// // 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<194>; .reg .b16 %rs<62>; .reg .f32 %f<1307>; .reg .b32 %r<456>; .reg .f64 %fd<5>; .reg .b64 %rd<101>; mov.u64 %SPL, __local_depot0; add.u64 %rd1, %SPL, 0; // begin inline asm call (%r58), _optix_get_launch_index_x, (); // end inline asm // begin inline asm call (%r59), _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 %rd26, [params+400]; cvta.to.global.u64 %rd27, %rd26; ld.const.u32 %r64, [params+392]; mad.lo.s32 %r65, %r64, %r59, %r58; mul.wide.u32 %rd28, %r65, 4; add.s64 %rd2, %rd27, %rd28; ld.global.v2.u8 {%rs7, %rs61}, [%rd2]; or.b16 %rs9, %rs7, %rs61; and.b16 %rs10, %rs9, 255; setp.eq.s16 %p10, %rs10, 0; @%p10 bra $L__BB0_2; ld.global.u8 %rs60, [%rd2+2]; bra.uni $L__BB0_3; $L__BB0_2: ld.global.u8 %rs60, [%rd2+2]; setp.eq.s16 %p11, %rs60, 0; mov.f32 %f1221, 0f00000000; mov.u16 %rs61, 0; mov.f32 %f1222, %f1221; mov.f32 %f1223, %f1221; @%p11 bra $L__BB0_4; $L__BB0_3: cvt.rn.f32.u16 %f247, %rs7; div.rn.f32 %f248, %f247, 0f437F0000; fma.rn.f32 %f249, %f248, 0f40000000, 0fBF800000; and.b16 %rs13, %rs61, 255; cvt.rn.f32.u16 %f250, %rs13; div.rn.f32 %f251, %f250, 0f437F0000; fma.rn.f32 %f252, %f251, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f253, %rs60; div.rn.f32 %f254, %f253, 0f437F0000; fma.rn.f32 %f255, %f254, 0f40000000, 0fBF800000; mul.f32 %f256, %f252, %f252; fma.rn.f32 %f257, %f249, %f249, %f256; fma.rn.f32 %f258, %f255, %f255, %f257; sqrt.rn.f32 %f259, %f258; rcp.rn.f32 %f260, %f259; mul.f32 %f1223, %f260, %f255; mul.f32 %f1222, %f260, %f252; mul.f32 %f1221, %f249, %f260; $L__BB0_4: ld.const.v2.u32 {%r66, %r67}, [params]; add.s32 %r6, %r66, %r58; add.s32 %r7, %r67, %r59; setp.eq.f32 %p12, %f1221, 0f00000000; setp.eq.f32 %p13, %f1222, 0f00000000; and.pred %p14, %p12, %p13; setp.eq.f32 %p15, %f1223, 0f00000000; and.pred %p16, %p15, %p14; @%p16 bra $L__BB0_150; bra.uni $L__BB0_5; $L__BB0_150: ld.const.u32 %r57, [params+104]; and.b32 %r426, %r57, 1; setp.eq.b32 %p185, %r426, 1; mov.pred %p186, 0; xor.pred %p187, %p185, %p186; not.pred %p188, %p187; @%p188 bra $L__BB0_152; ld.const.u64 %rd81, [params+144]; cvta.to.global.u64 %rd82, %rd81; ld.const.u32 %r427, [params+136]; mad.lo.s32 %r428, %r427, %r7, %r6; mul.wide.u32 %rd83, %r428, 4; add.s64 %rd84, %rd82, %rd83; mov.u16 %rs41, 0; st.global.v4.u8 [%rd84], {%rs41, %rs41, %rs41, %rs41}; $L__BB0_152: and.b32 %r429, %r57, 8; setp.eq.s32 %p189, %r429, 0; @%p189 bra $L__BB0_154; ld.const.u64 %rd85, [params+192]; cvta.to.global.u64 %rd86, %rd85; ld.const.u32 %r430, [params+184]; mad.lo.s32 %r431, %r430, %r7, %r6; mov.f32 %f1203, 0f00000000; cvt.rzi.u32.f32 %r432, %f1203; mul.wide.u32 %rd87, %r431, 2; add.s64 %rd88, %rd86, %rd87; mov.u16 %rs42, 0; cvt.u16.u32 %rs43, %r432; st.global.v2.u8 [%rd88], {%rs43, %rs42}; $L__BB0_154: and.b32 %r433, %r57, 4; setp.eq.s32 %p190, %r433, 0; @%p190 bra $L__BB0_158; ld.const.u32 %r434, [params+108]; setp.eq.s32 %p191, %r434, 0; ld.const.u64 %rd89, [params+224]; cvta.to.global.u64 %rd90, %rd89; ld.const.u32 %r435, [params+216]; mad.lo.s32 %r436, %r435, %r7, %r6; mul.wide.u32 %rd91, %r436, 8; add.s64 %rd24, %rd90, %rd91; @%p191 bra $L__BB0_157; ld.global.v4.u16 {%rs50, %rs51, %rs52, %rs53}, [%rd24]; // begin inline asm { cvt.f32.f16 %f1204, %rs50;} // end inline asm // begin inline asm { cvt.f32.f16 %f1205, %rs51;} // end inline asm // begin inline asm { cvt.f32.f16 %f1206, %rs52;} // end inline asm add.f32 %f1207, %f1204, 0f00000000; add.f32 %f1208, %f1205, 0f00000000; add.f32 %f1209, %f1206, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs49, %f1209;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs48, %f1208;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs47, %f1207;} // end inline asm mov.u16 %rs54, 0; st.global.v4.u16 [%rd24], {%rs47, %rs48, %rs49, %rs54}; bra.uni $L__BB0_158; $L__BB0_5: ld.const.u64 %rd29, [params+432]; cvta.to.global.u64 %rd30, %rd29; ld.const.u32 %r71, [params+424]; mad.lo.s32 %r72, %r71, %r59, %r58; mul.wide.u32 %rd31, %r72, 12; add.s64 %rd32, %rd30, %rd31; ld.global.f32 %f10, [%rd32]; mul.f32 %f268, %f10, 0f3456BF95; ld.global.f32 %f11, [%rd32+4]; mul.f32 %f269, %f11, 0f3456BF95; ld.global.f32 %f12, [%rd32+8]; mul.f32 %f270, %f12, 0f3456BF95; abs.f32 %f13, %f1221; div.rn.f32 %f271, %f268, %f13; abs.f32 %f272, %f1222; div.rn.f32 %f273, %f269, %f272; abs.f32 %f14, %f1223; div.rn.f32 %f274, %f270, %f14; abs.f32 %f275, %f271; abs.f32 %f276, %f273; abs.f32 %f277, %f274; mov.f32 %f278, 0f38D1B717; max.f32 %f279, %f275, %f278; max.f32 %f280, %f276, %f278; max.f32 %f281, %f277, %f278; fma.rn.f32 %f15, %f1221, %f279, %f10; fma.rn.f32 %f16, %f1222, %f280, %f11; fma.rn.f32 %f17, %f1223, %f281, %f12; ld.const.u64 %rd33, [params+128]; cvta.to.global.u64 %rd34, %rd33; ld.const.u32 %r73, [params+120]; mad.lo.s32 %r74, %r73, %r59, %r58; mul.wide.u32 %rd35, %r74, 4; add.s64 %rd36, %rd34, %rd35; ld.global.u32 %r448, [%rd36]; setp.lt.s32 %p18, %r4, 1; mov.pred %p17, 0; mov.u32 %r444, 0; mov.f32 %f1234, 0f00000000; mov.f32 %f1235, %f1234; mov.f32 %f1236, %f1234; mov.f32 %f1237, %f1234; mov.f32 %f1238, %f1234; mov.f32 %f1239, %f1234; mov.f32 %f1240, %f1234; mov.pred %p193, %p17; @%p18 bra $L__BB0_25; ld.const.f32 %f18, [params+620]; ld.const.u64 %rd37, [params+624]; cvta.to.global.u64 %rd3, %rd37; ld.const.u64 %rd38, [params+640]; cvta.to.global.u64 %rd39, %rd38; ld.const.u32 %r77, [params+632]; and.b32 %r78, %r59, 255; and.b32 %r79, %r58, 255; mad.lo.s32 %r80, %r77, %r78, %r79; mul.wide.u32 %rd40, %r80, 3; add.s64 %rd4, %rd39, %rd40; ld.const.f32 %f19, [params+660]; mul.f32 %f20, %f15, 0f3456BF95; mul.f32 %f21, %f16, 0f3456BF95; mul.f32 %f22, %f17, 0f3456BF95; ld.const.u64 %rd5, [params+96]; abs.f32 %f417, %f21; abs.f32 %f418, %f20; max.f32 %f419, %f418, %f417; abs.f32 %f420, %f22; max.f32 %f421, %f419, %f420; mov.u32 %r441, %r444; $L__BB0_7: mul.lo.s32 %r81, %r441, 3; mul.wide.s32 %rd41, %r81, 12; add.s64 %rd6, %rd3, %rd41; ld.global.f32 %f289, [%rd6]; sub.f32 %f290, %f289, %f10; ld.global.f32 %f291, [%rd6+4]; sub.f32 %f292, %f291, %f11; ld.global.f32 %f293, [%rd6+8]; sub.f32 %f294, %f293, %f12; mul.f32 %f295, %f292, %f292; fma.rn.f32 %f296, %f290, %f290, %f295; fma.rn.f32 %f297, %f294, %f294, %f296; sqrt.rn.f32 %f30, %f297; rcp.rn.f32 %f298, %f30; mul.f32 %f31, %f290, %f298; mul.f32 %f32, %f292, %f298; mul.f32 %f33, %f294, %f298; mul.f32 %f299, %f1222, %f32; fma.rn.f32 %f300, %f1221, %f31, %f299; fma.rn.f32 %f34, %f1223, %f33, %f300; setp.leu.f32 %p19, %f34, 0f00000000; @%p19 bra $L__BB0_24; setp.ne.s32 %p21, %r5, 0; mul.f32 %f301, %f30, %f30; div.rn.f32 %f35, %f18, %f301; ld.global.u8 %rs15, [%rd4]; cvt.rn.f32.u16 %f302, %rs15; div.rn.f32 %f303, %f302, 0fC37F0000; fma.rn.f32 %f304, %f303, 0f3F333333, 0f3F800000; setp.gt.f32 %p22, %f35, %f304; and.pred %p23, %p21, %p22; mov.pred %p193, -1; @%p23 bra $L__BB0_25; mul.f32 %f36, %f30, %f19; mov.f32 %f309, 0f40800000; abs.f32 %f38, %f36; setp.lt.f32 %p24, %f38, 0f00800000; mul.f32 %f311, %f38, 0f4B800000; selp.f32 %f312, %f311, %f38, %p24; selp.f32 %f313, 0fC3170000, 0fC2FE0000, %p24; mov.b32 %r82, %f312; and.b32 %r83, %r82, 8388607; or.b32 %r84, %r83, 1065353216; mov.b32 %f314, %r84; shr.u32 %r85, %r82, 23; cvt.rn.f32.u32 %f315, %r85; add.f32 %f316, %f313, %f315; setp.gt.f32 %p25, %f314, 0f3FB504F3; mul.f32 %f317, %f314, 0f3F000000; add.f32 %f318, %f316, 0f3F800000; selp.f32 %f319, %f318, %f316, %p25; selp.f32 %f320, %f317, %f314, %p25; add.f32 %f321, %f320, 0fBF800000; add.f32 %f322, %f320, 0f3F800000; rcp.approx.ftz.f32 %f323, %f322; add.f32 %f324, %f321, %f321; mul.f32 %f325, %f324, %f323; mul.f32 %f326, %f325, %f325; mov.f32 %f327, 0f3C4CAF63; mov.f32 %f328, 0f3B18F0FE; fma.rn.f32 %f329, %f328, %f326, %f327; mov.f32 %f330, 0f3DAAAABD; fma.rn.f32 %f331, %f329, %f326, %f330; mul.rn.f32 %f332, %f331, %f326; mul.rn.f32 %f333, %f332, %f325; sub.f32 %f334, %f321, %f325; add.f32 %f335, %f334, %f334; neg.f32 %f336, %f325; fma.rn.f32 %f337, %f336, %f321, %f335; mul.rn.f32 %f338, %f323, %f337; add.f32 %f339, %f333, %f325; sub.f32 %f340, %f325, %f339; add.f32 %f341, %f333, %f340; add.f32 %f342, %f338, %f341; add.f32 %f343, %f339, %f342; sub.f32 %f344, %f339, %f343; add.f32 %f345, %f342, %f344; mov.f32 %f346, 0f3F317200; mul.rn.f32 %f347, %f319, %f346; mov.f32 %f348, 0f35BFBE8E; mul.rn.f32 %f349, %f319, %f348; add.f32 %f350, %f347, %f343; sub.f32 %f351, %f347, %f350; add.f32 %f352, %f343, %f351; add.f32 %f353, %f345, %f352; add.f32 %f354, %f349, %f353; add.f32 %f355, %f350, %f354; sub.f32 %f356, %f350, %f355; add.f32 %f357, %f354, %f356; mul.rn.f32 %f358, %f309, %f355; neg.f32 %f359, %f358; fma.rn.f32 %f360, %f309, %f355, %f359; fma.rn.f32 %f361, %f309, %f357, %f360; mov.f32 %f362, 0f00000000; fma.rn.f32 %f363, %f362, %f355, %f361; add.rn.f32 %f364, %f358, %f363; neg.f32 %f365, %f364; add.rn.f32 %f366, %f358, %f365; add.rn.f32 %f367, %f366, %f363; mov.b32 %r86, %f364; setp.eq.s32 %p26, %r86, 1118925336; add.s32 %r87, %r86, -1; mov.b32 %f368, %r87; add.f32 %f369, %f367, 0f37000000; selp.f32 %f39, %f369, %f367, %p26; selp.f32 %f370, %f368, %f364, %p26; mov.f32 %f371, 0f3FB8AA3B; mul.rn.f32 %f372, %f370, %f371; cvt.rzi.f32.f32 %f373, %f372; abs.f32 %f374, %f373; setp.gt.f32 %p27, %f374, 0f42FC0000; mov.b32 %r88, %f373; and.b32 %r89, %r88, -2147483648; or.b32 %r90, %r89, 1123811328; mov.b32 %f375, %r90; selp.f32 %f376, %f375, %f373, %p27; mov.f32 %f377, 0fBF317218; fma.rn.f32 %f378, %f376, %f377, %f370; mov.f32 %f379, 0f3102E308; fma.rn.f32 %f380, %f376, %f379, %f378; mul.f32 %f381, %f380, 0f3FB8AA3B; add.f32 %f382, %f376, 0f4B40007F; mov.b32 %r91, %f382; shl.b32 %r92, %r91, 23; mov.b32 %f383, %r92; ex2.approx.ftz.f32 %f384, %f381; mul.f32 %f40, %f384, %f383; setp.eq.f32 %p28, %f40, 0f7F800000; mov.f32 %f1231, 0f7F800000; @%p28 bra $L__BB0_11; fma.rn.f32 %f1231, %f40, %f39, %f40; $L__BB0_11: mov.f32 %f1218, 0f40000000; cvt.rzi.f32.f32 %f1217, %f1218; add.f32 %f1216, %f1217, %f1217; mov.f32 %f1215, 0f40800000; sub.f32 %f1214, %f1215, %f1216; abs.f32 %f1213, %f1214; setp.lt.f32 %p29, %f36, 0f00000000; setp.eq.f32 %p30, %f1213, 0f3F800000; and.pred %p1, %p29, %p30; setp.eq.f32 %p31, %f36, 0f00000000; @%p31 bra $L__BB0_15; bra.uni $L__BB0_12; $L__BB0_15: add.f32 %f389, %f36, %f36; selp.f32 %f1233, %f389, 0f00000000, %p30; bra.uni $L__BB0_16; $L__BB0_12: mov.b32 %r93, %f1231; xor.b32 %r94, %r93, -2147483648; mov.b32 %f385, %r94; selp.f32 %f1233, %f385, %f1231, %p1; setp.geu.f32 %p32, %f36, 0f00000000; @%p32 bra $L__BB0_16; mov.f32 %f386, 0f40800000; cvt.rzi.f32.f32 %f387, %f386; setp.eq.f32 %p33, %f387, 0f40800000; @%p33 bra $L__BB0_16; mov.f32 %f1233, 0f7FFFFFFF; $L__BB0_16: add.f32 %f390, %f38, 0f40800000; mov.b32 %r95, %f390; setp.lt.s32 %p35, %r95, 2139095040; @%p35 bra $L__BB0_21; setp.gtu.f32 %p36, %f38, 0f7F800000; @%p36 bra $L__BB0_20; bra.uni $L__BB0_18; $L__BB0_20: add.f32 %f1233, %f36, 0f40800000; bra.uni $L__BB0_21; $L__BB0_18: setp.neu.f32 %p37, %f38, 0f7F800000; @%p37 bra $L__BB0_21; selp.f32 %f1233, 0fFF800000, 0f7F800000, %p1; $L__BB0_21: mov.f32 %f391, 0f3F800000; sub.f32 %f392, %f391, %f1233; setp.eq.f32 %p38, %f36, 0f3F800000; selp.f32 %f393, 0f00000000, %f392, %p38; cvt.sat.f32.f32 %f394, %f393; mul.f32 %f395, %f35, %f394; ld.global.f32 %f396, [%rd6+12]; mul.f32 %f397, %f31, %f396; ld.global.f32 %f398, [%rd6+16]; mul.f32 %f399, %f32, %f398; neg.f32 %f400, %f399; sub.f32 %f401, %f400, %f397; ld.global.f32 %f402, [%rd6+20]; mul.f32 %f403, %f33, %f402; sub.f32 %f404, %f401, %f403; cvt.sat.f32.f32 %f405, %f404; mul.f32 %f49, %f395, %f405; cvt.sat.f32.f32 %f50, %f34; mul.f32 %f406, %f49, %f50; setp.leu.f32 %p39, %f406, 0f3727C5AC; @%p39 bra $L__BB0_23; mul.f32 %f416, %f49, 0f3EA2F983; mov.f32 %f422, 0f38D1B717; max.f32 %f413, %f421, %f422; sub.f32 %f414, %f30, %f413; mov.f32 %f415, 0f00000000; mov.u32 %r132, 2; mov.u32 %r133, 1; mov.u32 %r134, 3; mov.u32 %r137, 1065353216; mov.u32 %r166, 0; // begin inline asm call(%r96,%r97,%r98,%r99,%r100,%r101,%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),_optix_trace_typed_32,(%r166,%rd5,%f15,%f16,%f17,%f31,%f32,%f33,%f413,%f414,%f415,%r133,%r166,%r133,%r132,%r133,%r134,%r137,%r137,%r137,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166,%r166); // end inline asm mov.b32 %f423, %r96; mov.b32 %f424, %r97; mov.b32 %f425, %r98; ld.global.f32 %f426, [%rd6+24]; mul.f32 %f427, %f426, %f423; ld.global.f32 %f428, [%rd6+28]; mul.f32 %f429, %f428, %f424; ld.global.f32 %f430, [%rd6+32]; mul.f32 %f431, %f430, %f425; mul.f32 %f432, %f416, %f427; mul.f32 %f433, %f416, %f429; mul.f32 %f434, %f416, %f431; fma.rn.f32 %f1237, %f50, %f432, %f1237; fma.rn.f32 %f1238, %f50, %f433, %f1238; fma.rn.f32 %f1239, %f50, %f434, %f1239; add.f32 %f1240, %f1240, %f423; fma.rn.f32 %f1234, %f31, %f423, %f1234; fma.rn.f32 %f1235, %f32, %f424, %f1235; fma.rn.f32 %f1236, %f33, %f425, %f1236; $L__BB0_23: add.s32 %r444, %r444, 1; $L__BB0_24: add.s32 %r441, %r441, 1; setp.lt.s32 %p41, %r441, %r4; mov.pred %p193, %p17; @%p41 bra $L__BB0_7; $L__BB0_25: cvt.rn.f32.s32 %f435, %r444; mov.f32 %f436, 0f3F800000; max.f32 %f437, %f435, %f436; rcp.rn.f32 %f438, %f437; mul.f32 %f1282, %f1237, %f438; mul.f32 %f1283, %f1238, %f438; mul.f32 %f1284, %f1239, %f438; div.rn.f32 %f1285, %f1240, %f437; not.pred %p42, %p193; @%p42 bra $L__BB0_56; abs.f32 %f1220, %f1223; abs.f32 %f1219, %f1221; setp.gt.f32 %p43, %f1219, %f1220; selp.f32 %f445, 0f00000000, %f1222, %p43; mov.f32 %f444, 0f00000000; neg.f32 %f446, %f1223; selp.f32 %f447, %f1221, %f446, %p43; neg.f32 %f448, %f1222; selp.f32 %f449, %f448, 0f00000000, %p43; mul.f32 %f450, %f447, %f447; fma.rn.f32 %f451, %f449, %f449, %f450; fma.rn.f32 %f452, %f445, %f445, %f451; sqrt.rn.f32 %f453, %f452; rcp.rn.f32 %f454, %f453; mul.f32 %f83, %f449, %f454; mul.f32 %f84, %f447, %f454; mul.f32 %f85, %f445, %f454; setp.lt.s32 %p44, %r5, 1; mov.f32 %f1234, %f444; mov.f32 %f1235, %f444; mov.f32 %f1236, %f444; mov.f32 %f1263, %f444; mov.f32 %f1262, %f444; mov.f32 %f1261, %f444; @%p44 bra $L__BB0_55; cvt.rn.f32.s32 %f461, %r5; rcp.rn.f32 %f86, %f461; mul.f32 %f87, %f15, 0f3456BF95; mul.f32 %f88, %f16, 0f3456BF95; mul.f32 %f89, %f17, 0f3456BF95; ld.const.u64 %rd7, [params+96]; mul.f32 %f462, %f1221, %f84; mul.f32 %f463, %f1222, %f83; sub.f32 %f90, %f463, %f462; mul.f32 %f464, %f1223, %f83; mul.f32 %f465, %f1221, %f85; sub.f32 %f91, %f465, %f464; mul.f32 %f466, %f1222, %f85; mul.f32 %f467, %f1223, %f84; sub.f32 %f92, %f467, %f466; mov.u32 %r167, 0; add.s64 %rd8, %rd1, 24; mov.u64 %rd43, __cudart_i2opi_f; abs.f32 %f530, %f88; abs.f32 %f531, %f87; max.f32 %f532, %f531, %f530; abs.f32 %f533, %f89; max.f32 %f534, %f532, %f533; mov.u32 %r445, %r167; $L__BB0_28: cvt.rn.f32.s32 %f99, %r445; mov.u32 %r447, %r167; $L__BB0_29: mad.lo.s32 %r169, %r448, 1664525, 1013904223; and.b32 %r170, %r169, 16777215; cvt.rn.f32.u32 %f468, %r170; fma.rn.f32 %f469, %f468, 0f33800000, %f99; mul.f32 %f106, %f86, %f469; mad.lo.s32 %r448, %r169, 1664525, 1013904223; and.b32 %r171, %r448, 16777215; cvt.rn.f32.u32 %f470, %r171; cvt.rn.f32.s32 %f471, %r447; fma.rn.f32 %f472, %f470, 0f33800000, %f471; mul.f32 %f473, %f86, %f472; mul.f32 %f474, %f106, %f106; mov.f32 %f475, 0f3F800000; sub.f32 %f476, %f475, %f474; mov.f32 %f477, 0f00000000; max.f32 %f478, %f477, %f476; sqrt.rn.f32 %f107, %f478; mul.f32 %f108, %f473, 0f40C90FDB; mul.f32 %f479, %f108, 0f3F22F983; cvt.rni.s32.f32 %r455, %f479; cvt.rn.f32.s32 %f480, %r455; mov.f32 %f481, 0fBFC90FDA; fma.rn.f32 %f482, %f480, %f481, %f108; mov.f32 %f483, 0fB3A22168; fma.rn.f32 %f484, %f480, %f483, %f482; mov.f32 %f485, 0fA7C234C5; fma.rn.f32 %f1270, %f480, %f485, %f484; abs.f32 %f110, %f108; setp.ltu.f32 %p45, %f110, 0f47CE4780; mov.u32 %r452, %r455; mov.f32 %f1267, %f1270; @%p45 bra $L__BB0_37; setp.eq.f32 %p46, %f110, 0f7F800000; @%p46 bra $L__BB0_36; bra.uni $L__BB0_31; $L__BB0_36: mov.f32 %f488, 0f00000000; mul.rn.f32 %f1267, %f108, %f488; mov.u32 %r452, 0; bra.uni $L__BB0_37; $L__BB0_31: mov.b32 %r21, %f108; bfe.u32 %r173, %r21, 23, 8; add.s32 %r22, %r173, -128; shl.b32 %r174, %r21, 8; or.b32 %r23, %r174, -2147483648; shr.u32 %r24, %r22, 5; mov.u64 %rd98, 0; mov.u32 %r449, 0; mov.u64 %rd96, %rd1; mov.u64 %rd97, %rd43; $L__BB0_32: .pragma "nounroll"; ld.global.nc.u32 %r175, [%rd97]; mad.wide.u32 %rd45, %r175, %r23, %rd98; shr.u64 %rd98, %rd45, 32; st.local.u32 [%rd96], %rd45; add.s64 %rd97, %rd97, 4; add.s64 %rd96, %rd96, 4; add.s32 %r449, %r449, 1; setp.ne.s32 %p47, %r449, 6; @%p47 bra $L__BB0_32; st.local.u32 [%rd8], %rd98; mov.u32 %r176, 4; sub.s32 %r27, %r176, %r24; mov.u32 %r177, 6; sub.s32 %r178, %r177, %r24; mul.wide.s32 %rd46, %r178, 4; add.s64 %rd47, %rd1, %rd46; ld.local.u32 %r450, [%rd47]; ld.local.u32 %r451, [%rd47+-4]; and.b32 %r30, %r22, 31; setp.eq.s32 %p48, %r30, 0; @%p48 bra $L__BB0_35; mov.u32 %r179, 32; sub.s32 %r180, %r179, %r30; shr.u32 %r181, %r451, %r180; shl.b32 %r182, %r450, %r30; add.s32 %r450, %r181, %r182; mul.wide.s32 %rd48, %r27, 4; add.s64 %rd49, %rd1, %rd48; ld.local.u32 %r183, [%rd49]; shr.u32 %r184, %r183, %r180; shl.b32 %r185, %r451, %r30; add.s32 %r451, %r184, %r185; $L__BB0_35: and.b32 %r186, %r21, -2147483648; shr.u32 %r187, %r451, 30; shl.b32 %r188, %r450, 2; or.b32 %r189, %r187, %r188; shr.u32 %r190, %r189, 31; shr.u32 %r191, %r450, 30; add.s32 %r192, %r190, %r191; neg.s32 %r193, %r192; setp.eq.s32 %p49, %r186, 0; selp.b32 %r452, %r192, %r193, %p49; setp.ne.s32 %p50, %r190, 0; xor.b32 %r194, %r186, -2147483648; selp.b32 %r195, %r194, %r186, %p50; selp.b32 %r196, -1, 0, %p50; xor.b32 %r197, %r189, %r196; shl.b32 %r198, %r451, 2; xor.b32 %r199, %r198, %r196; cvt.u64.u32 %rd50, %r197; cvt.u64.u32 %rd51, %r199; bfi.b64 %rd52, %rd50, %rd51, 32, 32; cvt.rn.f64.s64 %fd1, %rd52; mul.f64 %fd2, %fd1, 0d3BF921FB54442D19; cvt.rn.f32.f64 %f486, %fd2; setp.eq.s32 %p51, %r195, 0; neg.f32 %f487, %f486; selp.f32 %f1267, %f486, %f487, %p51; $L__BB0_37: add.s32 %r37, %r452, 1; and.b32 %r38, %r37, 1; setp.eq.s32 %p52, %r38, 0; selp.f32 %f114, %f1267, 0f3F800000, %p52; mul.rn.f32 %f115, %f1267, %f1267; mov.f32 %f1268, 0fB94D4153; @%p52 bra $L__BB0_39; mov.f32 %f490, 0fBAB607ED; mov.f32 %f491, 0f37CBAC00; fma.rn.f32 %f1268, %f491, %f115, %f490; $L__BB0_39: selp.f32 %f492, 0f3C0885E4, 0f3D2AAABB, %p52; fma.rn.f32 %f493, %f1268, %f115, %f492; selp.f32 %f494, 0fBE2AAAA8, 0fBEFFFFFF, %p52; fma.rn.f32 %f495, %f493, %f115, %f494; mov.f32 %f496, 0f00000000; fma.rn.f32 %f497, %f115, %f114, %f496; fma.rn.f32 %f1269, %f495, %f497, %f114; and.b32 %r201, %r37, 2; setp.eq.s32 %p54, %r201, 0; @%p54 bra $L__BB0_41; mov.f32 %f499, 0fBF800000; fma.rn.f32 %f1269, %f1269, %f499, %f496; $L__BB0_41: @%p45 bra $L__BB0_49; setp.eq.f32 %p56, %f110, 0f7F800000; @%p56 bra $L__BB0_48; bra.uni $L__BB0_43; $L__BB0_48: mov.f32 %f502, 0f00000000; mul.rn.f32 %f1270, %f108, %f502; mov.u32 %r455, 0; bra.uni $L__BB0_49; $L__BB0_43: mov.b32 %r39, %f108; bfe.u32 %r202, %r39, 23, 8; add.s32 %r40, %r202, -128; shl.b32 %r203, %r39, 8; or.b32 %r41, %r203, -2147483648; shr.u32 %r42, %r40, 5; mov.u64 %rd99, 0; mov.u64 %rd100, %rd99; $L__BB0_44: .pragma "nounroll"; shl.b64 %rd55, %rd99, 2; mov.u64 %rd56, __cudart_i2opi_f; add.s64 %rd57, %rd56, %rd55; ld.global.nc.u32 %r204, [%rd57]; mad.wide.u32 %rd58, %r204, %r41, %rd100; shr.u64 %rd100, %rd58, 32; add.s64 %rd59, %rd1, %rd55; st.local.u32 [%rd59], %rd58; cvt.u32.u64 %r205, %rd99; add.s32 %r206, %r205, 1; cvt.s64.s32 %rd99, %r206; setp.ne.s32 %p57, %r206, 6; @%p57 bra $L__BB0_44; st.local.u32 [%rd8], %rd100; mov.u32 %r207, 4; sub.s32 %r43, %r207, %r42; mov.u32 %r208, 6; sub.s32 %r209, %r208, %r42; mul.wide.s32 %rd60, %r209, 4; add.s64 %rd61, %rd1, %rd60; ld.local.u32 %r453, [%rd61]; ld.local.u32 %r454, [%rd61+-4]; and.b32 %r46, %r40, 31; setp.eq.s32 %p58, %r46, 0; @%p58 bra $L__BB0_47; mov.u32 %r210, 32; sub.s32 %r211, %r210, %r46; shr.u32 %r212, %r454, %r211; shl.b32 %r213, %r453, %r46; add.s32 %r453, %r212, %r213; mul.wide.s32 %rd62, %r43, 4; add.s64 %rd63, %rd1, %rd62; ld.local.u32 %r214, [%rd63]; shr.u32 %r215, %r214, %r211; shl.b32 %r216, %r454, %r46; add.s32 %r454, %r215, %r216; $L__BB0_47: and.b32 %r217, %r39, -2147483648; shr.u32 %r218, %r454, 30; shl.b32 %r219, %r453, 2; or.b32 %r220, %r218, %r219; shr.u32 %r221, %r220, 31; shr.u32 %r222, %r453, 30; add.s32 %r223, %r221, %r222; neg.s32 %r224, %r223; setp.eq.s32 %p59, %r217, 0; selp.b32 %r455, %r223, %r224, %p59; setp.ne.s32 %p60, %r221, 0; xor.b32 %r225, %r217, -2147483648; selp.b32 %r226, %r225, %r217, %p60; selp.b32 %r227, -1, 0, %p60; xor.b32 %r228, %r220, %r227; shl.b32 %r229, %r454, 2; xor.b32 %r230, %r229, %r227; cvt.u64.u32 %rd64, %r228; cvt.u64.u32 %rd65, %r230; bfi.b64 %rd66, %rd64, %rd65, 32, 32; cvt.rn.f64.s64 %fd3, %rd66; mul.f64 %fd4, %fd3, 0d3BF921FB54442D19; cvt.rn.f32.f64 %f500, %fd4; setp.eq.s32 %p61, %r226, 0; neg.f32 %f501, %f500; selp.f32 %f1270, %f500, %f501, %p61; $L__BB0_49: mul.f32 %f124, %f107, %f1269; and.b32 %r53, %r455, 1; setp.eq.s32 %p62, %r53, 0; selp.f32 %f125, %f1270, 0f3F800000, %p62; mul.rn.f32 %f126, %f1270, %f1270; mov.f32 %f1271, 0fB94D4153; @%p62 bra $L__BB0_51; mov.f32 %f504, 0fBAB607ED; mov.f32 %f505, 0f37CBAC00; fma.rn.f32 %f1271, %f505, %f126, %f504; $L__BB0_51: selp.f32 %f506, 0f3C0885E4, 0f3D2AAABB, %p62; fma.rn.f32 %f507, %f1271, %f126, %f506; selp.f32 %f508, 0fBE2AAAA8, 0fBEFFFFFF, %p62; fma.rn.f32 %f509, %f507, %f126, %f508; mov.f32 %f510, 0f00000000; fma.rn.f32 %f511, %f126, %f125, %f510; fma.rn.f32 %f1272, %f509, %f511, %f125; and.b32 %r232, %r455, 2; setp.eq.s32 %p64, %r232, 0; @%p64 bra $L__BB0_53; mov.f32 %f513, 0fBF800000; fma.rn.f32 %f1272, %f1272, %f513, %f510; $L__BB0_53: mul.f32 %f523, %f107, %f1272; mul.f32 %f524, %f83, %f523; mul.f32 %f525, %f84, %f523; mul.f32 %f526, %f85, %f523; fma.rn.f32 %f527, %f92, %f124, %f524; fma.rn.f32 %f528, %f91, %f124, %f525; fma.rn.f32 %f529, %f90, %f124, %f526; fma.rn.f32 %f517, %f1221, %f106, %f527; fma.rn.f32 %f518, %f1222, %f106, %f528; fma.rn.f32 %f519, %f1223, %f106, %f529; mov.f32 %f535, 0f38D1B717; max.f32 %f520, %f534, %f535; mov.f32 %f521, 0f6C4ECB8F; mov.u32 %r266, 1; mov.u32 %r269, 2; mov.u32 %r271, 4; mov.u32 %r274, 1065353216; mov.u32 %r275, 2139095039; mov.u32 %r303, 0; // begin inline asm call(%r233,%r234,%r235,%r236,%r237,%r238,%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),_optix_trace_typed_32,(%r303,%rd7,%f15,%f16,%f17,%f517,%f518,%f519,%f520,%f521,%f510,%r266,%r303,%r303,%r269,%r303,%r271,%r274,%r274,%r274,%r275,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303,%r303); // end inline asm mov.b32 %f536, %r233; mov.b32 %f537, %r234; mov.b32 %f538, %r235; mov.b32 %f539, %r236; max.f32 %f540, %f536, %f510; max.f32 %f541, %f537, %f510; max.f32 %f542, %f538, %f510; setp.ltu.f32 %p65, %f539, 0f00000000; selp.f32 %f543, %f542, 0f00000000, %p65; selp.f32 %f544, %f541, 0f00000000, %p65; selp.f32 %f545, %f540, 0f00000000, %p65; selp.f32 %f546, 0f3F800000, 0f00000000, %p65; fma.rn.f32 %f1234, %f517, %f546, %f1234; fma.rn.f32 %f1235, %f518, %f546, %f1235; fma.rn.f32 %f1236, %f519, %f546, %f1236; mul.f32 %f547, %f1222, %f518; fma.rn.f32 %f548, %f1221, %f517, %f547; fma.rn.f32 %f549, %f1223, %f519, %f548; cvt.sat.f32.f32 %f550, %f549; fma.rn.f32 %f1263, %f545, %f550, %f1263; fma.rn.f32 %f1262, %f544, %f550, %f1262; fma.rn.f32 %f1261, %f543, %f550, %f1261; add.s32 %r447, %r447, 1; setp.lt.s32 %p66, %r447, %r5; @%p66 bra $L__BB0_29; add.s32 %r445, %r445, 1; setp.lt.s32 %p67, %r445, %r5; @%p67 bra $L__BB0_28; $L__BB0_55: mul.lo.s32 %r304, %r5, %r5; cvt.rn.f32.s32 %f551, %r304; rcp.rn.f32 %f552, %f551; mul.f32 %f553, %f552, %f1263; mul.f32 %f554, %f552, %f1262; mul.f32 %f555, %f552, %f1261; div.rn.f32 %f1285, %f444, %f551; fma.rn.f32 %f1282, %f552, %f1263, %f553; fma.rn.f32 %f1283, %f552, %f1262, %f554; fma.rn.f32 %f1284, %f552, %f1261, %f555; $L__BB0_56: ld.const.u32 %r440, [params+616]; setp.lt.s32 %p68, %r440, 0; selp.f32 %f155, %f1282, %f1285, %p68; ld.const.u32 %r56, [params+104]; and.b32 %r305, %r56, 8; setp.eq.s32 %p69, %r305, 0; @%p69 bra $L__BB0_70; ld.const.u64 %rd68, [params+192]; cvta.to.global.u64 %rd19, %rd68; ld.const.u32 %r306, [params+184]; mad.lo.s32 %r307, %r306, %r7, %r6; cvt.u64.u32 %rd20, %r307; mov.f32 %f558, 0f3E68BA2E; cvt.rzi.f32.f32 %f559, %f558; add.f32 %f560, %f559, %f559; mov.f32 %f561, 0f3EE8BA2E; sub.f32 %f562, %f561, %f560; abs.f32 %f156, %f562; abs.f32 %f157, %f155; setp.lt.f32 %p70, %f157, 0f00800000; mul.f32 %f563, %f157, 0f4B800000; selp.f32 %f564, %f563, %f157, %p70; selp.f32 %f565, 0fC3170000, 0fC2FE0000, %p70; mov.b32 %r308, %f564; and.b32 %r309, %r308, 8388607; or.b32 %r310, %r309, 1065353216; mov.b32 %f566, %r310; shr.u32 %r311, %r308, 23; cvt.rn.f32.u32 %f567, %r311; add.f32 %f568, %f565, %f567; setp.gt.f32 %p71, %f566, 0f3FB504F3; mul.f32 %f569, %f566, 0f3F000000; add.f32 %f570, %f568, 0f3F800000; selp.f32 %f571, %f570, %f568, %p71; selp.f32 %f572, %f569, %f566, %p71; add.f32 %f573, %f572, 0fBF800000; add.f32 %f574, %f572, 0f3F800000; rcp.approx.ftz.f32 %f575, %f574; add.f32 %f576, %f573, %f573; mul.f32 %f577, %f576, %f575; mul.f32 %f578, %f577, %f577; mov.f32 %f579, 0f3C4CAF63; mov.f32 %f580, 0f3B18F0FE; fma.rn.f32 %f581, %f580, %f578, %f579; mov.f32 %f582, 0f3DAAAABD; fma.rn.f32 %f583, %f581, %f578, %f582; mul.rn.f32 %f584, %f583, %f578; mul.rn.f32 %f585, %f584, %f577; sub.f32 %f586, %f573, %f577; add.f32 %f587, %f586, %f586; neg.f32 %f588, %f577; fma.rn.f32 %f589, %f588, %f573, %f587; mul.rn.f32 %f590, %f575, %f589; add.f32 %f591, %f585, %f577; sub.f32 %f592, %f577, %f591; add.f32 %f593, %f585, %f592; add.f32 %f594, %f590, %f593; add.f32 %f595, %f591, %f594; sub.f32 %f596, %f591, %f595; add.f32 %f597, %f594, %f596; mov.f32 %f598, 0f3F317200; mul.rn.f32 %f599, %f571, %f598; mov.f32 %f600, 0f35BFBE8E; mul.rn.f32 %f601, %f571, %f600; add.f32 %f602, %f599, %f595; sub.f32 %f603, %f599, %f602; add.f32 %f604, %f595, %f603; add.f32 %f605, %f597, %f604; add.f32 %f606, %f601, %f605; add.f32 %f607, %f602, %f606; sub.f32 %f608, %f602, %f607; add.f32 %f609, %f606, %f608; mul.rn.f32 %f610, %f561, %f607; neg.f32 %f611, %f610; fma.rn.f32 %f612, %f561, %f607, %f611; fma.rn.f32 %f613, %f561, %f609, %f612; mov.f32 %f614, 0f00000000; fma.rn.f32 %f615, %f614, %f607, %f613; add.rn.f32 %f616, %f610, %f615; neg.f32 %f617, %f616; add.rn.f32 %f618, %f610, %f617; add.rn.f32 %f619, %f618, %f615; mov.b32 %r312, %f616; setp.eq.s32 %p72, %r312, 1118925336; add.s32 %r313, %r312, -1; mov.b32 %f620, %r313; add.f32 %f621, %f619, 0f37000000; selp.f32 %f158, %f621, %f619, %p72; selp.f32 %f622, %f620, %f616, %p72; mov.f32 %f623, 0f3FB8AA3B; mul.rn.f32 %f624, %f622, %f623; cvt.rzi.f32.f32 %f625, %f624; abs.f32 %f626, %f625; setp.gt.f32 %p73, %f626, 0f42FC0000; mov.b32 %r314, %f625; and.b32 %r315, %r314, -2147483648; or.b32 %r316, %r315, 1123811328; mov.b32 %f627, %r316; selp.f32 %f628, %f627, %f625, %p73; mov.f32 %f629, 0fBF317218; fma.rn.f32 %f630, %f628, %f629, %f622; mov.f32 %f631, 0f3102E308; fma.rn.f32 %f632, %f628, %f631, %f630; mul.f32 %f633, %f632, 0f3FB8AA3B; add.f32 %f634, %f628, 0f4B40007F; mov.b32 %r317, %f634; shl.b32 %r318, %r317, 23; mov.b32 %f635, %r318; ex2.approx.ftz.f32 %f636, %f633; mul.f32 %f159, %f636, %f635; setp.eq.f32 %p74, %f159, 0f7F800000; mov.f32 %f1286, 0f7F800000; @%p74 bra $L__BB0_59; fma.rn.f32 %f1286, %f159, %f158, %f159; $L__BB0_59: setp.lt.f32 %p75, %f155, 0f00000000; setp.eq.f32 %p76, %f156, 0f3F800000; and.pred %p3, %p75, %p76; setp.eq.f32 %p77, %f155, 0f00000000; @%p77 bra $L__BB0_63; bra.uni $L__BB0_60; $L__BB0_63: add.f32 %f641, %f155, %f155; selp.f32 %f1288, %f641, 0f00000000, %p76; bra.uni $L__BB0_64; $L__BB0_157: mov.f32 %f1212, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs57, %f1212;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs56, %f1212;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs55, %f1212;} // end inline asm mov.u16 %rs58, 0; st.global.v4.u16 [%rd24], {%rs55, %rs56, %rs57, %rs58}; $L__BB0_158: and.b32 %r437, %r57, 64; setp.eq.s32 %p192, %r437, 0; @%p192 bra $L__BB0_160; ld.const.u64 %rd92, [params+208]; cvta.to.global.u64 %rd93, %rd92; ld.const.u32 %r438, [params+200]; mad.lo.s32 %r439, %r438, %r7, %r6; mul.wide.u32 %rd94, %r439, 4; add.s64 %rd95, %rd93, %rd94; mov.u16 %rs59, 0; st.global.v4.u8 [%rd95], {%rs59, %rs59, %rs59, %rs59}; bra.uni $L__BB0_160; $L__BB0_60: mov.b32 %r319, %f1286; xor.b32 %r320, %r319, -2147483648; mov.b32 %f637, %r320; selp.f32 %f1288, %f637, %f1286, %p3; setp.geu.f32 %p78, %f155, 0f00000000; @%p78 bra $L__BB0_64; mov.f32 %f638, 0f3EE8BA2E; cvt.rzi.f32.f32 %f639, %f638; setp.eq.f32 %p79, %f639, 0f3EE8BA2E; @%p79 bra $L__BB0_64; mov.f32 %f1288, 0f7FFFFFFF; $L__BB0_64: add.f32 %f642, %f157, 0f3EE8BA2E; mov.b32 %r321, %f642; setp.lt.s32 %p81, %r321, 2139095040; @%p81 bra $L__BB0_69; setp.gtu.f32 %p82, %f157, 0f7F800000; @%p82 bra $L__BB0_68; bra.uni $L__BB0_66; $L__BB0_68: add.f32 %f1288, %f155, 0f3EE8BA2E; bra.uni $L__BB0_69; $L__BB0_66: setp.neu.f32 %p83, %f157, 0f7F800000; @%p83 bra $L__BB0_69; selp.f32 %f1288, 0fFF800000, 0f7F800000, %p3; $L__BB0_69: mul.f32 %f643, %f1288, 0f437F0000; setp.eq.f32 %p84, %f155, 0f3F800000; selp.f32 %f644, 0f437F0000, %f643, %p84; cvt.rzi.u32.f32 %r322, %f644; shl.b64 %rd69, %rd20, 1; add.s64 %rd70, %rd19, %rd69; cvt.u16.u32 %rs16, %r322; mov.u16 %rs17, 255; st.global.v2.u8 [%rd70], {%rs16, %rs17}; $L__BB0_70: ld.const.v2.f32 {%f645, %f646}, [params+648]; mul.f32 %f168, %f1282, %f645; mul.f32 %f169, %f1283, %f646; ld.const.f32 %f649, [params+656]; mul.f32 %f170, %f1284, %f649; and.b32 %r323, %r56, 1; setp.eq.b32 %p85, %r323, 1; mov.pred %p86, 0; xor.pred %p87, %p85, %p86; not.pred %p88, %p87; @%p88 bra $L__BB0_144; mov.f32 %f651, 0f3E666666; cvt.rzi.f32.f32 %f652, %f651; add.f32 %f653, %f652, %f652; mov.f32 %f654, 0f3EE66666; sub.f32 %f655, %f654, %f653; abs.f32 %f171, %f655; abs.f32 %f172, %f168; setp.lt.f32 %p89, %f172, 0f00800000; mul.f32 %f656, %f172, 0f4B800000; selp.f32 %f657, %f656, %f172, %p89; selp.f32 %f658, 0fC3170000, 0fC2FE0000, %p89; mov.b32 %r324, %f657; and.b32 %r325, %r324, 8388607; or.b32 %r326, %r325, 1065353216; mov.b32 %f659, %r326; shr.u32 %r327, %r324, 23; cvt.rn.f32.u32 %f660, %r327; add.f32 %f661, %f658, %f660; setp.gt.f32 %p90, %f659, 0f3FB504F3; mul.f32 %f662, %f659, 0f3F000000; add.f32 %f663, %f661, 0f3F800000; selp.f32 %f664, %f663, %f661, %p90; selp.f32 %f665, %f662, %f659, %p90; add.f32 %f666, %f665, 0fBF800000; add.f32 %f667, %f665, 0f3F800000; rcp.approx.ftz.f32 %f668, %f667; add.f32 %f669, %f666, %f666; mul.f32 %f670, %f669, %f668; mul.f32 %f671, %f670, %f670; mov.f32 %f672, 0f3C4CAF63; mov.f32 %f673, 0f3B18F0FE; fma.rn.f32 %f674, %f673, %f671, %f672; mov.f32 %f675, 0f3DAAAABD; fma.rn.f32 %f676, %f674, %f671, %f675; mul.rn.f32 %f677, %f676, %f671; mul.rn.f32 %f678, %f677, %f670; sub.f32 %f679, %f666, %f670; add.f32 %f680, %f679, %f679; neg.f32 %f681, %f670; fma.rn.f32 %f682, %f681, %f666, %f680; mul.rn.f32 %f683, %f668, %f682; add.f32 %f684, %f678, %f670; sub.f32 %f685, %f670, %f684; add.f32 %f686, %f678, %f685; add.f32 %f687, %f683, %f686; add.f32 %f688, %f684, %f687; sub.f32 %f689, %f684, %f688; add.f32 %f690, %f687, %f689; mov.f32 %f691, 0f3F317200; mul.rn.f32 %f692, %f664, %f691; mov.f32 %f693, 0f35BFBE8E; mul.rn.f32 %f694, %f664, %f693; add.f32 %f695, %f692, %f688; sub.f32 %f696, %f692, %f695; add.f32 %f697, %f688, %f696; add.f32 %f698, %f690, %f697; add.f32 %f699, %f694, %f698; add.f32 %f700, %f695, %f699; sub.f32 %f701, %f695, %f700; add.f32 %f702, %f699, %f701; mul.rn.f32 %f703, %f654, %f700; neg.f32 %f704, %f703; fma.rn.f32 %f705, %f654, %f700, %f704; fma.rn.f32 %f706, %f654, %f702, %f705; mov.f32 %f707, 0f00000000; fma.rn.f32 %f708, %f707, %f700, %f706; add.rn.f32 %f709, %f703, %f708; neg.f32 %f710, %f709; add.rn.f32 %f711, %f703, %f710; add.rn.f32 %f712, %f711, %f708; mov.b32 %r328, %f709; setp.eq.s32 %p91, %r328, 1118925336; add.s32 %r329, %r328, -1; mov.b32 %f713, %r329; add.f32 %f714, %f712, 0f37000000; selp.f32 %f173, %f714, %f712, %p91; selp.f32 %f715, %f713, %f709, %p91; mov.f32 %f716, 0f3FB8AA3B; mul.rn.f32 %f717, %f715, %f716; cvt.rzi.f32.f32 %f718, %f717; abs.f32 %f719, %f718; setp.gt.f32 %p92, %f719, 0f42FC0000; mov.b32 %r330, %f718; and.b32 %r331, %r330, -2147483648; or.b32 %r332, %r331, 1123811328; mov.b32 %f720, %r332; selp.f32 %f721, %f720, %f718, %p92; mov.f32 %f722, 0fBF317218; fma.rn.f32 %f723, %f721, %f722, %f715; mov.f32 %f724, 0f3102E308; fma.rn.f32 %f725, %f721, %f724, %f723; mul.f32 %f726, %f725, 0f3FB8AA3B; add.f32 %f727, %f721, 0f4B40007F; mov.b32 %r333, %f727; shl.b32 %r334, %r333, 23; mov.b32 %f728, %r334; ex2.approx.ftz.f32 %f729, %f726; mul.f32 %f174, %f729, %f728; setp.eq.f32 %p93, %f174, 0f7F800000; mov.f32 %f1289, 0f7F800000; @%p93 bra $L__BB0_73; fma.rn.f32 %f1289, %f174, %f173, %f174; $L__BB0_73: setp.lt.f32 %p94, %f168, 0f00000000; setp.eq.f32 %p95, %f171, 0f3F800000; and.pred %p4, %p94, %p95; setp.eq.f32 %p96, %f168, 0f00000000; @%p96 bra $L__BB0_77; bra.uni $L__BB0_74; $L__BB0_77: add.f32 %f734, %f168, %f168; selp.f32 %f1291, %f734, 0f00000000, %p95; bra.uni $L__BB0_78; $L__BB0_74: mov.b32 %r335, %f1289; xor.b32 %r336, %r335, -2147483648; mov.b32 %f730, %r336; selp.f32 %f1291, %f730, %f1289, %p4; setp.geu.f32 %p97, %f168, 0f00000000; @%p97 bra $L__BB0_78; mov.f32 %f731, 0f3EE66666; cvt.rzi.f32.f32 %f732, %f731; setp.eq.f32 %p98, %f732, 0f3EE66666; @%p98 bra $L__BB0_78; mov.f32 %f1291, 0f7FFFFFFF; $L__BB0_78: add.f32 %f735, %f172, 0f3EE66666; mov.b32 %r337, %f735; setp.lt.s32 %p100, %r337, 2139095040; @%p100 bra $L__BB0_83; setp.gtu.f32 %p101, %f172, 0f7F800000; @%p101 bra $L__BB0_82; bra.uni $L__BB0_80; $L__BB0_82: add.f32 %f1291, %f168, 0f3EE66666; bra.uni $L__BB0_83; $L__BB0_80: setp.neu.f32 %p102, %f172, 0f7F800000; @%p102 bra $L__BB0_83; selp.f32 %f1291, 0fFF800000, 0f7F800000, %p4; $L__BB0_83: setp.eq.f32 %p103, %f168, 0f3F800000; selp.f32 %f183, 0f3F800000, %f1291, %p103; abs.f32 %f184, %f169; setp.lt.f32 %p104, %f184, 0f00800000; mul.f32 %f737, %f184, 0f4B800000; selp.f32 %f738, %f737, %f184, %p104; selp.f32 %f739, 0fC3170000, 0fC2FE0000, %p104; mov.b32 %r338, %f738; and.b32 %r339, %r338, 8388607; or.b32 %r340, %r339, 1065353216; mov.b32 %f740, %r340; shr.u32 %r341, %r338, 23; cvt.rn.f32.u32 %f741, %r341; add.f32 %f742, %f739, %f741; setp.gt.f32 %p105, %f740, 0f3FB504F3; mul.f32 %f743, %f740, 0f3F000000; add.f32 %f744, %f742, 0f3F800000; selp.f32 %f745, %f744, %f742, %p105; selp.f32 %f746, %f743, %f740, %p105; add.f32 %f747, %f746, 0fBF800000; add.f32 %f748, %f746, 0f3F800000; rcp.approx.ftz.f32 %f749, %f748; add.f32 %f750, %f747, %f747; mul.f32 %f751, %f750, %f749; mul.f32 %f752, %f751, %f751; mov.f32 %f753, 0f3C4CAF63; mov.f32 %f754, 0f3B18F0FE; fma.rn.f32 %f755, %f754, %f752, %f753; mov.f32 %f756, 0f3DAAAABD; fma.rn.f32 %f757, %f755, %f752, %f756; mul.rn.f32 %f758, %f757, %f752; mul.rn.f32 %f759, %f758, %f751; sub.f32 %f760, %f747, %f751; add.f32 %f761, %f760, %f760; neg.f32 %f762, %f751; fma.rn.f32 %f763, %f762, %f747, %f761; mul.rn.f32 %f764, %f749, %f763; add.f32 %f765, %f759, %f751; sub.f32 %f766, %f751, %f765; add.f32 %f767, %f759, %f766; add.f32 %f768, %f764, %f767; add.f32 %f769, %f765, %f768; sub.f32 %f770, %f765, %f769; add.f32 %f771, %f768, %f770; mov.f32 %f772, 0f3F317200; mul.rn.f32 %f773, %f745, %f772; mov.f32 %f774, 0f35BFBE8E; mul.rn.f32 %f775, %f745, %f774; add.f32 %f776, %f773, %f769; sub.f32 %f777, %f773, %f776; add.f32 %f778, %f769, %f777; add.f32 %f779, %f771, %f778; add.f32 %f780, %f775, %f779; add.f32 %f781, %f776, %f780; sub.f32 %f782, %f776, %f781; add.f32 %f783, %f780, %f782; mov.f32 %f784, 0f3EE66666; mul.rn.f32 %f785, %f784, %f781; neg.f32 %f786, %f785; fma.rn.f32 %f787, %f784, %f781, %f786; fma.rn.f32 %f788, %f784, %f783, %f787; mov.f32 %f789, 0f00000000; fma.rn.f32 %f790, %f789, %f781, %f788; add.rn.f32 %f791, %f785, %f790; neg.f32 %f792, %f791; add.rn.f32 %f793, %f785, %f792; add.rn.f32 %f794, %f793, %f790; mov.b32 %r342, %f791; setp.eq.s32 %p106, %r342, 1118925336; add.s32 %r343, %r342, -1; mov.b32 %f795, %r343; add.f32 %f796, %f794, 0f37000000; selp.f32 %f185, %f796, %f794, %p106; selp.f32 %f797, %f795, %f791, %p106; mov.f32 %f798, 0f3FB8AA3B; mul.rn.f32 %f799, %f797, %f798; cvt.rzi.f32.f32 %f800, %f799; abs.f32 %f801, %f800; setp.gt.f32 %p107, %f801, 0f42FC0000; mov.b32 %r344, %f800; and.b32 %r345, %r344, -2147483648; or.b32 %r346, %r345, 1123811328; mov.b32 %f802, %r346; selp.f32 %f803, %f802, %f800, %p107; mov.f32 %f804, 0fBF317218; fma.rn.f32 %f805, %f803, %f804, %f797; mov.f32 %f806, 0f3102E308; fma.rn.f32 %f807, %f803, %f806, %f805; mul.f32 %f808, %f807, 0f3FB8AA3B; add.f32 %f809, %f803, 0f4B40007F; mov.b32 %r347, %f809; shl.b32 %r348, %r347, 23; mov.b32 %f810, %r348; ex2.approx.ftz.f32 %f811, %f808; mul.f32 %f186, %f811, %f810; setp.eq.f32 %p108, %f186, 0f7F800000; mov.f32 %f1292, 0f7F800000; @%p108 bra $L__BB0_85; fma.rn.f32 %f1292, %f186, %f185, %f186; $L__BB0_85: setp.lt.f32 %p109, %f169, 0f00000000; and.pred %p5, %p109, %p95; setp.eq.f32 %p111, %f169, 0f00000000; @%p111 bra $L__BB0_89; bra.uni $L__BB0_86; $L__BB0_89: add.f32 %f816, %f169, %f169; selp.f32 %f1294, %f816, 0f00000000, %p95; bra.uni $L__BB0_90; $L__BB0_86: mov.b32 %r349, %f1292; xor.b32 %r350, %r349, -2147483648; mov.b32 %f812, %r350; selp.f32 %f1294, %f812, %f1292, %p5; setp.geu.f32 %p112, %f169, 0f00000000; @%p112 bra $L__BB0_90; mov.f32 %f813, 0f3EE66666; cvt.rzi.f32.f32 %f814, %f813; setp.eq.f32 %p113, %f814, 0f3EE66666; @%p113 bra $L__BB0_90; mov.f32 %f1294, 0f7FFFFFFF; $L__BB0_90: add.f32 %f817, %f184, 0f3EE66666; mov.b32 %r351, %f817; setp.lt.s32 %p115, %r351, 2139095040; @%p115 bra $L__BB0_95; setp.gtu.f32 %p116, %f184, 0f7F800000; @%p116 bra $L__BB0_94; bra.uni $L__BB0_92; $L__BB0_94: add.f32 %f1294, %f169, 0f3EE66666; bra.uni $L__BB0_95; $L__BB0_92: setp.neu.f32 %p117, %f184, 0f7F800000; @%p117 bra $L__BB0_95; selp.f32 %f1294, 0fFF800000, 0f7F800000, %p5; $L__BB0_95: setp.eq.f32 %p118, %f169, 0f3F800000; selp.f32 %f195, 0f3F800000, %f1294, %p118; abs.f32 %f196, %f170; setp.lt.f32 %p119, %f196, 0f00800000; mul.f32 %f819, %f196, 0f4B800000; selp.f32 %f820, %f819, %f196, %p119; selp.f32 %f821, 0fC3170000, 0fC2FE0000, %p119; mov.b32 %r352, %f820; and.b32 %r353, %r352, 8388607; or.b32 %r354, %r353, 1065353216; mov.b32 %f822, %r354; shr.u32 %r355, %r352, 23; cvt.rn.f32.u32 %f823, %r355; add.f32 %f824, %f821, %f823; setp.gt.f32 %p120, %f822, 0f3FB504F3; mul.f32 %f825, %f822, 0f3F000000; add.f32 %f826, %f824, 0f3F800000; selp.f32 %f827, %f826, %f824, %p120; selp.f32 %f828, %f825, %f822, %p120; add.f32 %f829, %f828, 0fBF800000; add.f32 %f830, %f828, 0f3F800000; rcp.approx.ftz.f32 %f831, %f830; add.f32 %f832, %f829, %f829; mul.f32 %f833, %f832, %f831; mul.f32 %f834, %f833, %f833; mov.f32 %f835, 0f3C4CAF63; mov.f32 %f836, 0f3B18F0FE; fma.rn.f32 %f837, %f836, %f834, %f835; mov.f32 %f838, 0f3DAAAABD; fma.rn.f32 %f839, %f837, %f834, %f838; mul.rn.f32 %f840, %f839, %f834; mul.rn.f32 %f841, %f840, %f833; sub.f32 %f842, %f829, %f833; add.f32 %f843, %f842, %f842; neg.f32 %f844, %f833; fma.rn.f32 %f845, %f844, %f829, %f843; mul.rn.f32 %f846, %f831, %f845; add.f32 %f847, %f841, %f833; sub.f32 %f848, %f833, %f847; add.f32 %f849, %f841, %f848; add.f32 %f850, %f846, %f849; add.f32 %f851, %f847, %f850; sub.f32 %f852, %f847, %f851; add.f32 %f853, %f850, %f852; mov.f32 %f854, 0f3F317200; mul.rn.f32 %f855, %f827, %f854; mov.f32 %f856, 0f35BFBE8E; mul.rn.f32 %f857, %f827, %f856; add.f32 %f858, %f855, %f851; sub.f32 %f859, %f855, %f858; add.f32 %f860, %f851, %f859; add.f32 %f861, %f853, %f860; add.f32 %f862, %f857, %f861; add.f32 %f863, %f858, %f862; sub.f32 %f864, %f858, %f863; add.f32 %f865, %f862, %f864; mov.f32 %f866, 0f3EE66666; mul.rn.f32 %f867, %f866, %f863; neg.f32 %f868, %f867; fma.rn.f32 %f869, %f866, %f863, %f868; fma.rn.f32 %f870, %f866, %f865, %f869; mov.f32 %f871, 0f00000000; fma.rn.f32 %f872, %f871, %f863, %f870; add.rn.f32 %f873, %f867, %f872; neg.f32 %f874, %f873; add.rn.f32 %f875, %f867, %f874; add.rn.f32 %f876, %f875, %f872; mov.b32 %r356, %f873; setp.eq.s32 %p121, %r356, 1118925336; add.s32 %r357, %r356, -1; mov.b32 %f877, %r357; add.f32 %f878, %f876, 0f37000000; selp.f32 %f197, %f878, %f876, %p121; selp.f32 %f879, %f877, %f873, %p121; mov.f32 %f880, 0f3FB8AA3B; mul.rn.f32 %f881, %f879, %f880; cvt.rzi.f32.f32 %f882, %f881; abs.f32 %f883, %f882; setp.gt.f32 %p122, %f883, 0f42FC0000; mov.b32 %r358, %f882; and.b32 %r359, %r358, -2147483648; or.b32 %r360, %r359, 1123811328; mov.b32 %f884, %r360; selp.f32 %f885, %f884, %f882, %p122; mov.f32 %f886, 0fBF317218; fma.rn.f32 %f887, %f885, %f886, %f879; mov.f32 %f888, 0f3102E308; fma.rn.f32 %f889, %f885, %f888, %f887; mul.f32 %f890, %f889, 0f3FB8AA3B; add.f32 %f891, %f885, 0f4B40007F; mov.b32 %r361, %f891; shl.b32 %r362, %r361, 23; mov.b32 %f892, %r362; ex2.approx.ftz.f32 %f893, %f890; mul.f32 %f198, %f893, %f892; setp.eq.f32 %p123, %f198, 0f7F800000; mov.f32 %f1295, 0f7F800000; @%p123 bra $L__BB0_97; fma.rn.f32 %f1295, %f198, %f197, %f198; $L__BB0_97: setp.lt.f32 %p124, %f170, 0f00000000; and.pred %p6, %p124, %p95; setp.eq.f32 %p126, %f170, 0f00000000; @%p126 bra $L__BB0_101; bra.uni $L__BB0_98; $L__BB0_101: add.f32 %f898, %f170, %f170; selp.f32 %f1297, %f898, 0f00000000, %p95; bra.uni $L__BB0_102; $L__BB0_98: mov.b32 %r363, %f1295; xor.b32 %r364, %r363, -2147483648; mov.b32 %f894, %r364; selp.f32 %f1297, %f894, %f1295, %p6; setp.geu.f32 %p127, %f170, 0f00000000; @%p127 bra $L__BB0_102; mov.f32 %f895, 0f3EE66666; cvt.rzi.f32.f32 %f896, %f895; setp.eq.f32 %p128, %f896, 0f3EE66666; @%p128 bra $L__BB0_102; mov.f32 %f1297, 0f7FFFFFFF; $L__BB0_102: add.f32 %f899, %f196, 0f3EE66666; mov.b32 %r365, %f899; setp.lt.s32 %p130, %r365, 2139095040; @%p130 bra $L__BB0_107; setp.gtu.f32 %p131, %f196, 0f7F800000; @%p131 bra $L__BB0_106; bra.uni $L__BB0_104; $L__BB0_106: add.f32 %f1297, %f170, 0f3EE66666; bra.uni $L__BB0_107; $L__BB0_104: setp.neu.f32 %p132, %f196, 0f7F800000; @%p132 bra $L__BB0_107; selp.f32 %f1297, 0fFF800000, 0f7F800000, %p6; $L__BB0_107: setp.eq.f32 %p133, %f170, 0f3F800000; mov.f32 %f901, 0f3F800000; selp.f32 %f902, 0f3F800000, %f1297, %p133; ld.const.u64 %rd71, [params+144]; cvta.to.global.u64 %rd21, %rd71; ld.const.u32 %r366, [params+136]; mad.lo.s32 %r367, %r366, %r7, %r6; cvt.u64.u32 %rd22, %r367; min.f32 %f903, %f183, %f901; mov.f32 %f904, 0f00000000; max.f32 %f207, %f904, %f903; min.f32 %f905, %f195, %f901; max.f32 %f208, %f904, %f905; min.f32 %f906, %f902, %f901; max.f32 %f209, %f904, %f906; mov.f32 %f907, 0f3E555555; cvt.rzi.f32.f32 %f908, %f907; add.f32 %f909, %f908, %f908; mov.f32 %f910, 0f3ED55555; sub.f32 %f911, %f910, %f909; abs.f32 %f210, %f911; abs.f32 %f211, %f207; setp.lt.f32 %p134, %f211, 0f00800000; mul.f32 %f912, %f211, 0f4B800000; selp.f32 %f913, %f912, %f211, %p134; selp.f32 %f914, 0fC3170000, 0fC2FE0000, %p134; mov.b32 %r368, %f913; and.b32 %r369, %r368, 8388607; or.b32 %r370, %r369, 1065353216; mov.b32 %f915, %r370; shr.u32 %r371, %r368, 23; cvt.rn.f32.u32 %f916, %r371; add.f32 %f917, %f914, %f916; setp.gt.f32 %p135, %f915, 0f3FB504F3; mul.f32 %f918, %f915, 0f3F000000; add.f32 %f919, %f917, 0f3F800000; selp.f32 %f920, %f919, %f917, %p135; selp.f32 %f921, %f918, %f915, %p135; add.f32 %f922, %f921, 0fBF800000; add.f32 %f923, %f921, 0f3F800000; rcp.approx.ftz.f32 %f924, %f923; add.f32 %f925, %f922, %f922; mul.f32 %f926, %f925, %f924; mul.f32 %f927, %f926, %f926; mov.f32 %f928, 0f3C4CAF63; mov.f32 %f929, 0f3B18F0FE; fma.rn.f32 %f930, %f929, %f927, %f928; mov.f32 %f931, 0f3DAAAABD; fma.rn.f32 %f932, %f930, %f927, %f931; mul.rn.f32 %f933, %f932, %f927; mul.rn.f32 %f934, %f933, %f926; sub.f32 %f935, %f922, %f926; add.f32 %f936, %f935, %f935; neg.f32 %f937, %f926; fma.rn.f32 %f938, %f937, %f922, %f936; mul.rn.f32 %f939, %f924, %f938; add.f32 %f940, %f934, %f926; sub.f32 %f941, %f926, %f940; add.f32 %f942, %f934, %f941; add.f32 %f943, %f939, %f942; add.f32 %f944, %f940, %f943; sub.f32 %f945, %f940, %f944; add.f32 %f946, %f943, %f945; mov.f32 %f947, 0f3F317200; mul.rn.f32 %f948, %f920, %f947; mov.f32 %f949, 0f35BFBE8E; mul.rn.f32 %f950, %f920, %f949; add.f32 %f951, %f948, %f944; sub.f32 %f952, %f948, %f951; add.f32 %f953, %f944, %f952; add.f32 %f954, %f946, %f953; add.f32 %f955, %f950, %f954; add.f32 %f956, %f951, %f955; sub.f32 %f957, %f951, %f956; add.f32 %f958, %f955, %f957; mul.rn.f32 %f959, %f910, %f956; neg.f32 %f960, %f959; fma.rn.f32 %f961, %f910, %f956, %f960; fma.rn.f32 %f962, %f910, %f958, %f961; fma.rn.f32 %f963, %f904, %f956, %f962; add.rn.f32 %f964, %f959, %f963; neg.f32 %f965, %f964; add.rn.f32 %f966, %f959, %f965; add.rn.f32 %f967, %f966, %f963; mov.b32 %r372, %f964; setp.eq.s32 %p136, %r372, 1118925336; add.s32 %r373, %r372, -1; mov.b32 %f968, %r373; add.f32 %f969, %f967, 0f37000000; selp.f32 %f212, %f969, %f967, %p136; selp.f32 %f970, %f968, %f964, %p136; mov.f32 %f971, 0f3FB8AA3B; mul.rn.f32 %f972, %f970, %f971; cvt.rzi.f32.f32 %f973, %f972; abs.f32 %f974, %f973; setp.gt.f32 %p137, %f974, 0f42FC0000; mov.b32 %r374, %f973; and.b32 %r375, %r374, -2147483648; or.b32 %r376, %r375, 1123811328; mov.b32 %f975, %r376; selp.f32 %f976, %f975, %f973, %p137; mov.f32 %f977, 0fBF317218; fma.rn.f32 %f978, %f976, %f977, %f970; mov.f32 %f979, 0f3102E308; fma.rn.f32 %f980, %f976, %f979, %f978; mul.f32 %f981, %f980, 0f3FB8AA3B; add.f32 %f982, %f976, 0f4B40007F; mov.b32 %r377, %f982; shl.b32 %r378, %r377, 23; mov.b32 %f983, %r378; ex2.approx.ftz.f32 %f984, %f981; mul.f32 %f213, %f984, %f983; setp.eq.f32 %p138, %f213, 0f7F800000; mov.f32 %f1298, 0f7F800000; @%p138 bra $L__BB0_109; fma.rn.f32 %f1298, %f213, %f212, %f213; $L__BB0_109: setp.lt.f32 %p139, %f207, 0f00000000; setp.eq.f32 %p140, %f210, 0f3F800000; and.pred %p7, %p139, %p140; setp.eq.f32 %p141, %f207, 0f00000000; @%p141 bra $L__BB0_113; bra.uni $L__BB0_110; $L__BB0_113: add.f32 %f989, %f207, %f207; selp.f32 %f1300, %f989, 0f00000000, %p140; bra.uni $L__BB0_114; $L__BB0_110: mov.b32 %r379, %f1298; xor.b32 %r380, %r379, -2147483648; mov.b32 %f985, %r380; selp.f32 %f1300, %f985, %f1298, %p7; setp.geu.f32 %p142, %f207, 0f00000000; @%p142 bra $L__BB0_114; mov.f32 %f986, 0f3ED55555; cvt.rzi.f32.f32 %f987, %f986; setp.eq.f32 %p143, %f987, 0f3ED55555; @%p143 bra $L__BB0_114; mov.f32 %f1300, 0f7FFFFFFF; $L__BB0_114: add.f32 %f990, %f211, 0f3ED55555; mov.b32 %r381, %f990; setp.lt.s32 %p145, %r381, 2139095040; @%p145 bra $L__BB0_119; setp.gtu.f32 %p146, %f211, 0f7F800000; @%p146 bra $L__BB0_118; bra.uni $L__BB0_116; $L__BB0_118: add.f32 %f1300, %f207, 0f3ED55555; bra.uni $L__BB0_119; $L__BB0_116: setp.neu.f32 %p147, %f211, 0f7F800000; @%p147 bra $L__BB0_119; selp.f32 %f1300, 0fFF800000, 0f7F800000, %p7; $L__BB0_119: abs.f32 %f222, %f208; setp.lt.f32 %p148, %f222, 0f00800000; mul.f32 %f992, %f222, 0f4B800000; selp.f32 %f993, %f992, %f222, %p148; selp.f32 %f994, 0fC3170000, 0fC2FE0000, %p148; mov.b32 %r382, %f993; and.b32 %r383, %r382, 8388607; or.b32 %r384, %r383, 1065353216; mov.b32 %f995, %r384; shr.u32 %r385, %r382, 23; cvt.rn.f32.u32 %f996, %r385; add.f32 %f997, %f994, %f996; setp.gt.f32 %p149, %f995, 0f3FB504F3; mul.f32 %f998, %f995, 0f3F000000; add.f32 %f999, %f997, 0f3F800000; selp.f32 %f1000, %f999, %f997, %p149; selp.f32 %f1001, %f998, %f995, %p149; add.f32 %f1002, %f1001, 0fBF800000; add.f32 %f1003, %f1001, 0f3F800000; rcp.approx.ftz.f32 %f1004, %f1003; add.f32 %f1005, %f1002, %f1002; mul.f32 %f1006, %f1005, %f1004; mul.f32 %f1007, %f1006, %f1006; mov.f32 %f1008, 0f3C4CAF63; mov.f32 %f1009, 0f3B18F0FE; fma.rn.f32 %f1010, %f1009, %f1007, %f1008; mov.f32 %f1011, 0f3DAAAABD; fma.rn.f32 %f1012, %f1010, %f1007, %f1011; mul.rn.f32 %f1013, %f1012, %f1007; mul.rn.f32 %f1014, %f1013, %f1006; sub.f32 %f1015, %f1002, %f1006; add.f32 %f1016, %f1015, %f1015; neg.f32 %f1017, %f1006; fma.rn.f32 %f1018, %f1017, %f1002, %f1016; mul.rn.f32 %f1019, %f1004, %f1018; add.f32 %f1020, %f1014, %f1006; sub.f32 %f1021, %f1006, %f1020; add.f32 %f1022, %f1014, %f1021; add.f32 %f1023, %f1019, %f1022; add.f32 %f1024, %f1020, %f1023; sub.f32 %f1025, %f1020, %f1024; add.f32 %f1026, %f1023, %f1025; mov.f32 %f1027, 0f3F317200; mul.rn.f32 %f1028, %f1000, %f1027; mov.f32 %f1029, 0f35BFBE8E; mul.rn.f32 %f1030, %f1000, %f1029; add.f32 %f1031, %f1028, %f1024; sub.f32 %f1032, %f1028, %f1031; add.f32 %f1033, %f1024, %f1032; add.f32 %f1034, %f1026, %f1033; add.f32 %f1035, %f1030, %f1034; add.f32 %f1036, %f1031, %f1035; sub.f32 %f1037, %f1031, %f1036; add.f32 %f1038, %f1035, %f1037; mov.f32 %f1039, 0f3ED55555; mul.rn.f32 %f1040, %f1039, %f1036; neg.f32 %f1041, %f1040; fma.rn.f32 %f1042, %f1039, %f1036, %f1041; fma.rn.f32 %f1043, %f1039, %f1038, %f1042; mov.f32 %f1044, 0f00000000; fma.rn.f32 %f1045, %f1044, %f1036, %f1043; add.rn.f32 %f1046, %f1040, %f1045; neg.f32 %f1047, %f1046; add.rn.f32 %f1048, %f1040, %f1047; add.rn.f32 %f1049, %f1048, %f1045; mov.b32 %r386, %f1046; setp.eq.s32 %p150, %r386, 1118925336; add.s32 %r387, %r386, -1; mov.b32 %f1050, %r387; add.f32 %f1051, %f1049, 0f37000000; selp.f32 %f223, %f1051, %f1049, %p150; selp.f32 %f1052, %f1050, %f1046, %p150; mov.f32 %f1053, 0f3FB8AA3B; mul.rn.f32 %f1054, %f1052, %f1053; cvt.rzi.f32.f32 %f1055, %f1054; abs.f32 %f1056, %f1055; setp.gt.f32 %p151, %f1056, 0f42FC0000; mov.b32 %r388, %f1055; and.b32 %r389, %r388, -2147483648; or.b32 %r390, %r389, 1123811328; mov.b32 %f1057, %r390; selp.f32 %f1058, %f1057, %f1055, %p151; mov.f32 %f1059, 0fBF317218; fma.rn.f32 %f1060, %f1058, %f1059, %f1052; mov.f32 %f1061, 0f3102E308; fma.rn.f32 %f1062, %f1058, %f1061, %f1060; mul.f32 %f1063, %f1062, 0f3FB8AA3B; add.f32 %f1064, %f1058, 0f4B40007F; mov.b32 %r391, %f1064; shl.b32 %r392, %r391, 23; mov.b32 %f1065, %r392; ex2.approx.ftz.f32 %f1066, %f1063; mul.f32 %f224, %f1066, %f1065; setp.eq.f32 %p152, %f224, 0f7F800000; mov.f32 %f1301, 0f7F800000; @%p152 bra $L__BB0_121; fma.rn.f32 %f1301, %f224, %f223, %f224; $L__BB0_121: setp.lt.f32 %p153, %f208, 0f00000000; and.pred %p8, %p153, %p140; setp.eq.f32 %p155, %f208, 0f00000000; @%p155 bra $L__BB0_125; bra.uni $L__BB0_122; $L__BB0_125: add.f32 %f1071, %f208, %f208; selp.f32 %f1303, %f1071, 0f00000000, %p140; bra.uni $L__BB0_126; $L__BB0_122: mov.b32 %r393, %f1301; xor.b32 %r394, %r393, -2147483648; mov.b32 %f1067, %r394; selp.f32 %f1303, %f1067, %f1301, %p8; setp.geu.f32 %p156, %f208, 0f00000000; @%p156 bra $L__BB0_126; mov.f32 %f1068, 0f3ED55555; cvt.rzi.f32.f32 %f1069, %f1068; setp.eq.f32 %p157, %f1069, 0f3ED55555; @%p157 bra $L__BB0_126; mov.f32 %f1303, 0f7FFFFFFF; $L__BB0_126: add.f32 %f1072, %f222, 0f3ED55555; mov.b32 %r395, %f1072; setp.lt.s32 %p159, %r395, 2139095040; @%p159 bra $L__BB0_131; setp.gtu.f32 %p160, %f222, 0f7F800000; @%p160 bra $L__BB0_130; bra.uni $L__BB0_128; $L__BB0_130: add.f32 %f1303, %f208, 0f3ED55555; bra.uni $L__BB0_131; $L__BB0_128: setp.neu.f32 %p161, %f222, 0f7F800000; @%p161 bra $L__BB0_131; selp.f32 %f1303, 0fFF800000, 0f7F800000, %p8; $L__BB0_131: abs.f32 %f233, %f209; setp.lt.f32 %p162, %f233, 0f00800000; mul.f32 %f1074, %f233, 0f4B800000; selp.f32 %f1075, %f1074, %f233, %p162; selp.f32 %f1076, 0fC3170000, 0fC2FE0000, %p162; mov.b32 %r396, %f1075; and.b32 %r397, %r396, 8388607; or.b32 %r398, %r397, 1065353216; mov.b32 %f1077, %r398; shr.u32 %r399, %r396, 23; cvt.rn.f32.u32 %f1078, %r399; add.f32 %f1079, %f1076, %f1078; setp.gt.f32 %p163, %f1077, 0f3FB504F3; mul.f32 %f1080, %f1077, 0f3F000000; add.f32 %f1081, %f1079, 0f3F800000; selp.f32 %f1082, %f1081, %f1079, %p163; selp.f32 %f1083, %f1080, %f1077, %p163; add.f32 %f1084, %f1083, 0fBF800000; add.f32 %f1085, %f1083, 0f3F800000; rcp.approx.ftz.f32 %f1086, %f1085; add.f32 %f1087, %f1084, %f1084; mul.f32 %f1088, %f1087, %f1086; mul.f32 %f1089, %f1088, %f1088; mov.f32 %f1090, 0f3C4CAF63; mov.f32 %f1091, 0f3B18F0FE; fma.rn.f32 %f1092, %f1091, %f1089, %f1090; mov.f32 %f1093, 0f3DAAAABD; fma.rn.f32 %f1094, %f1092, %f1089, %f1093; mul.rn.f32 %f1095, %f1094, %f1089; mul.rn.f32 %f1096, %f1095, %f1088; sub.f32 %f1097, %f1084, %f1088; add.f32 %f1098, %f1097, %f1097; neg.f32 %f1099, %f1088; fma.rn.f32 %f1100, %f1099, %f1084, %f1098; mul.rn.f32 %f1101, %f1086, %f1100; add.f32 %f1102, %f1096, %f1088; sub.f32 %f1103, %f1088, %f1102; add.f32 %f1104, %f1096, %f1103; add.f32 %f1105, %f1101, %f1104; add.f32 %f1106, %f1102, %f1105; sub.f32 %f1107, %f1102, %f1106; add.f32 %f1108, %f1105, %f1107; mov.f32 %f1109, 0f3F317200; mul.rn.f32 %f1110, %f1082, %f1109; mov.f32 %f1111, 0f35BFBE8E; mul.rn.f32 %f1112, %f1082, %f1111; add.f32 %f1113, %f1110, %f1106; sub.f32 %f1114, %f1110, %f1113; add.f32 %f1115, %f1106, %f1114; add.f32 %f1116, %f1108, %f1115; add.f32 %f1117, %f1112, %f1116; add.f32 %f1118, %f1113, %f1117; sub.f32 %f1119, %f1113, %f1118; add.f32 %f1120, %f1117, %f1119; mov.f32 %f1121, 0f3ED55555; mul.rn.f32 %f1122, %f1121, %f1118; neg.f32 %f1123, %f1122; fma.rn.f32 %f1124, %f1121, %f1118, %f1123; fma.rn.f32 %f1125, %f1121, %f1120, %f1124; mov.f32 %f1126, 0f00000000; fma.rn.f32 %f1127, %f1126, %f1118, %f1125; add.rn.f32 %f1128, %f1122, %f1127; neg.f32 %f1129, %f1128; add.rn.f32 %f1130, %f1122, %f1129; add.rn.f32 %f1131, %f1130, %f1127; mov.b32 %r400, %f1128; setp.eq.s32 %p164, %r400, 1118925336; add.s32 %r401, %r400, -1; mov.b32 %f1132, %r401; add.f32 %f1133, %f1131, 0f37000000; selp.f32 %f234, %f1133, %f1131, %p164; selp.f32 %f1134, %f1132, %f1128, %p164; mov.f32 %f1135, 0f3FB8AA3B; mul.rn.f32 %f1136, %f1134, %f1135; cvt.rzi.f32.f32 %f1137, %f1136; abs.f32 %f1138, %f1137; setp.gt.f32 %p165, %f1138, 0f42FC0000; mov.b32 %r402, %f1137; and.b32 %r403, %r402, -2147483648; or.b32 %r404, %r403, 1123811328; mov.b32 %f1139, %r404; selp.f32 %f1140, %f1139, %f1137, %p165; mov.f32 %f1141, 0fBF317218; fma.rn.f32 %f1142, %f1140, %f1141, %f1134; mov.f32 %f1143, 0f3102E308; fma.rn.f32 %f1144, %f1140, %f1143, %f1142; mul.f32 %f1145, %f1144, 0f3FB8AA3B; add.f32 %f1146, %f1140, 0f4B40007F; mov.b32 %r405, %f1146; shl.b32 %r406, %r405, 23; mov.b32 %f1147, %r406; ex2.approx.ftz.f32 %f1148, %f1145; mul.f32 %f235, %f1148, %f1147; setp.eq.f32 %p166, %f235, 0f7F800000; mov.f32 %f1304, 0f7F800000; @%p166 bra $L__BB0_133; fma.rn.f32 %f1304, %f235, %f234, %f235; $L__BB0_133: setp.lt.f32 %p167, %f209, 0f00000000; and.pred %p9, %p167, %p140; setp.eq.f32 %p169, %f209, 0f00000000; @%p169 bra $L__BB0_137; bra.uni $L__BB0_134; $L__BB0_137: add.f32 %f1153, %f209, %f209; selp.f32 %f1306, %f1153, 0f00000000, %p140; bra.uni $L__BB0_138; $L__BB0_134: mov.b32 %r407, %f1304; xor.b32 %r408, %r407, -2147483648; mov.b32 %f1149, %r408; selp.f32 %f1306, %f1149, %f1304, %p9; setp.geu.f32 %p170, %f209, 0f00000000; @%p170 bra $L__BB0_138; mov.f32 %f1150, 0f3ED55555; cvt.rzi.f32.f32 %f1151, %f1150; setp.eq.f32 %p171, %f1151, 0f3ED55555; @%p171 bra $L__BB0_138; mov.f32 %f1306, 0f7FFFFFFF; $L__BB0_138: add.f32 %f1154, %f233, 0f3ED55555; mov.b32 %r409, %f1154; setp.lt.s32 %p173, %r409, 2139095040; @%p173 bra $L__BB0_143; setp.gtu.f32 %p174, %f233, 0f7F800000; @%p174 bra $L__BB0_142; bra.uni $L__BB0_140; $L__BB0_142: add.f32 %f1306, %f209, 0f3ED55555; bra.uni $L__BB0_143; $L__BB0_140: setp.neu.f32 %p175, %f233, 0f7F800000; @%p175 bra $L__BB0_143; selp.f32 %f1306, 0fFF800000, 0f7F800000, %p9; $L__BB0_143: fma.rn.f32 %f1155, %f1300, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p176, %f207, 0f3F800000; mov.f32 %f1156, 0f3F800000; selp.f32 %f1157, 0f3F7FFFFF, %f1155, %p176; mul.f32 %f1158, %f207, 0f414EB852; setp.lt.f32 %p177, %f207, 0f3B4D2E1C; selp.f32 %f1159, %f1158, %f1157, %p177; fma.rn.f32 %f1160, %f1303, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p178, %f208, 0f3F800000; selp.f32 %f1161, 0f3F7FFFFF, %f1160, %p178; mul.f32 %f1162, %f208, 0f414EB852; setp.lt.f32 %p179, %f208, 0f3B4D2E1C; selp.f32 %f1163, %f1162, %f1161, %p179; fma.rn.f32 %f1164, %f1306, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p180, %f209, 0f3F800000; selp.f32 %f1165, 0f3F7FFFFF, %f1164, %p180; mul.f32 %f1166, %f209, 0f414EB852; setp.lt.f32 %p181, %f209, 0f3B4D2E1C; selp.f32 %f1167, %f1166, %f1165, %p181; min.f32 %f1168, %f1159, %f1156; mov.f32 %f1169, 0f00000000; max.f32 %f1170, %f1169, %f1168; mul.f32 %f1171, %f1170, 0f43800000; cvt.rzi.u32.f32 %r410, %f1171; min.u32 %r411, %r410, 255; min.f32 %f1172, %f1163, %f1156; max.f32 %f1173, %f1169, %f1172; mul.f32 %f1174, %f1173, 0f43800000; cvt.rzi.u32.f32 %r412, %f1174; min.u32 %r413, %r412, 255; min.f32 %f1175, %f1167, %f1156; max.f32 %f1176, %f1169, %f1175; mul.f32 %f1177, %f1176, 0f43800000; cvt.rzi.u32.f32 %r414, %f1177; min.u32 %r415, %r414, 255; shl.b64 %rd72, %rd22, 2; add.s64 %rd73, %rd21, %rd72; cvt.u16.u32 %rs18, %r415; cvt.u16.u32 %rs19, %r413; cvt.u16.u32 %rs20, %r411; mov.u16 %rs21, 255; st.global.v4.u8 [%rd73], {%rs20, %rs19, %rs18, %rs21}; $L__BB0_144: and.b32 %r416, %r56, 4; setp.eq.s32 %p182, %r416, 0; @%p182 bra $L__BB0_148; ld.const.u32 %r417, [params+108]; setp.eq.s32 %p183, %r417, 0; ld.const.u64 %rd74, [params+224]; cvta.to.global.u64 %rd75, %rd74; ld.const.u32 %r418, [params+216]; mad.lo.s32 %r419, %r418, %r7, %r6; mul.wide.u32 %rd76, %r419, 8; add.s64 %rd23, %rd75, %rd76; @%p183 bra $L__BB0_147; ld.global.v4.u16 {%rs29, %rs30, %rs31, %rs32}, [%rd23]; // begin inline asm { cvt.f32.f16 %f1178, %rs29;} // end inline asm // begin inline asm { cvt.f32.f16 %f1179, %rs30;} // end inline asm // begin inline asm { cvt.f32.f16 %f1180, %rs31;} // end inline asm add.f32 %f1181, %f168, %f1178; add.f32 %f1182, %f169, %f1179; add.f32 %f1183, %f170, %f1180; mov.f32 %f1184, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs27, %f1183;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs26, %f1182;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs25, %f1181;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs28, %f1184;} // end inline asm st.global.v4.u16 [%rd23], {%rs25, %rs26, %rs27, %rs28}; bra.uni $L__BB0_148; $L__BB0_147: mov.f32 %f1188, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs36, %f1188;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs35, %f170;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs34, %f169;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs33, %f168;} // end inline asm st.global.v4.u16 [%rd23], {%rs33, %rs34, %rs35, %rs36}; $L__BB0_148: and.b32 %r420, %r56, 64; setp.eq.s32 %p184, %r420, 0; @%p184 bra $L__BB0_160; mul.f32 %f1189, %f1235, %f1235; fma.rn.f32 %f1190, %f1234, %f1234, %f1189; fma.rn.f32 %f1191, %f1236, %f1236, %f1190; sqrt.rn.f32 %f1192, %f1191; rcp.rn.f32 %f1193, %f1192; mul.f32 %f1194, %f1234, %f1193; mul.f32 %f1195, %f1235, %f1193; mul.f32 %f1196, %f1236, %f1193; ld.const.u64 %rd77, [params+208]; cvta.to.global.u64 %rd78, %rd77; ld.const.u32 %r421, [params+200]; mad.lo.s32 %r422, %r421, %r7, %r6; fma.rn.f32 %f1197, %f1194, 0f3F000000, 0f3F000000; mul.f32 %f1198, %f1197, 0f437F0000; cvt.rzi.u32.f32 %r423, %f1198; fma.rn.f32 %f1199, %f1195, 0f3F000000, 0f3F000000; mul.f32 %f1200, %f1199, 0f437F0000; cvt.rzi.u32.f32 %r424, %f1200; fma.rn.f32 %f1201, %f1196, 0f3F000000, 0f3F000000; mul.f32 %f1202, %f1201, 0f437F0000; cvt.rzi.u32.f32 %r425, %f1202; mul.wide.u32 %rd79, %r422, 4; add.s64 %rd80, %rd78, %rd79; cvt.u16.u32 %rs37, %r425; cvt.u16.u32 %rs38, %r424; cvt.u16.u32 %rs39, %r423; mov.u16 %rs40, 255; st.global.v4.u8 [%rd80], {%rs39, %rs38, %rs37, %rs40}; $L__BB0_160: ret; }