// // Generated by NVIDIA NVVM Compiler // // Compiler Build ID: CL-31968024 // Cuda compilation tools, release 12.0, V12.0.76 // Based on NVVM 7.0.1 // .version 8.0 .target sm_52 .address_size 64 // .globl __raygen__oxMain .const .align 16 .b8 params[1184]; .global .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162}; .visible .entry __raygen__oxMain() { .local .align 4 .b8 __local_depot0[28]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<205>; .reg .b16 %rs<180>; .reg .f32 %f<1703>; .reg .b32 %r<467>; .reg .f64 %fd<5>; .reg .b64 %rd<121>; mov.u64 %SPL, __local_depot0; add.u64 %rd1, %SPL, 0; // begin inline asm call (%r62), _optix_get_launch_index_x, (); // end inline asm // begin inline asm call (%r63), _optix_get_launch_index_y, (); // end inline asm ld.const.u32 %r3, [params+616]; abs.s32 %r4, %r3; ld.const.u32 %r5, [params+540]; ld.const.u64 %rd33, [params+400]; cvta.to.global.u64 %rd34, %rd33; ld.const.u32 %r68, [params+392]; mad.lo.s32 %r69, %r68, %r63, %r62; mul.wide.u32 %rd35, %r69, 4; add.s64 %rd2, %rd34, %rd35; ld.global.v2.u8 {%rs16, %rs175}, [%rd2]; or.b16 %rs18, %rs16, %rs175; and.b16 %rs19, %rs18, 255; setp.eq.s16 %p10, %rs19, 0; @%p10 bra $L__BB0_2; ld.global.u8 %rs174, [%rd2+2]; bra.uni $L__BB0_3; $L__BB0_2: ld.global.u8 %rs174, [%rd2+2]; setp.eq.s16 %p11, %rs174, 0; mov.f32 %f1560, 0f00000000; mov.u16 %rs175, 0; mov.f32 %f1561, %f1560; mov.f32 %f1562, %f1560; @%p11 bra $L__BB0_4; $L__BB0_3: cvt.rn.f32.u16 %f367, %rs16; div.rn.f32 %f368, %f367, 0f437F0000; fma.rn.f32 %f369, %f368, 0f40000000, 0fBF800000; and.b16 %rs22, %rs175, 255; cvt.rn.f32.u16 %f370, %rs22; div.rn.f32 %f371, %f370, 0f437F0000; fma.rn.f32 %f372, %f371, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f373, %rs174; div.rn.f32 %f374, %f373, 0f437F0000; fma.rn.f32 %f375, %f374, 0f40000000, 0fBF800000; mul.f32 %f376, %f372, %f372; fma.rn.f32 %f377, %f369, %f369, %f376; fma.rn.f32 %f378, %f375, %f375, %f377; sqrt.rn.f32 %f379, %f378; rcp.rn.f32 %f380, %f379; mul.f32 %f1562, %f380, %f375; mul.f32 %f1561, %f380, %f372; mul.f32 %f1560, %f369, %f380; $L__BB0_4: ld.const.v2.u32 {%r70, %r71}, [params]; add.s32 %r6, %r70, %r62; add.s32 %r7, %r71, %r63; setp.eq.f32 %p12, %f1560, 0f00000000; setp.eq.f32 %p13, %f1561, 0f00000000; and.pred %p14, %p12, %p13; setp.eq.f32 %p15, %f1562, 0f00000000; and.pred %p16, %p15, %p14; @%p16 bra $L__BB0_165; bra.uni $L__BB0_5; $L__BB0_165: ld.const.u32 %r59, [params+104]; and.b32 %r432, %r59, 1; setp.eq.b32 %p194, %r432, 1; mov.pred %p195, 0; xor.pred %p196, %p194, %p195; not.pred %p197, %p196; @%p197 bra $L__BB0_167; ld.const.u64 %rd96, [params+144]; cvta.to.global.u64 %rd97, %rd96; ld.const.u32 %r433, [params+136]; mad.lo.s32 %r434, %r433, %r7, %r6; mul.wide.u32 %rd98, %r434, 4; add.s64 %rd99, %rd97, %rd98; mov.u16 %rs111, 0; st.global.v4.u8 [%rd99], {%rs111, %rs111, %rs111, %rs111}; $L__BB0_167: and.b32 %r435, %r59, 8; setp.eq.s32 %p198, %r435, 0; @%p198 bra $L__BB0_169; ld.const.u64 %rd100, [params+192]; cvta.to.global.u64 %rd101, %rd100; ld.const.u32 %r436, [params+184]; mad.lo.s32 %r437, %r436, %r7, %r6; mov.f32 %f1515, 0f00000000; cvt.rzi.u32.f32 %r438, %f1515; mul.wide.u32 %rd102, %r437, 2; add.s64 %rd103, %rd101, %rd102; mov.u16 %rs112, 0; cvt.u16.u32 %rs113, %r438; st.global.v2.u8 [%rd103], {%rs113, %rs112}; $L__BB0_169: and.b32 %r439, %r59, 4; setp.eq.s32 %p199, %r439, 0; ld.const.u32 %r466, [params+108]; @%p199 bra $L__BB0_173; setp.eq.s32 %p200, %r466, 0; ld.const.u64 %rd104, [params+224]; cvta.to.global.u64 %rd105, %rd104; ld.const.u32 %r440, [params+216]; mad.lo.s32 %r441, %r440, %r7, %r6; mul.wide.u32 %rd106, %r441, 8; add.s64 %rd28, %rd105, %rd106; @%p200 bra $L__BB0_172; ld.global.v4.u16 {%rs120, %rs121, %rs122, %rs123}, [%rd28]; // begin inline asm { cvt.f32.f16 %f1516, %rs120;} // end inline asm // begin inline asm { cvt.f32.f16 %f1517, %rs121;} // end inline asm // begin inline asm { cvt.f32.f16 %f1518, %rs122;} // end inline asm add.f32 %f1519, %f1516, 0f00000000; add.f32 %f1520, %f1517, 0f00000000; add.f32 %f1521, %f1518, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs119, %f1521;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs118, %f1520;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs117, %f1519;} // end inline asm mov.u16 %rs124, 0; st.global.v4.u16 [%rd28], {%rs117, %rs118, %rs119, %rs124}; bra.uni $L__BB0_173; $L__BB0_5: ld.const.u64 %rd36, [params+432]; cvta.to.global.u64 %rd37, %rd36; ld.const.u32 %r74, [params+424]; mad.lo.s32 %r75, %r74, %r63, %r62; mul.wide.u32 %rd38, %r75, 12; add.s64 %rd39, %rd37, %rd38; ld.global.f32 %f10, [%rd39]; mul.f32 %f381, %f10, 0f3456BF95; ld.global.f32 %f11, [%rd39+4]; mul.f32 %f382, %f11, 0f3456BF95; ld.global.f32 %f12, [%rd39+8]; mul.f32 %f383, %f12, 0f3456BF95; abs.f32 %f13, %f1560; div.rn.f32 %f384, %f381, %f13; abs.f32 %f385, %f1561; div.rn.f32 %f386, %f382, %f385; abs.f32 %f14, %f1562; div.rn.f32 %f387, %f383, %f14; abs.f32 %f388, %f384; abs.f32 %f389, %f386; abs.f32 %f390, %f387; mov.f32 %f391, 0f38D1B717; max.f32 %f392, %f388, %f391; max.f32 %f393, %f389, %f391; max.f32 %f394, %f390, %f391; fma.rn.f32 %f15, %f1560, %f392, %f10; fma.rn.f32 %f16, %f1561, %f393, %f11; fma.rn.f32 %f17, %f1562, %f394, %f12; ld.const.u64 %rd40, [params+464]; cvta.to.global.u64 %rd41, %rd40; ld.const.u32 %r76, [params+456]; mad.lo.s32 %r77, %r76, %r63, %r62; mul.wide.u32 %rd42, %r77, 4; add.s64 %rd3, %rd41, %rd42; ld.global.v2.u8 {%rs24, %rs178}, [%rd3]; or.b16 %rs26, %rs24, %rs178; and.b16 %rs27, %rs26, 255; setp.eq.s16 %p17, %rs27, 0; @%p17 bra $L__BB0_7; ld.global.u8 %rs176, [%rd3+2]; bra.uni $L__BB0_8; $L__BB0_7: ld.global.u8 %rs176, [%rd3+2]; setp.eq.s16 %p18, %rs176, 0; mov.f32 %f1563, 0f00000000; mov.u16 %rs178, 0; mov.u16 %rs179, %rs178; mov.f32 %f1564, %f1563; mov.f32 %f1565, %f1563; @%p18 bra $L__BB0_9; $L__BB0_8: mov.u16 %rs179, %rs178; cvt.rn.f32.u16 %f398, %rs24; div.rn.f32 %f399, %f398, 0f437F0000; fma.rn.f32 %f400, %f399, 0f40000000, 0fBF800000; and.b16 %rs32, %rs179, 255; cvt.rn.f32.u16 %f401, %rs32; div.rn.f32 %f402, %f401, 0f437F0000; fma.rn.f32 %f403, %f402, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f404, %rs176; div.rn.f32 %f405, %f404, 0f437F0000; fma.rn.f32 %f406, %f405, 0f40000000, 0fBF800000; mul.f32 %f407, %f403, %f403; fma.rn.f32 %f408, %f400, %f400, %f407; fma.rn.f32 %f409, %f406, %f406, %f408; sqrt.rn.f32 %f410, %f409; rcp.rn.f32 %f411, %f410; mul.f32 %f1565, %f411, %f406; mul.f32 %f1564, %f411, %f403; mul.f32 %f1563, %f400, %f411; mov.u16 %rs178, %rs176; $L__BB0_9: mul.f32 %f415, %f1561, %f1565; mul.f32 %f416, %f1562, %f1564; sub.f32 %f417, %f415, %f416; mul.f32 %f418, %f1562, %f1563; mul.f32 %f419, %f1560, %f1565; sub.f32 %f420, %f418, %f419; mul.f32 %f421, %f1560, %f1564; mul.f32 %f422, %f1561, %f1563; sub.f32 %f423, %f421, %f422; ld.global.u8 %rs34, [%rd3+3]; setp.eq.s16 %p19, %rs34, 0; selp.f32 %f24, 0fBF800000, 0f3F800000, %p19; mul.f32 %f424, %f417, %f24; mul.f32 %f425, %f420, %f24; mul.f32 %f426, %f423, %f24; mul.f32 %f427, %f424, 0f00000000; mul.f32 %f428, %f425, 0f00000000; mul.f32 %f429, %f426, 0f00000000; fma.rn.f32 %f430, %f1563, 0f3F5105EC, %f427; fma.rn.f32 %f431, %f1564, 0f3F5105EC, %f428; fma.rn.f32 %f432, %f1565, 0f3F5105EC, %f429; mul.f32 %f25, %f1560, 0f3F13CD3A; add.f32 %f26, %f25, %f430; mul.f32 %f27, %f1561, 0f3F13CD3A; add.f32 %f28, %f27, %f431; mul.f32 %f29, %f1562, 0f3F13CD3A; add.f32 %f30, %f29, %f432; or.b16 %rs35, %rs24, %rs179; or.b16 %rs15, %rs35, %rs178; and.b16 %rs36, %rs15, 255; setp.eq.s16 %p20, %rs36, 0; mov.f32 %f1569, 0f00000000; mov.f32 %f1566, %f1569; mov.f32 %f1567, %f1569; mov.f32 %f1568, %f1569; @%p20 bra $L__BB0_11; cvt.rn.f32.u16 %f433, %rs24; div.rn.f32 %f434, %f433, 0f437F0000; fma.rn.f32 %f435, %f434, 0f40000000, 0fBF800000; and.b16 %rs38, %rs179, 255; cvt.rn.f32.u16 %f436, %rs38; div.rn.f32 %f437, %f436, 0f437F0000; fma.rn.f32 %f438, %f437, 0f40000000, 0fBF800000; and.b16 %rs39, %rs178, 255; cvt.rn.f32.u16 %f439, %rs39; div.rn.f32 %f440, %f439, 0f437F0000; fma.rn.f32 %f441, %f440, 0f40000000, 0fBF800000; mul.f32 %f442, %f438, %f438; fma.rn.f32 %f443, %f435, %f435, %f442; fma.rn.f32 %f444, %f441, %f441, %f443; sqrt.rn.f32 %f445, %f444; rcp.rn.f32 %f446, %f445; mul.f32 %f1568, %f446, %f441; mul.f32 %f1567, %f446, %f438; mul.f32 %f1566, %f435, %f446; $L__BB0_11: mul.f32 %f450, %f1561, %f1568; mul.f32 %f451, %f1562, %f1567; sub.f32 %f452, %f450, %f451; mul.f32 %f453, %f1562, %f1566; mul.f32 %f454, %f1560, %f1568; sub.f32 %f455, %f453, %f454; mul.f32 %f456, %f1560, %f1567; mul.f32 %f457, %f1561, %f1566; sub.f32 %f458, %f456, %f457; mul.f32 %f459, %f452, %f24; mul.f32 %f460, %f455, %f24; mul.f32 %f461, %f458, %f24; mul.f32 %f462, %f1566, 0f3ED105EC; mul.f32 %f463, %f1567, 0f3ED105EC; mul.f32 %f464, %f1568, 0f3ED105EC; mul.f32 %f465, %f459, 0f3F3504F3; mul.f32 %f466, %f460, 0f3F3504F3; mul.f32 %f467, %f461, 0f3F3504F3; sub.f32 %f468, %f465, %f462; sub.f32 %f469, %f466, %f463; sub.f32 %f470, %f467, %f464; add.f32 %f37, %f25, %f468; add.f32 %f38, %f27, %f469; add.f32 %f39, %f29, %f470; mov.f32 %f1570, %f1569; mov.f32 %f1571, %f1569; @%p20 bra $L__BB0_13; cvt.rn.f32.u16 %f471, %rs24; div.rn.f32 %f472, %f471, 0f437F0000; fma.rn.f32 %f473, %f472, 0f40000000, 0fBF800000; and.b16 %rs42, %rs179, 255; cvt.rn.f32.u16 %f474, %rs42; div.rn.f32 %f475, %f474, 0f437F0000; fma.rn.f32 %f476, %f475, 0f40000000, 0fBF800000; and.b16 %rs43, %rs178, 255; cvt.rn.f32.u16 %f477, %rs43; div.rn.f32 %f478, %f477, 0f437F0000; fma.rn.f32 %f479, %f478, 0f40000000, 0fBF800000; mul.f32 %f480, %f476, %f476; fma.rn.f32 %f481, %f473, %f473, %f480; fma.rn.f32 %f482, %f479, %f479, %f481; sqrt.rn.f32 %f483, %f482; rcp.rn.f32 %f484, %f483; mul.f32 %f1571, %f484, %f479; mul.f32 %f1570, %f484, %f476; mul.f32 %f1569, %f473, %f484; $L__BB0_13: mul.f32 %f498, %f1561, %f1571; mul.f32 %f499, %f1562, %f1570; sub.f32 %f500, %f498, %f499; mul.f32 %f501, %f1562, %f1569; mul.f32 %f502, %f1560, %f1571; sub.f32 %f503, %f501, %f502; mul.f32 %f504, %f1560, %f1570; mul.f32 %f505, %f1561, %f1569; sub.f32 %f506, %f504, %f505; mul.f32 %f507, %f500, %f24; mul.f32 %f508, %f503, %f24; mul.f32 %f509, %f506, %f24; mul.f32 %f510, %f1569, 0f3ED105EC; mul.f32 %f511, %f1570, 0f3ED105EC; mul.f32 %f512, %f1571, 0f3ED105EC; mul.f32 %f513, %f507, 0fBF3504F3; mul.f32 %f514, %f508, 0fBF3504F3; mul.f32 %f515, %f509, 0fBF3504F3; sub.f32 %f516, %f513, %f510; sub.f32 %f517, %f514, %f511; sub.f32 %f518, %f515, %f512; add.f32 %f46, %f25, %f516; add.f32 %f47, %f27, %f517; add.f32 %f48, %f29, %f518; ld.const.u64 %rd43, [params+128]; cvta.to.global.u64 %rd44, %rd43; ld.const.u32 %r79, [params+120]; mad.lo.s32 %r80, %r79, %r63, %r62; mul.wide.u32 %rd45, %r80, 4; add.s64 %rd46, %rd44, %rd45; ld.global.u32 %r457, [%rd46]; setp.lt.s32 %p23, %r4, 1; mov.pred %p22, 0; mov.u32 %r453, 0; mov.f32 %f1588, 0f00000000; mov.f32 %f1589, %f1588; mov.f32 %f1590, %f1588; mov.f32 %f1591, %f1588; mov.f32 %f1592, %f1588; mov.f32 %f1593, %f1588; mov.f32 %f1594, %f1588; mov.f32 %f1595, %f1588; mov.f32 %f1596, %f1588; mov.f32 %f1597, %f1588; mov.f32 %f1598, %f1588; mov.f32 %f1599, %f1588; mov.f32 %f1600, %f1588; mov.pred %p204, %p22; @%p23 bra $L__BB0_33; ld.const.f32 %f49, [params+620]; ld.const.u64 %rd47, [params+624]; cvta.to.global.u64 %rd4, %rd47; ld.const.u64 %rd48, [params+640]; cvta.to.global.u64 %rd49, %rd48; ld.const.u32 %r83, [params+632]; and.b32 %r84, %r63, 255; and.b32 %r85, %r62, 255; mad.lo.s32 %r86, %r83, %r84, %r85; mul.wide.u32 %rd50, %r86, 3; add.s64 %rd5, %rd49, %rd50; ld.const.f32 %f50, [params+660]; mul.f32 %f51, %f15, 0f3456BF95; mul.f32 %f52, %f16, 0f3456BF95; mul.f32 %f53, %f17, 0f3456BF95; ld.const.u64 %rd6, [params+96]; abs.f32 %f659, %f52; abs.f32 %f660, %f51; max.f32 %f661, %f660, %f659; abs.f32 %f662, %f53; max.f32 %f663, %f661, %f662; mov.u32 %r450, %r453; $L__BB0_15: mul.lo.s32 %r87, %r450, 3; mul.wide.s32 %rd51, %r87, 12; add.s64 %rd7, %rd4, %rd51; ld.global.f32 %f532, [%rd7]; sub.f32 %f533, %f532, %f10; ld.global.f32 %f534, [%rd7+4]; sub.f32 %f535, %f534, %f11; ld.global.f32 %f536, [%rd7+8]; sub.f32 %f537, %f536, %f12; mul.f32 %f538, %f535, %f535; fma.rn.f32 %f539, %f533, %f533, %f538; fma.rn.f32 %f540, %f537, %f537, %f539; sqrt.rn.f32 %f67, %f540; rcp.rn.f32 %f541, %f67; mul.f32 %f68, %f533, %f541; mul.f32 %f69, %f535, %f541; mul.f32 %f70, %f537, %f541; mul.f32 %f542, %f1561, %f69; fma.rn.f32 %f543, %f1560, %f68, %f542; fma.rn.f32 %f71, %f1562, %f70, %f543; setp.leu.f32 %p24, %f71, 0f00000000; @%p24 bra $L__BB0_32; setp.ne.s32 %p26, %r5, 0; mul.f32 %f544, %f67, %f67; div.rn.f32 %f72, %f49, %f544; ld.global.u8 %rs44, [%rd5]; cvt.rn.f32.u16 %f545, %rs44; div.rn.f32 %f546, %f545, 0fC37F0000; fma.rn.f32 %f547, %f546, 0f3F333333, 0f3F800000; setp.gt.f32 %p27, %f72, %f547; and.pred %p28, %p26, %p27; mov.pred %p204, -1; @%p28 bra $L__BB0_33; mul.f32 %f73, %f67, %f50; mov.f32 %f552, 0f40800000; abs.f32 %f75, %f73; setp.lt.f32 %p29, %f75, 0f00800000; mul.f32 %f554, %f75, 0f4B800000; selp.f32 %f555, %f554, %f75, %p29; selp.f32 %f556, 0fC3170000, 0fC2FE0000, %p29; mov.b32 %r88, %f555; and.b32 %r89, %r88, 8388607; or.b32 %r90, %r89, 1065353216; mov.b32 %f557, %r90; shr.u32 %r91, %r88, 23; cvt.rn.f32.u32 %f558, %r91; add.f32 %f559, %f556, %f558; setp.gt.f32 %p30, %f557, 0f3FB504F3; mul.f32 %f560, %f557, 0f3F000000; add.f32 %f561, %f559, 0f3F800000; selp.f32 %f562, %f561, %f559, %p30; selp.f32 %f563, %f560, %f557, %p30; add.f32 %f564, %f563, 0fBF800000; add.f32 %f565, %f563, 0f3F800000; rcp.approx.ftz.f32 %f566, %f565; add.f32 %f567, %f564, %f564; mul.f32 %f568, %f567, %f566; mul.f32 %f569, %f568, %f568; mov.f32 %f570, 0f3C4CAF63; mov.f32 %f571, 0f3B18F0FE; fma.rn.f32 %f572, %f571, %f569, %f570; mov.f32 %f573, 0f3DAAAABD; fma.rn.f32 %f574, %f572, %f569, %f573; mul.rn.f32 %f575, %f574, %f569; mul.rn.f32 %f576, %f575, %f568; sub.f32 %f577, %f564, %f568; add.f32 %f578, %f577, %f577; neg.f32 %f579, %f568; fma.rn.f32 %f580, %f579, %f564, %f578; mul.rn.f32 %f581, %f566, %f580; add.f32 %f582, %f576, %f568; sub.f32 %f583, %f568, %f582; add.f32 %f584, %f576, %f583; add.f32 %f585, %f581, %f584; add.f32 %f586, %f582, %f585; sub.f32 %f587, %f582, %f586; add.f32 %f588, %f585, %f587; mov.f32 %f589, 0f3F317200; mul.rn.f32 %f590, %f562, %f589; mov.f32 %f591, 0f35BFBE8E; mul.rn.f32 %f592, %f562, %f591; add.f32 %f593, %f590, %f586; sub.f32 %f594, %f590, %f593; add.f32 %f595, %f586, %f594; add.f32 %f596, %f588, %f595; add.f32 %f597, %f592, %f596; add.f32 %f598, %f593, %f597; sub.f32 %f599, %f593, %f598; add.f32 %f600, %f597, %f599; mul.rn.f32 %f601, %f552, %f598; neg.f32 %f602, %f601; fma.rn.f32 %f603, %f552, %f598, %f602; fma.rn.f32 %f604, %f552, %f600, %f603; mov.f32 %f605, 0f00000000; fma.rn.f32 %f606, %f605, %f598, %f604; add.rn.f32 %f607, %f601, %f606; neg.f32 %f608, %f607; add.rn.f32 %f609, %f601, %f608; add.rn.f32 %f610, %f609, %f606; mov.b32 %r92, %f607; setp.eq.s32 %p31, %r92, 1118925336; add.s32 %r93, %r92, -1; mov.b32 %f611, %r93; add.f32 %f612, %f610, 0f37000000; selp.f32 %f76, %f612, %f610, %p31; selp.f32 %f613, %f611, %f607, %p31; mov.f32 %f614, 0f3FB8AA3B; mul.rn.f32 %f615, %f613, %f614; cvt.rzi.f32.f32 %f616, %f615; abs.f32 %f617, %f616; setp.gt.f32 %p32, %f617, 0f42FC0000; mov.b32 %r94, %f616; and.b32 %r95, %r94, -2147483648; or.b32 %r96, %r95, 1123811328; mov.b32 %f618, %r96; selp.f32 %f619, %f618, %f616, %p32; mov.f32 %f620, 0fBF317218; fma.rn.f32 %f621, %f619, %f620, %f613; mov.f32 %f622, 0f3102E308; fma.rn.f32 %f623, %f619, %f622, %f621; mul.f32 %f624, %f623, 0f3FB8AA3B; add.f32 %f625, %f619, 0f4B40007F; mov.b32 %r97, %f625; shl.b32 %r98, %r97, 23; mov.b32 %f626, %r98; ex2.approx.ftz.f32 %f627, %f624; mul.f32 %f77, %f627, %f626; setp.eq.f32 %p33, %f77, 0f7F800000; mov.f32 %f1585, 0f7F800000; @%p33 bra $L__BB0_19; fma.rn.f32 %f1585, %f77, %f76, %f77; $L__BB0_19: mov.f32 %f1557, 0f40000000; cvt.rzi.f32.f32 %f1556, %f1557; add.f32 %f1555, %f1556, %f1556; mov.f32 %f1554, 0f40800000; sub.f32 %f1553, %f1554, %f1555; abs.f32 %f1552, %f1553; setp.lt.f32 %p34, %f73, 0f00000000; setp.eq.f32 %p35, %f1552, 0f3F800000; and.pred %p1, %p34, %p35; setp.eq.f32 %p36, %f73, 0f00000000; @%p36 bra $L__BB0_23; bra.uni $L__BB0_20; $L__BB0_23: add.f32 %f632, %f73, %f73; selp.f32 %f1587, %f632, 0f00000000, %p35; bra.uni $L__BB0_24; $L__BB0_20: mov.b32 %r99, %f1585; xor.b32 %r100, %r99, -2147483648; mov.b32 %f628, %r100; selp.f32 %f1587, %f628, %f1585, %p1; setp.geu.f32 %p37, %f73, 0f00000000; @%p37 bra $L__BB0_24; mov.f32 %f629, 0f40800000; cvt.rzi.f32.f32 %f630, %f629; setp.eq.f32 %p38, %f630, 0f40800000; @%p38 bra $L__BB0_24; mov.f32 %f1587, 0f7FFFFFFF; $L__BB0_24: add.f32 %f633, %f75, 0f40800000; mov.b32 %r101, %f633; setp.lt.s32 %p40, %r101, 2139095040; @%p40 bra $L__BB0_29; setp.gtu.f32 %p41, %f75, 0f7F800000; @%p41 bra $L__BB0_28; bra.uni $L__BB0_26; $L__BB0_28: add.f32 %f1587, %f73, 0f40800000; bra.uni $L__BB0_29; $L__BB0_26: setp.neu.f32 %p42, %f75, 0f7F800000; @%p42 bra $L__BB0_29; selp.f32 %f1587, 0fFF800000, 0f7F800000, %p1; $L__BB0_29: mov.f32 %f634, 0f3F800000; sub.f32 %f635, %f634, %f1587; setp.eq.f32 %p43, %f73, 0f3F800000; selp.f32 %f636, 0f00000000, %f635, %p43; cvt.sat.f32.f32 %f637, %f636; mul.f32 %f638, %f72, %f637; ld.global.f32 %f639, [%rd7+12]; mul.f32 %f640, %f68, %f639; ld.global.f32 %f641, [%rd7+16]; mul.f32 %f642, %f69, %f641; neg.f32 %f643, %f642; sub.f32 %f644, %f643, %f640; ld.global.f32 %f645, [%rd7+20]; mul.f32 %f646, %f70, %f645; sub.f32 %f647, %f644, %f646; cvt.sat.f32.f32 %f648, %f647; mul.f32 %f86, %f638, %f648; setp.leu.f32 %p44, %f86, 0f3727C5AC; @%p44 bra $L__BB0_31; cvt.sat.f32.f32 %f658, %f71; mov.f32 %f664, 0f38D1B717; max.f32 %f655, %f663, %f664; sub.f32 %f656, %f67, %f655; mov.f32 %f657, 0f00000000; mov.u32 %r138, 2; mov.u32 %r140, 1; mov.u32 %r141, 1065353216; mov.u32 %r172, 0; // begin inline asm call(%r102,%r103,%r104,%r105,%r106,%r107,%r108,%r109,%r110,%r111,%r112,%r113,%r114,%r115,%r116,%r117,%r118,%r119,%r120,%r121,%r122,%r123,%r124,%r125,%r126,%r127,%r128,%r129,%r130,%r131,%r132,%r133),_optix_trace_typed_32,(%r172,%rd6,%f15,%f16,%f17,%f68,%f69,%f70,%f655,%f656,%f657,%r140,%r172,%r140,%r138,%r140,%r140,%r141,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172,%r172); // end inline asm mov.b32 %f665, %r102; ld.global.f32 %f666, [%rd7+24]; mul.f32 %f667, %f666, %f665; ld.global.f32 %f668, [%rd7+28]; mul.f32 %f669, %f668, %f665; ld.global.f32 %f670, [%rd7+32]; mul.f32 %f671, %f670, %f665; mul.f32 %f672, %f86, 0f3EA2F983; mul.f32 %f673, %f672, %f667; mul.f32 %f674, %f672, %f669; mul.f32 %f675, %f672, %f671; fma.rn.f32 %f1597, %f658, %f673, %f1597; fma.rn.f32 %f1598, %f658, %f674, %f1598; fma.rn.f32 %f1599, %f658, %f675, %f1599; mul.f32 %f676, %f28, %f69; fma.rn.f32 %f677, %f26, %f68, %f676; fma.rn.f32 %f678, %f30, %f70, %f677; cvt.sat.f32.f32 %f679, %f678; fma.rn.f32 %f1594, %f673, %f679, %f1594; fma.rn.f32 %f1595, %f674, %f679, %f1595; fma.rn.f32 %f1596, %f679, %f675, %f1596; mul.f32 %f680, %f38, %f69; fma.rn.f32 %f681, %f37, %f68, %f680; fma.rn.f32 %f682, %f39, %f70, %f681; cvt.sat.f32.f32 %f683, %f682; fma.rn.f32 %f1591, %f673, %f683, %f1591; fma.rn.f32 %f1592, %f674, %f683, %f1592; fma.rn.f32 %f1593, %f675, %f683, %f1593; mul.f32 %f684, %f47, %f69; fma.rn.f32 %f685, %f46, %f68, %f684; fma.rn.f32 %f686, %f48, %f70, %f685; cvt.sat.f32.f32 %f687, %f686; fma.rn.f32 %f1588, %f673, %f687, %f1588; fma.rn.f32 %f1589, %f674, %f687, %f1589; fma.rn.f32 %f1590, %f675, %f687, %f1590; add.f32 %f1600, %f1600, %f665; $L__BB0_31: add.s32 %r453, %r453, 1; $L__BB0_32: add.s32 %r450, %r450, 1; setp.lt.s32 %p46, %r450, %r4; mov.pred %p204, %p22; @%p46 bra $L__BB0_15; $L__BB0_33: cvt.rn.f32.s32 %f688, %r453; mov.f32 %f689, 0f3F800000; max.f32 %f690, %f688, %f689; rcp.rn.f32 %f691, %f690; mul.f32 %f1678, %f1597, %f691; mul.f32 %f1679, %f1598, %f691; mul.f32 %f1680, %f1599, %f691; div.rn.f32 %f1681, %f1600, %f690; mul.f32 %f1675, %f1594, %f691; mul.f32 %f1676, %f1595, %f691; mul.f32 %f1677, %f1596, %f691; mul.f32 %f1672, %f1591, %f691; mul.f32 %f1673, %f1592, %f691; mul.f32 %f1674, %f1593, %f691; mul.f32 %f1669, %f1588, %f691; mul.f32 %f1670, %f1589, %f691; mul.f32 %f1671, %f1590, %f691; not.pred %p47, %p204; @%p47 bra $L__BB0_64; abs.f32 %f1559, %f1562; abs.f32 %f1558, %f1560; setp.gt.f32 %p48, %f1558, %f1559; selp.f32 %f704, 0f00000000, %f1561, %p48; mov.f32 %f703, 0f00000000; neg.f32 %f705, %f1562; selp.f32 %f706, %f1560, %f705, %p48; neg.f32 %f707, %f1561; selp.f32 %f708, %f707, 0f00000000, %p48; mul.f32 %f709, %f706, %f706; fma.rn.f32 %f710, %f708, %f708, %f709; fma.rn.f32 %f711, %f704, %f704, %f710; sqrt.rn.f32 %f712, %f711; rcp.rn.f32 %f713, %f712; mul.f32 %f152, %f708, %f713; mul.f32 %f153, %f706, %f713; mul.f32 %f154, %f704, %f713; setp.lt.s32 %p49, %r5, 1; mov.f32 %f1650, %f703; mov.f32 %f1649, %f703; mov.f32 %f1648, %f703; mov.f32 %f1647, %f703; mov.f32 %f1646, %f703; mov.f32 %f1645, %f703; mov.f32 %f1644, %f703; mov.f32 %f1643, %f703; mov.f32 %f1642, %f703; mov.f32 %f1641, %f703; mov.f32 %f1640, %f703; mov.f32 %f1639, %f703; @%p49 bra $L__BB0_63; cvt.rn.f32.s32 %f726, %r5; rcp.rn.f32 %f155, %f726; mul.f32 %f156, %f15, 0f3456BF95; mul.f32 %f157, %f16, 0f3456BF95; mul.f32 %f158, %f17, 0f3456BF95; ld.const.u64 %rd8, [params+96]; mul.f32 %f727, %f1560, %f153; mul.f32 %f728, %f1561, %f152; sub.f32 %f159, %f728, %f727; mul.f32 %f729, %f1562, %f152; mul.f32 %f730, %f1560, %f154; sub.f32 %f160, %f730, %f729; mul.f32 %f731, %f1561, %f154; mul.f32 %f732, %f1562, %f153; sub.f32 %f161, %f732, %f731; mov.u32 %r173, 0; add.s64 %rd9, %rd1, 24; mov.u64 %rd53, __cudart_i2opi_f; abs.f32 %f795, %f157; abs.f32 %f796, %f156; max.f32 %f797, %f796, %f795; abs.f32 %f798, %f158; max.f32 %f799, %f797, %f798; mov.u32 %r454, %r173; $L__BB0_36: cvt.rn.f32.s32 %f174, %r454; mov.u32 %r456, %r173; $L__BB0_37: mad.lo.s32 %r175, %r457, 1664525, 1013904223; and.b32 %r176, %r175, 16777215; cvt.rn.f32.u32 %f733, %r176; fma.rn.f32 %f734, %f733, 0f33800000, %f174; mul.f32 %f187, %f155, %f734; mad.lo.s32 %r457, %r175, 1664525, 1013904223; and.b32 %r177, %r457, 16777215; cvt.rn.f32.u32 %f735, %r177; cvt.rn.f32.s32 %f736, %r456; fma.rn.f32 %f737, %f735, 0f33800000, %f736; mul.f32 %f738, %f155, %f737; mul.f32 %f739, %f187, %f187; mov.f32 %f740, 0f3F800000; sub.f32 %f741, %f740, %f739; mov.f32 %f742, 0f00000000; max.f32 %f743, %f742, %f741; sqrt.rn.f32 %f188, %f743; mul.f32 %f189, %f738, 0f40C90FDB; mul.f32 %f744, %f189, 0f3F22F983; cvt.rni.s32.f32 %r464, %f744; cvt.rn.f32.s32 %f745, %r464; mov.f32 %f746, 0fBFC90FDA; fma.rn.f32 %f747, %f745, %f746, %f189; mov.f32 %f748, 0fB3A22168; fma.rn.f32 %f749, %f745, %f748, %f747; mov.f32 %f750, 0fA7C234C5; fma.rn.f32 %f1654, %f745, %f750, %f749; abs.f32 %f191, %f189; setp.ltu.f32 %p50, %f191, 0f47CE4780; mov.u32 %r461, %r464; mov.f32 %f1651, %f1654; @%p50 bra $L__BB0_45; setp.eq.f32 %p51, %f191, 0f7F800000; @%p51 bra $L__BB0_44; bra.uni $L__BB0_39; $L__BB0_44: mov.f32 %f753, 0f00000000; mul.rn.f32 %f1651, %f189, %f753; mov.u32 %r461, 0; bra.uni $L__BB0_45; $L__BB0_39: mov.b32 %r21, %f189; bfe.u32 %r179, %r21, 23, 8; add.s32 %r22, %r179, -128; shl.b32 %r180, %r21, 8; or.b32 %r23, %r180, -2147483648; shr.u32 %r24, %r22, 5; mov.u64 %rd118, 0; mov.u32 %r458, 0; mov.u64 %rd116, %rd1; mov.u64 %rd117, %rd53; $L__BB0_40: .pragma "nounroll"; ld.global.nc.u32 %r181, [%rd117]; mad.wide.u32 %rd55, %r181, %r23, %rd118; shr.u64 %rd118, %rd55, 32; st.local.u32 [%rd116], %rd55; add.s64 %rd117, %rd117, 4; add.s64 %rd116, %rd116, 4; add.s32 %r458, %r458, 1; setp.ne.s32 %p52, %r458, 6; @%p52 bra $L__BB0_40; st.local.u32 [%rd9], %rd118; mov.u32 %r182, 4; sub.s32 %r27, %r182, %r24; mov.u32 %r183, 6; sub.s32 %r184, %r183, %r24; mul.wide.s32 %rd56, %r184, 4; add.s64 %rd57, %rd1, %rd56; ld.local.u32 %r459, [%rd57]; ld.local.u32 %r460, [%rd57+-4]; and.b32 %r30, %r22, 31; setp.eq.s32 %p53, %r30, 0; @%p53 bra $L__BB0_43; mov.u32 %r185, 32; sub.s32 %r186, %r185, %r30; shr.u32 %r187, %r460, %r186; shl.b32 %r188, %r459, %r30; add.s32 %r459, %r187, %r188; mul.wide.s32 %rd58, %r27, 4; add.s64 %rd59, %rd1, %rd58; ld.local.u32 %r189, [%rd59]; shr.u32 %r190, %r189, %r186; shl.b32 %r191, %r460, %r30; add.s32 %r460, %r190, %r191; $L__BB0_43: and.b32 %r192, %r21, -2147483648; shr.u32 %r193, %r460, 30; shl.b32 %r194, %r459, 2; or.b32 %r195, %r193, %r194; shr.u32 %r196, %r195, 31; shr.u32 %r197, %r459, 30; add.s32 %r198, %r196, %r197; neg.s32 %r199, %r198; setp.eq.s32 %p54, %r192, 0; selp.b32 %r461, %r198, %r199, %p54; setp.ne.s32 %p55, %r196, 0; xor.b32 %r200, %r192, -2147483648; selp.b32 %r201, %r200, %r192, %p55; selp.b32 %r202, -1, 0, %p55; xor.b32 %r203, %r195, %r202; shl.b32 %r204, %r460, 2; xor.b32 %r205, %r204, %r202; cvt.u64.u32 %rd60, %r203; cvt.u64.u32 %rd61, %r205; bfi.b64 %rd62, %rd60, %rd61, 32, 32; cvt.rn.f64.s64 %fd1, %rd62; mul.f64 %fd2, %fd1, 0d3BF921FB54442D19; cvt.rn.f32.f64 %f751, %fd2; setp.eq.s32 %p56, %r201, 0; neg.f32 %f752, %f751; selp.f32 %f1651, %f751, %f752, %p56; $L__BB0_45: add.s32 %r37, %r461, 1; and.b32 %r38, %r37, 1; setp.eq.s32 %p57, %r38, 0; selp.f32 %f195, %f1651, 0f3F800000, %p57; mul.rn.f32 %f196, %f1651, %f1651; mov.f32 %f1652, 0fB94D4153; @%p57 bra $L__BB0_47; mov.f32 %f755, 0fBAB607ED; mov.f32 %f756, 0f37CBAC00; fma.rn.f32 %f1652, %f756, %f196, %f755; $L__BB0_47: selp.f32 %f757, 0f3C0885E4, 0f3D2AAABB, %p57; fma.rn.f32 %f758, %f1652, %f196, %f757; selp.f32 %f759, 0fBE2AAAA8, 0fBEFFFFFF, %p57; fma.rn.f32 %f760, %f758, %f196, %f759; mov.f32 %f761, 0f00000000; fma.rn.f32 %f762, %f196, %f195, %f761; fma.rn.f32 %f1653, %f760, %f762, %f195; and.b32 %r207, %r37, 2; setp.eq.s32 %p59, %r207, 0; @%p59 bra $L__BB0_49; mov.f32 %f764, 0fBF800000; fma.rn.f32 %f1653, %f1653, %f764, %f761; $L__BB0_49: @%p50 bra $L__BB0_57; setp.eq.f32 %p61, %f191, 0f7F800000; @%p61 bra $L__BB0_56; bra.uni $L__BB0_51; $L__BB0_56: mov.f32 %f767, 0f00000000; mul.rn.f32 %f1654, %f189, %f767; mov.u32 %r464, 0; bra.uni $L__BB0_57; $L__BB0_51: mov.b32 %r39, %f189; bfe.u32 %r208, %r39, 23, 8; add.s32 %r40, %r208, -128; shl.b32 %r209, %r39, 8; or.b32 %r41, %r209, -2147483648; shr.u32 %r42, %r40, 5; mov.u64 %rd119, 0; mov.u64 %rd120, %rd119; $L__BB0_52: .pragma "nounroll"; shl.b64 %rd65, %rd119, 2; mov.u64 %rd66, __cudart_i2opi_f; add.s64 %rd67, %rd66, %rd65; ld.global.nc.u32 %r210, [%rd67]; mad.wide.u32 %rd68, %r210, %r41, %rd120; shr.u64 %rd120, %rd68, 32; add.s64 %rd69, %rd1, %rd65; st.local.u32 [%rd69], %rd68; cvt.u32.u64 %r211, %rd119; add.s32 %r212, %r211, 1; cvt.s64.s32 %rd119, %r212; setp.ne.s32 %p62, %r212, 6; @%p62 bra $L__BB0_52; st.local.u32 [%rd9], %rd120; mov.u32 %r213, 4; sub.s32 %r43, %r213, %r42; mov.u32 %r214, 6; sub.s32 %r215, %r214, %r42; mul.wide.s32 %rd70, %r215, 4; add.s64 %rd71, %rd1, %rd70; ld.local.u32 %r462, [%rd71]; ld.local.u32 %r463, [%rd71+-4]; and.b32 %r46, %r40, 31; setp.eq.s32 %p63, %r46, 0; @%p63 bra $L__BB0_55; mov.u32 %r216, 32; sub.s32 %r217, %r216, %r46; shr.u32 %r218, %r463, %r217; shl.b32 %r219, %r462, %r46; add.s32 %r462, %r218, %r219; mul.wide.s32 %rd72, %r43, 4; add.s64 %rd73, %rd1, %rd72; ld.local.u32 %r220, [%rd73]; shr.u32 %r221, %r220, %r217; shl.b32 %r222, %r463, %r46; add.s32 %r463, %r221, %r222; $L__BB0_55: and.b32 %r223, %r39, -2147483648; shr.u32 %r224, %r463, 30; shl.b32 %r225, %r462, 2; or.b32 %r226, %r224, %r225; shr.u32 %r227, %r226, 31; shr.u32 %r228, %r462, 30; add.s32 %r229, %r227, %r228; neg.s32 %r230, %r229; setp.eq.s32 %p64, %r223, 0; selp.b32 %r464, %r229, %r230, %p64; setp.ne.s32 %p65, %r227, 0; xor.b32 %r231, %r223, -2147483648; selp.b32 %r232, %r231, %r223, %p65; selp.b32 %r233, -1, 0, %p65; xor.b32 %r234, %r226, %r233; shl.b32 %r235, %r463, 2; xor.b32 %r236, %r235, %r233; cvt.u64.u32 %rd74, %r234; cvt.u64.u32 %rd75, %r236; bfi.b64 %rd76, %rd74, %rd75, 32, 32; cvt.rn.f64.s64 %fd3, %rd76; mul.f64 %fd4, %fd3, 0d3BF921FB54442D19; cvt.rn.f32.f64 %f765, %fd4; setp.eq.s32 %p66, %r232, 0; neg.f32 %f766, %f765; selp.f32 %f1654, %f765, %f766, %p66; $L__BB0_57: mul.f32 %f205, %f188, %f1653; and.b32 %r53, %r464, 1; setp.eq.s32 %p67, %r53, 0; selp.f32 %f206, %f1654, 0f3F800000, %p67; mul.rn.f32 %f207, %f1654, %f1654; mov.f32 %f1655, 0fB94D4153; @%p67 bra $L__BB0_59; mov.f32 %f769, 0fBAB607ED; mov.f32 %f770, 0f37CBAC00; fma.rn.f32 %f1655, %f770, %f207, %f769; $L__BB0_59: selp.f32 %f771, 0f3C0885E4, 0f3D2AAABB, %p67; fma.rn.f32 %f772, %f1655, %f207, %f771; selp.f32 %f773, 0fBE2AAAA8, 0fBEFFFFFF, %p67; fma.rn.f32 %f774, %f772, %f207, %f773; mov.f32 %f775, 0f00000000; fma.rn.f32 %f776, %f207, %f206, %f775; fma.rn.f32 %f1656, %f774, %f776, %f206; and.b32 %r238, %r464, 2; setp.eq.s32 %p69, %r238, 0; @%p69 bra $L__BB0_61; mov.f32 %f778, 0fBF800000; fma.rn.f32 %f1656, %f1656, %f778, %f775; $L__BB0_61: mul.f32 %f788, %f188, %f1656; mul.f32 %f789, %f152, %f788; mul.f32 %f790, %f153, %f788; mul.f32 %f791, %f154, %f788; fma.rn.f32 %f792, %f161, %f205, %f789; fma.rn.f32 %f793, %f160, %f205, %f790; fma.rn.f32 %f794, %f159, %f205, %f791; fma.rn.f32 %f782, %f1560, %f187, %f792; fma.rn.f32 %f783, %f1561, %f187, %f793; fma.rn.f32 %f784, %f1562, %f187, %f794; mov.f32 %f800, 0f38D1B717; max.f32 %f785, %f799, %f800; mov.f32 %f786, 0f6C4ECB8F; mov.u32 %r272, 1; mov.u32 %r275, 2; mov.u32 %r277, 3; mov.u32 %r309, 0; // begin inline asm call(%r239,%r240,%r241,%r242,%r243,%r244,%r245,%r246,%r247,%r248,%r249,%r250,%r251,%r252,%r253,%r254,%r255,%r256,%r257,%r258,%r259,%r260,%r261,%r262,%r263,%r264,%r265,%r266,%r267,%r268,%r269,%r270),_optix_trace_typed_32,(%r309,%rd8,%f15,%f16,%f17,%f782,%f783,%f784,%f785,%f786,%f775,%r272,%r309,%r309,%r275,%r309,%r277,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309,%r309); // end inline asm mov.b32 %f801, %r239; mov.b32 %f802, %r240; mov.b32 %f803, %r241; max.f32 %f804, %f801, %f775; max.f32 %f805, %f802, %f775; max.f32 %f806, %f803, %f775; mul.f32 %f807, %f28, %f783; fma.rn.f32 %f808, %f26, %f782, %f807; fma.rn.f32 %f809, %f30, %f784, %f808; cvt.sat.f32.f32 %f810, %f809; fma.rn.f32 %f1644, %f804, %f810, %f1644; fma.rn.f32 %f1643, %f805, %f810, %f1643; fma.rn.f32 %f1642, %f806, %f810, %f1642; mul.f32 %f811, %f38, %f783; fma.rn.f32 %f812, %f37, %f782, %f811; fma.rn.f32 %f813, %f39, %f784, %f812; cvt.sat.f32.f32 %f814, %f813; fma.rn.f32 %f1647, %f804, %f814, %f1647; fma.rn.f32 %f1646, %f805, %f814, %f1646; fma.rn.f32 %f1645, %f806, %f814, %f1645; mul.f32 %f815, %f47, %f783; fma.rn.f32 %f816, %f46, %f782, %f815; fma.rn.f32 %f817, %f48, %f784, %f816; cvt.sat.f32.f32 %f818, %f817; fma.rn.f32 %f1650, %f804, %f818, %f1650; fma.rn.f32 %f1649, %f805, %f818, %f1649; fma.rn.f32 %f1648, %f806, %f818, %f1648; mul.f32 %f819, %f1561, %f783; fma.rn.f32 %f820, %f1560, %f782, %f819; fma.rn.f32 %f821, %f1562, %f784, %f820; cvt.sat.f32.f32 %f822, %f821; fma.rn.f32 %f1641, %f804, %f822, %f1641; fma.rn.f32 %f1640, %f805, %f822, %f1640; fma.rn.f32 %f1639, %f806, %f822, %f1639; add.s32 %r456, %r456, 1; setp.lt.s32 %p70, %r456, %r5; @%p70 bra $L__BB0_37; add.s32 %r454, %r454, 1; setp.lt.s32 %p71, %r454, %r5; @%p71 bra $L__BB0_36; $L__BB0_63: mul.lo.s32 %r310, %r5, %r5; cvt.rn.f32.s32 %f823, %r310; rcp.rn.f32 %f824, %f823; mul.f32 %f825, %f824, %f1641; mul.f32 %f826, %f824, %f1640; mul.f32 %f827, %f824, %f1639; div.rn.f32 %f1681, %f703, %f823; mul.f32 %f1675, %f824, %f1644; mul.f32 %f1676, %f824, %f1643; mul.f32 %f1677, %f824, %f1642; mul.f32 %f1672, %f824, %f1647; mul.f32 %f1673, %f824, %f1646; mul.f32 %f1674, %f824, %f1645; mul.f32 %f1669, %f824, %f1650; mul.f32 %f1670, %f824, %f1649; mul.f32 %f1671, %f824, %f1648; fma.rn.f32 %f1678, %f824, %f1641, %f825; fma.rn.f32 %f1679, %f824, %f1640, %f826; fma.rn.f32 %f1680, %f824, %f1639, %f827; $L__BB0_64: ld.const.u32 %r449, [params+616]; setp.lt.s32 %p72, %r449, 0; selp.f32 %f263, %f1678, %f1681, %p72; ld.const.u32 %r56, [params+104]; and.b32 %r311, %r56, 8; setp.eq.s32 %p73, %r311, 0; @%p73 bra $L__BB0_78; ld.const.u64 %rd78, [params+192]; cvta.to.global.u64 %rd20, %rd78; ld.const.u32 %r312, [params+184]; mad.lo.s32 %r313, %r312, %r7, %r6; cvt.u64.u32 %rd21, %r313; mov.f32 %f830, 0f3E68BA2E; cvt.rzi.f32.f32 %f831, %f830; add.f32 %f832, %f831, %f831; mov.f32 %f833, 0f3EE8BA2E; sub.f32 %f834, %f833, %f832; abs.f32 %f264, %f834; abs.f32 %f265, %f263; setp.lt.f32 %p74, %f265, 0f00800000; mul.f32 %f835, %f265, 0f4B800000; selp.f32 %f836, %f835, %f265, %p74; selp.f32 %f837, 0fC3170000, 0fC2FE0000, %p74; mov.b32 %r314, %f836; and.b32 %r315, %r314, 8388607; or.b32 %r316, %r315, 1065353216; mov.b32 %f838, %r316; shr.u32 %r317, %r314, 23; cvt.rn.f32.u32 %f839, %r317; add.f32 %f840, %f837, %f839; setp.gt.f32 %p75, %f838, 0f3FB504F3; mul.f32 %f841, %f838, 0f3F000000; add.f32 %f842, %f840, 0f3F800000; selp.f32 %f843, %f842, %f840, %p75; selp.f32 %f844, %f841, %f838, %p75; add.f32 %f845, %f844, 0fBF800000; add.f32 %f846, %f844, 0f3F800000; rcp.approx.ftz.f32 %f847, %f846; add.f32 %f848, %f845, %f845; mul.f32 %f849, %f848, %f847; mul.f32 %f850, %f849, %f849; mov.f32 %f851, 0f3C4CAF63; mov.f32 %f852, 0f3B18F0FE; fma.rn.f32 %f853, %f852, %f850, %f851; mov.f32 %f854, 0f3DAAAABD; fma.rn.f32 %f855, %f853, %f850, %f854; mul.rn.f32 %f856, %f855, %f850; mul.rn.f32 %f857, %f856, %f849; sub.f32 %f858, %f845, %f849; add.f32 %f859, %f858, %f858; neg.f32 %f860, %f849; fma.rn.f32 %f861, %f860, %f845, %f859; mul.rn.f32 %f862, %f847, %f861; add.f32 %f863, %f857, %f849; sub.f32 %f864, %f849, %f863; add.f32 %f865, %f857, %f864; add.f32 %f866, %f862, %f865; add.f32 %f867, %f863, %f866; sub.f32 %f868, %f863, %f867; add.f32 %f869, %f866, %f868; mov.f32 %f870, 0f3F317200; mul.rn.f32 %f871, %f843, %f870; mov.f32 %f872, 0f35BFBE8E; mul.rn.f32 %f873, %f843, %f872; add.f32 %f874, %f871, %f867; sub.f32 %f875, %f871, %f874; add.f32 %f876, %f867, %f875; add.f32 %f877, %f869, %f876; add.f32 %f878, %f873, %f877; add.f32 %f879, %f874, %f878; sub.f32 %f880, %f874, %f879; add.f32 %f881, %f878, %f880; mul.rn.f32 %f882, %f833, %f879; neg.f32 %f883, %f882; fma.rn.f32 %f884, %f833, %f879, %f883; fma.rn.f32 %f885, %f833, %f881, %f884; mov.f32 %f886, 0f00000000; fma.rn.f32 %f887, %f886, %f879, %f885; add.rn.f32 %f888, %f882, %f887; neg.f32 %f889, %f888; add.rn.f32 %f890, %f882, %f889; add.rn.f32 %f891, %f890, %f887; mov.b32 %r318, %f888; setp.eq.s32 %p76, %r318, 1118925336; add.s32 %r319, %r318, -1; mov.b32 %f892, %r319; add.f32 %f893, %f891, 0f37000000; selp.f32 %f266, %f893, %f891, %p76; selp.f32 %f894, %f892, %f888, %p76; mov.f32 %f895, 0f3FB8AA3B; mul.rn.f32 %f896, %f894, %f895; cvt.rzi.f32.f32 %f897, %f896; abs.f32 %f898, %f897; setp.gt.f32 %p77, %f898, 0f42FC0000; mov.b32 %r320, %f897; and.b32 %r321, %r320, -2147483648; or.b32 %r322, %r321, 1123811328; mov.b32 %f899, %r322; selp.f32 %f900, %f899, %f897, %p77; mov.f32 %f901, 0fBF317218; fma.rn.f32 %f902, %f900, %f901, %f894; mov.f32 %f903, 0f3102E308; fma.rn.f32 %f904, %f900, %f903, %f902; mul.f32 %f905, %f904, 0f3FB8AA3B; add.f32 %f906, %f900, 0f4B40007F; mov.b32 %r323, %f906; shl.b32 %r324, %r323, 23; mov.b32 %f907, %r324; ex2.approx.ftz.f32 %f908, %f905; mul.f32 %f267, %f908, %f907; setp.eq.f32 %p78, %f267, 0f7F800000; mov.f32 %f1682, 0f7F800000; @%p78 bra $L__BB0_67; fma.rn.f32 %f1682, %f267, %f266, %f267; $L__BB0_67: setp.lt.f32 %p79, %f263, 0f00000000; setp.eq.f32 %p80, %f264, 0f3F800000; and.pred %p3, %p79, %p80; setp.eq.f32 %p81, %f263, 0f00000000; @%p81 bra $L__BB0_71; bra.uni $L__BB0_68; $L__BB0_71: add.f32 %f913, %f263, %f263; selp.f32 %f1684, %f913, 0f00000000, %p80; bra.uni $L__BB0_72; $L__BB0_172: mov.f32 %f1524, 0f00000000; mov.u32 %r466, 0; // begin inline asm { cvt.rn.f16.f32 %rs127, %f1524;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs126, %f1524;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs125, %f1524;} // end inline asm mov.u16 %rs128, 0; st.global.v4.u16 [%rd28], {%rs125, %rs126, %rs127, %rs128}; $L__BB0_173: ld.const.u64 %rd107, [params+256]; cvta.to.global.u64 %rd108, %rd107; ld.const.u32 %r443, [params+248]; mad.lo.s32 %r444, %r443, %r7, %r6; mul.wide.u32 %rd109, %r444, 8; add.s64 %rd29, %rd108, %rd109; setp.eq.s32 %p201, %r466, 0; @%p201 bra $L__BB0_175; ld.global.v4.u16 {%rs135, %rs136, %rs137, %rs138}, [%rd29]; // begin inline asm { cvt.f32.f16 %f1525, %rs135;} // end inline asm // begin inline asm { cvt.f32.f16 %f1526, %rs136;} // end inline asm // begin inline asm { cvt.f32.f16 %f1527, %rs137;} // end inline asm add.f32 %f1528, %f1525, 0f00000000; add.f32 %f1529, %f1526, 0f00000000; add.f32 %f1530, %f1527, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs134, %f1530;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs133, %f1529;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs132, %f1528;} // end inline asm mov.u16 %rs139, 0; st.global.v4.u16 [%rd29], {%rs132, %rs133, %rs134, %rs139}; bra.uni $L__BB0_176; $L__BB0_175: mov.f32 %f1533, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs142, %f1533;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs141, %f1533;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs140, %f1533;} // end inline asm mov.u16 %rs143, 0; st.global.v4.u16 [%rd29], {%rs140, %rs141, %rs142, %rs143}; $L__BB0_176: ld.const.u64 %rd110, [params+272]; cvta.to.global.u64 %rd111, %rd110; ld.const.u32 %r445, [params+264]; mad.lo.s32 %r446, %r445, %r7, %r6; mul.wide.u32 %rd112, %r446, 8; add.s64 %rd30, %rd111, %rd112; @%p201 bra $L__BB0_178; ld.global.v4.u16 {%rs150, %rs151, %rs152, %rs153}, [%rd30]; // begin inline asm { cvt.f32.f16 %f1534, %rs150;} // end inline asm // begin inline asm { cvt.f32.f16 %f1535, %rs151;} // end inline asm // begin inline asm { cvt.f32.f16 %f1536, %rs152;} // end inline asm add.f32 %f1537, %f1534, 0f00000000; add.f32 %f1538, %f1535, 0f00000000; add.f32 %f1539, %f1536, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs149, %f1539;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs148, %f1538;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs147, %f1537;} // end inline asm mov.u16 %rs154, 0; st.global.v4.u16 [%rd30], {%rs147, %rs148, %rs149, %rs154}; bra.uni $L__BB0_179; $L__BB0_178: mov.f32 %f1542, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs157, %f1542;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs156, %f1542;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs155, %f1542;} // end inline asm mov.u16 %rs158, 0; st.global.v4.u16 [%rd30], {%rs155, %rs156, %rs157, %rs158}; $L__BB0_179: ld.const.u64 %rd113, [params+288]; cvta.to.global.u64 %rd114, %rd113; ld.const.u32 %r447, [params+280]; mad.lo.s32 %r448, %r447, %r7, %r6; mul.wide.u32 %rd115, %r448, 8; add.s64 %rd31, %rd114, %rd115; @%p201 bra $L__BB0_181; ld.global.v4.u16 {%rs165, %rs166, %rs167, %rs168}, [%rd31]; // begin inline asm { cvt.f32.f16 %f1543, %rs165;} // end inline asm // begin inline asm { cvt.f32.f16 %f1544, %rs166;} // end inline asm // begin inline asm { cvt.f32.f16 %f1545, %rs167;} // end inline asm add.f32 %f1546, %f1543, 0f00000000; add.f32 %f1547, %f1544, 0f00000000; add.f32 %f1548, %f1545, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs164, %f1548;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs163, %f1547;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs162, %f1546;} // end inline asm mov.u16 %rs169, 0; st.global.v4.u16 [%rd31], {%rs162, %rs163, %rs164, %rs169}; bra.uni $L__BB0_182; $L__BB0_181: mov.f32 %f1551, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs172, %f1551;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs171, %f1551;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs170, %f1551;} // end inline asm mov.u16 %rs173, 0; st.global.v4.u16 [%rd31], {%rs170, %rs171, %rs172, %rs173}; bra.uni $L__BB0_182; $L__BB0_68: mov.b32 %r325, %f1682; xor.b32 %r326, %r325, -2147483648; mov.b32 %f909, %r326; selp.f32 %f1684, %f909, %f1682, %p3; setp.geu.f32 %p82, %f263, 0f00000000; @%p82 bra $L__BB0_72; mov.f32 %f910, 0f3EE8BA2E; cvt.rzi.f32.f32 %f911, %f910; setp.eq.f32 %p83, %f911, 0f3EE8BA2E; @%p83 bra $L__BB0_72; mov.f32 %f1684, 0f7FFFFFFF; $L__BB0_72: add.f32 %f914, %f265, 0f3EE8BA2E; mov.b32 %r327, %f914; setp.lt.s32 %p85, %r327, 2139095040; @%p85 bra $L__BB0_77; setp.gtu.f32 %p86, %f265, 0f7F800000; @%p86 bra $L__BB0_76; bra.uni $L__BB0_74; $L__BB0_76: add.f32 %f1684, %f263, 0f3EE8BA2E; bra.uni $L__BB0_77; $L__BB0_74: setp.neu.f32 %p87, %f265, 0f7F800000; @%p87 bra $L__BB0_77; selp.f32 %f1684, 0fFF800000, 0f7F800000, %p3; $L__BB0_77: mul.f32 %f915, %f1684, 0f437F0000; setp.eq.f32 %p88, %f263, 0f3F800000; selp.f32 %f916, 0f437F0000, %f915, %p88; cvt.rzi.u32.f32 %r328, %f916; shl.b64 %rd79, %rd21, 1; add.s64 %rd80, %rd20, %rd79; cvt.u16.u32 %rs45, %r328; mov.u16 %rs46, 255; st.global.v2.u8 [%rd80], {%rs45, %rs46}; $L__BB0_78: ld.const.v2.f32 {%f917, %f918}, [params+648]; mul.f32 %f278, %f1678, %f917; mul.f32 %f279, %f1679, %f918; ld.const.f32 %f280, [params+656]; mul.f32 %f281, %f1680, %f280; and.b32 %r329, %r56, 1; setp.eq.b32 %p89, %r329, 1; mov.pred %p90, 0; xor.pred %p91, %p89, %p90; not.pred %p92, %p91; @%p92 bra $L__BB0_152; mov.f32 %f920, 0f3E666666; cvt.rzi.f32.f32 %f921, %f920; add.f32 %f922, %f921, %f921; mov.f32 %f923, 0f3EE66666; sub.f32 %f924, %f923, %f922; abs.f32 %f282, %f924; abs.f32 %f283, %f278; setp.lt.f32 %p93, %f283, 0f00800000; mul.f32 %f925, %f283, 0f4B800000; selp.f32 %f926, %f925, %f283, %p93; selp.f32 %f927, 0fC3170000, 0fC2FE0000, %p93; mov.b32 %r330, %f926; and.b32 %r331, %r330, 8388607; or.b32 %r332, %r331, 1065353216; mov.b32 %f928, %r332; shr.u32 %r333, %r330, 23; cvt.rn.f32.u32 %f929, %r333; add.f32 %f930, %f927, %f929; setp.gt.f32 %p94, %f928, 0f3FB504F3; mul.f32 %f931, %f928, 0f3F000000; add.f32 %f932, %f930, 0f3F800000; selp.f32 %f933, %f932, %f930, %p94; selp.f32 %f934, %f931, %f928, %p94; add.f32 %f935, %f934, 0fBF800000; add.f32 %f936, %f934, 0f3F800000; rcp.approx.ftz.f32 %f937, %f936; add.f32 %f938, %f935, %f935; mul.f32 %f939, %f938, %f937; mul.f32 %f940, %f939, %f939; mov.f32 %f941, 0f3C4CAF63; mov.f32 %f942, 0f3B18F0FE; fma.rn.f32 %f943, %f942, %f940, %f941; mov.f32 %f944, 0f3DAAAABD; fma.rn.f32 %f945, %f943, %f940, %f944; mul.rn.f32 %f946, %f945, %f940; mul.rn.f32 %f947, %f946, %f939; sub.f32 %f948, %f935, %f939; add.f32 %f949, %f948, %f948; neg.f32 %f950, %f939; fma.rn.f32 %f951, %f950, %f935, %f949; mul.rn.f32 %f952, %f937, %f951; add.f32 %f953, %f947, %f939; sub.f32 %f954, %f939, %f953; add.f32 %f955, %f947, %f954; add.f32 %f956, %f952, %f955; add.f32 %f957, %f953, %f956; sub.f32 %f958, %f953, %f957; add.f32 %f959, %f956, %f958; mov.f32 %f960, 0f3F317200; mul.rn.f32 %f961, %f933, %f960; mov.f32 %f962, 0f35BFBE8E; mul.rn.f32 %f963, %f933, %f962; add.f32 %f964, %f961, %f957; sub.f32 %f965, %f961, %f964; add.f32 %f966, %f957, %f965; add.f32 %f967, %f959, %f966; add.f32 %f968, %f963, %f967; add.f32 %f969, %f964, %f968; sub.f32 %f970, %f964, %f969; add.f32 %f971, %f968, %f970; mul.rn.f32 %f972, %f923, %f969; neg.f32 %f973, %f972; fma.rn.f32 %f974, %f923, %f969, %f973; fma.rn.f32 %f975, %f923, %f971, %f974; mov.f32 %f976, 0f00000000; fma.rn.f32 %f977, %f976, %f969, %f975; add.rn.f32 %f978, %f972, %f977; neg.f32 %f979, %f978; add.rn.f32 %f980, %f972, %f979; add.rn.f32 %f981, %f980, %f977; mov.b32 %r334, %f978; setp.eq.s32 %p95, %r334, 1118925336; add.s32 %r335, %r334, -1; mov.b32 %f982, %r335; add.f32 %f983, %f981, 0f37000000; selp.f32 %f284, %f983, %f981, %p95; selp.f32 %f984, %f982, %f978, %p95; mov.f32 %f985, 0f3FB8AA3B; mul.rn.f32 %f986, %f984, %f985; cvt.rzi.f32.f32 %f987, %f986; abs.f32 %f988, %f987; setp.gt.f32 %p96, %f988, 0f42FC0000; mov.b32 %r336, %f987; and.b32 %r337, %r336, -2147483648; or.b32 %r338, %r337, 1123811328; mov.b32 %f989, %r338; selp.f32 %f990, %f989, %f987, %p96; mov.f32 %f991, 0fBF317218; fma.rn.f32 %f992, %f990, %f991, %f984; mov.f32 %f993, 0f3102E308; fma.rn.f32 %f994, %f990, %f993, %f992; mul.f32 %f995, %f994, 0f3FB8AA3B; add.f32 %f996, %f990, 0f4B40007F; mov.b32 %r339, %f996; shl.b32 %r340, %r339, 23; mov.b32 %f997, %r340; ex2.approx.ftz.f32 %f998, %f995; mul.f32 %f285, %f998, %f997; setp.eq.f32 %p97, %f285, 0f7F800000; mov.f32 %f1685, 0f7F800000; @%p97 bra $L__BB0_81; fma.rn.f32 %f1685, %f285, %f284, %f285; $L__BB0_81: setp.lt.f32 %p98, %f278, 0f00000000; setp.eq.f32 %p99, %f282, 0f3F800000; and.pred %p4, %p98, %p99; setp.eq.f32 %p100, %f278, 0f00000000; @%p100 bra $L__BB0_85; bra.uni $L__BB0_82; $L__BB0_85: add.f32 %f1003, %f278, %f278; selp.f32 %f1687, %f1003, 0f00000000, %p99; bra.uni $L__BB0_86; $L__BB0_82: mov.b32 %r341, %f1685; xor.b32 %r342, %r341, -2147483648; mov.b32 %f999, %r342; selp.f32 %f1687, %f999, %f1685, %p4; setp.geu.f32 %p101, %f278, 0f00000000; @%p101 bra $L__BB0_86; mov.f32 %f1000, 0f3EE66666; cvt.rzi.f32.f32 %f1001, %f1000; setp.eq.f32 %p102, %f1001, 0f3EE66666; @%p102 bra $L__BB0_86; mov.f32 %f1687, 0f7FFFFFFF; $L__BB0_86: add.f32 %f1004, %f283, 0f3EE66666; mov.b32 %r343, %f1004; setp.lt.s32 %p104, %r343, 2139095040; @%p104 bra $L__BB0_91; setp.gtu.f32 %p105, %f283, 0f7F800000; @%p105 bra $L__BB0_90; bra.uni $L__BB0_88; $L__BB0_90: add.f32 %f1687, %f278, 0f3EE66666; bra.uni $L__BB0_91; $L__BB0_88: setp.neu.f32 %p106, %f283, 0f7F800000; @%p106 bra $L__BB0_91; selp.f32 %f1687, 0fFF800000, 0f7F800000, %p4; $L__BB0_91: setp.eq.f32 %p107, %f278, 0f3F800000; selp.f32 %f294, 0f3F800000, %f1687, %p107; abs.f32 %f295, %f279; setp.lt.f32 %p108, %f295, 0f00800000; mul.f32 %f1006, %f295, 0f4B800000; selp.f32 %f1007, %f1006, %f295, %p108; selp.f32 %f1008, 0fC3170000, 0fC2FE0000, %p108; mov.b32 %r344, %f1007; and.b32 %r345, %r344, 8388607; or.b32 %r346, %r345, 1065353216; mov.b32 %f1009, %r346; shr.u32 %r347, %r344, 23; cvt.rn.f32.u32 %f1010, %r347; add.f32 %f1011, %f1008, %f1010; setp.gt.f32 %p109, %f1009, 0f3FB504F3; mul.f32 %f1012, %f1009, 0f3F000000; add.f32 %f1013, %f1011, 0f3F800000; selp.f32 %f1014, %f1013, %f1011, %p109; selp.f32 %f1015, %f1012, %f1009, %p109; add.f32 %f1016, %f1015, 0fBF800000; add.f32 %f1017, %f1015, 0f3F800000; rcp.approx.ftz.f32 %f1018, %f1017; add.f32 %f1019, %f1016, %f1016; mul.f32 %f1020, %f1019, %f1018; mul.f32 %f1021, %f1020, %f1020; mov.f32 %f1022, 0f3C4CAF63; mov.f32 %f1023, 0f3B18F0FE; fma.rn.f32 %f1024, %f1023, %f1021, %f1022; mov.f32 %f1025, 0f3DAAAABD; fma.rn.f32 %f1026, %f1024, %f1021, %f1025; mul.rn.f32 %f1027, %f1026, %f1021; mul.rn.f32 %f1028, %f1027, %f1020; sub.f32 %f1029, %f1016, %f1020; add.f32 %f1030, %f1029, %f1029; neg.f32 %f1031, %f1020; fma.rn.f32 %f1032, %f1031, %f1016, %f1030; mul.rn.f32 %f1033, %f1018, %f1032; add.f32 %f1034, %f1028, %f1020; sub.f32 %f1035, %f1020, %f1034; add.f32 %f1036, %f1028, %f1035; add.f32 %f1037, %f1033, %f1036; add.f32 %f1038, %f1034, %f1037; sub.f32 %f1039, %f1034, %f1038; add.f32 %f1040, %f1037, %f1039; mov.f32 %f1041, 0f3F317200; mul.rn.f32 %f1042, %f1014, %f1041; mov.f32 %f1043, 0f35BFBE8E; mul.rn.f32 %f1044, %f1014, %f1043; add.f32 %f1045, %f1042, %f1038; sub.f32 %f1046, %f1042, %f1045; add.f32 %f1047, %f1038, %f1046; add.f32 %f1048, %f1040, %f1047; add.f32 %f1049, %f1044, %f1048; add.f32 %f1050, %f1045, %f1049; sub.f32 %f1051, %f1045, %f1050; add.f32 %f1052, %f1049, %f1051; mov.f32 %f1053, 0f3EE66666; mul.rn.f32 %f1054, %f1053, %f1050; neg.f32 %f1055, %f1054; fma.rn.f32 %f1056, %f1053, %f1050, %f1055; fma.rn.f32 %f1057, %f1053, %f1052, %f1056; mov.f32 %f1058, 0f00000000; fma.rn.f32 %f1059, %f1058, %f1050, %f1057; add.rn.f32 %f1060, %f1054, %f1059; neg.f32 %f1061, %f1060; add.rn.f32 %f1062, %f1054, %f1061; add.rn.f32 %f1063, %f1062, %f1059; mov.b32 %r348, %f1060; setp.eq.s32 %p110, %r348, 1118925336; add.s32 %r349, %r348, -1; mov.b32 %f1064, %r349; add.f32 %f1065, %f1063, 0f37000000; selp.f32 %f296, %f1065, %f1063, %p110; selp.f32 %f1066, %f1064, %f1060, %p110; mov.f32 %f1067, 0f3FB8AA3B; mul.rn.f32 %f1068, %f1066, %f1067; cvt.rzi.f32.f32 %f1069, %f1068; abs.f32 %f1070, %f1069; setp.gt.f32 %p111, %f1070, 0f42FC0000; mov.b32 %r350, %f1069; and.b32 %r351, %r350, -2147483648; or.b32 %r352, %r351, 1123811328; mov.b32 %f1071, %r352; selp.f32 %f1072, %f1071, %f1069, %p111; mov.f32 %f1073, 0fBF317218; fma.rn.f32 %f1074, %f1072, %f1073, %f1066; mov.f32 %f1075, 0f3102E308; fma.rn.f32 %f1076, %f1072, %f1075, %f1074; mul.f32 %f1077, %f1076, 0f3FB8AA3B; add.f32 %f1078, %f1072, 0f4B40007F; mov.b32 %r353, %f1078; shl.b32 %r354, %r353, 23; mov.b32 %f1079, %r354; ex2.approx.ftz.f32 %f1080, %f1077; mul.f32 %f297, %f1080, %f1079; setp.eq.f32 %p112, %f297, 0f7F800000; mov.f32 %f1688, 0f7F800000; @%p112 bra $L__BB0_93; fma.rn.f32 %f1688, %f297, %f296, %f297; $L__BB0_93: setp.lt.f32 %p113, %f279, 0f00000000; and.pred %p5, %p113, %p99; setp.eq.f32 %p115, %f279, 0f00000000; @%p115 bra $L__BB0_97; bra.uni $L__BB0_94; $L__BB0_97: add.f32 %f1085, %f279, %f279; selp.f32 %f1690, %f1085, 0f00000000, %p99; bra.uni $L__BB0_98; $L__BB0_94: mov.b32 %r355, %f1688; xor.b32 %r356, %r355, -2147483648; mov.b32 %f1081, %r356; selp.f32 %f1690, %f1081, %f1688, %p5; setp.geu.f32 %p116, %f279, 0f00000000; @%p116 bra $L__BB0_98; mov.f32 %f1082, 0f3EE66666; cvt.rzi.f32.f32 %f1083, %f1082; setp.eq.f32 %p117, %f1083, 0f3EE66666; @%p117 bra $L__BB0_98; mov.f32 %f1690, 0f7FFFFFFF; $L__BB0_98: add.f32 %f1086, %f295, 0f3EE66666; mov.b32 %r357, %f1086; setp.lt.s32 %p119, %r357, 2139095040; @%p119 bra $L__BB0_103; setp.gtu.f32 %p120, %f295, 0f7F800000; @%p120 bra $L__BB0_102; bra.uni $L__BB0_100; $L__BB0_102: add.f32 %f1690, %f279, 0f3EE66666; bra.uni $L__BB0_103; $L__BB0_100: setp.neu.f32 %p121, %f295, 0f7F800000; @%p121 bra $L__BB0_103; selp.f32 %f1690, 0fFF800000, 0f7F800000, %p5; $L__BB0_103: setp.eq.f32 %p122, %f279, 0f3F800000; selp.f32 %f306, 0f3F800000, %f1690, %p122; abs.f32 %f307, %f281; setp.lt.f32 %p123, %f307, 0f00800000; mul.f32 %f1088, %f307, 0f4B800000; selp.f32 %f1089, %f1088, %f307, %p123; selp.f32 %f1090, 0fC3170000, 0fC2FE0000, %p123; mov.b32 %r358, %f1089; and.b32 %r359, %r358, 8388607; or.b32 %r360, %r359, 1065353216; mov.b32 %f1091, %r360; shr.u32 %r361, %r358, 23; cvt.rn.f32.u32 %f1092, %r361; add.f32 %f1093, %f1090, %f1092; setp.gt.f32 %p124, %f1091, 0f3FB504F3; mul.f32 %f1094, %f1091, 0f3F000000; add.f32 %f1095, %f1093, 0f3F800000; selp.f32 %f1096, %f1095, %f1093, %p124; selp.f32 %f1097, %f1094, %f1091, %p124; add.f32 %f1098, %f1097, 0fBF800000; add.f32 %f1099, %f1097, 0f3F800000; rcp.approx.ftz.f32 %f1100, %f1099; add.f32 %f1101, %f1098, %f1098; mul.f32 %f1102, %f1101, %f1100; mul.f32 %f1103, %f1102, %f1102; mov.f32 %f1104, 0f3C4CAF63; mov.f32 %f1105, 0f3B18F0FE; fma.rn.f32 %f1106, %f1105, %f1103, %f1104; mov.f32 %f1107, 0f3DAAAABD; fma.rn.f32 %f1108, %f1106, %f1103, %f1107; mul.rn.f32 %f1109, %f1108, %f1103; mul.rn.f32 %f1110, %f1109, %f1102; sub.f32 %f1111, %f1098, %f1102; add.f32 %f1112, %f1111, %f1111; neg.f32 %f1113, %f1102; fma.rn.f32 %f1114, %f1113, %f1098, %f1112; mul.rn.f32 %f1115, %f1100, %f1114; add.f32 %f1116, %f1110, %f1102; sub.f32 %f1117, %f1102, %f1116; add.f32 %f1118, %f1110, %f1117; add.f32 %f1119, %f1115, %f1118; add.f32 %f1120, %f1116, %f1119; sub.f32 %f1121, %f1116, %f1120; add.f32 %f1122, %f1119, %f1121; mov.f32 %f1123, 0f3F317200; mul.rn.f32 %f1124, %f1096, %f1123; mov.f32 %f1125, 0f35BFBE8E; mul.rn.f32 %f1126, %f1096, %f1125; add.f32 %f1127, %f1124, %f1120; sub.f32 %f1128, %f1124, %f1127; add.f32 %f1129, %f1120, %f1128; add.f32 %f1130, %f1122, %f1129; add.f32 %f1131, %f1126, %f1130; add.f32 %f1132, %f1127, %f1131; sub.f32 %f1133, %f1127, %f1132; add.f32 %f1134, %f1131, %f1133; mov.f32 %f1135, 0f3EE66666; mul.rn.f32 %f1136, %f1135, %f1132; neg.f32 %f1137, %f1136; fma.rn.f32 %f1138, %f1135, %f1132, %f1137; fma.rn.f32 %f1139, %f1135, %f1134, %f1138; mov.f32 %f1140, 0f00000000; fma.rn.f32 %f1141, %f1140, %f1132, %f1139; add.rn.f32 %f1142, %f1136, %f1141; neg.f32 %f1143, %f1142; add.rn.f32 %f1144, %f1136, %f1143; add.rn.f32 %f1145, %f1144, %f1141; mov.b32 %r362, %f1142; setp.eq.s32 %p125, %r362, 1118925336; add.s32 %r363, %r362, -1; mov.b32 %f1146, %r363; add.f32 %f1147, %f1145, 0f37000000; selp.f32 %f308, %f1147, %f1145, %p125; selp.f32 %f1148, %f1146, %f1142, %p125; mov.f32 %f1149, 0f3FB8AA3B; mul.rn.f32 %f1150, %f1148, %f1149; cvt.rzi.f32.f32 %f1151, %f1150; abs.f32 %f1152, %f1151; setp.gt.f32 %p126, %f1152, 0f42FC0000; mov.b32 %r364, %f1151; and.b32 %r365, %r364, -2147483648; or.b32 %r366, %r365, 1123811328; mov.b32 %f1153, %r366; selp.f32 %f1154, %f1153, %f1151, %p126; mov.f32 %f1155, 0fBF317218; fma.rn.f32 %f1156, %f1154, %f1155, %f1148; mov.f32 %f1157, 0f3102E308; fma.rn.f32 %f1158, %f1154, %f1157, %f1156; mul.f32 %f1159, %f1158, 0f3FB8AA3B; add.f32 %f1160, %f1154, 0f4B40007F; mov.b32 %r367, %f1160; shl.b32 %r368, %r367, 23; mov.b32 %f1161, %r368; ex2.approx.ftz.f32 %f1162, %f1159; mul.f32 %f309, %f1162, %f1161; setp.eq.f32 %p127, %f309, 0f7F800000; mov.f32 %f1691, 0f7F800000; @%p127 bra $L__BB0_105; fma.rn.f32 %f1691, %f309, %f308, %f309; $L__BB0_105: setp.lt.f32 %p128, %f281, 0f00000000; and.pred %p6, %p128, %p99; setp.eq.f32 %p130, %f281, 0f00000000; @%p130 bra $L__BB0_109; bra.uni $L__BB0_106; $L__BB0_109: add.f32 %f1167, %f281, %f281; selp.f32 %f1693, %f1167, 0f00000000, %p99; bra.uni $L__BB0_110; $L__BB0_106: mov.b32 %r369, %f1691; xor.b32 %r370, %r369, -2147483648; mov.b32 %f1163, %r370; selp.f32 %f1693, %f1163, %f1691, %p6; setp.geu.f32 %p131, %f281, 0f00000000; @%p131 bra $L__BB0_110; mov.f32 %f1164, 0f3EE66666; cvt.rzi.f32.f32 %f1165, %f1164; setp.eq.f32 %p132, %f1165, 0f3EE66666; @%p132 bra $L__BB0_110; mov.f32 %f1693, 0f7FFFFFFF; $L__BB0_110: add.f32 %f1168, %f307, 0f3EE66666; mov.b32 %r371, %f1168; setp.lt.s32 %p134, %r371, 2139095040; @%p134 bra $L__BB0_115; setp.gtu.f32 %p135, %f307, 0f7F800000; @%p135 bra $L__BB0_114; bra.uni $L__BB0_112; $L__BB0_114: add.f32 %f1693, %f281, 0f3EE66666; bra.uni $L__BB0_115; $L__BB0_112: setp.neu.f32 %p136, %f307, 0f7F800000; @%p136 bra $L__BB0_115; selp.f32 %f1693, 0fFF800000, 0f7F800000, %p6; $L__BB0_115: setp.eq.f32 %p137, %f281, 0f3F800000; mov.f32 %f1170, 0f3F800000; selp.f32 %f1171, 0f3F800000, %f1693, %p137; ld.const.u64 %rd81, [params+144]; cvta.to.global.u64 %rd22, %rd81; ld.const.u32 %r372, [params+136]; mad.lo.s32 %r373, %r372, %r7, %r6; cvt.u64.u32 %rd23, %r373; min.f32 %f1172, %f294, %f1170; mov.f32 %f1173, 0f00000000; max.f32 %f318, %f1173, %f1172; min.f32 %f1174, %f306, %f1170; max.f32 %f319, %f1173, %f1174; min.f32 %f1175, %f1171, %f1170; max.f32 %f320, %f1173, %f1175; mov.f32 %f1176, 0f3E555555; cvt.rzi.f32.f32 %f1177, %f1176; add.f32 %f1178, %f1177, %f1177; mov.f32 %f1179, 0f3ED55555; sub.f32 %f1180, %f1179, %f1178; abs.f32 %f321, %f1180; abs.f32 %f322, %f318; setp.lt.f32 %p138, %f322, 0f00800000; mul.f32 %f1181, %f322, 0f4B800000; selp.f32 %f1182, %f1181, %f322, %p138; selp.f32 %f1183, 0fC3170000, 0fC2FE0000, %p138; mov.b32 %r374, %f1182; and.b32 %r375, %r374, 8388607; or.b32 %r376, %r375, 1065353216; mov.b32 %f1184, %r376; shr.u32 %r377, %r374, 23; cvt.rn.f32.u32 %f1185, %r377; add.f32 %f1186, %f1183, %f1185; setp.gt.f32 %p139, %f1184, 0f3FB504F3; mul.f32 %f1187, %f1184, 0f3F000000; add.f32 %f1188, %f1186, 0f3F800000; selp.f32 %f1189, %f1188, %f1186, %p139; selp.f32 %f1190, %f1187, %f1184, %p139; add.f32 %f1191, %f1190, 0fBF800000; add.f32 %f1192, %f1190, 0f3F800000; rcp.approx.ftz.f32 %f1193, %f1192; add.f32 %f1194, %f1191, %f1191; mul.f32 %f1195, %f1194, %f1193; mul.f32 %f1196, %f1195, %f1195; mov.f32 %f1197, 0f3C4CAF63; mov.f32 %f1198, 0f3B18F0FE; fma.rn.f32 %f1199, %f1198, %f1196, %f1197; mov.f32 %f1200, 0f3DAAAABD; fma.rn.f32 %f1201, %f1199, %f1196, %f1200; mul.rn.f32 %f1202, %f1201, %f1196; mul.rn.f32 %f1203, %f1202, %f1195; sub.f32 %f1204, %f1191, %f1195; add.f32 %f1205, %f1204, %f1204; neg.f32 %f1206, %f1195; fma.rn.f32 %f1207, %f1206, %f1191, %f1205; mul.rn.f32 %f1208, %f1193, %f1207; add.f32 %f1209, %f1203, %f1195; sub.f32 %f1210, %f1195, %f1209; add.f32 %f1211, %f1203, %f1210; add.f32 %f1212, %f1208, %f1211; add.f32 %f1213, %f1209, %f1212; sub.f32 %f1214, %f1209, %f1213; add.f32 %f1215, %f1212, %f1214; mov.f32 %f1216, 0f3F317200; mul.rn.f32 %f1217, %f1189, %f1216; mov.f32 %f1218, 0f35BFBE8E; mul.rn.f32 %f1219, %f1189, %f1218; add.f32 %f1220, %f1217, %f1213; sub.f32 %f1221, %f1217, %f1220; add.f32 %f1222, %f1213, %f1221; add.f32 %f1223, %f1215, %f1222; add.f32 %f1224, %f1219, %f1223; add.f32 %f1225, %f1220, %f1224; sub.f32 %f1226, %f1220, %f1225; add.f32 %f1227, %f1224, %f1226; mul.rn.f32 %f1228, %f1179, %f1225; neg.f32 %f1229, %f1228; fma.rn.f32 %f1230, %f1179, %f1225, %f1229; fma.rn.f32 %f1231, %f1179, %f1227, %f1230; fma.rn.f32 %f1232, %f1173, %f1225, %f1231; add.rn.f32 %f1233, %f1228, %f1232; neg.f32 %f1234, %f1233; add.rn.f32 %f1235, %f1228, %f1234; add.rn.f32 %f1236, %f1235, %f1232; mov.b32 %r378, %f1233; setp.eq.s32 %p140, %r378, 1118925336; add.s32 %r379, %r378, -1; mov.b32 %f1237, %r379; add.f32 %f1238, %f1236, 0f37000000; selp.f32 %f323, %f1238, %f1236, %p140; selp.f32 %f1239, %f1237, %f1233, %p140; mov.f32 %f1240, 0f3FB8AA3B; mul.rn.f32 %f1241, %f1239, %f1240; cvt.rzi.f32.f32 %f1242, %f1241; abs.f32 %f1243, %f1242; setp.gt.f32 %p141, %f1243, 0f42FC0000; mov.b32 %r380, %f1242; and.b32 %r381, %r380, -2147483648; or.b32 %r382, %r381, 1123811328; mov.b32 %f1244, %r382; selp.f32 %f1245, %f1244, %f1242, %p141; mov.f32 %f1246, 0fBF317218; fma.rn.f32 %f1247, %f1245, %f1246, %f1239; mov.f32 %f1248, 0f3102E308; fma.rn.f32 %f1249, %f1245, %f1248, %f1247; mul.f32 %f1250, %f1249, 0f3FB8AA3B; add.f32 %f1251, %f1245, 0f4B40007F; mov.b32 %r383, %f1251; shl.b32 %r384, %r383, 23; mov.b32 %f1252, %r384; ex2.approx.ftz.f32 %f1253, %f1250; mul.f32 %f324, %f1253, %f1252; setp.eq.f32 %p142, %f324, 0f7F800000; mov.f32 %f1694, 0f7F800000; @%p142 bra $L__BB0_117; fma.rn.f32 %f1694, %f324, %f323, %f324; $L__BB0_117: setp.lt.f32 %p143, %f318, 0f00000000; setp.eq.f32 %p144, %f321, 0f3F800000; and.pred %p7, %p143, %p144; setp.eq.f32 %p145, %f318, 0f00000000; @%p145 bra $L__BB0_121; bra.uni $L__BB0_118; $L__BB0_121: add.f32 %f1258, %f318, %f318; selp.f32 %f1696, %f1258, 0f00000000, %p144; bra.uni $L__BB0_122; $L__BB0_118: mov.b32 %r385, %f1694; xor.b32 %r386, %r385, -2147483648; mov.b32 %f1254, %r386; selp.f32 %f1696, %f1254, %f1694, %p7; setp.geu.f32 %p146, %f318, 0f00000000; @%p146 bra $L__BB0_122; mov.f32 %f1255, 0f3ED55555; cvt.rzi.f32.f32 %f1256, %f1255; setp.eq.f32 %p147, %f1256, 0f3ED55555; @%p147 bra $L__BB0_122; mov.f32 %f1696, 0f7FFFFFFF; $L__BB0_122: add.f32 %f1259, %f322, 0f3ED55555; mov.b32 %r387, %f1259; setp.lt.s32 %p149, %r387, 2139095040; @%p149 bra $L__BB0_127; setp.gtu.f32 %p150, %f322, 0f7F800000; @%p150 bra $L__BB0_126; bra.uni $L__BB0_124; $L__BB0_126: add.f32 %f1696, %f318, 0f3ED55555; bra.uni $L__BB0_127; $L__BB0_124: setp.neu.f32 %p151, %f322, 0f7F800000; @%p151 bra $L__BB0_127; selp.f32 %f1696, 0fFF800000, 0f7F800000, %p7; $L__BB0_127: abs.f32 %f333, %f319; setp.lt.f32 %p152, %f333, 0f00800000; mul.f32 %f1261, %f333, 0f4B800000; selp.f32 %f1262, %f1261, %f333, %p152; selp.f32 %f1263, 0fC3170000, 0fC2FE0000, %p152; mov.b32 %r388, %f1262; and.b32 %r389, %r388, 8388607; or.b32 %r390, %r389, 1065353216; mov.b32 %f1264, %r390; shr.u32 %r391, %r388, 23; cvt.rn.f32.u32 %f1265, %r391; add.f32 %f1266, %f1263, %f1265; setp.gt.f32 %p153, %f1264, 0f3FB504F3; mul.f32 %f1267, %f1264, 0f3F000000; add.f32 %f1268, %f1266, 0f3F800000; selp.f32 %f1269, %f1268, %f1266, %p153; selp.f32 %f1270, %f1267, %f1264, %p153; add.f32 %f1271, %f1270, 0fBF800000; add.f32 %f1272, %f1270, 0f3F800000; rcp.approx.ftz.f32 %f1273, %f1272; add.f32 %f1274, %f1271, %f1271; mul.f32 %f1275, %f1274, %f1273; mul.f32 %f1276, %f1275, %f1275; mov.f32 %f1277, 0f3C4CAF63; mov.f32 %f1278, 0f3B18F0FE; fma.rn.f32 %f1279, %f1278, %f1276, %f1277; mov.f32 %f1280, 0f3DAAAABD; fma.rn.f32 %f1281, %f1279, %f1276, %f1280; mul.rn.f32 %f1282, %f1281, %f1276; mul.rn.f32 %f1283, %f1282, %f1275; sub.f32 %f1284, %f1271, %f1275; add.f32 %f1285, %f1284, %f1284; neg.f32 %f1286, %f1275; fma.rn.f32 %f1287, %f1286, %f1271, %f1285; mul.rn.f32 %f1288, %f1273, %f1287; add.f32 %f1289, %f1283, %f1275; sub.f32 %f1290, %f1275, %f1289; add.f32 %f1291, %f1283, %f1290; add.f32 %f1292, %f1288, %f1291; add.f32 %f1293, %f1289, %f1292; sub.f32 %f1294, %f1289, %f1293; add.f32 %f1295, %f1292, %f1294; mov.f32 %f1296, 0f3F317200; mul.rn.f32 %f1297, %f1269, %f1296; mov.f32 %f1298, 0f35BFBE8E; mul.rn.f32 %f1299, %f1269, %f1298; add.f32 %f1300, %f1297, %f1293; sub.f32 %f1301, %f1297, %f1300; add.f32 %f1302, %f1293, %f1301; add.f32 %f1303, %f1295, %f1302; add.f32 %f1304, %f1299, %f1303; add.f32 %f1305, %f1300, %f1304; sub.f32 %f1306, %f1300, %f1305; add.f32 %f1307, %f1304, %f1306; mov.f32 %f1308, 0f3ED55555; mul.rn.f32 %f1309, %f1308, %f1305; neg.f32 %f1310, %f1309; fma.rn.f32 %f1311, %f1308, %f1305, %f1310; fma.rn.f32 %f1312, %f1308, %f1307, %f1311; mov.f32 %f1313, 0f00000000; fma.rn.f32 %f1314, %f1313, %f1305, %f1312; add.rn.f32 %f1315, %f1309, %f1314; neg.f32 %f1316, %f1315; add.rn.f32 %f1317, %f1309, %f1316; add.rn.f32 %f1318, %f1317, %f1314; mov.b32 %r392, %f1315; setp.eq.s32 %p154, %r392, 1118925336; add.s32 %r393, %r392, -1; mov.b32 %f1319, %r393; add.f32 %f1320, %f1318, 0f37000000; selp.f32 %f334, %f1320, %f1318, %p154; selp.f32 %f1321, %f1319, %f1315, %p154; mov.f32 %f1322, 0f3FB8AA3B; mul.rn.f32 %f1323, %f1321, %f1322; cvt.rzi.f32.f32 %f1324, %f1323; abs.f32 %f1325, %f1324; setp.gt.f32 %p155, %f1325, 0f42FC0000; mov.b32 %r394, %f1324; and.b32 %r395, %r394, -2147483648; or.b32 %r396, %r395, 1123811328; mov.b32 %f1326, %r396; selp.f32 %f1327, %f1326, %f1324, %p155; mov.f32 %f1328, 0fBF317218; fma.rn.f32 %f1329, %f1327, %f1328, %f1321; mov.f32 %f1330, 0f3102E308; fma.rn.f32 %f1331, %f1327, %f1330, %f1329; mul.f32 %f1332, %f1331, 0f3FB8AA3B; add.f32 %f1333, %f1327, 0f4B40007F; mov.b32 %r397, %f1333; shl.b32 %r398, %r397, 23; mov.b32 %f1334, %r398; ex2.approx.ftz.f32 %f1335, %f1332; mul.f32 %f335, %f1335, %f1334; setp.eq.f32 %p156, %f335, 0f7F800000; mov.f32 %f1697, 0f7F800000; @%p156 bra $L__BB0_129; fma.rn.f32 %f1697, %f335, %f334, %f335; $L__BB0_129: setp.lt.f32 %p157, %f319, 0f00000000; and.pred %p8, %p157, %p144; setp.eq.f32 %p159, %f319, 0f00000000; @%p159 bra $L__BB0_133; bra.uni $L__BB0_130; $L__BB0_133: add.f32 %f1340, %f319, %f319; selp.f32 %f1699, %f1340, 0f00000000, %p144; bra.uni $L__BB0_134; $L__BB0_130: mov.b32 %r399, %f1697; xor.b32 %r400, %r399, -2147483648; mov.b32 %f1336, %r400; selp.f32 %f1699, %f1336, %f1697, %p8; setp.geu.f32 %p160, %f319, 0f00000000; @%p160 bra $L__BB0_134; mov.f32 %f1337, 0f3ED55555; cvt.rzi.f32.f32 %f1338, %f1337; setp.eq.f32 %p161, %f1338, 0f3ED55555; @%p161 bra $L__BB0_134; mov.f32 %f1699, 0f7FFFFFFF; $L__BB0_134: add.f32 %f1341, %f333, 0f3ED55555; mov.b32 %r401, %f1341; setp.lt.s32 %p163, %r401, 2139095040; @%p163 bra $L__BB0_139; setp.gtu.f32 %p164, %f333, 0f7F800000; @%p164 bra $L__BB0_138; bra.uni $L__BB0_136; $L__BB0_138: add.f32 %f1699, %f319, 0f3ED55555; bra.uni $L__BB0_139; $L__BB0_136: setp.neu.f32 %p165, %f333, 0f7F800000; @%p165 bra $L__BB0_139; selp.f32 %f1699, 0fFF800000, 0f7F800000, %p8; $L__BB0_139: abs.f32 %f344, %f320; setp.lt.f32 %p166, %f344, 0f00800000; mul.f32 %f1343, %f344, 0f4B800000; selp.f32 %f1344, %f1343, %f344, %p166; selp.f32 %f1345, 0fC3170000, 0fC2FE0000, %p166; mov.b32 %r402, %f1344; and.b32 %r403, %r402, 8388607; or.b32 %r404, %r403, 1065353216; mov.b32 %f1346, %r404; shr.u32 %r405, %r402, 23; cvt.rn.f32.u32 %f1347, %r405; add.f32 %f1348, %f1345, %f1347; setp.gt.f32 %p167, %f1346, 0f3FB504F3; mul.f32 %f1349, %f1346, 0f3F000000; add.f32 %f1350, %f1348, 0f3F800000; selp.f32 %f1351, %f1350, %f1348, %p167; selp.f32 %f1352, %f1349, %f1346, %p167; add.f32 %f1353, %f1352, 0fBF800000; add.f32 %f1354, %f1352, 0f3F800000; rcp.approx.ftz.f32 %f1355, %f1354; add.f32 %f1356, %f1353, %f1353; mul.f32 %f1357, %f1356, %f1355; mul.f32 %f1358, %f1357, %f1357; mov.f32 %f1359, 0f3C4CAF63; mov.f32 %f1360, 0f3B18F0FE; fma.rn.f32 %f1361, %f1360, %f1358, %f1359; mov.f32 %f1362, 0f3DAAAABD; fma.rn.f32 %f1363, %f1361, %f1358, %f1362; mul.rn.f32 %f1364, %f1363, %f1358; mul.rn.f32 %f1365, %f1364, %f1357; sub.f32 %f1366, %f1353, %f1357; add.f32 %f1367, %f1366, %f1366; neg.f32 %f1368, %f1357; fma.rn.f32 %f1369, %f1368, %f1353, %f1367; mul.rn.f32 %f1370, %f1355, %f1369; add.f32 %f1371, %f1365, %f1357; sub.f32 %f1372, %f1357, %f1371; add.f32 %f1373, %f1365, %f1372; add.f32 %f1374, %f1370, %f1373; add.f32 %f1375, %f1371, %f1374; sub.f32 %f1376, %f1371, %f1375; add.f32 %f1377, %f1374, %f1376; mov.f32 %f1378, 0f3F317200; mul.rn.f32 %f1379, %f1351, %f1378; mov.f32 %f1380, 0f35BFBE8E; mul.rn.f32 %f1381, %f1351, %f1380; add.f32 %f1382, %f1379, %f1375; sub.f32 %f1383, %f1379, %f1382; add.f32 %f1384, %f1375, %f1383; add.f32 %f1385, %f1377, %f1384; add.f32 %f1386, %f1381, %f1385; add.f32 %f1387, %f1382, %f1386; sub.f32 %f1388, %f1382, %f1387; add.f32 %f1389, %f1386, %f1388; mov.f32 %f1390, 0f3ED55555; mul.rn.f32 %f1391, %f1390, %f1387; neg.f32 %f1392, %f1391; fma.rn.f32 %f1393, %f1390, %f1387, %f1392; fma.rn.f32 %f1394, %f1390, %f1389, %f1393; mov.f32 %f1395, 0f00000000; fma.rn.f32 %f1396, %f1395, %f1387, %f1394; add.rn.f32 %f1397, %f1391, %f1396; neg.f32 %f1398, %f1397; add.rn.f32 %f1399, %f1391, %f1398; add.rn.f32 %f1400, %f1399, %f1396; mov.b32 %r406, %f1397; setp.eq.s32 %p168, %r406, 1118925336; add.s32 %r407, %r406, -1; mov.b32 %f1401, %r407; add.f32 %f1402, %f1400, 0f37000000; selp.f32 %f345, %f1402, %f1400, %p168; selp.f32 %f1403, %f1401, %f1397, %p168; mov.f32 %f1404, 0f3FB8AA3B; mul.rn.f32 %f1405, %f1403, %f1404; cvt.rzi.f32.f32 %f1406, %f1405; abs.f32 %f1407, %f1406; setp.gt.f32 %p169, %f1407, 0f42FC0000; mov.b32 %r408, %f1406; and.b32 %r409, %r408, -2147483648; or.b32 %r410, %r409, 1123811328; mov.b32 %f1408, %r410; selp.f32 %f1409, %f1408, %f1406, %p169; mov.f32 %f1410, 0fBF317218; fma.rn.f32 %f1411, %f1409, %f1410, %f1403; mov.f32 %f1412, 0f3102E308; fma.rn.f32 %f1413, %f1409, %f1412, %f1411; mul.f32 %f1414, %f1413, 0f3FB8AA3B; add.f32 %f1415, %f1409, 0f4B40007F; mov.b32 %r411, %f1415; shl.b32 %r412, %r411, 23; mov.b32 %f1416, %r412; ex2.approx.ftz.f32 %f1417, %f1414; mul.f32 %f346, %f1417, %f1416; setp.eq.f32 %p170, %f346, 0f7F800000; mov.f32 %f1700, 0f7F800000; @%p170 bra $L__BB0_141; fma.rn.f32 %f1700, %f346, %f345, %f346; $L__BB0_141: setp.lt.f32 %p171, %f320, 0f00000000; and.pred %p9, %p171, %p144; setp.eq.f32 %p173, %f320, 0f00000000; @%p173 bra $L__BB0_145; bra.uni $L__BB0_142; $L__BB0_145: add.f32 %f1422, %f320, %f320; selp.f32 %f1702, %f1422, 0f00000000, %p144; bra.uni $L__BB0_146; $L__BB0_142: mov.b32 %r413, %f1700; xor.b32 %r414, %r413, -2147483648; mov.b32 %f1418, %r414; selp.f32 %f1702, %f1418, %f1700, %p9; setp.geu.f32 %p174, %f320, 0f00000000; @%p174 bra $L__BB0_146; mov.f32 %f1419, 0f3ED55555; cvt.rzi.f32.f32 %f1420, %f1419; setp.eq.f32 %p175, %f1420, 0f3ED55555; @%p175 bra $L__BB0_146; mov.f32 %f1702, 0f7FFFFFFF; $L__BB0_146: add.f32 %f1423, %f344, 0f3ED55555; mov.b32 %r415, %f1423; setp.lt.s32 %p177, %r415, 2139095040; @%p177 bra $L__BB0_151; setp.gtu.f32 %p178, %f344, 0f7F800000; @%p178 bra $L__BB0_150; bra.uni $L__BB0_148; $L__BB0_150: add.f32 %f1702, %f320, 0f3ED55555; bra.uni $L__BB0_151; $L__BB0_148: setp.neu.f32 %p179, %f344, 0f7F800000; @%p179 bra $L__BB0_151; selp.f32 %f1702, 0fFF800000, 0f7F800000, %p9; $L__BB0_151: fma.rn.f32 %f1424, %f1696, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p180, %f318, 0f3F800000; mov.f32 %f1425, 0f3F800000; selp.f32 %f1426, 0f3F7FFFFF, %f1424, %p180; mul.f32 %f1427, %f318, 0f414EB852; setp.lt.f32 %p181, %f318, 0f3B4D2E1C; selp.f32 %f1428, %f1427, %f1426, %p181; fma.rn.f32 %f1429, %f1699, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p182, %f319, 0f3F800000; selp.f32 %f1430, 0f3F7FFFFF, %f1429, %p182; mul.f32 %f1431, %f319, 0f414EB852; setp.lt.f32 %p183, %f319, 0f3B4D2E1C; selp.f32 %f1432, %f1431, %f1430, %p183; fma.rn.f32 %f1433, %f1702, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p184, %f320, 0f3F800000; selp.f32 %f1434, 0f3F7FFFFF, %f1433, %p184; mul.f32 %f1435, %f320, 0f414EB852; setp.lt.f32 %p185, %f320, 0f3B4D2E1C; selp.f32 %f1436, %f1435, %f1434, %p185; min.f32 %f1437, %f1428, %f1425; mov.f32 %f1438, 0f00000000; max.f32 %f1439, %f1438, %f1437; mul.f32 %f1440, %f1439, 0f43800000; cvt.rzi.u32.f32 %r416, %f1440; min.u32 %r417, %r416, 255; min.f32 %f1441, %f1432, %f1425; max.f32 %f1442, %f1438, %f1441; mul.f32 %f1443, %f1442, 0f43800000; cvt.rzi.u32.f32 %r418, %f1443; min.u32 %r419, %r418, 255; min.f32 %f1444, %f1436, %f1425; max.f32 %f1445, %f1438, %f1444; mul.f32 %f1446, %f1445, 0f43800000; cvt.rzi.u32.f32 %r420, %f1446; min.u32 %r421, %r420, 255; shl.b64 %rd82, %rd23, 2; add.s64 %rd83, %rd22, %rd82; cvt.u16.u32 %rs47, %r421; cvt.u16.u32 %rs48, %r419; cvt.u16.u32 %rs49, %r417; mov.u16 %rs50, 255; st.global.v4.u8 [%rd83], {%rs49, %rs48, %rs47, %rs50}; $L__BB0_152: and.b32 %r422, %r56, 4; setp.eq.s32 %p186, %r422, 0; ld.const.u32 %r465, [params+108]; @%p186 bra $L__BB0_156; setp.eq.s32 %p187, %r465, 0; ld.const.u64 %rd84, [params+224]; cvta.to.global.u64 %rd85, %rd84; ld.const.u32 %r423, [params+216]; mad.lo.s32 %r424, %r423, %r7, %r6; mul.wide.u32 %rd86, %r424, 8; add.s64 %rd24, %rd85, %rd86; @%p187 bra $L__BB0_155; ld.global.v4.u16 {%rs58, %rs59, %rs60, %rs61}, [%rd24]; // begin inline asm { cvt.f32.f16 %f1447, %rs58;} // end inline asm // begin inline asm { cvt.f32.f16 %f1448, %rs59;} // end inline asm // begin inline asm { cvt.f32.f16 %f1449, %rs60;} // end inline asm add.f32 %f1450, %f278, %f1447; add.f32 %f1451, %f279, %f1448; add.f32 %f1452, %f281, %f1449; mov.f32 %f1453, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs56, %f1452;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs55, %f1451;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs54, %f1450;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs57, %f1453;} // end inline asm st.global.v4.u16 [%rd24], {%rs54, %rs55, %rs56, %rs57}; bra.uni $L__BB0_156; $L__BB0_155: mov.f32 %f1457, 0f3F800000; mov.u32 %r465, 0; // begin inline asm { cvt.rn.f16.f32 %rs65, %f1457;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs64, %f281;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs63, %f279;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs62, %f278;} // end inline asm st.global.v4.u16 [%rd24], {%rs62, %rs63, %rs64, %rs65}; $L__BB0_156: mul.f32 %f1458, %f1672, %f917; mul.f32 %f1459, %f1675, %f917; add.f32 %f1460, %f1459, %f1458; mul.f32 %f1461, %f1673, %f918; mul.f32 %f1462, %f1676, %f918; add.f32 %f1463, %f1462, %f1461; mul.f32 %f1464, %f1674, %f280; mul.f32 %f1465, %f1677, %f280; add.f32 %f1466, %f1465, %f1464; mul.f32 %f1467, %f1669, %f917; add.f32 %f1468, %f1460, %f1467; mul.f32 %f1469, %f1670, %f918; add.f32 %f1470, %f1463, %f1469; mul.f32 %f1471, %f1671, %f280; add.f32 %f1472, %f1466, %f1471; mul.f32 %f1473, %f1468, 0f3F13CD3A; mul.f32 %f1474, %f1470, 0f3F13CD3A; mul.f32 %f1475, %f1472, 0f3F13CD3A; div.rn.f32 %f1476, %f278, %f1473; div.rn.f32 %f1477, %f279, %f1474; div.rn.f32 %f1478, %f281, %f1475; setp.eq.f32 %p188, %f278, 0f00000000; selp.f32 %f1479, 0f00000000, %f1476, %p188; setp.eq.f32 %p189, %f279, 0f00000000; selp.f32 %f1480, 0f00000000, %f1477, %p189; setp.eq.f32 %p190, %f281, 0f00000000; selp.f32 %f1481, 0f00000000, %f1478, %p190; mul.f32 %f355, %f1459, %f1479; mul.f32 %f356, %f1462, %f1480; mul.f32 %f357, %f1465, %f1481; mul.f32 %f358, %f1458, %f1479; mul.f32 %f359, %f1461, %f1480; mul.f32 %f360, %f1464, %f1481; mul.f32 %f361, %f1467, %f1479; mul.f32 %f362, %f1469, %f1480; mul.f32 %f363, %f1471, %f1481; ld.const.u64 %rd87, [params+256]; cvta.to.global.u64 %rd88, %rd87; ld.const.u32 %r426, [params+248]; mad.lo.s32 %r427, %r426, %r7, %r6; mul.wide.u32 %rd89, %r427, 8; add.s64 %rd25, %rd88, %rd89; setp.eq.s32 %p191, %r465, 0; @%p191 bra $L__BB0_158; ld.global.v4.u16 {%rs73, %rs74, %rs75, %rs76}, [%rd25]; // begin inline asm { cvt.f32.f16 %f1482, %rs73;} // end inline asm // begin inline asm { cvt.f32.f16 %f1483, %rs74;} // end inline asm // begin inline asm { cvt.f32.f16 %f1484, %rs75;} // end inline asm add.f32 %f1485, %f355, %f1482; add.f32 %f1486, %f356, %f1483; add.f32 %f1487, %f357, %f1484; mov.f32 %f1488, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs71, %f1487;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs70, %f1486;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs69, %f1485;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs72, %f1488;} // end inline asm st.global.v4.u16 [%rd25], {%rs69, %rs70, %rs71, %rs72}; bra.uni $L__BB0_159; $L__BB0_158: mov.f32 %f1492, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs80, %f1492;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs79, %f357;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs78, %f356;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs77, %f355;} // end inline asm st.global.v4.u16 [%rd25], {%rs77, %rs78, %rs79, %rs80}; $L__BB0_159: ld.const.u64 %rd90, [params+272]; cvta.to.global.u64 %rd91, %rd90; ld.const.u32 %r428, [params+264]; mad.lo.s32 %r429, %r428, %r7, %r6; mul.wide.u32 %rd92, %r429, 8; add.s64 %rd26, %rd91, %rd92; @%p191 bra $L__BB0_161; ld.global.v4.u16 {%rs88, %rs89, %rs90, %rs91}, [%rd26]; // begin inline asm { cvt.f32.f16 %f1493, %rs88;} // end inline asm // begin inline asm { cvt.f32.f16 %f1494, %rs89;} // end inline asm // begin inline asm { cvt.f32.f16 %f1495, %rs90;} // end inline asm add.f32 %f1496, %f358, %f1493; add.f32 %f1497, %f359, %f1494; add.f32 %f1498, %f360, %f1495; mov.f32 %f1499, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs86, %f1498;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs85, %f1497;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs84, %f1496;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs87, %f1499;} // end inline asm st.global.v4.u16 [%rd26], {%rs84, %rs85, %rs86, %rs87}; bra.uni $L__BB0_162; $L__BB0_161: mov.f32 %f1503, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs95, %f1503;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs94, %f360;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs93, %f359;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs92, %f358;} // end inline asm st.global.v4.u16 [%rd26], {%rs92, %rs93, %rs94, %rs95}; $L__BB0_162: ld.const.u64 %rd93, [params+288]; cvta.to.global.u64 %rd94, %rd93; ld.const.u32 %r430, [params+280]; mad.lo.s32 %r431, %r430, %r7, %r6; mul.wide.u32 %rd95, %r431, 8; add.s64 %rd27, %rd94, %rd95; @%p191 bra $L__BB0_164; ld.global.v4.u16 {%rs103, %rs104, %rs105, %rs106}, [%rd27]; // begin inline asm { cvt.f32.f16 %f1504, %rs103;} // end inline asm // begin inline asm { cvt.f32.f16 %f1505, %rs104;} // end inline asm // begin inline asm { cvt.f32.f16 %f1506, %rs105;} // end inline asm add.f32 %f1507, %f361, %f1504; add.f32 %f1508, %f362, %f1505; add.f32 %f1509, %f363, %f1506; mov.f32 %f1510, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs101, %f1509;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs100, %f1508;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs99, %f1507;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs102, %f1510;} // end inline asm st.global.v4.u16 [%rd27], {%rs99, %rs100, %rs101, %rs102}; bra.uni $L__BB0_182; $L__BB0_164: mov.f32 %f1514, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs110, %f1514;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs109, %f363;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs108, %f362;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs107, %f361;} // end inline asm st.global.v4.u16 [%rd27], {%rs107, %rs108, %rs109, %rs110}; $L__BB0_182: ret; }