// // 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<207>; .reg .b16 %rs<180>; .reg .f32 %f<1718>; .reg .b32 %r<500>; .reg .f64 %fd<5>; .reg .b64 %rd<123>; mov.u64 %SPL, __local_depot0; add.u64 %rd1, %SPL, 0; // begin inline asm call (%r94), _optix_get_launch_index_x, (); // end inline asm // begin inline asm call (%r95), _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 %r100, [params+392]; mad.lo.s32 %r101, %r100, %r95, %r94; mul.wide.u32 %rd35, %r101, 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 %f1569, 0f00000000; mov.u16 %rs175, 0; mov.f32 %f1570, %f1569; mov.f32 %f1571, %f1569; @%p11 bra $L__BB0_4; $L__BB0_3: cvt.rn.f32.u16 %f381, %rs16; div.rn.f32 %f382, %f381, 0f437F0000; fma.rn.f32 %f383, %f382, 0f40000000, 0fBF800000; and.b16 %rs22, %rs175, 255; cvt.rn.f32.u16 %f384, %rs22; div.rn.f32 %f385, %f384, 0f437F0000; fma.rn.f32 %f386, %f385, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f387, %rs174; div.rn.f32 %f388, %f387, 0f437F0000; fma.rn.f32 %f389, %f388, 0f40000000, 0fBF800000; mul.f32 %f390, %f386, %f386; fma.rn.f32 %f391, %f383, %f383, %f390; fma.rn.f32 %f392, %f389, %f389, %f391; sqrt.rn.f32 %f393, %f392; rcp.rn.f32 %f394, %f393; mul.f32 %f1571, %f394, %f389; mul.f32 %f1570, %f394, %f386; mul.f32 %f1569, %f383, %f394; $L__BB0_4: ld.const.v2.u32 {%r102, %r103}, [params]; add.s32 %r6, %r102, %r94; add.s32 %r7, %r103, %r95; setp.eq.f32 %p12, %f1569, 0f00000000; setp.eq.f32 %p13, %f1570, 0f00000000; and.pred %p14, %p12, %p13; setp.eq.f32 %p15, %f1571, 0f00000000; and.pred %p16, %p15, %p14; @%p16 bra $L__BB0_167; bra.uni $L__BB0_5; $L__BB0_167: ld.const.u32 %r91, [params+104]; and.b32 %r464, %r91, 1; setp.eq.b32 %p196, %r464, 1; mov.pred %p197, 0; xor.pred %p198, %p196, %p197; not.pred %p199, %p198; @%p199 bra $L__BB0_169; ld.const.u64 %rd96, [params+144]; cvta.to.global.u64 %rd97, %rd96; ld.const.u32 %r465, [params+136]; mad.lo.s32 %r466, %r465, %r7, %r6; mul.wide.u32 %rd98, %r466, 4; add.s64 %rd99, %rd97, %rd98; mov.u16 %rs111, 0; st.global.v4.u8 [%rd99], {%rs111, %rs111, %rs111, %rs111}; $L__BB0_169: and.b32 %r467, %r91, 8; setp.eq.s32 %p200, %r467, 0; @%p200 bra $L__BB0_171; ld.const.u64 %rd100, [params+192]; cvta.to.global.u64 %rd101, %rd100; ld.const.u32 %r468, [params+184]; mad.lo.s32 %r469, %r468, %r7, %r6; mov.f32 %f1524, 0f00000000; cvt.rzi.u32.f32 %r470, %f1524; mul.wide.u32 %rd102, %r469, 2; add.s64 %rd103, %rd101, %rd102; mov.u16 %rs112, 0; cvt.u16.u32 %rs113, %r470; st.global.v2.u8 [%rd103], {%rs113, %rs112}; $L__BB0_171: and.b32 %r471, %r91, 4; setp.eq.s32 %p201, %r471, 0; ld.const.u32 %r499, [params+108]; @%p201 bra $L__BB0_175; setp.eq.s32 %p202, %r499, 0; ld.const.u64 %rd104, [params+224]; cvta.to.global.u64 %rd105, %rd104; ld.const.u32 %r472, [params+216]; mad.lo.s32 %r473, %r472, %r7, %r6; mul.wide.u32 %rd106, %r473, 8; add.s64 %rd28, %rd105, %rd106; @%p202 bra $L__BB0_174; ld.global.v4.u16 {%rs120, %rs121, %rs122, %rs123}, [%rd28]; // begin inline asm { cvt.f32.f16 %f1525, %rs120;} // end inline asm // begin inline asm { cvt.f32.f16 %f1526, %rs121;} // end inline asm // begin inline asm { cvt.f32.f16 %f1527, %rs122;} // 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 %rs119, %f1530;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs118, %f1529;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs117, %f1528;} // end inline asm mov.u16 %rs124, 0; st.global.v4.u16 [%rd28], {%rs117, %rs118, %rs119, %rs124}; bra.uni $L__BB0_175; $L__BB0_5: ld.const.u64 %rd36, [params+432]; cvta.to.global.u64 %rd37, %rd36; ld.const.u32 %r106, [params+424]; mad.lo.s32 %r107, %r106, %r95, %r94; mul.wide.u32 %rd38, %r107, 12; add.s64 %rd39, %rd37, %rd38; ld.global.f32 %f10, [%rd39]; mul.f32 %f395, %f10, 0f3456BF95; ld.global.f32 %f11, [%rd39+4]; mul.f32 %f396, %f11, 0f3456BF95; ld.global.f32 %f12, [%rd39+8]; mul.f32 %f397, %f12, 0f3456BF95; abs.f32 %f13, %f1569; div.rn.f32 %f398, %f395, %f13; abs.f32 %f399, %f1570; div.rn.f32 %f400, %f396, %f399; abs.f32 %f14, %f1571; div.rn.f32 %f401, %f397, %f14; abs.f32 %f402, %f398; abs.f32 %f403, %f400; abs.f32 %f404, %f401; mov.f32 %f405, 0f38D1B717; max.f32 %f406, %f402, %f405; max.f32 %f407, %f403, %f405; max.f32 %f408, %f404, %f405; fma.rn.f32 %f15, %f1569, %f406, %f10; fma.rn.f32 %f16, %f1570, %f407, %f11; fma.rn.f32 %f17, %f1571, %f408, %f12; ld.const.u64 %rd40, [params+464]; cvta.to.global.u64 %rd41, %rd40; ld.const.u32 %r108, [params+456]; mad.lo.s32 %r109, %r108, %r95, %r94; mul.wide.u32 %rd42, %r109, 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 %f1572, 0f00000000; mov.u16 %rs178, 0; mov.u16 %rs179, %rs178; mov.f32 %f1573, %f1572; mov.f32 %f1574, %f1572; @%p18 bra $L__BB0_9; $L__BB0_8: mov.u16 %rs179, %rs178; cvt.rn.f32.u16 %f412, %rs24; div.rn.f32 %f413, %f412, 0f437F0000; fma.rn.f32 %f414, %f413, 0f40000000, 0fBF800000; and.b16 %rs32, %rs179, 255; cvt.rn.f32.u16 %f415, %rs32; div.rn.f32 %f416, %f415, 0f437F0000; fma.rn.f32 %f417, %f416, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f418, %rs176; div.rn.f32 %f419, %f418, 0f437F0000; fma.rn.f32 %f420, %f419, 0f40000000, 0fBF800000; mul.f32 %f421, %f417, %f417; fma.rn.f32 %f422, %f414, %f414, %f421; fma.rn.f32 %f423, %f420, %f420, %f422; sqrt.rn.f32 %f424, %f423; rcp.rn.f32 %f425, %f424; mul.f32 %f1574, %f425, %f420; mul.f32 %f1573, %f425, %f417; mul.f32 %f1572, %f414, %f425; mov.u16 %rs178, %rs176; $L__BB0_9: mul.f32 %f429, %f1570, %f1574; mul.f32 %f430, %f1571, %f1573; sub.f32 %f431, %f429, %f430; mul.f32 %f432, %f1571, %f1572; mul.f32 %f433, %f1569, %f1574; sub.f32 %f434, %f432, %f433; mul.f32 %f435, %f1569, %f1573; mul.f32 %f436, %f1570, %f1572; sub.f32 %f437, %f435, %f436; ld.global.u8 %rs34, [%rd3+3]; setp.eq.s16 %p19, %rs34, 0; selp.f32 %f24, 0fBF800000, 0f3F800000, %p19; mul.f32 %f438, %f431, %f24; mul.f32 %f439, %f434, %f24; mul.f32 %f440, %f437, %f24; mul.f32 %f441, %f438, 0f00000000; mul.f32 %f442, %f439, 0f00000000; mul.f32 %f443, %f440, 0f00000000; fma.rn.f32 %f444, %f1572, 0f3F5105EC, %f441; fma.rn.f32 %f445, %f1573, 0f3F5105EC, %f442; fma.rn.f32 %f446, %f1574, 0f3F5105EC, %f443; mul.f32 %f25, %f1569, 0f3F13CD3A; add.f32 %f26, %f25, %f444; mul.f32 %f27, %f1570, 0f3F13CD3A; add.f32 %f28, %f27, %f445; mul.f32 %f29, %f1571, 0f3F13CD3A; add.f32 %f30, %f29, %f446; or.b16 %rs35, %rs24, %rs179; or.b16 %rs15, %rs35, %rs178; and.b16 %rs36, %rs15, 255; setp.eq.s16 %p20, %rs36, 0; mov.f32 %f1578, 0f00000000; mov.f32 %f1575, %f1578; mov.f32 %f1576, %f1578; mov.f32 %f1577, %f1578; @%p20 bra $L__BB0_11; cvt.rn.f32.u16 %f447, %rs24; div.rn.f32 %f448, %f447, 0f437F0000; fma.rn.f32 %f449, %f448, 0f40000000, 0fBF800000; and.b16 %rs38, %rs179, 255; cvt.rn.f32.u16 %f450, %rs38; div.rn.f32 %f451, %f450, 0f437F0000; fma.rn.f32 %f452, %f451, 0f40000000, 0fBF800000; and.b16 %rs39, %rs178, 255; cvt.rn.f32.u16 %f453, %rs39; div.rn.f32 %f454, %f453, 0f437F0000; fma.rn.f32 %f455, %f454, 0f40000000, 0fBF800000; mul.f32 %f456, %f452, %f452; fma.rn.f32 %f457, %f449, %f449, %f456; fma.rn.f32 %f458, %f455, %f455, %f457; sqrt.rn.f32 %f459, %f458; rcp.rn.f32 %f460, %f459; mul.f32 %f1577, %f460, %f455; mul.f32 %f1576, %f460, %f452; mul.f32 %f1575, %f449, %f460; $L__BB0_11: mul.f32 %f464, %f1570, %f1577; mul.f32 %f465, %f1571, %f1576; sub.f32 %f466, %f464, %f465; mul.f32 %f467, %f1571, %f1575; mul.f32 %f468, %f1569, %f1577; sub.f32 %f469, %f467, %f468; mul.f32 %f470, %f1569, %f1576; mul.f32 %f471, %f1570, %f1575; sub.f32 %f472, %f470, %f471; mul.f32 %f473, %f466, %f24; mul.f32 %f474, %f469, %f24; mul.f32 %f475, %f472, %f24; mul.f32 %f476, %f1575, 0f3ED105EC; mul.f32 %f477, %f1576, 0f3ED105EC; mul.f32 %f478, %f1577, 0f3ED105EC; mul.f32 %f479, %f473, 0f3F3504F3; mul.f32 %f480, %f474, 0f3F3504F3; mul.f32 %f481, %f475, 0f3F3504F3; sub.f32 %f482, %f479, %f476; sub.f32 %f483, %f480, %f477; sub.f32 %f484, %f481, %f478; add.f32 %f37, %f25, %f482; add.f32 %f38, %f27, %f483; add.f32 %f39, %f29, %f484; mov.f32 %f1579, %f1578; mov.f32 %f1580, %f1578; @%p20 bra $L__BB0_13; cvt.rn.f32.u16 %f485, %rs24; div.rn.f32 %f486, %f485, 0f437F0000; fma.rn.f32 %f487, %f486, 0f40000000, 0fBF800000; and.b16 %rs42, %rs179, 255; cvt.rn.f32.u16 %f488, %rs42; div.rn.f32 %f489, %f488, 0f437F0000; fma.rn.f32 %f490, %f489, 0f40000000, 0fBF800000; and.b16 %rs43, %rs178, 255; cvt.rn.f32.u16 %f491, %rs43; div.rn.f32 %f492, %f491, 0f437F0000; fma.rn.f32 %f493, %f492, 0f40000000, 0fBF800000; mul.f32 %f494, %f490, %f490; fma.rn.f32 %f495, %f487, %f487, %f494; fma.rn.f32 %f496, %f493, %f493, %f495; sqrt.rn.f32 %f497, %f496; rcp.rn.f32 %f498, %f497; mul.f32 %f1580, %f498, %f493; mul.f32 %f1579, %f498, %f490; mul.f32 %f1578, %f487, %f498; $L__BB0_13: mul.f32 %f512, %f1570, %f1580; mul.f32 %f513, %f1571, %f1579; sub.f32 %f514, %f512, %f513; mul.f32 %f515, %f1571, %f1578; mul.f32 %f516, %f1569, %f1580; sub.f32 %f517, %f515, %f516; mul.f32 %f518, %f1569, %f1579; mul.f32 %f519, %f1570, %f1578; sub.f32 %f520, %f518, %f519; mul.f32 %f521, %f514, %f24; mul.f32 %f522, %f517, %f24; mul.f32 %f523, %f520, %f24; mul.f32 %f524, %f1578, 0f3ED105EC; mul.f32 %f525, %f1579, 0f3ED105EC; mul.f32 %f526, %f1580, 0f3ED105EC; mul.f32 %f527, %f521, 0fBF3504F3; mul.f32 %f528, %f522, 0fBF3504F3; mul.f32 %f529, %f523, 0fBF3504F3; sub.f32 %f530, %f527, %f524; sub.f32 %f531, %f528, %f525; sub.f32 %f532, %f529, %f526; add.f32 %f46, %f25, %f530; add.f32 %f47, %f27, %f531; add.f32 %f48, %f29, %f532; ld.const.u64 %rd43, [params+128]; cvta.to.global.u64 %rd44, %rd43; ld.const.u32 %r111, [params+120]; mad.lo.s32 %r112, %r111, %r95, %r94; mul.wide.u32 %rd45, %r112, 4; add.s64 %rd46, %rd44, %rd45; ld.global.u32 %r490, [%rd46]; setp.lt.s32 %p23, %r4, 1; mov.pred %p22, 0; mov.u32 %r486, 0; mov.f32 %f1597, 0f00000000; mov.f32 %f1598, %f1597; mov.f32 %f1599, %f1597; mov.f32 %f1600, %f1597; mov.f32 %f1601, %f1597; mov.f32 %f1602, %f1597; mov.f32 %f1603, %f1597; mov.f32 %f1604, %f1597; mov.f32 %f1605, %f1597; mov.f32 %f1606, %f1597; mov.f32 %f1607, %f1597; mov.f32 %f1608, %f1597; mov.f32 %f1609, %f1597; mov.pred %p206, %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 %r115, [params+632]; and.b32 %r116, %r95, 255; and.b32 %r117, %r94, 255; mad.lo.s32 %r118, %r115, %r116, %r117; mul.wide.u32 %rd50, %r118, 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 %f673, %f52; abs.f32 %f674, %f51; max.f32 %f675, %f674, %f673; abs.f32 %f676, %f53; max.f32 %f677, %f675, %f676; mov.u32 %r483, %r486; $L__BB0_15: shl.b32 %r119, %r483, 1; mul.wide.s32 %rd51, %r119, 12; add.s64 %rd7, %rd4, %rd51; ld.global.f32 %f546, [%rd7]; sub.f32 %f547, %f546, %f10; ld.global.f32 %f548, [%rd7+4]; sub.f32 %f549, %f548, %f11; ld.global.f32 %f550, [%rd7+8]; sub.f32 %f551, %f550, %f12; mul.f32 %f552, %f549, %f549; fma.rn.f32 %f553, %f547, %f547, %f552; fma.rn.f32 %f554, %f551, %f551, %f553; sqrt.rn.f32 %f67, %f554; rcp.rn.f32 %f555, %f67; mul.f32 %f68, %f547, %f555; mul.f32 %f69, %f549, %f555; mul.f32 %f70, %f551, %f555; mul.f32 %f556, %f1570, %f69; fma.rn.f32 %f557, %f1569, %f68, %f556; fma.rn.f32 %f71, %f1571, %f70, %f557; setp.leu.f32 %p24, %f71, 0f00000000; @%p24 bra $L__BB0_32; setp.ne.s32 %p26, %r5, 0; mul.f32 %f558, %f67, %f67; div.rn.f32 %f72, %f49, %f558; ld.global.u8 %rs44, [%rd5]; cvt.rn.f32.u16 %f559, %rs44; div.rn.f32 %f560, %f559, 0fC37F0000; fma.rn.f32 %f561, %f560, 0f3F333333, 0f3F800000; setp.gt.f32 %p27, %f72, %f561; and.pred %p28, %p26, %p27; mov.pred %p206, -1; @%p28 bra $L__BB0_33; mul.f32 %f73, %f67, %f50; mov.f32 %f566, 0f40800000; abs.f32 %f75, %f73; setp.lt.f32 %p29, %f75, 0f00800000; mul.f32 %f568, %f75, 0f4B800000; selp.f32 %f569, %f568, %f75, %p29; selp.f32 %f570, 0fC3170000, 0fC2FE0000, %p29; mov.b32 %r120, %f569; and.b32 %r121, %r120, 8388607; or.b32 %r122, %r121, 1065353216; mov.b32 %f571, %r122; shr.u32 %r123, %r120, 23; cvt.rn.f32.u32 %f572, %r123; add.f32 %f573, %f570, %f572; setp.gt.f32 %p30, %f571, 0f3FB504F3; mul.f32 %f574, %f571, 0f3F000000; add.f32 %f575, %f573, 0f3F800000; selp.f32 %f576, %f575, %f573, %p30; selp.f32 %f577, %f574, %f571, %p30; add.f32 %f578, %f577, 0fBF800000; add.f32 %f579, %f577, 0f3F800000; rcp.approx.ftz.f32 %f580, %f579; add.f32 %f581, %f578, %f578; mul.f32 %f582, %f581, %f580; mul.f32 %f583, %f582, %f582; mov.f32 %f584, 0f3C4CAF63; mov.f32 %f585, 0f3B18F0FE; fma.rn.f32 %f586, %f585, %f583, %f584; mov.f32 %f587, 0f3DAAAABD; fma.rn.f32 %f588, %f586, %f583, %f587; mul.rn.f32 %f589, %f588, %f583; mul.rn.f32 %f590, %f589, %f582; sub.f32 %f591, %f578, %f582; add.f32 %f592, %f591, %f591; neg.f32 %f593, %f582; fma.rn.f32 %f594, %f593, %f578, %f592; mul.rn.f32 %f595, %f580, %f594; add.f32 %f596, %f590, %f582; sub.f32 %f597, %f582, %f596; add.f32 %f598, %f590, %f597; add.f32 %f599, %f595, %f598; add.f32 %f600, %f596, %f599; sub.f32 %f601, %f596, %f600; add.f32 %f602, %f599, %f601; mov.f32 %f603, 0f3F317200; mul.rn.f32 %f604, %f576, %f603; mov.f32 %f605, 0f35BFBE8E; mul.rn.f32 %f606, %f576, %f605; add.f32 %f607, %f604, %f600; sub.f32 %f608, %f604, %f607; add.f32 %f609, %f600, %f608; add.f32 %f610, %f602, %f609; add.f32 %f611, %f606, %f610; add.f32 %f612, %f607, %f611; sub.f32 %f613, %f607, %f612; add.f32 %f614, %f611, %f613; mul.rn.f32 %f615, %f566, %f612; neg.f32 %f616, %f615; fma.rn.f32 %f617, %f566, %f612, %f616; fma.rn.f32 %f618, %f566, %f614, %f617; mov.f32 %f619, 0f00000000; fma.rn.f32 %f620, %f619, %f612, %f618; add.rn.f32 %f621, %f615, %f620; neg.f32 %f622, %f621; add.rn.f32 %f623, %f615, %f622; add.rn.f32 %f624, %f623, %f620; mov.b32 %r124, %f621; setp.eq.s32 %p31, %r124, 1118925336; add.s32 %r125, %r124, -1; mov.b32 %f625, %r125; add.f32 %f626, %f624, 0f37000000; selp.f32 %f76, %f626, %f624, %p31; selp.f32 %f627, %f625, %f621, %p31; mov.f32 %f628, 0f3FB8AA3B; mul.rn.f32 %f629, %f627, %f628; cvt.rzi.f32.f32 %f630, %f629; abs.f32 %f631, %f630; setp.gt.f32 %p32, %f631, 0f42FC0000; mov.b32 %r126, %f630; and.b32 %r127, %r126, -2147483648; or.b32 %r128, %r127, 1123811328; mov.b32 %f632, %r128; selp.f32 %f633, %f632, %f630, %p32; mov.f32 %f634, 0fBF317218; fma.rn.f32 %f635, %f633, %f634, %f627; mov.f32 %f636, 0f3102E308; fma.rn.f32 %f637, %f633, %f636, %f635; mul.f32 %f638, %f637, 0f3FB8AA3B; add.f32 %f639, %f633, 0f4B40007F; mov.b32 %r129, %f639; shl.b32 %r130, %r129, 23; mov.b32 %f640, %r130; ex2.approx.ftz.f32 %f641, %f638; mul.f32 %f77, %f641, %f640; setp.eq.f32 %p33, %f77, 0f7F800000; mov.f32 %f1594, 0f7F800000; @%p33 bra $L__BB0_19; fma.rn.f32 %f1594, %f77, %f76, %f77; $L__BB0_19: mov.f32 %f1566, 0f40000000; cvt.rzi.f32.f32 %f1565, %f1566; add.f32 %f1564, %f1565, %f1565; mov.f32 %f1563, 0f40800000; sub.f32 %f1562, %f1563, %f1564; abs.f32 %f1561, %f1562; setp.lt.f32 %p34, %f73, 0f00000000; setp.eq.f32 %p35, %f1561, 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 %f646, %f73, %f73; selp.f32 %f1596, %f646, 0f00000000, %p35; bra.uni $L__BB0_24; $L__BB0_20: mov.b32 %r131, %f1594; xor.b32 %r132, %r131, -2147483648; mov.b32 %f642, %r132; selp.f32 %f1596, %f642, %f1594, %p1; setp.geu.f32 %p37, %f73, 0f00000000; @%p37 bra $L__BB0_24; mov.f32 %f643, 0f40800000; cvt.rzi.f32.f32 %f644, %f643; setp.eq.f32 %p38, %f644, 0f40800000; @%p38 bra $L__BB0_24; mov.f32 %f1596, 0f7FFFFFFF; $L__BB0_24: add.f32 %f647, %f75, 0f40800000; mov.b32 %r133, %f647; setp.lt.s32 %p40, %r133, 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 %f1596, %f73, 0f40800000; bra.uni $L__BB0_29; $L__BB0_26: setp.neu.f32 %p42, %f75, 0f7F800000; @%p42 bra $L__BB0_29; selp.f32 %f1596, 0fFF800000, 0f7F800000, %p1; $L__BB0_29: shl.b32 %r481, %r483, 1; mul.wide.s32 %rd117, %r481, 12; add.s64 %rd116, %rd4, %rd117; mov.f32 %f648, 0f3F800000; sub.f32 %f649, %f648, %f1596; setp.eq.f32 %p43, %f73, 0f3F800000; selp.f32 %f650, 0f00000000, %f649, %p43; cvt.sat.f32.f32 %f651, %f650; mul.f32 %f652, %f72, %f651; ld.global.f32 %f653, [%rd116+12]; mul.f32 %f654, %f68, %f653; ld.global.f32 %f655, [%rd116+16]; mul.f32 %f656, %f69, %f655; neg.f32 %f657, %f656; sub.f32 %f658, %f657, %f654; ld.global.f32 %f659, [%rd116+20]; mul.f32 %f660, %f70, %f659; sub.f32 %f661, %f658, %f660; cvt.sat.f32.f32 %f662, %f661; mul.f32 %f86, %f652, %f662; setp.leu.f32 %p44, %f86, 0f3727C5AC; @%p44 bra $L__BB0_31; cvt.sat.f32.f32 %f672, %f71; mov.f32 %f678, 0f38D1B717; max.f32 %f669, %f677, %f678; sub.f32 %f670, %f67, %f669; mov.f32 %f671, 0f00000000; mov.u32 %r170, 2; mov.u32 %r171, 1; mov.u32 %r172, 3; mov.u32 %r175, 1065353216; mov.u32 %r204, 0; // begin inline asm call(%r134,%r135,%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),_optix_trace_typed_32,(%r204,%rd6,%f15,%f16,%f17,%f68,%f69,%f70,%f669,%f670,%f671,%r171,%r204,%r171,%r170,%r171,%r172,%r175,%r175,%r175,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204,%r204); // end inline asm mov.b32 %f679, %r134; mov.b32 %f680, %r135; mov.b32 %f681, %r136; mul.f32 %f682, %f86, 0f3EA2F983; mul.f32 %f683, %f682, %f679; mul.f32 %f684, %f682, %f680; mul.f32 %f685, %f682, %f681; fma.rn.f32 %f1606, %f672, %f683, %f1606; fma.rn.f32 %f1607, %f672, %f684, %f1607; fma.rn.f32 %f1608, %f672, %f685, %f1608; mul.f32 %f686, %f28, %f69; fma.rn.f32 %f687, %f26, %f68, %f686; fma.rn.f32 %f688, %f30, %f70, %f687; cvt.sat.f32.f32 %f689, %f688; fma.rn.f32 %f1603, %f689, %f683, %f1603; fma.rn.f32 %f1604, %f689, %f684, %f1604; fma.rn.f32 %f1605, %f689, %f685, %f1605; mul.f32 %f690, %f38, %f69; fma.rn.f32 %f691, %f37, %f68, %f690; fma.rn.f32 %f692, %f39, %f70, %f691; cvt.sat.f32.f32 %f693, %f692; fma.rn.f32 %f1600, %f693, %f683, %f1600; fma.rn.f32 %f1601, %f693, %f684, %f1601; fma.rn.f32 %f1602, %f693, %f685, %f1602; mul.f32 %f694, %f47, %f69; fma.rn.f32 %f695, %f46, %f68, %f694; fma.rn.f32 %f696, %f48, %f70, %f695; cvt.sat.f32.f32 %f697, %f696; fma.rn.f32 %f1597, %f683, %f697, %f1597; fma.rn.f32 %f1598, %f684, %f697, %f1598; fma.rn.f32 %f1599, %f685, %f697, %f1599; add.f32 %f1609, %f1609, %f679; $L__BB0_31: add.s32 %r486, %r486, 1; $L__BB0_32: add.s32 %r483, %r483, 1; setp.lt.s32 %p46, %r483, %r4; mov.pred %p206, %p22; @%p46 bra $L__BB0_15; $L__BB0_33: cvt.rn.f32.s32 %f698, %r486; mov.f32 %f699, 0f3F800000; max.f32 %f700, %f698, %f699; rcp.rn.f32 %f701, %f700; mul.f32 %f1693, %f1606, %f701; mul.f32 %f1694, %f1607, %f701; mul.f32 %f1695, %f1608, %f701; div.rn.f32 %f1696, %f1609, %f700; mul.f32 %f1690, %f1603, %f701; mul.f32 %f1691, %f1604, %f701; mul.f32 %f1692, %f1605, %f701; mul.f32 %f1687, %f1600, %f701; mul.f32 %f1688, %f1601, %f701; mul.f32 %f1689, %f1602, %f701; mul.f32 %f1684, %f1597, %f701; mul.f32 %f1685, %f1598, %f701; mul.f32 %f1686, %f1599, %f701; not.pred %p47, %p206; @%p47 bra $L__BB0_66; abs.f32 %f1568, %f1571; abs.f32 %f1567, %f1569; setp.gt.f32 %p48, %f1567, %f1568; selp.f32 %f715, 0f00000000, %f1570, %p48; mov.f32 %f1661, 0f00000000; neg.f32 %f716, %f1571; selp.f32 %f717, %f1569, %f716, %p48; neg.f32 %f718, %f1570; selp.f32 %f719, %f718, 0f00000000, %p48; mul.f32 %f720, %f717, %f717; fma.rn.f32 %f721, %f719, %f719, %f720; fma.rn.f32 %f722, %f715, %f715, %f721; sqrt.rn.f32 %f723, %f722; rcp.rn.f32 %f724, %f723; mul.f32 %f152, %f719, %f724; mul.f32 %f153, %f717, %f724; mul.f32 %f154, %f715, %f724; setp.lt.s32 %p49, %r5, 1; mov.f32 %f1660, %f1661; mov.f32 %f1659, %f1661; mov.f32 %f1658, %f1661; mov.f32 %f1657, %f1661; mov.f32 %f1656, %f1661; mov.f32 %f1655, %f1661; mov.f32 %f1654, %f1661; mov.f32 %f1653, %f1661; mov.f32 %f1652, %f1661; mov.f32 %f1651, %f1661; mov.f32 %f1650, %f1661; mov.f32 %f1649, %f1661; @%p49 bra $L__BB0_65; cvt.rn.f32.s32 %f738, %r5; rcp.rn.f32 %f155, %f738; mul.f32 %f156, %f15, 0f3456BF95; mul.f32 %f157, %f16, 0f3456BF95; mul.f32 %f158, %f17, 0f3456BF95; ld.const.u64 %rd8, [params+96]; mul.f32 %f739, %f1569, %f153; mul.f32 %f740, %f1570, %f152; sub.f32 %f159, %f740, %f739; mul.f32 %f741, %f1571, %f152; mul.f32 %f742, %f1569, %f154; sub.f32 %f160, %f742, %f741; mul.f32 %f743, %f1570, %f154; mul.f32 %f744, %f1571, %f153; sub.f32 %f161, %f744, %f743; mov.u32 %r205, 0; add.s64 %rd9, %rd1, 24; mov.u64 %rd53, __cudart_i2opi_f; abs.f32 %f810, %f157; abs.f32 %f811, %f156; max.f32 %f812, %f811, %f810; abs.f32 %f813, %f158; max.f32 %f814, %f812, %f813; mov.u32 %r487, %r205; $L__BB0_36: cvt.rn.f32.s32 %f175, %r487; mov.u32 %r489, %r205; $L__BB0_37: mad.lo.s32 %r207, %r490, 1664525, 1013904223; and.b32 %r208, %r207, 16777215; cvt.rn.f32.u32 %f745, %r208; fma.rn.f32 %f746, %f745, 0f33800000, %f175; mul.f32 %f189, %f155, %f746; mad.lo.s32 %r490, %r207, 1664525, 1013904223; and.b32 %r209, %r490, 16777215; cvt.rn.f32.u32 %f747, %r209; cvt.rn.f32.s32 %f748, %r489; fma.rn.f32 %f749, %f747, 0f33800000, %f748; mul.f32 %f750, %f155, %f749; mul.f32 %f751, %f189, %f189; mov.f32 %f752, 0f3F800000; sub.f32 %f753, %f752, %f751; mov.f32 %f754, 0f00000000; max.f32 %f755, %f754, %f753; sqrt.rn.f32 %f190, %f755; mul.f32 %f191, %f750, 0f40C90FDB; mul.f32 %f756, %f191, 0f3F22F983; cvt.rni.s32.f32 %r497, %f756; cvt.rn.f32.s32 %f757, %r497; mov.f32 %f758, 0fBFC90FDA; fma.rn.f32 %f759, %f757, %f758, %f191; mov.f32 %f760, 0fB3A22168; fma.rn.f32 %f761, %f757, %f760, %f759; mov.f32 %f762, 0fA7C234C5; fma.rn.f32 %f1665, %f757, %f762, %f761; abs.f32 %f193, %f191; setp.ltu.f32 %p50, %f193, 0f47CE4780; mov.u32 %r494, %r497; mov.f32 %f1662, %f1665; @%p50 bra $L__BB0_45; setp.eq.f32 %p51, %f193, 0f7F800000; @%p51 bra $L__BB0_44; bra.uni $L__BB0_39; $L__BB0_44: mov.f32 %f765, 0f00000000; mul.rn.f32 %f1662, %f191, %f765; mov.u32 %r494, 0; bra.uni $L__BB0_45; $L__BB0_39: mov.b32 %r21, %f191; bfe.u32 %r211, %r21, 23, 8; add.s32 %r22, %r211, -128; shl.b32 %r212, %r21, 8; or.b32 %r23, %r212, -2147483648; shr.u32 %r24, %r22, 5; mov.u64 %rd120, 0; mov.u32 %r491, 0; mov.u64 %rd118, %rd1; mov.u64 %rd119, %rd53; $L__BB0_40: .pragma "nounroll"; ld.global.nc.u32 %r213, [%rd119]; mad.wide.u32 %rd55, %r213, %r23, %rd120; shr.u64 %rd120, %rd55, 32; st.local.u32 [%rd118], %rd55; add.s64 %rd119, %rd119, 4; add.s64 %rd118, %rd118, 4; add.s32 %r491, %r491, 1; setp.ne.s32 %p52, %r491, 6; @%p52 bra $L__BB0_40; st.local.u32 [%rd9], %rd120; mov.u32 %r214, 4; sub.s32 %r27, %r214, %r24; mov.u32 %r215, 6; sub.s32 %r216, %r215, %r24; mul.wide.s32 %rd56, %r216, 4; add.s64 %rd57, %rd1, %rd56; ld.local.u32 %r492, [%rd57]; ld.local.u32 %r493, [%rd57+-4]; and.b32 %r30, %r22, 31; setp.eq.s32 %p53, %r30, 0; @%p53 bra $L__BB0_43; mov.u32 %r217, 32; sub.s32 %r218, %r217, %r30; shr.u32 %r219, %r493, %r218; shl.b32 %r220, %r492, %r30; add.s32 %r492, %r219, %r220; mul.wide.s32 %rd58, %r27, 4; add.s64 %rd59, %rd1, %rd58; ld.local.u32 %r221, [%rd59]; shr.u32 %r222, %r221, %r218; shl.b32 %r223, %r493, %r30; add.s32 %r493, %r222, %r223; $L__BB0_43: and.b32 %r224, %r21, -2147483648; shr.u32 %r225, %r493, 30; shl.b32 %r226, %r492, 2; or.b32 %r227, %r225, %r226; shr.u32 %r228, %r227, 31; shr.u32 %r229, %r492, 30; add.s32 %r230, %r228, %r229; neg.s32 %r231, %r230; setp.eq.s32 %p54, %r224, 0; selp.b32 %r494, %r230, %r231, %p54; setp.ne.s32 %p55, %r228, 0; xor.b32 %r232, %r224, -2147483648; selp.b32 %r233, %r232, %r224, %p55; selp.b32 %r234, -1, 0, %p55; xor.b32 %r235, %r227, %r234; shl.b32 %r236, %r493, 2; xor.b32 %r237, %r236, %r234; cvt.u64.u32 %rd60, %r235; cvt.u64.u32 %rd61, %r237; bfi.b64 %rd62, %rd60, %rd61, 32, 32; cvt.rn.f64.s64 %fd1, %rd62; mul.f64 %fd2, %fd1, 0d3BF921FB54442D19; cvt.rn.f32.f64 %f763, %fd2; setp.eq.s32 %p56, %r233, 0; neg.f32 %f764, %f763; selp.f32 %f1662, %f763, %f764, %p56; $L__BB0_45: add.s32 %r37, %r494, 1; and.b32 %r38, %r37, 1; setp.eq.s32 %p57, %r38, 0; selp.f32 %f197, %f1662, 0f3F800000, %p57; mul.rn.f32 %f198, %f1662, %f1662; mov.f32 %f1663, 0fB94D4153; @%p57 bra $L__BB0_47; mov.f32 %f767, 0fBAB607ED; mov.f32 %f768, 0f37CBAC00; fma.rn.f32 %f1663, %f768, %f198, %f767; $L__BB0_47: selp.f32 %f769, 0f3C0885E4, 0f3D2AAABB, %p57; fma.rn.f32 %f770, %f1663, %f198, %f769; selp.f32 %f771, 0fBE2AAAA8, 0fBEFFFFFF, %p57; fma.rn.f32 %f772, %f770, %f198, %f771; mov.f32 %f773, 0f00000000; fma.rn.f32 %f774, %f198, %f197, %f773; fma.rn.f32 %f1664, %f772, %f774, %f197; and.b32 %r239, %r37, 2; setp.eq.s32 %p59, %r239, 0; @%p59 bra $L__BB0_49; mov.f32 %f776, 0fBF800000; fma.rn.f32 %f1664, %f1664, %f776, %f773; $L__BB0_49: @%p50 bra $L__BB0_57; setp.eq.f32 %p61, %f193, 0f7F800000; @%p61 bra $L__BB0_56; bra.uni $L__BB0_51; $L__BB0_56: mov.f32 %f779, 0f00000000; mul.rn.f32 %f1665, %f191, %f779; mov.u32 %r497, 0; bra.uni $L__BB0_57; $L__BB0_51: mov.b32 %r39, %f191; bfe.u32 %r240, %r39, 23, 8; add.s32 %r40, %r240, -128; shl.b32 %r241, %r39, 8; or.b32 %r41, %r241, -2147483648; shr.u32 %r42, %r40, 5; mov.u64 %rd121, 0; mov.u64 %rd122, %rd121; $L__BB0_52: .pragma "nounroll"; shl.b64 %rd65, %rd121, 2; mov.u64 %rd66, __cudart_i2opi_f; add.s64 %rd67, %rd66, %rd65; ld.global.nc.u32 %r242, [%rd67]; mad.wide.u32 %rd68, %r242, %r41, %rd122; shr.u64 %rd122, %rd68, 32; add.s64 %rd69, %rd1, %rd65; st.local.u32 [%rd69], %rd68; cvt.u32.u64 %r243, %rd121; add.s32 %r244, %r243, 1; cvt.s64.s32 %rd121, %r244; setp.ne.s32 %p62, %r244, 6; @%p62 bra $L__BB0_52; st.local.u32 [%rd9], %rd122; mov.u32 %r245, 4; sub.s32 %r43, %r245, %r42; mov.u32 %r246, 6; sub.s32 %r247, %r246, %r42; mul.wide.s32 %rd70, %r247, 4; add.s64 %rd71, %rd1, %rd70; ld.local.u32 %r495, [%rd71]; ld.local.u32 %r496, [%rd71+-4]; and.b32 %r46, %r40, 31; setp.eq.s32 %p63, %r46, 0; @%p63 bra $L__BB0_55; mov.u32 %r248, 32; sub.s32 %r249, %r248, %r46; shr.u32 %r250, %r496, %r249; shl.b32 %r251, %r495, %r46; add.s32 %r495, %r250, %r251; mul.wide.s32 %rd72, %r43, 4; add.s64 %rd73, %rd1, %rd72; ld.local.u32 %r252, [%rd73]; shr.u32 %r253, %r252, %r249; shl.b32 %r254, %r496, %r46; add.s32 %r496, %r253, %r254; $L__BB0_55: and.b32 %r255, %r39, -2147483648; shr.u32 %r256, %r496, 30; shl.b32 %r257, %r495, 2; or.b32 %r258, %r256, %r257; shr.u32 %r259, %r258, 31; shr.u32 %r260, %r495, 30; add.s32 %r261, %r259, %r260; neg.s32 %r262, %r261; setp.eq.s32 %p64, %r255, 0; selp.b32 %r497, %r261, %r262, %p64; setp.ne.s32 %p65, %r259, 0; xor.b32 %r263, %r255, -2147483648; selp.b32 %r264, %r263, %r255, %p65; selp.b32 %r265, -1, 0, %p65; xor.b32 %r266, %r258, %r265; shl.b32 %r267, %r496, 2; xor.b32 %r268, %r267, %r265; cvt.u64.u32 %rd74, %r266; cvt.u64.u32 %rd75, %r268; bfi.b64 %rd76, %rd74, %rd75, 32, 32; cvt.rn.f64.s64 %fd3, %rd76; mul.f64 %fd4, %fd3, 0d3BF921FB54442D19; cvt.rn.f32.f64 %f777, %fd4; setp.eq.s32 %p66, %r264, 0; neg.f32 %f778, %f777; selp.f32 %f1665, %f777, %f778, %p66; $L__BB0_57: mul.f32 %f207, %f190, %f1664; and.b32 %r53, %r497, 1; setp.eq.s32 %p67, %r53, 0; selp.f32 %f208, %f1665, 0f3F800000, %p67; mul.rn.f32 %f209, %f1665, %f1665; mov.f32 %f1666, 0fB94D4153; @%p67 bra $L__BB0_59; mov.f32 %f781, 0fBAB607ED; mov.f32 %f782, 0f37CBAC00; fma.rn.f32 %f1666, %f782, %f209, %f781; $L__BB0_59: selp.f32 %f783, 0f3C0885E4, 0f3D2AAABB, %p67; fma.rn.f32 %f784, %f1666, %f209, %f783; selp.f32 %f785, 0fBE2AAAA8, 0fBEFFFFFF, %p67; fma.rn.f32 %f786, %f784, %f209, %f785; mov.f32 %f1668, 0f00000000; fma.rn.f32 %f788, %f209, %f208, %f1668; fma.rn.f32 %f1667, %f786, %f788, %f208; and.b32 %r270, %r497, 2; setp.eq.s32 %p69, %r270, 0; @%p69 bra $L__BB0_61; mov.f32 %f789, 0f00000000; mov.f32 %f790, 0fBF800000; fma.rn.f32 %f1667, %f1667, %f790, %f789; $L__BB0_61: mul.f32 %f803, %f190, %f1667; mul.f32 %f804, %f152, %f803; mul.f32 %f805, %f153, %f803; mul.f32 %f806, %f154, %f803; fma.rn.f32 %f807, %f161, %f207, %f804; fma.rn.f32 %f808, %f160, %f207, %f805; fma.rn.f32 %f809, %f159, %f207, %f806; fma.rn.f32 %f215, %f1569, %f189, %f807; fma.rn.f32 %f216, %f1570, %f189, %f808; fma.rn.f32 %f217, %f1571, %f189, %f809; mov.f32 %f815, 0f38D1B717; max.f32 %f797, %f814, %f815; mov.f32 %f798, 0f6C4ECB8F; mov.u32 %r304, 1; mov.u32 %r307, 2; mov.u32 %r309, 4; mov.u32 %r312, 1065353216; mov.u32 %r313, 2139095039; mov.u32 %r341, 0; // begin inline asm call(%r271,%r272,%r273,%r274,%r275,%r276,%r277,%r278,%r279,%r280,%r281,%r282,%r283,%r284,%r285,%r286,%r287,%r288,%r289,%r290,%r291,%r292,%r293,%r294,%r295,%r296,%r297,%r298,%r299,%r300,%r301,%r302),_optix_trace_typed_32,(%r341,%rd8,%f15,%f16,%f17,%f215,%f216,%f217,%f797,%f798,%f1668,%r304,%r341,%r341,%r307,%r341,%r309,%r312,%r312,%r312,%r313,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341,%r341); // end inline asm mov.b32 %f218, %r274; setp.ge.f32 %p70, %f218, 0f00000000; mov.f32 %f1669, %f1668; mov.f32 %f1670, %f1668; @%p70 bra $L__BB0_63; mov.b32 %f1670, %r271; mov.b32 %f1669, %r272; mov.b32 %f1668, %r273; $L__BB0_63: setp.ltu.f32 %p71, %f218, 0f00000000; selp.f32 %f816, 0f3F800000, 0f00000000, %p71; add.f32 %f1649, %f1649, %f816; mul.f32 %f817, %f28, %f216; fma.rn.f32 %f818, %f26, %f215, %f817; fma.rn.f32 %f819, %f30, %f217, %f818; cvt.sat.f32.f32 %f820, %f819; fma.rn.f32 %f1655, %f1670, %f820, %f1655; fma.rn.f32 %f1654, %f1669, %f820, %f1654; fma.rn.f32 %f1653, %f1668, %f820, %f1653; mul.f32 %f821, %f38, %f216; fma.rn.f32 %f822, %f37, %f215, %f821; fma.rn.f32 %f823, %f39, %f217, %f822; cvt.sat.f32.f32 %f824, %f823; fma.rn.f32 %f1658, %f1670, %f824, %f1658; fma.rn.f32 %f1657, %f1669, %f824, %f1657; fma.rn.f32 %f1656, %f1668, %f824, %f1656; mul.f32 %f825, %f47, %f216; fma.rn.f32 %f826, %f46, %f215, %f825; fma.rn.f32 %f827, %f48, %f217, %f826; cvt.sat.f32.f32 %f828, %f827; fma.rn.f32 %f1661, %f1670, %f828, %f1661; fma.rn.f32 %f1660, %f1669, %f828, %f1660; fma.rn.f32 %f1659, %f1668, %f828, %f1659; mul.f32 %f829, %f1570, %f216; fma.rn.f32 %f830, %f1569, %f215, %f829; fma.rn.f32 %f831, %f1571, %f217, %f830; cvt.sat.f32.f32 %f832, %f831; fma.rn.f32 %f1652, %f1670, %f832, %f1652; fma.rn.f32 %f1651, %f1669, %f832, %f1651; fma.rn.f32 %f1650, %f1668, %f832, %f1650; add.s32 %r489, %r489, 1; setp.lt.s32 %p72, %r489, %r5; @%p72 bra $L__BB0_37; add.s32 %r487, %r487, 1; setp.lt.s32 %p73, %r487, %r5; @%p73 bra $L__BB0_36; $L__BB0_65: mul.lo.s32 %r342, %r5, %r5; cvt.rn.f32.s32 %f833, %r342; rcp.rn.f32 %f834, %f833; mul.f32 %f835, %f834, %f1652; mul.f32 %f836, %f834, %f1651; mul.f32 %f837, %f834, %f1650; div.rn.f32 %f1696, %f1649, %f833; mul.f32 %f1690, %f834, %f1655; mul.f32 %f1691, %f834, %f1654; mul.f32 %f1692, %f834, %f1653; mul.f32 %f1687, %f834, %f1658; mul.f32 %f1688, %f834, %f1657; mul.f32 %f1689, %f834, %f1656; mul.f32 %f1684, %f834, %f1661; mul.f32 %f1685, %f834, %f1660; mul.f32 %f1686, %f834, %f1659; fma.rn.f32 %f1693, %f834, %f1652, %f835; fma.rn.f32 %f1694, %f834, %f1651, %f836; fma.rn.f32 %f1695, %f834, %f1650, %f837; $L__BB0_66: ld.const.u32 %r482, [params+616]; setp.lt.s32 %p74, %r482, 0; selp.f32 %f277, %f1693, %f1696, %p74; ld.const.u32 %r88, [params+104]; and.b32 %r343, %r88, 8; setp.eq.s32 %p75, %r343, 0; @%p75 bra $L__BB0_80; ld.const.u64 %rd78, [params+192]; cvta.to.global.u64 %rd20, %rd78; ld.const.u32 %r344, [params+184]; mad.lo.s32 %r345, %r344, %r7, %r6; cvt.u64.u32 %rd21, %r345; mov.f32 %f839, 0f3E68BA2E; cvt.rzi.f32.f32 %f840, %f839; add.f32 %f841, %f840, %f840; mov.f32 %f842, 0f3EE8BA2E; sub.f32 %f843, %f842, %f841; abs.f32 %f278, %f843; abs.f32 %f279, %f277; setp.lt.f32 %p76, %f279, 0f00800000; mul.f32 %f844, %f279, 0f4B800000; selp.f32 %f845, %f844, %f279, %p76; selp.f32 %f846, 0fC3170000, 0fC2FE0000, %p76; mov.b32 %r346, %f845; and.b32 %r347, %r346, 8388607; or.b32 %r348, %r347, 1065353216; mov.b32 %f847, %r348; shr.u32 %r349, %r346, 23; cvt.rn.f32.u32 %f848, %r349; add.f32 %f849, %f846, %f848; setp.gt.f32 %p77, %f847, 0f3FB504F3; mul.f32 %f850, %f847, 0f3F000000; add.f32 %f851, %f849, 0f3F800000; selp.f32 %f852, %f851, %f849, %p77; selp.f32 %f853, %f850, %f847, %p77; add.f32 %f854, %f853, 0fBF800000; add.f32 %f855, %f853, 0f3F800000; rcp.approx.ftz.f32 %f856, %f855; add.f32 %f857, %f854, %f854; mul.f32 %f858, %f857, %f856; mul.f32 %f859, %f858, %f858; mov.f32 %f860, 0f3C4CAF63; mov.f32 %f861, 0f3B18F0FE; fma.rn.f32 %f862, %f861, %f859, %f860; mov.f32 %f863, 0f3DAAAABD; fma.rn.f32 %f864, %f862, %f859, %f863; mul.rn.f32 %f865, %f864, %f859; mul.rn.f32 %f866, %f865, %f858; sub.f32 %f867, %f854, %f858; add.f32 %f868, %f867, %f867; neg.f32 %f869, %f858; fma.rn.f32 %f870, %f869, %f854, %f868; mul.rn.f32 %f871, %f856, %f870; add.f32 %f872, %f866, %f858; sub.f32 %f873, %f858, %f872; add.f32 %f874, %f866, %f873; add.f32 %f875, %f871, %f874; add.f32 %f876, %f872, %f875; sub.f32 %f877, %f872, %f876; add.f32 %f878, %f875, %f877; mov.f32 %f879, 0f3F317200; mul.rn.f32 %f880, %f852, %f879; mov.f32 %f881, 0f35BFBE8E; mul.rn.f32 %f882, %f852, %f881; add.f32 %f883, %f880, %f876; sub.f32 %f884, %f880, %f883; add.f32 %f885, %f876, %f884; add.f32 %f886, %f878, %f885; add.f32 %f887, %f882, %f886; add.f32 %f888, %f883, %f887; sub.f32 %f889, %f883, %f888; add.f32 %f890, %f887, %f889; mul.rn.f32 %f891, %f842, %f888; neg.f32 %f892, %f891; fma.rn.f32 %f893, %f842, %f888, %f892; fma.rn.f32 %f894, %f842, %f890, %f893; mov.f32 %f895, 0f00000000; fma.rn.f32 %f896, %f895, %f888, %f894; add.rn.f32 %f897, %f891, %f896; neg.f32 %f898, %f897; add.rn.f32 %f899, %f891, %f898; add.rn.f32 %f900, %f899, %f896; mov.b32 %r350, %f897; setp.eq.s32 %p78, %r350, 1118925336; add.s32 %r351, %r350, -1; mov.b32 %f901, %r351; add.f32 %f902, %f900, 0f37000000; selp.f32 %f280, %f902, %f900, %p78; selp.f32 %f903, %f901, %f897, %p78; mov.f32 %f904, 0f3FB8AA3B; mul.rn.f32 %f905, %f903, %f904; cvt.rzi.f32.f32 %f906, %f905; abs.f32 %f907, %f906; setp.gt.f32 %p79, %f907, 0f42FC0000; mov.b32 %r352, %f906; and.b32 %r353, %r352, -2147483648; or.b32 %r354, %r353, 1123811328; mov.b32 %f908, %r354; selp.f32 %f909, %f908, %f906, %p79; mov.f32 %f910, 0fBF317218; fma.rn.f32 %f911, %f909, %f910, %f903; mov.f32 %f912, 0f3102E308; fma.rn.f32 %f913, %f909, %f912, %f911; mul.f32 %f914, %f913, 0f3FB8AA3B; add.f32 %f915, %f909, 0f4B40007F; mov.b32 %r355, %f915; shl.b32 %r356, %r355, 23; mov.b32 %f916, %r356; ex2.approx.ftz.f32 %f917, %f914; mul.f32 %f281, %f917, %f916; setp.eq.f32 %p80, %f281, 0f7F800000; mov.f32 %f1697, 0f7F800000; @%p80 bra $L__BB0_69; fma.rn.f32 %f1697, %f281, %f280, %f281; $L__BB0_69: setp.lt.f32 %p81, %f277, 0f00000000; setp.eq.f32 %p82, %f278, 0f3F800000; and.pred %p3, %p81, %p82; setp.eq.f32 %p83, %f277, 0f00000000; @%p83 bra $L__BB0_73; bra.uni $L__BB0_70; $L__BB0_73: add.f32 %f922, %f277, %f277; selp.f32 %f1699, %f922, 0f00000000, %p82; bra.uni $L__BB0_74; $L__BB0_174: mov.f32 %f1533, 0f00000000; mov.u32 %r499, 0; // begin inline asm { cvt.rn.f16.f32 %rs127, %f1533;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs126, %f1533;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs125, %f1533;} // end inline asm mov.u16 %rs128, 0; st.global.v4.u16 [%rd28], {%rs125, %rs126, %rs127, %rs128}; $L__BB0_175: ld.const.u64 %rd107, [params+256]; cvta.to.global.u64 %rd108, %rd107; ld.const.u32 %r475, [params+248]; mad.lo.s32 %r476, %r475, %r7, %r6; mul.wide.u32 %rd109, %r476, 8; add.s64 %rd29, %rd108, %rd109; setp.eq.s32 %p203, %r499, 0; @%p203 bra $L__BB0_177; ld.global.v4.u16 {%rs135, %rs136, %rs137, %rs138}, [%rd29]; // begin inline asm { cvt.f32.f16 %f1534, %rs135;} // end inline asm // begin inline asm { cvt.f32.f16 %f1535, %rs136;} // end inline asm // begin inline asm { cvt.f32.f16 %f1536, %rs137;} // 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 %rs134, %f1539;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs133, %f1538;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs132, %f1537;} // end inline asm mov.u16 %rs139, 0; st.global.v4.u16 [%rd29], {%rs132, %rs133, %rs134, %rs139}; bra.uni $L__BB0_178; $L__BB0_177: mov.f32 %f1542, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs142, %f1542;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs141, %f1542;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs140, %f1542;} // end inline asm mov.u16 %rs143, 0; st.global.v4.u16 [%rd29], {%rs140, %rs141, %rs142, %rs143}; $L__BB0_178: ld.const.u64 %rd110, [params+272]; cvta.to.global.u64 %rd111, %rd110; ld.const.u32 %r477, [params+264]; mad.lo.s32 %r478, %r477, %r7, %r6; mul.wide.u32 %rd112, %r478, 8; add.s64 %rd30, %rd111, %rd112; @%p203 bra $L__BB0_180; ld.global.v4.u16 {%rs150, %rs151, %rs152, %rs153}, [%rd30]; // begin inline asm { cvt.f32.f16 %f1543, %rs150;} // end inline asm // begin inline asm { cvt.f32.f16 %f1544, %rs151;} // end inline asm // begin inline asm { cvt.f32.f16 %f1545, %rs152;} // 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 %rs149, %f1548;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs148, %f1547;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs147, %f1546;} // end inline asm mov.u16 %rs154, 0; st.global.v4.u16 [%rd30], {%rs147, %rs148, %rs149, %rs154}; bra.uni $L__BB0_181; $L__BB0_180: mov.f32 %f1551, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs157, %f1551;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs156, %f1551;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs155, %f1551;} // end inline asm mov.u16 %rs158, 0; st.global.v4.u16 [%rd30], {%rs155, %rs156, %rs157, %rs158}; $L__BB0_181: ld.const.u64 %rd113, [params+288]; cvta.to.global.u64 %rd114, %rd113; ld.const.u32 %r479, [params+280]; mad.lo.s32 %r480, %r479, %r7, %r6; mul.wide.u32 %rd115, %r480, 8; add.s64 %rd31, %rd114, %rd115; @%p203 bra $L__BB0_183; ld.global.v4.u16 {%rs165, %rs166, %rs167, %rs168}, [%rd31]; // begin inline asm { cvt.f32.f16 %f1552, %rs165;} // end inline asm // begin inline asm { cvt.f32.f16 %f1553, %rs166;} // end inline asm // begin inline asm { cvt.f32.f16 %f1554, %rs167;} // end inline asm add.f32 %f1555, %f1552, 0f00000000; add.f32 %f1556, %f1553, 0f00000000; add.f32 %f1557, %f1554, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs164, %f1557;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs163, %f1556;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs162, %f1555;} // end inline asm mov.u16 %rs169, 0; st.global.v4.u16 [%rd31], {%rs162, %rs163, %rs164, %rs169}; bra.uni $L__BB0_184; $L__BB0_183: mov.f32 %f1560, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs172, %f1560;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs171, %f1560;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs170, %f1560;} // end inline asm mov.u16 %rs173, 0; st.global.v4.u16 [%rd31], {%rs170, %rs171, %rs172, %rs173}; bra.uni $L__BB0_184; $L__BB0_70: mov.b32 %r357, %f1697; xor.b32 %r358, %r357, -2147483648; mov.b32 %f918, %r358; selp.f32 %f1699, %f918, %f1697, %p3; setp.geu.f32 %p84, %f277, 0f00000000; @%p84 bra $L__BB0_74; mov.f32 %f919, 0f3EE8BA2E; cvt.rzi.f32.f32 %f920, %f919; setp.eq.f32 %p85, %f920, 0f3EE8BA2E; @%p85 bra $L__BB0_74; mov.f32 %f1699, 0f7FFFFFFF; $L__BB0_74: add.f32 %f923, %f279, 0f3EE8BA2E; mov.b32 %r359, %f923; setp.lt.s32 %p87, %r359, 2139095040; @%p87 bra $L__BB0_79; setp.gtu.f32 %p88, %f279, 0f7F800000; @%p88 bra $L__BB0_78; bra.uni $L__BB0_76; $L__BB0_78: add.f32 %f1699, %f277, 0f3EE8BA2E; bra.uni $L__BB0_79; $L__BB0_76: setp.neu.f32 %p89, %f279, 0f7F800000; @%p89 bra $L__BB0_79; selp.f32 %f1699, 0fFF800000, 0f7F800000, %p3; $L__BB0_79: mul.f32 %f924, %f1699, 0f437F0000; setp.eq.f32 %p90, %f277, 0f3F800000; selp.f32 %f925, 0f437F0000, %f924, %p90; cvt.rzi.u32.f32 %r360, %f925; shl.b64 %rd79, %rd21, 1; add.s64 %rd80, %rd20, %rd79; cvt.u16.u32 %rs45, %r360; mov.u16 %rs46, 255; st.global.v2.u8 [%rd80], {%rs45, %rs46}; $L__BB0_80: ld.const.v2.f32 {%f926, %f927}, [params+648]; mul.f32 %f292, %f1693, %f926; mul.f32 %f293, %f1694, %f927; ld.const.f32 %f294, [params+656]; mul.f32 %f295, %f1695, %f294; and.b32 %r361, %r88, 1; setp.eq.b32 %p91, %r361, 1; mov.pred %p92, 0; xor.pred %p93, %p91, %p92; not.pred %p94, %p93; @%p94 bra $L__BB0_154; mov.f32 %f929, 0f3E666666; cvt.rzi.f32.f32 %f930, %f929; add.f32 %f931, %f930, %f930; mov.f32 %f932, 0f3EE66666; sub.f32 %f933, %f932, %f931; abs.f32 %f296, %f933; abs.f32 %f297, %f292; setp.lt.f32 %p95, %f297, 0f00800000; mul.f32 %f934, %f297, 0f4B800000; selp.f32 %f935, %f934, %f297, %p95; selp.f32 %f936, 0fC3170000, 0fC2FE0000, %p95; mov.b32 %r362, %f935; and.b32 %r363, %r362, 8388607; or.b32 %r364, %r363, 1065353216; mov.b32 %f937, %r364; shr.u32 %r365, %r362, 23; cvt.rn.f32.u32 %f938, %r365; add.f32 %f939, %f936, %f938; setp.gt.f32 %p96, %f937, 0f3FB504F3; mul.f32 %f940, %f937, 0f3F000000; add.f32 %f941, %f939, 0f3F800000; selp.f32 %f942, %f941, %f939, %p96; selp.f32 %f943, %f940, %f937, %p96; add.f32 %f944, %f943, 0fBF800000; add.f32 %f945, %f943, 0f3F800000; rcp.approx.ftz.f32 %f946, %f945; add.f32 %f947, %f944, %f944; mul.f32 %f948, %f947, %f946; mul.f32 %f949, %f948, %f948; mov.f32 %f950, 0f3C4CAF63; mov.f32 %f951, 0f3B18F0FE; fma.rn.f32 %f952, %f951, %f949, %f950; mov.f32 %f953, 0f3DAAAABD; fma.rn.f32 %f954, %f952, %f949, %f953; mul.rn.f32 %f955, %f954, %f949; mul.rn.f32 %f956, %f955, %f948; sub.f32 %f957, %f944, %f948; add.f32 %f958, %f957, %f957; neg.f32 %f959, %f948; fma.rn.f32 %f960, %f959, %f944, %f958; mul.rn.f32 %f961, %f946, %f960; add.f32 %f962, %f956, %f948; sub.f32 %f963, %f948, %f962; add.f32 %f964, %f956, %f963; add.f32 %f965, %f961, %f964; add.f32 %f966, %f962, %f965; sub.f32 %f967, %f962, %f966; add.f32 %f968, %f965, %f967; mov.f32 %f969, 0f3F317200; mul.rn.f32 %f970, %f942, %f969; mov.f32 %f971, 0f35BFBE8E; mul.rn.f32 %f972, %f942, %f971; add.f32 %f973, %f970, %f966; sub.f32 %f974, %f970, %f973; add.f32 %f975, %f966, %f974; add.f32 %f976, %f968, %f975; add.f32 %f977, %f972, %f976; add.f32 %f978, %f973, %f977; sub.f32 %f979, %f973, %f978; add.f32 %f980, %f977, %f979; mul.rn.f32 %f981, %f932, %f978; neg.f32 %f982, %f981; fma.rn.f32 %f983, %f932, %f978, %f982; fma.rn.f32 %f984, %f932, %f980, %f983; mov.f32 %f985, 0f00000000; fma.rn.f32 %f986, %f985, %f978, %f984; add.rn.f32 %f987, %f981, %f986; neg.f32 %f988, %f987; add.rn.f32 %f989, %f981, %f988; add.rn.f32 %f990, %f989, %f986; mov.b32 %r366, %f987; setp.eq.s32 %p97, %r366, 1118925336; add.s32 %r367, %r366, -1; mov.b32 %f991, %r367; add.f32 %f992, %f990, 0f37000000; selp.f32 %f298, %f992, %f990, %p97; selp.f32 %f993, %f991, %f987, %p97; mov.f32 %f994, 0f3FB8AA3B; mul.rn.f32 %f995, %f993, %f994; cvt.rzi.f32.f32 %f996, %f995; abs.f32 %f997, %f996; setp.gt.f32 %p98, %f997, 0f42FC0000; mov.b32 %r368, %f996; and.b32 %r369, %r368, -2147483648; or.b32 %r370, %r369, 1123811328; mov.b32 %f998, %r370; selp.f32 %f999, %f998, %f996, %p98; mov.f32 %f1000, 0fBF317218; fma.rn.f32 %f1001, %f999, %f1000, %f993; mov.f32 %f1002, 0f3102E308; fma.rn.f32 %f1003, %f999, %f1002, %f1001; mul.f32 %f1004, %f1003, 0f3FB8AA3B; add.f32 %f1005, %f999, 0f4B40007F; mov.b32 %r371, %f1005; shl.b32 %r372, %r371, 23; mov.b32 %f1006, %r372; ex2.approx.ftz.f32 %f1007, %f1004; mul.f32 %f299, %f1007, %f1006; setp.eq.f32 %p99, %f299, 0f7F800000; mov.f32 %f1700, 0f7F800000; @%p99 bra $L__BB0_83; fma.rn.f32 %f1700, %f299, %f298, %f299; $L__BB0_83: setp.lt.f32 %p100, %f292, 0f00000000; setp.eq.f32 %p101, %f296, 0f3F800000; and.pred %p4, %p100, %p101; setp.eq.f32 %p102, %f292, 0f00000000; @%p102 bra $L__BB0_87; bra.uni $L__BB0_84; $L__BB0_87: add.f32 %f1012, %f292, %f292; selp.f32 %f1702, %f1012, 0f00000000, %p101; bra.uni $L__BB0_88; $L__BB0_84: mov.b32 %r373, %f1700; xor.b32 %r374, %r373, -2147483648; mov.b32 %f1008, %r374; selp.f32 %f1702, %f1008, %f1700, %p4; setp.geu.f32 %p103, %f292, 0f00000000; @%p103 bra $L__BB0_88; mov.f32 %f1009, 0f3EE66666; cvt.rzi.f32.f32 %f1010, %f1009; setp.eq.f32 %p104, %f1010, 0f3EE66666; @%p104 bra $L__BB0_88; mov.f32 %f1702, 0f7FFFFFFF; $L__BB0_88: add.f32 %f1013, %f297, 0f3EE66666; mov.b32 %r375, %f1013; setp.lt.s32 %p106, %r375, 2139095040; @%p106 bra $L__BB0_93; setp.gtu.f32 %p107, %f297, 0f7F800000; @%p107 bra $L__BB0_92; bra.uni $L__BB0_90; $L__BB0_92: add.f32 %f1702, %f292, 0f3EE66666; bra.uni $L__BB0_93; $L__BB0_90: setp.neu.f32 %p108, %f297, 0f7F800000; @%p108 bra $L__BB0_93; selp.f32 %f1702, 0fFF800000, 0f7F800000, %p4; $L__BB0_93: setp.eq.f32 %p109, %f292, 0f3F800000; selp.f32 %f308, 0f3F800000, %f1702, %p109; abs.f32 %f309, %f293; setp.lt.f32 %p110, %f309, 0f00800000; mul.f32 %f1015, %f309, 0f4B800000; selp.f32 %f1016, %f1015, %f309, %p110; selp.f32 %f1017, 0fC3170000, 0fC2FE0000, %p110; mov.b32 %r376, %f1016; and.b32 %r377, %r376, 8388607; or.b32 %r378, %r377, 1065353216; mov.b32 %f1018, %r378; shr.u32 %r379, %r376, 23; cvt.rn.f32.u32 %f1019, %r379; add.f32 %f1020, %f1017, %f1019; setp.gt.f32 %p111, %f1018, 0f3FB504F3; mul.f32 %f1021, %f1018, 0f3F000000; add.f32 %f1022, %f1020, 0f3F800000; selp.f32 %f1023, %f1022, %f1020, %p111; selp.f32 %f1024, %f1021, %f1018, %p111; add.f32 %f1025, %f1024, 0fBF800000; add.f32 %f1026, %f1024, 0f3F800000; rcp.approx.ftz.f32 %f1027, %f1026; add.f32 %f1028, %f1025, %f1025; mul.f32 %f1029, %f1028, %f1027; mul.f32 %f1030, %f1029, %f1029; mov.f32 %f1031, 0f3C4CAF63; mov.f32 %f1032, 0f3B18F0FE; fma.rn.f32 %f1033, %f1032, %f1030, %f1031; mov.f32 %f1034, 0f3DAAAABD; fma.rn.f32 %f1035, %f1033, %f1030, %f1034; mul.rn.f32 %f1036, %f1035, %f1030; mul.rn.f32 %f1037, %f1036, %f1029; sub.f32 %f1038, %f1025, %f1029; add.f32 %f1039, %f1038, %f1038; neg.f32 %f1040, %f1029; fma.rn.f32 %f1041, %f1040, %f1025, %f1039; mul.rn.f32 %f1042, %f1027, %f1041; add.f32 %f1043, %f1037, %f1029; sub.f32 %f1044, %f1029, %f1043; add.f32 %f1045, %f1037, %f1044; add.f32 %f1046, %f1042, %f1045; add.f32 %f1047, %f1043, %f1046; sub.f32 %f1048, %f1043, %f1047; add.f32 %f1049, %f1046, %f1048; mov.f32 %f1050, 0f3F317200; mul.rn.f32 %f1051, %f1023, %f1050; mov.f32 %f1052, 0f35BFBE8E; mul.rn.f32 %f1053, %f1023, %f1052; add.f32 %f1054, %f1051, %f1047; sub.f32 %f1055, %f1051, %f1054; add.f32 %f1056, %f1047, %f1055; add.f32 %f1057, %f1049, %f1056; add.f32 %f1058, %f1053, %f1057; add.f32 %f1059, %f1054, %f1058; sub.f32 %f1060, %f1054, %f1059; add.f32 %f1061, %f1058, %f1060; mov.f32 %f1062, 0f3EE66666; mul.rn.f32 %f1063, %f1062, %f1059; neg.f32 %f1064, %f1063; fma.rn.f32 %f1065, %f1062, %f1059, %f1064; fma.rn.f32 %f1066, %f1062, %f1061, %f1065; mov.f32 %f1067, 0f00000000; fma.rn.f32 %f1068, %f1067, %f1059, %f1066; add.rn.f32 %f1069, %f1063, %f1068; neg.f32 %f1070, %f1069; add.rn.f32 %f1071, %f1063, %f1070; add.rn.f32 %f1072, %f1071, %f1068; mov.b32 %r380, %f1069; setp.eq.s32 %p112, %r380, 1118925336; add.s32 %r381, %r380, -1; mov.b32 %f1073, %r381; add.f32 %f1074, %f1072, 0f37000000; selp.f32 %f310, %f1074, %f1072, %p112; selp.f32 %f1075, %f1073, %f1069, %p112; mov.f32 %f1076, 0f3FB8AA3B; mul.rn.f32 %f1077, %f1075, %f1076; cvt.rzi.f32.f32 %f1078, %f1077; abs.f32 %f1079, %f1078; setp.gt.f32 %p113, %f1079, 0f42FC0000; mov.b32 %r382, %f1078; and.b32 %r383, %r382, -2147483648; or.b32 %r384, %r383, 1123811328; mov.b32 %f1080, %r384; selp.f32 %f1081, %f1080, %f1078, %p113; mov.f32 %f1082, 0fBF317218; fma.rn.f32 %f1083, %f1081, %f1082, %f1075; mov.f32 %f1084, 0f3102E308; fma.rn.f32 %f1085, %f1081, %f1084, %f1083; mul.f32 %f1086, %f1085, 0f3FB8AA3B; add.f32 %f1087, %f1081, 0f4B40007F; mov.b32 %r385, %f1087; shl.b32 %r386, %r385, 23; mov.b32 %f1088, %r386; ex2.approx.ftz.f32 %f1089, %f1086; mul.f32 %f311, %f1089, %f1088; setp.eq.f32 %p114, %f311, 0f7F800000; mov.f32 %f1703, 0f7F800000; @%p114 bra $L__BB0_95; fma.rn.f32 %f1703, %f311, %f310, %f311; $L__BB0_95: setp.lt.f32 %p115, %f293, 0f00000000; and.pred %p5, %p115, %p101; setp.eq.f32 %p117, %f293, 0f00000000; @%p117 bra $L__BB0_99; bra.uni $L__BB0_96; $L__BB0_99: add.f32 %f1094, %f293, %f293; selp.f32 %f1705, %f1094, 0f00000000, %p101; bra.uni $L__BB0_100; $L__BB0_96: mov.b32 %r387, %f1703; xor.b32 %r388, %r387, -2147483648; mov.b32 %f1090, %r388; selp.f32 %f1705, %f1090, %f1703, %p5; setp.geu.f32 %p118, %f293, 0f00000000; @%p118 bra $L__BB0_100; mov.f32 %f1091, 0f3EE66666; cvt.rzi.f32.f32 %f1092, %f1091; setp.eq.f32 %p119, %f1092, 0f3EE66666; @%p119 bra $L__BB0_100; mov.f32 %f1705, 0f7FFFFFFF; $L__BB0_100: add.f32 %f1095, %f309, 0f3EE66666; mov.b32 %r389, %f1095; setp.lt.s32 %p121, %r389, 2139095040; @%p121 bra $L__BB0_105; setp.gtu.f32 %p122, %f309, 0f7F800000; @%p122 bra $L__BB0_104; bra.uni $L__BB0_102; $L__BB0_104: add.f32 %f1705, %f293, 0f3EE66666; bra.uni $L__BB0_105; $L__BB0_102: setp.neu.f32 %p123, %f309, 0f7F800000; @%p123 bra $L__BB0_105; selp.f32 %f1705, 0fFF800000, 0f7F800000, %p5; $L__BB0_105: setp.eq.f32 %p124, %f293, 0f3F800000; selp.f32 %f320, 0f3F800000, %f1705, %p124; abs.f32 %f321, %f295; setp.lt.f32 %p125, %f321, 0f00800000; mul.f32 %f1097, %f321, 0f4B800000; selp.f32 %f1098, %f1097, %f321, %p125; selp.f32 %f1099, 0fC3170000, 0fC2FE0000, %p125; mov.b32 %r390, %f1098; and.b32 %r391, %r390, 8388607; or.b32 %r392, %r391, 1065353216; mov.b32 %f1100, %r392; shr.u32 %r393, %r390, 23; cvt.rn.f32.u32 %f1101, %r393; add.f32 %f1102, %f1099, %f1101; setp.gt.f32 %p126, %f1100, 0f3FB504F3; mul.f32 %f1103, %f1100, 0f3F000000; add.f32 %f1104, %f1102, 0f3F800000; selp.f32 %f1105, %f1104, %f1102, %p126; selp.f32 %f1106, %f1103, %f1100, %p126; add.f32 %f1107, %f1106, 0fBF800000; add.f32 %f1108, %f1106, 0f3F800000; rcp.approx.ftz.f32 %f1109, %f1108; add.f32 %f1110, %f1107, %f1107; mul.f32 %f1111, %f1110, %f1109; mul.f32 %f1112, %f1111, %f1111; mov.f32 %f1113, 0f3C4CAF63; mov.f32 %f1114, 0f3B18F0FE; fma.rn.f32 %f1115, %f1114, %f1112, %f1113; mov.f32 %f1116, 0f3DAAAABD; fma.rn.f32 %f1117, %f1115, %f1112, %f1116; mul.rn.f32 %f1118, %f1117, %f1112; mul.rn.f32 %f1119, %f1118, %f1111; sub.f32 %f1120, %f1107, %f1111; add.f32 %f1121, %f1120, %f1120; neg.f32 %f1122, %f1111; fma.rn.f32 %f1123, %f1122, %f1107, %f1121; mul.rn.f32 %f1124, %f1109, %f1123; add.f32 %f1125, %f1119, %f1111; sub.f32 %f1126, %f1111, %f1125; add.f32 %f1127, %f1119, %f1126; add.f32 %f1128, %f1124, %f1127; add.f32 %f1129, %f1125, %f1128; sub.f32 %f1130, %f1125, %f1129; add.f32 %f1131, %f1128, %f1130; mov.f32 %f1132, 0f3F317200; mul.rn.f32 %f1133, %f1105, %f1132; mov.f32 %f1134, 0f35BFBE8E; mul.rn.f32 %f1135, %f1105, %f1134; add.f32 %f1136, %f1133, %f1129; sub.f32 %f1137, %f1133, %f1136; add.f32 %f1138, %f1129, %f1137; add.f32 %f1139, %f1131, %f1138; add.f32 %f1140, %f1135, %f1139; add.f32 %f1141, %f1136, %f1140; sub.f32 %f1142, %f1136, %f1141; add.f32 %f1143, %f1140, %f1142; mov.f32 %f1144, 0f3EE66666; mul.rn.f32 %f1145, %f1144, %f1141; neg.f32 %f1146, %f1145; fma.rn.f32 %f1147, %f1144, %f1141, %f1146; fma.rn.f32 %f1148, %f1144, %f1143, %f1147; mov.f32 %f1149, 0f00000000; fma.rn.f32 %f1150, %f1149, %f1141, %f1148; add.rn.f32 %f1151, %f1145, %f1150; neg.f32 %f1152, %f1151; add.rn.f32 %f1153, %f1145, %f1152; add.rn.f32 %f1154, %f1153, %f1150; mov.b32 %r394, %f1151; setp.eq.s32 %p127, %r394, 1118925336; add.s32 %r395, %r394, -1; mov.b32 %f1155, %r395; add.f32 %f1156, %f1154, 0f37000000; selp.f32 %f322, %f1156, %f1154, %p127; selp.f32 %f1157, %f1155, %f1151, %p127; mov.f32 %f1158, 0f3FB8AA3B; mul.rn.f32 %f1159, %f1157, %f1158; cvt.rzi.f32.f32 %f1160, %f1159; abs.f32 %f1161, %f1160; setp.gt.f32 %p128, %f1161, 0f42FC0000; mov.b32 %r396, %f1160; and.b32 %r397, %r396, -2147483648; or.b32 %r398, %r397, 1123811328; mov.b32 %f1162, %r398; selp.f32 %f1163, %f1162, %f1160, %p128; mov.f32 %f1164, 0fBF317218; fma.rn.f32 %f1165, %f1163, %f1164, %f1157; mov.f32 %f1166, 0f3102E308; fma.rn.f32 %f1167, %f1163, %f1166, %f1165; mul.f32 %f1168, %f1167, 0f3FB8AA3B; add.f32 %f1169, %f1163, 0f4B40007F; mov.b32 %r399, %f1169; shl.b32 %r400, %r399, 23; mov.b32 %f1170, %r400; ex2.approx.ftz.f32 %f1171, %f1168; mul.f32 %f323, %f1171, %f1170; setp.eq.f32 %p129, %f323, 0f7F800000; mov.f32 %f1706, 0f7F800000; @%p129 bra $L__BB0_107; fma.rn.f32 %f1706, %f323, %f322, %f323; $L__BB0_107: setp.lt.f32 %p130, %f295, 0f00000000; and.pred %p6, %p130, %p101; setp.eq.f32 %p132, %f295, 0f00000000; @%p132 bra $L__BB0_111; bra.uni $L__BB0_108; $L__BB0_111: add.f32 %f1176, %f295, %f295; selp.f32 %f1708, %f1176, 0f00000000, %p101; bra.uni $L__BB0_112; $L__BB0_108: mov.b32 %r401, %f1706; xor.b32 %r402, %r401, -2147483648; mov.b32 %f1172, %r402; selp.f32 %f1708, %f1172, %f1706, %p6; setp.geu.f32 %p133, %f295, 0f00000000; @%p133 bra $L__BB0_112; mov.f32 %f1173, 0f3EE66666; cvt.rzi.f32.f32 %f1174, %f1173; setp.eq.f32 %p134, %f1174, 0f3EE66666; @%p134 bra $L__BB0_112; mov.f32 %f1708, 0f7FFFFFFF; $L__BB0_112: add.f32 %f1177, %f321, 0f3EE66666; mov.b32 %r403, %f1177; setp.lt.s32 %p136, %r403, 2139095040; @%p136 bra $L__BB0_117; setp.gtu.f32 %p137, %f321, 0f7F800000; @%p137 bra $L__BB0_116; bra.uni $L__BB0_114; $L__BB0_116: add.f32 %f1708, %f295, 0f3EE66666; bra.uni $L__BB0_117; $L__BB0_114: setp.neu.f32 %p138, %f321, 0f7F800000; @%p138 bra $L__BB0_117; selp.f32 %f1708, 0fFF800000, 0f7F800000, %p6; $L__BB0_117: setp.eq.f32 %p139, %f295, 0f3F800000; mov.f32 %f1179, 0f3F800000; selp.f32 %f1180, 0f3F800000, %f1708, %p139; ld.const.u64 %rd81, [params+144]; cvta.to.global.u64 %rd22, %rd81; ld.const.u32 %r404, [params+136]; mad.lo.s32 %r405, %r404, %r7, %r6; cvt.u64.u32 %rd23, %r405; min.f32 %f1181, %f308, %f1179; mov.f32 %f1182, 0f00000000; max.f32 %f332, %f1182, %f1181; min.f32 %f1183, %f320, %f1179; max.f32 %f333, %f1182, %f1183; min.f32 %f1184, %f1180, %f1179; max.f32 %f334, %f1182, %f1184; mov.f32 %f1185, 0f3E555555; cvt.rzi.f32.f32 %f1186, %f1185; add.f32 %f1187, %f1186, %f1186; mov.f32 %f1188, 0f3ED55555; sub.f32 %f1189, %f1188, %f1187; abs.f32 %f335, %f1189; abs.f32 %f336, %f332; setp.lt.f32 %p140, %f336, 0f00800000; mul.f32 %f1190, %f336, 0f4B800000; selp.f32 %f1191, %f1190, %f336, %p140; selp.f32 %f1192, 0fC3170000, 0fC2FE0000, %p140; mov.b32 %r406, %f1191; and.b32 %r407, %r406, 8388607; or.b32 %r408, %r407, 1065353216; mov.b32 %f1193, %r408; shr.u32 %r409, %r406, 23; cvt.rn.f32.u32 %f1194, %r409; add.f32 %f1195, %f1192, %f1194; setp.gt.f32 %p141, %f1193, 0f3FB504F3; mul.f32 %f1196, %f1193, 0f3F000000; add.f32 %f1197, %f1195, 0f3F800000; selp.f32 %f1198, %f1197, %f1195, %p141; selp.f32 %f1199, %f1196, %f1193, %p141; add.f32 %f1200, %f1199, 0fBF800000; add.f32 %f1201, %f1199, 0f3F800000; rcp.approx.ftz.f32 %f1202, %f1201; add.f32 %f1203, %f1200, %f1200; mul.f32 %f1204, %f1203, %f1202; mul.f32 %f1205, %f1204, %f1204; mov.f32 %f1206, 0f3C4CAF63; mov.f32 %f1207, 0f3B18F0FE; fma.rn.f32 %f1208, %f1207, %f1205, %f1206; mov.f32 %f1209, 0f3DAAAABD; fma.rn.f32 %f1210, %f1208, %f1205, %f1209; mul.rn.f32 %f1211, %f1210, %f1205; mul.rn.f32 %f1212, %f1211, %f1204; sub.f32 %f1213, %f1200, %f1204; add.f32 %f1214, %f1213, %f1213; neg.f32 %f1215, %f1204; fma.rn.f32 %f1216, %f1215, %f1200, %f1214; mul.rn.f32 %f1217, %f1202, %f1216; add.f32 %f1218, %f1212, %f1204; sub.f32 %f1219, %f1204, %f1218; add.f32 %f1220, %f1212, %f1219; add.f32 %f1221, %f1217, %f1220; add.f32 %f1222, %f1218, %f1221; sub.f32 %f1223, %f1218, %f1222; add.f32 %f1224, %f1221, %f1223; mov.f32 %f1225, 0f3F317200; mul.rn.f32 %f1226, %f1198, %f1225; mov.f32 %f1227, 0f35BFBE8E; mul.rn.f32 %f1228, %f1198, %f1227; add.f32 %f1229, %f1226, %f1222; sub.f32 %f1230, %f1226, %f1229; add.f32 %f1231, %f1222, %f1230; add.f32 %f1232, %f1224, %f1231; add.f32 %f1233, %f1228, %f1232; add.f32 %f1234, %f1229, %f1233; sub.f32 %f1235, %f1229, %f1234; add.f32 %f1236, %f1233, %f1235; mul.rn.f32 %f1237, %f1188, %f1234; neg.f32 %f1238, %f1237; fma.rn.f32 %f1239, %f1188, %f1234, %f1238; fma.rn.f32 %f1240, %f1188, %f1236, %f1239; fma.rn.f32 %f1241, %f1182, %f1234, %f1240; add.rn.f32 %f1242, %f1237, %f1241; neg.f32 %f1243, %f1242; add.rn.f32 %f1244, %f1237, %f1243; add.rn.f32 %f1245, %f1244, %f1241; mov.b32 %r410, %f1242; setp.eq.s32 %p142, %r410, 1118925336; add.s32 %r411, %r410, -1; mov.b32 %f1246, %r411; add.f32 %f1247, %f1245, 0f37000000; selp.f32 %f337, %f1247, %f1245, %p142; selp.f32 %f1248, %f1246, %f1242, %p142; mov.f32 %f1249, 0f3FB8AA3B; mul.rn.f32 %f1250, %f1248, %f1249; cvt.rzi.f32.f32 %f1251, %f1250; abs.f32 %f1252, %f1251; setp.gt.f32 %p143, %f1252, 0f42FC0000; mov.b32 %r412, %f1251; and.b32 %r413, %r412, -2147483648; or.b32 %r414, %r413, 1123811328; mov.b32 %f1253, %r414; selp.f32 %f1254, %f1253, %f1251, %p143; mov.f32 %f1255, 0fBF317218; fma.rn.f32 %f1256, %f1254, %f1255, %f1248; mov.f32 %f1257, 0f3102E308; fma.rn.f32 %f1258, %f1254, %f1257, %f1256; mul.f32 %f1259, %f1258, 0f3FB8AA3B; add.f32 %f1260, %f1254, 0f4B40007F; mov.b32 %r415, %f1260; shl.b32 %r416, %r415, 23; mov.b32 %f1261, %r416; ex2.approx.ftz.f32 %f1262, %f1259; mul.f32 %f338, %f1262, %f1261; setp.eq.f32 %p144, %f338, 0f7F800000; mov.f32 %f1709, 0f7F800000; @%p144 bra $L__BB0_119; fma.rn.f32 %f1709, %f338, %f337, %f338; $L__BB0_119: setp.lt.f32 %p145, %f332, 0f00000000; setp.eq.f32 %p146, %f335, 0f3F800000; and.pred %p7, %p145, %p146; setp.eq.f32 %p147, %f332, 0f00000000; @%p147 bra $L__BB0_123; bra.uni $L__BB0_120; $L__BB0_123: add.f32 %f1267, %f332, %f332; selp.f32 %f1711, %f1267, 0f00000000, %p146; bra.uni $L__BB0_124; $L__BB0_120: mov.b32 %r417, %f1709; xor.b32 %r418, %r417, -2147483648; mov.b32 %f1263, %r418; selp.f32 %f1711, %f1263, %f1709, %p7; setp.geu.f32 %p148, %f332, 0f00000000; @%p148 bra $L__BB0_124; mov.f32 %f1264, 0f3ED55555; cvt.rzi.f32.f32 %f1265, %f1264; setp.eq.f32 %p149, %f1265, 0f3ED55555; @%p149 bra $L__BB0_124; mov.f32 %f1711, 0f7FFFFFFF; $L__BB0_124: add.f32 %f1268, %f336, 0f3ED55555; mov.b32 %r419, %f1268; setp.lt.s32 %p151, %r419, 2139095040; @%p151 bra $L__BB0_129; setp.gtu.f32 %p152, %f336, 0f7F800000; @%p152 bra $L__BB0_128; bra.uni $L__BB0_126; $L__BB0_128: add.f32 %f1711, %f332, 0f3ED55555; bra.uni $L__BB0_129; $L__BB0_126: setp.neu.f32 %p153, %f336, 0f7F800000; @%p153 bra $L__BB0_129; selp.f32 %f1711, 0fFF800000, 0f7F800000, %p7; $L__BB0_129: abs.f32 %f347, %f333; setp.lt.f32 %p154, %f347, 0f00800000; mul.f32 %f1270, %f347, 0f4B800000; selp.f32 %f1271, %f1270, %f347, %p154; selp.f32 %f1272, 0fC3170000, 0fC2FE0000, %p154; mov.b32 %r420, %f1271; and.b32 %r421, %r420, 8388607; or.b32 %r422, %r421, 1065353216; mov.b32 %f1273, %r422; shr.u32 %r423, %r420, 23; cvt.rn.f32.u32 %f1274, %r423; add.f32 %f1275, %f1272, %f1274; setp.gt.f32 %p155, %f1273, 0f3FB504F3; mul.f32 %f1276, %f1273, 0f3F000000; add.f32 %f1277, %f1275, 0f3F800000; selp.f32 %f1278, %f1277, %f1275, %p155; selp.f32 %f1279, %f1276, %f1273, %p155; add.f32 %f1280, %f1279, 0fBF800000; add.f32 %f1281, %f1279, 0f3F800000; rcp.approx.ftz.f32 %f1282, %f1281; add.f32 %f1283, %f1280, %f1280; mul.f32 %f1284, %f1283, %f1282; mul.f32 %f1285, %f1284, %f1284; mov.f32 %f1286, 0f3C4CAF63; mov.f32 %f1287, 0f3B18F0FE; fma.rn.f32 %f1288, %f1287, %f1285, %f1286; mov.f32 %f1289, 0f3DAAAABD; fma.rn.f32 %f1290, %f1288, %f1285, %f1289; mul.rn.f32 %f1291, %f1290, %f1285; mul.rn.f32 %f1292, %f1291, %f1284; sub.f32 %f1293, %f1280, %f1284; add.f32 %f1294, %f1293, %f1293; neg.f32 %f1295, %f1284; fma.rn.f32 %f1296, %f1295, %f1280, %f1294; mul.rn.f32 %f1297, %f1282, %f1296; add.f32 %f1298, %f1292, %f1284; sub.f32 %f1299, %f1284, %f1298; add.f32 %f1300, %f1292, %f1299; add.f32 %f1301, %f1297, %f1300; add.f32 %f1302, %f1298, %f1301; sub.f32 %f1303, %f1298, %f1302; add.f32 %f1304, %f1301, %f1303; mov.f32 %f1305, 0f3F317200; mul.rn.f32 %f1306, %f1278, %f1305; mov.f32 %f1307, 0f35BFBE8E; mul.rn.f32 %f1308, %f1278, %f1307; add.f32 %f1309, %f1306, %f1302; sub.f32 %f1310, %f1306, %f1309; add.f32 %f1311, %f1302, %f1310; add.f32 %f1312, %f1304, %f1311; add.f32 %f1313, %f1308, %f1312; add.f32 %f1314, %f1309, %f1313; sub.f32 %f1315, %f1309, %f1314; add.f32 %f1316, %f1313, %f1315; mov.f32 %f1317, 0f3ED55555; mul.rn.f32 %f1318, %f1317, %f1314; neg.f32 %f1319, %f1318; fma.rn.f32 %f1320, %f1317, %f1314, %f1319; fma.rn.f32 %f1321, %f1317, %f1316, %f1320; mov.f32 %f1322, 0f00000000; fma.rn.f32 %f1323, %f1322, %f1314, %f1321; add.rn.f32 %f1324, %f1318, %f1323; neg.f32 %f1325, %f1324; add.rn.f32 %f1326, %f1318, %f1325; add.rn.f32 %f1327, %f1326, %f1323; mov.b32 %r424, %f1324; setp.eq.s32 %p156, %r424, 1118925336; add.s32 %r425, %r424, -1; mov.b32 %f1328, %r425; add.f32 %f1329, %f1327, 0f37000000; selp.f32 %f348, %f1329, %f1327, %p156; selp.f32 %f1330, %f1328, %f1324, %p156; mov.f32 %f1331, 0f3FB8AA3B; mul.rn.f32 %f1332, %f1330, %f1331; cvt.rzi.f32.f32 %f1333, %f1332; abs.f32 %f1334, %f1333; setp.gt.f32 %p157, %f1334, 0f42FC0000; mov.b32 %r426, %f1333; and.b32 %r427, %r426, -2147483648; or.b32 %r428, %r427, 1123811328; mov.b32 %f1335, %r428; selp.f32 %f1336, %f1335, %f1333, %p157; mov.f32 %f1337, 0fBF317218; fma.rn.f32 %f1338, %f1336, %f1337, %f1330; mov.f32 %f1339, 0f3102E308; fma.rn.f32 %f1340, %f1336, %f1339, %f1338; mul.f32 %f1341, %f1340, 0f3FB8AA3B; add.f32 %f1342, %f1336, 0f4B40007F; mov.b32 %r429, %f1342; shl.b32 %r430, %r429, 23; mov.b32 %f1343, %r430; ex2.approx.ftz.f32 %f1344, %f1341; mul.f32 %f349, %f1344, %f1343; setp.eq.f32 %p158, %f349, 0f7F800000; mov.f32 %f1712, 0f7F800000; @%p158 bra $L__BB0_131; fma.rn.f32 %f1712, %f349, %f348, %f349; $L__BB0_131: setp.lt.f32 %p159, %f333, 0f00000000; and.pred %p8, %p159, %p146; setp.eq.f32 %p161, %f333, 0f00000000; @%p161 bra $L__BB0_135; bra.uni $L__BB0_132; $L__BB0_135: add.f32 %f1349, %f333, %f333; selp.f32 %f1714, %f1349, 0f00000000, %p146; bra.uni $L__BB0_136; $L__BB0_132: mov.b32 %r431, %f1712; xor.b32 %r432, %r431, -2147483648; mov.b32 %f1345, %r432; selp.f32 %f1714, %f1345, %f1712, %p8; setp.geu.f32 %p162, %f333, 0f00000000; @%p162 bra $L__BB0_136; mov.f32 %f1346, 0f3ED55555; cvt.rzi.f32.f32 %f1347, %f1346; setp.eq.f32 %p163, %f1347, 0f3ED55555; @%p163 bra $L__BB0_136; mov.f32 %f1714, 0f7FFFFFFF; $L__BB0_136: add.f32 %f1350, %f347, 0f3ED55555; mov.b32 %r433, %f1350; setp.lt.s32 %p165, %r433, 2139095040; @%p165 bra $L__BB0_141; setp.gtu.f32 %p166, %f347, 0f7F800000; @%p166 bra $L__BB0_140; bra.uni $L__BB0_138; $L__BB0_140: add.f32 %f1714, %f333, 0f3ED55555; bra.uni $L__BB0_141; $L__BB0_138: setp.neu.f32 %p167, %f347, 0f7F800000; @%p167 bra $L__BB0_141; selp.f32 %f1714, 0fFF800000, 0f7F800000, %p8; $L__BB0_141: abs.f32 %f358, %f334; setp.lt.f32 %p168, %f358, 0f00800000; mul.f32 %f1352, %f358, 0f4B800000; selp.f32 %f1353, %f1352, %f358, %p168; selp.f32 %f1354, 0fC3170000, 0fC2FE0000, %p168; mov.b32 %r434, %f1353; and.b32 %r435, %r434, 8388607; or.b32 %r436, %r435, 1065353216; mov.b32 %f1355, %r436; shr.u32 %r437, %r434, 23; cvt.rn.f32.u32 %f1356, %r437; add.f32 %f1357, %f1354, %f1356; setp.gt.f32 %p169, %f1355, 0f3FB504F3; mul.f32 %f1358, %f1355, 0f3F000000; add.f32 %f1359, %f1357, 0f3F800000; selp.f32 %f1360, %f1359, %f1357, %p169; selp.f32 %f1361, %f1358, %f1355, %p169; add.f32 %f1362, %f1361, 0fBF800000; add.f32 %f1363, %f1361, 0f3F800000; rcp.approx.ftz.f32 %f1364, %f1363; add.f32 %f1365, %f1362, %f1362; mul.f32 %f1366, %f1365, %f1364; mul.f32 %f1367, %f1366, %f1366; mov.f32 %f1368, 0f3C4CAF63; mov.f32 %f1369, 0f3B18F0FE; fma.rn.f32 %f1370, %f1369, %f1367, %f1368; mov.f32 %f1371, 0f3DAAAABD; fma.rn.f32 %f1372, %f1370, %f1367, %f1371; mul.rn.f32 %f1373, %f1372, %f1367; mul.rn.f32 %f1374, %f1373, %f1366; sub.f32 %f1375, %f1362, %f1366; add.f32 %f1376, %f1375, %f1375; neg.f32 %f1377, %f1366; fma.rn.f32 %f1378, %f1377, %f1362, %f1376; mul.rn.f32 %f1379, %f1364, %f1378; add.f32 %f1380, %f1374, %f1366; sub.f32 %f1381, %f1366, %f1380; add.f32 %f1382, %f1374, %f1381; add.f32 %f1383, %f1379, %f1382; add.f32 %f1384, %f1380, %f1383; sub.f32 %f1385, %f1380, %f1384; add.f32 %f1386, %f1383, %f1385; mov.f32 %f1387, 0f3F317200; mul.rn.f32 %f1388, %f1360, %f1387; mov.f32 %f1389, 0f35BFBE8E; mul.rn.f32 %f1390, %f1360, %f1389; add.f32 %f1391, %f1388, %f1384; sub.f32 %f1392, %f1388, %f1391; add.f32 %f1393, %f1384, %f1392; add.f32 %f1394, %f1386, %f1393; add.f32 %f1395, %f1390, %f1394; add.f32 %f1396, %f1391, %f1395; sub.f32 %f1397, %f1391, %f1396; add.f32 %f1398, %f1395, %f1397; mov.f32 %f1399, 0f3ED55555; mul.rn.f32 %f1400, %f1399, %f1396; neg.f32 %f1401, %f1400; fma.rn.f32 %f1402, %f1399, %f1396, %f1401; fma.rn.f32 %f1403, %f1399, %f1398, %f1402; mov.f32 %f1404, 0f00000000; fma.rn.f32 %f1405, %f1404, %f1396, %f1403; add.rn.f32 %f1406, %f1400, %f1405; neg.f32 %f1407, %f1406; add.rn.f32 %f1408, %f1400, %f1407; add.rn.f32 %f1409, %f1408, %f1405; mov.b32 %r438, %f1406; setp.eq.s32 %p170, %r438, 1118925336; add.s32 %r439, %r438, -1; mov.b32 %f1410, %r439; add.f32 %f1411, %f1409, 0f37000000; selp.f32 %f359, %f1411, %f1409, %p170; selp.f32 %f1412, %f1410, %f1406, %p170; mov.f32 %f1413, 0f3FB8AA3B; mul.rn.f32 %f1414, %f1412, %f1413; cvt.rzi.f32.f32 %f1415, %f1414; abs.f32 %f1416, %f1415; setp.gt.f32 %p171, %f1416, 0f42FC0000; mov.b32 %r440, %f1415; and.b32 %r441, %r440, -2147483648; or.b32 %r442, %r441, 1123811328; mov.b32 %f1417, %r442; selp.f32 %f1418, %f1417, %f1415, %p171; mov.f32 %f1419, 0fBF317218; fma.rn.f32 %f1420, %f1418, %f1419, %f1412; mov.f32 %f1421, 0f3102E308; fma.rn.f32 %f1422, %f1418, %f1421, %f1420; mul.f32 %f1423, %f1422, 0f3FB8AA3B; add.f32 %f1424, %f1418, 0f4B40007F; mov.b32 %r443, %f1424; shl.b32 %r444, %r443, 23; mov.b32 %f1425, %r444; ex2.approx.ftz.f32 %f1426, %f1423; mul.f32 %f360, %f1426, %f1425; setp.eq.f32 %p172, %f360, 0f7F800000; mov.f32 %f1715, 0f7F800000; @%p172 bra $L__BB0_143; fma.rn.f32 %f1715, %f360, %f359, %f360; $L__BB0_143: setp.lt.f32 %p173, %f334, 0f00000000; and.pred %p9, %p173, %p146; setp.eq.f32 %p175, %f334, 0f00000000; @%p175 bra $L__BB0_147; bra.uni $L__BB0_144; $L__BB0_147: add.f32 %f1431, %f334, %f334; selp.f32 %f1717, %f1431, 0f00000000, %p146; bra.uni $L__BB0_148; $L__BB0_144: mov.b32 %r445, %f1715; xor.b32 %r446, %r445, -2147483648; mov.b32 %f1427, %r446; selp.f32 %f1717, %f1427, %f1715, %p9; setp.geu.f32 %p176, %f334, 0f00000000; @%p176 bra $L__BB0_148; mov.f32 %f1428, 0f3ED55555; cvt.rzi.f32.f32 %f1429, %f1428; setp.eq.f32 %p177, %f1429, 0f3ED55555; @%p177 bra $L__BB0_148; mov.f32 %f1717, 0f7FFFFFFF; $L__BB0_148: add.f32 %f1432, %f358, 0f3ED55555; mov.b32 %r447, %f1432; setp.lt.s32 %p179, %r447, 2139095040; @%p179 bra $L__BB0_153; setp.gtu.f32 %p180, %f358, 0f7F800000; @%p180 bra $L__BB0_152; bra.uni $L__BB0_150; $L__BB0_152: add.f32 %f1717, %f334, 0f3ED55555; bra.uni $L__BB0_153; $L__BB0_150: setp.neu.f32 %p181, %f358, 0f7F800000; @%p181 bra $L__BB0_153; selp.f32 %f1717, 0fFF800000, 0f7F800000, %p9; $L__BB0_153: fma.rn.f32 %f1433, %f1711, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p182, %f332, 0f3F800000; mov.f32 %f1434, 0f3F800000; selp.f32 %f1435, 0f3F7FFFFF, %f1433, %p182; mul.f32 %f1436, %f332, 0f414EB852; setp.lt.f32 %p183, %f332, 0f3B4D2E1C; selp.f32 %f1437, %f1436, %f1435, %p183; fma.rn.f32 %f1438, %f1714, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p184, %f333, 0f3F800000; selp.f32 %f1439, 0f3F7FFFFF, %f1438, %p184; mul.f32 %f1440, %f333, 0f414EB852; setp.lt.f32 %p185, %f333, 0f3B4D2E1C; selp.f32 %f1441, %f1440, %f1439, %p185; fma.rn.f32 %f1442, %f1717, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p186, %f334, 0f3F800000; selp.f32 %f1443, 0f3F7FFFFF, %f1442, %p186; mul.f32 %f1444, %f334, 0f414EB852; setp.lt.f32 %p187, %f334, 0f3B4D2E1C; selp.f32 %f1445, %f1444, %f1443, %p187; min.f32 %f1446, %f1437, %f1434; mov.f32 %f1447, 0f00000000; max.f32 %f1448, %f1447, %f1446; mul.f32 %f1449, %f1448, 0f43800000; cvt.rzi.u32.f32 %r448, %f1449; min.u32 %r449, %r448, 255; min.f32 %f1450, %f1441, %f1434; max.f32 %f1451, %f1447, %f1450; mul.f32 %f1452, %f1451, 0f43800000; cvt.rzi.u32.f32 %r450, %f1452; min.u32 %r451, %r450, 255; min.f32 %f1453, %f1445, %f1434; max.f32 %f1454, %f1447, %f1453; mul.f32 %f1455, %f1454, 0f43800000; cvt.rzi.u32.f32 %r452, %f1455; min.u32 %r453, %r452, 255; shl.b64 %rd82, %rd23, 2; add.s64 %rd83, %rd22, %rd82; cvt.u16.u32 %rs47, %r453; cvt.u16.u32 %rs48, %r451; cvt.u16.u32 %rs49, %r449; mov.u16 %rs50, 255; st.global.v4.u8 [%rd83], {%rs49, %rs48, %rs47, %rs50}; $L__BB0_154: and.b32 %r454, %r88, 4; setp.eq.s32 %p188, %r454, 0; ld.const.u32 %r498, [params+108]; @%p188 bra $L__BB0_158; setp.eq.s32 %p189, %r498, 0; ld.const.u64 %rd84, [params+224]; cvta.to.global.u64 %rd85, %rd84; ld.const.u32 %r455, [params+216]; mad.lo.s32 %r456, %r455, %r7, %r6; mul.wide.u32 %rd86, %r456, 8; add.s64 %rd24, %rd85, %rd86; @%p189 bra $L__BB0_157; ld.global.v4.u16 {%rs58, %rs59, %rs60, %rs61}, [%rd24]; // begin inline asm { cvt.f32.f16 %f1456, %rs58;} // end inline asm // begin inline asm { cvt.f32.f16 %f1457, %rs59;} // end inline asm // begin inline asm { cvt.f32.f16 %f1458, %rs60;} // end inline asm add.f32 %f1459, %f292, %f1456; add.f32 %f1460, %f293, %f1457; add.f32 %f1461, %f295, %f1458; mov.f32 %f1462, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs56, %f1461;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs55, %f1460;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs54, %f1459;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs57, %f1462;} // end inline asm st.global.v4.u16 [%rd24], {%rs54, %rs55, %rs56, %rs57}; bra.uni $L__BB0_158; $L__BB0_157: mov.f32 %f1466, 0f3F800000; mov.u32 %r498, 0; // begin inline asm { cvt.rn.f16.f32 %rs65, %f1466;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs64, %f295;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs63, %f293;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs62, %f292;} // end inline asm st.global.v4.u16 [%rd24], {%rs62, %rs63, %rs64, %rs65}; $L__BB0_158: add.f32 %f1467, %f1687, %f1690; add.f32 %f1468, %f1684, %f1467; add.f32 %f1469, %f1688, %f1691; add.f32 %f1470, %f1685, %f1469; add.f32 %f1471, %f1689, %f1692; add.f32 %f1472, %f1686, %f1471; mul.f32 %f1473, %f1468, 0f3F13CD3A; mul.f32 %f1474, %f1470, 0f3F13CD3A; mul.f32 %f1475, %f1472, 0f3F13CD3A; div.rn.f32 %f1476, %f1693, %f1473; div.rn.f32 %f1477, %f1694, %f1474; div.rn.f32 %f1478, %f1695, %f1475; setp.eq.f32 %p190, %f1693, 0f00000000; selp.f32 %f1479, 0f00000000, %f1476, %p190; setp.eq.f32 %p191, %f1694, 0f00000000; selp.f32 %f1480, 0f00000000, %f1477, %p191; setp.eq.f32 %p192, %f1695, 0f00000000; selp.f32 %f1481, 0f00000000, %f1478, %p192; mul.f32 %f1482, %f1690, %f1479; mul.f32 %f1483, %f1691, %f1480; mul.f32 %f1484, %f1692, %f1481; mul.f32 %f1485, %f1687, %f1479; mul.f32 %f1486, %f1688, %f1480; mul.f32 %f1487, %f1689, %f1481; mul.f32 %f1488, %f1684, %f1479; mul.f32 %f1489, %f1685, %f1480; mul.f32 %f1490, %f1686, %f1481; mul.f32 %f369, %f1482, %f926; mul.f32 %f370, %f1483, %f927; mul.f32 %f371, %f1484, %f294; mul.f32 %f372, %f1485, %f926; mul.f32 %f373, %f1486, %f927; mul.f32 %f374, %f1487, %f294; mul.f32 %f375, %f1488, %f926; mul.f32 %f376, %f1489, %f927; mul.f32 %f377, %f1490, %f294; ld.const.u64 %rd87, [params+256]; cvta.to.global.u64 %rd88, %rd87; ld.const.u32 %r458, [params+248]; mad.lo.s32 %r459, %r458, %r7, %r6; mul.wide.u32 %rd89, %r459, 8; add.s64 %rd25, %rd88, %rd89; setp.eq.s32 %p193, %r498, 0; @%p193 bra $L__BB0_160; ld.global.v4.u16 {%rs73, %rs74, %rs75, %rs76}, [%rd25]; // begin inline asm { cvt.f32.f16 %f1491, %rs73;} // end inline asm // begin inline asm { cvt.f32.f16 %f1492, %rs74;} // end inline asm // begin inline asm { cvt.f32.f16 %f1493, %rs75;} // end inline asm add.f32 %f1494, %f369, %f1491; add.f32 %f1495, %f370, %f1492; add.f32 %f1496, %f371, %f1493; mov.f32 %f1497, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs71, %f1496;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs70, %f1495;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs69, %f1494;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs72, %f1497;} // end inline asm st.global.v4.u16 [%rd25], {%rs69, %rs70, %rs71, %rs72}; bra.uni $L__BB0_161; $L__BB0_160: mov.f32 %f1501, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs80, %f1501;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs79, %f371;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs78, %f370;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs77, %f369;} // end inline asm st.global.v4.u16 [%rd25], {%rs77, %rs78, %rs79, %rs80}; $L__BB0_161: ld.const.u64 %rd90, [params+272]; cvta.to.global.u64 %rd91, %rd90; ld.const.u32 %r460, [params+264]; mad.lo.s32 %r461, %r460, %r7, %r6; mul.wide.u32 %rd92, %r461, 8; add.s64 %rd26, %rd91, %rd92; @%p193 bra $L__BB0_163; ld.global.v4.u16 {%rs88, %rs89, %rs90, %rs91}, [%rd26]; // begin inline asm { cvt.f32.f16 %f1502, %rs88;} // end inline asm // begin inline asm { cvt.f32.f16 %f1503, %rs89;} // end inline asm // begin inline asm { cvt.f32.f16 %f1504, %rs90;} // end inline asm add.f32 %f1505, %f372, %f1502; add.f32 %f1506, %f373, %f1503; add.f32 %f1507, %f374, %f1504; mov.f32 %f1508, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs86, %f1507;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs85, %f1506;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs84, %f1505;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs87, %f1508;} // end inline asm st.global.v4.u16 [%rd26], {%rs84, %rs85, %rs86, %rs87}; bra.uni $L__BB0_164; $L__BB0_163: mov.f32 %f1512, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs95, %f1512;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs94, %f374;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs93, %f373;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs92, %f372;} // end inline asm st.global.v4.u16 [%rd26], {%rs92, %rs93, %rs94, %rs95}; $L__BB0_164: ld.const.u64 %rd93, [params+288]; cvta.to.global.u64 %rd94, %rd93; ld.const.u32 %r462, [params+280]; mad.lo.s32 %r463, %r462, %r7, %r6; mul.wide.u32 %rd95, %r463, 8; add.s64 %rd27, %rd94, %rd95; @%p193 bra $L__BB0_166; ld.global.v4.u16 {%rs103, %rs104, %rs105, %rs106}, [%rd27]; // begin inline asm { cvt.f32.f16 %f1513, %rs103;} // end inline asm // begin inline asm { cvt.f32.f16 %f1514, %rs104;} // end inline asm // begin inline asm { cvt.f32.f16 %f1515, %rs105;} // end inline asm add.f32 %f1516, %f375, %f1513; add.f32 %f1517, %f376, %f1514; add.f32 %f1518, %f377, %f1515; mov.f32 %f1519, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs101, %f1518;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs100, %f1517;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs99, %f1516;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs102, %f1519;} // end inline asm st.global.v4.u16 [%rd27], {%rs99, %rs100, %rs101, %rs102}; bra.uni $L__BB0_184; $L__BB0_166: mov.f32 %f1523, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs110, %f1523;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs109, %f377;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs108, %f376;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs107, %f375;} // end inline asm st.global.v4.u16 [%rd27], {%rs107, %rs108, %rs109, %rs110}; $L__BB0_184: ret; }