma.rn.f32 %f883, %f882, 0f40000000, 0fBF800000; mul.f32 %f884, %f19, %f883; mul.f32 %f885, %f20, %f883; mul.f32 %f886, %f21, %f883; fma.rn.f32 %f887, %f22, %f880, %f884; fma.rn.f32 %f888, %f23, %f880, %f885; fma.rn.f32 %f889, %f24, %f880, %f886; mul.f32 %f890, %f888, %f888; fma.rn.f32 %f891, %f887, %f887, %f890; fma.rn.f32 %f892, %f889, %f889, %f891; sqrt.rn.f32 %f893, %f892; rcp.rn.f32 %f894, %f893; mul.f32 %f895, %f887, %f894; mul.f32 %f896, %f888, %f894; mul.f32 %f897, %f889, %f894; mul.f32 %f898, %f1620, %f897; mul.f32 %f899, %f1621, %f896; sub.f32 %f900, %f898, %f899; mul.f32 %f901, %f1621, %f895; mul.f32 %f902, %f1619, %f897; sub.f32 %f903, %f901, %f902; mul.f32 %f904, %f1619, %f896; mul.f32 %f905, %f1620, %f895; sub.f32 %f906, %f904, %f905; mul.f32 %f907, %f895, 0f3F61E1D2; mul.f32 %f908, %f896, 0f3F61E1D2; mul.f32 %f909, %f897, 0f3F61E1D2; fma.rn.f32 %f910, %f25, %f907, %f16; fma.rn.f32 %f911, %f25, %f908, %f17; fma.rn.f32 %f912, %f25, %f909, %f18; mul.f32 %f913, %f900, 0fBF61E1E3; mul.f32 %f914, %f903, 0fBF61E1E3; mul.f32 %f915, %f906, 0fBF61E1E3; fma.rn.f32 %f121, %f25, %f913, %f910; fma.rn.f32 %f122, %f25, %f914, %f911; fma.rn.f32 %f123, %f25, %f915, %f912; sub.f32 %f916, %f121, %f16; sub.f32 %f917, %f122, %f17; sub.f32 %f918, %f123, %f18; mul.f32 %f919, %f917, %f917; fma.rn.f32 %f920, %f916, %f916, %f919; fma.rn.f32 %f921, %f918, %f918, %f920; sqrt.rn.f32 %f876, %f921; rcp.rn.f32 %f922, %f876; mul.f32 %f872, %f922, %f916; mul.f32 %f873, %f922, %f917; mul.f32 %f874, %f922, %f918; mov.f32 %f877, 0f00000000; // begin inline asm call(%r1169,%r1170,%r1171,%r1172,%r1173,%r1174,%r1175,%r1176,%r1177,%r1178,%r1179,%r1180,%r1181,%r1182,%r1183,%r1184,%r1185,%r1186,%r1187,%r1188,%r1189,%r1190,%r1191,%r1192,%r1193,%r1194,%r1195,%r1196,%r1197,%r1198,%r1199,%r1200),_optix_trace_typed_32,(%r1094,%rd4,%f16,%f17,%f18,%f872,%f873,%f874,%f382,%f876,%f877,%r1062,%r1094,%r1062,%r1060,%r1062,%r1062,%r1063,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094,%r1094); // end inline asm mov.b32 %f923, %r1169; setp.neu.f32 %p71, %f923, 0f3F800000; @%p71 bra $L__BB0_50; mov.f32 %f931, 0f6C4ECB8F; mov.u32 %r1279, 2; mov.u32 %r1281, 1; mov.u32 %r1282, 1065353216; mov.u32 %r1313, 0; // begin inline asm call(%r1243,%r1244,%r1245,%r1246,%r1247,%r1248,%r1249,%r1250,%r1251,%r1252,%r1253,%r1254,%r1255,%r1256,%r1257,%r1258,%r1259,%r1260,%r1261,%r1262,%r1263,%r1264,%r1265,%r1266,%r1267,%r1268,%r1269,%r1270,%r1271,%r1272,%r1273,%r1274),_optix_trace_typed_32,(%r1313,%rd4,%f121,%f122,%f123,%f68,%f69,%f70,%f382,%f931,%f877,%r1281,%r1313,%r1281,%r1279,%r1281,%r1281,%r1282,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313,%r1313); // end inline asm mov.b32 %f933, %r1243; add.f32 %f1632, %f1632, %f933; add.f32 %f1633, %f1633, 0f3F800000; mov.u16 %rs102, 1; $L__BB0_50: setp.ne.s16 %p72, %rs102, 0; @%p72 bra $L__BB0_52; mov.f32 %f941, 0f6C4ECB8F; mov.f32 %f942, 0f00000000; mov.u32 %r1350, 2; mov.u32 %r1352, 1; mov.u32 %r1353, 1065353216; mov.u32 %r1384, 0; // begin inline asm call(%r1314,%r1315,%r1316,%r1317,%r1318,%r1319,%r1320,%r1321,%r1322,%r1323,%r1324,%r1325,%r1326,%r1327,%r1328,%r1329,%r1330,%r1331,%r1332,%r1333,%r1334,%r1335,%r1336,%r1337,%r1338,%r1339,%r1340,%r1341,%r1342,%r1343,%r1344,%r1345),_optix_trace_typed_32,(%r1384,%rd4,%f16,%f17,%f18,%f68,%f69,%f70,%f382,%f941,%f942,%r1352,%r1384,%r1352,%r1350,%r1352,%r1352,%r1353,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384,%r1384); // end inline asm mov.b32 %f943, %r1314; add.f32 %f1632, %f1632, %f943; add.f32 %f1633, %f1633, 0f3F800000; $L__BB0_52: add.s32 %r1537, %r1537, 1; setp.lt.s32 %p73, %r1537, %r9; @%p73 bra $L__BB0_10; add.s32 %r1535, %r1535, 1; setp.lt.s32 %p74, %r1535, %r9; @%p74 bra $L__BB0_9; $L__BB0_54: setp.eq.s32 %p75, %r9, 0; mov.f32 %f1652, 0f3F800000; @%p75 bra $L__BB0_56; mov.f32 %f945, 0f3F800000; max.f32 %f946, %f1633, %f945; div.rn.f32 %f1652, %f1632, %f946; $L__BB0_56: ld.const.u32 %r1534, [params+544]; ld.const.f32 %f947, [params+524]; mul.f32 %f948, %f1652, %f947; ld.const.v2.f32 {%f949, %f950}, [params+528]; mul.f32 %f953, %f1652, %f949; mul.f32 %f954, %f1652, %f950; cvt.sat.f32.f32 %f955, %f15; mul.f32 %f956, %f948, %f955; mul.f32 %f957, %f953, %f955; mul.f32 %f958, %f954, %f955; setp.eq.s32 %p76, %r1534, 0; selp.f32 %f136, %f956, %f948, %p76; selp.f32 %f137, %f957, %f953, %p76; selp.f32 %f138, %f958, %f954, %p76; ld.const.u32 %r60, [params+104]; and.b32 %r1385, %r60, 8; setp.eq.s32 %p77, %r1385, 0; @%p77 bra $L__BB0_70; ld.const.u64 %rd79, [params+192]; cvta.to.global.u64 %rd16, %rd79; ld.const.u32 %r1386, [params+184]; mad.lo.s32 %r1387, %r1386, %r6, %r5; cvt.u64.u32 %rd17, %r1387; mov.f32 %f960, 0f3E68BA2E; cvt.rzi.f32.f32 %f961, %f960; add.f32 %f962, %f961, %f961; mov.f32 %f963, 0f3EE8BA2E; sub.f32 %f964, %f963, %f962; abs.f32 %f139, %f964; abs.f32 %f140, %f1652; setp.lt.f32 %p78, %f140, 0f00800000; mul.f32 %f965, %f140, 0f4B800000; selp.f32 %f966, %f965, %f140, %p78; selp.f32 %f967, 0fC3170000, 0fC2FE0000, %p78; mov.b32 %r1388, %f966; and.b32 %r1389, %r1388, 8388607; or.b32 %r1390, %r1389, 1065353216; mov.b32 %f968, %r1390; shr.u32 %r1391, %r1388, 23; cvt.rn.f32.u32 %f969, %r1391; add.f32 %f970, %f967, %f969; setp.gt.f32 %p79, %f968, 0f3FB504F3; mul.f32 %f971, %f968, 0f3F000000; add.f32 %f972, %f970, 0f3F800000; selp.f32 %f973, %f972, %f970, %p79; selp.f32 %f974, %f971, %f968, %p79; add.f32 %f975, %f974, 0fBF800000; add.f32 %f976, %f974, 0f3F800000; rcp.approx.ftz.f32 %f977, %f976; add.f32 %f978, %f975, %f975; mul.f32 %f979, %f978, %f977; mul.f32 %f980, %f979, %f979; mov.f32 %f981, 0f3C4CAF63; mov.f32 %f982, 0f3B18F0FE; fma.rn.f32 %f983, %f982, %f980, %f981; mov.f32 %f984, 0f3DAAAABD; fma.rn.f32 %f985, %f983, %f980, %f984; mul.rn.f32 %f986, %f985, %f980; mul.rn.f32 %f987, %f986, %f979; sub.f32 %f988, %f975, %f979; add.f32 %f989, %f988, %f988; neg.f32 %f990, %f979; fma.rn.f32 %f991, %f990, %f975, %f989; mul.rn.f32 %f992, %f977, %f991; add.f32 %f993, %f987, %f979; sub.f32 %f994, %f979, %f993; add.f32 %f995, %f987, %f994; add.f32 %f996, %f992, %f995; add.f32 %f997, %f993, %f996; sub.f32 %f998, %f993, %f997; add.f32 %f999, %f996, %f998; mov.f32 %f1000, 0f3F317200; mul.rn.f32 %f1001, %f973, %f1000; mov.f32 %f1002, 0f35BFBE8E; mul.rn.f32 %f1003, %f973, %f1002; add.f32 %f1004, %f1001, %f997; sub.f32 %f1005, %f1001, %f1004; add.f32 %f1006, %f997, %f1005; add.f32 %f1007, %f999, %f1006; add.f32 %f1008, %f1003, %f1007; add.f32 %f1009, %f1004, %f1008; sub.f32 %f1010, %f1004, %f1009; add.f32 %f1011, %f1008, %f1010; mul.rn.f32 %f1012, %f963, %f1009; neg.f32 %f1013, %f1012; fma.rn.f32 %f1014, %f963, %f1009, %f1013; fma.rn.f32 %f1015, %f963, %f1011, %f1014; mov.f32 %f1016, 0f00000000; fma.rn.f32 %f1017, %f1016, %f1009, %f1015; add.rn.f32 %f1018, %f1012, %f1017; neg.f32 %f1019, %f1018; add.rn.f32 %f1020, %f1012, %f1019; add.rn.f32 %f1021, %f1020, %f1017; mov.b32 %r1392, %f1018; setp.eq.s32 %p80, %r1392, 1118925336; add.s32 %r1393, %r1392, -1; mov.b32 %f1022, %r1393; add.f32 %f1023, %f1021, 0f37000000; selp.f32 %f141, %f1023, %f1021, %p80; selp.f32 %f1024, %f1022, %f1018, %p80; mov.f32 %f1025, 0f3FB8AA3B; mul.rn.f32 %f1026, %f1024, %f1025; cvt.rzi.f32.f32 %f1027, %f1026; abs.f32 %f1028, %f1027; setp.gt.f32 %p81, %f1028, 0f42FC0000; mov.b32 %r1394, %f1027; and.b32 %r1395, %r1394, -2147483648; or.b32 %r1396, %r1395, 1123811328; mov.b32 %f1029, %r1396; selp.f32 %f1030, %f1029, %f1027, %p81; mov.f32 %f1031, 0fBF317218; fma.rn.f32 %f1032, %f1030, %f1031, %f1024; mov.f32 %f1033, 0f3102E308; fma.rn.f32 %f1034, %f1030, %f1033, %f1032; mul.f32 %f1035, %f1034, 0f3FB8AA3B; add.f32 %f1036, %f1030, 0f4B40007F; mov.b32 %r1397, %f1036; shl.b32 %r1398, %r1397, 23; mov.b32 %f1037, %r1398; ex2.approx.ftz.f32 %f1038, %f1035; mul.f32 %f142, %f1038, %f1037; setp.eq.f32 %p82, %f142, 0f7F800000; mov.f32 %f1653, 0f7F800000; @%p82 bra $L__BB0_59; fma.rn.f32 %f1653, %f142, %f141, %f142; $L__BB0_59: setp.lt.f32 %p83, %f1652, 0f00000000; setp.eq.f32 %p84, %f139, 0f3F800000; and.pred %p1, %p83, %p84; setp.eq.f32 %p85, %f1652, 0f00000000; @%p85 bra $L__BB0_63; bra.uni $L__BB0_60; $L__BB0_63: add.f32 %f1043, %f1652, %f1652; selp.f32 %f1655, %f1043, 0f00000000, %p84; bra.uni $L__BB0_64; $L__BB0_167: mov.f32 %f1617, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs96, %f1617;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs95, %f1617;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs94, %f1617;} // end inline asm mov.u16 %rs97, 0; st.global.v4.u16 [%rd22], {%rs94, %rs95, %rs96, %rs97}; $L__BB0_168: and.b32 %r1531, %r61, 64; setp.eq.s32 %p208, %r1531, 0; @%p208 bra $L__BB0_170; ld.const.u64 %rd118, [params+208]; cvta.to.global.u64 %rd119, %rd118; ld.const.u32 %r1532, [params+200]; mad.lo.s32 %r1533, %r1532, %r6, %r5; mul.wide.u32 %rd120, %r1533, 4; add.s64 %rd121, %rd119, %rd120; mov.u16 %rs98, 0; st.global.v4.u8 [%rd121], {%rs98, %rs98, %rs98, %rs98}; bra.uni $L__BB0_170; $L__BB0_60: mov.b32 %r1399, %f1653; xor.b32 %r1400, %r1399, -2147483648; mov.b32 %f1039, %r1400; selp.f32 %f1655, %f1039, %f1653, %p1; setp.geu.f32 %p86, %f1652, 0f00000000; @%p86 bra $L__BB0_64; mov.f32 %f1040, 0f3EE8BA2E; cvt.rzi.f32.f32 %f1041, %f1040; setp.eq.f32 %p87, %f1041, 0f3EE8BA2E; @%p87 bra $L__BB0_64; mov.f32 %f1655, 0f7FFFFFFF; $L__BB0_64: add.f32 %f1044, %f140, 0f3EE8BA2E; mov.b32 %r1401, %f1044; setp.lt.s32 %p89, %r1401, 2139095040; @%p89 bra $L__BB0_69; setp.gtu.f32 %p90, %f140, 0f7F800000; @%p90 bra $L__BB0_68; bra.uni $L__BB0_66; $L__BB0_68: add.f32 %f1655, %f1652, 0f3EE8BA2E; bra.uni $L__BB0_69; $L__BB0_66: setp.neu.f32 %p91, %f140, 0f7F800000; @%p91 bra $L__BB0_69; selp.f32 %f1655, 0fFF800000, 0f7F800000, %p1; $L__BB0_69: mul.f32 %f1045, %f1655, 0f437F0000; setp.eq.f32 %p92, %f1652, 0f3F800000; selp.f32 %f1046, 0f437F0000, %f1045, %p92; cvt.rzi.u32.f32 %r1402, %f1046; shl.b64 %rd80, %rd17, 1; add.s64 %rd81, %rd16, %rd80; cvt.u16.u32 %rs34, %r1402; mov.u16 %rs35, 255; st.global.v2.u8 [%rd81], {%rs34, %rs35}; $L__BB0_70: and.b32 %r1403, %r60, 1; setp.eq.b32 %p93, %r1403, 1; mov.pred %p94, 0; xor.pred %p95, %p93, %p94; not.pred %p96, %p95; @%p96 bra $L__BB0_144; mov.f32 %f1048, 0f3E666666; cvt.rzi.f32.f32 %f1049, %f1048; add.f32 %f1050, %f1049, %f1049; mov.f32 %f1051, 0f3EE66666; sub.f32 %f1052, %f1051, %f1050; abs.f32 %f151, %f1052; abs.f32 %f152, %f136; setp.lt.f32 %p97, %f152, 0f00800000; mul.f32 %f1053, %f152, 0f4B800000; selp.f32 %f1054, %f1053, %f152, %p97; selp.f32 %f1055, 0fC3170000, 0fC2FE0000, %p97; mov.b32 %r1404, %f1054; and.b32 %r1405, %r1404, 8388607; or.b32 %r1406, %r1405, 1065353216; mov.b32 %f1056, %r1406; shr.u32 %r1407, %r1404, 23; cvt.rn.f32.u32 %f1057, %r1407; add.f32 %f1058, %f1055, %f1057; setp.gt.f32 %p98, %f1056, 0f3FB504F3; mul.f32 %f1059, %f1056, 0f3F000000; add.f32 %f1060, %f1058, 0f3F800000; selp.f32 %f1061, %f1060, %f1058, %p98; selp.f32 %f1062, %f1059, %f1056, %p98; add.f32 %f1063, %f1062, 0fBF800000; add.f32 %f1064, %f1062, 0f3F800000; rcp.approx.ftz.f32 %f1065, %f1064; add.f32 %f1066, %f1063, %f1063; mul.f32 %f1067, %f1066, %f1065; mul.f32 %f1068, %f1067, %f1067; mov.f32 %f1069, 0f3C4CAF63; mov.f32 %f1070, 0f3B18F0FE; fma.rn.f32 %f1071, %f1070, %f1068, %f1069; mov.f32 %f1072, 0f3DAAAABD; fma.rn.f32 %f1073, %f1071, %f1068, %f1072; mul.rn.f32 %f1074, %f1073, %f1068; mul.rn.f32 %f1075, %f1074, %f1067; sub.f32 %f1076, %f1063, %f1067; add.f32 %f1077, %f1076, %f1076; neg.f32 %f1078, %f1067; fma.rn.f32 %f1079, %f1078, %f1063, %f1077; mul.rn.f32 %f1080, %f1065, %f1079; add.f32 %f1081, %f1075, %f1067; sub.f32 %f1082, %f1067, %f1081; add.f32 %f1083, %f1075, %f1082; add.f32 %f1084, %f1080, %f1083; add.f32 %f1085, %f1081, %f1084; sub.f32 %f1086, %f1081, %f1085; add.f32 %f1087, %f1084, %f1086; mov.f32 %f1088, 0f3F317200; mul.rn.f32 %f1089, %f1061, %f1088; mov.f32 %f1090, 0f35BFBE8E; mul.rn.f32 %f1091, %f1061, %f1090; add.f32 %f1092, %f1089, %f1085; sub.f32 %f1093, %f1089, %f1092; add.f32 %f1094, %f1085, %f1093; add.f32 %f1095, %f1087, %f1094; add.f32 %f1096, %f1091, %f1095; add.f32 %f1097, %f1092, %f1096; sub.f32 %f1098, %f1092, %f1097; add.f32 %f1099, %f1096, %f1098; mul.rn.f32 %f1100, %f1051, %f1097; neg.f32 %f1101, %f1100; fma.rn.f32 %f1102, %f1051, %f1097, %f1101; fma.rn.f32 %f1103, %f1051, %f1099, %f1102; mov.f32 %f1104, 0f00000000; fma.rn.f32 %f1105, %f1104, %f1097, %f1103; add.rn.f32 %f1106, %f1100, %f1105; neg.f32 %f1107, %f1106; add.rn.f32 %f1108, %f1100, %f1107; add.rn.f32 %f1109, %f1108, %f1105; mov.b32 %r1408, %f1106; setp.eq.s32 %p99, %r1408, 1118925336; add.s32 %r1409, %r1408, -1; mov.b32 %f1110, %r1409; add.f32 %f1111, %f1109, 0f37000000; selp.f32 %f153, %f1111, %f1109, %p99; selp.f32 %f1112, %f1110, %f1106, %p99; mov.f32 %f1113, 0f3FB8AA3B; mul.rn.f32 %f1114, %f1112, %f1113; cvt.rzi.f32.f32 %f1115, %f1114; abs.f32 %f1116, %f1115; setp.gt.f32 %p100, %f1116, 0f42FC0000; mov.b32 %r1410, %f1115; and.b32 %r1411, %r1410, -2147483648; or.b32 %r1412, %r1411, 1123811328; mov.b32 %f1117, %r1412; selp.f32 %f1118, %f1117, %f1115, %p100; mov.f32 %f1119, 0fBF317218; fma.rn.f32 %f1120, %f1118, %f1119, %f1112; mov.f32 %f1121, 0f3102E308; fma.rn.f32 %f1122, %f1118, %f1121, %f1120; mul.f32 %f1123, %f1122, 0f3FB8AA3B; add.f32 %f1124, %f1118, 0f4B40007F; mov.b32 %r1413, %f1124; shl.b32 %r1414, %r1413, 23; mov.b32 %f1125, %r1414; ex2.approx.ftz.f32 %f1126, %f1123; mul.f32 %f154, %f1126, %f1125; setp.eq.f32 %p101, %f154, 0f7F800000; mov.f32 %f1656, 0f7F800000; @%p101 bra $L__BB0_73; fma.rn.f32 %f1656, %f154, %f153, %f154; $L__BB0_73: setp.lt.f32 %p102, %f136, 0f00000000; setp.eq.f32 %p103, %f151, 0f3F800000; and.pred %p2, %p102, %p103; setp.eq.f32 %p104, %f136, 0f00000000; @%p104 bra $L__BB0_77; bra.uni $L__BB0_74; $L__BB0_77: add.f32 %f1131, %f136, %f136; selp.f32 %f1658, %f1131, 0f00000000, %p103; bra.uni $L__BB0_78; $L__BB0_74: mov.b32 %r1415, %f1656; xor.b32 %r1416, %r1415, -2147483648; mov.b32 %f1127, %r1416; selp.f32 %f1658, %f1127, %f1656, %p2; setp.geu.f32 %p105, %f136, 0f00000000; @%p105 bra $L__BB0_78; mov.f32 %f1128, 0f3EE66666; cvt.rzi.f32.f32 %f1129, %f1128; setp.eq.f32 %p106, %f1129, 0f3EE66666; @%p106 bra $L__BB0_78; mov.f32 %f1658, 0f7FFFFFFF; $L__BB0_78: add.f32 %f1132, %f152, 0f3EE66666; mov.b32 %r1417, %f1132; setp.lt.s32 %p108, %r1417, 2139095040; @%p108 bra $L__BB0_83; setp.gtu.f32 %p109, %f152, 0f7F800000; @%p109 bra $L__BB0_82; bra.uni $L__BB0_80; $L__BB0_82: add.f32 %f1658, %f136, 0f3EE66666; bra.uni $L__BB0_83; $L__BB0_80: setp.neu.f32 %p110, %f152, 0f7F800000; @%p110 bra $L__BB0_83; selp.f32 %f1658, 0fFF800000, 0f7F800000, %p2; $L__BB0_83: setp.eq.f32 %p111, %f136, 0f3F800000; selp.f32 %f163, 0f3F800000, %f1658, %p111; abs.f32 %f164, %f137; setp.lt.f32 %p112, %f164, 0f00800000; mul.f32 %f1134, %f164, 0f4B800000; selp.f32 %f1135, %f1134, %f164, %p112; selp.f32 %f1136, 0fC3170000, 0fC2FE0000, %p112; mov.b32 %r1418, %f1135; and.b32 %r1419, %r1418, 8388607; or.b32 %r1420, %r1419, 1065353216; mov.b32 %f1137, %r1420; shr.u32 %r1421, %r1418, 23; cvt.rn.f32.u32 %f1138, %r1421; add.f32 %f1139, %f1136, %f1138; setp.gt.f32 %p113, %f1137, 0f3FB504F3; mul.f32 %f1140, %f1137, 0f3F000000; add.f32 %f1141, %f1139, 0f3F800000; selp.f32 %f1142, %f1141, %f1139, %p113; selp.f32 %f1143, %f1140, %f1137, %p113; add.f32 %f1144, %f1143, 0fBF800000; add.f32 %f1145, %f1143, 0f3F800000; rcp.approx.ftz.f32 %f1146, %f1145; add.f32 %f1147, %f1144, %f1144; mul.f32 %f1148, %f1147, %f1146; mul.f32 %f1149, %f1148, %f1148; mov.f32 %f1150, 0f3C4CAF63; mov.f32 %f1151, 0f3B18F0FE; fma.rn.f32 %f1152, %f1151, %f1149, %f1150; mov.f32 %f1153, 0f3DAAAABD; fma.rn.f32 %f1154, %f1152, %f1149, %f1153; mul.rn.f32 %f1155, %f1154, %f1149; mul.rn.f32 %f1156, %f1155, %f1148; sub.f32 %f1157, %f1144, %f1148; add.f32 %f1158, %f1157, %f1157; neg.f32 %f1159, %f1148; fma.rn.f32 %f1160, %f1159, %f1144, %f1158; mul.rn.f32 %f1161, %f1146, %f1160; add.f32 %f1162, %f1156, %f1148; sub.f32 %f1163, %f1148, %f1162; add.f32 %f1164, %f1156, %f1163; add.f32 %f1165, %f1161, %f1164; add.f32 %f1166, %f1162, %f1165; sub.f32 %f1167, %f1162, %f1166; add.f32 %f1168, %f1165, %f1167; mov.f32 %f1169, 0f3F317200; mul.rn.f32 %f1170, %f1142, %f1169; mov.f32 %f1171, 0f35BFBE8E; mul.rn.f32 %f1172, %f1142, %f1171; add.f32 %f1173, %f1170, %f1166; sub.f32 %f1174, %f1170, %f1173; add.f32 %f1175, %f1166, %f1174; add.f32 %f1176, %f1168, %f1175; add.f32 %f1177, %f1172, %f1176; add.f32 %f1178, %f1173, %f1177; sub.f32 %f1179, %f1173, %f1178; add.f32 %f1180, %f1177, %f1179; mov.f32 %f1181, 0f3EE66666; mul.rn.f32 %f1182, %f1181, %f1178; neg.f32 %f1183, %f1182; fma.rn.f32 %f1184, %f1181, %f1178, %f1183; fma.rn.f32 %f1185, %f1181, %f1180, %f1184; mov.f32 %f1186, 0f00000000; fma.rn.f32 %f1187, %f1186, %f1178, %f1185; add.rn.f32 %f1188, %f1182, %f1187; neg.f32 %f1189, %f1188; add.rn.f32 %f1190, %f1182, %f1189; add.rn.f32 %f1191, %f1190, %f1187; mov.b32 %r1422, %f1188; setp.eq.s32 %p114, %r1422, 1118925336; add.s32 %r1423, %r1422, -1; mov.b32 %f1192, %r1423; add.f32 %f1193, %f1191, 0f37000000; selp.f32 %f165, %f1193, %f1191, %p114; selp.f32 %f1194, %f1192, %f1188, %p114; mov.f32 %f1195, 0f3FB8AA3B; mul.rn.f32 %f1196, %f1194, %f1195; cvt.rzi.f32.f32 %f1197, %f1196; abs.f32 %f1198, %f1197; setp.gt.f32 %p115, %f1198, 0f42FC0000; mov.b32 %r1424, %f1197; and.b32 %r1425, %r1424, -2147483648; or.b32 %r1426, %r1425, 1123811328; mov.b32 %f1199, %r1426; selp.f32 %f1200, %f1199, %f1197, %p115; mov.f32 %f1201, 0fBF317218; fma.rn.f32 %f1202, %f1200, %f1201, %f1194; mov.f32 %f1203, 0f3102E308; fma.rn.f32 %f1204, %f1200, %f1203, %f1202; mul.f32 %f1205, %f1204, 0f3FB8AA3B; add.f32 %f1206, %f1200, 0f4B40007F; mov.b32 %r1427, %f1206; shl.b32 %r1428, %r1427, 23; mov.b32 %f1207, %r1428; ex2.approx.ftz.f32 %f1208, %f1205; mul.f32 %f166, %f1208, %f1207; setp.eq.f32 %p116, %f166, 0f7F800000; mov.f32 %f1659, 0f7F800000; @%p116 bra $L__BB0_85; fma.rn.f32 %f1659, %f166, %f165, %f166; $L__BB0_85: setp.lt.f32 %p117, %f137, 0f00000000; and.pred %p3, %p117, %p103; setp.eq.f32 %p119, %f137, 0f00000000; @%p119 bra $L__BB0_89; bra.uni $L__BB0_86; $L__BB0_89: add.f32 %f1213, %f137, %f137; selp.f32 %f1661, %f1213, 0f00000000, %p103; bra.uni $L__BB0_90; $L__BB0_86: mov.b32 %r1429, %f1659; xor.b32 %r1430, %r1429, -2147483648; mov.b32 %f1209, %r1430; selp.f32 %f1661, %f1209, %f1659, %p3; setp.geu.f32 %p120, %f137, 0f00000000; @%p120 bra $L__BB0_90; mov.f32 %f1210, 0f3EE66666; cvt.rzi.f32.f32 %f1211, %f1210; setp.eq.f32 %p121, %f1211, 0f3EE66666; @%p121 bra $L__BB0_90; mov.f32 %f1661, 0f7FFFFFFF; $L__BB0_90: add.f32 %f1214, %f164, 0f3EE66666; mov.b32 %r1431, %f1214; setp.lt.s32 %p123, %r1431, 2139095040; @%p123 bra $L__BB0_95; setp.gtu.f32 %p124, %f164, 0f7F800000; @%p124 bra $L__BB0_94; bra.uni $L__BB0_92; $L__BB0_94: add.f32 %f1661, %f137, 0f3EE66666; bra.uni $L__BB0_95; $L__BB0_92: setp.neu.f32 %p125, %f164, 0f7F800000; @%p125 bra $L__BB0_95; selp.f32 %f1661, 0fFF800000, 0f7F800000, %p3; $L__BB0_95: setp.eq.f32 %p126, %f137, 0f3F800000; selp.f32 %f175, 0f3F800000, %f1661, %p126; abs.f32 %f176, %f138; setp.lt.f32 %p127, %f176, 0f00800000; mul.f32 %f1216, %f176, 0f4B800000; selp.f32 %f1217, %f1216, %f176, %p127; selp.f32 %f1218, 0fC3170000, 0fC2FE0000, %p127; mov.b32 %r1432, %f1217; and.b32 %r1433, %r1432, 8388607; or.b32 %r1434, %r1433, 1065353216; mov.b32 %f1219, %r1434; shr.u32 %r1435, %r1432, 23; cvt.rn.f32.u32 %f1220, %r1435; add.f32 %f1221, %f1218, %f1220; setp.gt.f32 %p128, %f1219, 0f3FB504F3; mul.f32 %f1222, %f1219, 0f3F000000; add.f32 %f1223, %f1221, 0f3F800000; selp.f32 %f1224, %f1223, %f1221, %p128; selp.f32 %f1225, %f1222, %f1219, %p128; add.f32 %f1226, %f1225, 0fBF800000; add.f32 %f1227, %f1225, 0f3F800000; rcp.approx.ftz.f32 %f1228, %f1227; add.f32 %f1229, %f1226, %f1226; mul.f32 %f1230, %f1229, %f1228; mul.f32 %f1231, %f1230, %f1230; mov.f32 %f1232, 0f3C4CAF63; mov.f32 %f1233, 0f3B18F0FE; fma.rn.f32 %f1234, %f1233, %f1231, %f1232; mov.f32 %f1235, 0f3DAAAABD; fma.rn.f32 %f1236, %f1234, %f1231, %f1235; mul.rn.f32 %f1237, %f1236, %f1231; mul.rn.f32 %f1238, %f1237, %f1230; sub.f32 %f1239, %f1226, %f1230; add.f32 %f1240, %f1239, %f1239; neg.f32 %f1241, %f1230; fma.rn.f32 %f1242, %f1241, %f1226, %f1240; mul.rn.f32 %f1243, %f1228, %f1242; add.f32 %f1244, %f1238, %f1230; sub.f32 %f1245, %f1230, %f1244; add.f32 %f1246, %f1238, %f1245; add.f32 %f1247, %f1243, %f1246; add.f32 %f1248, %f1244, %f1247; sub.f32 %f1249, %f1244, %f1248; add.f32 %f1250, %f1247, %f1249; mov.f32 %f1251, 0f3F317200; mul.rn.f32 %f1252, %f1224, %f1251; mov.f32 %f1253, 0f35BFBE8E; mul.rn.f32 %f1254, %f1224, %f1253; add.f32 %f1255, %f1252, %f1248; sub.f32 %f1256, %f1252, %f1255; add.f32 %f1257, %f1248, %f1256; add.f32 %f1258, %f1250, %f1257; add.f32 %f1259, %f1254, %f1258; add.f32 %f1260, %f1255, %f1259; sub.f32 %f1261, %f1255, %f1260; add.f32 %f1262, %f1259, %f1261; mov.f32 %f1263, 0f3EE66666; mul.rn.f32 %f1264, %f1263, %f1260; neg.f32 %f1265, %f1264; fma.rn.f32 %f1266, %f1263, %f1260, %f1265; fma.rn.f32 %f1267, %f1263, %f1262, %f1266; mov.f32 %f1268, 0f00000000; fma.rn.f32 %f1269, %f1268, %f1260, %f1267; add.rn.f32 %f1270, %f1264, %f1269; neg.f32 %f1271, %f1270; add.rn.f32 %f1272, %f1264, %f1271; add.rn.f32 %f1273, %f1272, %f1269; mov.b32 %r1436, %f1270; setp.eq.s32 %p129, %r1436, 1118925336; add.s32 %r1437, %r1436, -1; mov.b32 %f1274, %r1437; add.f32 %f1275, %f1273, 0f37000000; selp.f32 %f177, %f1275, %f1273, %p129; selp.f32 %f1276, %f1274, %f1270, %p129; mov.f32 %f1277, 0f3FB8AA3B; mul.rn.f32 %f1278, %f1276, %f1277; cvt.rzi.f32.f32 %f1279, %f1278; abs.f32 %f1280, %f1279; setp.gt.f32 %p130, %f1280, 0f42FC0000; mov.b32 %r1438, %f1279; and.b32 %r1439, %r1438, -2147483648; or.b32 %r1440, %r1439, 1123811328; mov.b32 %f1281, %r1440; selp.f32 %f1282, %f1281, %f1279, %p130; mov.f32 %f1283, 0fBF317218; fma.rn.f32 %f1284, %f1282, %f1283, %f1276; mov.f32 %f1285, 0f3102E308; fma.rn.f32 %f1286, %f1282, %f1285, %f1284; mul.f32 %f1287, %f1286, 0f3FB8AA3B; add.f32 %f1288, %f1282, 0f4B40007F; mov.b32 %r1441, %f1288; shl.b32 %r1442, %r1441, 23; mov.b32 %f1289, %r1442; ex2.approx.ftz.f32 %f1290, %f1287; mul.f32 %f178, %f1290, %f1289; setp.eq.f32 %p131, %f178, 0f7F800000; mov.f32 %f1662, 0f7F800000; @%p131 bra $L__BB0_97; fma.rn.f32 %f1662, %f178, %f177, %f178; $L__BB0_97: setp.lt.f32 %p132, %f138, 0f00000000; and.pred %p4, %p132, %p103; setp.eq.f32 %p134, %f138, 0f00000000; @%p134 bra $L__BB0_101; bra.uni $L__BB0_98; $L__BB0_101: add.f32 %f1295, %f138, %f138; selp.f32 %f1664, %f1295, 0f00000000, %p103; bra.uni $L__BB0_102; $L__BB0_98: mov.b32 %r1443, %f1662; xor.b32 %r1444, %r1443, -2147483648; mov.b32 %f1291, %r1444; selp.f32 %f1664, %f1291, %f1662, %p4; setp.geu.f32 %p135, %f138, 0f00000000; @%p135 bra $L__BB0_102; mov.f32 %f1292, 0f3EE66666; cvt.rzi.f32.f32 %f1293, %f1292; setp.eq.f32 %p136, %f1293, 0f3EE66666; @%p136 bra $L__BB0_102; mov.f32 %f1664, 0f7FFFFFFF; $L__BB0_102: add.f32 %f1296, %f176, 0f3EE66666; mov.b32 %r1445, %f1296; setp.lt.s32 %p138, %r1445, 2139095040; @%p138 bra $L__BB0_107; setp.gtu.f32 %p139, %f176, 0f7F800000; @%p139 bra $L__BB0_106; bra.uni $L__BB0_104; $L__BB0_106: add.f32 %f1664, %f138, 0f3EE66666; bra.uni $L__BB0_107; $L__BB0_104: setp.neu.f32 %p140, %f176, 0f7F800000; @%p140 bra $L__BB0_107; selp.f32 %f1664, 0fFF800000, 0f7F800000, %p4; $L__BB0_107: setp.eq.f32 %p141, %f138, 0f3F800000; mov.f32 %f1298, 0f3F800000; selp.f32 %f1299, 0f3F800000, %f1664, %p141; ld.const.u64 %rd82, [params+144]; cvta.to.global.u64 %rd18, %rd82; ld.const.u32 %r1446, [params+136]; mad.lo.s32 %r1447, %r1446, %r6, %r5; cvt.u64.u32 %rd19, %r1447; min.f32 %f1300, %f163, %f1298; mov.f32 %f1301, 0f00000000; max.f32 %f187, %f1301, %f1300; min.f32 %f1302, %f175, %f1298; max.f32 %f188, %f1301, %f1302; min.f32 %f1303, %f1299, %f1298; max.f32 %f189, %f1301, %f1303; mov.f32 %f1304, 0f3E555555; cvt.rzi.f32.f32 %f1305, %f1304; add.f32 %f1306, %f1305, %f1305; mov.f32 %f1307, 0f3ED55555; sub.f32 %f1308, %f1307, %f1306; abs.f32 %f190, %f1308; abs.f32 %f191, %f187; setp.lt.f32 %p142, %f191, 0f00800000; mul.f32 %f1309, %f191, 0f4B800000; selp.f32 %f1310, %f1309, %f191, %p142; selp.f32 %f1311, 0fC3170000, 0fC2FE0000, %p142; mov.b32 %r1448, %f1310; and.b32 %r1449, %r1448, 8388607; or.b32 %r1450, %r1449, 1065353216; mov.b32 %f1312, %r1450; shr.u32 %r1451, %r1448, 23; cvt.rn.f32.u32 %f1313, %r1451; add.f32 %f1314, %f1311, %f1313; setp.gt.f32 %p143, %f1312, 0f3FB504F3; mul.f32 %f1315, %f1312, 0f3F000000; add.f32 %f1316, %f1314, 0f3F800000; selp.f32 %f1317, %f1316, %f1314, %p143; selp.f32 %f1318, %f1315, %f1312, %p143; add.f32 %f1319, %f1318, 0fBF800000; add.f32 %f1320, %f1318, 0f3F800000; rcp.approx.ftz.f32 %f1321, %f1320; add.f32 %f1322, %f1319, %f1319; mul.f32 %f1323, %f1322, %f1321; mul.f32 %f1324, %f1323, %f1323; mov.f32 %f1325, 0f3C4CAF63; mov.f32 %f1326, 0f3B18F0FE; fma.rn.f32 %f1327, %f1326, %f1324, %f1325; mov.f32 %f1328, 0f3DAAAABD; fma.rn.f32 %f1329, %f1327, %f1324, %f1328; mul.rn.f32 %f1330, %f1329, %f1324; mul.rn.f32 %f1331, %f1330, %f1323; sub.f32 %f1332, %f1319, %f1323; add.f32 %f1333, %f1332, %f1332; neg.f32 %f1334, %f1323; fma.rn.f32 %f1335, %f1334, %f1319, %f1333; mul.rn.f32 %f1336, %f1321, %f1335; add.f32 %f1337, %f1331, %f1323; sub.f32 %f1338, %f1323, %f1337; add.f32 %f1339, %f1331, %f1338; add.f32 %f1340, %f1336, %f1339; add.f32 %f1341, %f1337, %f1340; sub.f32 %f1342, %f1337, %f1341; add.f32 %f1343, %f1340, %f1342; mov.f32 %f1344, 0f3F317200; mul.rn.f32 %f1345, %f1317, %f1344; mov.f32 %f1346, 0f35BFBE8E; mul.rn.f32 %f1347, %f1317, %f1346; add.f32 %f1348, %f1345, %f1341; sub.f32 %f1349, %f1345, %f1348; add.f32 %f1350, %f1341, %f1349; add.f32 %f1351, %f1343, %f1350; add.f32 %f1352, %f1347, %f1351; add.f32 %f1353, %f1348, %f1352; sub.f32 %f1354, %f1348, %f1353; add.f32 %f1355, %f1352, %f1354; mul.rn.f32 %f1356, %f1307, %f1353; neg.f32 %f1357, %f1356; fma.rn.f32 %f1358, %f1307, %f1353, %f1357; fma.rn.f32 %f1359, %f1307, %f1355, %f1358; fma.rn.f32 %f1360, %f1301, %f1353, %f1359; add.rn.f32 %f1361, %f1356, %f1360; neg.f32 %f1362, %f1361; add.rn.f32 %f1363, %f1356, %f1362; add.rn.f32 %f1364, %f1363, %f1360; mov.b32 %r1452, %f1361; setp.eq.s32 %p144, %r1452, 1118925336; add.s32 %r1453, %r1452, -1; mov.b32 %f1365, %r1453; add.f32 %f1366, %f1364, 0f37000000; selp.f32 %f192, %f1366, %f1364, %p144; selp.f32 %f1367, %f1365, %f1361, %p144; mov.f32 %f1368, 0f3FB8AA3B; mul.rn.f32 %f1369, %f1367, %f1368; cvt.rzi.f32.f32 %f1370, %f1369; abs.f32 %f1371, %f1370; setp.gt.f32 %p145, %f1371, 0f42FC0000; mov.b32 %r1454, %f1370; and.b32 %r1455, %r1454, -2147483648; or.b32 %r1456, %r1455, 1123811328; mov.b32 %f1372, %r1456; selp.f32 %f1373, %f1372, %f1370, %p145; mov.f32 %f1374, 0fBF317218; fma.rn.f32 %f1375, %f1373, %f1374, %f1367; mov.f32 %f1376, 0f3102E308; fma.rn.f32 %f1377, %f1373, %f1376, %f1375; mul.f32 %f1378, %f1377, 0f3FB8AA3B; add.f32 %f1379, %f1373, 0f4B40007F; mov.b32 %r1457, %f1379; shl.b32 %r1458, %r1457, 23; mov.b32 %f1380, %r1458; ex2.approx.ftz.f32 %f1381, %f1378; mul.f32 %f193, %f1381, %f1380; setp.eq.f32 %p146, %f193, 0f7F800000; mov.f32 %f1665, 0f7F800000; @%p146 bra $L__BB0_109; fma.rn.f32 %f1665, %f193, %f192, %f193; $L__BB0_109: setp.lt.f32 %p147, %f187, 0f00000000; setp.eq.f32 %p148, %f190, 0f3F800000; and.pred %p5, %p147, %p148; setp.eq.f32 %p149, %f187, 0f00000000; @%p149 bra $L__BB0_113; bra.uni $L__BB0_110; $L__BB0_113: add.f32 %f1386, %f187, %f187; selp.f32 %f1667, %f1386, 0f00000000, %p148; bra.uni $L__BB0_114; $L__BB0_110: mov.b32 %r1459, %f1665; xor.b32 %r1460, %r1459, -2147483648; mov.b32 %f1382, %r1460; selp.f32 %f1667, %f1382, %f1665, %p5; setp.geu.f32 %p150, %f187, 0f00000000; @%p150 bra $L__BB0_114; mov.f32 %f1383, 0f3ED55555; cvt.rzi.f32.f32 %f1384, %f1383; setp.eq.f32 %p151, %f1384, 0f3ED55555; @%p151 bra $L__BB0_114; mov.f32 %f1667, 0f7FFFFFFF; $L__BB0_114: add.f32 %f1387, %f191, 0f3ED55555; mov.b32 %r1461, %f1387; setp.lt.s32 %p153, %r1461, 2139095040; @%p153 bra $L__BB0_119; setp.gtu.f32 %p154, %f191, 0f7F800000; @%p154 bra $L__BB0_118; bra.uni $L__BB0_116; $L__BB0_118: add.f32 %f1667, %f187, 0f3ED55555; bra.uni $L__BB0_119; $L__BB0_116: setp.neu.f32 %p155, %f191, 0f7F800000; @%p155 bra $L__BB0_119; selp.f32 %f1667, 0fFF800000, 0f7F800000, %p5; $L__BB0_119: abs.f32 %f202, %f188; setp.lt.f32 %p156, %f202, 0f00800000; mul.f32 %f1389, %f202, 0f4B800000; selp.f32 %f1390, %f1389, %f202, %p156; selp.f32 %f1391, 0fC3170000, 0fC2FE0000, %p156; mov.b32 %r1462, %f1390; and.b32 %r1463, %r1462, 8388607; or.b32 %r1464, %r1463, 1065353216; mov.b32 %f1392, %r1464; shr.u32 %r1465, %r1462, 23; cvt.rn.f32.u32 %f1393, %r1465; add.f32 %f1394, %f1391, %f1393; setp.gt.f32 %p157, %f1392, 0f3FB504F3; mul.f32 %f1395, %f1392, 0f3F000000; add.f32 %f1396, %f1394, 0f3F800000; selp.f32 %f1397, %f1396, %f1394, %p157; selp.f32 %f1398, %f1395, %f1392, %p157; add.f32 %f1399, %f1398, 0fBF800000; add.f32 %f1400, %f1398, 0f3F800000; rcp.approx.ftz.f32 %f1401, %f1400; add.f32 %f1402, %f1399, %f1399; mul.f32 %f1403, %f1402, %f1401; mul.f32 %f1404, %f1403, %f1403; mov.f32 %f1405, 0f3C4CAF63; mov.f32 %f1406, 0f3B18F0FE; fma.rn.f32 %f1407, %f1406, %f1404, %f1405; mov.f32 %f1408, 0f3DAAAABD; fma.rn.f32 %f1409, %f1407, %f1404, %f1408; mul.rn.f32 %f1410, %f1409, %f1404; mul.rn.f32 %f1411, %f1410, %f1403; sub.f32 %f1412, %f1399, %f1403; add.f32 %f1413, %f1412, %f1412; neg.f32 %f1414, %f1403; fma.rn.f32 %f1415, %f1414, %f1399, %f1413; mul.rn.f32 %f1416, %f1401, %f1415; add.f32 %f1417, %f1411, %f1403; sub.f32 %f1418, %f1403, %f1417; add.f32 %f1419, %f1411, %f1418; add.f32 %f1420, %f1416, %f1419; add.f32 %f1421, %f1417, %f1420; sub.f32 %f1422, %f1417, %f1421; add.f32 %f1423, %f1420, %f1422; mov.f32 %f1424, 0f3F317200; mul.rn.f32 %f1425, %f1397, %f1424; mov.f32 %f1426, 0f35BFBE8E; mul.rn.f32 %f1427, %f1397, %f1426; add.f32 %f1428, %f1425, %f1421; sub.f32 %f1429, %f1425, %f1428; add.f32 %f1430, %f1421, %f1429; add.f32 %f1431, %f1423, %f1430; add.f32 %f1432, %f1427, %f1431; add.f32 %f1433, %f1428, %f1432; sub.f32 %f1434, %f1428, %f1433; add.f32 %f1435, %f1432, %f1434; mov.f32 %f1436, 0f3ED55555; mul.rn.f32 %f1437, %f1436, %f1433; neg.f32 %f1438, %f1437; fma.rn.f32 %f1439, %f1436, %f1433, %f1438; fma.rn.f32 %f1440, %f1436, %f1435, %f1439; mov.f32 %f1441, 0f00000000; fma.rn.f32 %f1442, %f1441, %f1433, %f1440; add.rn.f32 %f1443, %f1437, %f1442; neg.f32 %f1444, %f1443; add.rn.f32 %f1445, %f1437, %f1444; add.rn.f32 %f1446, %f1445, %f1442; mov.b32 %r1466, %f1443; setp.eq.s32 %p158, %r1466, 1118925336; add.s32 %r1467, %r1466, -1; mov.b32 %f1447, %r1467; add.f32 %f1448, %f1446, 0f37000000; selp.f32 %f203, %f1448, %f1446, %p158; selp.f32 %f1449, %f1447, %f1443, %p158; mov.f32 %f1450, 0f3FB8AA3B; mul.rn.f32 %f1451, %f1449, %f1450; cvt.rzi.f32.f32 %f1452, %f1451; abs.f32 %f1453, %f1452; setp.gt.f32 %p159, %f1453, 0f42FC0000; mov.b32 %r1468, %f1452; and.b32 %r1469, %r1468, -2147483648; or.b32 %r1470, %r1469, 1123811328; mov.b32 %f1454, %r1470; selp.f32 %f1455, %f1454, %f1452, %p159; mov.f32 %f1456, 0fBF317218; fma.rn.f32 %f1457, %f1455, %f1456, %f1449; mov.f32 %f1458, 0f3102E308; fma.rn.f32 %f1459, %f1455, %f1458, %f1457; mul.f32 %f1460, %f1459, 0f3FB8AA3B; add.f32 %f1461, %f1455, 0f4B40007F; mov.b32 %r1471, %f1461; shl.b32 %r1472, %r1471, 23; mov.b32 %f1462, %r1472; ex2.approx.ftz.f32 %f1463, %f1460; mul.f32 %f204, %f1463, %f1462; setp.eq.f32 %p160, %f204, 0f7F800000; mov.f32 %f1668, 0f7F800000; @%p160 bra $L__BB0_121; fma.rn.f32 %f1668, %f204, %f203, %f204; $L__BB0_121: setp.lt.f32 %p161, %f188, 0f00000000; and.pred %p6, %p161, %p148; setp.eq.f32 %p163, %f188, 0f00000000; @%p163 bra $L__BB0_125; bra.uni $L__BB0_122; $L__BB0_125: add.f32 %f1468, %f188, %f188; selp.f32 %f1670, %f1468, 0f00000000, %p148; bra.uni $L__BB0_126; $L__BB0_122: mov.b32 %r1473, %f1668; xor.b32 %r1474, %r1473, -2147483648; mov.b32 %f1464, %r1474; selp.f32 %f1670, %f1464, %f1668, %p6; setp.geu.f32 %p164, %f188, 0f00000000; @%p164 bra $L__BB0_126; mov.f32 %f1465, 0f3ED55555; cvt.rzi.f32.f32 %f1466, %f1465; setp.eq.f32 %p165, %f1466, 0f3ED55555; @%p165 bra $L__BB0_126; mov.f32 %f1670, 0f7FFFFFFF; $L__BB0_126: add.f32 %f1469, %f202, 0f3ED55555; mov.b32 %r1475, %f1469; setp.lt.s32 %p167, %r1475, 2139095040; @%p167 bra $L__BB0_131; setp.gtu.f32 %p168, %f202, 0f7F800000; @%p168 bra $L__BB0_130; bra.uni $L__BB0_128; $L__BB0_130: add.f32 %f1670, %f188, 0f3ED55555; bra.uni $L__BB0_131; $L__BB0_128: setp.neu.f32 %p169, %f202, 0f7F800000; @%p169 bra $L__BB0_131; selp.f32 %f1670, 0fFF800000, 0f7F800000, %p6; $L__BB0_131: abs.f32 %f213, %f189; setp.lt.f32 %p170, %f213, 0f00800000; mul.f32 %f1471, %f213, 0f4B800000; selp.f32 %f1472, %f1471, %f213, %p170; selp.f32 %f1473, 0fC3170000, 0fC2FE0000, %p170; mov.b32 %r1476, %f1472; and.b32 %r1477, %r1476, 8388607; or.b32 %r1478, %r1477, 1065353216; mov.b32 %f1474, %r1478; shr.u32 %r1479, %r1476, 23; cvt.rn.f32.u32 %f1475, %r1479; add.f32 %f1476, %f1473, %f1475; setp.gt.f32 %p171, %f1474, 0f3FB504F3; mul.f32 %f1477, %f1474, 0f3F000000; add.f32 %f1478, %f1476, 0f3F800000; selp.f32 %f1479, %f1478, %f1476, %p171; selp.f32 %f1480, %f1477, %f1474, %p171; add.f32 %f1481, %f1480, 0fBF800000; add.f32 %f1482, %f1480, 0f3F800000; rcp.approx.ftz.f32 %f1483, %f1482; add.f32 %f1484, %f1481, %f1481; mul.f32 %f1485, %f1484, %f1483; mul.f32 %f1486, %f1485, %f1485; mov.f32 %f1487, 0f3C4CAF63; mov.f32 %f1488, 0f3B18F0FE; fma.rn.f32 %f1489, %f1488, %f1486, %f1487; mov.f32 %f1490, 0f3DAAAABD; fma.rn.f32 %f1491, %f1489, %f1486, %f1490; mul.rn.f32 %f1492, %f1491, %f1486; mul.rn.f32 %f1493, %f1492, %f1485; sub.f32 %f1494, %f1481, %f1485; add.f32 %f1495, %f1494, %f1494; neg.f32 %f1496, %f1485; fma.rn.f32 %f1497, %f1496, %f1481, %f1495; mul.rn.f32 %f1498, %f1483, %f1497; add.f32 %f1499, %f1493, %f1485; sub.f32 %f1500, %f1485, %f1499; add.f32 %f1501, %f1493, %f1500; add.f32 %f1502, %f1498, %f1501; add.f32 %f1503, %f1499, %f1502; sub.f32 %f1504, %f1499, %f1503; add.f32 %f1505, %f1502, %f1504; mov.f32 %f1506, 0f3F317200; mul.rn.f32 %f1507, %f1479, %f1506; mov.f32 %f1508, 0f35BFBE8E; mul.rn.f32 %f1509, %f1479, %f1508; add.f32 %f1510, %f1507, %f1503; sub.f32 %f1511, %f1507, %f1510; add.f32 %f1512, %f1503, %f1511; add.f32 %f1513, %f1505, %f1512; add.f32 %f1514, %f1509, %f1513; add.f32 %f1515, %f1510, %f1514; sub.f32 %f1516, %f1510, %f1515; add.f32 %f1517, %f1514, %f1516; mov.f32 %f1518, 0f3ED55555; mul.rn.f32 %f1519, %f1518, %f1515; neg.f32 %f1520, %f1519; fma.rn.f32 %f1521, %f1518, %f1515, %f1520; fma.rn.f32 %f1522, %f1518, %f1517, %f1521; mov.f32 %f1523, 0f00000000; fma.rn.f32 %f1524, %f1523, %f1515, %f1522; add.rn.f32 %f1525, %f1519, %f1524; neg.f32 %f1526, %f1525; add.rn.f32 %f1527, %f1519, %f1526; add.rn.f32 %f1528, %f1527, %f1524; mov.b32 %r1480, %f1525; setp.eq.s32 %p172, %r1480, 1118925336; add.s32 %r1481, %r1480, -1; mov.b32 %f1529, %r1481; add.f32 %f1530, %f1528, 0f37000000; selp.f32 %f214, %f1530, %f1528, %p172; selp.f32 %f1531, %f1529, %f1525, %p172; mov.f32 %f1532, 0f3FB8AA3B; mul.rn.f32 %f1533, %f1531, %f1532; cvt.rzi.f32.f32 %f1534, %f1533; abs.f32 %f1535, %f1534; setp.gt.f32 %p173, %f1535, 0f42FC0000; mov.b32 %r1482, %f1534; and.b32 %r1483, %r1482, -2147483648; or.b32 %r1484, %r1483, 1123811328; mov.b32 %f1536, %r1484; selp.f32 %f1537, %f1536, %f1534, %p173; mov.f32 %f1538, 0fBF317218; fma.rn.f32 %f1539, %f1537, %f1538, %f1531; mov.f32 %f1540, 0f3102E308; fma.rn.f32 %f1541, %f1537, %f1540, %f1539; mul.f32 %f1542, %f1541, 0f3FB8AA3B; add.f32 %f1543, %f1537, 0f4B40007F; mov.b32 %r1485, %f1543; shl.b32 %r1486, %r1485, 23; mov.b32 %f1544, %r1486; ex2.approx.ftz.f32 %f1545, %f1542; mul.f32 %f215, %f1545, %f1544; setp.eq.f32 %p174, %f215, 0f7F800000; mov.f32 %f1671, 0f7F800000; @%p174 bra $L__BB0_133; fma.rn.f32 %f1671, %f215, %f214, %f215; $L__BB0_133: setp.lt.f32 %p175, %f189, 0f00000000; and.pred %p7, %p175, %p148; setp.eq.f32 %p177, %f189, 0f00000000; @%p177 bra $L__BB0_137; bra.uni $L__BB0_134; $L__BB0_137: add.f32 %f1550, %f189, %f189; selp.f32 %f1673, %f1550, 0f00000000, %p148; bra.uni $L__BB0_138; $L__BB0_134: mov.b32 %r1487, %f1671; xor.b32 %r1488, %r1487, -2147483648; mov.b32 %f1546, %r1488; selp.f32 %f1673, %f1546, %f1671, %p7; setp.geu.f32 %p178, %f189, 0f00000000; @%p178 bra $L__BB0_138; mov.f32 %f1547, 0f3ED55555; cvt.rzi.f32.f32 %f1548, %f1547; setp.eq.f32 %p179, %f1548, 0f3ED55555; @%p179 bra $L__BB0_138; mov.f32 %f1673, 0f7FFFFFFF; $L__BB0_138: add.f32 %f1551, %f213, 0f3ED55555; mov.b32 %r1489, %f1551; setp.lt.s32 %p181, %r1489, 2139095040; @%p181 bra $L__BB0_143; setp.gtu.f32 %p182, %f213, 0f7F800000; @%p182 bra $L__BB0_142; bra.uni $L__BB0_140; $L__BB0_142: add.f32 %f1673, %f189, 0f3ED55555; bra.uni $L__BB0_143; $L__BB0_140: setp.neu.f32 %p183, %f213, 0f7F800000; @%p183 bra $L__BB0_143; selp.f32 %f1673, 0fFF800000, 0f7F800000, %p7; $L__BB0_143: fma.rn.f32 %f1552, %f1667, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p184, %f187, 0f3F800000; mov.f32 %f1553, 0f3F800000; selp.f32 %f1554, 0f3F7FFFFF, %f1552, %p184; mul.f32 %f1555, %f187, 0f414EB852; setp.lt.f32 %p185, %f187, 0f3B4D2E1C; selp.f32 %f1556, %f1555, %f1554, %p185; fma.rn.f32 %f1557, %f1670, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p186, %f188, 0f3F800000; selp.f32 %f1558, 0f3F7FFFFF, %f1557, %p186; mul.f32 %f1559, %f188, 0f414EB852; setp.lt.f32 %p187, %f188, 0f3B4D2E1C; selp.f32 %f1560, %f1559, %f1558, %p187; fma.rn.f32 %f1561, %f1673, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p188, %f189, 0f3F800000; selp.f32 %f1562, 0f3F7FFFFF, %f1561, %p188; mul.f32 %f1563, %f189, 0f414EB852; setp.lt.f32 %p189, %f189, 0f3B4D2E1C; selp.f32 %f1564, %f1563, %f1562, %p189; min.f32 %f1565, %f1556, %f1553; mov.f32 %f1566, 0f00000000; max.f32 %f1567, %f1566, %f1565; mul.f32 %f1568, %f1567, 0f43800000; cvt.rzi.u32.f32 %r1490, %f1568; min.u32 %r1491, %r1490, 255; min.f32 %f1569, %f1560, %f1553; max.f32 %f1570, %f1566, %f1569; mul.f32 %f1571, %f1570, 0f43800000; cvt.rzi.u32.f32 %r1492, %f1571; min.u32 %r1493, %r1492, 255; min.f32 %f1572, %f1564, %f1553; max.f32 %f1573, %f1566, %f1572; mul.f32 %f1574, %f1573, 0f43800000; cvt.rzi.u32.f32 %r1494, %f1574; min.u32 %r1495, %r1494, 255; shl.b64 %rd83, %rd19, 2; add.s64 %rd84, %rd18, %rd83; cvt.u16.u32 %rs36, %r1495; cvt.u16.u32 %rs37, %r1493; cvt.u16.u32 %rs38, %r1491; mov.u16 %rs39, 255; st.global.v4.u8 [%rd84], {%rs38, %rs37, %rs36, %rs39}; $L__BB0_144: and.b32 %r1496, %r60, 4; setp.eq.s32 %p190, %r1496, 0; @%p190 bra $L__BB0_148; ld.const.u32 %r1497, [params+108]; setp.eq.s32 %p191, %r1497, 0; ld.const.u64 %rd85, [params+224]; cvta.to.global.u64 %rd86, %rd85; ld.const.u32 %r1498, [params+216]; mad.lo.s32 %r1499, %r1498, %r6, %r5; mul.wide.u32 %rd87, %r1499, 8; add.s64 %rd20, %rd86, %rd87; @%p191 bra $L__BB0_147; ld.global.v4.u16 {%rs47, %rs48, %rs49, %rs50}, [%rd20]; // begin inline asm { cvt.f32.f16 %f1575, %rs47;} // end inline asm // begin inline asm { cvt.f32.f16 %f1576, %rs48;} // end inline asm // begin inline asm { cvt.f32.f16 %f1577, %rs49;} // end inline asm add.f32 %f1578, %f136, %f1575; add.f32 %f1579, %f137, %f1576; add.f32 %f1580, %f138, %f1577; mov.f32 %f1581, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs45, %f1580;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs44, %f1579;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs43, %f1578;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs46, %f1581;} // end inline asm st.global.v4.u16 [%rd20], {%rs43, %rs44, %rs45, %rs46}; bra.uni $L__BB0_148; $L__BB0_147: mov.f32 %f1585, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs54, %f1585;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs53, %f138;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs52, %f137;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs51, %f136;} // end inline asm st.global.v4.u16 [%rd20], {%rs51, %rs52, %rs53, %rs54}; $L__BB0_148: and.b32 %r1500, %r60, 64; setp.eq.s32 %p192, %r1500, 0; @%p192 bra $L__BB0_170; ld.const.u64 %rd88, [params+208]; cvta.to.global.u64 %rd89, %rd88; ld.const.u32 %r1501, [params+200]; mad.lo.s32 %r1502, %r1501, %r6, %r5; mul.f32 %f1586, %f241, 0f3F000000; mov.f32 %f1587, 0f3F000000; sub.f32 %f1588, %f1587, %f1586; mul.f32 %f1589, %f1588, 0f437F0000; cvt.rzi.u32.f32 %r1503, %f1589; mul.f32 %f1590, %f242, 0f3F000000; sub.f32 %f1591, %f1587, %f1590; mul.f32 %f1592, %f1591, 0f437F0000; cvt.rzi.u32.f32 %r1504, %f1592; mul.f32 %f1593, %f243, 0f3F000000; sub.f32 %f1594, %f1587, %f1593; mul.f32 %f1595, %f1594, 0f437F0000; cvt.rzi.u32.f32 %r1505, %f1595; mul.wide.u32 %rd90, %r1502, 4; add.s64 %rd91, %rd89, %rd90; cvt.u16.u32 %rs55, %r1505; cvt.u16.u32 %rs56, %r1504; cvt.u16.u32 %rs57, %r1503; mov.u16 %rs58, 255; st.global.v4.u8 [%rd91], {%rs57, %rs56, %rs55, %rs58}; bra.uni $L__BB0_170; $L__BB0_157: mov.f32 %f1606, 0f00000000; mov.f32 %f1607, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs77, %f1607;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs76, %f1606;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs75, %f1606;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs74, %f1606;} // end inline asm st.global.v4.u16 [%rd21], {%rs74, %rs75, %rs76, %rs77}; $L__BB0_158: and.b32 %r1517, %r8, 64; setp.eq.s32 %p200, %r1517, 0; @%p200 bra $L__BB0_170; ld.const.u64 %rd103, [params+208]; cvta.to.global.u64 %rd104, %rd103; ld.const.u32 %r1518, [params+200]; mad.lo.s32 %r1519, %r1518, %r6, %r5; mul.wide.u32 %rd105, %r1519, 4; add.s64 %rd106, %rd104, %rd105; mov.u16 %rs78, 0; mov.u16 %rs79, 255; st.global.v4.u8 [%rd106], {%rs78, %rs78, %rs78, %rs79}; $L__BB0_170: ret; }