// // 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<1740>; .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 %f1629, 0f00000000; mov.u16 %rs177, 0; mov.f32 %f1630, %f1629; mov.f32 %f1631, %f1629; @%p11 bra $L__BB0_4; $L__BB0_3: cvt.rn.f32.u16 %f317, %rs18; div.rn.f32 %f318, %f317, 0f437F0000; fma.rn.f32 %f319, %f318, 0f40000000, 0fBF800000; and.b16 %rs24, %rs177, 255; cvt.rn.f32.u16 %f320, %rs24; div.rn.f32 %f321, %f320, 0f437F0000; fma.rn.f32 %f322, %f321, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f323, %rs176; div.rn.f32 %f324, %f323, 0f437F0000; fma.rn.f32 %f325, %f324, 0f40000000, 0fBF800000; mul.f32 %f326, %f322, %f322; fma.rn.f32 %f327, %f319, %f319, %f326; fma.rn.f32 %f328, %f325, %f325, %f327; sqrt.rn.f32 %f329, %f328; rcp.rn.f32 %f330, %f329; mul.f32 %f1631, %f330, %f325; mul.f32 %f1630, %f330, %f322; mul.f32 %f1629, %f319, %f330; $L__BB0_4: ld.const.v2.u32 {%r34, %r35}, [params]; add.s32 %r3, %r34, %r26; add.s32 %r4, %r35, %r27; setp.eq.f32 %p12, %f1629, 0f00000000; setp.eq.f32 %p13, %f1630, 0f00000000; and.pred %p14, %p12, %p13; setp.eq.f32 %p15, %f1631, 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 %f1589, 0f00000000; cvt.rzi.u32.f32 %r410, %f1589; 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 %f1590, %rs122;} // end inline asm // begin inline asm { cvt.f32.f16 %f1591, %rs123;} // end inline asm // begin inline asm { cvt.f32.f16 %f1592, %rs124;} // end inline asm add.f32 %f1593, %f1590, 0f00000000; add.f32 %f1594, %f1591, 0f00000000; add.f32 %f1595, %f1592, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs121, %f1595;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs120, %f1594;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs119, %f1593;} // 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 %f331, %f10, 0f3456BF95; ld.global.f32 %f11, [%rd26+4]; mul.f32 %f332, %f11, 0f3456BF95; ld.global.f32 %f12, [%rd26+8]; mul.f32 %f333, %f12, 0f3456BF95; abs.f32 %f334, %f1629; div.rn.f32 %f335, %f331, %f334; abs.f32 %f336, %f1630; div.rn.f32 %f337, %f332, %f336; abs.f32 %f338, %f1631; div.rn.f32 %f339, %f333, %f338; abs.f32 %f340, %f335; abs.f32 %f341, %f337; abs.f32 %f342, %f339; mov.f32 %f343, 0f38D1B717; max.f32 %f344, %f340, %f343; max.f32 %f345, %f341, %f343; max.f32 %f346, %f342, %f343; fma.rn.f32 %f13, %f1629, %f344, %f10; fma.rn.f32 %f14, %f1630, %f345, %f11; fma.rn.f32 %f15, %f1631, %f346, %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 %f1632, 0f00000000; mov.u16 %rs180, 0; mov.u16 %rs181, %rs180; mov.f32 %f1633, %f1632; mov.f32 %f1634, %f1632; @%p18 bra $L__BB0_9; $L__BB0_8: mov.u16 %rs181, %rs180; cvt.rn.f32.u16 %f350, %rs26; div.rn.f32 %f351, %f350, 0f437F0000; fma.rn.f32 %f352, %f351, 0f40000000, 0fBF800000; and.b16 %rs34, %rs181, 255; cvt.rn.f32.u16 %f353, %rs34; div.rn.f32 %f354, %f353, 0f437F0000; fma.rn.f32 %f355, %f354, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f356, %rs178; div.rn.f32 %f357, %f356, 0f437F0000; fma.rn.f32 %f358, %f357, 0f40000000, 0fBF800000; mul.f32 %f359, %f355, %f355; fma.rn.f32 %f360, %f352, %f352, %f359; fma.rn.f32 %f361, %f358, %f358, %f360; sqrt.rn.f32 %f362, %f361; rcp.rn.f32 %f363, %f362; mul.f32 %f1634, %f363, %f358; mul.f32 %f1633, %f363, %f355; mul.f32 %f1632, %f352, %f363; mov.u16 %rs180, %rs178; $L__BB0_9: mul.f32 %f367, %f1630, %f1634; mul.f32 %f368, %f1631, %f1633; sub.f32 %f369, %f367, %f368; mul.f32 %f370, %f1631, %f1632; mul.f32 %f371, %f1629, %f1634; sub.f32 %f372, %f370, %f371; mul.f32 %f373, %f1629, %f1633; mul.f32 %f374, %f1630, %f1632; sub.f32 %f375, %f373, %f374; ld.global.u8 %rs36, [%rd2+3]; setp.eq.s16 %p19, %rs36, 0; selp.f32 %f24, 0fBF800000, 0f3F800000, %p19; mul.f32 %f376, %f369, %f24; mul.f32 %f377, %f372, %f24; mul.f32 %f378, %f375, %f24; mul.f32 %f379, %f376, 0f00000000; mul.f32 %f380, %f377, 0f00000000; mul.f32 %f381, %f378, 0f00000000; fma.rn.f32 %f382, %f1632, 0f3F5105EC, %f379; fma.rn.f32 %f383, %f1633, 0f3F5105EC, %f380; fma.rn.f32 %f384, %f1634, 0f3F5105EC, %f381; mul.f32 %f25, %f1629, 0f3F13CD3A; add.f32 %f26, %f25, %f382; mul.f32 %f27, %f1630, 0f3F13CD3A; add.f32 %f28, %f27, %f383; mul.f32 %f29, %f1631, 0f3F13CD3A; add.f32 %f30, %f29, %f384; or.b16 %rs37, %rs26, %rs181; or.b16 %rs15, %rs37, %rs180; and.b16 %rs38, %rs15, 255; setp.eq.s16 %p20, %rs38, 0; mov.f32 %f1638, 0f00000000; mov.f32 %f1635, %f1638; mov.f32 %f1636, %f1638; mov.f32 %f1637, %f1638; @%p20 bra $L__BB0_11; cvt.rn.f32.u16 %f385, %rs26; div.rn.f32 %f386, %f385, 0f437F0000; fma.rn.f32 %f387, %f386, 0f40000000, 0fBF800000; and.b16 %rs40, %rs181, 255; cvt.rn.f32.u16 %f388, %rs40; div.rn.f32 %f389, %f388, 0f437F0000; fma.rn.f32 %f390, %f389, 0f40000000, 0fBF800000; and.b16 %rs41, %rs180, 255; cvt.rn.f32.u16 %f391, %rs41; div.rn.f32 %f392, %f391, 0f437F0000; fma.rn.f32 %f393, %f392, 0f40000000, 0fBF800000; mul.f32 %f394, %f390, %f390; fma.rn.f32 %f395, %f387, %f387, %f394; fma.rn.f32 %f396, %f393, %f393, %f395; sqrt.rn.f32 %f397, %f396; rcp.rn.f32 %f398, %f397; mul.f32 %f1637, %f398, %f393; mul.f32 %f1636, %f398, %f390; mul.f32 %f1635, %f387, %f398; $L__BB0_11: mul.f32 %f402, %f1630, %f1637; mul.f32 %f403, %f1631, %f1636; sub.f32 %f404, %f402, %f403; mul.f32 %f405, %f1631, %f1635; mul.f32 %f406, %f1629, %f1637; sub.f32 %f407, %f405, %f406; mul.f32 %f408, %f1629, %f1636; mul.f32 %f409, %f1630, %f1635; sub.f32 %f410, %f408, %f409; mul.f32 %f411, %f404, %f24; mul.f32 %f412, %f407, %f24; mul.f32 %f413, %f410, %f24; mul.f32 %f414, %f1635, 0f3ED105EC; mul.f32 %f415, %f1636, 0f3ED105EC; mul.f32 %f416, %f1637, 0f3ED105EC; mul.f32 %f417, %f411, 0f3F3504F3; mul.f32 %f418, %f412, 0f3F3504F3; mul.f32 %f419, %f413, 0f3F3504F3; sub.f32 %f420, %f417, %f414; sub.f32 %f421, %f418, %f415; sub.f32 %f422, %f419, %f416; add.f32 %f37, %f25, %f420; add.f32 %f38, %f27, %f421; add.f32 %f39, %f29, %f422; mov.f32 %f1639, %f1638; mov.f32 %f1640, %f1638; @%p20 bra $L__BB0_13; cvt.rn.f32.u16 %f423, %rs26; div.rn.f32 %f424, %f423, 0f437F0000; fma.rn.f32 %f425, %f424, 0f40000000, 0fBF800000; and.b16 %rs44, %rs181, 255; cvt.rn.f32.u16 %f426, %rs44; div.rn.f32 %f427, %f426, 0f437F0000; fma.rn.f32 %f428, %f427, 0f40000000, 0fBF800000; and.b16 %rs45, %rs180, 255; cvt.rn.f32.u16 %f429, %rs45; div.rn.f32 %f430, %f429, 0f437F0000; fma.rn.f32 %f431, %f430, 0f40000000, 0fBF800000; mul.f32 %f432, %f428, %f428; fma.rn.f32 %f433, %f425, %f425, %f432; fma.rn.f32 %f434, %f431, %f431, %f433; sqrt.rn.f32 %f435, %f434; rcp.rn.f32 %f436, %f435; mul.f32 %f1640, %f436, %f431; mul.f32 %f1639, %f436, %f428; mul.f32 %f1638, %f425, %f436; $L__BB0_13: mul.f32 %f437, %f1630, %f1640; mul.f32 %f438, %f1631, %f1639; sub.f32 %f439, %f437, %f438; mul.f32 %f440, %f1631, %f1638; mul.f32 %f441, %f1629, %f1640; sub.f32 %f442, %f440, %f441; mul.f32 %f443, %f1629, %f1639; mul.f32 %f444, %f1630, %f1638; sub.f32 %f445, %f443, %f444; mul.f32 %f446, %f439, %f24; mul.f32 %f447, %f442, %f24; mul.f32 %f448, %f445, %f24; mul.f32 %f449, %f1638, 0f3ED105EC; mul.f32 %f450, %f1639, 0f3ED105EC; mul.f32 %f451, %f1640, 0f3ED105EC; mul.f32 %f452, %f446, 0fBF3504F3; mul.f32 %f453, %f447, 0fBF3504F3; mul.f32 %f454, %f448, 0fBF3504F3; sub.f32 %f455, %f452, %f449; sub.f32 %f456, %f453, %f450; sub.f32 %f457, %f454, %f451; add.f32 %f46, %f25, %f455; add.f32 %f47, %f27, %f456; add.f32 %f48, %f29, %f457; 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 %f471, 0f40000000; cvt.rzi.f32.f32 %f472, %f471; add.f32 %f473, %f472, %f472; mov.f32 %f474, 0f40800000; sub.f32 %f475, %f474, %f473; abs.f32 %f49, %f475; 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 %f470, 0f00000000; mov.u32 %r421, 0; ld.const.u64 %rd4, [params+96]; abs.f32 %f714, %f50; abs.f32 %f715, %f51; max.f32 %f716, %f714, %f715; abs.f32 %f717, %f52; max.f32 %f718, %f716, %f717; mov.f32 %f1705, %f470; mov.f32 %f1704, %f470; mov.f32 %f1703, %f470; mov.f32 %f1702, %f470; mov.f32 %f1701, %f470; mov.f32 %f1700, %f470; mov.f32 %f1699, %f470; mov.f32 %f1698, %f470; mov.f32 %f1697, %f470; mov.f32 %f1696, %f470; mov.f32 %f1695, %f470; mov.f32 %f1694, %f470; mov.f32 %f1693, %f470; $L__BB0_15: mul.wide.u32 %rd31, %r421, 112; add.s64 %rd6, %rd3, %rd31; ld.global.v4.f32 {%f477, %f478, %f479, %f480}, [%rd6]; ld.global.v4.f32 {%f481, %f482, %f483, %f484}, [%rd6+16]; ld.global.v4.f32 {%f487, %f1662, %f1663, %f490}, [%rd6+32]; ld.global.v4.f32 {%f491, %f492, %f493, %f1660}, [%rd6+48]; ld.global.v4.f32 {%f495, %f496, %f497, %f498}, [%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 %f499, %f478, %f10; sub.f32 %f500, %f479, %f11; sub.f32 %f501, %f480, %f12; mul.f32 %f502, %f500, %f500; fma.rn.f32 %f503, %f499, %f499, %f502; fma.rn.f32 %f504, %f501, %f501, %f503; sqrt.rn.f32 %f505, %f504; rcp.rn.f32 %f506, %f505; mul.f32 %f87, %f499, %f506; mul.f32 %f88, %f500, %f506; mul.f32 %f89, %f501, %f506; mul.f32 %f90, %f505, %f482; mul.f32 %f91, %f505, %f484; abs.f32 %f92, %f91; setp.lt.f32 %p23, %f92, 0f00800000; mul.f32 %f507, %f92, 0f4B800000; selp.f32 %f508, %f507, %f92, %p23; selp.f32 %f509, 0fC3170000, 0fC2FE0000, %p23; mov.b32 %r50, %f508; and.b32 %r51, %r50, 8388607; or.b32 %r52, %r51, 1065353216; mov.b32 %f510, %r52; shr.u32 %r53, %r50, 23; cvt.rn.f32.u32 %f511, %r53; add.f32 %f512, %f509, %f511; setp.gt.f32 %p24, %f510, 0f3FB504F3; mul.f32 %f513, %f510, 0f3F000000; add.f32 %f514, %f512, 0f3F800000; selp.f32 %f515, %f514, %f512, %p24; selp.f32 %f516, %f513, %f510, %p24; add.f32 %f517, %f516, 0fBF800000; add.f32 %f518, %f516, 0f3F800000; rcp.approx.ftz.f32 %f519, %f518; add.f32 %f520, %f517, %f517; mul.f32 %f521, %f520, %f519; mul.f32 %f522, %f521, %f521; mov.f32 %f523, 0f3C4CAF63; mov.f32 %f524, 0f3B18F0FE; fma.rn.f32 %f525, %f524, %f522, %f523; mov.f32 %f526, 0f3DAAAABD; fma.rn.f32 %f527, %f525, %f522, %f526; mul.rn.f32 %f528, %f527, %f522; mul.rn.f32 %f529, %f528, %f521; sub.f32 %f530, %f517, %f521; add.f32 %f531, %f530, %f530; neg.f32 %f532, %f521; fma.rn.f32 %f533, %f532, %f517, %f531; mul.rn.f32 %f534, %f519, %f533; add.f32 %f535, %f529, %f521; sub.f32 %f536, %f521, %f535; add.f32 %f537, %f529, %f536; add.f32 %f538, %f534, %f537; add.f32 %f539, %f535, %f538; sub.f32 %f540, %f535, %f539; add.f32 %f541, %f538, %f540; mov.f32 %f542, 0f3F317200; mul.rn.f32 %f543, %f515, %f542; mov.f32 %f544, 0f35BFBE8E; mul.rn.f32 %f545, %f515, %f544; add.f32 %f546, %f543, %f539; sub.f32 %f547, %f543, %f546; add.f32 %f548, %f539, %f547; add.f32 %f549, %f541, %f548; add.f32 %f550, %f545, %f549; add.f32 %f551, %f546, %f550; sub.f32 %f552, %f546, %f551; add.f32 %f553, %f550, %f552; mul.rn.f32 %f555, %f474, %f551; neg.f32 %f556, %f555; fma.rn.f32 %f557, %f474, %f551, %f556; fma.rn.f32 %f558, %f474, %f553, %f557; fma.rn.f32 %f560, %f470, %f551, %f558; add.rn.f32 %f561, %f555, %f560; neg.f32 %f562, %f561; add.rn.f32 %f563, %f555, %f562; add.rn.f32 %f564, %f563, %f560; mov.b32 %r54, %f561; setp.eq.s32 %p25, %r54, 1118925336; add.s32 %r55, %r54, -1; mov.b32 %f565, %r55; add.f32 %f566, %f564, 0f37000000; selp.f32 %f93, %f566, %f564, %p25; selp.f32 %f567, %f565, %f561, %p25; mov.f32 %f568, 0f3FB8AA3B; mul.rn.f32 %f569, %f567, %f568; cvt.rzi.f32.f32 %f570, %f569; abs.f32 %f571, %f570; setp.gt.f32 %p26, %f571, 0f42FC0000; mov.b32 %r56, %f570; and.b32 %r57, %r56, -2147483648; or.b32 %r58, %r57, 1123811328; mov.b32 %f572, %r58; selp.f32 %f573, %f572, %f570, %p26; mov.f32 %f574, 0fBF317218; fma.rn.f32 %f575, %f573, %f574, %f567; mov.f32 %f576, 0f3102E308; fma.rn.f32 %f577, %f573, %f576, %f575; mul.f32 %f578, %f577, 0f3FB8AA3B; add.f32 %f579, %f573, 0f4B40007F; mov.b32 %r59, %f579; shl.b32 %r60, %r59, 23; mov.b32 %f580, %r60; ex2.approx.ftz.f32 %f581, %f578; mul.f32 %f94, %f581, %f580; setp.eq.f32 %p27, %f94, 0f7F800000; mov.f32 %f1654, 0f7F800000; @%p27 bra $L__BB0_17; fma.rn.f32 %f1654, %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 %f586, %f91, %f91; selp.f32 %f1656, %f586, 0f00000000, %p29; bra.uni $L__BB0_22; $L__BB0_18: mov.b32 %r61, %f1654; xor.b32 %r62, %r61, -2147483648; mov.b32 %f582, %r62; selp.f32 %f1656, %f582, %f1654, %p1; setp.geu.f32 %p31, %f91, 0f00000000; @%p31 bra $L__BB0_22; mov.f32 %f583, 0f40800000; cvt.rzi.f32.f32 %f584, %f583; setp.eq.f32 %p32, %f584, 0f40800000; @%p32 bra $L__BB0_22; mov.f32 %f1656, 0f7FFFFFFF; $L__BB0_22: abs.f32 %f1626, %f91; add.f32 %f587, %f1626, 0f40800000; mov.b32 %r63, %f587; setp.lt.s32 %p34, %r63, 2139095040; @%p34 bra $L__BB0_27; abs.f32 %f1627, %f91; setp.gtu.f32 %p35, %f1627, 0f7F800000; @%p35 bra $L__BB0_26; bra.uni $L__BB0_24; $L__BB0_26: add.f32 %f1656, %f91, 0f40800000; bra.uni $L__BB0_27; $L__BB0_24: abs.f32 %f1628, %f91; setp.neu.f32 %p36, %f1628, 0f7F800000; @%p36 bra $L__BB0_27; selp.f32 %f1656, 0fFF800000, 0f7F800000, %p1; $L__BB0_27: mov.f32 %f588, 0f3F800000; sub.f32 %f589, %f588, %f1656; setp.eq.f32 %p37, %f91, 0f3F800000; selp.f32 %f590, 0f00000000, %f589, %p37; cvt.sat.f32.f32 %f591, %f590; fma.rn.f32 %f592, %f90, %f90, %f483; div.rn.f32 %f103, %f591, %f592; mul.f32 %f593, %f1630, %f88; fma.rn.f32 %f594, %f1629, %f87, %f593; fma.rn.f32 %f104, %f1631, %f89, %f594; setp.eq.f32 %p38, %f477, 0f3F800000; @%p38 bra $L__BB0_33; bra.uni $L__BB0_28; $L__BB0_33: setp.eq.s64 %p42, %rd7, 0; mov.f32 %f1661, %f487; mov.f32 %f1664, %f103; @%p42 bra $L__BB0_37; setp.geu.f32 %p43, %f1660, 0f00000000; mov.f32 %f1657, %f87; mov.f32 %f1658, %f88; mov.f32 %f1659, %f89; @%p43 bra $L__BB0_36; mul.f32 %f619, %f84, %f500; fma.rn.f32 %f620, %f83, %f499, %f619; fma.rn.f32 %f622, %f85, %f501, %f620; rcp.rn.f32 %f623, %f622; mul.f32 %f1657, %f499, %f623; mul.f32 %f1658, %f500, %f623; mul.f32 %f1659, %f501, %f623; neg.f32 %f1660, %f1660; $L__BB0_36: mul.f32 %f624, %f85, %f89; mul.f32 %f625, %f84, %f88; neg.f32 %f626, %f625; mul.f32 %f627, %f83, %f87; sub.f32 %f628, %f626, %f627; sub.f32 %f629, %f628, %f624; setp.gt.f32 %p44, %f629, 0f00000000; selp.f32 %f630, 0f3F800000, 0f00000000, %p44; mov.f32 %f631, 0f3F800000; mul.f32 %f632, %f492, %f1658; mul.f32 %f633, %f496, %f1658; fma.rn.f32 %f634, %f491, %f1657, %f632; fma.rn.f32 %f635, %f495, %f1657, %f633; fma.rn.f32 %f636, %f493, %f1659, %f634; fma.rn.f32 %f637, %f497, %f1659, %f635; fma.rn.f32 %f638, %f1660, %f636, 0f3F000000; sub.f32 %f639, %f631, %f638; fma.rn.f32 %f640, %f1660, %f637, 0f3F000000; tex.2d.v4.f32.f32 {%f641, %f642, %f643, %f644}, [%rd7, {%f639, %f640}]; mul.f32 %f645, %f630, %f641; mul.f32 %f646, %f630, %f642; mul.f32 %f647, %f630, %f643; mul.f32 %f1661, %f487, %f645; mul.f32 %f1662, %f1662, %f646; mul.f32 %f1663, %f1663, %f647; mov.f32 %f1664, %f103; bra.uni $L__BB0_37; $L__BB0_28: setp.eq.f32 %p39, %f477, 0f40000000; @%p39 bra $L__BB0_31; bra.uni $L__BB0_29; $L__BB0_31: setp.eq.s64 %p41, %rd7, 0; mov.f32 %f1661, %f487; mov.f32 %f1664, %f103; @%p41 bra $L__BB0_37; mul.f32 %f604, %f492, %f88; mul.f32 %f605, %f496, %f88; mul.f32 %f606, %f84, %f88; fma.rn.f32 %f607, %f491, %f87, %f604; fma.rn.f32 %f608, %f495, %f87, %f605; fma.rn.f32 %f609, %f83, %f87, %f606; fma.rn.f32 %f610, %f493, %f89, %f607; fma.rn.f32 %f611, %f497, %f89, %f608; fma.rn.f32 %f612, %f85, %f89, %f609; tex.cube.v4.f32.f32 {%f613, %f614, %f615, %f616}, [%rd7, {%f610, %f611, %f612, %f612}]; mul.f32 %f1661, %f487, %f613; mul.f32 %f1662, %f1662, %f614; mul.f32 %f1663, %f1663, %f615; mov.f32 %f1664, %f103; bra.uni $L__BB0_37; $L__BB0_29: setp.neu.f32 %p40, %f477, 0f40800000; mov.f32 %f1661, %f487; mov.f32 %f1664, %f103; @%p40 bra $L__BB0_37; mul.f32 %f595, %f83, %f87; mul.f32 %f596, %f84, %f88; neg.f32 %f597, %f596; sub.f32 %f598, %f597, %f595; mul.f32 %f599, %f85, %f89; sub.f32 %f600, %f598, %f599; fma.rn.f32 %f601, %f490, %f600, %f1660; cvt.sat.f32.f32 %f602, %f601; mul.f32 %f603, %f602, %f602; mul.f32 %f1664, %f103, %f603; mov.f32 %f1661, %f487; $L__BB0_37: max.f32 %f660, %f1661, %f1662; max.f32 %f661, %f660, %f1663; setp.eq.s32 %p46, %r6, 0; selp.f32 %f662, %f104, 0f3F800000, %p46; cvt.sat.f32.f32 %f663, %f662; mul.f32 %f137, %f663, %f1664; mul.f32 %f664, %f137, %f661; setp.lt.f32 %p47, %f664, 0f3727C5AC; mov.pred %p188, -1; mov.f32 %f1665, 0f00000000; mov.f32 %f1666, %f1665; mov.f32 %f1667, %f1665; mov.f32 %f1668, %f1665; mov.f32 %f1669, %f1665; mov.f32 %f1670, %f1665; mov.f32 %f1671, %f1665; mov.f32 %f1672, %f1665; mov.f32 %f1673, %f1665; mov.f32 %f1674, %f1665; mov.f32 %f1675, %f1665; mov.f32 %f1676, %f1665; @%p47 bra $L__BB0_39; mul.f32 %f1665, %f1661, %f137; mul.f32 %f665, %f28, %f88; fma.rn.f32 %f666, %f26, %f87, %f665; fma.rn.f32 %f667, %f30, %f89, %f666; cvt.sat.f32.f32 %f668, %f667; mul.f32 %f669, %f1661, %f1664; mul.f32 %f670, %f669, %f668; mul.f32 %f671, %f1662, %f1664; mul.f32 %f672, %f671, %f668; mul.f32 %f673, %f1663, %f1664; mul.f32 %f674, %f673, %f668; mul.f32 %f675, %f38, %f88; fma.rn.f32 %f676, %f37, %f87, %f675; fma.rn.f32 %f677, %f39, %f89, %f676; cvt.sat.f32.f32 %f678, %f677; mul.f32 %f679, %f669, %f678; mul.f32 %f680, %f671, %f678; mul.f32 %f681, %f673, %f678; mul.f32 %f682, %f47, %f88; fma.rn.f32 %f683, %f46, %f87, %f682; fma.rn.f32 %f684, %f48, %f89, %f683; cvt.sat.f32.f32 %f685, %f684; mul.f32 %f686, %f669, %f685; mul.f32 %f687, %f671, %f685; mul.f32 %f688, %f673, %f685; add.f32 %f689, %f670, %f679; add.f32 %f690, %f672, %f680; add.f32 %f691, %f674, %f681; add.f32 %f692, %f689, %f686; add.f32 %f693, %f690, %f687; add.f32 %f694, %f691, %f688; mul.f32 %f695, %f692, 0f3F13CD3A; mul.f32 %f696, %f693, 0f3F13CD3A; mul.f32 %f697, %f694, 0f3F13CD3A; div.rn.f32 %f698, %f1665, %f695; mul.f32 %f1666, %f1662, %f137; div.rn.f32 %f699, %f1666, %f696; mul.f32 %f1667, %f1663, %f137; div.rn.f32 %f700, %f1667, %f697; setp.eq.f32 %p49, %f1665, 0f00000000; selp.f32 %f701, 0f00000000, %f698, %p49; setp.eq.f32 %p50, %f1666, 0f00000000; selp.f32 %f702, 0f00000000, %f699, %p50; setp.eq.f32 %p51, %f1667, 0f00000000; selp.f32 %f703, 0f00000000, %f700, %p51; mul.f32 %f1668, %f670, %f701; mul.f32 %f1669, %f672, %f702; mul.f32 %f1670, %f674, %f703; mul.f32 %f1671, %f679, %f701; mul.f32 %f1672, %f680, %f702; mul.f32 %f1673, %f681, %f703; mul.f32 %f1674, %f686, %f701; mul.f32 %f1675, %f687, %f702; mul.f32 %f1676, %f688, %f703; mov.pred %p188, 0; $L__BB0_39: @%p188 bra $L__BB0_51; setp.eq.s32 %p52, %r46, 0; mov.u16 %rs182, 0; mov.f32 %f1692, 0f3F800000; mov.f32 %f1690, %f1692; mov.f32 %f1691, %f1692; @%p52 bra $L__BB0_48; abs.s32 %r10, %r46; setp.lt.s32 %p53, %r10, 1; mov.f32 %f1680, 0f00000000; mov.f32 %f1681, %f1680; mov.f32 %f1682, %f1680; @%p53 bra $L__BB0_47; mov.f32 %f719, 0f38D1B717; max.f32 %f720, %f718, %f719; and.b32 %r65, %r425, 32; setp.eq.s32 %p54, %r65, 0; mov.u32 %r424, 0; selp.f32 %f721, 0f3F800000, 0f41200000, %p54; mul.f32 %f162, %f721, %f720; and.b32 %r11, %r10, 1; setp.eq.s32 %p55, %r10, 1; mov.f32 %f1682, 0f00000000; mov.f32 %f1681, %f1682; mov.f32 %f1680, %f1682; @%p55 bra $L__BB0_45; sub.s32 %r423, %r10, %r11; $L__BB0_44: cvt.rn.f32.s32 %f743, %r424; add.f32 %f744, %f16, %f743; sub.f32 %f745, %f17, %f743; mul.f32 %f746, %f744, 0f3DD32618; cvt.rmi.f32.f32 %f747, %f746; sub.f32 %f748, %f746, %f747; mul.f32 %f749, %f745, 0f3DD2F1AA; cvt.rmi.f32.f32 %f750, %f749; sub.f32 %f751, %f749, %f750; mul.f32 %f752, %f743, 0f3DC74539; cvt.rmi.f32.f32 %f753, %f752; sub.f32 %f754, %f752, %f753; add.f32 %f755, %f751, 0f420551EC; add.f32 %f756, %f748, 0f420551EC; add.f32 %f757, %f754, 0f420551EC; mul.f32 %f758, %f748, %f755; fma.rn.f32 %f759, %f751, %f756, %f758; fma.rn.f32 %f760, %f754, %f757, %f759; add.f32 %f761, %f748, %f760; add.f32 %f762, %f751, %f760; add.f32 %f763, %f754, %f760; add.f32 %f764, %f761, %f762; mul.f32 %f765, %f763, %f764; cvt.rmi.f32.f32 %f766, %f765; sub.f32 %f767, %f765, %f766; add.f32 %f768, %f761, %f761; mul.f32 %f769, %f762, %f768; cvt.rmi.f32.f32 %f770, %f769; sub.f32 %f771, %f769, %f770; mul.f32 %f772, %f761, %f764; cvt.rmi.f32.f32 %f773, %f772; sub.f32 %f774, %f772, %f773; fma.rn.f32 %f775, %f767, 0f40000000, 0fBF800000; fma.rn.f32 %f776, %f771, 0f40000000, 0fBF800000; fma.rn.f32 %f777, %f774, 0f40000000, 0fBF800000; ld.global.f32 %f778, [%rd6+16]; ld.global.f32 %f779, [%rd6+4]; fma.rn.f32 %f780, %f778, %f775, %f779; ld.global.f32 %f781, [%rd6+8]; fma.rn.f32 %f782, %f778, %f776, %f781; ld.global.f32 %f783, [%rd6+12]; fma.rn.f32 %f784, %f778, %f777, %f783; sub.f32 %f785, %f780, %f10; sub.f32 %f786, %f782, %f11; sub.f32 %f787, %f784, %f12; mul.f32 %f788, %f786, %f786; fma.rn.f32 %f789, %f785, %f785, %f788; fma.rn.f32 %f790, %f787, %f787, %f789; sqrt.rn.f32 %f732, %f790; rcp.rn.f32 %f791, %f732; mul.f32 %f728, %f791, %f785; mul.f32 %f729, %f791, %f786; mul.f32 %f730, %f791, %f787; mov.f32 %f742, 0f00000000; mov.u32 %r174, 2; mov.u32 %r175, 1; mov.u32 %r176, 3; mov.u32 %r179, 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,%f728,%f729,%f730,%f162,%f732,%f742,%r175,%r208,%r175,%r174,%r175,%r176,%r179,%r179,%r179,%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 %f792, %r67; mov.b32 %f793, %r68; mov.b32 %f794, %r69; add.f32 %f795, %f1680, %f792; add.f32 %f796, %f1681, %f793; add.f32 %f797, %f1682, %f794; add.s32 %r209, %r424, 1; cvt.rn.f32.s32 %f798, %r209; add.f32 %f799, %f16, %f798; sub.f32 %f800, %f17, %f798; mul.f32 %f801, %f799, 0f3DD32618; cvt.rmi.f32.f32 %f802, %f801; sub.f32 %f803, %f801, %f802; mul.f32 %f804, %f800, 0f3DD2F1AA; cvt.rmi.f32.f32 %f805, %f804; sub.f32 %f806, %f804, %f805; mul.f32 %f807, %f798, 0f3DC74539; cvt.rmi.f32.f32 %f808, %f807; sub.f32 %f809, %f807, %f808; add.f32 %f810, %f806, 0f420551EC; add.f32 %f811, %f803, 0f420551EC; add.f32 %f812, %f809, 0f420551EC; mul.f32 %f813, %f803, %f810; fma.rn.f32 %f814, %f806, %f811, %f813; fma.rn.f32 %f815, %f809, %f812, %f814; add.f32 %f816, %f803, %f815; add.f32 %f817, %f806, %f815; add.f32 %f818, %f809, %f815; add.f32 %f819, %f816, %f817; mul.f32 %f820, %f818, %f819; cvt.rmi.f32.f32 %f821, %f820; sub.f32 %f822, %f820, %f821; add.f32 %f823, %f816, %f816; mul.f32 %f824, %f817, %f823; cvt.rmi.f32.f32 %f825, %f824; sub.f32 %f826, %f824, %f825; mul.f32 %f827, %f816, %f819; cvt.rmi.f32.f32 %f828, %f827; sub.f32 %f829, %f827, %f828; fma.rn.f32 %f830, %f822, 0f40000000, 0fBF800000; fma.rn.f32 %f831, %f826, 0f40000000, 0fBF800000; fma.rn.f32 %f832, %f829, 0f40000000, 0fBF800000; ld.global.f32 %f833, [%rd6+16]; ld.global.f32 %f834, [%rd6+4]; fma.rn.f32 %f835, %f833, %f830, %f834; ld.global.f32 %f836, [%rd6+8]; fma.rn.f32 %f837, %f833, %f831, %f836; ld.global.f32 %f838, [%rd6+12]; fma.rn.f32 %f839, %f833, %f832, %f838; sub.f32 %f840, %f835, %f10; sub.f32 %f841, %f837, %f11; sub.f32 %f842, %f839, %f12; mul.f32 %f843, %f841, %f841; fma.rn.f32 %f844, %f840, %f840, %f843; fma.rn.f32 %f845, %f842, %f842, %f844; sqrt.rn.f32 %f741, %f845; rcp.rn.f32 %f846, %f741; mul.f32 %f737, %f846, %f840; mul.f32 %f738, %f846, %f841; mul.f32 %f739, %f846, %f842; // 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,%f737,%f738,%f739,%f162,%f741,%f742,%r175,%r208,%r175,%r174,%r175,%r176,%r179,%r179,%r179,%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 %f847, %r138; mov.b32 %f848, %r139; mov.b32 %f849, %r140; add.f32 %f1680, %f795, %f847; add.f32 %f1681, %f796, %f848; add.f32 %f1682, %f797, %f849; 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 %f859, %r424; add.f32 %f860, %f16, %f859; sub.f32 %f861, %f17, %f859; mul.f32 %f862, %f860, 0f3DD32618; cvt.rmi.f32.f32 %f863, %f862; sub.f32 %f864, %f862, %f863; mul.f32 %f865, %f861, 0f3DD2F1AA; cvt.rmi.f32.f32 %f866, %f865; sub.f32 %f867, %f865, %f866; mul.f32 %f868, %f859, 0f3DC74539; cvt.rmi.f32.f32 %f869, %f868; sub.f32 %f870, %f868, %f869; add.f32 %f871, %f867, 0f420551EC; add.f32 %f872, %f864, 0f420551EC; add.f32 %f873, %f870, 0f420551EC; mul.f32 %f874, %f864, %f871; fma.rn.f32 %f875, %f867, %f872, %f874; fma.rn.f32 %f876, %f870, %f873, %f875; add.f32 %f877, %f864, %f876; add.f32 %f878, %f867, %f876; add.f32 %f879, %f870, %f876; add.f32 %f880, %f877, %f878; mul.f32 %f881, %f879, %f880; cvt.rmi.f32.f32 %f882, %f881; sub.f32 %f883, %f881, %f882; add.f32 %f884, %f877, %f877; mul.f32 %f885, %f878, %f884; cvt.rmi.f32.f32 %f886, %f885; sub.f32 %f887, %f885, %f886; mul.f32 %f888, %f877, %f880; cvt.rmi.f32.f32 %f889, %f888; sub.f32 %f890, %f888, %f889; fma.rn.f32 %f891, %f883, 0f40000000, 0fBF800000; fma.rn.f32 %f892, %f887, 0f40000000, 0fBF800000; fma.rn.f32 %f893, %f890, 0f40000000, 0fBF800000; ld.global.f32 %f894, [%rd6+16]; ld.global.f32 %f895, [%rd6+4]; fma.rn.f32 %f896, %f894, %f891, %f895; ld.global.f32 %f897, [%rd6+8]; fma.rn.f32 %f898, %f894, %f892, %f897; ld.global.f32 %f899, [%rd6+12]; fma.rn.f32 %f900, %f894, %f893, %f899; sub.f32 %f901, %f896, %f10; sub.f32 %f902, %f898, %f11; sub.f32 %f903, %f900, %f12; mul.f32 %f904, %f902, %f902; fma.rn.f32 %f905, %f901, %f901, %f904; fma.rn.f32 %f906, %f903, %f903, %f905; sqrt.rn.f32 %f857, %f906; rcp.rn.f32 %f907, %f857; mul.f32 %f853, %f907, %f901; mul.f32 %f854, %f907, %f902; mul.f32 %f855, %f907, %f903; mov.f32 %f858, 0f00000000; mov.u32 %r246, 2; mov.u32 %r247, 1; mov.u32 %r248, 3; mov.u32 %r251, 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,%f853,%f854,%f855,%f162,%f857,%f858,%r247,%r280,%r247,%r246,%r247,%r248,%r251,%r251,%r251,%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 %f908, %r210; mov.b32 %f909, %r211; mov.b32 %f910, %r212; add.f32 %f1680, %f1680, %f908; add.f32 %f1681, %f1681, %f909; add.f32 %f1682, %f1682, %f910; $L__BB0_47: cvt.rn.f32.s32 %f911, %r10; rcp.rn.f32 %f912, %f911; mul.f32 %f1692, %f912, %f1680; mul.f32 %f1690, %f912, %f1681; mul.f32 %f1691, %f912, %f1682; shr.u32 %r281, %r46, 31; cvt.u16.u32 %rs182, %r281; $L__BB0_48: fma.rn.f32 %f1693, %f1665, %f1692, %f1693; fma.rn.f32 %f1694, %f1666, %f1690, %f1694; fma.rn.f32 %f1695, %f1667, %f1691, %f1695; fma.rn.f32 %f1696, %f1668, %f1692, %f1696; fma.rn.f32 %f1697, %f1669, %f1690, %f1697; fma.rn.f32 %f1698, %f1670, %f1691, %f1698; fma.rn.f32 %f1699, %f1671, %f1692, %f1699; fma.rn.f32 %f1700, %f1672, %f1690, %f1700; fma.rn.f32 %f1701, %f1673, %f1691, %f1701; fma.rn.f32 %f1702, %f1674, %f1692, %f1702; fma.rn.f32 %f1703, %f1675, %f1690, %f1703; fma.rn.f32 %f1704, %f1676, %f1691, %f1704; setp.eq.s16 %p58, %rs182, 0; @%p58 bra $L__BB0_50; div.rn.f32 %f913, %f1665, %f487; div.rn.f32 %f914, %f913, %f103; cvt.sat.f32.f32 %f915, %f914; mul.f32 %f1692, %f1692, %f915; $L__BB0_50: add.f32 %f1705, %f1705, %f1692; $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 %f1693, 0f00000000; ld.const.u32 %r425, [params+104]; mov.f32 %f1694, %f1693; mov.f32 %f1695, %f1693; mov.f32 %f1696, %f1693; mov.f32 %f1697, %f1693; mov.f32 %f1698, %f1693; mov.f32 %f1699, %f1693; mov.f32 %f1700, %f1693; mov.f32 %f1701, %f1693; mov.f32 %f1702, %f1693; mov.f32 %f1703, %f1693; mov.f32 %f1704, %f1693; mov.f32 %f1705, %f1693; $L__BB0_53: and.b32 %r283, %r425, 8; setp.eq.s32 %p60, %r283, 0; @%p60 bra $L__BB0_67; cvt.sat.f32.f32 %f228, %f1705; 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 %f930, 0f3E68BA2E; cvt.rzi.f32.f32 %f931, %f930; add.f32 %f932, %f931, %f931; mov.f32 %f933, 0f3EE8BA2E; sub.f32 %f934, %f933, %f932; abs.f32 %f229, %f934; abs.f32 %f230, %f228; setp.lt.f32 %p61, %f230, 0f00800000; mul.f32 %f935, %f230, 0f4B800000; selp.f32 %f936, %f935, %f230, %p61; selp.f32 %f937, 0fC3170000, 0fC2FE0000, %p61; mov.b32 %r286, %f936; and.b32 %r287, %r286, 8388607; or.b32 %r288, %r287, 1065353216; mov.b32 %f938, %r288; shr.u32 %r289, %r286, 23; cvt.rn.f32.u32 %f939, %r289; add.f32 %f940, %f937, %f939; setp.gt.f32 %p62, %f938, 0f3FB504F3; mul.f32 %f941, %f938, 0f3F000000; add.f32 %f942, %f940, 0f3F800000; selp.f32 %f943, %f942, %f940, %p62; selp.f32 %f944, %f941, %f938, %p62; add.f32 %f945, %f944, 0fBF800000; add.f32 %f946, %f944, 0f3F800000; rcp.approx.ftz.f32 %f947, %f946; add.f32 %f948, %f945, %f945; mul.f32 %f949, %f948, %f947; mul.f32 %f950, %f949, %f949; mov.f32 %f951, 0f3C4CAF63; mov.f32 %f952, 0f3B18F0FE; fma.rn.f32 %f953, %f952, %f950, %f951; mov.f32 %f954, 0f3DAAAABD; fma.rn.f32 %f955, %f953, %f950, %f954; mul.rn.f32 %f956, %f955, %f950; mul.rn.f32 %f957, %f956, %f949; sub.f32 %f958, %f945, %f949; add.f32 %f959, %f958, %f958; neg.f32 %f960, %f949; fma.rn.f32 %f961, %f960, %f945, %f959; mul.rn.f32 %f962, %f947, %f961; add.f32 %f963, %f957, %f949; sub.f32 %f964, %f949, %f963; add.f32 %f965, %f957, %f964; add.f32 %f966, %f962, %f965; add.f32 %f967, %f963, %f966; sub.f32 %f968, %f963, %f967; add.f32 %f969, %f966, %f968; mov.f32 %f970, 0f3F317200; mul.rn.f32 %f971, %f943, %f970; mov.f32 %f972, 0f35BFBE8E; mul.rn.f32 %f973, %f943, %f972; add.f32 %f974, %f971, %f967; sub.f32 %f975, %f971, %f974; add.f32 %f976, %f967, %f975; add.f32 %f977, %f969, %f976; add.f32 %f978, %f973, %f977; add.f32 %f979, %f974, %f978; sub.f32 %f980, %f974, %f979; add.f32 %f981, %f978, %f980; mul.rn.f32 %f982, %f933, %f979; neg.f32 %f983, %f982; fma.rn.f32 %f984, %f933, %f979, %f983; fma.rn.f32 %f985, %f933, %f981, %f984; mov.f32 %f986, 0f00000000; fma.rn.f32 %f987, %f986, %f979, %f985; add.rn.f32 %f988, %f982, %f987; neg.f32 %f989, %f988; add.rn.f32 %f990, %f982, %f989; add.rn.f32 %f991, %f990, %f987; mov.b32 %r290, %f988; setp.eq.s32 %p63, %r290, 1118925336; add.s32 %r291, %r290, -1; mov.b32 %f992, %r291; add.f32 %f993, %f991, 0f37000000; selp.f32 %f231, %f993, %f991, %p63; selp.f32 %f994, %f992, %f988, %p63; mov.f32 %f995, 0f3FB8AA3B; mul.rn.f32 %f996, %f994, %f995; cvt.rzi.f32.f32 %f997, %f996; abs.f32 %f998, %f997; setp.gt.f32 %p64, %f998, 0f42FC0000; mov.b32 %r292, %f997; and.b32 %r293, %r292, -2147483648; or.b32 %r294, %r293, 1123811328; mov.b32 %f999, %r294; selp.f32 %f1000, %f999, %f997, %p64; mov.f32 %f1001, 0fBF317218; fma.rn.f32 %f1002, %f1000, %f1001, %f994; mov.f32 %f1003, 0f3102E308; fma.rn.f32 %f1004, %f1000, %f1003, %f1002; mul.f32 %f1005, %f1004, 0f3FB8AA3B; add.f32 %f1006, %f1000, 0f4B40007F; mov.b32 %r295, %f1006; shl.b32 %r296, %r295, 23; mov.b32 %f1007, %r296; ex2.approx.ftz.f32 %f1008, %f1005; mul.f32 %f232, %f1008, %f1007; setp.eq.f32 %p65, %f232, 0f7F800000; mov.f32 %f1719, 0f7F800000; @%p65 bra $L__BB0_56; fma.rn.f32 %f1719, %f232, %f231, %f232; $L__BB0_56: setp.lt.f32 %p66, %f228, 0f00000000; setp.eq.f32 %p67, %f229, 0f3F800000; and.pred %p3, %p66, %p67; setp.eq.f32 %p68, %f228, 0f00000000; @%p68 bra $L__BB0_60; bra.uni $L__BB0_57; $L__BB0_60: add.f32 %f1013, %f228, %f228; selp.f32 %f1721, %f1013, 0f00000000, %p67; bra.uni $L__BB0_61; $L__BB0_161: mov.f32 %f1598, 0f00000000; mov.u32 %r427, 0; // begin inline asm { cvt.rn.f16.f32 %rs129, %f1598;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs128, %f1598;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs127, %f1598;} // 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 %f1599, %rs137;} // end inline asm // begin inline asm { cvt.f32.f16 %f1600, %rs138;} // end inline asm // begin inline asm { cvt.f32.f16 %f1601, %rs139;} // end inline asm add.f32 %f1602, %f1599, 0f00000000; add.f32 %f1603, %f1600, 0f00000000; add.f32 %f1604, %f1601, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs136, %f1604;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs135, %f1603;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs134, %f1602;} // 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 %f1607, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs144, %f1607;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs143, %f1607;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs142, %f1607;} // 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 %f1608, %rs152;} // end inline asm // begin inline asm { cvt.f32.f16 %f1609, %rs153;} // end inline asm // begin inline asm { cvt.f32.f16 %f1610, %rs154;} // end inline asm add.f32 %f1611, %f1608, 0f00000000; add.f32 %f1612, %f1609, 0f00000000; add.f32 %f1613, %f1610, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs151, %f1613;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs150, %f1612;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs149, %f1611;} // 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 %f1616, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs159, %f1616;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs158, %f1616;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs157, %f1616;} // 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 %f1617, %rs167;} // end inline asm // begin inline asm { cvt.f32.f16 %f1618, %rs168;} // end inline asm // begin inline asm { cvt.f32.f16 %f1619, %rs169;} // end inline asm add.f32 %f1620, %f1617, 0f00000000; add.f32 %f1621, %f1618, 0f00000000; add.f32 %f1622, %f1619, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs166, %f1622;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs165, %f1621;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs164, %f1620;} // 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 %f1625, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs174, %f1625;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs173, %f1625;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs172, %f1625;} // 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, %f1719; xor.b32 %r298, %r297, -2147483648; mov.b32 %f1009, %r298; selp.f32 %f1721, %f1009, %f1719, %p3; setp.geu.f32 %p69, %f228, 0f00000000; @%p69 bra $L__BB0_61; mov.f32 %f1010, 0f3EE8BA2E; cvt.rzi.f32.f32 %f1011, %f1010; setp.eq.f32 %p70, %f1011, 0f3EE8BA2E; @%p70 bra $L__BB0_61; mov.f32 %f1721, 0f7FFFFFFF; $L__BB0_61: add.f32 %f1014, %f230, 0f3EE8BA2E; mov.b32 %r299, %f1014; setp.lt.s32 %p72, %r299, 2139095040; @%p72 bra $L__BB0_66; setp.gtu.f32 %p73, %f230, 0f7F800000; @%p73 bra $L__BB0_65; bra.uni $L__BB0_63; $L__BB0_65: add.f32 %f1721, %f228, 0f3EE8BA2E; bra.uni $L__BB0_66; $L__BB0_63: setp.neu.f32 %p74, %f230, 0f7F800000; @%p74 bra $L__BB0_66; selp.f32 %f1721, 0fFF800000, 0f7F800000, %p3; $L__BB0_66: mul.f32 %f1015, %f1721, 0f437F0000; setp.eq.f32 %p75, %f228, 0f3F800000; selp.f32 %f1016, 0f437F0000, %f1015, %p75; cvt.rzi.u32.f32 %r300, %f1016; 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 %f1018, 0f3E666666; cvt.rzi.f32.f32 %f1019, %f1018; add.f32 %f1020, %f1019, %f1019; mov.f32 %f1021, 0f3EE66666; sub.f32 %f1022, %f1021, %f1020; abs.f32 %f241, %f1022; abs.f32 %f242, %f1693; setp.lt.f32 %p80, %f242, 0f00800000; mul.f32 %f1023, %f242, 0f4B800000; selp.f32 %f1024, %f1023, %f242, %p80; selp.f32 %f1025, 0fC3170000, 0fC2FE0000, %p80; mov.b32 %r302, %f1024; and.b32 %r303, %r302, 8388607; or.b32 %r304, %r303, 1065353216; mov.b32 %f1026, %r304; shr.u32 %r305, %r302, 23; cvt.rn.f32.u32 %f1027, %r305; add.f32 %f1028, %f1025, %f1027; setp.gt.f32 %p81, %f1026, 0f3FB504F3; mul.f32 %f1029, %f1026, 0f3F000000; add.f32 %f1030, %f1028, 0f3F800000; selp.f32 %f1031, %f1030, %f1028, %p81; selp.f32 %f1032, %f1029, %f1026, %p81; add.f32 %f1033, %f1032, 0fBF800000; add.f32 %f1034, %f1032, 0f3F800000; rcp.approx.ftz.f32 %f1035, %f1034; add.f32 %f1036, %f1033, %f1033; mul.f32 %f1037, %f1036, %f1035; mul.f32 %f1038, %f1037, %f1037; mov.f32 %f1039, 0f3C4CAF63; mov.f32 %f1040, 0f3B18F0FE; fma.rn.f32 %f1041, %f1040, %f1038, %f1039; mov.f32 %f1042, 0f3DAAAABD; fma.rn.f32 %f1043, %f1041, %f1038, %f1042; mul.rn.f32 %f1044, %f1043, %f1038; mul.rn.f32 %f1045, %f1044, %f1037; sub.f32 %f1046, %f1033, %f1037; add.f32 %f1047, %f1046, %f1046; neg.f32 %f1048, %f1037; fma.rn.f32 %f1049, %f1048, %f1033, %f1047; mul.rn.f32 %f1050, %f1035, %f1049; add.f32 %f1051, %f1045, %f1037; sub.f32 %f1052, %f1037, %f1051; add.f32 %f1053, %f1045, %f1052; add.f32 %f1054, %f1050, %f1053; add.f32 %f1055, %f1051, %f1054; sub.f32 %f1056, %f1051, %f1055; add.f32 %f1057, %f1054, %f1056; mov.f32 %f1058, 0f3F317200; mul.rn.f32 %f1059, %f1031, %f1058; mov.f32 %f1060, 0f35BFBE8E; mul.rn.f32 %f1061, %f1031, %f1060; add.f32 %f1062, %f1059, %f1055; sub.f32 %f1063, %f1059, %f1062; add.f32 %f1064, %f1055, %f1063; add.f32 %f1065, %f1057, %f1064; add.f32 %f1066, %f1061, %f1065; add.f32 %f1067, %f1062, %f1066; sub.f32 %f1068, %f1062, %f1067; add.f32 %f1069, %f1066, %f1068; mul.rn.f32 %f1070, %f1021, %f1067; neg.f32 %f1071, %f1070; fma.rn.f32 %f1072, %f1021, %f1067, %f1071; fma.rn.f32 %f1073, %f1021, %f1069, %f1072; mov.f32 %f1074, 0f00000000; fma.rn.f32 %f1075, %f1074, %f1067, %f1073; add.rn.f32 %f1076, %f1070, %f1075; neg.f32 %f1077, %f1076; add.rn.f32 %f1078, %f1070, %f1077; add.rn.f32 %f1079, %f1078, %f1075; mov.b32 %r306, %f1076; setp.eq.s32 %p82, %r306, 1118925336; add.s32 %r307, %r306, -1; mov.b32 %f1080, %r307; add.f32 %f1081, %f1079, 0f37000000; selp.f32 %f243, %f1081, %f1079, %p82; selp.f32 %f1082, %f1080, %f1076, %p82; mov.f32 %f1083, 0f3FB8AA3B; mul.rn.f32 %f1084, %f1082, %f1083; cvt.rzi.f32.f32 %f1085, %f1084; abs.f32 %f1086, %f1085; setp.gt.f32 %p83, %f1086, 0f42FC0000; mov.b32 %r308, %f1085; and.b32 %r309, %r308, -2147483648; or.b32 %r310, %r309, 1123811328; mov.b32 %f1087, %r310; selp.f32 %f1088, %f1087, %f1085, %p83; mov.f32 %f1089, 0fBF317218; fma.rn.f32 %f1090, %f1088, %f1089, %f1082; mov.f32 %f1091, 0f3102E308; fma.rn.f32 %f1092, %f1088, %f1091, %f1090; mul.f32 %f1093, %f1092, 0f3FB8AA3B; add.f32 %f1094, %f1088, 0f4B40007F; mov.b32 %r311, %f1094; shl.b32 %r312, %r311, 23; mov.b32 %f1095, %r312; ex2.approx.ftz.f32 %f1096, %f1093; mul.f32 %f244, %f1096, %f1095; setp.eq.f32 %p84, %f244, 0f7F800000; mov.f32 %f1722, 0f7F800000; @%p84 bra $L__BB0_70; fma.rn.f32 %f1722, %f244, %f243, %f244; $L__BB0_70: setp.lt.f32 %p85, %f1693, 0f00000000; setp.eq.f32 %p86, %f241, 0f3F800000; and.pred %p4, %p85, %p86; setp.eq.f32 %p87, %f1693, 0f00000000; @%p87 bra $L__BB0_74; bra.uni $L__BB0_71; $L__BB0_74: add.f32 %f1101, %f1693, %f1693; selp.f32 %f1724, %f1101, 0f00000000, %p86; bra.uni $L__BB0_75; $L__BB0_71: mov.b32 %r313, %f1722; xor.b32 %r314, %r313, -2147483648; mov.b32 %f1097, %r314; selp.f32 %f1724, %f1097, %f1722, %p4; setp.geu.f32 %p88, %f1693, 0f00000000; @%p88 bra $L__BB0_75; mov.f32 %f1098, 0f3EE66666; cvt.rzi.f32.f32 %f1099, %f1098; setp.eq.f32 %p89, %f1099, 0f3EE66666; @%p89 bra $L__BB0_75; mov.f32 %f1724, 0f7FFFFFFF; $L__BB0_75: add.f32 %f1102, %f242, 0f3EE66666; mov.b32 %r315, %f1102; setp.lt.s32 %p91, %r315, 2139095040; @%p91 bra $L__BB0_80; setp.gtu.f32 %p92, %f242, 0f7F800000; @%p92 bra $L__BB0_79; bra.uni $L__BB0_77; $L__BB0_79: add.f32 %f1724, %f1693, 0f3EE66666; bra.uni $L__BB0_80; $L__BB0_77: setp.neu.f32 %p93, %f242, 0f7F800000; @%p93 bra $L__BB0_80; selp.f32 %f1724, 0fFF800000, 0f7F800000, %p4; $L__BB0_80: setp.eq.f32 %p94, %f1693, 0f3F800000; selp.f32 %f253, 0f3F800000, %f1724, %p94; abs.f32 %f254, %f1694; setp.lt.f32 %p95, %f254, 0f00800000; mul.f32 %f1104, %f254, 0f4B800000; selp.f32 %f1105, %f1104, %f254, %p95; selp.f32 %f1106, 0fC3170000, 0fC2FE0000, %p95; mov.b32 %r316, %f1105; and.b32 %r317, %r316, 8388607; or.b32 %r318, %r317, 1065353216; mov.b32 %f1107, %r318; shr.u32 %r319, %r316, 23; cvt.rn.f32.u32 %f1108, %r319; add.f32 %f1109, %f1106, %f1108; setp.gt.f32 %p96, %f1107, 0f3FB504F3; mul.f32 %f1110, %f1107, 0f3F000000; add.f32 %f1111, %f1109, 0f3F800000; selp.f32 %f1112, %f1111, %f1109, %p96; selp.f32 %f1113, %f1110, %f1107, %p96; add.f32 %f1114, %f1113, 0fBF800000; add.f32 %f1115, %f1113, 0f3F800000; rcp.approx.ftz.f32 %f1116, %f1115; add.f32 %f1117, %f1114, %f1114; mul.f32 %f1118, %f1117, %f1116; mul.f32 %f1119, %f1118, %f1118; mov.f32 %f1120, 0f3C4CAF63; mov.f32 %f1121, 0f3B18F0FE; fma.rn.f32 %f1122, %f1121, %f1119, %f1120; mov.f32 %f1123, 0f3DAAAABD; fma.rn.f32 %f1124, %f1122, %f1119, %f1123; mul.rn.f32 %f1125, %f1124, %f1119; mul.rn.f32 %f1126, %f1125, %f1118; sub.f32 %f1127, %f1114, %f1118; add.f32 %f1128, %f1127, %f1127; neg.f32 %f1129, %f1118; fma.rn.f32 %f1130, %f1129, %f1114, %f1128; mul.rn.f32 %f1131, %f1116, %f1130; add.f32 %f1132, %f1126, %f1118; sub.f32 %f1133, %f1118, %f1132; add.f32 %f1134, %f1126, %f1133; add.f32 %f1135, %f1131, %f1134; add.f32 %f1136, %f1132, %f1135; sub.f32 %f1137, %f1132, %f1136; add.f32 %f1138, %f1135, %f1137; mov.f32 %f1139, 0f3F317200; mul.rn.f32 %f1140, %f1112, %f1139; mov.f32 %f1141, 0f35BFBE8E; mul.rn.f32 %f1142, %f1112, %f1141; add.f32 %f1143, %f1140, %f1136; sub.f32 %f1144, %f1140, %f1143; add.f32 %f1145, %f1136, %f1144; add.f32 %f1146, %f1138, %f1145; add.f32 %f1147, %f1142, %f1146; add.f32 %f1148, %f1143, %f1147; sub.f32 %f1149, %f1143, %f1148; add.f32 %f1150, %f1147, %f1149; mov.f32 %f1151, 0f3EE66666; mul.rn.f32 %f1152, %f1151, %f1148; neg.f32 %f1153, %f1152; fma.rn.f32 %f1154, %f1151, %f1148, %f1153; fma.rn.f32 %f1155, %f1151, %f1150, %f1154; mov.f32 %f1156, 0f00000000; fma.rn.f32 %f1157, %f1156, %f1148, %f1155; add.rn.f32 %f1158, %f1152, %f1157; neg.f32 %f1159, %f1158; add.rn.f32 %f1160, %f1152, %f1159; add.rn.f32 %f1161, %f1160, %f1157; mov.b32 %r320, %f1158; setp.eq.s32 %p97, %r320, 1118925336; add.s32 %r321, %r320, -1; mov.b32 %f1162, %r321; add.f32 %f1163, %f1161, 0f37000000; selp.f32 %f255, %f1163, %f1161, %p97; selp.f32 %f1164, %f1162, %f1158, %p97; mov.f32 %f1165, 0f3FB8AA3B; mul.rn.f32 %f1166, %f1164, %f1165; cvt.rzi.f32.f32 %f1167, %f1166; abs.f32 %f1168, %f1167; setp.gt.f32 %p98, %f1168, 0f42FC0000; mov.b32 %r322, %f1167; and.b32 %r323, %r322, -2147483648; or.b32 %r324, %r323, 1123811328; mov.b32 %f1169, %r324; selp.f32 %f1170, %f1169, %f1167, %p98; mov.f32 %f1171, 0fBF317218; fma.rn.f32 %f1172, %f1170, %f1171, %f1164; mov.f32 %f1173, 0f3102E308; fma.rn.f32 %f1174, %f1170, %f1173, %f1172; mul.f32 %f1175, %f1174, 0f3FB8AA3B; add.f32 %f1176, %f1170, 0f4B40007F; mov.b32 %r325, %f1176; shl.b32 %r326, %r325, 23; mov.b32 %f1177, %r326; ex2.approx.ftz.f32 %f1178, %f1175; mul.f32 %f256, %f1178, %f1177; setp.eq.f32 %p99, %f256, 0f7F800000; mov.f32 %f1725, 0f7F800000; @%p99 bra $L__BB0_82; fma.rn.f32 %f1725, %f256, %f255, %f256; $L__BB0_82: setp.lt.f32 %p100, %f1694, 0f00000000; and.pred %p5, %p100, %p86; setp.eq.f32 %p102, %f1694, 0f00000000; @%p102 bra $L__BB0_86; bra.uni $L__BB0_83; $L__BB0_86: add.f32 %f1183, %f1694, %f1694; selp.f32 %f1727, %f1183, 0f00000000, %p86; bra.uni $L__BB0_87; $L__BB0_83: mov.b32 %r327, %f1725; xor.b32 %r328, %r327, -2147483648; mov.b32 %f1179, %r328; selp.f32 %f1727, %f1179, %f1725, %p5; setp.geu.f32 %p103, %f1694, 0f00000000; @%p103 bra $L__BB0_87; mov.f32 %f1180, 0f3EE66666; cvt.rzi.f32.f32 %f1181, %f1180; setp.eq.f32 %p104, %f1181, 0f3EE66666; @%p104 bra $L__BB0_87; mov.f32 %f1727, 0f7FFFFFFF; $L__BB0_87: add.f32 %f1184, %f254, 0f3EE66666; mov.b32 %r329, %f1184; setp.lt.s32 %p106, %r329, 2139095040; @%p106 bra $L__BB0_92; setp.gtu.f32 %p107, %f254, 0f7F800000; @%p107 bra $L__BB0_91; bra.uni $L__BB0_89; $L__BB0_91: add.f32 %f1727, %f1694, 0f3EE66666; bra.uni $L__BB0_92; $L__BB0_89: setp.neu.f32 %p108, %f254, 0f7F800000; @%p108 bra $L__BB0_92; selp.f32 %f1727, 0fFF800000, 0f7F800000, %p5; $L__BB0_92: setp.eq.f32 %p109, %f1694, 0f3F800000; selp.f32 %f265, 0f3F800000, %f1727, %p109; abs.f32 %f266, %f1695; setp.lt.f32 %p110, %f266, 0f00800000; mul.f32 %f1186, %f266, 0f4B800000; selp.f32 %f1187, %f1186, %f266, %p110; selp.f32 %f1188, 0fC3170000, 0fC2FE0000, %p110; mov.b32 %r330, %f1187; and.b32 %r331, %r330, 8388607; or.b32 %r332, %r331, 1065353216; mov.b32 %f1189, %r332; shr.u32 %r333, %r330, 23; cvt.rn.f32.u32 %f1190, %r333; add.f32 %f1191, %f1188, %f1190; setp.gt.f32 %p111, %f1189, 0f3FB504F3; mul.f32 %f1192, %f1189, 0f3F000000; add.f32 %f1193, %f1191, 0f3F800000; selp.f32 %f1194, %f1193, %f1191, %p111; selp.f32 %f1195, %f1192, %f1189, %p111; add.f32 %f1196, %f1195, 0fBF800000; add.f32 %f1197, %f1195, 0f3F800000; rcp.approx.ftz.f32 %f1198, %f1197; add.f32 %f1199, %f1196, %f1196; mul.f32 %f1200, %f1199, %f1198; mul.f32 %f1201, %f1200, %f1200; mov.f32 %f1202, 0f3C4CAF63; mov.f32 %f1203, 0f3B18F0FE; fma.rn.f32 %f1204, %f1203, %f1201, %f1202; mov.f32 %f1205, 0f3DAAAABD; fma.rn.f32 %f1206, %f1204, %f1201, %f1205; mul.rn.f32 %f1207, %f1206, %f1201; mul.rn.f32 %f1208, %f1207, %f1200; sub.f32 %f1209, %f1196, %f1200; add.f32 %f1210, %f1209, %f1209; neg.f32 %f1211, %f1200; fma.rn.f32 %f1212, %f1211, %f1196, %f1210; mul.rn.f32 %f1213, %f1198, %f1212; add.f32 %f1214, %f1208, %f1200; sub.f32 %f1215, %f1200, %f1214; add.f32 %f1216, %f1208, %f1215; add.f32 %f1217, %f1213, %f1216; add.f32 %f1218, %f1214, %f1217; sub.f32 %f1219, %f1214, %f1218; add.f32 %f1220, %f1217, %f1219; mov.f32 %f1221, 0f3F317200; mul.rn.f32 %f1222, %f1194, %f1221; mov.f32 %f1223, 0f35BFBE8E; mul.rn.f32 %f1224, %f1194, %f1223; add.f32 %f1225, %f1222, %f1218; sub.f32 %f1226, %f1222, %f1225; add.f32 %f1227, %f1218, %f1226; add.f32 %f1228, %f1220, %f1227; add.f32 %f1229, %f1224, %f1228; add.f32 %f1230, %f1225, %f1229; sub.f32 %f1231, %f1225, %f1230; add.f32 %f1232, %f1229, %f1231; mov.f32 %f1233, 0f3EE66666; mul.rn.f32 %f1234, %f1233, %f1230; neg.f32 %f1235, %f1234; fma.rn.f32 %f1236, %f1233, %f1230, %f1235; fma.rn.f32 %f1237, %f1233, %f1232, %f1236; mov.f32 %f1238, 0f00000000; fma.rn.f32 %f1239, %f1238, %f1230, %f1237; add.rn.f32 %f1240, %f1234, %f1239; neg.f32 %f1241, %f1240; add.rn.f32 %f1242, %f1234, %f1241; add.rn.f32 %f1243, %f1242, %f1239; mov.b32 %r334, %f1240; setp.eq.s32 %p112, %r334, 1118925336; add.s32 %r335, %r334, -1; mov.b32 %f1244, %r335; add.f32 %f1245, %f1243, 0f37000000; selp.f32 %f267, %f1245, %f1243, %p112; selp.f32 %f1246, %f1244, %f1240, %p112; mov.f32 %f1247, 0f3FB8AA3B; mul.rn.f32 %f1248, %f1246, %f1247; cvt.rzi.f32.f32 %f1249, %f1248; abs.f32 %f1250, %f1249; setp.gt.f32 %p113, %f1250, 0f42FC0000; mov.b32 %r336, %f1249; and.b32 %r337, %r336, -2147483648; or.b32 %r338, %r337, 1123811328; mov.b32 %f1251, %r338; selp.f32 %f1252, %f1251, %f1249, %p113; mov.f32 %f1253, 0fBF317218; fma.rn.f32 %f1254, %f1252, %f1253, %f1246; mov.f32 %f1255, 0f3102E308; fma.rn.f32 %f1256, %f1252, %f1255, %f1254; mul.f32 %f1257, %f1256, 0f3FB8AA3B; add.f32 %f1258, %f1252, 0f4B40007F; mov.b32 %r339, %f1258; shl.b32 %r340, %r339, 23; mov.b32 %f1259, %r340; ex2.approx.ftz.f32 %f1260, %f1257; mul.f32 %f268, %f1260, %f1259; setp.eq.f32 %p114, %f268, 0f7F800000; mov.f32 %f1728, 0f7F800000; @%p114 bra $L__BB0_94; fma.rn.f32 %f1728, %f268, %f267, %f268; $L__BB0_94: setp.lt.f32 %p115, %f1695, 0f00000000; and.pred %p6, %p115, %p86; setp.eq.f32 %p117, %f1695, 0f00000000; @%p117 bra $L__BB0_98; bra.uni $L__BB0_95; $L__BB0_98: add.f32 %f1265, %f1695, %f1695; selp.f32 %f1730, %f1265, 0f00000000, %p86; bra.uni $L__BB0_99; $L__BB0_95: mov.b32 %r341, %f1728; xor.b32 %r342, %r341, -2147483648; mov.b32 %f1261, %r342; selp.f32 %f1730, %f1261, %f1728, %p6; setp.geu.f32 %p118, %f1695, 0f00000000; @%p118 bra $L__BB0_99; mov.f32 %f1262, 0f3EE66666; cvt.rzi.f32.f32 %f1263, %f1262; setp.eq.f32 %p119, %f1263, 0f3EE66666; @%p119 bra $L__BB0_99; mov.f32 %f1730, 0f7FFFFFFF; $L__BB0_99: add.f32 %f1266, %f266, 0f3EE66666; mov.b32 %r343, %f1266; setp.lt.s32 %p121, %r343, 2139095040; @%p121 bra $L__BB0_104; setp.gtu.f32 %p122, %f266, 0f7F800000; @%p122 bra $L__BB0_103; bra.uni $L__BB0_101; $L__BB0_103: add.f32 %f1730, %f1695, 0f3EE66666; bra.uni $L__BB0_104; $L__BB0_101: setp.neu.f32 %p123, %f266, 0f7F800000; @%p123 bra $L__BB0_104; selp.f32 %f1730, 0fFF800000, 0f7F800000, %p6; $L__BB0_104: setp.eq.f32 %p124, %f1695, 0f3F800000; mov.f32 %f1268, 0f3F800000; selp.f32 %f1269, 0f3F800000, %f1730, %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 %f1270, %f253, %f1268; mov.f32 %f1271, 0f00000000; max.f32 %f277, %f1271, %f1270; min.f32 %f1272, %f265, %f1268; max.f32 %f278, %f1271, %f1272; min.f32 %f1273, %f1269, %f1268; max.f32 %f279, %f1271, %f1273; mov.f32 %f1274, 0f3E555555; cvt.rzi.f32.f32 %f1275, %f1274; add.f32 %f1276, %f1275, %f1275; mov.f32 %f1277, 0f3ED55555; sub.f32 %f1278, %f1277, %f1276; abs.f32 %f280, %f1278; abs.f32 %f281, %f277; setp.lt.f32 %p125, %f281, 0f00800000; mul.f32 %f1279, %f281, 0f4B800000; selp.f32 %f1280, %f1279, %f281, %p125; selp.f32 %f1281, 0fC3170000, 0fC2FE0000, %p125; mov.b32 %r346, %f1280; and.b32 %r347, %r346, 8388607; or.b32 %r348, %r347, 1065353216; mov.b32 %f1282, %r348; shr.u32 %r349, %r346, 23; cvt.rn.f32.u32 %f1283, %r349; add.f32 %f1284, %f1281, %f1283; setp.gt.f32 %p126, %f1282, 0f3FB504F3; mul.f32 %f1285, %f1282, 0f3F000000; add.f32 %f1286, %f1284, 0f3F800000; selp.f32 %f1287, %f1286, %f1284, %p126; selp.f32 %f1288, %f1285, %f1282, %p126; add.f32 %f1289, %f1288, 0fBF800000; add.f32 %f1290, %f1288, 0f3F800000; rcp.approx.ftz.f32 %f1291, %f1290; add.f32 %f1292, %f1289, %f1289; mul.f32 %f1293, %f1292, %f1291; mul.f32 %f1294, %f1293, %f1293; mov.f32 %f1295, 0f3C4CAF63; mov.f32 %f1296, 0f3B18F0FE; fma.rn.f32 %f1297, %f1296, %f1294, %f1295; mov.f32 %f1298, 0f3DAAAABD; fma.rn.f32 %f1299, %f1297, %f1294, %f1298; mul.rn.f32 %f1300, %f1299, %f1294; mul.rn.f32 %f1301, %f1300, %f1293; sub.f32 %f1302, %f1289, %f1293; add.f32 %f1303, %f1302, %f1302; neg.f32 %f1304, %f1293; fma.rn.f32 %f1305, %f1304, %f1289, %f1303; mul.rn.f32 %f1306, %f1291, %f1305; add.f32 %f1307, %f1301, %f1293; sub.f32 %f1308, %f1293, %f1307; add.f32 %f1309, %f1301, %f1308; add.f32 %f1310, %f1306, %f1309; add.f32 %f1311, %f1307, %f1310; sub.f32 %f1312, %f1307, %f1311; add.f32 %f1313, %f1310, %f1312; mov.f32 %f1314, 0f3F317200; mul.rn.f32 %f1315, %f1287, %f1314; mov.f32 %f1316, 0f35BFBE8E; mul.rn.f32 %f1317, %f1287, %f1316; add.f32 %f1318, %f1315, %f1311; sub.f32 %f1319, %f1315, %f1318; add.f32 %f1320, %f1311, %f1319; add.f32 %f1321, %f1313, %f1320; add.f32 %f1322, %f1317, %f1321; add.f32 %f1323, %f1318, %f1322; sub.f32 %f1324, %f1318, %f1323; add.f32 %f1325, %f1322, %f1324; mul.rn.f32 %f1326, %f1277, %f1323; neg.f32 %f1327, %f1326; fma.rn.f32 %f1328, %f1277, %f1323, %f1327; fma.rn.f32 %f1329, %f1277, %f1325, %f1328; fma.rn.f32 %f1330, %f1271, %f1323, %f1329; add.rn.f32 %f1331, %f1326, %f1330; neg.f32 %f1332, %f1331; add.rn.f32 %f1333, %f1326, %f1332; add.rn.f32 %f1334, %f1333, %f1330; mov.b32 %r350, %f1331; setp.eq.s32 %p127, %r350, 1118925336; add.s32 %r351, %r350, -1; mov.b32 %f1335, %r351; add.f32 %f1336, %f1334, 0f37000000; selp.f32 %f282, %f1336, %f1334, %p127; selp.f32 %f1337, %f1335, %f1331, %p127; mov.f32 %f1338, 0f3FB8AA3B; mul.rn.f32 %f1339, %f1337, %f1338; cvt.rzi.f32.f32 %f1340, %f1339; abs.f32 %f1341, %f1340; setp.gt.f32 %p128, %f1341, 0f42FC0000; mov.b32 %r352, %f1340; and.b32 %r353, %r352, -2147483648; or.b32 %r354, %r353, 1123811328; mov.b32 %f1342, %r354; selp.f32 %f1343, %f1342, %f1340, %p128; mov.f32 %f1344, 0fBF317218; fma.rn.f32 %f1345, %f1343, %f1344, %f1337; mov.f32 %f1346, 0f3102E308; fma.rn.f32 %f1347, %f1343, %f1346, %f1345; mul.f32 %f1348, %f1347, 0f3FB8AA3B; add.f32 %f1349, %f1343, 0f4B40007F; mov.b32 %r355, %f1349; shl.b32 %r356, %r355, 23; mov.b32 %f1350, %r356; ex2.approx.ftz.f32 %f1351, %f1348; mul.f32 %f283, %f1351, %f1350; setp.eq.f32 %p129, %f283, 0f7F800000; mov.f32 %f1731, 0f7F800000; @%p129 bra $L__BB0_106; fma.rn.f32 %f1731, %f283, %f282, %f283; $L__BB0_106: setp.lt.f32 %p130, %f277, 0f00000000; setp.eq.f32 %p131, %f280, 0f3F800000; and.pred %p7, %p130, %p131; setp.eq.f32 %p132, %f277, 0f00000000; @%p132 bra $L__BB0_110; bra.uni $L__BB0_107; $L__BB0_110: add.f32 %f1356, %f277, %f277; selp.f32 %f1733, %f1356, 0f00000000, %p131; bra.uni $L__BB0_111; $L__BB0_107: mov.b32 %r357, %f1731; xor.b32 %r358, %r357, -2147483648; mov.b32 %f1352, %r358; selp.f32 %f1733, %f1352, %f1731, %p7; setp.geu.f32 %p133, %f277, 0f00000000; @%p133 bra $L__BB0_111; mov.f32 %f1353, 0f3ED55555; cvt.rzi.f32.f32 %f1354, %f1353; setp.eq.f32 %p134, %f1354, 0f3ED55555; @%p134 bra $L__BB0_111; mov.f32 %f1733, 0f7FFFFFFF; $L__BB0_111: add.f32 %f1357, %f281, 0f3ED55555; mov.b32 %r359, %f1357; setp.lt.s32 %p136, %r359, 2139095040; @%p136 bra $L__BB0_116; setp.gtu.f32 %p137, %f281, 0f7F800000; @%p137 bra $L__BB0_115; bra.uni $L__BB0_113; $L__BB0_115: add.f32 %f1733, %f277, 0f3ED55555; bra.uni $L__BB0_116; $L__BB0_113: setp.neu.f32 %p138, %f281, 0f7F800000; @%p138 bra $L__BB0_116; selp.f32 %f1733, 0fFF800000, 0f7F800000, %p7; $L__BB0_116: abs.f32 %f292, %f278; setp.lt.f32 %p139, %f292, 0f00800000; mul.f32 %f1359, %f292, 0f4B800000; selp.f32 %f1360, %f1359, %f292, %p139; selp.f32 %f1361, 0fC3170000, 0fC2FE0000, %p139; mov.b32 %r360, %f1360; and.b32 %r361, %r360, 8388607; or.b32 %r362, %r361, 1065353216; mov.b32 %f1362, %r362; shr.u32 %r363, %r360, 23; cvt.rn.f32.u32 %f1363, %r363; add.f32 %f1364, %f1361, %f1363; setp.gt.f32 %p140, %f1362, 0f3FB504F3; mul.f32 %f1365, %f1362, 0f3F000000; add.f32 %f1366, %f1364, 0f3F800000; selp.f32 %f1367, %f1366, %f1364, %p140; selp.f32 %f1368, %f1365, %f1362, %p140; add.f32 %f1369, %f1368, 0fBF800000; add.f32 %f1370, %f1368, 0f3F800000; rcp.approx.ftz.f32 %f1371, %f1370; add.f32 %f1372, %f1369, %f1369; mul.f32 %f1373, %f1372, %f1371; mul.f32 %f1374, %f1373, %f1373; mov.f32 %f1375, 0f3C4CAF63; mov.f32 %f1376, 0f3B18F0FE; fma.rn.f32 %f1377, %f1376, %f1374, %f1375; mov.f32 %f1378, 0f3DAAAABD; fma.rn.f32 %f1379, %f1377, %f1374, %f1378; mul.rn.f32 %f1380, %f1379, %f1374; mul.rn.f32 %f1381, %f1380, %f1373; sub.f32 %f1382, %f1369, %f1373; add.f32 %f1383, %f1382, %f1382; neg.f32 %f1384, %f1373; fma.rn.f32 %f1385, %f1384, %f1369, %f1383; mul.rn.f32 %f1386, %f1371, %f1385; add.f32 %f1387, %f1381, %f1373; sub.f32 %f1388, %f1373, %f1387; add.f32 %f1389, %f1381, %f1388; add.f32 %f1390, %f1386, %f1389; add.f32 %f1391, %f1387, %f1390; sub.f32 %f1392, %f1387, %f1391; add.f32 %f1393, %f1390, %f1392; mov.f32 %f1394, 0f3F317200; mul.rn.f32 %f1395, %f1367, %f1394; mov.f32 %f1396, 0f35BFBE8E; mul.rn.f32 %f1397, %f1367, %f1396; add.f32 %f1398, %f1395, %f1391; sub.f32 %f1399, %f1395, %f1398; add.f32 %f1400, %f1391, %f1399; add.f32 %f1401, %f1393, %f1400; add.f32 %f1402, %f1397, %f1401; add.f32 %f1403, %f1398, %f1402; sub.f32 %f1404, %f1398, %f1403; add.f32 %f1405, %f1402, %f1404; mov.f32 %f1406, 0f3ED55555; mul.rn.f32 %f1407, %f1406, %f1403; neg.f32 %f1408, %f1407; fma.rn.f32 %f1409, %f1406, %f1403, %f1408; fma.rn.f32 %f1410, %f1406, %f1405, %f1409; mov.f32 %f1411, 0f00000000; fma.rn.f32 %f1412, %f1411, %f1403, %f1410; add.rn.f32 %f1413, %f1407, %f1412; neg.f32 %f1414, %f1413; add.rn.f32 %f1415, %f1407, %f1414; add.rn.f32 %f1416, %f1415, %f1412; mov.b32 %r364, %f1413; setp.eq.s32 %p141, %r364, 1118925336; add.s32 %r365, %r364, -1; mov.b32 %f1417, %r365; add.f32 %f1418, %f1416, 0f37000000; selp.f32 %f293, %f1418, %f1416, %p141; selp.f32 %f1419, %f1417, %f1413, %p141; mov.f32 %f1420, 0f3FB8AA3B; mul.rn.f32 %f1421, %f1419, %f1420; cvt.rzi.f32.f32 %f1422, %f1421; abs.f32 %f1423, %f1422; setp.gt.f32 %p142, %f1423, 0f42FC0000; mov.b32 %r366, %f1422; and.b32 %r367, %r366, -2147483648; or.b32 %r368, %r367, 1123811328; mov.b32 %f1424, %r368; selp.f32 %f1425, %f1424, %f1422, %p142; mov.f32 %f1426, 0fBF317218; fma.rn.f32 %f1427, %f1425, %f1426, %f1419; mov.f32 %f1428, 0f3102E308; fma.rn.f32 %f1429, %f1425, %f1428, %f1427; mul.f32 %f1430, %f1429, 0f3FB8AA3B; add.f32 %f1431, %f1425, 0f4B40007F; mov.b32 %r369, %f1431; shl.b32 %r370, %r369, 23; mov.b32 %f1432, %r370; ex2.approx.ftz.f32 %f1433, %f1430; mul.f32 %f294, %f1433, %f1432; setp.eq.f32 %p143, %f294, 0f7F800000; mov.f32 %f1734, 0f7F800000; @%p143 bra $L__BB0_118; fma.rn.f32 %f1734, %f294, %f293, %f294; $L__BB0_118: setp.lt.f32 %p144, %f278, 0f00000000; and.pred %p8, %p144, %p131; setp.eq.f32 %p146, %f278, 0f00000000; @%p146 bra $L__BB0_122; bra.uni $L__BB0_119; $L__BB0_122: add.f32 %f1438, %f278, %f278; selp.f32 %f1736, %f1438, 0f00000000, %p131; bra.uni $L__BB0_123; $L__BB0_119: mov.b32 %r371, %f1734; xor.b32 %r372, %r371, -2147483648; mov.b32 %f1434, %r372; selp.f32 %f1736, %f1434, %f1734, %p8; setp.geu.f32 %p147, %f278, 0f00000000; @%p147 bra $L__BB0_123; mov.f32 %f1435, 0f3ED55555; cvt.rzi.f32.f32 %f1436, %f1435; setp.eq.f32 %p148, %f1436, 0f3ED55555; @%p148 bra $L__BB0_123; mov.f32 %f1736, 0f7FFFFFFF; $L__BB0_123: add.f32 %f1439, %f292, 0f3ED55555; mov.b32 %r373, %f1439; setp.lt.s32 %p150, %r373, 2139095040; @%p150 bra $L__BB0_128; setp.gtu.f32 %p151, %f292, 0f7F800000; @%p151 bra $L__BB0_127; bra.uni $L__BB0_125; $L__BB0_127: add.f32 %f1736, %f278, 0f3ED55555; bra.uni $L__BB0_128; $L__BB0_125: setp.neu.f32 %p152, %f292, 0f7F800000; @%p152 bra $L__BB0_128; selp.f32 %f1736, 0fFF800000, 0f7F800000, %p8; $L__BB0_128: abs.f32 %f303, %f279; setp.lt.f32 %p153, %f303, 0f00800000; mul.f32 %f1441, %f303, 0f4B800000; selp.f32 %f1442, %f1441, %f303, %p153; selp.f32 %f1443, 0fC3170000, 0fC2FE0000, %p153; mov.b32 %r374, %f1442; and.b32 %r375, %r374, 8388607; or.b32 %r376, %r375, 1065353216; mov.b32 %f1444, %r376; shr.u32 %r377, %r374, 23; cvt.rn.f32.u32 %f1445, %r377; add.f32 %f1446, %f1443, %f1445; setp.gt.f32 %p154, %f1444, 0f3FB504F3; mul.f32 %f1447, %f1444, 0f3F000000; add.f32 %f1448, %f1446, 0f3F800000; selp.f32 %f1449, %f1448, %f1446, %p154; selp.f32 %f1450, %f1447, %f1444, %p154; add.f32 %f1451, %f1450, 0fBF800000; add.f32 %f1452, %f1450, 0f3F800000; rcp.approx.ftz.f32 %f1453, %f1452; add.f32 %f1454, %f1451, %f1451; mul.f32 %f1455, %f1454, %f1453; mul.f32 %f1456, %f1455, %f1455; mov.f32 %f1457, 0f3C4CAF63; mov.f32 %f1458, 0f3B18F0FE; fma.rn.f32 %f1459, %f1458, %f1456, %f1457; mov.f32 %f1460, 0f3DAAAABD; fma.rn.f32 %f1461, %f1459, %f1456, %f1460; mul.rn.f32 %f1462, %f1461, %f1456; mul.rn.f32 %f1463, %f1462, %f1455; sub.f32 %f1464, %f1451, %f1455; add.f32 %f1465, %f1464, %f1464; neg.f32 %f1466, %f1455; fma.rn.f32 %f1467, %f1466, %f1451, %f1465; mul.rn.f32 %f1468, %f1453, %f1467; add.f32 %f1469, %f1463, %f1455; sub.f32 %f1470, %f1455, %f1469; add.f32 %f1471, %f1463, %f1470; add.f32 %f1472, %f1468, %f1471; add.f32 %f1473, %f1469, %f1472; sub.f32 %f1474, %f1469, %f1473; add.f32 %f1475, %f1472, %f1474; mov.f32 %f1476, 0f3F317200; mul.rn.f32 %f1477, %f1449, %f1476; mov.f32 %f1478, 0f35BFBE8E; mul.rn.f32 %f1479, %f1449, %f1478; add.f32 %f1480, %f1477, %f1473; sub.f32 %f1481, %f1477, %f1480; add.f32 %f1482, %f1473, %f1481; add.f32 %f1483, %f1475, %f1482; add.f32 %f1484, %f1479, %f1483; add.f32 %f1485, %f1480, %f1484; sub.f32 %f1486, %f1480, %f1485; add.f32 %f1487, %f1484, %f1486; mov.f32 %f1488, 0f3ED55555; mul.rn.f32 %f1489, %f1488, %f1485; neg.f32 %f1490, %f1489; fma.rn.f32 %f1491, %f1488, %f1485, %f1490; fma.rn.f32 %f1492, %f1488, %f1487, %f1491; mov.f32 %f1493, 0f00000000; fma.rn.f32 %f1494, %f1493, %f1485, %f1492; add.rn.f32 %f1495, %f1489, %f1494; neg.f32 %f1496, %f1495; add.rn.f32 %f1497, %f1489, %f1496; add.rn.f32 %f1498, %f1497, %f1494; mov.b32 %r378, %f1495; setp.eq.s32 %p155, %r378, 1118925336; add.s32 %r379, %r378, -1; mov.b32 %f1499, %r379; add.f32 %f1500, %f1498, 0f37000000; selp.f32 %f304, %f1500, %f1498, %p155; selp.f32 %f1501, %f1499, %f1495, %p155; mov.f32 %f1502, 0f3FB8AA3B; mul.rn.f32 %f1503, %f1501, %f1502; cvt.rzi.f32.f32 %f1504, %f1503; abs.f32 %f1505, %f1504; setp.gt.f32 %p156, %f1505, 0f42FC0000; mov.b32 %r380, %f1504; and.b32 %r381, %r380, -2147483648; or.b32 %r382, %r381, 1123811328; mov.b32 %f1506, %r382; selp.f32 %f1507, %f1506, %f1504, %p156; mov.f32 %f1508, 0fBF317218; fma.rn.f32 %f1509, %f1507, %f1508, %f1501; mov.f32 %f1510, 0f3102E308; fma.rn.f32 %f1511, %f1507, %f1510, %f1509; mul.f32 %f1512, %f1511, 0f3FB8AA3B; add.f32 %f1513, %f1507, 0f4B40007F; mov.b32 %r383, %f1513; shl.b32 %r384, %r383, 23; mov.b32 %f1514, %r384; ex2.approx.ftz.f32 %f1515, %f1512; mul.f32 %f305, %f1515, %f1514; setp.eq.f32 %p157, %f305, 0f7F800000; mov.f32 %f1737, 0f7F800000; @%p157 bra $L__BB0_130; fma.rn.f32 %f1737, %f305, %f304, %f305; $L__BB0_130: setp.lt.f32 %p158, %f279, 0f00000000; and.pred %p9, %p158, %p131; setp.eq.f32 %p160, %f279, 0f00000000; @%p160 bra $L__BB0_134; bra.uni $L__BB0_131; $L__BB0_134: add.f32 %f1520, %f279, %f279; selp.f32 %f1739, %f1520, 0f00000000, %p131; bra.uni $L__BB0_135; $L__BB0_131: mov.b32 %r385, %f1737; xor.b32 %r386, %r385, -2147483648; mov.b32 %f1516, %r386; selp.f32 %f1739, %f1516, %f1737, %p9; setp.geu.f32 %p161, %f279, 0f00000000; @%p161 bra $L__BB0_135; mov.f32 %f1517, 0f3ED55555; cvt.rzi.f32.f32 %f1518, %f1517; setp.eq.f32 %p162, %f1518, 0f3ED55555; @%p162 bra $L__BB0_135; mov.f32 %f1739, 0f7FFFFFFF; $L__BB0_135: add.f32 %f1521, %f303, 0f3ED55555; mov.b32 %r387, %f1521; setp.lt.s32 %p164, %r387, 2139095040; @%p164 bra $L__BB0_140; setp.gtu.f32 %p165, %f303, 0f7F800000; @%p165 bra $L__BB0_139; bra.uni $L__BB0_137; $L__BB0_139: add.f32 %f1739, %f279, 0f3ED55555; bra.uni $L__BB0_140; $L__BB0_137: setp.neu.f32 %p166, %f303, 0f7F800000; @%p166 bra $L__BB0_140; selp.f32 %f1739, 0fFF800000, 0f7F800000, %p9; $L__BB0_140: fma.rn.f32 %f1522, %f1733, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p167, %f277, 0f3F800000; mov.f32 %f1523, 0f3F800000; selp.f32 %f1524, 0f3F7FFFFF, %f1522, %p167; mul.f32 %f1525, %f277, 0f414EB852; setp.lt.f32 %p168, %f277, 0f3B4D2E1C; selp.f32 %f1526, %f1525, %f1524, %p168; fma.rn.f32 %f1527, %f1736, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p169, %f278, 0f3F800000; selp.f32 %f1528, 0f3F7FFFFF, %f1527, %p169; mul.f32 %f1529, %f278, 0f414EB852; setp.lt.f32 %p170, %f278, 0f3B4D2E1C; selp.f32 %f1530, %f1529, %f1528, %p170; fma.rn.f32 %f1531, %f1739, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p171, %f279, 0f3F800000; selp.f32 %f1532, 0f3F7FFFFF, %f1531, %p171; mul.f32 %f1533, %f279, 0f414EB852; setp.lt.f32 %p172, %f279, 0f3B4D2E1C; selp.f32 %f1534, %f1533, %f1532, %p172; min.f32 %f1535, %f1526, %f1523; mov.f32 %f1536, 0f00000000; max.f32 %f1537, %f1536, %f1535; mul.f32 %f1538, %f1537, 0f43800000; cvt.rzi.u32.f32 %r388, %f1538; min.u32 %r389, %r388, 255; min.f32 %f1539, %f1530, %f1523; max.f32 %f1540, %f1536, %f1539; mul.f32 %f1541, %f1540, 0f43800000; cvt.rzi.u32.f32 %r390, %f1541; min.u32 %r391, %r390, 255; min.f32 %f1542, %f1534, %f1523; max.f32 %f1543, %f1536, %f1542; mul.f32 %f1544, %f1543, 0f43800000; cvt.rzi.u32.f32 %r392, %f1544; 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 %f1545, %rs60;} // end inline asm // begin inline asm { cvt.f32.f16 %f1546, %rs61;} // end inline asm // begin inline asm { cvt.f32.f16 %f1547, %rs62;} // end inline asm add.f32 %f1548, %f1693, %f1545; add.f32 %f1549, %f1694, %f1546; add.f32 %f1550, %f1695, %f1547; mov.f32 %f1551, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs58, %f1550;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs57, %f1549;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs56, %f1548;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs59, %f1551;} // end inline asm st.global.v4.u16 [%rd12], {%rs56, %rs57, %rs58, %rs59}; bra.uni $L__BB0_145; $L__BB0_144: mov.f32 %f1555, 0f3F800000; mov.u32 %r426, 0; // begin inline asm { cvt.rn.f16.f32 %rs67, %f1555;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs66, %f1695;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs65, %f1694;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs64, %f1693;} // 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 %f1556, %rs75;} // end inline asm // begin inline asm { cvt.f32.f16 %f1557, %rs76;} // end inline asm // begin inline asm { cvt.f32.f16 %f1558, %rs77;} // end inline asm add.f32 %f1559, %f1696, %f1556; add.f32 %f1560, %f1697, %f1557; add.f32 %f1561, %f1698, %f1558; mov.f32 %f1562, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs73, %f1561;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs72, %f1560;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs71, %f1559;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs74, %f1562;} // end inline asm st.global.v4.u16 [%rd13], {%rs71, %rs72, %rs73, %rs74}; bra.uni $L__BB0_148; $L__BB0_147: mov.f32 %f1566, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs82, %f1566;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs81, %f1698;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs80, %f1697;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs79, %f1696;} // 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 %f1567, %rs90;} // end inline asm // begin inline asm { cvt.f32.f16 %f1568, %rs91;} // end inline asm // begin inline asm { cvt.f32.f16 %f1569, %rs92;} // end inline asm add.f32 %f1570, %f1699, %f1567; add.f32 %f1571, %f1700, %f1568; add.f32 %f1572, %f1701, %f1569; mov.f32 %f1573, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs88, %f1572;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs87, %f1571;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs86, %f1570;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs89, %f1573;} // end inline asm st.global.v4.u16 [%rd14], {%rs86, %rs87, %rs88, %rs89}; bra.uni $L__BB0_151; $L__BB0_150: mov.f32 %f1577, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs97, %f1577;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs96, %f1701;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs95, %f1700;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs94, %f1699;} // 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 %f1578, %rs105;} // end inline asm // begin inline asm { cvt.f32.f16 %f1579, %rs106;} // end inline asm // begin inline asm { cvt.f32.f16 %f1580, %rs107;} // end inline asm add.f32 %f1581, %f1702, %f1578; add.f32 %f1582, %f1703, %f1579; add.f32 %f1583, %f1704, %f1580; mov.f32 %f1584, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs103, %f1583;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs102, %f1582;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs101, %f1581;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs104, %f1584;} // end inline asm st.global.v4.u16 [%rd15], {%rs101, %rs102, %rs103, %rs104}; bra.uni $L__BB0_171; $L__BB0_153: mov.f32 %f1588, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs112, %f1588;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs111, %f1704;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs110, %f1703;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs109, %f1702;} // end inline asm st.global.v4.u16 [%rd15], {%rs109, %rs110, %rs111, %rs112}; $L__BB0_171: ret; }