// // 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<210>; .reg .b16 %rs<180>; .reg .f32 %f<1721>; .reg .b32 %r<454>; .reg .b64 %rd<78>; // 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 %rd21, [params+400]; cvta.to.global.u64 %rd22, %rd21; ld.const.u32 %r32, [params+392]; mad.lo.s32 %r33, %r32, %r27, %r26; mul.wide.u32 %rd23, %r33, 4; add.s64 %rd1, %rd22, %rd23; ld.global.v2.u8 {%rs9, %rs178}, [%rd1]; or.b16 %rs11, %rs9, %rs178; and.b16 %rs12, %rs11, 255; setp.eq.s16 %p11, %rs12, 0; @%p11 bra $L__BB0_2; ld.global.u8 %rs177, [%rd1+2]; bra.uni $L__BB0_3; $L__BB0_2: ld.global.u8 %rs177, [%rd1+2]; setp.eq.s16 %p12, %rs177, 0; mov.f32 %f1614, 0f00000000; mov.u16 %rs178, 0; mov.f32 %f1615, %f1614; mov.f32 %f1616, %f1614; @%p12 bra $L__BB0_4; $L__BB0_3: cvt.rn.f32.u16 %f313, %rs9; div.rn.f32 %f314, %f313, 0f437F0000; fma.rn.f32 %f315, %f314, 0f40000000, 0fBF800000; and.b16 %rs15, %rs178, 255; cvt.rn.f32.u16 %f316, %rs15; div.rn.f32 %f317, %f316, 0f437F0000; fma.rn.f32 %f318, %f317, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f319, %rs177; div.rn.f32 %f320, %f319, 0f437F0000; fma.rn.f32 %f321, %f320, 0f40000000, 0fBF800000; mul.f32 %f322, %f318, %f318; fma.rn.f32 %f323, %f315, %f315, %f322; fma.rn.f32 %f324, %f321, %f321, %f323; sqrt.rn.f32 %f325, %f324; rcp.rn.f32 %f326, %f325; mul.f32 %f1616, %f326, %f321; mul.f32 %f1615, %f326, %f318; mul.f32 %f1614, %f315, %f326; $L__BB0_4: ld.const.v2.u32 {%r34, %r35}, [params]; add.s32 %r3, %r34, %r26; add.s32 %r4, %r35, %r27; setp.eq.f32 %p13, %f1614, 0f00000000; setp.eq.f32 %p14, %f1615, 0f00000000; and.pred %p15, %p13, %p14; setp.eq.f32 %p16, %f1616, 0f00000000; and.pred %p17, %p16, %p15; @%p17 bra $L__BB0_163; bra.uni $L__BB0_5; $L__BB0_163: ld.const.u32 %r23, [params+104]; and.b32 %r428, %r23, 1; setp.eq.b32 %p198, %r428, 1; mov.pred %p199, 0; xor.pred %p200, %p198, %p199; not.pred %p201, %p200; @%p201 bra $L__BB0_165; ld.const.u64 %rd54, [params+144]; cvta.to.global.u64 %rd55, %rd54; ld.const.u32 %r429, [params+136]; mad.lo.s32 %r430, %r429, %r4, %r3; mul.wide.u32 %rd56, %r430, 4; add.s64 %rd57, %rd55, %rd56; mov.u16 %rs99, 0; st.global.v4.u8 [%rd57], {%rs99, %rs99, %rs99, %rs99}; $L__BB0_165: and.b32 %r431, %r23, 8; setp.eq.s32 %p202, %r431, 0; @%p202 bra $L__BB0_167; ld.const.u64 %rd58, [params+192]; cvta.to.global.u64 %rd59, %rd58; ld.const.u32 %r432, [params+184]; mad.lo.s32 %r433, %r432, %r4, %r3; mov.f32 %f1565, 0f00000000; cvt.rzi.u32.f32 %r434, %f1565; mul.wide.u32 %rd60, %r433, 2; add.s64 %rd61, %rd59, %rd60; mov.u16 %rs100, 0; cvt.u16.u32 %rs101, %r434; st.global.v2.u8 [%rd61], {%rs101, %rs100}; $L__BB0_167: and.b32 %r435, %r23, 4; setp.eq.s32 %p203, %r435, 0; ld.const.u32 %r453, [params+108]; @%p203 bra $L__BB0_171; setp.eq.s32 %p204, %r453, 0; ld.const.u64 %rd62, [params+224]; cvta.to.global.u64 %rd63, %rd62; ld.const.u32 %r436, [params+216]; mad.lo.s32 %r437, %r436, %r4, %r3; mul.wide.u32 %rd64, %r437, 8; add.s64 %rd16, %rd63, %rd64; @%p204 bra $L__BB0_170; ld.global.v4.u16 {%rs108, %rs109, %rs110, %rs111}, [%rd16]; // begin inline asm { cvt.f32.f16 %f1566, %rs108;} // end inline asm // begin inline asm { cvt.f32.f16 %f1567, %rs109;} // end inline asm // begin inline asm { cvt.f32.f16 %f1568, %rs110;} // 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 %rs107, %f1571;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs106, %f1570;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs105, %f1569;} // end inline asm mov.u16 %rs112, 0; st.global.v4.u16 [%rd16], {%rs105, %rs106, %rs107, %rs112}; bra.uni $L__BB0_171; $L__BB0_5: ld.const.u64 %rd24, [params+432]; cvta.to.global.u64 %rd25, %rd24; ld.const.u32 %r38, [params+424]; mad.lo.s32 %r39, %r38, %r27, %r26; mul.wide.u32 %rd26, %r39, 12; add.s64 %rd27, %rd25, %rd26; ld.global.f32 %f10, [%rd27]; mul.f32 %f327, %f10, 0f3456BF95; ld.global.f32 %f11, [%rd27+4]; mul.f32 %f328, %f11, 0f3456BF95; ld.global.f32 %f12, [%rd27+8]; mul.f32 %f329, %f12, 0f3456BF95; abs.f32 %f330, %f1614; div.rn.f32 %f331, %f327, %f330; abs.f32 %f332, %f1615; div.rn.f32 %f333, %f328, %f332; abs.f32 %f334, %f1616; div.rn.f32 %f335, %f329, %f334; abs.f32 %f336, %f331; abs.f32 %f337, %f333; abs.f32 %f338, %f335; mov.f32 %f339, 0f38D1B717; max.f32 %f340, %f336, %f339; max.f32 %f341, %f337, %f339; max.f32 %f342, %f338, %f339; fma.rn.f32 %f13, %f1614, %f340, %f10; fma.rn.f32 %f14, %f1615, %f341, %f11; fma.rn.f32 %f15, %f1616, %f342, %f12; ld.const.u32 %r5, [params+600]; setp.eq.s32 %p18, %r5, 0; @%p18 bra $L__BB0_58; ld.const.u64 %rd28, [params+608]; cvta.to.global.u64 %rd2, %rd28; mov.f32 %f359, 0f40000000; cvt.rzi.f32.f32 %f360, %f359; add.f32 %f361, %f360, %f360; mov.f32 %f362, 0f40800000; sub.f32 %f363, %f362, %f361; abs.f32 %f16, %f363; ld.const.u32 %r6, [params+544]; ld.const.u32 %r451, [params+104]; mul.f32 %f17, %f13, 0f3456BF95; mul.f32 %f18, %f14, 0f3456BF95; mul.f32 %f19, %f15, 0f3456BF95; mov.f32 %f358, 0f00000000; mov.u32 %r447, 0; ld.const.u64 %rd3, [params+96]; abs.f32 %f565, %f17; abs.f32 %f566, %f18; max.f32 %f567, %f565, %f566; abs.f32 %f568, %f19; max.f32 %f569, %f567, %f568; mov.f32 %f1683, %f358; mov.f32 %f1682, %f358; mov.f32 %f1681, %f358; mov.f32 %f1680, %f358; mov.f32 %f1679, %f358; mov.f32 %f1678, %f358; mov.f32 %f1677, %f358; mov.f32 %f1676, %f358; mov.f32 %f1675, %f358; mov.f32 %f1674, %f358; mov.f32 %f1673, %f358; mov.f32 %f1672, %f358; mov.f32 %f1671, %f358; mov.f32 %f1670, %f358; mov.f32 %f1669, %f358; mov.f32 %f1668, %f358; $L__BB0_7: mul.wide.u32 %rd29, %r447, 112; add.s64 %rd5, %rd2, %rd29; ld.global.v4.f32 {%f365, %f366, %f367, %f368}, [%rd5]; ld.global.v4.f32 {%f369, %f370, %f371, %f372}, [%rd5+16]; ld.global.v4.f32 {%f375, %f1641, %f1642, %f378}, [%rd5+32]; ld.global.v4.f32 {%f379, %f380, %f381, %f1639}, [%rd5+48]; ld.global.v4.f32 {%f383, %f384, %f385, %f386}, [%rd5+64]; ld.global.v4.u32 {%r41, %r42, %r43, %r44}, [%rd5+80]; mov.b32 %f55, %r43; mov.b32 %f54, %r42; mov.b32 %f53, %r41; ld.global.u64 %rd6, [%rd5+96]; sub.f32 %f387, %f366, %f10; sub.f32 %f388, %f367, %f11; sub.f32 %f389, %f368, %f12; mul.f32 %f390, %f388, %f388; fma.rn.f32 %f391, %f387, %f387, %f390; fma.rn.f32 %f392, %f389, %f389, %f391; sqrt.rn.f32 %f393, %f392; rcp.rn.f32 %f394, %f393; mul.f32 %f57, %f387, %f394; mul.f32 %f58, %f388, %f394; mul.f32 %f59, %f389, %f394; mul.f32 %f60, %f393, %f370; mul.f32 %f61, %f393, %f372; abs.f32 %f62, %f61; setp.lt.f32 %p19, %f62, 0f00800000; mul.f32 %f395, %f62, 0f4B800000; selp.f32 %f396, %f395, %f62, %p19; selp.f32 %f397, 0fC3170000, 0fC2FE0000, %p19; mov.b32 %r48, %f396; and.b32 %r49, %r48, 8388607; or.b32 %r50, %r49, 1065353216; mov.b32 %f398, %r50; shr.u32 %r51, %r48, 23; cvt.rn.f32.u32 %f399, %r51; add.f32 %f400, %f397, %f399; setp.gt.f32 %p20, %f398, 0f3FB504F3; mul.f32 %f401, %f398, 0f3F000000; add.f32 %f402, %f400, 0f3F800000; selp.f32 %f403, %f402, %f400, %p20; selp.f32 %f404, %f401, %f398, %p20; add.f32 %f405, %f404, 0fBF800000; add.f32 %f406, %f404, 0f3F800000; rcp.approx.ftz.f32 %f407, %f406; add.f32 %f408, %f405, %f405; mul.f32 %f409, %f408, %f407; mul.f32 %f410, %f409, %f409; mov.f32 %f411, 0f3C4CAF63; mov.f32 %f412, 0f3B18F0FE; fma.rn.f32 %f413, %f412, %f410, %f411; mov.f32 %f414, 0f3DAAAABD; fma.rn.f32 %f415, %f413, %f410, %f414; mul.rn.f32 %f416, %f415, %f410; mul.rn.f32 %f417, %f416, %f409; sub.f32 %f418, %f405, %f409; add.f32 %f419, %f418, %f418; neg.f32 %f420, %f409; fma.rn.f32 %f421, %f420, %f405, %f419; mul.rn.f32 %f422, %f407, %f421; add.f32 %f423, %f417, %f409; sub.f32 %f424, %f409, %f423; add.f32 %f425, %f417, %f424; add.f32 %f426, %f422, %f425; add.f32 %f427, %f423, %f426; sub.f32 %f428, %f423, %f427; add.f32 %f429, %f426, %f428; mov.f32 %f430, 0f3F317200; mul.rn.f32 %f431, %f403, %f430; mov.f32 %f432, 0f35BFBE8E; mul.rn.f32 %f433, %f403, %f432; add.f32 %f434, %f431, %f427; sub.f32 %f435, %f431, %f434; add.f32 %f436, %f427, %f435; add.f32 %f437, %f429, %f436; add.f32 %f438, %f433, %f437; add.f32 %f439, %f434, %f438; sub.f32 %f440, %f434, %f439; add.f32 %f441, %f438, %f440; mul.rn.f32 %f443, %f362, %f439; neg.f32 %f444, %f443; fma.rn.f32 %f445, %f362, %f439, %f444; fma.rn.f32 %f446, %f362, %f441, %f445; fma.rn.f32 %f448, %f358, %f439, %f446; add.rn.f32 %f449, %f443, %f448; neg.f32 %f450, %f449; add.rn.f32 %f451, %f443, %f450; add.rn.f32 %f452, %f451, %f448; mov.b32 %r52, %f449; setp.eq.s32 %p21, %r52, 1118925336; add.s32 %r53, %r52, -1; mov.b32 %f453, %r53; add.f32 %f454, %f452, 0f37000000; selp.f32 %f63, %f454, %f452, %p21; selp.f32 %f455, %f453, %f449, %p21; mov.f32 %f456, 0f3FB8AA3B; mul.rn.f32 %f457, %f455, %f456; cvt.rzi.f32.f32 %f458, %f457; abs.f32 %f459, %f458; setp.gt.f32 %p22, %f459, 0f42FC0000; mov.b32 %r54, %f458; and.b32 %r55, %r54, -2147483648; or.b32 %r56, %r55, 1123811328; mov.b32 %f460, %r56; selp.f32 %f461, %f460, %f458, %p22; mov.f32 %f462, 0fBF317218; fma.rn.f32 %f463, %f461, %f462, %f455; mov.f32 %f464, 0f3102E308; fma.rn.f32 %f465, %f461, %f464, %f463; mul.f32 %f466, %f465, 0f3FB8AA3B; add.f32 %f467, %f461, 0f4B40007F; mov.b32 %r57, %f467; shl.b32 %r58, %r57, 23; mov.b32 %f468, %r58; ex2.approx.ftz.f32 %f469, %f466; mul.f32 %f64, %f469, %f468; setp.eq.f32 %p23, %f64, 0f7F800000; mov.f32 %f1633, 0f7F800000; @%p23 bra $L__BB0_9; fma.rn.f32 %f1633, %f64, %f63, %f64; $L__BB0_9: setp.lt.f32 %p24, %f61, 0f00000000; setp.eq.f32 %p25, %f16, 0f3F800000; and.pred %p1, %p24, %p25; setp.eq.f32 %p26, %f61, 0f00000000; @%p26 bra $L__BB0_13; bra.uni $L__BB0_10; $L__BB0_13: add.f32 %f474, %f61, %f61; selp.f32 %f1635, %f474, 0f00000000, %p25; bra.uni $L__BB0_14; $L__BB0_10: mov.b32 %r59, %f1633; xor.b32 %r60, %r59, -2147483648; mov.b32 %f470, %r60; selp.f32 %f1635, %f470, %f1633, %p1; setp.geu.f32 %p27, %f61, 0f00000000; @%p27 bra $L__BB0_14; mov.f32 %f471, 0f40800000; cvt.rzi.f32.f32 %f472, %f471; setp.eq.f32 %p28, %f472, 0f40800000; @%p28 bra $L__BB0_14; mov.f32 %f1635, 0f7FFFFFFF; $L__BB0_14: abs.f32 %f1611, %f61; add.f32 %f475, %f1611, 0f40800000; mov.b32 %r61, %f475; setp.lt.s32 %p30, %r61, 2139095040; @%p30 bra $L__BB0_19; abs.f32 %f1612, %f61; setp.gtu.f32 %p31, %f1612, 0f7F800000; @%p31 bra $L__BB0_18; bra.uni $L__BB0_16; $L__BB0_18: add.f32 %f1635, %f61, 0f40800000; bra.uni $L__BB0_19; $L__BB0_16: abs.f32 %f1613, %f61; setp.neu.f32 %p32, %f1613, 0f7F800000; @%p32 bra $L__BB0_19; selp.f32 %f1635, 0fFF800000, 0f7F800000, %p1; $L__BB0_19: mov.f32 %f476, 0f3F800000; sub.f32 %f477, %f476, %f1635; setp.eq.f32 %p33, %f61, 0f3F800000; selp.f32 %f478, 0f00000000, %f477, %p33; cvt.sat.f32.f32 %f479, %f478; fma.rn.f32 %f480, %f60, %f60, %f371; div.rn.f32 %f73, %f479, %f480; mul.f32 %f481, %f1615, %f58; fma.rn.f32 %f482, %f1614, %f57, %f481; fma.rn.f32 %f483, %f1616, %f59, %f482; setp.eq.s32 %p34, %r6, 0; selp.f32 %f74, %f483, 0f3F800000, %p34; mul.f32 %f484, %f74, 0f40800000; cvt.sat.f32.f32 %f75, %f484; setp.eq.f32 %p35, %f365, 0f3F800000; @%p35 bra $L__BB0_25; bra.uni $L__BB0_20; $L__BB0_25: setp.eq.s64 %p39, %rd6, 0; mov.f32 %f1640, %f375; mov.f32 %f1643, %f73; @%p39 bra $L__BB0_29; setp.geu.f32 %p40, %f1639, 0f00000000; mov.f32 %f1636, %f57; mov.f32 %f1637, %f58; mov.f32 %f1638, %f59; @%p40 bra $L__BB0_28; mul.f32 %f509, %f54, %f388; fma.rn.f32 %f510, %f53, %f387, %f509; fma.rn.f32 %f512, %f55, %f389, %f510; rcp.rn.f32 %f513, %f512; mul.f32 %f1636, %f387, %f513; mul.f32 %f1637, %f388, %f513; mul.f32 %f1638, %f389, %f513; neg.f32 %f1639, %f1639; $L__BB0_28: mul.f32 %f514, %f55, %f59; mul.f32 %f515, %f54, %f58; neg.f32 %f516, %f515; mul.f32 %f517, %f53, %f57; sub.f32 %f518, %f516, %f517; sub.f32 %f519, %f518, %f514; setp.gt.f32 %p41, %f519, 0f00000000; selp.f32 %f520, 0f3F800000, 0f00000000, %p41; mov.f32 %f521, 0f3F800000; mul.f32 %f522, %f380, %f1637; mul.f32 %f523, %f384, %f1637; fma.rn.f32 %f524, %f379, %f1636, %f522; fma.rn.f32 %f525, %f383, %f1636, %f523; fma.rn.f32 %f526, %f381, %f1638, %f524; fma.rn.f32 %f527, %f385, %f1638, %f525; fma.rn.f32 %f528, %f1639, %f526, 0f3F000000; sub.f32 %f529, %f521, %f528; fma.rn.f32 %f530, %f1639, %f527, 0f3F000000; tex.2d.v4.f32.f32 {%f531, %f532, %f533, %f534}, [%rd6, {%f529, %f530}]; mul.f32 %f535, %f520, %f531; mul.f32 %f536, %f520, %f532; mul.f32 %f537, %f520, %f533; mul.f32 %f1640, %f375, %f535; mul.f32 %f1641, %f1641, %f536; mul.f32 %f1642, %f1642, %f537; mov.f32 %f1643, %f73; bra.uni $L__BB0_29; $L__BB0_20: setp.eq.f32 %p36, %f365, 0f40000000; @%p36 bra $L__BB0_23; bra.uni $L__BB0_21; $L__BB0_23: setp.eq.s64 %p38, %rd6, 0; mov.f32 %f1640, %f375; mov.f32 %f1643, %f73; @%p38 bra $L__BB0_29; mul.f32 %f494, %f380, %f58; mul.f32 %f495, %f384, %f58; mul.f32 %f496, %f54, %f58; fma.rn.f32 %f497, %f379, %f57, %f494; fma.rn.f32 %f498, %f383, %f57, %f495; fma.rn.f32 %f499, %f53, %f57, %f496; fma.rn.f32 %f500, %f381, %f59, %f497; fma.rn.f32 %f501, %f385, %f59, %f498; fma.rn.f32 %f502, %f55, %f59, %f499; tex.cube.v4.f32.f32 {%f503, %f504, %f505, %f506}, [%rd6, {%f500, %f501, %f502, %f502}]; mul.f32 %f1640, %f375, %f503; mul.f32 %f1641, %f1641, %f504; mul.f32 %f1642, %f1642, %f505; mov.f32 %f1643, %f73; bra.uni $L__BB0_29; $L__BB0_21: setp.neu.f32 %p37, %f365, 0f40800000; mov.f32 %f1640, %f375; mov.f32 %f1643, %f73; @%p37 bra $L__BB0_29; mul.f32 %f485, %f53, %f57; mul.f32 %f486, %f54, %f58; neg.f32 %f487, %f486; sub.f32 %f488, %f487, %f485; mul.f32 %f489, %f55, %f59; sub.f32 %f490, %f488, %f489; fma.rn.f32 %f491, %f378, %f490, %f1639; cvt.sat.f32.f32 %f492, %f491; mul.f32 %f493, %f492, %f492; mul.f32 %f1643, %f73, %f493; mov.f32 %f1640, %f375; $L__BB0_29: max.f32 %f553, %f1640, %f1641; max.f32 %f554, %f553, %f1642; mul.f32 %f555, %f75, %f1643; mul.f32 %f556, %f555, %f554; setp.lt.f32 %p43, %f556, 0f3727C5AC; mov.pred %p209, -1; mov.f32 %f1644, 0f00000000; mov.f32 %f1645, %f1644; mov.f32 %f1646, %f1644; mov.f32 %f1647, %f1644; mov.f32 %f1648, %f1644; mov.f32 %f1649, %f1644; mov.f32 %f1650, %f1644; mov.f32 %f1651, %f1644; mov.f32 %f1652, %f1644; mov.f32 %f1653, %f1644; mov.f32 %f1654, %f1644; mov.f32 %f1655, %f1644; mov.f32 %f1656, %f1644; mov.f32 %f1657, %f1644; mov.f32 %f1658, %f1644; @%p43 bra $L__BB0_31; cvt.sat.f32.f32 %f557, %f74; mul.f32 %f558, %f1643, %f557; mul.f32 %f1645, %f1640, %f558; mul.f32 %f1644, %f1641, %f558; mul.f32 %f1646, %f1642, %f558; mul.f32 %f559, %f1643, 0f3E800000; mul.f32 %f560, %f75, %f559; mul.f32 %f1647, %f1640, %f560; mul.f32 %f1648, %f1641, %f560; mul.f32 %f1649, %f1642, %f560; mul.f32 %f1650, %f57, %f1647; mul.f32 %f1651, %f57, %f1648; mul.f32 %f1652, %f57, %f1649; mul.f32 %f1653, %f58, %f1647; mul.f32 %f1654, %f58, %f1648; mul.f32 %f1655, %f58, %f1649; mul.f32 %f1656, %f59, %f1647; mul.f32 %f1657, %f59, %f1648; mul.f32 %f1658, %f59, %f1649; mov.pred %p209, 0; $L__BB0_31: @%p209 bra $L__BB0_57; setp.eq.s32 %p45, %r44, 0; mov.u16 %rs179, 0; mov.f32 %f1667, 0f3F800000; @%p45 bra $L__BB0_40; abs.s32 %r10, %r44; setp.lt.s32 %p46, %r10, 1; mov.f32 %f1660, 0f00000000; @%p46 bra $L__BB0_39; mov.f32 %f570, 0f38D1B717; max.f32 %f571, %f569, %f570; and.b32 %r63, %r451, 32; setp.eq.s32 %p47, %r63, 0; mov.u32 %r450, 0; selp.f32 %f572, 0f3F800000, 0f41200000, %p47; mul.f32 %f138, %f572, %f571; and.b32 %r11, %r10, 1; setp.eq.s32 %p48, %r10, 1; mov.f32 %f1660, 0f00000000; @%p48 bra $L__BB0_37; sub.s32 %r449, %r10, %r11; $L__BB0_36: cvt.rn.f32.s32 %f592, %r450; mul.f32 %f593, %f592, 0f3DD32618; cvt.rmi.f32.f32 %f594, %f593; sub.f32 %f595, %f593, %f594; mul.f32 %f596, %f592, 0f3DD2F1AA; cvt.rmi.f32.f32 %f597, %f596; sub.f32 %f598, %f596, %f597; mul.f32 %f599, %f592, 0f3DC74539; cvt.rmi.f32.f32 %f600, %f599; sub.f32 %f601, %f599, %f600; add.f32 %f602, %f598, 0f4199851F; add.f32 %f603, %f601, 0f4199851F; add.f32 %f604, %f595, 0f4199851F; mul.f32 %f605, %f598, %f603; fma.rn.f32 %f606, %f595, %f602, %f605; fma.rn.f32 %f607, %f604, %f601, %f606; add.f32 %f608, %f595, %f607; add.f32 %f609, %f598, %f607; add.f32 %f610, %f601, %f607; add.f32 %f611, %f608, %f609; mul.f32 %f612, %f610, %f611; cvt.rmi.f32.f32 %f613, %f612; sub.f32 %f614, %f612, %f613; add.f32 %f615, %f608, %f610; mul.f32 %f616, %f609, %f615; cvt.rmi.f32.f32 %f617, %f616; sub.f32 %f618, %f616, %f617; add.f32 %f619, %f609, %f610; mul.f32 %f620, %f608, %f619; cvt.rmi.f32.f32 %f621, %f620; sub.f32 %f622, %f620, %f621; fma.rn.f32 %f623, %f614, 0f40000000, 0fBF800000; fma.rn.f32 %f624, %f618, 0f40000000, 0fBF800000; fma.rn.f32 %f625, %f622, 0f40000000, 0fBF800000; ld.global.f32 %f626, [%rd5+16]; ld.global.f32 %f627, [%rd5+4]; fma.rn.f32 %f628, %f626, %f623, %f627; ld.global.f32 %f629, [%rd5+8]; fma.rn.f32 %f630, %f626, %f624, %f629; ld.global.f32 %f631, [%rd5+12]; fma.rn.f32 %f632, %f626, %f625, %f631; sub.f32 %f633, %f628, %f10; sub.f32 %f634, %f630, %f11; sub.f32 %f635, %f632, %f12; mul.f32 %f636, %f634, %f634; fma.rn.f32 %f637, %f633, %f633, %f636; fma.rn.f32 %f638, %f635, %f635, %f637; sqrt.rn.f32 %f581, %f638; rcp.rn.f32 %f639, %f581; mul.f32 %f577, %f639, %f633; mul.f32 %f578, %f639, %f634; mul.f32 %f579, %f639, %f635; mov.f32 %f591, 0f00000000; mov.u32 %r172, 2; mov.u32 %r174, 1; mov.u32 %r175, 1065353216; mov.u32 %r206, 0; // begin inline asm call(%r65,%r66,%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),_optix_trace_typed_32,(%r206,%rd3,%f13,%f14,%f15,%f577,%f578,%f579,%f138,%f581,%f591,%r174,%r206,%r174,%r172,%r174,%r174,%r175,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206); // end inline asm mov.b32 %f640, %r65; add.f32 %f641, %f1660, %f640; add.s32 %r207, %r450, 1; cvt.rn.f32.s32 %f642, %r207; mul.f32 %f643, %f642, 0f3DD32618; cvt.rmi.f32.f32 %f644, %f643; sub.f32 %f645, %f643, %f644; mul.f32 %f646, %f642, 0f3DD2F1AA; cvt.rmi.f32.f32 %f647, %f646; sub.f32 %f648, %f646, %f647; mul.f32 %f649, %f642, 0f3DC74539; cvt.rmi.f32.f32 %f650, %f649; sub.f32 %f651, %f649, %f650; add.f32 %f652, %f648, 0f4199851F; add.f32 %f653, %f651, 0f4199851F; add.f32 %f654, %f645, 0f4199851F; mul.f32 %f655, %f648, %f653; fma.rn.f32 %f656, %f645, %f652, %f655; fma.rn.f32 %f657, %f654, %f651, %f656; add.f32 %f658, %f645, %f657; add.f32 %f659, %f648, %f657; add.f32 %f660, %f651, %f657; add.f32 %f661, %f658, %f659; mul.f32 %f662, %f660, %f661; cvt.rmi.f32.f32 %f663, %f662; sub.f32 %f664, %f662, %f663; add.f32 %f665, %f658, %f660; mul.f32 %f666, %f659, %f665; cvt.rmi.f32.f32 %f667, %f666; sub.f32 %f668, %f666, %f667; add.f32 %f669, %f659, %f660; mul.f32 %f670, %f658, %f669; cvt.rmi.f32.f32 %f671, %f670; sub.f32 %f672, %f670, %f671; fma.rn.f32 %f673, %f664, 0f40000000, 0fBF800000; fma.rn.f32 %f674, %f668, 0f40000000, 0fBF800000; fma.rn.f32 %f675, %f672, 0f40000000, 0fBF800000; ld.global.f32 %f676, [%rd5+16]; ld.global.f32 %f677, [%rd5+4]; fma.rn.f32 %f678, %f676, %f673, %f677; ld.global.f32 %f679, [%rd5+8]; fma.rn.f32 %f680, %f676, %f674, %f679; ld.global.f32 %f681, [%rd5+12]; fma.rn.f32 %f682, %f676, %f675, %f681; sub.f32 %f683, %f678, %f10; sub.f32 %f684, %f680, %f11; sub.f32 %f685, %f682, %f12; mul.f32 %f686, %f684, %f684; fma.rn.f32 %f687, %f683, %f683, %f686; fma.rn.f32 %f688, %f685, %f685, %f687; sqrt.rn.f32 %f590, %f688; rcp.rn.f32 %f689, %f590; mul.f32 %f586, %f689, %f683; mul.f32 %f587, %f689, %f684; mul.f32 %f588, %f689, %f685; // begin inline asm call(%r136,%r137,%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),_optix_trace_typed_32,(%r206,%rd3,%f13,%f14,%f15,%f586,%f587,%f588,%f138,%f590,%f591,%r174,%r206,%r174,%r172,%r174,%r174,%r175,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206,%r206); // end inline asm mov.b32 %f690, %r136; add.f32 %f1660, %f641, %f690; add.s32 %r450, %r450, 2; add.s32 %r449, %r449, -2; setp.ne.s32 %p49, %r449, 0; @%p49 bra $L__BB0_36; $L__BB0_37: setp.eq.s32 %p50, %r11, 0; @%p50 bra $L__BB0_39; cvt.rn.f32.s32 %f700, %r450; mul.f32 %f701, %f700, 0f3DD32618; cvt.rmi.f32.f32 %f702, %f701; sub.f32 %f703, %f701, %f702; mul.f32 %f704, %f700, 0f3DD2F1AA; cvt.rmi.f32.f32 %f705, %f704; sub.f32 %f706, %f704, %f705; mul.f32 %f707, %f700, 0f3DC74539; cvt.rmi.f32.f32 %f708, %f707; sub.f32 %f709, %f707, %f708; add.f32 %f710, %f706, 0f4199851F; add.f32 %f711, %f709, 0f4199851F; add.f32 %f712, %f703, 0f4199851F; mul.f32 %f713, %f706, %f711; fma.rn.f32 %f714, %f703, %f710, %f713; fma.rn.f32 %f715, %f712, %f709, %f714; add.f32 %f716, %f703, %f715; add.f32 %f717, %f706, %f715; add.f32 %f718, %f709, %f715; add.f32 %f719, %f716, %f717; mul.f32 %f720, %f718, %f719; cvt.rmi.f32.f32 %f721, %f720; sub.f32 %f722, %f720, %f721; add.f32 %f723, %f716, %f718; mul.f32 %f724, %f717, %f723; cvt.rmi.f32.f32 %f725, %f724; sub.f32 %f726, %f724, %f725; add.f32 %f727, %f717, %f718; mul.f32 %f728, %f716, %f727; cvt.rmi.f32.f32 %f729, %f728; sub.f32 %f730, %f728, %f729; fma.rn.f32 %f731, %f722, 0f40000000, 0fBF800000; fma.rn.f32 %f732, %f726, 0f40000000, 0fBF800000; fma.rn.f32 %f733, %f730, 0f40000000, 0fBF800000; ld.global.f32 %f734, [%rd5+16]; ld.global.f32 %f735, [%rd5+4]; fma.rn.f32 %f736, %f734, %f731, %f735; ld.global.f32 %f737, [%rd5+8]; fma.rn.f32 %f738, %f734, %f732, %f737; ld.global.f32 %f739, [%rd5+12]; fma.rn.f32 %f740, %f734, %f733, %f739; sub.f32 %f741, %f736, %f10; sub.f32 %f742, %f738, %f11; sub.f32 %f743, %f740, %f12; mul.f32 %f744, %f742, %f742; fma.rn.f32 %f745, %f741, %f741, %f744; fma.rn.f32 %f746, %f743, %f743, %f745; sqrt.rn.f32 %f698, %f746; rcp.rn.f32 %f747, %f698; mul.f32 %f694, %f747, %f741; mul.f32 %f695, %f747, %f742; mul.f32 %f696, %f747, %f743; mov.f32 %f699, 0f00000000; mov.u32 %r244, 2; mov.u32 %r246, 1; mov.u32 %r247, 1065353216; mov.u32 %r278, 0; // begin inline asm call(%r208,%r209,%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),_optix_trace_typed_32,(%r278,%rd3,%f13,%f14,%f15,%f694,%f695,%f696,%f138,%f698,%f699,%r246,%r278,%r246,%r244,%r246,%r246,%r247,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278,%r278); // end inline asm mov.b32 %f748, %r208; add.f32 %f1660, %f1660, %f748; $L__BB0_39: cvt.rn.f32.s32 %f749, %r10; div.rn.f32 %f1667, %f1660, %f749; shr.u32 %r279, %r44, 31; cvt.u16.u32 %rs179, %r279; $L__BB0_40: fma.rn.f32 %f1668, %f1645, %f1667, %f1668; fma.rn.f32 %f1669, %f1644, %f1667, %f1669; fma.rn.f32 %f1670, %f1646, %f1667, %f1670; fma.rn.f32 %f1671, %f1647, %f1667, %f1671; fma.rn.f32 %f1672, %f1648, %f1667, %f1672; fma.rn.f32 %f1673, %f1649, %f1667, %f1673; fma.rn.f32 %f1674, %f1650, %f1667, %f1674; fma.rn.f32 %f1675, %f1651, %f1667, %f1675; fma.rn.f32 %f1676, %f1652, %f1667, %f1676; fma.rn.f32 %f1677, %f1653, %f1667, %f1677; fma.rn.f32 %f1678, %f1654, %f1667, %f1678; fma.rn.f32 %f1679, %f1655, %f1667, %f1679; fma.rn.f32 %f1680, %f1656, %f1667, %f1680; fma.rn.f32 %f1681, %f1657, %f1667, %f1681; fma.rn.f32 %f1682, %f1658, %f1667, %f1682; setp.eq.s16 %p51, %rs179, 0; @%p51 bra $L__BB0_56; mul.f32 %f751, %f1650, 0f3F000000; mul.f32 %f752, %f1653, 0f3F000000; mul.f32 %f753, %f752, %f752; fma.rn.f32 %f754, %f751, %f751, %f753; mul.f32 %f755, %f1656, 0f3F000000; fma.rn.f32 %f756, %f755, %f755, %f754; sqrt.rn.f32 %f757, %f756; rcp.rn.f32 %f758, %f757; mov.f32 %f759, 0f3F800000; mul.f32 %f760, %f751, %f758; mul.f32 %f761, %f752, %f758; mul.f32 %f762, %f755, %f758; mul.f32 %f763, %f1615, %f761; fma.rn.f32 %f764, %f1614, %f760, %f763; fma.rn.f32 %f765, %f1616, %f762, %f764; fma.rn.f32 %f162, %f765, 0f3F000000, 0f3F000000; add.f32 %f766, %f757, %f757; div.rn.f32 %f767, %f766, %f1647; add.f32 %f163, %f767, 0f3F800000; div.rn.f32 %f768, %f757, %f1647; sub.f32 %f769, %f759, %f768; add.f32 %f770, %f768, 0f3F800000; div.rn.f32 %f164, %f769, %f770; sub.f32 %f771, %f759, %f164; add.f32 %f772, %f163, 0f3F800000; mul.f32 %f165, %f771, %f772; mul.f32 %f773, %f163, 0f3F000000; cvt.rzi.f32.f32 %f774, %f773; add.f32 %f775, %f774, %f774; sub.f32 %f776, %f163, %f775; abs.f32 %f166, %f776; abs.f32 %f167, %f162; setp.lt.f32 %p52, %f167, 0f00800000; mul.f32 %f777, %f167, 0f4B800000; selp.f32 %f778, %f777, %f167, %p52; selp.f32 %f779, 0fC3170000, 0fC2FE0000, %p52; mov.b32 %r280, %f778; and.b32 %r281, %r280, 8388607; or.b32 %r282, %r281, 1065353216; mov.b32 %f780, %r282; shr.u32 %r283, %r280, 23; cvt.rn.f32.u32 %f781, %r283; add.f32 %f782, %f779, %f781; setp.gt.f32 %p53, %f780, 0f3FB504F3; mul.f32 %f783, %f780, 0f3F000000; add.f32 %f784, %f782, 0f3F800000; selp.f32 %f785, %f784, %f782, %p53; selp.f32 %f786, %f783, %f780, %p53; add.f32 %f787, %f786, 0fBF800000; add.f32 %f788, %f786, 0f3F800000; rcp.approx.ftz.f32 %f789, %f788; add.f32 %f790, %f787, %f787; mul.f32 %f791, %f790, %f789; mul.f32 %f792, %f791, %f791; mov.f32 %f793, 0f3C4CAF63; mov.f32 %f794, 0f3B18F0FE; fma.rn.f32 %f795, %f794, %f792, %f793; mov.f32 %f796, 0f3DAAAABD; fma.rn.f32 %f797, %f795, %f792, %f796; mul.rn.f32 %f798, %f797, %f792; mul.rn.f32 %f799, %f798, %f791; sub.f32 %f800, %f787, %f791; add.f32 %f801, %f800, %f800; neg.f32 %f802, %f791; fma.rn.f32 %f803, %f802, %f787, %f801; mul.rn.f32 %f804, %f789, %f803; add.f32 %f805, %f799, %f791; sub.f32 %f806, %f791, %f805; add.f32 %f807, %f799, %f806; add.f32 %f808, %f804, %f807; add.f32 %f809, %f805, %f808; sub.f32 %f810, %f805, %f809; add.f32 %f811, %f808, %f810; mov.f32 %f812, 0f3F317200; mul.rn.f32 %f813, %f785, %f812; mov.f32 %f814, 0f35BFBE8E; mul.rn.f32 %f815, %f785, %f814; add.f32 %f816, %f813, %f809; sub.f32 %f817, %f813, %f816; add.f32 %f818, %f809, %f817; add.f32 %f819, %f811, %f818; add.f32 %f820, %f815, %f819; add.f32 %f821, %f816, %f820; sub.f32 %f822, %f816, %f821; add.f32 %f823, %f820, %f822; abs.f32 %f168, %f163; setp.gt.f32 %p54, %f168, 0f77F684DF; mul.f32 %f824, %f163, 0f39000000; selp.f32 %f825, %f824, %f163, %p54; mul.rn.f32 %f826, %f825, %f821; neg.f32 %f827, %f826; fma.rn.f32 %f828, %f825, %f821, %f827; fma.rn.f32 %f829, %f825, %f823, %f828; mov.f32 %f830, 0f00000000; fma.rn.f32 %f831, %f830, %f821, %f829; add.rn.f32 %f832, %f826, %f831; neg.f32 %f833, %f832; add.rn.f32 %f834, %f826, %f833; add.rn.f32 %f835, %f834, %f831; mov.b32 %r284, %f832; setp.eq.s32 %p55, %r284, 1118925336; add.s32 %r285, %r284, -1; mov.b32 %f836, %r285; add.f32 %f837, %f835, 0f37000000; selp.f32 %f169, %f837, %f835, %p55; selp.f32 %f838, %f836, %f832, %p55; mov.f32 %f839, 0f3FB8AA3B; mul.rn.f32 %f840, %f838, %f839; cvt.rzi.f32.f32 %f841, %f840; abs.f32 %f842, %f841; setp.gt.f32 %p56, %f842, 0f42FC0000; mov.b32 %r286, %f841; and.b32 %r287, %r286, -2147483648; or.b32 %r288, %r287, 1123811328; mov.b32 %f843, %r288; selp.f32 %f844, %f843, %f841, %p56; mov.f32 %f845, 0fBF317218; fma.rn.f32 %f846, %f844, %f845, %f838; mov.f32 %f847, 0f3102E308; fma.rn.f32 %f848, %f844, %f847, %f846; mul.f32 %f849, %f848, 0f3FB8AA3B; add.f32 %f850, %f844, 0f4B40007F; mov.b32 %r289, %f850; shl.b32 %r290, %r289, 23; mov.b32 %f851, %r290; ex2.approx.ftz.f32 %f852, %f849; mul.f32 %f170, %f852, %f851; setp.eq.f32 %p57, %f170, 0f7F800000; mov.f32 %f1664, 0f7F800000; @%p57 bra $L__BB0_43; fma.rn.f32 %f1664, %f170, %f169, %f170; $L__BB0_43: setp.lt.f32 %p58, %f162, 0f00000000; setp.eq.f32 %p59, %f166, 0f3F800000; and.pred %p3, %p59, %p58; setp.eq.f32 %p60, %f162, 0f00000000; @%p60 bra $L__BB0_47; bra.uni $L__BB0_44; $L__BB0_47: add.f32 %f856, %f162, %f162; mov.b32 %r293, %f856; selp.b32 %r294, %r293, 0, %p59; or.b32 %r295, %r294, 2139095040; setp.lt.f32 %p64, %f163, 0f00000000; selp.b32 %r296, %r295, %r294, %p64; mov.b32 %f1666, %r296; bra.uni $L__BB0_48; $L__BB0_44: mov.b32 %r291, %f1664; xor.b32 %r292, %r291, -2147483648; mov.b32 %f853, %r292; selp.f32 %f1666, %f853, %f1664, %p3; setp.geu.f32 %p61, %f162, 0f00000000; @%p61 bra $L__BB0_48; cvt.rzi.f32.f32 %f854, %f163; setp.eq.f32 %p62, %f854, %f163; @%p62 bra $L__BB0_48; mov.f32 %f1666, 0f7FFFFFFF; $L__BB0_48: add.f32 %f857, %f167, %f168; mov.b32 %r297, %f857; setp.lt.s32 %p65, %r297, 2139095040; @%p65 bra $L__BB0_55; setp.gtu.f32 %p66, %f167, 0f7F800000; setp.gtu.f32 %p67, %f168, 0f7F800000; or.pred %p68, %p66, %p67; @%p68 bra $L__BB0_54; bra.uni $L__BB0_50; $L__BB0_54: add.f32 %f1666, %f163, %f162; bra.uni $L__BB0_55; $L__BB0_50: setp.eq.f32 %p69, %f168, 0f7F800000; @%p69 bra $L__BB0_53; bra.uni $L__BB0_51; $L__BB0_53: setp.gt.f32 %p72, %f167, 0f3F800000; selp.b32 %r301, 2139095040, 0, %p72; xor.b32 %r302, %r301, 2139095040; setp.lt.f32 %p73, %f163, 0f00000000; selp.b32 %r303, %r302, %r301, %p73; mov.b32 %f858, %r303; setp.eq.f32 %p74, %f162, 0fBF800000; selp.f32 %f1666, 0f3F800000, %f858, %p74; bra.uni $L__BB0_55; $L__BB0_51: setp.neu.f32 %p70, %f167, 0f7F800000; @%p70 bra $L__BB0_55; setp.ge.f32 %p71, %f163, 0f00000000; selp.b32 %r298, 2139095040, 0, %p71; or.b32 %r299, %r298, -2147483648; selp.b32 %r300, %r299, %r298, %p3; mov.b32 %f1666, %r300; $L__BB0_55: setp.eq.f32 %p75, %f163, 0f00000000; setp.eq.f32 %p76, %f162, 0f3F800000; or.pred %p77, %p75, %p76; selp.f32 %f859, 0f3F800000, %f1666, %p77; fma.rn.f32 %f860, %f165, %f859, %f164; mul.f32 %f861, %f1647, %f860; div.rn.f32 %f862, %f861, %f375; div.rn.f32 %f863, %f862, %f73; cvt.sat.f32.f32 %f864, %f863; mul.f32 %f1667, %f1667, %f864; $L__BB0_56: add.f32 %f1683, %f1683, %f1667; $L__BB0_57: cvt.u64.u32 %rd77, %r447; cvt.u32.u64 %r304, %rd77; add.s32 %r447, %r304, 1; setp.lt.u32 %p78, %r447, %r5; @%p78 bra $L__BB0_7; bra.uni $L__BB0_59; $L__BB0_58: mov.f32 %f1668, 0f00000000; ld.const.u32 %r451, [params+104]; mov.f32 %f1669, %f1668; mov.f32 %f1670, %f1668; mov.f32 %f1671, %f1668; mov.f32 %f1672, %f1668; mov.f32 %f1673, %f1668; mov.f32 %f1674, %f1668; mov.f32 %f1675, %f1668; mov.f32 %f1676, %f1668; mov.f32 %f1677, %f1668; mov.f32 %f1678, %f1668; mov.f32 %f1679, %f1668; mov.f32 %f1680, %f1668; mov.f32 %f1681, %f1668; mov.f32 %f1682, %f1668; mov.f32 %f1683, %f1668; $L__BB0_59: and.b32 %r305, %r451, 8; setp.eq.s32 %p79, %r305, 0; @%p79 bra $L__BB0_73; cvt.sat.f32.f32 %f215, %f1683; ld.const.u64 %rd33, [params+192]; cvta.to.global.u64 %rd7, %rd33; ld.const.u32 %r306, [params+184]; mad.lo.s32 %r307, %r306, %r4, %r3; cvt.u64.u32 %rd8, %r307; mov.f32 %f882, 0f3E68BA2E; cvt.rzi.f32.f32 %f883, %f882; add.f32 %f884, %f883, %f883; mov.f32 %f885, 0f3EE8BA2E; sub.f32 %f886, %f885, %f884; abs.f32 %f216, %f886; abs.f32 %f217, %f215; setp.lt.f32 %p80, %f217, 0f00800000; mul.f32 %f887, %f217, 0f4B800000; selp.f32 %f888, %f887, %f217, %p80; selp.f32 %f889, 0fC3170000, 0fC2FE0000, %p80; mov.b32 %r308, %f888; and.b32 %r309, %r308, 8388607; or.b32 %r310, %r309, 1065353216; mov.b32 %f890, %r310; shr.u32 %r311, %r308, 23; cvt.rn.f32.u32 %f891, %r311; add.f32 %f892, %f889, %f891; setp.gt.f32 %p81, %f890, 0f3FB504F3; mul.f32 %f893, %f890, 0f3F000000; add.f32 %f894, %f892, 0f3F800000; selp.f32 %f895, %f894, %f892, %p81; selp.f32 %f896, %f893, %f890, %p81; add.f32 %f897, %f896, 0fBF800000; add.f32 %f898, %f896, 0f3F800000; rcp.approx.ftz.f32 %f899, %f898; add.f32 %f900, %f897, %f897; mul.f32 %f901, %f900, %f899; mul.f32 %f902, %f901, %f901; mov.f32 %f903, 0f3C4CAF63; mov.f32 %f904, 0f3B18F0FE; fma.rn.f32 %f905, %f904, %f902, %f903; mov.f32 %f906, 0f3DAAAABD; fma.rn.f32 %f907, %f905, %f902, %f906; mul.rn.f32 %f908, %f907, %f902; mul.rn.f32 %f909, %f908, %f901; sub.f32 %f910, %f897, %f901; add.f32 %f911, %f910, %f910; neg.f32 %f912, %f901; fma.rn.f32 %f913, %f912, %f897, %f911; mul.rn.f32 %f914, %f899, %f913; add.f32 %f915, %f909, %f901; sub.f32 %f916, %f901, %f915; add.f32 %f917, %f909, %f916; add.f32 %f918, %f914, %f917; add.f32 %f919, %f915, %f918; sub.f32 %f920, %f915, %f919; add.f32 %f921, %f918, %f920; mov.f32 %f922, 0f3F317200; mul.rn.f32 %f923, %f895, %f922; mov.f32 %f924, 0f35BFBE8E; mul.rn.f32 %f925, %f895, %f924; add.f32 %f926, %f923, %f919; sub.f32 %f927, %f923, %f926; add.f32 %f928, %f919, %f927; add.f32 %f929, %f921, %f928; add.f32 %f930, %f925, %f929; add.f32 %f931, %f926, %f930; sub.f32 %f932, %f926, %f931; add.f32 %f933, %f930, %f932; mul.rn.f32 %f934, %f885, %f931; neg.f32 %f935, %f934; fma.rn.f32 %f936, %f885, %f931, %f935; fma.rn.f32 %f937, %f885, %f933, %f936; mov.f32 %f938, 0f00000000; fma.rn.f32 %f939, %f938, %f931, %f937; add.rn.f32 %f940, %f934, %f939; neg.f32 %f941, %f940; add.rn.f32 %f942, %f934, %f941; add.rn.f32 %f943, %f942, %f939; mov.b32 %r312, %f940; setp.eq.s32 %p82, %r312, 1118925336; add.s32 %r313, %r312, -1; mov.b32 %f944, %r313; add.f32 %f945, %f943, 0f37000000; selp.f32 %f218, %f945, %f943, %p82; selp.f32 %f946, %f944, %f940, %p82; mov.f32 %f947, 0f3FB8AA3B; mul.rn.f32 %f948, %f946, %f947; cvt.rzi.f32.f32 %f949, %f948; abs.f32 %f950, %f949; setp.gt.f32 %p83, %f950, 0f42FC0000; mov.b32 %r314, %f949; and.b32 %r315, %r314, -2147483648; or.b32 %r316, %r315, 1123811328; mov.b32 %f951, %r316; selp.f32 %f952, %f951, %f949, %p83; mov.f32 %f953, 0fBF317218; fma.rn.f32 %f954, %f952, %f953, %f946; mov.f32 %f955, 0f3102E308; fma.rn.f32 %f956, %f952, %f955, %f954; mul.f32 %f957, %f956, 0f3FB8AA3B; add.f32 %f958, %f952, 0f4B40007F; mov.b32 %r317, %f958; shl.b32 %r318, %r317, 23; mov.b32 %f959, %r318; ex2.approx.ftz.f32 %f960, %f957; mul.f32 %f219, %f960, %f959; setp.eq.f32 %p84, %f219, 0f7F800000; mov.f32 %f1700, 0f7F800000; @%p84 bra $L__BB0_62; fma.rn.f32 %f1700, %f219, %f218, %f219; $L__BB0_62: setp.lt.f32 %p85, %f215, 0f00000000; setp.eq.f32 %p86, %f216, 0f3F800000; and.pred %p4, %p85, %p86; setp.eq.f32 %p87, %f215, 0f00000000; @%p87 bra $L__BB0_66; bra.uni $L__BB0_63; $L__BB0_66: add.f32 %f965, %f215, %f215; selp.f32 %f1702, %f965, 0f00000000, %p86; bra.uni $L__BB0_67; $L__BB0_170: mov.f32 %f1574, 0f00000000; mov.u32 %r453, 0; // begin inline asm { cvt.rn.f16.f32 %rs115, %f1574;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs114, %f1574;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs113, %f1574;} // end inline asm mov.u16 %rs116, 0; st.global.v4.u16 [%rd16], {%rs113, %rs114, %rs115, %rs116}; $L__BB0_171: ld.const.u64 %rd65, [params+256]; cvta.to.global.u64 %rd66, %rd65; ld.const.u32 %r439, [params+248]; mad.lo.s32 %r440, %r439, %r4, %r3; mul.wide.u32 %rd67, %r440, 8; add.s64 %rd17, %rd66, %rd67; setp.eq.s32 %p205, %r453, 0; @%p205 bra $L__BB0_173; ld.global.v4.u16 {%rs123, %rs124, %rs125, %rs126}, [%rd17]; // begin inline asm { cvt.f32.f16 %f1575, %rs123;} // end inline asm // begin inline asm { cvt.f32.f16 %f1576, %rs124;} // end inline asm // begin inline asm { cvt.f32.f16 %f1577, %rs125;} // 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 %rs122, %f1580;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs121, %f1579;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs120, %f1578;} // end inline asm mov.u16 %rs127, 0; st.global.v4.u16 [%rd17], {%rs120, %rs121, %rs122, %rs127}; bra.uni $L__BB0_174; $L__BB0_173: mov.f32 %f1583, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs130, %f1583;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs129, %f1583;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs128, %f1583;} // end inline asm mov.u16 %rs131, 0; st.global.v4.u16 [%rd17], {%rs128, %rs129, %rs130, %rs131}; $L__BB0_174: ld.const.u64 %rd68, [params+272]; cvta.to.global.u64 %rd69, %rd68; ld.const.u32 %r441, [params+264]; mad.lo.s32 %r442, %r441, %r4, %r3; mul.wide.u32 %rd70, %r442, 8; add.s64 %rd18, %rd69, %rd70; @%p205 bra $L__BB0_176; ld.global.v4.u16 {%rs138, %rs139, %rs140, %rs141}, [%rd18]; // begin inline asm { cvt.f32.f16 %f1584, %rs138;} // end inline asm // begin inline asm { cvt.f32.f16 %f1585, %rs139;} // end inline asm // begin inline asm { cvt.f32.f16 %f1586, %rs140;} // 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 %rs137, %f1589;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs136, %f1588;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs135, %f1587;} // end inline asm mov.u16 %rs142, 0; st.global.v4.u16 [%rd18], {%rs135, %rs136, %rs137, %rs142}; bra.uni $L__BB0_177; $L__BB0_176: mov.f32 %f1592, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs145, %f1592;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs144, %f1592;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs143, %f1592;} // end inline asm mov.u16 %rs146, 0; st.global.v4.u16 [%rd18], {%rs143, %rs144, %rs145, %rs146}; $L__BB0_177: ld.const.u64 %rd71, [params+288]; cvta.to.global.u64 %rd72, %rd71; ld.const.u32 %r443, [params+280]; mad.lo.s32 %r444, %r443, %r4, %r3; mul.wide.u32 %rd73, %r444, 8; add.s64 %rd19, %rd72, %rd73; @%p205 bra $L__BB0_179; ld.global.v4.u16 {%rs153, %rs154, %rs155, %rs156}, [%rd19]; // begin inline asm { cvt.f32.f16 %f1593, %rs153;} // end inline asm // begin inline asm { cvt.f32.f16 %f1594, %rs154;} // end inline asm // begin inline asm { cvt.f32.f16 %f1595, %rs155;} // end inline asm add.f32 %f1596, %f1593, 0f00000000; add.f32 %f1597, %f1594, 0f00000000; add.f32 %f1598, %f1595, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs152, %f1598;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs151, %f1597;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs150, %f1596;} // end inline asm mov.u16 %rs157, 0; st.global.v4.u16 [%rd19], {%rs150, %rs151, %rs152, %rs157}; bra.uni $L__BB0_180; $L__BB0_179: mov.f32 %f1601, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs160, %f1601;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs159, %f1601;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs158, %f1601;} // end inline asm mov.u16 %rs161, 0; st.global.v4.u16 [%rd19], {%rs158, %rs159, %rs160, %rs161}; $L__BB0_180: ld.const.u64 %rd74, [params+304]; cvta.to.global.u64 %rd75, %rd74; ld.const.u32 %r445, [params+296]; mad.lo.s32 %r446, %r445, %r4, %r3; mul.wide.u32 %rd76, %r446, 8; add.s64 %rd20, %rd75, %rd76; @%p205 bra $L__BB0_182; ld.global.v4.u16 {%rs168, %rs169, %rs170, %rs171}, [%rd20]; // begin inline asm { cvt.f32.f16 %f1602, %rs168;} // end inline asm // begin inline asm { cvt.f32.f16 %f1603, %rs169;} // end inline asm // begin inline asm { cvt.f32.f16 %f1604, %rs170;} // end inline asm add.f32 %f1605, %f1602, 0f00000000; add.f32 %f1606, %f1603, 0f00000000; add.f32 %f1607, %f1604, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs167, %f1607;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs166, %f1606;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs165, %f1605;} // end inline asm mov.u16 %rs172, 0; st.global.v4.u16 [%rd20], {%rs165, %rs166, %rs167, %rs172}; bra.uni $L__BB0_183; $L__BB0_182: mov.f32 %f1610, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs175, %f1610;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs174, %f1610;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs173, %f1610;} // end inline asm mov.u16 %rs176, 0; st.global.v4.u16 [%rd20], {%rs173, %rs174, %rs175, %rs176}; bra.uni $L__BB0_183; $L__BB0_63: mov.b32 %r319, %f1700; xor.b32 %r320, %r319, -2147483648; mov.b32 %f961, %r320; selp.f32 %f1702, %f961, %f1700, %p4; setp.geu.f32 %p88, %f215, 0f00000000; @%p88 bra $L__BB0_67; mov.f32 %f962, 0f3EE8BA2E; cvt.rzi.f32.f32 %f963, %f962; setp.eq.f32 %p89, %f963, 0f3EE8BA2E; @%p89 bra $L__BB0_67; mov.f32 %f1702, 0f7FFFFFFF; $L__BB0_67: add.f32 %f966, %f217, 0f3EE8BA2E; mov.b32 %r321, %f966; setp.lt.s32 %p91, %r321, 2139095040; @%p91 bra $L__BB0_72; setp.gtu.f32 %p92, %f217, 0f7F800000; @%p92 bra $L__BB0_71; bra.uni $L__BB0_69; $L__BB0_71: add.f32 %f1702, %f215, 0f3EE8BA2E; bra.uni $L__BB0_72; $L__BB0_69: setp.neu.f32 %p93, %f217, 0f7F800000; @%p93 bra $L__BB0_72; selp.f32 %f1702, 0fFF800000, 0f7F800000, %p4; $L__BB0_72: mul.f32 %f967, %f1702, 0f437F0000; setp.eq.f32 %p94, %f215, 0f3F800000; selp.f32 %f968, 0f437F0000, %f967, %p94; cvt.rzi.u32.f32 %r322, %f968; shl.b64 %rd34, %rd8, 1; add.s64 %rd35, %rd7, %rd34; cvt.u16.u32 %rs18, %r322; mov.u16 %rs19, 255; st.global.v2.u8 [%rd35], {%rs18, %rs19}; $L__BB0_73: and.b32 %r323, %r451, 1; setp.eq.b32 %p95, %r323, 1; mov.pred %p96, 0; xor.pred %p97, %p95, %p96; not.pred %p98, %p97; @%p98 bra $L__BB0_147; mov.f32 %f970, 0f3E666666; cvt.rzi.f32.f32 %f971, %f970; add.f32 %f972, %f971, %f971; mov.f32 %f973, 0f3EE66666; sub.f32 %f974, %f973, %f972; abs.f32 %f228, %f974; abs.f32 %f229, %f1668; setp.lt.f32 %p99, %f229, 0f00800000; mul.f32 %f975, %f229, 0f4B800000; selp.f32 %f976, %f975, %f229, %p99; selp.f32 %f977, 0fC3170000, 0fC2FE0000, %p99; mov.b32 %r324, %f976; and.b32 %r325, %r324, 8388607; or.b32 %r326, %r325, 1065353216; mov.b32 %f978, %r326; shr.u32 %r327, %r324, 23; cvt.rn.f32.u32 %f979, %r327; add.f32 %f980, %f977, %f979; setp.gt.f32 %p100, %f978, 0f3FB504F3; mul.f32 %f981, %f978, 0f3F000000; add.f32 %f982, %f980, 0f3F800000; selp.f32 %f983, %f982, %f980, %p100; selp.f32 %f984, %f981, %f978, %p100; add.f32 %f985, %f984, 0fBF800000; add.f32 %f986, %f984, 0f3F800000; rcp.approx.ftz.f32 %f987, %f986; add.f32 %f988, %f985, %f985; mul.f32 %f989, %f988, %f987; mul.f32 %f990, %f989, %f989; mov.f32 %f991, 0f3C4CAF63; mov.f32 %f992, 0f3B18F0FE; fma.rn.f32 %f993, %f992, %f990, %f991; mov.f32 %f994, 0f3DAAAABD; fma.rn.f32 %f995, %f993, %f990, %f994; mul.rn.f32 %f996, %f995, %f990; mul.rn.f32 %f997, %f996, %f989; sub.f32 %f998, %f985, %f989; add.f32 %f999, %f998, %f998; neg.f32 %f1000, %f989; fma.rn.f32 %f1001, %f1000, %f985, %f999; mul.rn.f32 %f1002, %f987, %f1001; add.f32 %f1003, %f997, %f989; sub.f32 %f1004, %f989, %f1003; add.f32 %f1005, %f997, %f1004; add.f32 %f1006, %f1002, %f1005; add.f32 %f1007, %f1003, %f1006; sub.f32 %f1008, %f1003, %f1007; add.f32 %f1009, %f1006, %f1008; mov.f32 %f1010, 0f3F317200; mul.rn.f32 %f1011, %f983, %f1010; mov.f32 %f1012, 0f35BFBE8E; mul.rn.f32 %f1013, %f983, %f1012; add.f32 %f1014, %f1011, %f1007; sub.f32 %f1015, %f1011, %f1014; add.f32 %f1016, %f1007, %f1015; add.f32 %f1017, %f1009, %f1016; add.f32 %f1018, %f1013, %f1017; add.f32 %f1019, %f1014, %f1018; sub.f32 %f1020, %f1014, %f1019; add.f32 %f1021, %f1018, %f1020; mul.rn.f32 %f1022, %f973, %f1019; neg.f32 %f1023, %f1022; fma.rn.f32 %f1024, %f973, %f1019, %f1023; fma.rn.f32 %f1025, %f973, %f1021, %f1024; mov.f32 %f1026, 0f00000000; fma.rn.f32 %f1027, %f1026, %f1019, %f1025; add.rn.f32 %f1028, %f1022, %f1027; neg.f32 %f1029, %f1028; add.rn.f32 %f1030, %f1022, %f1029; add.rn.f32 %f1031, %f1030, %f1027; mov.b32 %r328, %f1028; setp.eq.s32 %p101, %r328, 1118925336; add.s32 %r329, %r328, -1; mov.b32 %f1032, %r329; add.f32 %f1033, %f1031, 0f37000000; selp.f32 %f230, %f1033, %f1031, %p101; selp.f32 %f1034, %f1032, %f1028, %p101; mov.f32 %f1035, 0f3FB8AA3B; mul.rn.f32 %f1036, %f1034, %f1035; cvt.rzi.f32.f32 %f1037, %f1036; abs.f32 %f1038, %f1037; setp.gt.f32 %p102, %f1038, 0f42FC0000; mov.b32 %r330, %f1037; and.b32 %r331, %r330, -2147483648; or.b32 %r332, %r331, 1123811328; mov.b32 %f1039, %r332; selp.f32 %f1040, %f1039, %f1037, %p102; mov.f32 %f1041, 0fBF317218; fma.rn.f32 %f1042, %f1040, %f1041, %f1034; mov.f32 %f1043, 0f3102E308; fma.rn.f32 %f1044, %f1040, %f1043, %f1042; mul.f32 %f1045, %f1044, 0f3FB8AA3B; add.f32 %f1046, %f1040, 0f4B40007F; mov.b32 %r333, %f1046; shl.b32 %r334, %r333, 23; mov.b32 %f1047, %r334; ex2.approx.ftz.f32 %f1048, %f1045; mul.f32 %f231, %f1048, %f1047; setp.eq.f32 %p103, %f231, 0f7F800000; mov.f32 %f1703, 0f7F800000; @%p103 bra $L__BB0_76; fma.rn.f32 %f1703, %f231, %f230, %f231; $L__BB0_76: setp.lt.f32 %p104, %f1668, 0f00000000; setp.eq.f32 %p105, %f228, 0f3F800000; and.pred %p5, %p104, %p105; setp.eq.f32 %p106, %f1668, 0f00000000; @%p106 bra $L__BB0_80; bra.uni $L__BB0_77; $L__BB0_80: add.f32 %f1053, %f1668, %f1668; selp.f32 %f1705, %f1053, 0f00000000, %p105; bra.uni $L__BB0_81; $L__BB0_77: mov.b32 %r335, %f1703; xor.b32 %r336, %r335, -2147483648; mov.b32 %f1049, %r336; selp.f32 %f1705, %f1049, %f1703, %p5; setp.geu.f32 %p107, %f1668, 0f00000000; @%p107 bra $L__BB0_81; mov.f32 %f1050, 0f3EE66666; cvt.rzi.f32.f32 %f1051, %f1050; setp.eq.f32 %p108, %f1051, 0f3EE66666; @%p108 bra $L__BB0_81; mov.f32 %f1705, 0f7FFFFFFF; $L__BB0_81: add.f32 %f1054, %f229, 0f3EE66666; mov.b32 %r337, %f1054; setp.lt.s32 %p110, %r337, 2139095040; @%p110 bra $L__BB0_86; setp.gtu.f32 %p111, %f229, 0f7F800000; @%p111 bra $L__BB0_85; bra.uni $L__BB0_83; $L__BB0_85: add.f32 %f1705, %f1668, 0f3EE66666; bra.uni $L__BB0_86; $L__BB0_83: setp.neu.f32 %p112, %f229, 0f7F800000; @%p112 bra $L__BB0_86; selp.f32 %f1705, 0fFF800000, 0f7F800000, %p5; $L__BB0_86: setp.eq.f32 %p113, %f1668, 0f3F800000; selp.f32 %f240, 0f3F800000, %f1705, %p113; abs.f32 %f241, %f1669; setp.lt.f32 %p114, %f241, 0f00800000; mul.f32 %f1056, %f241, 0f4B800000; selp.f32 %f1057, %f1056, %f241, %p114; selp.f32 %f1058, 0fC3170000, 0fC2FE0000, %p114; mov.b32 %r338, %f1057; and.b32 %r339, %r338, 8388607; or.b32 %r340, %r339, 1065353216; mov.b32 %f1059, %r340; shr.u32 %r341, %r338, 23; cvt.rn.f32.u32 %f1060, %r341; add.f32 %f1061, %f1058, %f1060; setp.gt.f32 %p115, %f1059, 0f3FB504F3; mul.f32 %f1062, %f1059, 0f3F000000; add.f32 %f1063, %f1061, 0f3F800000; selp.f32 %f1064, %f1063, %f1061, %p115; selp.f32 %f1065, %f1062, %f1059, %p115; add.f32 %f1066, %f1065, 0fBF800000; add.f32 %f1067, %f1065, 0f3F800000; rcp.approx.ftz.f32 %f1068, %f1067; add.f32 %f1069, %f1066, %f1066; mul.f32 %f1070, %f1069, %f1068; mul.f32 %f1071, %f1070, %f1070; mov.f32 %f1072, 0f3C4CAF63; mov.f32 %f1073, 0f3B18F0FE; fma.rn.f32 %f1074, %f1073, %f1071, %f1072; mov.f32 %f1075, 0f3DAAAABD; fma.rn.f32 %f1076, %f1074, %f1071, %f1075; mul.rn.f32 %f1077, %f1076, %f1071; mul.rn.f32 %f1078, %f1077, %f1070; sub.f32 %f1079, %f1066, %f1070; add.f32 %f1080, %f1079, %f1079; neg.f32 %f1081, %f1070; fma.rn.f32 %f1082, %f1081, %f1066, %f1080; mul.rn.f32 %f1083, %f1068, %f1082; add.f32 %f1084, %f1078, %f1070; sub.f32 %f1085, %f1070, %f1084; add.f32 %f1086, %f1078, %f1085; add.f32 %f1087, %f1083, %f1086; add.f32 %f1088, %f1084, %f1087; sub.f32 %f1089, %f1084, %f1088; add.f32 %f1090, %f1087, %f1089; mov.f32 %f1091, 0f3F317200; mul.rn.f32 %f1092, %f1064, %f1091; mov.f32 %f1093, 0f35BFBE8E; mul.rn.f32 %f1094, %f1064, %f1093; add.f32 %f1095, %f1092, %f1088; sub.f32 %f1096, %f1092, %f1095; add.f32 %f1097, %f1088, %f1096; add.f32 %f1098, %f1090, %f1097; add.f32 %f1099, %f1094, %f1098; add.f32 %f1100, %f1095, %f1099; sub.f32 %f1101, %f1095, %f1100; add.f32 %f1102, %f1099, %f1101; mov.f32 %f1103, 0f3EE66666; mul.rn.f32 %f1104, %f1103, %f1100; neg.f32 %f1105, %f1104; fma.rn.f32 %f1106, %f1103, %f1100, %f1105; fma.rn.f32 %f1107, %f1103, %f1102, %f1106; mov.f32 %f1108, 0f00000000; fma.rn.f32 %f1109, %f1108, %f1100, %f1107; add.rn.f32 %f1110, %f1104, %f1109; neg.f32 %f1111, %f1110; add.rn.f32 %f1112, %f1104, %f1111; add.rn.f32 %f1113, %f1112, %f1109; mov.b32 %r342, %f1110; setp.eq.s32 %p116, %r342, 1118925336; add.s32 %r343, %r342, -1; mov.b32 %f1114, %r343; add.f32 %f1115, %f1113, 0f37000000; selp.f32 %f242, %f1115, %f1113, %p116; selp.f32 %f1116, %f1114, %f1110, %p116; mov.f32 %f1117, 0f3FB8AA3B; mul.rn.f32 %f1118, %f1116, %f1117; cvt.rzi.f32.f32 %f1119, %f1118; abs.f32 %f1120, %f1119; setp.gt.f32 %p117, %f1120, 0f42FC0000; mov.b32 %r344, %f1119; and.b32 %r345, %r344, -2147483648; or.b32 %r346, %r345, 1123811328; mov.b32 %f1121, %r346; selp.f32 %f1122, %f1121, %f1119, %p117; mov.f32 %f1123, 0fBF317218; fma.rn.f32 %f1124, %f1122, %f1123, %f1116; mov.f32 %f1125, 0f3102E308; fma.rn.f32 %f1126, %f1122, %f1125, %f1124; mul.f32 %f1127, %f1126, 0f3FB8AA3B; add.f32 %f1128, %f1122, 0f4B40007F; mov.b32 %r347, %f1128; shl.b32 %r348, %r347, 23; mov.b32 %f1129, %r348; ex2.approx.ftz.f32 %f1130, %f1127; mul.f32 %f243, %f1130, %f1129; setp.eq.f32 %p118, %f243, 0f7F800000; mov.f32 %f1706, 0f7F800000; @%p118 bra $L__BB0_88; fma.rn.f32 %f1706, %f243, %f242, %f243; $L__BB0_88: setp.lt.f32 %p119, %f1669, 0f00000000; and.pred %p6, %p119, %p105; setp.eq.f32 %p121, %f1669, 0f00000000; @%p121 bra $L__BB0_92; bra.uni $L__BB0_89; $L__BB0_92: add.f32 %f1135, %f1669, %f1669; selp.f32 %f1708, %f1135, 0f00000000, %p105; bra.uni $L__BB0_93; $L__BB0_89: mov.b32 %r349, %f1706; xor.b32 %r350, %r349, -2147483648; mov.b32 %f1131, %r350; selp.f32 %f1708, %f1131, %f1706, %p6; setp.geu.f32 %p122, %f1669, 0f00000000; @%p122 bra $L__BB0_93; mov.f32 %f1132, 0f3EE66666; cvt.rzi.f32.f32 %f1133, %f1132; setp.eq.f32 %p123, %f1133, 0f3EE66666; @%p123 bra $L__BB0_93; mov.f32 %f1708, 0f7FFFFFFF; $L__BB0_93: add.f32 %f1136, %f241, 0f3EE66666; mov.b32 %r351, %f1136; setp.lt.s32 %p125, %r351, 2139095040; @%p125 bra $L__BB0_98; setp.gtu.f32 %p126, %f241, 0f7F800000; @%p126 bra $L__BB0_97; bra.uni $L__BB0_95; $L__BB0_97: add.f32 %f1708, %f1669, 0f3EE66666; bra.uni $L__BB0_98; $L__BB0_95: setp.neu.f32 %p127, %f241, 0f7F800000; @%p127 bra $L__BB0_98; selp.f32 %f1708, 0fFF800000, 0f7F800000, %p6; $L__BB0_98: setp.eq.f32 %p128, %f1669, 0f3F800000; selp.f32 %f252, 0f3F800000, %f1708, %p128; abs.f32 %f253, %f1670; setp.lt.f32 %p129, %f253, 0f00800000; mul.f32 %f1138, %f253, 0f4B800000; selp.f32 %f1139, %f1138, %f253, %p129; selp.f32 %f1140, 0fC3170000, 0fC2FE0000, %p129; mov.b32 %r352, %f1139; and.b32 %r353, %r352, 8388607; or.b32 %r354, %r353, 1065353216; mov.b32 %f1141, %r354; shr.u32 %r355, %r352, 23; cvt.rn.f32.u32 %f1142, %r355; add.f32 %f1143, %f1140, %f1142; setp.gt.f32 %p130, %f1141, 0f3FB504F3; mul.f32 %f1144, %f1141, 0f3F000000; add.f32 %f1145, %f1143, 0f3F800000; selp.f32 %f1146, %f1145, %f1143, %p130; selp.f32 %f1147, %f1144, %f1141, %p130; add.f32 %f1148, %f1147, 0fBF800000; add.f32 %f1149, %f1147, 0f3F800000; rcp.approx.ftz.f32 %f1150, %f1149; add.f32 %f1151, %f1148, %f1148; mul.f32 %f1152, %f1151, %f1150; mul.f32 %f1153, %f1152, %f1152; mov.f32 %f1154, 0f3C4CAF63; mov.f32 %f1155, 0f3B18F0FE; fma.rn.f32 %f1156, %f1155, %f1153, %f1154; mov.f32 %f1157, 0f3DAAAABD; fma.rn.f32 %f1158, %f1156, %f1153, %f1157; mul.rn.f32 %f1159, %f1158, %f1153; mul.rn.f32 %f1160, %f1159, %f1152; sub.f32 %f1161, %f1148, %f1152; add.f32 %f1162, %f1161, %f1161; neg.f32 %f1163, %f1152; fma.rn.f32 %f1164, %f1163, %f1148, %f1162; mul.rn.f32 %f1165, %f1150, %f1164; add.f32 %f1166, %f1160, %f1152; sub.f32 %f1167, %f1152, %f1166; add.f32 %f1168, %f1160, %f1167; add.f32 %f1169, %f1165, %f1168; add.f32 %f1170, %f1166, %f1169; sub.f32 %f1171, %f1166, %f1170; add.f32 %f1172, %f1169, %f1171; mov.f32 %f1173, 0f3F317200; mul.rn.f32 %f1174, %f1146, %f1173; mov.f32 %f1175, 0f35BFBE8E; mul.rn.f32 %f1176, %f1146, %f1175; add.f32 %f1177, %f1174, %f1170; sub.f32 %f1178, %f1174, %f1177; add.f32 %f1179, %f1170, %f1178; add.f32 %f1180, %f1172, %f1179; add.f32 %f1181, %f1176, %f1180; add.f32 %f1182, %f1177, %f1181; sub.f32 %f1183, %f1177, %f1182; add.f32 %f1184, %f1181, %f1183; mov.f32 %f1185, 0f3EE66666; mul.rn.f32 %f1186, %f1185, %f1182; neg.f32 %f1187, %f1186; fma.rn.f32 %f1188, %f1185, %f1182, %f1187; fma.rn.f32 %f1189, %f1185, %f1184, %f1188; mov.f32 %f1190, 0f00000000; fma.rn.f32 %f1191, %f1190, %f1182, %f1189; add.rn.f32 %f1192, %f1186, %f1191; neg.f32 %f1193, %f1192; add.rn.f32 %f1194, %f1186, %f1193; add.rn.f32 %f1195, %f1194, %f1191; mov.b32 %r356, %f1192; setp.eq.s32 %p131, %r356, 1118925336; add.s32 %r357, %r356, -1; mov.b32 %f1196, %r357; add.f32 %f1197, %f1195, 0f37000000; selp.f32 %f254, %f1197, %f1195, %p131; selp.f32 %f1198, %f1196, %f1192, %p131; mov.f32 %f1199, 0f3FB8AA3B; mul.rn.f32 %f1200, %f1198, %f1199; cvt.rzi.f32.f32 %f1201, %f1200; abs.f32 %f1202, %f1201; setp.gt.f32 %p132, %f1202, 0f42FC0000; mov.b32 %r358, %f1201; and.b32 %r359, %r358, -2147483648; or.b32 %r360, %r359, 1123811328; mov.b32 %f1203, %r360; selp.f32 %f1204, %f1203, %f1201, %p132; mov.f32 %f1205, 0fBF317218; fma.rn.f32 %f1206, %f1204, %f1205, %f1198; mov.f32 %f1207, 0f3102E308; fma.rn.f32 %f1208, %f1204, %f1207, %f1206; mul.f32 %f1209, %f1208, 0f3FB8AA3B; add.f32 %f1210, %f1204, 0f4B40007F; mov.b32 %r361, %f1210; shl.b32 %r362, %r361, 23; mov.b32 %f1211, %r362; ex2.approx.ftz.f32 %f1212, %f1209; mul.f32 %f255, %f1212, %f1211; setp.eq.f32 %p133, %f255, 0f7F800000; mov.f32 %f1709, 0f7F800000; @%p133 bra $L__BB0_100; fma.rn.f32 %f1709, %f255, %f254, %f255; $L__BB0_100: setp.lt.f32 %p134, %f1670, 0f00000000; and.pred %p7, %p134, %p105; setp.eq.f32 %p136, %f1670, 0f00000000; @%p136 bra $L__BB0_104; bra.uni $L__BB0_101; $L__BB0_104: add.f32 %f1217, %f1670, %f1670; selp.f32 %f1711, %f1217, 0f00000000, %p105; bra.uni $L__BB0_105; $L__BB0_101: mov.b32 %r363, %f1709; xor.b32 %r364, %r363, -2147483648; mov.b32 %f1213, %r364; selp.f32 %f1711, %f1213, %f1709, %p7; setp.geu.f32 %p137, %f1670, 0f00000000; @%p137 bra $L__BB0_105; mov.f32 %f1214, 0f3EE66666; cvt.rzi.f32.f32 %f1215, %f1214; setp.eq.f32 %p138, %f1215, 0f3EE66666; @%p138 bra $L__BB0_105; mov.f32 %f1711, 0f7FFFFFFF; $L__BB0_105: add.f32 %f1218, %f253, 0f3EE66666; mov.b32 %r365, %f1218; setp.lt.s32 %p140, %r365, 2139095040; @%p140 bra $L__BB0_110; setp.gtu.f32 %p141, %f253, 0f7F800000; @%p141 bra $L__BB0_109; bra.uni $L__BB0_107; $L__BB0_109: add.f32 %f1711, %f1670, 0f3EE66666; bra.uni $L__BB0_110; $L__BB0_107: setp.neu.f32 %p142, %f253, 0f7F800000; @%p142 bra $L__BB0_110; selp.f32 %f1711, 0fFF800000, 0f7F800000, %p7; $L__BB0_110: setp.eq.f32 %p143, %f1670, 0f3F800000; mov.f32 %f1220, 0f3F800000; selp.f32 %f1221, 0f3F800000, %f1711, %p143; ld.const.u64 %rd36, [params+144]; cvta.to.global.u64 %rd9, %rd36; ld.const.u32 %r366, [params+136]; mad.lo.s32 %r367, %r366, %r4, %r3; cvt.u64.u32 %rd10, %r367; min.f32 %f1222, %f240, %f1220; mov.f32 %f1223, 0f00000000; max.f32 %f264, %f1223, %f1222; min.f32 %f1224, %f252, %f1220; max.f32 %f265, %f1223, %f1224; min.f32 %f1225, %f1221, %f1220; max.f32 %f266, %f1223, %f1225; mov.f32 %f1226, 0f3E555555; cvt.rzi.f32.f32 %f1227, %f1226; add.f32 %f1228, %f1227, %f1227; mov.f32 %f1229, 0f3ED55555; sub.f32 %f1230, %f1229, %f1228; abs.f32 %f267, %f1230; abs.f32 %f268, %f264; setp.lt.f32 %p144, %f268, 0f00800000; mul.f32 %f1231, %f268, 0f4B800000; selp.f32 %f1232, %f1231, %f268, %p144; selp.f32 %f1233, 0fC3170000, 0fC2FE0000, %p144; mov.b32 %r368, %f1232; and.b32 %r369, %r368, 8388607; or.b32 %r370, %r369, 1065353216; mov.b32 %f1234, %r370; shr.u32 %r371, %r368, 23; cvt.rn.f32.u32 %f1235, %r371; add.f32 %f1236, %f1233, %f1235; setp.gt.f32 %p145, %f1234, 0f3FB504F3; mul.f32 %f1237, %f1234, 0f3F000000; add.f32 %f1238, %f1236, 0f3F800000; selp.f32 %f1239, %f1238, %f1236, %p145; selp.f32 %f1240, %f1237, %f1234, %p145; add.f32 %f1241, %f1240, 0fBF800000; add.f32 %f1242, %f1240, 0f3F800000; rcp.approx.ftz.f32 %f1243, %f1242; add.f32 %f1244, %f1241, %f1241; mul.f32 %f1245, %f1244, %f1243; mul.f32 %f1246, %f1245, %f1245; mov.f32 %f1247, 0f3C4CAF63; mov.f32 %f1248, 0f3B18F0FE; fma.rn.f32 %f1249, %f1248, %f1246, %f1247; mov.f32 %f1250, 0f3DAAAABD; fma.rn.f32 %f1251, %f1249, %f1246, %f1250; mul.rn.f32 %f1252, %f1251, %f1246; mul.rn.f32 %f1253, %f1252, %f1245; sub.f32 %f1254, %f1241, %f1245; add.f32 %f1255, %f1254, %f1254; neg.f32 %f1256, %f1245; fma.rn.f32 %f1257, %f1256, %f1241, %f1255; mul.rn.f32 %f1258, %f1243, %f1257; add.f32 %f1259, %f1253, %f1245; sub.f32 %f1260, %f1245, %f1259; add.f32 %f1261, %f1253, %f1260; add.f32 %f1262, %f1258, %f1261; add.f32 %f1263, %f1259, %f1262; sub.f32 %f1264, %f1259, %f1263; add.f32 %f1265, %f1262, %f1264; mov.f32 %f1266, 0f3F317200; mul.rn.f32 %f1267, %f1239, %f1266; mov.f32 %f1268, 0f35BFBE8E; mul.rn.f32 %f1269, %f1239, %f1268; add.f32 %f1270, %f1267, %f1263; sub.f32 %f1271, %f1267, %f1270; add.f32 %f1272, %f1263, %f1271; add.f32 %f1273, %f1265, %f1272; add.f32 %f1274, %f1269, %f1273; add.f32 %f1275, %f1270, %f1274; sub.f32 %f1276, %f1270, %f1275; add.f32 %f1277, %f1274, %f1276; mul.rn.f32 %f1278, %f1229, %f1275; neg.f32 %f1279, %f1278; fma.rn.f32 %f1280, %f1229, %f1275, %f1279; fma.rn.f32 %f1281, %f1229, %f1277, %f1280; fma.rn.f32 %f1282, %f1223, %f1275, %f1281; add.rn.f32 %f1283, %f1278, %f1282; neg.f32 %f1284, %f1283; add.rn.f32 %f1285, %f1278, %f1284; add.rn.f32 %f1286, %f1285, %f1282; mov.b32 %r372, %f1283; setp.eq.s32 %p146, %r372, 1118925336; add.s32 %r373, %r372, -1; mov.b32 %f1287, %r373; add.f32 %f1288, %f1286, 0f37000000; selp.f32 %f269, %f1288, %f1286, %p146; selp.f32 %f1289, %f1287, %f1283, %p146; mov.f32 %f1290, 0f3FB8AA3B; mul.rn.f32 %f1291, %f1289, %f1290; cvt.rzi.f32.f32 %f1292, %f1291; abs.f32 %f1293, %f1292; setp.gt.f32 %p147, %f1293, 0f42FC0000; mov.b32 %r374, %f1292; and.b32 %r375, %r374, -2147483648; or.b32 %r376, %r375, 1123811328; mov.b32 %f1294, %r376; selp.f32 %f1295, %f1294, %f1292, %p147; mov.f32 %f1296, 0fBF317218; fma.rn.f32 %f1297, %f1295, %f1296, %f1289; mov.f32 %f1298, 0f3102E308; fma.rn.f32 %f1299, %f1295, %f1298, %f1297; mul.f32 %f1300, %f1299, 0f3FB8AA3B; add.f32 %f1301, %f1295, 0f4B40007F; mov.b32 %r377, %f1301; shl.b32 %r378, %r377, 23; mov.b32 %f1302, %r378; ex2.approx.ftz.f32 %f1303, %f1300; mul.f32 %f270, %f1303, %f1302; setp.eq.f32 %p148, %f270, 0f7F800000; mov.f32 %f1712, 0f7F800000; @%p148 bra $L__BB0_112; fma.rn.f32 %f1712, %f270, %f269, %f270; $L__BB0_112: setp.lt.f32 %p149, %f264, 0f00000000; setp.eq.f32 %p150, %f267, 0f3F800000; and.pred %p8, %p149, %p150; setp.eq.f32 %p151, %f264, 0f00000000; @%p151 bra $L__BB0_116; bra.uni $L__BB0_113; $L__BB0_116: add.f32 %f1308, %f264, %f264; selp.f32 %f1714, %f1308, 0f00000000, %p150; bra.uni $L__BB0_117; $L__BB0_113: mov.b32 %r379, %f1712; xor.b32 %r380, %r379, -2147483648; mov.b32 %f1304, %r380; selp.f32 %f1714, %f1304, %f1712, %p8; setp.geu.f32 %p152, %f264, 0f00000000; @%p152 bra $L__BB0_117; mov.f32 %f1305, 0f3ED55555; cvt.rzi.f32.f32 %f1306, %f1305; setp.eq.f32 %p153, %f1306, 0f3ED55555; @%p153 bra $L__BB0_117; mov.f32 %f1714, 0f7FFFFFFF; $L__BB0_117: add.f32 %f1309, %f268, 0f3ED55555; mov.b32 %r381, %f1309; setp.lt.s32 %p155, %r381, 2139095040; @%p155 bra $L__BB0_122; setp.gtu.f32 %p156, %f268, 0f7F800000; @%p156 bra $L__BB0_121; bra.uni $L__BB0_119; $L__BB0_121: add.f32 %f1714, %f264, 0f3ED55555; bra.uni $L__BB0_122; $L__BB0_119: setp.neu.f32 %p157, %f268, 0f7F800000; @%p157 bra $L__BB0_122; selp.f32 %f1714, 0fFF800000, 0f7F800000, %p8; $L__BB0_122: abs.f32 %f279, %f265; setp.lt.f32 %p158, %f279, 0f00800000; mul.f32 %f1311, %f279, 0f4B800000; selp.f32 %f1312, %f1311, %f279, %p158; selp.f32 %f1313, 0fC3170000, 0fC2FE0000, %p158; mov.b32 %r382, %f1312; and.b32 %r383, %r382, 8388607; or.b32 %r384, %r383, 1065353216; mov.b32 %f1314, %r384; shr.u32 %r385, %r382, 23; cvt.rn.f32.u32 %f1315, %r385; add.f32 %f1316, %f1313, %f1315; setp.gt.f32 %p159, %f1314, 0f3FB504F3; mul.f32 %f1317, %f1314, 0f3F000000; add.f32 %f1318, %f1316, 0f3F800000; selp.f32 %f1319, %f1318, %f1316, %p159; selp.f32 %f1320, %f1317, %f1314, %p159; add.f32 %f1321, %f1320, 0fBF800000; add.f32 %f1322, %f1320, 0f3F800000; rcp.approx.ftz.f32 %f1323, %f1322; add.f32 %f1324, %f1321, %f1321; mul.f32 %f1325, %f1324, %f1323; mul.f32 %f1326, %f1325, %f1325; mov.f32 %f1327, 0f3C4CAF63; mov.f32 %f1328, 0f3B18F0FE; fma.rn.f32 %f1329, %f1328, %f1326, %f1327; mov.f32 %f1330, 0f3DAAAABD; fma.rn.f32 %f1331, %f1329, %f1326, %f1330; mul.rn.f32 %f1332, %f1331, %f1326; mul.rn.f32 %f1333, %f1332, %f1325; sub.f32 %f1334, %f1321, %f1325; add.f32 %f1335, %f1334, %f1334; neg.f32 %f1336, %f1325; fma.rn.f32 %f1337, %f1336, %f1321, %f1335; mul.rn.f32 %f1338, %f1323, %f1337; add.f32 %f1339, %f1333, %f1325; sub.f32 %f1340, %f1325, %f1339; add.f32 %f1341, %f1333, %f1340; add.f32 %f1342, %f1338, %f1341; add.f32 %f1343, %f1339, %f1342; sub.f32 %f1344, %f1339, %f1343; add.f32 %f1345, %f1342, %f1344; mov.f32 %f1346, 0f3F317200; mul.rn.f32 %f1347, %f1319, %f1346; mov.f32 %f1348, 0f35BFBE8E; mul.rn.f32 %f1349, %f1319, %f1348; add.f32 %f1350, %f1347, %f1343; sub.f32 %f1351, %f1347, %f1350; add.f32 %f1352, %f1343, %f1351; add.f32 %f1353, %f1345, %f1352; add.f32 %f1354, %f1349, %f1353; add.f32 %f1355, %f1350, %f1354; sub.f32 %f1356, %f1350, %f1355; add.f32 %f1357, %f1354, %f1356; mov.f32 %f1358, 0f3ED55555; mul.rn.f32 %f1359, %f1358, %f1355; neg.f32 %f1360, %f1359; fma.rn.f32 %f1361, %f1358, %f1355, %f1360; fma.rn.f32 %f1362, %f1358, %f1357, %f1361; mov.f32 %f1363, 0f00000000; fma.rn.f32 %f1364, %f1363, %f1355, %f1362; add.rn.f32 %f1365, %f1359, %f1364; neg.f32 %f1366, %f1365; add.rn.f32 %f1367, %f1359, %f1366; add.rn.f32 %f1368, %f1367, %f1364; mov.b32 %r386, %f1365; setp.eq.s32 %p160, %r386, 1118925336; add.s32 %r387, %r386, -1; mov.b32 %f1369, %r387; add.f32 %f1370, %f1368, 0f37000000; selp.f32 %f280, %f1370, %f1368, %p160; selp.f32 %f1371, %f1369, %f1365, %p160; mov.f32 %f1372, 0f3FB8AA3B; mul.rn.f32 %f1373, %f1371, %f1372; cvt.rzi.f32.f32 %f1374, %f1373; abs.f32 %f1375, %f1374; setp.gt.f32 %p161, %f1375, 0f42FC0000; mov.b32 %r388, %f1374; and.b32 %r389, %r388, -2147483648; or.b32 %r390, %r389, 1123811328; mov.b32 %f1376, %r390; selp.f32 %f1377, %f1376, %f1374, %p161; mov.f32 %f1378, 0fBF317218; fma.rn.f32 %f1379, %f1377, %f1378, %f1371; mov.f32 %f1380, 0f3102E308; fma.rn.f32 %f1381, %f1377, %f1380, %f1379; mul.f32 %f1382, %f1381, 0f3FB8AA3B; add.f32 %f1383, %f1377, 0f4B40007F; mov.b32 %r391, %f1383; shl.b32 %r392, %r391, 23; mov.b32 %f1384, %r392; ex2.approx.ftz.f32 %f1385, %f1382; mul.f32 %f281, %f1385, %f1384; setp.eq.f32 %p162, %f281, 0f7F800000; mov.f32 %f1715, 0f7F800000; @%p162 bra $L__BB0_124; fma.rn.f32 %f1715, %f281, %f280, %f281; $L__BB0_124: setp.lt.f32 %p163, %f265, 0f00000000; and.pred %p9, %p163, %p150; setp.eq.f32 %p165, %f265, 0f00000000; @%p165 bra $L__BB0_128; bra.uni $L__BB0_125; $L__BB0_128: add.f32 %f1390, %f265, %f265; selp.f32 %f1717, %f1390, 0f00000000, %p150; bra.uni $L__BB0_129; $L__BB0_125: mov.b32 %r393, %f1715; xor.b32 %r394, %r393, -2147483648; mov.b32 %f1386, %r394; selp.f32 %f1717, %f1386, %f1715, %p9; setp.geu.f32 %p166, %f265, 0f00000000; @%p166 bra $L__BB0_129; mov.f32 %f1387, 0f3ED55555; cvt.rzi.f32.f32 %f1388, %f1387; setp.eq.f32 %p167, %f1388, 0f3ED55555; @%p167 bra $L__BB0_129; mov.f32 %f1717, 0f7FFFFFFF; $L__BB0_129: add.f32 %f1391, %f279, 0f3ED55555; mov.b32 %r395, %f1391; setp.lt.s32 %p169, %r395, 2139095040; @%p169 bra $L__BB0_134; setp.gtu.f32 %p170, %f279, 0f7F800000; @%p170 bra $L__BB0_133; bra.uni $L__BB0_131; $L__BB0_133: add.f32 %f1717, %f265, 0f3ED55555; bra.uni $L__BB0_134; $L__BB0_131: setp.neu.f32 %p171, %f279, 0f7F800000; @%p171 bra $L__BB0_134; selp.f32 %f1717, 0fFF800000, 0f7F800000, %p9; $L__BB0_134: abs.f32 %f290, %f266; setp.lt.f32 %p172, %f290, 0f00800000; mul.f32 %f1393, %f290, 0f4B800000; selp.f32 %f1394, %f1393, %f290, %p172; selp.f32 %f1395, 0fC3170000, 0fC2FE0000, %p172; mov.b32 %r396, %f1394; and.b32 %r397, %r396, 8388607; or.b32 %r398, %r397, 1065353216; mov.b32 %f1396, %r398; shr.u32 %r399, %r396, 23; cvt.rn.f32.u32 %f1397, %r399; add.f32 %f1398, %f1395, %f1397; setp.gt.f32 %p173, %f1396, 0f3FB504F3; mul.f32 %f1399, %f1396, 0f3F000000; add.f32 %f1400, %f1398, 0f3F800000; selp.f32 %f1401, %f1400, %f1398, %p173; selp.f32 %f1402, %f1399, %f1396, %p173; add.f32 %f1403, %f1402, 0fBF800000; add.f32 %f1404, %f1402, 0f3F800000; rcp.approx.ftz.f32 %f1405, %f1404; add.f32 %f1406, %f1403, %f1403; mul.f32 %f1407, %f1406, %f1405; mul.f32 %f1408, %f1407, %f1407; mov.f32 %f1409, 0f3C4CAF63; mov.f32 %f1410, 0f3B18F0FE; fma.rn.f32 %f1411, %f1410, %f1408, %f1409; mov.f32 %f1412, 0f3DAAAABD; fma.rn.f32 %f1413, %f1411, %f1408, %f1412; mul.rn.f32 %f1414, %f1413, %f1408; mul.rn.f32 %f1415, %f1414, %f1407; sub.f32 %f1416, %f1403, %f1407; add.f32 %f1417, %f1416, %f1416; neg.f32 %f1418, %f1407; fma.rn.f32 %f1419, %f1418, %f1403, %f1417; mul.rn.f32 %f1420, %f1405, %f1419; add.f32 %f1421, %f1415, %f1407; sub.f32 %f1422, %f1407, %f1421; add.f32 %f1423, %f1415, %f1422; add.f32 %f1424, %f1420, %f1423; add.f32 %f1425, %f1421, %f1424; sub.f32 %f1426, %f1421, %f1425; add.f32 %f1427, %f1424, %f1426; mov.f32 %f1428, 0f3F317200; mul.rn.f32 %f1429, %f1401, %f1428; mov.f32 %f1430, 0f35BFBE8E; mul.rn.f32 %f1431, %f1401, %f1430; add.f32 %f1432, %f1429, %f1425; sub.f32 %f1433, %f1429, %f1432; add.f32 %f1434, %f1425, %f1433; add.f32 %f1435, %f1427, %f1434; add.f32 %f1436, %f1431, %f1435; add.f32 %f1437, %f1432, %f1436; sub.f32 %f1438, %f1432, %f1437; add.f32 %f1439, %f1436, %f1438; mov.f32 %f1440, 0f3ED55555; mul.rn.f32 %f1441, %f1440, %f1437; neg.f32 %f1442, %f1441; fma.rn.f32 %f1443, %f1440, %f1437, %f1442; fma.rn.f32 %f1444, %f1440, %f1439, %f1443; mov.f32 %f1445, 0f00000000; fma.rn.f32 %f1446, %f1445, %f1437, %f1444; add.rn.f32 %f1447, %f1441, %f1446; neg.f32 %f1448, %f1447; add.rn.f32 %f1449, %f1441, %f1448; add.rn.f32 %f1450, %f1449, %f1446; mov.b32 %r400, %f1447; setp.eq.s32 %p174, %r400, 1118925336; add.s32 %r401, %r400, -1; mov.b32 %f1451, %r401; add.f32 %f1452, %f1450, 0f37000000; selp.f32 %f291, %f1452, %f1450, %p174; selp.f32 %f1453, %f1451, %f1447, %p174; mov.f32 %f1454, 0f3FB8AA3B; mul.rn.f32 %f1455, %f1453, %f1454; cvt.rzi.f32.f32 %f1456, %f1455; abs.f32 %f1457, %f1456; setp.gt.f32 %p175, %f1457, 0f42FC0000; mov.b32 %r402, %f1456; and.b32 %r403, %r402, -2147483648; or.b32 %r404, %r403, 1123811328; mov.b32 %f1458, %r404; selp.f32 %f1459, %f1458, %f1456, %p175; mov.f32 %f1460, 0fBF317218; fma.rn.f32 %f1461, %f1459, %f1460, %f1453; mov.f32 %f1462, 0f3102E308; fma.rn.f32 %f1463, %f1459, %f1462, %f1461; mul.f32 %f1464, %f1463, 0f3FB8AA3B; add.f32 %f1465, %f1459, 0f4B40007F; mov.b32 %r405, %f1465; shl.b32 %r406, %r405, 23; mov.b32 %f1466, %r406; ex2.approx.ftz.f32 %f1467, %f1464; mul.f32 %f292, %f1467, %f1466; setp.eq.f32 %p176, %f292, 0f7F800000; mov.f32 %f1718, 0f7F800000; @%p176 bra $L__BB0_136; fma.rn.f32 %f1718, %f292, %f291, %f292; $L__BB0_136: setp.lt.f32 %p177, %f266, 0f00000000; and.pred %p10, %p177, %p150; setp.eq.f32 %p179, %f266, 0f00000000; @%p179 bra $L__BB0_140; bra.uni $L__BB0_137; $L__BB0_140: add.f32 %f1472, %f266, %f266; selp.f32 %f1720, %f1472, 0f00000000, %p150; bra.uni $L__BB0_141; $L__BB0_137: mov.b32 %r407, %f1718; xor.b32 %r408, %r407, -2147483648; mov.b32 %f1468, %r408; selp.f32 %f1720, %f1468, %f1718, %p10; setp.geu.f32 %p180, %f266, 0f00000000; @%p180 bra $L__BB0_141; mov.f32 %f1469, 0f3ED55555; cvt.rzi.f32.f32 %f1470, %f1469; setp.eq.f32 %p181, %f1470, 0f3ED55555; @%p181 bra $L__BB0_141; mov.f32 %f1720, 0f7FFFFFFF; $L__BB0_141: add.f32 %f1473, %f290, 0f3ED55555; mov.b32 %r409, %f1473; setp.lt.s32 %p183, %r409, 2139095040; @%p183 bra $L__BB0_146; setp.gtu.f32 %p184, %f290, 0f7F800000; @%p184 bra $L__BB0_145; bra.uni $L__BB0_143; $L__BB0_145: add.f32 %f1720, %f266, 0f3ED55555; bra.uni $L__BB0_146; $L__BB0_143: setp.neu.f32 %p185, %f290, 0f7F800000; @%p185 bra $L__BB0_146; selp.f32 %f1720, 0fFF800000, 0f7F800000, %p10; $L__BB0_146: fma.rn.f32 %f1474, %f1714, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p186, %f264, 0f3F800000; mov.f32 %f1475, 0f3F800000; selp.f32 %f1476, 0f3F7FFFFF, %f1474, %p186; mul.f32 %f1477, %f264, 0f414EB852; setp.lt.f32 %p187, %f264, 0f3B4D2E1C; selp.f32 %f1478, %f1477, %f1476, %p187; fma.rn.f32 %f1479, %f1717, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p188, %f265, 0f3F800000; selp.f32 %f1480, 0f3F7FFFFF, %f1479, %p188; mul.f32 %f1481, %f265, 0f414EB852; setp.lt.f32 %p189, %f265, 0f3B4D2E1C; selp.f32 %f1482, %f1481, %f1480, %p189; fma.rn.f32 %f1483, %f1720, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p190, %f266, 0f3F800000; selp.f32 %f1484, 0f3F7FFFFF, %f1483, %p190; mul.f32 %f1485, %f266, 0f414EB852; setp.lt.f32 %p191, %f266, 0f3B4D2E1C; selp.f32 %f1486, %f1485, %f1484, %p191; min.f32 %f1487, %f1478, %f1475; mov.f32 %f1488, 0f00000000; max.f32 %f1489, %f1488, %f1487; mul.f32 %f1490, %f1489, 0f43800000; cvt.rzi.u32.f32 %r410, %f1490; min.u32 %r411, %r410, 255; min.f32 %f1491, %f1482, %f1475; max.f32 %f1492, %f1488, %f1491; mul.f32 %f1493, %f1492, 0f43800000; cvt.rzi.u32.f32 %r412, %f1493; min.u32 %r413, %r412, 255; min.f32 %f1494, %f1486, %f1475; max.f32 %f1495, %f1488, %f1494; mul.f32 %f1496, %f1495, 0f43800000; cvt.rzi.u32.f32 %r414, %f1496; min.u32 %r415, %r414, 255; shl.b64 %rd37, %rd10, 2; add.s64 %rd38, %rd9, %rd37; cvt.u16.u32 %rs20, %r415; cvt.u16.u32 %rs21, %r413; cvt.u16.u32 %rs22, %r411; mov.u16 %rs23, 255; st.global.v4.u8 [%rd38], {%rs22, %rs21, %rs20, %rs23}; $L__BB0_147: and.b32 %r416, %r451, 4; setp.eq.s32 %p192, %r416, 0; ld.const.u32 %r452, [params+108]; @%p192 bra $L__BB0_151; setp.eq.s32 %p193, %r452, 0; ld.const.u64 %rd39, [params+224]; cvta.to.global.u64 %rd40, %rd39; ld.const.u32 %r417, [params+216]; mad.lo.s32 %r418, %r417, %r4, %r3; mul.wide.u32 %rd41, %r418, 8; add.s64 %rd11, %rd40, %rd41; @%p193 bra $L__BB0_150; ld.global.v4.u16 {%rs31, %rs32, %rs33, %rs34}, [%rd11]; // begin inline asm { cvt.f32.f16 %f1497, %rs31;} // end inline asm // begin inline asm { cvt.f32.f16 %f1498, %rs32;} // end inline asm // begin inline asm { cvt.f32.f16 %f1499, %rs33;} // end inline asm add.f32 %f1500, %f1668, %f1497; add.f32 %f1501, %f1669, %f1498; add.f32 %f1502, %f1670, %f1499; mov.f32 %f1503, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs29, %f1502;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs28, %f1501;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs27, %f1500;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs30, %f1503;} // end inline asm st.global.v4.u16 [%rd11], {%rs27, %rs28, %rs29, %rs30}; bra.uni $L__BB0_151; $L__BB0_150: mov.f32 %f1507, 0f3F800000; mov.u32 %r452, 0; // begin inline asm { cvt.rn.f16.f32 %rs38, %f1507;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs37, %f1670;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs36, %f1669;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs35, %f1668;} // end inline asm st.global.v4.u16 [%rd11], {%rs35, %rs36, %rs37, %rs38}; $L__BB0_151: mov.f32 %f1508, 0f34000000; max.f32 %f1509, %f1671, %f1508; div.rn.f32 %f1510, %f1674, %f1509; max.f32 %f1511, %f1672, %f1508; div.rn.f32 %f1512, %f1675, %f1511; max.f32 %f1513, %f1673, %f1508; div.rn.f32 %f1514, %f1676, %f1513; fma.rn.f32 %f301, %f1510, 0f3F000000, 0f3F000000; fma.rn.f32 %f302, %f1512, 0f3F000000, 0f3F000000; fma.rn.f32 %f303, %f1514, 0f3F000000, 0f3F000000; div.rn.f32 %f1515, %f1677, %f1509; div.rn.f32 %f1516, %f1678, %f1511; div.rn.f32 %f1517, %f1679, %f1513; fma.rn.f32 %f304, %f1515, 0f3F000000, 0f3F000000; fma.rn.f32 %f305, %f1516, 0f3F000000, 0f3F000000; fma.rn.f32 %f306, %f1517, 0f3F000000, 0f3F000000; div.rn.f32 %f1518, %f1680, %f1509; div.rn.f32 %f1519, %f1681, %f1511; div.rn.f32 %f1520, %f1682, %f1513; fma.rn.f32 %f307, %f1518, 0f3F000000, 0f3F000000; fma.rn.f32 %f308, %f1519, 0f3F000000, 0f3F000000; fma.rn.f32 %f309, %f1520, 0f3F000000, 0f3F000000; ld.const.u64 %rd42, [params+256]; cvta.to.global.u64 %rd43, %rd42; ld.const.u32 %r420, [params+248]; mad.lo.s32 %r421, %r420, %r4, %r3; mul.wide.u32 %rd44, %r421, 8; add.s64 %rd12, %rd43, %rd44; setp.eq.s32 %p194, %r452, 0; @%p194 bra $L__BB0_153; ld.global.v4.u16 {%rs46, %rs47, %rs48, %rs49}, [%rd12]; // begin inline asm { cvt.f32.f16 %f1521, %rs46;} // end inline asm // begin inline asm { cvt.f32.f16 %f1522, %rs47;} // end inline asm // begin inline asm { cvt.f32.f16 %f1523, %rs48;} // end inline asm add.f32 %f1524, %f1671, %f1521; add.f32 %f1525, %f1672, %f1522; add.f32 %f1526, %f1673, %f1523; mov.f32 %f1527, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs44, %f1526;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs43, %f1525;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs42, %f1524;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs45, %f1527;} // end inline asm st.global.v4.u16 [%rd12], {%rs42, %rs43, %rs44, %rs45}; bra.uni $L__BB0_154; $L__BB0_153: mov.f32 %f1531, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs53, %f1531;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs52, %f1673;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs51, %f1672;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs50, %f1671;} // end inline asm st.global.v4.u16 [%rd12], {%rs50, %rs51, %rs52, %rs53}; $L__BB0_154: ld.const.u64 %rd45, [params+272]; cvta.to.global.u64 %rd46, %rd45; ld.const.u32 %r422, [params+264]; mad.lo.s32 %r423, %r422, %r4, %r3; mul.wide.u32 %rd47, %r423, 8; add.s64 %rd13, %rd46, %rd47; @%p194 bra $L__BB0_156; ld.global.v4.u16 {%rs61, %rs62, %rs63, %rs64}, [%rd13]; // begin inline asm { cvt.f32.f16 %f1532, %rs61;} // end inline asm // begin inline asm { cvt.f32.f16 %f1533, %rs62;} // end inline asm // begin inline asm { cvt.f32.f16 %f1534, %rs63;} // end inline asm add.f32 %f1535, %f301, %f1532; add.f32 %f1536, %f302, %f1533; add.f32 %f1537, %f303, %f1534; mov.f32 %f1538, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs59, %f1537;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs58, %f1536;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs57, %f1535;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs60, %f1538;} // end inline asm st.global.v4.u16 [%rd13], {%rs57, %rs58, %rs59, %rs60}; bra.uni $L__BB0_157; $L__BB0_156: mov.f32 %f1542, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs68, %f1542;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs67, %f303;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs66, %f302;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs65, %f301;} // end inline asm st.global.v4.u16 [%rd13], {%rs65, %rs66, %rs67, %rs68}; $L__BB0_157: ld.const.u64 %rd48, [params+288]; cvta.to.global.u64 %rd49, %rd48; ld.const.u32 %r424, [params+280]; mad.lo.s32 %r425, %r424, %r4, %r3; mul.wide.u32 %rd50, %r425, 8; add.s64 %rd14, %rd49, %rd50; @%p194 bra $L__BB0_159; ld.global.v4.u16 {%rs76, %rs77, %rs78, %rs79}, [%rd14]; // begin inline asm { cvt.f32.f16 %f1543, %rs76;} // end inline asm // begin inline asm { cvt.f32.f16 %f1544, %rs77;} // end inline asm // begin inline asm { cvt.f32.f16 %f1545, %rs78;} // end inline asm add.f32 %f1546, %f304, %f1543; add.f32 %f1547, %f305, %f1544; add.f32 %f1548, %f306, %f1545; mov.f32 %f1549, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs74, %f1548;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs73, %f1547;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs72, %f1546;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs75, %f1549;} // end inline asm st.global.v4.u16 [%rd14], {%rs72, %rs73, %rs74, %rs75}; bra.uni $L__BB0_160; $L__BB0_159: mov.f32 %f1553, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs83, %f1553;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs82, %f306;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs81, %f305;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs80, %f304;} // end inline asm st.global.v4.u16 [%rd14], {%rs80, %rs81, %rs82, %rs83}; $L__BB0_160: ld.const.u64 %rd51, [params+304]; cvta.to.global.u64 %rd52, %rd51; ld.const.u32 %r426, [params+296]; mad.lo.s32 %r427, %r426, %r4, %r3; mul.wide.u32 %rd53, %r427, 8; add.s64 %rd15, %rd52, %rd53; @%p194 bra $L__BB0_162; ld.global.v4.u16 {%rs91, %rs92, %rs93, %rs94}, [%rd15]; // begin inline asm { cvt.f32.f16 %f1554, %rs91;} // end inline asm // begin inline asm { cvt.f32.f16 %f1555, %rs92;} // end inline asm // begin inline asm { cvt.f32.f16 %f1556, %rs93;} // end inline asm add.f32 %f1557, %f307, %f1554; add.f32 %f1558, %f308, %f1555; add.f32 %f1559, %f309, %f1556; mov.f32 %f1560, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs89, %f1559;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs88, %f1558;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs87, %f1557;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs90, %f1560;} // end inline asm st.global.v4.u16 [%rd15], {%rs87, %rs88, %rs89, %rs90}; bra.uni $L__BB0_183; $L__BB0_162: mov.f32 %f1564, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs98, %f1564;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs97, %f309;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs96, %f308;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs95, %f307;} // end inline asm st.global.v4.u16 [%rd15], {%rs95, %rs96, %rs97, %rs98}; $L__BB0_183: ret; }