// // 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]; .visible .entry __raygen__oxMain() { .reg .pred %p<189>; .reg .b16 %rs<183>; .reg .f32 %f<1697>; .reg .b32 %r<428>; .reg .b64 %rd<74>; // begin inline asm call (%r26), _optix_get_launch_index_x, (); // end inline asm // begin inline asm call (%r27), _optix_get_launch_index_y, (); // end inline asm ld.const.u64 %rd20, [params+400]; cvta.to.global.u64 %rd21, %rd20; ld.const.u32 %r32, [params+392]; mad.lo.s32 %r33, %r32, %r27, %r26; mul.wide.u32 %rd22, %r33, 4; add.s64 %rd1, %rd21, %rd22; ld.global.v2.u8 {%rs18, %rs177}, [%rd1]; or.b16 %rs20, %rs18, %rs177; and.b16 %rs21, %rs20, 255; setp.eq.s16 %p10, %rs21, 0; @%p10 bra $L__BB0_2; ld.global.u8 %rs176, [%rd1+2]; bra.uni $L__BB0_3; $L__BB0_2: ld.global.u8 %rs176, [%rd1+2]; setp.eq.s16 %p11, %rs176, 0; mov.f32 %f1596, 0f00000000; mov.u16 %rs177, 0; mov.f32 %f1597, %f1596; mov.f32 %f1598, %f1596; @%p11 bra $L__BB0_4; $L__BB0_3: cvt.rn.f32.u16 %f301, %rs18; div.rn.f32 %f302, %f301, 0f437F0000; fma.rn.f32 %f303, %f302, 0f40000000, 0fBF800000; and.b16 %rs24, %rs177, 255; cvt.rn.f32.u16 %f304, %rs24; div.rn.f32 %f305, %f304, 0f437F0000; fma.rn.f32 %f306, %f305, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f307, %rs176; div.rn.f32 %f308, %f307, 0f437F0000; fma.rn.f32 %f309, %f308, 0f40000000, 0fBF800000; mul.f32 %f310, %f306, %f306; fma.rn.f32 %f311, %f303, %f303, %f310; fma.rn.f32 %f312, %f309, %f309, %f311; sqrt.rn.f32 %f313, %f312; rcp.rn.f32 %f314, %f313; mul.f32 %f1598, %f314, %f309; mul.f32 %f1597, %f314, %f306; mul.f32 %f1596, %f303, %f314; $L__BB0_4: ld.const.v2.u32 {%r34, %r35}, [params]; add.s32 %r3, %r34, %r26; add.s32 %r4, %r35, %r27; setp.eq.f32 %p12, %f1596, 0f00000000; setp.eq.f32 %p13, %f1597, 0f00000000; and.pred %p14, %p12, %p13; setp.eq.f32 %p15, %f1598, 0f00000000; and.pred %p16, %p15, %p14; @%p16 bra $L__BB0_154; bra.uni $L__BB0_5; $L__BB0_154: ld.const.u32 %r23, [params+104]; and.b32 %r404, %r23, 1; setp.eq.b32 %p178, %r404, 1; mov.pred %p179, 0; xor.pred %p180, %p178, %p179; not.pred %p181, %p180; @%p181 bra $L__BB0_156; ld.const.u64 %rd53, [params+144]; cvta.to.global.u64 %rd54, %rd53; ld.const.u32 %r405, [params+136]; mad.lo.s32 %r406, %r405, %r4, %r3; mul.wide.u32 %rd55, %r406, 4; add.s64 %rd56, %rd54, %rd55; mov.u16 %rs113, 0; st.global.v4.u8 [%rd56], {%rs113, %rs113, %rs113, %rs113}; $L__BB0_156: and.b32 %r407, %r23, 8; setp.eq.s32 %p182, %r407, 0; @%p182 bra $L__BB0_158; ld.const.u64 %rd57, [params+192]; cvta.to.global.u64 %rd58, %rd57; ld.const.u32 %r408, [params+184]; mad.lo.s32 %r409, %r408, %r4, %r3; mov.f32 %f1556, 0f00000000; cvt.rzi.u32.f32 %r410, %f1556; mul.wide.u32 %rd59, %r409, 2; add.s64 %rd60, %rd58, %rd59; mov.u16 %rs114, 0; cvt.u16.u32 %rs115, %r410; st.global.v2.u8 [%rd60], {%rs115, %rs114}; $L__BB0_158: and.b32 %r411, %r23, 4; setp.eq.s32 %p183, %r411, 0; ld.const.u32 %r427, [params+108]; @%p183 bra $L__BB0_162; setp.eq.s32 %p184, %r427, 0; ld.const.u64 %rd61, [params+224]; cvta.to.global.u64 %rd62, %rd61; ld.const.u32 %r412, [params+216]; mad.lo.s32 %r413, %r412, %r4, %r3; mul.wide.u32 %rd63, %r413, 8; add.s64 %rd16, %rd62, %rd63; @%p184 bra $L__BB0_161; ld.global.v4.u16 {%rs122, %rs123, %rs124, %rs125}, [%rd16]; // begin inline asm { cvt.f32.f16 %f1557, %rs122;} // end inline asm // begin inline asm { cvt.f32.f16 %f1558, %rs123;} // end inline asm // begin inline asm { cvt.f32.f16 %f1559, %rs124;} // end inline asm add.f32 %f1560, %f1557, 0f00000000; add.f32 %f1561, %f1558, 0f00000000; add.f32 %f1562, %f1559, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs121, %f1562;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs120, %f1561;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs119, %f1560;} // end inline asm mov.u16 %rs126, 0; st.global.v4.u16 [%rd16], {%rs119, %rs120, %rs121, %rs126}; bra.uni $L__BB0_162; $L__BB0_5: ld.const.u64 %rd23, [params+432]; cvta.to.global.u64 %rd24, %rd23; ld.const.u32 %r38, [params+424]; mad.lo.s32 %r39, %r38, %r27, %r26; mul.wide.u32 %rd25, %r39, 12; add.s64 %rd26, %rd24, %rd25; ld.global.f32 %f10, [%rd26]; mul.f32 %f315, %f10, 0f3456BF95; ld.global.f32 %f11, [%rd26+4]; mul.f32 %f316, %f11, 0f3456BF95; ld.global.f32 %f12, [%rd26+8]; mul.f32 %f317, %f12, 0f3456BF95; abs.f32 %f318, %f1596; div.rn.f32 %f319, %f315, %f318; abs.f32 %f320, %f1597; div.rn.f32 %f321, %f316, %f320; abs.f32 %f322, %f1598; div.rn.f32 %f323, %f317, %f322; abs.f32 %f324, %f319; abs.f32 %f325, %f321; abs.f32 %f326, %f323; mov.f32 %f327, 0f38D1B717; max.f32 %f328, %f324, %f327; max.f32 %f329, %f325, %f327; max.f32 %f330, %f326, %f327; fma.rn.f32 %f13, %f1596, %f328, %f10; fma.rn.f32 %f14, %f1597, %f329, %f11; fma.rn.f32 %f15, %f1598, %f330, %f12; cvt.rn.f32.u32 %f16, %r26; cvt.rn.f32.u32 %f17, %r27; ld.const.u64 %rd27, [params+464]; cvta.to.global.u64 %rd28, %rd27; ld.const.u32 %r40, [params+456]; mad.lo.s32 %r41, %r40, %r27, %r26; mul.wide.u32 %rd29, %r41, 4; add.s64 %rd2, %rd28, %rd29; ld.global.v2.u8 {%rs26, %rs180}, [%rd2]; or.b16 %rs28, %rs26, %rs180; and.b16 %rs29, %rs28, 255; setp.eq.s16 %p17, %rs29, 0; @%p17 bra $L__BB0_7; ld.global.u8 %rs178, [%rd2+2]; bra.uni $L__BB0_8; $L__BB0_7: ld.global.u8 %rs178, [%rd2+2]; setp.eq.s16 %p18, %rs178, 0; mov.f32 %f1599, 0f00000000; mov.u16 %rs180, 0; mov.u16 %rs181, %rs180; mov.f32 %f1600, %f1599; mov.f32 %f1601, %f1599; @%p18 bra $L__BB0_9; $L__BB0_8: mov.u16 %rs181, %rs180; cvt.rn.f32.u16 %f334, %rs26; div.rn.f32 %f335, %f334, 0f437F0000; fma.rn.f32 %f336, %f335, 0f40000000, 0fBF800000; and.b16 %rs34, %rs181, 255; cvt.rn.f32.u16 %f337, %rs34; div.rn.f32 %f338, %f337, 0f437F0000; fma.rn.f32 %f339, %f338, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f340, %rs178; div.rn.f32 %f341, %f340, 0f437F0000; fma.rn.f32 %f342, %f341, 0f40000000, 0fBF800000; mul.f32 %f343, %f339, %f339; fma.rn.f32 %f344, %f336, %f336, %f343; fma.rn.f32 %f345, %f342, %f342, %f344; sqrt.rn.f32 %f346, %f345; rcp.rn.f32 %f347, %f346; mul.f32 %f1601, %f347, %f342; mul.f32 %f1600, %f347, %f339; mul.f32 %f1599, %f336, %f347; mov.u16 %rs180, %rs178; $L__BB0_9: mul.f32 %f351, %f1597, %f1601; mul.f32 %f352, %f1598, %f1600; sub.f32 %f353, %f351, %f352; mul.f32 %f354, %f1598, %f1599; mul.f32 %f355, %f1596, %f1601; sub.f32 %f356, %f354, %f355; mul.f32 %f357, %f1596, %f1600; mul.f32 %f358, %f1597, %f1599; sub.f32 %f359, %f357, %f358; ld.global.u8 %rs36, [%rd2+3]; setp.eq.s16 %p19, %rs36, 0; selp.f32 %f24, 0fBF800000, 0f3F800000, %p19; mul.f32 %f360, %f353, %f24; mul.f32 %f361, %f356, %f24; mul.f32 %f362, %f359, %f24; mul.f32 %f363, %f360, 0f00000000; mul.f32 %f364, %f361, 0f00000000; mul.f32 %f365, %f362, 0f00000000; fma.rn.f32 %f366, %f1599, 0f3F5105EC, %f363; fma.rn.f32 %f367, %f1600, 0f3F5105EC, %f364; fma.rn.f32 %f368, %f1601, 0f3F5105EC, %f365; mul.f32 %f25, %f1596, 0f3F13CD3A; add.f32 %f26, %f25, %f366; mul.f32 %f27, %f1597, 0f3F13CD3A; add.f32 %f28, %f27, %f367; mul.f32 %f29, %f1598, 0f3F13CD3A; add.f32 %f30, %f29, %f368; or.b16 %rs37, %rs26, %rs181; or.b16 %rs15, %rs37, %rs180; and.b16 %rs38, %rs15, 255; setp.eq.s16 %p20, %rs38, 0; mov.f32 %f1605, 0f00000000; mov.f32 %f1602, %f1605; mov.f32 %f1603, %f1605; mov.f32 %f1604, %f1605; @%p20 bra $L__BB0_11; cvt.rn.f32.u16 %f369, %rs26; div.rn.f32 %f370, %f369, 0f437F0000; fma.rn.f32 %f371, %f370, 0f40000000, 0fBF800000; and.b16 %rs40, %rs181, 255; cvt.rn.f32.u16 %f372, %rs40; div.rn.f32 %f373, %f372, 0f437F0000; fma.rn.f32 %f374, %f373, 0f40000000, 0fBF800000; and.b16 %rs41, %rs180, 255; cvt.rn.f32.u16 %f375, %rs41; div.rn.f32 %f376, %f375, 0f437F0000; fma.rn.f32 %f377, %f376, 0f40000000, 0fBF800000; mul.f32 %f378, %f374, %f374; fma.rn.f32 %f379, %f371, %f371, %f378; fma.rn.f32 %f380, %f377, %f377, %f379; sqrt.rn.f32 %f381, %f380; rcp.rn.f32 %f382, %f381; mul.f32 %f1604, %f382, %f377; mul.f32 %f1603, %f382, %f374; mul.f32 %f1602, %f371, %f382; $L__BB0_11: mul.f32 %f386, %f1597, %f1604; mul.f32 %f387, %f1598, %f1603; sub.f32 %f388, %f386, %f387; mul.f32 %f389, %f1598, %f1602; mul.f32 %f390, %f1596, %f1604; sub.f32 %f391, %f389, %f390; mul.f32 %f392, %f1596, %f1603; mul.f32 %f393, %f1597, %f1602; sub.f32 %f394, %f392, %f393; mul.f32 %f395, %f388, %f24; mul.f32 %f396, %f391, %f24; mul.f32 %f397, %f394, %f24; mul.f32 %f398, %f1602, 0f3ED105EC; mul.f32 %f399, %f1603, 0f3ED105EC; mul.f32 %f400, %f1604, 0f3ED105EC; mul.f32 %f401, %f395, 0f3F3504F3; mul.f32 %f402, %f396, 0f3F3504F3; mul.f32 %f403, %f397, 0f3F3504F3; sub.f32 %f404, %f401, %f398; sub.f32 %f405, %f402, %f399; sub.f32 %f406, %f403, %f400; add.f32 %f37, %f25, %f404; add.f32 %f38, %f27, %f405; add.f32 %f39, %f29, %f406; mov.f32 %f1606, %f1605; mov.f32 %f1607, %f1605; @%p20 bra $L__BB0_13; cvt.rn.f32.u16 %f407, %rs26; div.rn.f32 %f408, %f407, 0f437F0000; fma.rn.f32 %f409, %f408, 0f40000000, 0fBF800000; and.b16 %rs44, %rs181, 255; cvt.rn.f32.u16 %f410, %rs44; div.rn.f32 %f411, %f410, 0f437F0000; fma.rn.f32 %f412, %f411, 0f40000000, 0fBF800000; and.b16 %rs45, %rs180, 255; cvt.rn.f32.u16 %f413, %rs45; div.rn.f32 %f414, %f413, 0f437F0000; fma.rn.f32 %f415, %f414, 0f40000000, 0fBF800000; mul.f32 %f416, %f412, %f412; fma.rn.f32 %f417, %f409, %f409, %f416; fma.rn.f32 %f418, %f415, %f415, %f417; sqrt.rn.f32 %f419, %f418; rcp.rn.f32 %f420, %f419; mul.f32 %f1607, %f420, %f415; mul.f32 %f1606, %f420, %f412; mul.f32 %f1605, %f409, %f420; $L__BB0_13: mul.f32 %f421, %f1597, %f1607; mul.f32 %f422, %f1598, %f1606; sub.f32 %f423, %f421, %f422; mul.f32 %f424, %f1598, %f1605; mul.f32 %f425, %f1596, %f1607; sub.f32 %f426, %f424, %f425; mul.f32 %f427, %f1596, %f1606; mul.f32 %f428, %f1597, %f1605; sub.f32 %f429, %f427, %f428; mul.f32 %f430, %f423, %f24; mul.f32 %f431, %f426, %f24; mul.f32 %f432, %f429, %f24; mul.f32 %f433, %f1605, 0f3ED105EC; mul.f32 %f434, %f1606, 0f3ED105EC; mul.f32 %f435, %f1607, 0f3ED105EC; mul.f32 %f436, %f430, 0fBF3504F3; mul.f32 %f437, %f431, 0fBF3504F3; mul.f32 %f438, %f432, 0fBF3504F3; sub.f32 %f439, %f436, %f433; sub.f32 %f440, %f437, %f434; sub.f32 %f441, %f438, %f435; add.f32 %f46, %f25, %f439; add.f32 %f47, %f27, %f440; add.f32 %f48, %f29, %f441; ld.const.u32 %r5, [params+600]; setp.eq.s32 %p22, %r5, 0; @%p22 bra $L__BB0_52; ld.const.u64 %rd30, [params+608]; cvta.to.global.u64 %rd3, %rd30; mov.f32 %f455, 0f40000000; cvt.rzi.f32.f32 %f456, %f455; add.f32 %f457, %f456, %f456; mov.f32 %f458, 0f40800000; sub.f32 %f459, %f458, %f457; abs.f32 %f49, %f459; ld.const.u32 %r6, [params+544]; ld.const.u32 %r425, [params+104]; mul.f32 %f50, %f13, 0f3456BF95; mul.f32 %f51, %f14, 0f3456BF95; mul.f32 %f52, %f15, 0f3456BF95; mov.f32 %f454, 0f00000000; mov.u32 %r421, 0; ld.const.u64 %rd4, [params+96]; abs.f32 %f692, %f50; abs.f32 %f693, %f51; max.f32 %f694, %f692, %f693; abs.f32 %f695, %f52; max.f32 %f696, %f694, %f695; mov.f32 %f1662, %f454; mov.f32 %f1661, %f454; mov.f32 %f1660, %f454; mov.f32 %f1659, %f454; mov.f32 %f1658, %f454; mov.f32 %f1657, %f454; mov.f32 %f1656, %f454; mov.f32 %f1655, %f454; mov.f32 %f1654, %f454; mov.f32 %f1653, %f454; mov.f32 %f1652, %f454; mov.f32 %f1651, %f454; mov.f32 %f1650, %f454; $L__BB0_15: mul.wide.u32 %rd31, %r421, 112; add.s64 %rd6, %rd3, %rd31; ld.global.v4.f32 {%f461, %f462, %f463, %f464}, [%rd6]; ld.global.v4.f32 {%f465, %f466, %f467, %f468}, [%rd6+16]; ld.global.v4.f32 {%f471, %f1629, %f1630, %f474}, [%rd6+32]; ld.global.v4.f32 {%f475, %f476, %f477, %f1627}, [%rd6+48]; ld.global.v4.f32 {%f479, %f480, %f481, %f482}, [%rd6+64]; ld.global.v4.u32 {%r43, %r44, %r45, %r46}, [%rd6+80]; mov.b32 %f85, %r45; mov.b32 %f84, %r44; mov.b32 %f83, %r43; ld.global.u64 %rd7, [%rd6+96]; sub.f32 %f483, %f462, %f10; sub.f32 %f484, %f463, %f11; sub.f32 %f485, %f464, %f12; mul.f32 %f486, %f484, %f484; fma.rn.f32 %f487, %f483, %f483, %f486; fma.rn.f32 %f488, %f485, %f485, %f487; sqrt.rn.f32 %f489, %f488; rcp.rn.f32 %f490, %f489; mul.f32 %f87, %f483, %f490; mul.f32 %f88, %f484, %f490; mul.f32 %f89, %f485, %f490; mul.f32 %f90, %f489, %f466; mul.f32 %f91, %f489, %f468; abs.f32 %f92, %f91; setp.lt.f32 %p23, %f92, 0f00800000; mul.f32 %f491, %f92, 0f4B800000; selp.f32 %f492, %f491, %f92, %p23; selp.f32 %f493, 0fC3170000, 0fC2FE0000, %p23; mov.b32 %r50, %f492; and.b32 %r51, %r50, 8388607; or.b32 %r52, %r51, 1065353216; mov.b32 %f494, %r52; shr.u32 %r53, %r50, 23; cvt.rn.f32.u32 %f495, %r53; add.f32 %f496, %f493, %f495; setp.gt.f32 %p24, %f494, 0f3FB504F3; mul.f32 %f497, %f494, 0f3F000000; add.f32 %f498, %f496, 0f3F800000; selp.f32 %f499, %f498, %f496, %p24; selp.f32 %f500, %f497, %f494, %p24; add.f32 %f501, %f500, 0fBF800000; add.f32 %f502, %f500, 0f3F800000; rcp.approx.ftz.f32 %f503, %f502; add.f32 %f504, %f501, %f501; mul.f32 %f505, %f504, %f503; mul.f32 %f506, %f505, %f505; mov.f32 %f507, 0f3C4CAF63; mov.f32 %f508, 0f3B18F0FE; fma.rn.f32 %f509, %f508, %f506, %f507; mov.f32 %f510, 0f3DAAAABD; fma.rn.f32 %f511, %f509, %f506, %f510; mul.rn.f32 %f512, %f511, %f506; mul.rn.f32 %f513, %f512, %f505; sub.f32 %f514, %f501, %f505; add.f32 %f515, %f514, %f514; neg.f32 %f516, %f505; fma.rn.f32 %f517, %f516, %f501, %f515; mul.rn.f32 %f518, %f503, %f517; add.f32 %f519, %f513, %f505; sub.f32 %f520, %f505, %f519; add.f32 %f521, %f513, %f520; add.f32 %f522, %f518, %f521; add.f32 %f523, %f519, %f522; sub.f32 %f524, %f519, %f523; add.f32 %f525, %f522, %f524; mov.f32 %f526, 0f3F317200; mul.rn.f32 %f527, %f499, %f526; mov.f32 %f528, 0f35BFBE8E; mul.rn.f32 %f529, %f499, %f528; add.f32 %f530, %f527, %f523; sub.f32 %f531, %f527, %f530; add.f32 %f532, %f523, %f531; add.f32 %f533, %f525, %f532; add.f32 %f534, %f529, %f533; add.f32 %f535, %f530, %f534; sub.f32 %f536, %f530, %f535; add.f32 %f537, %f534, %f536; mul.rn.f32 %f539, %f458, %f535; neg.f32 %f540, %f539; fma.rn.f32 %f541, %f458, %f535, %f540; fma.rn.f32 %f542, %f458, %f537, %f541; fma.rn.f32 %f544, %f454, %f535, %f542; add.rn.f32 %f545, %f539, %f544; neg.f32 %f546, %f545; add.rn.f32 %f547, %f539, %f546; add.rn.f32 %f548, %f547, %f544; mov.b32 %r54, %f545; setp.eq.s32 %p25, %r54, 1118925336; add.s32 %r55, %r54, -1; mov.b32 %f549, %r55; add.f32 %f550, %f548, 0f37000000; selp.f32 %f93, %f550, %f548, %p25; selp.f32 %f551, %f549, %f545, %p25; mov.f32 %f552, 0f3FB8AA3B; mul.rn.f32 %f553, %f551, %f552; cvt.rzi.f32.f32 %f554, %f553; abs.f32 %f555, %f554; setp.gt.f32 %p26, %f555, 0f42FC0000; mov.b32 %r56, %f554; and.b32 %r57, %r56, -2147483648; or.b32 %r58, %r57, 1123811328; mov.b32 %f556, %r58; selp.f32 %f557, %f556, %f554, %p26; mov.f32 %f558, 0fBF317218; fma.rn.f32 %f559, %f557, %f558, %f551; mov.f32 %f560, 0f3102E308; fma.rn.f32 %f561, %f557, %f560, %f559; mul.f32 %f562, %f561, 0f3FB8AA3B; add.f32 %f563, %f557, 0f4B40007F; mov.b32 %r59, %f563; shl.b32 %r60, %r59, 23; mov.b32 %f564, %r60; ex2.approx.ftz.f32 %f565, %f562; mul.f32 %f94, %f565, %f564; setp.eq.f32 %p27, %f94, 0f7F800000; mov.f32 %f1621, 0f7F800000; @%p27 bra $L__BB0_17; fma.rn.f32 %f1621, %f94, %f93, %f94; $L__BB0_17: setp.lt.f32 %p28, %f91, 0f00000000; setp.eq.f32 %p29, %f49, 0f3F800000; and.pred %p1, %p28, %p29; setp.eq.f32 %p30, %f91, 0f00000000; @%p30 bra $L__BB0_21; bra.uni $L__BB0_18; $L__BB0_21: add.f32 %f570, %f91, %f91; selp.f32 %f1623, %f570, 0f00000000, %p29; bra.uni $L__BB0_22; $L__BB0_18: mov.b32 %r61, %f1621; xor.b32 %r62, %r61, -2147483648; mov.b32 %f566, %r62; selp.f32 %f1623, %f566, %f1621, %p1; setp.geu.f32 %p31, %f91, 0f00000000; @%p31 bra $L__BB0_22; mov.f32 %f567, 0f40800000; cvt.rzi.f32.f32 %f568, %f567; setp.eq.f32 %p32, %f568, 0f40800000; @%p32 bra $L__BB0_22; mov.f32 %f1623, 0f7FFFFFFF; $L__BB0_22: abs.f32 %f1593, %f91; add.f32 %f571, %f1593, 0f40800000; mov.b32 %r63, %f571; setp.lt.s32 %p34, %r63, 2139095040; @%p34 bra $L__BB0_27; abs.f32 %f1594, %f91; setp.gtu.f32 %p35, %f1594, 0f7F800000; @%p35 bra $L__BB0_26; bra.uni $L__BB0_24; $L__BB0_26: add.f32 %f1623, %f91, 0f40800000; bra.uni $L__BB0_27; $L__BB0_24: abs.f32 %f1595, %f91; setp.neu.f32 %p36, %f1595, 0f7F800000; @%p36 bra $L__BB0_27; selp.f32 %f1623, 0fFF800000, 0f7F800000, %p1; $L__BB0_27: mov.f32 %f572, 0f3F800000; sub.f32 %f573, %f572, %f1623; setp.eq.f32 %p37, %f91, 0f3F800000; selp.f32 %f574, 0f00000000, %f573, %p37; cvt.sat.f32.f32 %f575, %f574; fma.rn.f32 %f576, %f90, %f90, %f467; div.rn.f32 %f103, %f575, %f576; mul.f32 %f577, %f1597, %f88; fma.rn.f32 %f578, %f1596, %f87, %f577; fma.rn.f32 %f104, %f1598, %f89, %f578; setp.eq.f32 %p38, %f461, 0f3F800000; @%p38 bra $L__BB0_33; bra.uni $L__BB0_28; $L__BB0_33: setp.eq.s64 %p42, %rd7, 0; mov.f32 %f1628, %f471; mov.f32 %f1631, %f103; @%p42 bra $L__BB0_37; setp.geu.f32 %p43, %f1627, 0f00000000; mov.f32 %f1624, %f87; mov.f32 %f1625, %f88; mov.f32 %f1626, %f89; @%p43 bra $L__BB0_36; mul.f32 %f603, %f84, %f484; fma.rn.f32 %f604, %f83, %f483, %f603; fma.rn.f32 %f606, %f85, %f485, %f604; rcp.rn.f32 %f607, %f606; mul.f32 %f1624, %f483, %f607; mul.f32 %f1625, %f484, %f607; mul.f32 %f1626, %f485, %f607; neg.f32 %f1627, %f1627; $L__BB0_36: mul.f32 %f608, %f85, %f89; mul.f32 %f609, %f84, %f88; neg.f32 %f610, %f609; mul.f32 %f611, %f83, %f87; sub.f32 %f612, %f610, %f611; sub.f32 %f613, %f612, %f608; setp.gt.f32 %p44, %f613, 0f00000000; selp.f32 %f614, 0f3F800000, 0f00000000, %p44; mov.f32 %f615, 0f3F800000; mul.f32 %f616, %f476, %f1625; mul.f32 %f617, %f480, %f1625; fma.rn.f32 %f618, %f475, %f1624, %f616; fma.rn.f32 %f619, %f479, %f1624, %f617; fma.rn.f32 %f620, %f477, %f1626, %f618; fma.rn.f32 %f621, %f481, %f1626, %f619; fma.rn.f32 %f622, %f1627, %f620, 0f3F000000; sub.f32 %f623, %f615, %f622; fma.rn.f32 %f624, %f1627, %f621, 0f3F000000; tex.2d.v4.f32.f32 {%f625, %f626, %f627, %f628}, [%rd7, {%f623, %f624}]; mul.f32 %f629, %f614, %f625; mul.f32 %f630, %f614, %f626; mul.f32 %f631, %f614, %f627; mul.f32 %f1628, %f471, %f629; mul.f32 %f1629, %f1629, %f630; mul.f32 %f1630, %f1630, %f631; mov.f32 %f1631, %f103; bra.uni $L__BB0_37; $L__BB0_28: setp.eq.f32 %p39, %f461, 0f40000000; @%p39 bra $L__BB0_31; bra.uni $L__BB0_29; $L__BB0_31: setp.eq.s64 %p41, %rd7, 0; mov.f32 %f1628, %f471; mov.f32 %f1631, %f103; @%p41 bra $L__BB0_37; mul.f32 %f588, %f476, %f88; mul.f32 %f589, %f480, %f88; mul.f32 %f590, %f84, %f88; fma.rn.f32 %f591, %f475, %f87, %f588; fma.rn.f32 %f592, %f479, %f87, %f589; fma.rn.f32 %f593, %f83, %f87, %f590; fma.rn.f32 %f594, %f477, %f89, %f591; fma.rn.f32 %f595, %f481, %f89, %f592; fma.rn.f32 %f596, %f85, %f89, %f593; tex.cube.v4.f32.f32 {%f597, %f598, %f599, %f600}, [%rd7, {%f594, %f595, %f596, %f596}]; mul.f32 %f1628, %f471, %f597; mul.f32 %f1629, %f1629, %f598; mul.f32 %f1630, %f1630, %f599; mov.f32 %f1631, %f103; bra.uni $L__BB0_37; $L__BB0_29: setp.neu.f32 %p40, %f461, 0f40800000; mov.f32 %f1628, %f471; mov.f32 %f1631, %f103; @%p40 bra $L__BB0_37; mul.f32 %f579, %f83, %f87; mul.f32 %f580, %f84, %f88; neg.f32 %f581, %f580; sub.f32 %f582, %f581, %f579; mul.f32 %f583, %f85, %f89; sub.f32 %f584, %f582, %f583; fma.rn.f32 %f585, %f474, %f584, %f1627; cvt.sat.f32.f32 %f586, %f585; mul.f32 %f587, %f586, %f586; mul.f32 %f1631, %f103, %f587; mov.f32 %f1628, %f471; $L__BB0_37: max.f32 %f644, %f1628, %f1629; max.f32 %f645, %f644, %f1630; setp.eq.s32 %p46, %r6, 0; selp.f32 %f646, %f104, 0f3F800000, %p46; cvt.sat.f32.f32 %f647, %f646; mul.f32 %f137, %f647, %f1631; mul.f32 %f648, %f137, %f645; setp.lt.f32 %p47, %f648, 0f3727C5AC; mov.pred %p188, -1; mov.f32 %f1632, 0f00000000; mov.f32 %f1633, %f1632; mov.f32 %f1634, %f1632; mov.f32 %f1635, %f1632; mov.f32 %f1636, %f1632; mov.f32 %f1637, %f1632; mov.f32 %f1638, %f1632; mov.f32 %f1639, %f1632; mov.f32 %f1640, %f1632; mov.f32 %f1641, %f1632; mov.f32 %f1642, %f1632; mov.f32 %f1643, %f1632; @%p47 bra $L__BB0_39; mul.f32 %f1632, %f1628, %f137; mul.f32 %f649, %f28, %f88; fma.rn.f32 %f650, %f26, %f87, %f649; fma.rn.f32 %f651, %f30, %f89, %f650; cvt.sat.f32.f32 %f652, %f651; mul.f32 %f653, %f1628, %f1631; mul.f32 %f654, %f653, %f652; mul.f32 %f655, %f1629, %f1631; mul.f32 %f656, %f655, %f652; mul.f32 %f657, %f1630, %f1631; mul.f32 %f658, %f657, %f652; mul.f32 %f659, %f38, %f88; fma.rn.f32 %f660, %f37, %f87, %f659; fma.rn.f32 %f661, %f39, %f89, %f660; cvt.sat.f32.f32 %f662, %f661; mul.f32 %f663, %f653, %f662; mul.f32 %f664, %f655, %f662; mul.f32 %f665, %f657, %f662; mul.f32 %f666, %f47, %f88; fma.rn.f32 %f667, %f46, %f87, %f666; fma.rn.f32 %f668, %f48, %f89, %f667; cvt.sat.f32.f32 %f669, %f668; mul.f32 %f670, %f653, %f669; mul.f32 %f671, %f655, %f669; mul.f32 %f672, %f657, %f669; add.f32 %f673, %f654, %f663; add.f32 %f674, %f656, %f664; add.f32 %f675, %f658, %f665; add.f32 %f676, %f673, %f670; add.f32 %f677, %f674, %f671; add.f32 %f678, %f675, %f672; mul.f32 %f679, %f676, 0f3F13CD3A; mul.f32 %f680, %f677, 0f3F13CD3A; mul.f32 %f681, %f678, 0f3F13CD3A; div.rn.f32 %f682, %f1632, %f679; mul.f32 %f1633, %f1629, %f137; div.rn.f32 %f683, %f1633, %f680; mul.f32 %f1634, %f1630, %f137; div.rn.f32 %f684, %f1634, %f681; setp.eq.f32 %p49, %f1632, 0f00000000; selp.f32 %f685, 0f00000000, %f682, %p49; setp.eq.f32 %p50, %f1633, 0f00000000; selp.f32 %f686, 0f00000000, %f683, %p50; setp.eq.f32 %p51, %f1634, 0f00000000; selp.f32 %f687, 0f00000000, %f684, %p51; mul.f32 %f1635, %f654, %f685; mul.f32 %f1636, %f656, %f686; mul.f32 %f1637, %f658, %f687; mul.f32 %f1638, %f663, %f685; mul.f32 %f1639, %f664, %f686; mul.f32 %f1640, %f665, %f687; mul.f32 %f1641, %f670, %f685; mul.f32 %f1642, %f671, %f686; mul.f32 %f1643, %f672, %f687; mov.pred %p188, 0; $L__BB0_39: @%p188 bra $L__BB0_51; setp.eq.s32 %p52, %r46, 0; mov.u16 %rs182, 0; mov.f32 %f1649, 0f3F800000; @%p52 bra $L__BB0_48; abs.s32 %r10, %r46; setp.lt.s32 %p53, %r10, 1; mov.f32 %f1645, 0f00000000; @%p53 bra $L__BB0_47; mov.f32 %f697, 0f38D1B717; max.f32 %f698, %f696, %f697; and.b32 %r65, %r425, 32; setp.eq.s32 %p54, %r65, 0; mov.u32 %r424, 0; selp.f32 %f699, 0f3F800000, 0f41200000, %p54; mul.f32 %f162, %f699, %f698; and.b32 %r11, %r10, 1; setp.eq.s32 %p55, %r10, 1; mov.f32 %f1645, 0f00000000; @%p55 bra $L__BB0_45; sub.s32 %r423, %r10, %r11; $L__BB0_44: cvt.rn.f32.s32 %f719, %r424; add.f32 %f720, %f16, %f719; sub.f32 %f721, %f17, %f719; mul.f32 %f722, %f720, 0f3DD32618; cvt.rmi.f32.f32 %f723, %f722; sub.f32 %f724, %f722, %f723; mul.f32 %f725, %f721, 0f3DD2F1AA; cvt.rmi.f32.f32 %f726, %f725; sub.f32 %f727, %f725, %f726; mul.f32 %f728, %f719, 0f3DC74539; cvt.rmi.f32.f32 %f729, %f728; sub.f32 %f730, %f728, %f729; add.f32 %f731, %f727, 0f420551EC; add.f32 %f732, %f724, 0f420551EC; add.f32 %f733, %f730, 0f420551EC; mul.f32 %f734, %f724, %f731; fma.rn.f32 %f735, %f727, %f732, %f734; fma.rn.f32 %f736, %f730, %f733, %f735; add.f32 %f737, %f724, %f736; add.f32 %f738, %f727, %f736; add.f32 %f739, %f730, %f736; add.f32 %f740, %f737, %f738; mul.f32 %f741, %f739, %f740; cvt.rmi.f32.f32 %f742, %f741; sub.f32 %f743, %f741, %f742; add.f32 %f744, %f737, %f737; mul.f32 %f745, %f738, %f744; cvt.rmi.f32.f32 %f746, %f745; sub.f32 %f747, %f745, %f746; mul.f32 %f748, %f737, %f740; cvt.rmi.f32.f32 %f749, %f748; sub.f32 %f750, %f748, %f749; fma.rn.f32 %f751, %f743, 0f40000000, 0fBF800000; fma.rn.f32 %f752, %f747, 0f40000000, 0fBF800000; fma.rn.f32 %f753, %f750, 0f40000000, 0fBF800000; ld.global.f32 %f754, [%rd6+16]; ld.global.f32 %f755, [%rd6+4]; fma.rn.f32 %f756, %f754, %f751, %f755; ld.global.f32 %f757, [%rd6+8]; fma.rn.f32 %f758, %f754, %f752, %f757; ld.global.f32 %f759, [%rd6+12]; fma.rn.f32 %f760, %f754, %f753, %f759; sub.f32 %f761, %f756, %f10; sub.f32 %f762, %f758, %f11; sub.f32 %f763, %f760, %f12; mul.f32 %f764, %f762, %f762; fma.rn.f32 %f765, %f761, %f761, %f764; fma.rn.f32 %f766, %f763, %f763, %f765; sqrt.rn.f32 %f708, %f766; rcp.rn.f32 %f767, %f708; mul.f32 %f704, %f767, %f761; mul.f32 %f705, %f767, %f762; mul.f32 %f706, %f767, %f763; mov.f32 %f718, 0f00000000; mov.u32 %r174, 2; mov.u32 %r176, 1; mov.u32 %r177, 1065353216; mov.u32 %r208, 0; // begin inline asm call(%r67,%r68,%r69,%r70,%r71,%r72,%r73,%r74,%r75,%r76,%r77,%r78,%r79,%r80,%r81,%r82,%r83,%r84,%r85,%r86,%r87,%r88,%r89,%r90,%r91,%r92,%r93,%r94,%r95,%r96,%r97,%r98),_optix_trace_typed_32,(%r208,%rd4,%f13,%f14,%f15,%f704,%f705,%f706,%f162,%f708,%f718,%r176,%r208,%r176,%r174,%r176,%r176,%r177,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208); // end inline asm mov.b32 %f768, %r67; add.f32 %f769, %f1645, %f768; add.s32 %r209, %r424, 1; cvt.rn.f32.s32 %f770, %r209; add.f32 %f771, %f16, %f770; sub.f32 %f772, %f17, %f770; mul.f32 %f773, %f771, 0f3DD32618; cvt.rmi.f32.f32 %f774, %f773; sub.f32 %f775, %f773, %f774; mul.f32 %f776, %f772, 0f3DD2F1AA; cvt.rmi.f32.f32 %f777, %f776; sub.f32 %f778, %f776, %f777; mul.f32 %f779, %f770, 0f3DC74539; cvt.rmi.f32.f32 %f780, %f779; sub.f32 %f781, %f779, %f780; add.f32 %f782, %f778, 0f420551EC; add.f32 %f783, %f775, 0f420551EC; add.f32 %f784, %f781, 0f420551EC; mul.f32 %f785, %f775, %f782; fma.rn.f32 %f786, %f778, %f783, %f785; fma.rn.f32 %f787, %f781, %f784, %f786; add.f32 %f788, %f775, %f787; add.f32 %f789, %f778, %f787; add.f32 %f790, %f781, %f787; add.f32 %f791, %f788, %f789; mul.f32 %f792, %f790, %f791; cvt.rmi.f32.f32 %f793, %f792; sub.f32 %f794, %f792, %f793; add.f32 %f795, %f788, %f788; mul.f32 %f796, %f789, %f795; cvt.rmi.f32.f32 %f797, %f796; sub.f32 %f798, %f796, %f797; mul.f32 %f799, %f788, %f791; cvt.rmi.f32.f32 %f800, %f799; sub.f32 %f801, %f799, %f800; fma.rn.f32 %f802, %f794, 0f40000000, 0fBF800000; fma.rn.f32 %f803, %f798, 0f40000000, 0fBF800000; fma.rn.f32 %f804, %f801, 0f40000000, 0fBF800000; ld.global.f32 %f805, [%rd6+16]; ld.global.f32 %f806, [%rd6+4]; fma.rn.f32 %f807, %f805, %f802, %f806; ld.global.f32 %f808, [%rd6+8]; fma.rn.f32 %f809, %f805, %f803, %f808; ld.global.f32 %f810, [%rd6+12]; fma.rn.f32 %f811, %f805, %f804, %f810; sub.f32 %f812, %f807, %f10; sub.f32 %f813, %f809, %f11; sub.f32 %f814, %f811, %f12; mul.f32 %f815, %f813, %f813; fma.rn.f32 %f816, %f812, %f812, %f815; fma.rn.f32 %f817, %f814, %f814, %f816; sqrt.rn.f32 %f717, %f817; rcp.rn.f32 %f818, %f717; mul.f32 %f713, %f818, %f812; mul.f32 %f714, %f818, %f813; mul.f32 %f715, %f818, %f814; // begin inline asm call(%r138,%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164,%r165,%r166,%r167,%r168,%r169),_optix_trace_typed_32,(%r208,%rd4,%f13,%f14,%f15,%f713,%f714,%f715,%f162,%f717,%f718,%r176,%r208,%r176,%r174,%r176,%r176,%r177,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208,%r208); // end inline asm mov.b32 %f819, %r138; add.f32 %f1645, %f769, %f819; add.s32 %r424, %r424, 2; add.s32 %r423, %r423, -2; setp.ne.s32 %p56, %r423, 0; @%p56 bra $L__BB0_44; $L__BB0_45: setp.eq.s32 %p57, %r11, 0; @%p57 bra $L__BB0_47; cvt.rn.f32.s32 %f829, %r424; add.f32 %f830, %f16, %f829; sub.f32 %f831, %f17, %f829; mul.f32 %f832, %f830, 0f3DD32618; cvt.rmi.f32.f32 %f833, %f832; sub.f32 %f834, %f832, %f833; mul.f32 %f835, %f831, 0f3DD2F1AA; cvt.rmi.f32.f32 %f836, %f835; sub.f32 %f837, %f835, %f836; mul.f32 %f838, %f829, 0f3DC74539; cvt.rmi.f32.f32 %f839, %f838; sub.f32 %f840, %f838, %f839; add.f32 %f841, %f837, 0f420551EC; add.f32 %f842, %f834, 0f420551EC; add.f32 %f843, %f840, 0f420551EC; mul.f32 %f844, %f834, %f841; fma.rn.f32 %f845, %f837, %f842, %f844; fma.rn.f32 %f846, %f840, %f843, %f845; add.f32 %f847, %f834, %f846; add.f32 %f848, %f837, %f846; add.f32 %f849, %f840, %f846; add.f32 %f850, %f847, %f848; mul.f32 %f851, %f849, %f850; cvt.rmi.f32.f32 %f852, %f851; sub.f32 %f853, %f851, %f852; add.f32 %f854, %f847, %f847; mul.f32 %f855, %f848, %f854; cvt.rmi.f32.f32 %f856, %f855; sub.f32 %f857, %f855, %f856; mul.f32 %f858, %f847, %f850; cvt.rmi.f32.f32 %f859, %f858; sub.f32 %f860, %f858, %f859; fma.rn.f32 %f861, %f853, 0f40000000, 0fBF800000; fma.rn.f32 %f862, %f857, 0f40000000, 0fBF800000; fma.rn.f32 %f863, %f860, 0f40000000, 0fBF800000; ld.global.f32 %f864, [%rd6+16]; ld.global.f32 %f865, [%rd6+4]; fma.rn.f32 %f866, %f864, %f861, %f865; ld.global.f32 %f867, [%rd6+8]; fma.rn.f32 %f868, %f864, %f862, %f867; ld.global.f32 %f869, [%rd6+12]; fma.rn.f32 %f870, %f864, %f863, %f869; sub.f32 %f871, %f866, %f10; sub.f32 %f872, %f868, %f11; sub.f32 %f873, %f870, %f12; mul.f32 %f874, %f872, %f872; fma.rn.f32 %f875, %f871, %f871, %f874; fma.rn.f32 %f876, %f873, %f873, %f875; sqrt.rn.f32 %f827, %f876; rcp.rn.f32 %f877, %f827; mul.f32 %f823, %f877, %f871; mul.f32 %f824, %f877, %f872; mul.f32 %f825, %f877, %f873; mov.f32 %f828, 0f00000000; mov.u32 %r246, 2; mov.u32 %r248, 1; mov.u32 %r249, 1065353216; mov.u32 %r280, 0; // begin inline asm call(%r210,%r211,%r212,%r213,%r214,%r215,%r216,%r217,%r218,%r219,%r220,%r221,%r222,%r223,%r224,%r225,%r226,%r227,%r228,%r229,%r230,%r231,%r232,%r233,%r234,%r235,%r236,%r237,%r238,%r239,%r240,%r241),_optix_trace_typed_32,(%r280,%rd4,%f13,%f14,%f15,%f823,%f824,%f825,%f162,%f827,%f828,%r248,%r280,%r248,%r246,%r248,%r248,%r249,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280,%r280); // end inline asm mov.b32 %f878, %r210; add.f32 %f1645, %f1645, %f878; $L__BB0_47: cvt.rn.f32.s32 %f879, %r10; div.rn.f32 %f1649, %f1645, %f879; shr.u32 %r281, %r46, 31; cvt.u16.u32 %rs182, %r281; $L__BB0_48: fma.rn.f32 %f1650, %f1632, %f1649, %f1650; fma.rn.f32 %f1651, %f1633, %f1649, %f1651; fma.rn.f32 %f1652, %f1634, %f1649, %f1652; fma.rn.f32 %f1653, %f1635, %f1649, %f1653; fma.rn.f32 %f1654, %f1636, %f1649, %f1654; fma.rn.f32 %f1655, %f1637, %f1649, %f1655; fma.rn.f32 %f1656, %f1638, %f1649, %f1656; fma.rn.f32 %f1657, %f1639, %f1649, %f1657; fma.rn.f32 %f1658, %f1640, %f1649, %f1658; fma.rn.f32 %f1659, %f1641, %f1649, %f1659; fma.rn.f32 %f1660, %f1642, %f1649, %f1660; fma.rn.f32 %f1661, %f1643, %f1649, %f1661; setp.eq.s16 %p58, %rs182, 0; @%p58 bra $L__BB0_50; div.rn.f32 %f880, %f1632, %f471; div.rn.f32 %f881, %f880, %f103; cvt.sat.f32.f32 %f882, %f881; mul.f32 %f1649, %f1649, %f882; $L__BB0_50: add.f32 %f1662, %f1662, %f1649; $L__BB0_51: cvt.u64.u32 %rd73, %r421; cvt.u32.u64 %r282, %rd73; add.s32 %r421, %r282, 1; setp.lt.u32 %p59, %r421, %r5; @%p59 bra $L__BB0_15; bra.uni $L__BB0_53; $L__BB0_52: mov.f32 %f1650, 0f00000000; ld.const.u32 %r425, [params+104]; mov.f32 %f1651, %f1650; mov.f32 %f1652, %f1650; mov.f32 %f1653, %f1650; mov.f32 %f1654, %f1650; mov.f32 %f1655, %f1650; mov.f32 %f1656, %f1650; mov.f32 %f1657, %f1650; mov.f32 %f1658, %f1650; mov.f32 %f1659, %f1650; mov.f32 %f1660, %f1650; mov.f32 %f1661, %f1650; mov.f32 %f1662, %f1650; $L__BB0_53: and.b32 %r283, %r425, 8; setp.eq.s32 %p60, %r283, 0; @%p60 bra $L__BB0_67; cvt.sat.f32.f32 %f212, %f1662; ld.const.u64 %rd35, [params+192]; cvta.to.global.u64 %rd8, %rd35; ld.const.u32 %r284, [params+184]; mad.lo.s32 %r285, %r284, %r4, %r3; cvt.u64.u32 %rd9, %r285; mov.f32 %f897, 0f3E68BA2E; cvt.rzi.f32.f32 %f898, %f897; add.f32 %f899, %f898, %f898; mov.f32 %f900, 0f3EE8BA2E; sub.f32 %f901, %f900, %f899; abs.f32 %f213, %f901; abs.f32 %f214, %f212; setp.lt.f32 %p61, %f214, 0f00800000; mul.f32 %f902, %f214, 0f4B800000; selp.f32 %f903, %f902, %f214, %p61; selp.f32 %f904, 0fC3170000, 0fC2FE0000, %p61; mov.b32 %r286, %f903; and.b32 %r287, %r286, 8388607; or.b32 %r288, %r287, 1065353216; mov.b32 %f905, %r288; shr.u32 %r289, %r286, 23; cvt.rn.f32.u32 %f906, %r289; add.f32 %f907, %f904, %f906; setp.gt.f32 %p62, %f905, 0f3FB504F3; mul.f32 %f908, %f905, 0f3F000000; add.f32 %f909, %f907, 0f3F800000; selp.f32 %f910, %f909, %f907, %p62; selp.f32 %f911, %f908, %f905, %p62; add.f32 %f912, %f911, 0fBF800000; add.f32 %f913, %f911, 0f3F800000; rcp.approx.ftz.f32 %f914, %f913; add.f32 %f915, %f912, %f912; mul.f32 %f916, %f915, %f914; mul.f32 %f917, %f916, %f916; mov.f32 %f918, 0f3C4CAF63; mov.f32 %f919, 0f3B18F0FE; fma.rn.f32 %f920, %f919, %f917, %f918; mov.f32 %f921, 0f3DAAAABD; fma.rn.f32 %f922, %f920, %f917, %f921; mul.rn.f32 %f923, %f922, %f917; mul.rn.f32 %f924, %f923, %f916; sub.f32 %f925, %f912, %f916; add.f32 %f926, %f925, %f925; neg.f32 %f927, %f916; fma.rn.f32 %f928, %f927, %f912, %f926; mul.rn.f32 %f929, %f914, %f928; add.f32 %f930, %f924, %f916; sub.f32 %f931, %f916, %f930; add.f32 %f932, %f924, %f931; add.f32 %f933, %f929, %f932; add.f32 %f934, %f930, %f933; sub.f32 %f935, %f930, %f934; add.f32 %f936, %f933, %f935; mov.f32 %f937, 0f3F317200; mul.rn.f32 %f938, %f910, %f937; mov.f32 %f939, 0f35BFBE8E; mul.rn.f32 %f940, %f910, %f939; add.f32 %f941, %f938, %f934; sub.f32 %f942, %f938, %f941; add.f32 %f943, %f934, %f942; add.f32 %f944, %f936, %f943; add.f32 %f945, %f940, %f944; add.f32 %f946, %f941, %f945; sub.f32 %f947, %f941, %f946; add.f32 %f948, %f945, %f947; mul.rn.f32 %f949, %f900, %f946; neg.f32 %f950, %f949; fma.rn.f32 %f951, %f900, %f946, %f950; fma.rn.f32 %f952, %f900, %f948, %f951; mov.f32 %f953, 0f00000000; fma.rn.f32 %f954, %f953, %f946, %f952; add.rn.f32 %f955, %f949, %f954; neg.f32 %f956, %f955; add.rn.f32 %f957, %f949, %f956; add.rn.f32 %f958, %f957, %f954; mov.b32 %r290, %f955; setp.eq.s32 %p63, %r290, 1118925336; add.s32 %r291, %r290, -1; mov.b32 %f959, %r291; add.f32 %f960, %f958, 0f37000000; selp.f32 %f215, %f960, %f958, %p63; selp.f32 %f961, %f959, %f955, %p63; mov.f32 %f962, 0f3FB8AA3B; mul.rn.f32 %f963, %f961, %f962; cvt.rzi.f32.f32 %f964, %f963; abs.f32 %f965, %f964; setp.gt.f32 %p64, %f965, 0f42FC0000; mov.b32 %r292, %f964; and.b32 %r293, %r292, -2147483648; or.b32 %r294, %r293, 1123811328; mov.b32 %f966, %r294; selp.f32 %f967, %f966, %f964, %p64; mov.f32 %f968, 0fBF317218; fma.rn.f32 %f969, %f967, %f968, %f961; mov.f32 %f970, 0f3102E308; fma.rn.f32 %f971, %f967, %f970, %f969; mul.f32 %f972, %f971, 0f3FB8AA3B; add.f32 %f973, %f967, 0f4B40007F; mov.b32 %r295, %f973; shl.b32 %r296, %r295, 23; mov.b32 %f974, %r296; ex2.approx.ftz.f32 %f975, %f972; mul.f32 %f216, %f975, %f974; setp.eq.f32 %p65, %f216, 0f7F800000; mov.f32 %f1676, 0f7F800000; @%p65 bra $L__BB0_56; fma.rn.f32 %f1676, %f216, %f215, %f216; $L__BB0_56: setp.lt.f32 %p66, %f212, 0f00000000; setp.eq.f32 %p67, %f213, 0f3F800000; and.pred %p3, %p66, %p67; setp.eq.f32 %p68, %f212, 0f00000000; @%p68 bra $L__BB0_60; bra.uni $L__BB0_57; $L__BB0_60: add.f32 %f980, %f212, %f212; selp.f32 %f1678, %f980, 0f00000000, %p67; bra.uni $L__BB0_61; $L__BB0_161: mov.f32 %f1565, 0f00000000; mov.u32 %r427, 0; // begin inline asm { cvt.rn.f16.f32 %rs129, %f1565;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs128, %f1565;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs127, %f1565;} // end inline asm mov.u16 %rs130, 0; st.global.v4.u16 [%rd16], {%rs127, %rs128, %rs129, %rs130}; $L__BB0_162: ld.const.u64 %rd64, [params+256]; cvta.to.global.u64 %rd65, %rd64; ld.const.u32 %r415, [params+248]; mad.lo.s32 %r416, %r415, %r4, %r3; mul.wide.u32 %rd66, %r416, 8; add.s64 %rd17, %rd65, %rd66; setp.eq.s32 %p185, %r427, 0; @%p185 bra $L__BB0_164; ld.global.v4.u16 {%rs137, %rs138, %rs139, %rs140}, [%rd17]; // begin inline asm { cvt.f32.f16 %f1566, %rs137;} // end inline asm // begin inline asm { cvt.f32.f16 %f1567, %rs138;} // end inline asm // begin inline asm { cvt.f32.f16 %f1568, %rs139;} // end inline asm add.f32 %f1569, %f1566, 0f00000000; add.f32 %f1570, %f1567, 0f00000000; add.f32 %f1571, %f1568, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs136, %f1571;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs135, %f1570;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs134, %f1569;} // end inline asm mov.u16 %rs141, 0; st.global.v4.u16 [%rd17], {%rs134, %rs135, %rs136, %rs141}; bra.uni $L__BB0_165; $L__BB0_164: mov.f32 %f1574, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs144, %f1574;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs143, %f1574;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs142, %f1574;} // end inline asm mov.u16 %rs145, 0; st.global.v4.u16 [%rd17], {%rs142, %rs143, %rs144, %rs145}; $L__BB0_165: ld.const.u64 %rd67, [params+272]; cvta.to.global.u64 %rd68, %rd67; ld.const.u32 %r417, [params+264]; mad.lo.s32 %r418, %r417, %r4, %r3; mul.wide.u32 %rd69, %r418, 8; add.s64 %rd18, %rd68, %rd69; @%p185 bra $L__BB0_167; ld.global.v4.u16 {%rs152, %rs153, %rs154, %rs155}, [%rd18]; // begin inline asm { cvt.f32.f16 %f1575, %rs152;} // end inline asm // begin inline asm { cvt.f32.f16 %f1576, %rs153;} // end inline asm // begin inline asm { cvt.f32.f16 %f1577, %rs154;} // end inline asm add.f32 %f1578, %f1575, 0f00000000; add.f32 %f1579, %f1576, 0f00000000; add.f32 %f1580, %f1577, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs151, %f1580;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs150, %f1579;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs149, %f1578;} // end inline asm mov.u16 %rs156, 0; st.global.v4.u16 [%rd18], {%rs149, %rs150, %rs151, %rs156}; bra.uni $L__BB0_168; $L__BB0_167: mov.f32 %f1583, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs159, %f1583;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs158, %f1583;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs157, %f1583;} // end inline asm mov.u16 %rs160, 0; st.global.v4.u16 [%rd18], {%rs157, %rs158, %rs159, %rs160}; $L__BB0_168: ld.const.u64 %rd70, [params+288]; cvta.to.global.u64 %rd71, %rd70; ld.const.u32 %r419, [params+280]; mad.lo.s32 %r420, %r419, %r4, %r3; mul.wide.u32 %rd72, %r420, 8; add.s64 %rd19, %rd71, %rd72; @%p185 bra $L__BB0_170; ld.global.v4.u16 {%rs167, %rs168, %rs169, %rs170}, [%rd19]; // begin inline asm { cvt.f32.f16 %f1584, %rs167;} // end inline asm // begin inline asm { cvt.f32.f16 %f1585, %rs168;} // end inline asm // begin inline asm { cvt.f32.f16 %f1586, %rs169;} // end inline asm add.f32 %f1587, %f1584, 0f00000000; add.f32 %f1588, %f1585, 0f00000000; add.f32 %f1589, %f1586, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs166, %f1589;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs165, %f1588;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs164, %f1587;} // end inline asm mov.u16 %rs171, 0; st.global.v4.u16 [%rd19], {%rs164, %rs165, %rs166, %rs171}; bra.uni $L__BB0_171; $L__BB0_170: mov.f32 %f1592, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs174, %f1592;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs173, %f1592;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs172, %f1592;} // end inline asm mov.u16 %rs175, 0; st.global.v4.u16 [%rd19], {%rs172, %rs173, %rs174, %rs175}; bra.uni $L__BB0_171; $L__BB0_57: mov.b32 %r297, %f1676; xor.b32 %r298, %r297, -2147483648; mov.b32 %f976, %r298; selp.f32 %f1678, %f976, %f1676, %p3; setp.geu.f32 %p69, %f212, 0f00000000; @%p69 bra $L__BB0_61; mov.f32 %f977, 0f3EE8BA2E; cvt.rzi.f32.f32 %f978, %f977; setp.eq.f32 %p70, %f978, 0f3EE8BA2E; @%p70 bra $L__BB0_61; mov.f32 %f1678, 0f7FFFFFFF; $L__BB0_61: add.f32 %f981, %f214, 0f3EE8BA2E; mov.b32 %r299, %f981; setp.lt.s32 %p72, %r299, 2139095040; @%p72 bra $L__BB0_66; setp.gtu.f32 %p73, %f214, 0f7F800000; @%p73 bra $L__BB0_65; bra.uni $L__BB0_63; $L__BB0_65: add.f32 %f1678, %f212, 0f3EE8BA2E; bra.uni $L__BB0_66; $L__BB0_63: setp.neu.f32 %p74, %f214, 0f7F800000; @%p74 bra $L__BB0_66; selp.f32 %f1678, 0fFF800000, 0f7F800000, %p3; $L__BB0_66: mul.f32 %f982, %f1678, 0f437F0000; setp.eq.f32 %p75, %f212, 0f3F800000; selp.f32 %f983, 0f437F0000, %f982, %p75; cvt.rzi.u32.f32 %r300, %f983; shl.b64 %rd36, %rd9, 1; add.s64 %rd37, %rd8, %rd36; cvt.u16.u32 %rs47, %r300; mov.u16 %rs48, 255; st.global.v2.u8 [%rd37], {%rs47, %rs48}; $L__BB0_67: and.b32 %r301, %r425, 1; setp.eq.b32 %p76, %r301, 1; mov.pred %p77, 0; xor.pred %p78, %p76, %p77; not.pred %p79, %p78; @%p79 bra $L__BB0_141; mov.f32 %f985, 0f3E666666; cvt.rzi.f32.f32 %f986, %f985; add.f32 %f987, %f986, %f986; mov.f32 %f988, 0f3EE66666; sub.f32 %f989, %f988, %f987; abs.f32 %f225, %f989; abs.f32 %f226, %f1650; setp.lt.f32 %p80, %f226, 0f00800000; mul.f32 %f990, %f226, 0f4B800000; selp.f32 %f991, %f990, %f226, %p80; selp.f32 %f992, 0fC3170000, 0fC2FE0000, %p80; mov.b32 %r302, %f991; and.b32 %r303, %r302, 8388607; or.b32 %r304, %r303, 1065353216; mov.b32 %f993, %r304; shr.u32 %r305, %r302, 23; cvt.rn.f32.u32 %f994, %r305; add.f32 %f995, %f992, %f994; setp.gt.f32 %p81, %f993, 0f3FB504F3; mul.f32 %f996, %f993, 0f3F000000; add.f32 %f997, %f995, 0f3F800000; selp.f32 %f998, %f997, %f995, %p81; selp.f32 %f999, %f996, %f993, %p81; add.f32 %f1000, %f999, 0fBF800000; add.f32 %f1001, %f999, 0f3F800000; rcp.approx.ftz.f32 %f1002, %f1001; add.f32 %f1003, %f1000, %f1000; mul.f32 %f1004, %f1003, %f1002; mul.f32 %f1005, %f1004, %f1004; mov.f32 %f1006, 0f3C4CAF63; mov.f32 %f1007, 0f3B18F0FE; fma.rn.f32 %f1008, %f1007, %f1005, %f1006; mov.f32 %f1009, 0f3DAAAABD; fma.rn.f32 %f1010, %f1008, %f1005, %f1009; mul.rn.f32 %f1011, %f1010, %f1005; mul.rn.f32 %f1012, %f1011, %f1004; sub.f32 %f1013, %f1000, %f1004; add.f32 %f1014, %f1013, %f1013; neg.f32 %f1015, %f1004; fma.rn.f32 %f1016, %f1015, %f1000, %f1014; mul.rn.f32 %f1017, %f1002, %f1016; add.f32 %f1018, %f1012, %f1004; sub.f32 %f1019, %f1004, %f1018; add.f32 %f1020, %f1012, %f1019; add.f32 %f1021, %f1017, %f1020; add.f32 %f1022, %f1018, %f1021; sub.f32 %f1023, %f1018, %f1022; add.f32 %f1024, %f1021, %f1023; mov.f32 %f1025, 0f3F317200; mul.rn.f32 %f1026, %f998, %f1025; mov.f32 %f1027, 0f35BFBE8E; mul.rn.f32 %f1028, %f998, %f1027; add.f32 %f1029, %f1026, %f1022; sub.f32 %f1030, %f1026, %f1029; add.f32 %f1031, %f1022, %f1030; add.f32 %f1032, %f1024, %f1031; add.f32 %f1033, %f1028, %f1032; add.f32 %f1034, %f1029, %f1033; sub.f32 %f1035, %f1029, %f1034; add.f32 %f1036, %f1033, %f1035; mul.rn.f32 %f1037, %f988, %f1034; neg.f32 %f1038, %f1037; fma.rn.f32 %f1039, %f988, %f1034, %f1038; fma.rn.f32 %f1040, %f988, %f1036, %f1039; mov.f32 %f1041, 0f00000000; fma.rn.f32 %f1042, %f1041, %f1034, %f1040; add.rn.f32 %f1043, %f1037, %f1042; neg.f32 %f1044, %f1043; add.rn.f32 %f1045, %f1037, %f1044; add.rn.f32 %f1046, %f1045, %f1042; mov.b32 %r306, %f1043; setp.eq.s32 %p82, %r306, 1118925336; add.s32 %r307, %r306, -1; mov.b32 %f1047, %r307; add.f32 %f1048, %f1046, 0f37000000; selp.f32 %f227, %f1048, %f1046, %p82; selp.f32 %f1049, %f1047, %f1043, %p82; mov.f32 %f1050, 0f3FB8AA3B; mul.rn.f32 %f1051, %f1049, %f1050; cvt.rzi.f32.f32 %f1052, %f1051; abs.f32 %f1053, %f1052; setp.gt.f32 %p83, %f1053, 0f42FC0000; mov.b32 %r308, %f1052; and.b32 %r309, %r308, -2147483648; or.b32 %r310, %r309, 1123811328; mov.b32 %f1054, %r310; selp.f32 %f1055, %f1054, %f1052, %p83; mov.f32 %f1056, 0fBF317218; fma.rn.f32 %f1057, %f1055, %f1056, %f1049; mov.f32 %f1058, 0f3102E308; fma.rn.f32 %f1059, %f1055, %f1058, %f1057; mul.f32 %f1060, %f1059, 0f3FB8AA3B; add.f32 %f1061, %f1055, 0f4B40007F; mov.b32 %r311, %f1061; shl.b32 %r312, %r311, 23; mov.b32 %f1062, %r312; ex2.approx.ftz.f32 %f1063, %f1060; mul.f32 %f228, %f1063, %f1062; setp.eq.f32 %p84, %f228, 0f7F800000; mov.f32 %f1679, 0f7F800000; @%p84 bra $L__BB0_70; fma.rn.f32 %f1679, %f228, %f227, %f228; $L__BB0_70: setp.lt.f32 %p85, %f1650, 0f00000000; setp.eq.f32 %p86, %f225, 0f3F800000; and.pred %p4, %p85, %p86; setp.eq.f32 %p87, %f1650, 0f00000000; @%p87 bra $L__BB0_74; bra.uni $L__BB0_71; $L__BB0_74: add.f32 %f1068, %f1650, %f1650; selp.f32 %f1681, %f1068, 0f00000000, %p86; bra.uni $L__BB0_75; $L__BB0_71: mov.b32 %r313, %f1679; xor.b32 %r314, %r313, -2147483648; mov.b32 %f1064, %r314; selp.f32 %f1681, %f1064, %f1679, %p4; setp.geu.f32 %p88, %f1650, 0f00000000; @%p88 bra $L__BB0_75; mov.f32 %f1065, 0f3EE66666; cvt.rzi.f32.f32 %f1066, %f1065; setp.eq.f32 %p89, %f1066, 0f3EE66666; @%p89 bra $L__BB0_75; mov.f32 %f1681, 0f7FFFFFFF; $L__BB0_75: add.f32 %f1069, %f226, 0f3EE66666; mov.b32 %r315, %f1069; setp.lt.s32 %p91, %r315, 2139095040; @%p91 bra $L__BB0_80; setp.gtu.f32 %p92, %f226, 0f7F800000; @%p92 bra $L__BB0_79; bra.uni $L__BB0_77; $L__BB0_79: add.f32 %f1681, %f1650, 0f3EE66666; bra.uni $L__BB0_80; $L__BB0_77: setp.neu.f32 %p93, %f226, 0f7F800000; @%p93 bra $L__BB0_80; selp.f32 %f1681, 0fFF800000, 0f7F800000, %p4; $L__BB0_80: setp.eq.f32 %p94, %f1650, 0f3F800000; selp.f32 %f237, 0f3F800000, %f1681, %p94; abs.f32 %f238, %f1651; setp.lt.f32 %p95, %f238, 0f00800000; mul.f32 %f1071, %f238, 0f4B800000; selp.f32 %f1072, %f1071, %f238, %p95; selp.f32 %f1073, 0fC3170000, 0fC2FE0000, %p95; mov.b32 %r316, %f1072; and.b32 %r317, %r316, 8388607; or.b32 %r318, %r317, 1065353216; mov.b32 %f1074, %r318; shr.u32 %r319, %r316, 23; cvt.rn.f32.u32 %f1075, %r319; add.f32 %f1076, %f1073, %f1075; setp.gt.f32 %p96, %f1074, 0f3FB504F3; mul.f32 %f1077, %f1074, 0f3F000000; add.f32 %f1078, %f1076, 0f3F800000; selp.f32 %f1079, %f1078, %f1076, %p96; selp.f32 %f1080, %f1077, %f1074, %p96; add.f32 %f1081, %f1080, 0fBF800000; add.f32 %f1082, %f1080, 0f3F800000; rcp.approx.ftz.f32 %f1083, %f1082; add.f32 %f1084, %f1081, %f1081; mul.f32 %f1085, %f1084, %f1083; mul.f32 %f1086, %f1085, %f1085; mov.f32 %f1087, 0f3C4CAF63; mov.f32 %f1088, 0f3B18F0FE; fma.rn.f32 %f1089, %f1088, %f1086, %f1087; mov.f32 %f1090, 0f3DAAAABD; fma.rn.f32 %f1091, %f1089, %f1086, %f1090; mul.rn.f32 %f1092, %f1091, %f1086; mul.rn.f32 %f1093, %f1092, %f1085; sub.f32 %f1094, %f1081, %f1085; add.f32 %f1095, %f1094, %f1094; neg.f32 %f1096, %f1085; fma.rn.f32 %f1097, %f1096, %f1081, %f1095; mul.rn.f32 %f1098, %f1083, %f1097; add.f32 %f1099, %f1093, %f1085; sub.f32 %f1100, %f1085, %f1099; add.f32 %f1101, %f1093, %f1100; add.f32 %f1102, %f1098, %f1101; add.f32 %f1103, %f1099, %f1102; sub.f32 %f1104, %f1099, %f1103; add.f32 %f1105, %f1102, %f1104; mov.f32 %f1106, 0f3F317200; mul.rn.f32 %f1107, %f1079, %f1106; mov.f32 %f1108, 0f35BFBE8E; mul.rn.f32 %f1109, %f1079, %f1108; add.f32 %f1110, %f1107, %f1103; sub.f32 %f1111, %f1107, %f1110; add.f32 %f1112, %f1103, %f1111; add.f32 %f1113, %f1105, %f1112; add.f32 %f1114, %f1109, %f1113; add.f32 %f1115, %f1110, %f1114; sub.f32 %f1116, %f1110, %f1115; add.f32 %f1117, %f1114, %f1116; mov.f32 %f1118, 0f3EE66666; mul.rn.f32 %f1119, %f1118, %f1115; neg.f32 %f1120, %f1119; fma.rn.f32 %f1121, %f1118, %f1115, %f1120; fma.rn.f32 %f1122, %f1118, %f1117, %f1121; mov.f32 %f1123, 0f00000000; fma.rn.f32 %f1124, %f1123, %f1115, %f1122; add.rn.f32 %f1125, %f1119, %f1124; neg.f32 %f1126, %f1125; add.rn.f32 %f1127, %f1119, %f1126; add.rn.f32 %f1128, %f1127, %f1124; mov.b32 %r320, %f1125; setp.eq.s32 %p97, %r320, 1118925336; add.s32 %r321, %r320, -1; mov.b32 %f1129, %r321; add.f32 %f1130, %f1128, 0f37000000; selp.f32 %f239, %f1130, %f1128, %p97; selp.f32 %f1131, %f1129, %f1125, %p97; mov.f32 %f1132, 0f3FB8AA3B; mul.rn.f32 %f1133, %f1131, %f1132; cvt.rzi.f32.f32 %f1134, %f1133; abs.f32 %f1135, %f1134; setp.gt.f32 %p98, %f1135, 0f42FC0000; mov.b32 %r322, %f1134; and.b32 %r323, %r322, -2147483648; or.b32 %r324, %r323, 1123811328; mov.b32 %f1136, %r324; selp.f32 %f1137, %f1136, %f1134, %p98; mov.f32 %f1138, 0fBF317218; fma.rn.f32 %f1139, %f1137, %f1138, %f1131; mov.f32 %f1140, 0f3102E308; fma.rn.f32 %f1141, %f1137, %f1140, %f1139; mul.f32 %f1142, %f1141, 0f3FB8AA3B; add.f32 %f1143, %f1137, 0f4B40007F; mov.b32 %r325, %f1143; shl.b32 %r326, %r325, 23; mov.b32 %f1144, %r326; ex2.approx.ftz.f32 %f1145, %f1142; mul.f32 %f240, %f1145, %f1144; setp.eq.f32 %p99, %f240, 0f7F800000; mov.f32 %f1682, 0f7F800000; @%p99 bra $L__BB0_82; fma.rn.f32 %f1682, %f240, %f239, %f240; $L__BB0_82: setp.lt.f32 %p100, %f1651, 0f00000000; and.pred %p5, %p100, %p86; setp.eq.f32 %p102, %f1651, 0f00000000; @%p102 bra $L__BB0_86; bra.uni $L__BB0_83; $L__BB0_86: add.f32 %f1150, %f1651, %f1651; selp.f32 %f1684, %f1150, 0f00000000, %p86; bra.uni $L__BB0_87; $L__BB0_83: mov.b32 %r327, %f1682; xor.b32 %r328, %r327, -2147483648; mov.b32 %f1146, %r328; selp.f32 %f1684, %f1146, %f1682, %p5; setp.geu.f32 %p103, %f1651, 0f00000000; @%p103 bra $L__BB0_87; mov.f32 %f1147, 0f3EE66666; cvt.rzi.f32.f32 %f1148, %f1147; setp.eq.f32 %p104, %f1148, 0f3EE66666; @%p104 bra $L__BB0_87; mov.f32 %f1684, 0f7FFFFFFF; $L__BB0_87: add.f32 %f1151, %f238, 0f3EE66666; mov.b32 %r329, %f1151; setp.lt.s32 %p106, %r329, 2139095040; @%p106 bra $L__BB0_92; setp.gtu.f32 %p107, %f238, 0f7F800000; @%p107 bra $L__BB0_91; bra.uni $L__BB0_89; $L__BB0_91: add.f32 %f1684, %f1651, 0f3EE66666; bra.uni $L__BB0_92; $L__BB0_89: setp.neu.f32 %p108, %f238, 0f7F800000; @%p108 bra $L__BB0_92; selp.f32 %f1684, 0fFF800000, 0f7F800000, %p5; $L__BB0_92: setp.eq.f32 %p109, %f1651, 0f3F800000; selp.f32 %f249, 0f3F800000, %f1684, %p109; abs.f32 %f250, %f1652; setp.lt.f32 %p110, %f250, 0f00800000; mul.f32 %f1153, %f250, 0f4B800000; selp.f32 %f1154, %f1153, %f250, %p110; selp.f32 %f1155, 0fC3170000, 0fC2FE0000, %p110; mov.b32 %r330, %f1154; and.b32 %r331, %r330, 8388607; or.b32 %r332, %r331, 1065353216; mov.b32 %f1156, %r332; shr.u32 %r333, %r330, 23; cvt.rn.f32.u32 %f1157, %r333; add.f32 %f1158, %f1155, %f1157; setp.gt.f32 %p111, %f1156, 0f3FB504F3; mul.f32 %f1159, %f1156, 0f3F000000; add.f32 %f1160, %f1158, 0f3F800000; selp.f32 %f1161, %f1160, %f1158, %p111; selp.f32 %f1162, %f1159, %f1156, %p111; add.f32 %f1163, %f1162, 0fBF800000; add.f32 %f1164, %f1162, 0f3F800000; rcp.approx.ftz.f32 %f1165, %f1164; add.f32 %f1166, %f1163, %f1163; mul.f32 %f1167, %f1166, %f1165; mul.f32 %f1168, %f1167, %f1167; mov.f32 %f1169, 0f3C4CAF63; mov.f32 %f1170, 0f3B18F0FE; fma.rn.f32 %f1171, %f1170, %f1168, %f1169; mov.f32 %f1172, 0f3DAAAABD; fma.rn.f32 %f1173, %f1171, %f1168, %f1172; mul.rn.f32 %f1174, %f1173, %f1168; mul.rn.f32 %f1175, %f1174, %f1167; sub.f32 %f1176, %f1163, %f1167; add.f32 %f1177, %f1176, %f1176; neg.f32 %f1178, %f1167; fma.rn.f32 %f1179, %f1178, %f1163, %f1177; mul.rn.f32 %f1180, %f1165, %f1179; add.f32 %f1181, %f1175, %f1167; sub.f32 %f1182, %f1167, %f1181; add.f32 %f1183, %f1175, %f1182; add.f32 %f1184, %f1180, %f1183; add.f32 %f1185, %f1181, %f1184; sub.f32 %f1186, %f1181, %f1185; add.f32 %f1187, %f1184, %f1186; mov.f32 %f1188, 0f3F317200; mul.rn.f32 %f1189, %f1161, %f1188; mov.f32 %f1190, 0f35BFBE8E; mul.rn.f32 %f1191, %f1161, %f1190; add.f32 %f1192, %f1189, %f1185; sub.f32 %f1193, %f1189, %f1192; add.f32 %f1194, %f1185, %f1193; add.f32 %f1195, %f1187, %f1194; add.f32 %f1196, %f1191, %f1195; add.f32 %f1197, %f1192, %f1196; sub.f32 %f1198, %f1192, %f1197; add.f32 %f1199, %f1196, %f1198; mov.f32 %f1200, 0f3EE66666; mul.rn.f32 %f1201, %f1200, %f1197; neg.f32 %f1202, %f1201; fma.rn.f32 %f1203, %f1200, %f1197, %f1202; fma.rn.f32 %f1204, %f1200, %f1199, %f1203; mov.f32 %f1205, 0f00000000; fma.rn.f32 %f1206, %f1205, %f1197, %f1204; add.rn.f32 %f1207, %f1201, %f1206; neg.f32 %f1208, %f1207; add.rn.f32 %f1209, %f1201, %f1208; add.rn.f32 %f1210, %f1209, %f1206; mov.b32 %r334, %f1207; setp.eq.s32 %p112, %r334, 1118925336; add.s32 %r335, %r334, -1; mov.b32 %f1211, %r335; add.f32 %f1212, %f1210, 0f37000000; selp.f32 %f251, %f1212, %f1210, %p112; selp.f32 %f1213, %f1211, %f1207, %p112; mov.f32 %f1214, 0f3FB8AA3B; mul.rn.f32 %f1215, %f1213, %f1214; cvt.rzi.f32.f32 %f1216, %f1215; abs.f32 %f1217, %f1216; setp.gt.f32 %p113, %f1217, 0f42FC0000; mov.b32 %r336, %f1216; and.b32 %r337, %r336, -2147483648; or.b32 %r338, %r337, 1123811328; mov.b32 %f1218, %r338; selp.f32 %f1219, %f1218, %f1216, %p113; mov.f32 %f1220, 0fBF317218; fma.rn.f32 %f1221, %f1219, %f1220, %f1213; mov.f32 %f1222, 0f3102E308; fma.rn.f32 %f1223, %f1219, %f1222, %f1221; mul.f32 %f1224, %f1223, 0f3FB8AA3B; add.f32 %f1225, %f1219, 0f4B40007F; mov.b32 %r339, %f1225; shl.b32 %r340, %r339, 23; mov.b32 %f1226, %r340; ex2.approx.ftz.f32 %f1227, %f1224; mul.f32 %f252, %f1227, %f1226; setp.eq.f32 %p114, %f252, 0f7F800000; mov.f32 %f1685, 0f7F800000; @%p114 bra $L__BB0_94; fma.rn.f32 %f1685, %f252, %f251, %f252; $L__BB0_94: setp.lt.f32 %p115, %f1652, 0f00000000; and.pred %p6, %p115, %p86; setp.eq.f32 %p117, %f1652, 0f00000000; @%p117 bra $L__BB0_98; bra.uni $L__BB0_95; $L__BB0_98: add.f32 %f1232, %f1652, %f1652; selp.f32 %f1687, %f1232, 0f00000000, %p86; bra.uni $L__BB0_99; $L__BB0_95: mov.b32 %r341, %f1685; xor.b32 %r342, %r341, -2147483648; mov.b32 %f1228, %r342; selp.f32 %f1687, %f1228, %f1685, %p6; setp.geu.f32 %p118, %f1652, 0f00000000; @%p118 bra $L__BB0_99; mov.f32 %f1229, 0f3EE66666; cvt.rzi.f32.f32 %f1230, %f1229; setp.eq.f32 %p119, %f1230, 0f3EE66666; @%p119 bra $L__BB0_99; mov.f32 %f1687, 0f7FFFFFFF; $L__BB0_99: add.f32 %f1233, %f250, 0f3EE66666; mov.b32 %r343, %f1233; setp.lt.s32 %p121, %r343, 2139095040; @%p121 bra $L__BB0_104; setp.gtu.f32 %p122, %f250, 0f7F800000; @%p122 bra $L__BB0_103; bra.uni $L__BB0_101; $L__BB0_103: add.f32 %f1687, %f1652, 0f3EE66666; bra.uni $L__BB0_104; $L__BB0_101: setp.neu.f32 %p123, %f250, 0f7F800000; @%p123 bra $L__BB0_104; selp.f32 %f1687, 0fFF800000, 0f7F800000, %p6; $L__BB0_104: setp.eq.f32 %p124, %f1652, 0f3F800000; mov.f32 %f1235, 0f3F800000; selp.f32 %f1236, 0f3F800000, %f1687, %p124; ld.const.u64 %rd38, [params+144]; cvta.to.global.u64 %rd10, %rd38; ld.const.u32 %r344, [params+136]; mad.lo.s32 %r345, %r344, %r4, %r3; cvt.u64.u32 %rd11, %r345; min.f32 %f1237, %f237, %f1235; mov.f32 %f1238, 0f00000000; max.f32 %f261, %f1238, %f1237; min.f32 %f1239, %f249, %f1235; max.f32 %f262, %f1238, %f1239; min.f32 %f1240, %f1236, %f1235; max.f32 %f263, %f1238, %f1240; mov.f32 %f1241, 0f3E555555; cvt.rzi.f32.f32 %f1242, %f1241; add.f32 %f1243, %f1242, %f1242; mov.f32 %f1244, 0f3ED55555; sub.f32 %f1245, %f1244, %f1243; abs.f32 %f264, %f1245; abs.f32 %f265, %f261; setp.lt.f32 %p125, %f265, 0f00800000; mul.f32 %f1246, %f265, 0f4B800000; selp.f32 %f1247, %f1246, %f265, %p125; selp.f32 %f1248, 0fC3170000, 0fC2FE0000, %p125; mov.b32 %r346, %f1247; and.b32 %r347, %r346, 8388607; or.b32 %r348, %r347, 1065353216; mov.b32 %f1249, %r348; shr.u32 %r349, %r346, 23; cvt.rn.f32.u32 %f1250, %r349; add.f32 %f1251, %f1248, %f1250; setp.gt.f32 %p126, %f1249, 0f3FB504F3; mul.f32 %f1252, %f1249, 0f3F000000; add.f32 %f1253, %f1251, 0f3F800000; selp.f32 %f1254, %f1253, %f1251, %p126; selp.f32 %f1255, %f1252, %f1249, %p126; add.f32 %f1256, %f1255, 0fBF800000; add.f32 %f1257, %f1255, 0f3F800000; rcp.approx.ftz.f32 %f1258, %f1257; add.f32 %f1259, %f1256, %f1256; mul.f32 %f1260, %f1259, %f1258; mul.f32 %f1261, %f1260, %f1260; mov.f32 %f1262, 0f3C4CAF63; mov.f32 %f1263, 0f3B18F0FE; fma.rn.f32 %f1264, %f1263, %f1261, %f1262; mov.f32 %f1265, 0f3DAAAABD; fma.rn.f32 %f1266, %f1264, %f1261, %f1265; mul.rn.f32 %f1267, %f1266, %f1261; mul.rn.f32 %f1268, %f1267, %f1260; sub.f32 %f1269, %f1256, %f1260; add.f32 %f1270, %f1269, %f1269; neg.f32 %f1271, %f1260; fma.rn.f32 %f1272, %f1271, %f1256, %f1270; mul.rn.f32 %f1273, %f1258, %f1272; add.f32 %f1274, %f1268, %f1260; sub.f32 %f1275, %f1260, %f1274; add.f32 %f1276, %f1268, %f1275; add.f32 %f1277, %f1273, %f1276; add.f32 %f1278, %f1274, %f1277; sub.f32 %f1279, %f1274, %f1278; add.f32 %f1280, %f1277, %f1279; mov.f32 %f1281, 0f3F317200; mul.rn.f32 %f1282, %f1254, %f1281; mov.f32 %f1283, 0f35BFBE8E; mul.rn.f32 %f1284, %f1254, %f1283; add.f32 %f1285, %f1282, %f1278; sub.f32 %f1286, %f1282, %f1285; add.f32 %f1287, %f1278, %f1286; add.f32 %f1288, %f1280, %f1287; add.f32 %f1289, %f1284, %f1288; add.f32 %f1290, %f1285, %f1289; sub.f32 %f1291, %f1285, %f1290; add.f32 %f1292, %f1289, %f1291; mul.rn.f32 %f1293, %f1244, %f1290; neg.f32 %f1294, %f1293; fma.rn.f32 %f1295, %f1244, %f1290, %f1294; fma.rn.f32 %f1296, %f1244, %f1292, %f1295; fma.rn.f32 %f1297, %f1238, %f1290, %f1296; add.rn.f32 %f1298, %f1293, %f1297; neg.f32 %f1299, %f1298; add.rn.f32 %f1300, %f1293, %f1299; add.rn.f32 %f1301, %f1300, %f1297; mov.b32 %r350, %f1298; setp.eq.s32 %p127, %r350, 1118925336; add.s32 %r351, %r350, -1; mov.b32 %f1302, %r351; add.f32 %f1303, %f1301, 0f37000000; selp.f32 %f266, %f1303, %f1301, %p127; selp.f32 %f1304, %f1302, %f1298, %p127; mov.f32 %f1305, 0f3FB8AA3B; mul.rn.f32 %f1306, %f1304, %f1305; cvt.rzi.f32.f32 %f1307, %f1306; abs.f32 %f1308, %f1307; setp.gt.f32 %p128, %f1308, 0f42FC0000; mov.b32 %r352, %f1307; and.b32 %r353, %r352, -2147483648; or.b32 %r354, %r353, 1123811328; mov.b32 %f1309, %r354; selp.f32 %f1310, %f1309, %f1307, %p128; mov.f32 %f1311, 0fBF317218; fma.rn.f32 %f1312, %f1310, %f1311, %f1304; mov.f32 %f1313, 0f3102E308; fma.rn.f32 %f1314, %f1310, %f1313, %f1312; mul.f32 %f1315, %f1314, 0f3FB8AA3B; add.f32 %f1316, %f1310, 0f4B40007F; mov.b32 %r355, %f1316; shl.b32 %r356, %r355, 23; mov.b32 %f1317, %r356; ex2.approx.ftz.f32 %f1318, %f1315; mul.f32 %f267, %f1318, %f1317; setp.eq.f32 %p129, %f267, 0f7F800000; mov.f32 %f1688, 0f7F800000; @%p129 bra $L__BB0_106; fma.rn.f32 %f1688, %f267, %f266, %f267; $L__BB0_106: setp.lt.f32 %p130, %f261, 0f00000000; setp.eq.f32 %p131, %f264, 0f3F800000; and.pred %p7, %p130, %p131; setp.eq.f32 %p132, %f261, 0f00000000; @%p132 bra $L__BB0_110; bra.uni $L__BB0_107; $L__BB0_110: add.f32 %f1323, %f261, %f261; selp.f32 %f1690, %f1323, 0f00000000, %p131; bra.uni $L__BB0_111; $L__BB0_107: mov.b32 %r357, %f1688; xor.b32 %r358, %r357, -2147483648; mov.b32 %f1319, %r358; selp.f32 %f1690, %f1319, %f1688, %p7; setp.geu.f32 %p133, %f261, 0f00000000; @%p133 bra $L__BB0_111; mov.f32 %f1320, 0f3ED55555; cvt.rzi.f32.f32 %f1321, %f1320; setp.eq.f32 %p134, %f1321, 0f3ED55555; @%p134 bra $L__BB0_111; mov.f32 %f1690, 0f7FFFFFFF; $L__BB0_111: add.f32 %f1324, %f265, 0f3ED55555; mov.b32 %r359, %f1324; setp.lt.s32 %p136, %r359, 2139095040; @%p136 bra $L__BB0_116; setp.gtu.f32 %p137, %f265, 0f7F800000; @%p137 bra $L__BB0_115; bra.uni $L__BB0_113; $L__BB0_115: add.f32 %f1690, %f261, 0f3ED55555; bra.uni $L__BB0_116; $L__BB0_113: setp.neu.f32 %p138, %f265, 0f7F800000; @%p138 bra $L__BB0_116; selp.f32 %f1690, 0fFF800000, 0f7F800000, %p7; $L__BB0_116: abs.f32 %f276, %f262; setp.lt.f32 %p139, %f276, 0f00800000; mul.f32 %f1326, %f276, 0f4B800000; selp.f32 %f1327, %f1326, %f276, %p139; selp.f32 %f1328, 0fC3170000, 0fC2FE0000, %p139; mov.b32 %r360, %f1327; and.b32 %r361, %r360, 8388607; or.b32 %r362, %r361, 1065353216; mov.b32 %f1329, %r362; shr.u32 %r363, %r360, 23; cvt.rn.f32.u32 %f1330, %r363; add.f32 %f1331, %f1328, %f1330; setp.gt.f32 %p140, %f1329, 0f3FB504F3; mul.f32 %f1332, %f1329, 0f3F000000; add.f32 %f1333, %f1331, 0f3F800000; selp.f32 %f1334, %f1333, %f1331, %p140; selp.f32 %f1335, %f1332, %f1329, %p140; add.f32 %f1336, %f1335, 0fBF800000; add.f32 %f1337, %f1335, 0f3F800000; rcp.approx.ftz.f32 %f1338, %f1337; add.f32 %f1339, %f1336, %f1336; mul.f32 %f1340, %f1339, %f1338; mul.f32 %f1341, %f1340, %f1340; mov.f32 %f1342, 0f3C4CAF63; mov.f32 %f1343, 0f3B18F0FE; fma.rn.f32 %f1344, %f1343, %f1341, %f1342; mov.f32 %f1345, 0f3DAAAABD; fma.rn.f32 %f1346, %f1344, %f1341, %f1345; mul.rn.f32 %f1347, %f1346, %f1341; mul.rn.f32 %f1348, %f1347, %f1340; sub.f32 %f1349, %f1336, %f1340; add.f32 %f1350, %f1349, %f1349; neg.f32 %f1351, %f1340; fma.rn.f32 %f1352, %f1351, %f1336, %f1350; mul.rn.f32 %f1353, %f1338, %f1352; add.f32 %f1354, %f1348, %f1340; sub.f32 %f1355, %f1340, %f1354; add.f32 %f1356, %f1348, %f1355; add.f32 %f1357, %f1353, %f1356; add.f32 %f1358, %f1354, %f1357; sub.f32 %f1359, %f1354, %f1358; add.f32 %f1360, %f1357, %f1359; mov.f32 %f1361, 0f3F317200; mul.rn.f32 %f1362, %f1334, %f1361; mov.f32 %f1363, 0f35BFBE8E; mul.rn.f32 %f1364, %f1334, %f1363; add.f32 %f1365, %f1362, %f1358; sub.f32 %f1366, %f1362, %f1365; add.f32 %f1367, %f1358, %f1366; add.f32 %f1368, %f1360, %f1367; add.f32 %f1369, %f1364, %f1368; add.f32 %f1370, %f1365, %f1369; sub.f32 %f1371, %f1365, %f1370; add.f32 %f1372, %f1369, %f1371; mov.f32 %f1373, 0f3ED55555; mul.rn.f32 %f1374, %f1373, %f1370; neg.f32 %f1375, %f1374; fma.rn.f32 %f1376, %f1373, %f1370, %f1375; fma.rn.f32 %f1377, %f1373, %f1372, %f1376; mov.f32 %f1378, 0f00000000; fma.rn.f32 %f1379, %f1378, %f1370, %f1377; add.rn.f32 %f1380, %f1374, %f1379; neg.f32 %f1381, %f1380; add.rn.f32 %f1382, %f1374, %f1381; add.rn.f32 %f1383, %f1382, %f1379; mov.b32 %r364, %f1380; setp.eq.s32 %p141, %r364, 1118925336; add.s32 %r365, %r364, -1; mov.b32 %f1384, %r365; add.f32 %f1385, %f1383, 0f37000000; selp.f32 %f277, %f1385, %f1383, %p141; selp.f32 %f1386, %f1384, %f1380, %p141; mov.f32 %f1387, 0f3FB8AA3B; mul.rn.f32 %f1388, %f1386, %f1387; cvt.rzi.f32.f32 %f1389, %f1388; abs.f32 %f1390, %f1389; setp.gt.f32 %p142, %f1390, 0f42FC0000; mov.b32 %r366, %f1389; and.b32 %r367, %r366, -2147483648; or.b32 %r368, %r367, 1123811328; mov.b32 %f1391, %r368; selp.f32 %f1392, %f1391, %f1389, %p142; mov.f32 %f1393, 0fBF317218; fma.rn.f32 %f1394, %f1392, %f1393, %f1386; mov.f32 %f1395, 0f3102E308; fma.rn.f32 %f1396, %f1392, %f1395, %f1394; mul.f32 %f1397, %f1396, 0f3FB8AA3B; add.f32 %f1398, %f1392, 0f4B40007F; mov.b32 %r369, %f1398; shl.b32 %r370, %r369, 23; mov.b32 %f1399, %r370; ex2.approx.ftz.f32 %f1400, %f1397; mul.f32 %f278, %f1400, %f1399; setp.eq.f32 %p143, %f278, 0f7F800000; mov.f32 %f1691, 0f7F800000; @%p143 bra $L__BB0_118; fma.rn.f32 %f1691, %f278, %f277, %f278; $L__BB0_118: setp.lt.f32 %p144, %f262, 0f00000000; and.pred %p8, %p144, %p131; setp.eq.f32 %p146, %f262, 0f00000000; @%p146 bra $L__BB0_122; bra.uni $L__BB0_119; $L__BB0_122: add.f32 %f1405, %f262, %f262; selp.f32 %f1693, %f1405, 0f00000000, %p131; bra.uni $L__BB0_123; $L__BB0_119: mov.b32 %r371, %f1691; xor.b32 %r372, %r371, -2147483648; mov.b32 %f1401, %r372; selp.f32 %f1693, %f1401, %f1691, %p8; setp.geu.f32 %p147, %f262, 0f00000000; @%p147 bra $L__BB0_123; mov.f32 %f1402, 0f3ED55555; cvt.rzi.f32.f32 %f1403, %f1402; setp.eq.f32 %p148, %f1403, 0f3ED55555; @%p148 bra $L__BB0_123; mov.f32 %f1693, 0f7FFFFFFF; $L__BB0_123: add.f32 %f1406, %f276, 0f3ED55555; mov.b32 %r373, %f1406; setp.lt.s32 %p150, %r373, 2139095040; @%p150 bra $L__BB0_128; setp.gtu.f32 %p151, %f276, 0f7F800000; @%p151 bra $L__BB0_127; bra.uni $L__BB0_125; $L__BB0_127: add.f32 %f1693, %f262, 0f3ED55555; bra.uni $L__BB0_128; $L__BB0_125: setp.neu.f32 %p152, %f276, 0f7F800000; @%p152 bra $L__BB0_128; selp.f32 %f1693, 0fFF800000, 0f7F800000, %p8; $L__BB0_128: abs.f32 %f287, %f263; setp.lt.f32 %p153, %f287, 0f00800000; mul.f32 %f1408, %f287, 0f4B800000; selp.f32 %f1409, %f1408, %f287, %p153; selp.f32 %f1410, 0fC3170000, 0fC2FE0000, %p153; mov.b32 %r374, %f1409; and.b32 %r375, %r374, 8388607; or.b32 %r376, %r375, 1065353216; mov.b32 %f1411, %r376; shr.u32 %r377, %r374, 23; cvt.rn.f32.u32 %f1412, %r377; add.f32 %f1413, %f1410, %f1412; setp.gt.f32 %p154, %f1411, 0f3FB504F3; mul.f32 %f1414, %f1411, 0f3F000000; add.f32 %f1415, %f1413, 0f3F800000; selp.f32 %f1416, %f1415, %f1413, %p154; selp.f32 %f1417, %f1414, %f1411, %p154; add.f32 %f1418, %f1417, 0fBF800000; add.f32 %f1419, %f1417, 0f3F800000; rcp.approx.ftz.f32 %f1420, %f1419; add.f32 %f1421, %f1418, %f1418; mul.f32 %f1422, %f1421, %f1420; mul.f32 %f1423, %f1422, %f1422; mov.f32 %f1424, 0f3C4CAF63; mov.f32 %f1425, 0f3B18F0FE; fma.rn.f32 %f1426, %f1425, %f1423, %f1424; mov.f32 %f1427, 0f3DAAAABD; fma.rn.f32 %f1428, %f1426, %f1423, %f1427; mul.rn.f32 %f1429, %f1428, %f1423; mul.rn.f32 %f1430, %f1429, %f1422; sub.f32 %f1431, %f1418, %f1422; add.f32 %f1432, %f1431, %f1431; neg.f32 %f1433, %f1422; fma.rn.f32 %f1434, %f1433, %f1418, %f1432; mul.rn.f32 %f1435, %f1420, %f1434; add.f32 %f1436, %f1430, %f1422; sub.f32 %f1437, %f1422, %f1436; add.f32 %f1438, %f1430, %f1437; add.f32 %f1439, %f1435, %f1438; add.f32 %f1440, %f1436, %f1439; sub.f32 %f1441, %f1436, %f1440; add.f32 %f1442, %f1439, %f1441; mov.f32 %f1443, 0f3F317200; mul.rn.f32 %f1444, %f1416, %f1443; mov.f32 %f1445, 0f35BFBE8E; mul.rn.f32 %f1446, %f1416, %f1445; add.f32 %f1447, %f1444, %f1440; sub.f32 %f1448, %f1444, %f1447; add.f32 %f1449, %f1440, %f1448; add.f32 %f1450, %f1442, %f1449; add.f32 %f1451, %f1446, %f1450; add.f32 %f1452, %f1447, %f1451; sub.f32 %f1453, %f1447, %f1452; add.f32 %f1454, %f1451, %f1453; mov.f32 %f1455, 0f3ED55555; mul.rn.f32 %f1456, %f1455, %f1452; neg.f32 %f1457, %f1456; fma.rn.f32 %f1458, %f1455, %f1452, %f1457; fma.rn.f32 %f1459, %f1455, %f1454, %f1458; mov.f32 %f1460, 0f00000000; fma.rn.f32 %f1461, %f1460, %f1452, %f1459; add.rn.f32 %f1462, %f1456, %f1461; neg.f32 %f1463, %f1462; add.rn.f32 %f1464, %f1456, %f1463; add.rn.f32 %f1465, %f1464, %f1461; mov.b32 %r378, %f1462; setp.eq.s32 %p155, %r378, 1118925336; add.s32 %r379, %r378, -1; mov.b32 %f1466, %r379; add.f32 %f1467, %f1465, 0f37000000; selp.f32 %f288, %f1467, %f1465, %p155; selp.f32 %f1468, %f1466, %f1462, %p155; mov.f32 %f1469, 0f3FB8AA3B; mul.rn.f32 %f1470, %f1468, %f1469; cvt.rzi.f32.f32 %f1471, %f1470; abs.f32 %f1472, %f1471; setp.gt.f32 %p156, %f1472, 0f42FC0000; mov.b32 %r380, %f1471; and.b32 %r381, %r380, -2147483648; or.b32 %r382, %r381, 1123811328; mov.b32 %f1473, %r382; selp.f32 %f1474, %f1473, %f1471, %p156; mov.f32 %f1475, 0fBF317218; fma.rn.f32 %f1476, %f1474, %f1475, %f1468; mov.f32 %f1477, 0f3102E308; fma.rn.f32 %f1478, %f1474, %f1477, %f1476; mul.f32 %f1479, %f1478, 0f3FB8AA3B; add.f32 %f1480, %f1474, 0f4B40007F; mov.b32 %r383, %f1480; shl.b32 %r384, %r383, 23; mov.b32 %f1481, %r384; ex2.approx.ftz.f32 %f1482, %f1479; mul.f32 %f289, %f1482, %f1481; setp.eq.f32 %p157, %f289, 0f7F800000; mov.f32 %f1694, 0f7F800000; @%p157 bra $L__BB0_130; fma.rn.f32 %f1694, %f289, %f288, %f289; $L__BB0_130: setp.lt.f32 %p158, %f263, 0f00000000; and.pred %p9, %p158, %p131; setp.eq.f32 %p160, %f263, 0f00000000; @%p160 bra $L__BB0_134; bra.uni $L__BB0_131; $L__BB0_134: add.f32 %f1487, %f263, %f263; selp.f32 %f1696, %f1487, 0f00000000, %p131; bra.uni $L__BB0_135; $L__BB0_131: mov.b32 %r385, %f1694; xor.b32 %r386, %r385, -2147483648; mov.b32 %f1483, %r386; selp.f32 %f1696, %f1483, %f1694, %p9; setp.geu.f32 %p161, %f263, 0f00000000; @%p161 bra $L__BB0_135; mov.f32 %f1484, 0f3ED55555; cvt.rzi.f32.f32 %f1485, %f1484; setp.eq.f32 %p162, %f1485, 0f3ED55555; @%p162 bra $L__BB0_135; mov.f32 %f1696, 0f7FFFFFFF; $L__BB0_135: add.f32 %f1488, %f287, 0f3ED55555; mov.b32 %r387, %f1488; setp.lt.s32 %p164, %r387, 2139095040; @%p164 bra $L__BB0_140; setp.gtu.f32 %p165, %f287, 0f7F800000; @%p165 bra $L__BB0_139; bra.uni $L__BB0_137; $L__BB0_139: add.f32 %f1696, %f263, 0f3ED55555; bra.uni $L__BB0_140; $L__BB0_137: setp.neu.f32 %p166, %f287, 0f7F800000; @%p166 bra $L__BB0_140; selp.f32 %f1696, 0fFF800000, 0f7F800000, %p9; $L__BB0_140: fma.rn.f32 %f1489, %f1690, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p167, %f261, 0f3F800000; mov.f32 %f1490, 0f3F800000; selp.f32 %f1491, 0f3F7FFFFF, %f1489, %p167; mul.f32 %f1492, %f261, 0f414EB852; setp.lt.f32 %p168, %f261, 0f3B4D2E1C; selp.f32 %f1493, %f1492, %f1491, %p168; fma.rn.f32 %f1494, %f1693, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p169, %f262, 0f3F800000; selp.f32 %f1495, 0f3F7FFFFF, %f1494, %p169; mul.f32 %f1496, %f262, 0f414EB852; setp.lt.f32 %p170, %f262, 0f3B4D2E1C; selp.f32 %f1497, %f1496, %f1495, %p170; fma.rn.f32 %f1498, %f1696, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p171, %f263, 0f3F800000; selp.f32 %f1499, 0f3F7FFFFF, %f1498, %p171; mul.f32 %f1500, %f263, 0f414EB852; setp.lt.f32 %p172, %f263, 0f3B4D2E1C; selp.f32 %f1501, %f1500, %f1499, %p172; min.f32 %f1502, %f1493, %f1490; mov.f32 %f1503, 0f00000000; max.f32 %f1504, %f1503, %f1502; mul.f32 %f1505, %f1504, 0f43800000; cvt.rzi.u32.f32 %r388, %f1505; min.u32 %r389, %r388, 255; min.f32 %f1506, %f1497, %f1490; max.f32 %f1507, %f1503, %f1506; mul.f32 %f1508, %f1507, 0f43800000; cvt.rzi.u32.f32 %r390, %f1508; min.u32 %r391, %r390, 255; min.f32 %f1509, %f1501, %f1490; max.f32 %f1510, %f1503, %f1509; mul.f32 %f1511, %f1510, 0f43800000; cvt.rzi.u32.f32 %r392, %f1511; min.u32 %r393, %r392, 255; shl.b64 %rd39, %rd11, 2; add.s64 %rd40, %rd10, %rd39; cvt.u16.u32 %rs49, %r393; cvt.u16.u32 %rs50, %r391; cvt.u16.u32 %rs51, %r389; mov.u16 %rs52, 255; st.global.v4.u8 [%rd40], {%rs51, %rs50, %rs49, %rs52}; $L__BB0_141: and.b32 %r394, %r425, 4; setp.eq.s32 %p173, %r394, 0; ld.const.u32 %r426, [params+108]; @%p173 bra $L__BB0_145; setp.eq.s32 %p174, %r426, 0; ld.const.u64 %rd41, [params+224]; cvta.to.global.u64 %rd42, %rd41; ld.const.u32 %r395, [params+216]; mad.lo.s32 %r396, %r395, %r4, %r3; mul.wide.u32 %rd43, %r396, 8; add.s64 %rd12, %rd42, %rd43; @%p174 bra $L__BB0_144; ld.global.v4.u16 {%rs60, %rs61, %rs62, %rs63}, [%rd12]; // begin inline asm { cvt.f32.f16 %f1512, %rs60;} // end inline asm // begin inline asm { cvt.f32.f16 %f1513, %rs61;} // end inline asm // begin inline asm { cvt.f32.f16 %f1514, %rs62;} // end inline asm add.f32 %f1515, %f1650, %f1512; add.f32 %f1516, %f1651, %f1513; add.f32 %f1517, %f1652, %f1514; mov.f32 %f1518, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs58, %f1517;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs57, %f1516;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs56, %f1515;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs59, %f1518;} // end inline asm st.global.v4.u16 [%rd12], {%rs56, %rs57, %rs58, %rs59}; bra.uni $L__BB0_145; $L__BB0_144: mov.f32 %f1522, 0f3F800000; mov.u32 %r426, 0; // begin inline asm { cvt.rn.f16.f32 %rs67, %f1522;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs66, %f1652;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs65, %f1651;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs64, %f1650;} // end inline asm st.global.v4.u16 [%rd12], {%rs64, %rs65, %rs66, %rs67}; $L__BB0_145: ld.const.u64 %rd44, [params+256]; cvta.to.global.u64 %rd45, %rd44; ld.const.u32 %r398, [params+248]; mad.lo.s32 %r399, %r398, %r4, %r3; mul.wide.u32 %rd46, %r399, 8; add.s64 %rd13, %rd45, %rd46; setp.eq.s32 %p175, %r426, 0; @%p175 bra $L__BB0_147; ld.global.v4.u16 {%rs75, %rs76, %rs77, %rs78}, [%rd13]; // begin inline asm { cvt.f32.f16 %f1523, %rs75;} // end inline asm // begin inline asm { cvt.f32.f16 %f1524, %rs76;} // end inline asm // begin inline asm { cvt.f32.f16 %f1525, %rs77;} // end inline asm add.f32 %f1526, %f1653, %f1523; add.f32 %f1527, %f1654, %f1524; add.f32 %f1528, %f1655, %f1525; mov.f32 %f1529, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs73, %f1528;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs72, %f1527;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs71, %f1526;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs74, %f1529;} // end inline asm st.global.v4.u16 [%rd13], {%rs71, %rs72, %rs73, %rs74}; bra.uni $L__BB0_148; $L__BB0_147: mov.f32 %f1533, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs82, %f1533;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs81, %f1655;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs80, %f1654;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs79, %f1653;} // end inline asm st.global.v4.u16 [%rd13], {%rs79, %rs80, %rs81, %rs82}; $L__BB0_148: ld.const.u64 %rd47, [params+272]; cvta.to.global.u64 %rd48, %rd47; ld.const.u32 %r400, [params+264]; mad.lo.s32 %r401, %r400, %r4, %r3; mul.wide.u32 %rd49, %r401, 8; add.s64 %rd14, %rd48, %rd49; @%p175 bra $L__BB0_150; ld.global.v4.u16 {%rs90, %rs91, %rs92, %rs93}, [%rd14]; // begin inline asm { cvt.f32.f16 %f1534, %rs90;} // end inline asm // begin inline asm { cvt.f32.f16 %f1535, %rs91;} // end inline asm // begin inline asm { cvt.f32.f16 %f1536, %rs92;} // end inline asm add.f32 %f1537, %f1656, %f1534; add.f32 %f1538, %f1657, %f1535; add.f32 %f1539, %f1658, %f1536; mov.f32 %f1540, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs88, %f1539;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs87, %f1538;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs86, %f1537;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs89, %f1540;} // end inline asm st.global.v4.u16 [%rd14], {%rs86, %rs87, %rs88, %rs89}; bra.uni $L__BB0_151; $L__BB0_150: mov.f32 %f1544, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs97, %f1544;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs96, %f1658;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs95, %f1657;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs94, %f1656;} // end inline asm st.global.v4.u16 [%rd14], {%rs94, %rs95, %rs96, %rs97}; $L__BB0_151: ld.const.u64 %rd50, [params+288]; cvta.to.global.u64 %rd51, %rd50; ld.const.u32 %r402, [params+280]; mad.lo.s32 %r403, %r402, %r4, %r3; mul.wide.u32 %rd52, %r403, 8; add.s64 %rd15, %rd51, %rd52; @%p175 bra $L__BB0_153; ld.global.v4.u16 {%rs105, %rs106, %rs107, %rs108}, [%rd15]; // begin inline asm { cvt.f32.f16 %f1545, %rs105;} // end inline asm // begin inline asm { cvt.f32.f16 %f1546, %rs106;} // end inline asm // begin inline asm { cvt.f32.f16 %f1547, %rs107;} // end inline asm add.f32 %f1548, %f1659, %f1545; add.f32 %f1549, %f1660, %f1546; add.f32 %f1550, %f1661, %f1547; mov.f32 %f1551, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs103, %f1550;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs102, %f1549;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs101, %f1548;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs104, %f1551;} // end inline asm st.global.v4.u16 [%rd15], {%rs101, %rs102, %rs103, %rs104}; bra.uni $L__BB0_171; $L__BB0_153: mov.f32 %f1555, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs112, %f1555;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs111, %f1661;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs110, %f1660;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs109, %f1659;} // end inline asm st.global.v4.u16 [%rd15], {%rs109, %rs110, %rs111, %rs112}; $L__BB0_171: ret; }