// // 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<198>; .reg .b16 %rs<177>; .reg .f32 %f<1579>; .reg .b32 %r<464>; .reg .f64 %fd<5>; .reg .b64 %rd<125>; mov.u64 %SPL, __local_depot0; add.u64 %rd1, %SPL, 0; // begin inline asm call (%r60), _optix_get_launch_index_x, (); // end inline asm // begin inline asm call (%r61), _optix_get_launch_index_y, (); // end inline asm ld.const.u32 %r3, [params+616]; abs.s32 %r4, %r3; ld.const.u32 %r5, [params+540]; shl.b32 %r6, %r5, 1; ld.const.u64 %rd34, [params+400]; cvta.to.global.u64 %rd35, %rd34; ld.const.u32 %r66, [params+392]; mad.lo.s32 %r67, %r66, %r61, %r60; mul.wide.u32 %rd36, %r67, 4; add.s64 %rd2, %rd35, %rd36; ld.global.v2.u8 {%rs7, %rs176}, [%rd2]; or.b16 %rs9, %rs7, %rs176; and.b16 %rs10, %rs9, 255; setp.eq.s16 %p10, %rs10, 0; @%p10 bra $L__BB0_2; ld.global.u8 %rs175, [%rd2+2]; bra.uni $L__BB0_3; $L__BB0_2: ld.global.u8 %rs175, [%rd2+2]; setp.eq.s16 %p11, %rs175, 0; mov.f32 %f1437, 0f00000000; mov.u16 %rs176, 0; mov.f32 %f1438, %f1437; mov.f32 %f1439, %f1437; @%p11 bra $L__BB0_4; $L__BB0_3: cvt.rn.f32.u16 %f360, %rs7; div.rn.f32 %f361, %f360, 0f437F0000; fma.rn.f32 %f362, %f361, 0f40000000, 0fBF800000; and.b16 %rs13, %rs176, 255; cvt.rn.f32.u16 %f363, %rs13; div.rn.f32 %f364, %f363, 0f437F0000; fma.rn.f32 %f365, %f364, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f366, %rs175; div.rn.f32 %f367, %f366, 0f437F0000; fma.rn.f32 %f368, %f367, 0f40000000, 0fBF800000; mul.f32 %f369, %f365, %f365; fma.rn.f32 %f370, %f362, %f362, %f369; fma.rn.f32 %f371, %f368, %f368, %f370; sqrt.rn.f32 %f372, %f371; rcp.rn.f32 %f373, %f372; mul.f32 %f1439, %f373, %f368; mul.f32 %f1438, %f373, %f365; mul.f32 %f1437, %f362, %f373; $L__BB0_4: ld.const.v2.u32 {%r68, %r69}, [params]; add.s32 %r7, %r68, %r60; add.s32 %r8, %r69, %r61; setp.eq.f32 %p12, %f1437, 0f00000000; setp.eq.f32 %p13, %f1438, 0f00000000; and.pred %p14, %p12, %p13; setp.eq.f32 %p15, %f1439, 0f00000000; and.pred %p16, %p15, %p14; @%p16 bra $L__BB0_158; bra.uni $L__BB0_5; $L__BB0_158: ld.const.u32 %r57, [params+104]; and.b32 %r429, %r57, 1; setp.eq.b32 %p186, %r429, 1; mov.pred %p187, 0; xor.pred %p188, %p186, %p187; not.pred %p189, %p188; @%p189 bra $L__BB0_160; ld.const.u64 %rd97, [params+144]; cvta.to.global.u64 %rd98, %rd97; ld.const.u32 %r430, [params+136]; mad.lo.s32 %r431, %r430, %r8, %r7; mul.wide.u32 %rd99, %r431, 4; add.s64 %rd100, %rd98, %rd99; mov.u16 %rs97, 0; st.global.v4.u8 [%rd100], {%rs97, %rs97, %rs97, %rs97}; $L__BB0_160: and.b32 %r432, %r57, 8; setp.eq.s32 %p190, %r432, 0; @%p190 bra $L__BB0_162; ld.const.u64 %rd101, [params+192]; cvta.to.global.u64 %rd102, %rd101; ld.const.u32 %r433, [params+184]; mad.lo.s32 %r434, %r433, %r8, %r7; mov.f32 %f1385, 0f00000000; cvt.rzi.u32.f32 %r435, %f1385; mul.wide.u32 %rd103, %r434, 2; add.s64 %rd104, %rd102, %rd103; mov.u16 %rs98, 0; cvt.u16.u32 %rs99, %r435; st.global.v2.u8 [%rd104], {%rs99, %rs98}; $L__BB0_162: and.b32 %r436, %r57, 4; setp.eq.s32 %p191, %r436, 0; ld.const.u32 %r463, [params+108]; @%p191 bra $L__BB0_166; setp.eq.s32 %p192, %r463, 0; ld.const.u64 %rd105, [params+224]; cvta.to.global.u64 %rd106, %rd105; ld.const.u32 %r437, [params+216]; mad.lo.s32 %r438, %r437, %r8, %r7; mul.wide.u32 %rd107, %r438, 8; add.s64 %rd28, %rd106, %rd107; @%p192 bra $L__BB0_165; ld.global.v4.u16 {%rs106, %rs107, %rs108, %rs109}, [%rd28]; // begin inline asm { cvt.f32.f16 %f1386, %rs106;} // end inline asm // begin inline asm { cvt.f32.f16 %f1387, %rs107;} // end inline asm // begin inline asm { cvt.f32.f16 %f1388, %rs108;} // end inline asm add.f32 %f1389, %f1386, 0f00000000; add.f32 %f1390, %f1387, 0f00000000; add.f32 %f1391, %f1388, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs105, %f1391;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs104, %f1390;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs103, %f1389;} // end inline asm mov.u16 %rs110, 0; st.global.v4.u16 [%rd28], {%rs103, %rs104, %rs105, %rs110}; bra.uni $L__BB0_166; $L__BB0_5: ld.const.u64 %rd37, [params+432]; cvta.to.global.u64 %rd38, %rd37; ld.const.u32 %r73, [params+424]; mad.lo.s32 %r74, %r73, %r61, %r60; mul.wide.u32 %rd39, %r74, 12; add.s64 %rd40, %rd38, %rd39; ld.global.f32 %f10, [%rd40]; mul.f32 %f390, %f10, 0f3456BF95; ld.global.f32 %f11, [%rd40+4]; mul.f32 %f391, %f11, 0f3456BF95; ld.global.f32 %f12, [%rd40+8]; mul.f32 %f392, %f12, 0f3456BF95; abs.f32 %f393, %f1437; div.rn.f32 %f394, %f390, %f393; abs.f32 %f395, %f1438; div.rn.f32 %f396, %f391, %f395; abs.f32 %f397, %f1439; div.rn.f32 %f398, %f392, %f397; abs.f32 %f399, %f394; abs.f32 %f400, %f396; abs.f32 %f401, %f398; mov.f32 %f402, 0f38D1B717; max.f32 %f403, %f399, %f402; max.f32 %f404, %f400, %f402; max.f32 %f405, %f401, %f402; fma.rn.f32 %f13, %f1437, %f403, %f10; fma.rn.f32 %f14, %f1438, %f404, %f11; fma.rn.f32 %f15, %f1439, %f405, %f12; ld.const.u64 %rd41, [params+128]; cvta.to.global.u64 %rd42, %rd41; ld.const.u32 %r75, [params+120]; mad.lo.s32 %r76, %r75, %r61, %r60; mul.wide.u32 %rd43, %r76, 4; add.s64 %rd44, %rd42, %rd43; ld.global.u32 %r454, [%rd44]; setp.lt.s32 %p18, %r4, 1; mov.pred %p17, 0; mov.f32 %f1459, 0f00000000; mov.u32 %r450, 0; mov.f32 %f1460, %f1459; mov.f32 %f1461, %f1459; mov.f32 %f1462, %f1459; mov.f32 %f1463, %f1459; mov.f32 %f1464, %f1459; mov.f32 %f1465, %f1459; mov.f32 %f1466, %f1459; mov.f32 %f1467, %f1459; mov.f32 %f1468, %f1459; mov.f32 %f1469, %f1459; mov.f32 %f1470, %f1459; mov.f32 %f1471, %f1459; mov.f32 %f1472, %f1459; mov.f32 %f1473, %f1459; mov.f32 %f1474, %f1459; mov.pred %p197, %p17; @%p18 bra $L__BB0_23; ld.const.f32 %f16, [params+620]; ld.const.u64 %rd45, [params+624]; cvta.to.global.u64 %rd3, %rd45; ld.const.u64 %rd46, [params+640]; cvta.to.global.u64 %rd47, %rd46; ld.const.u32 %r78, [params+632]; and.b32 %r79, %r61, 255; and.b32 %r80, %r60, 255; mad.lo.s32 %r81, %r78, %r79, %r80; mul.wide.u32 %rd48, %r81, 3; add.s64 %rd4, %rd47, %rd48; ld.const.f32 %f17, [params+660]; mul.f32 %f18, %f13, 0f3456BF95; mul.f32 %f19, %f14, 0f3456BF95; mul.f32 %f20, %f15, 0f3456BF95; ld.const.u64 %rd5, [params+96]; abs.f32 %f549, %f19; abs.f32 %f550, %f18; max.f32 %f551, %f550, %f549; abs.f32 %f552, %f20; max.f32 %f553, %f551, %f552; $L__BB0_7: mul.lo.s32 %r82, %r450, 3; mul.wide.s32 %rd49, %r82, 12; add.s64 %rd6, %rd3, %rd49; ld.global.f32 %f422, [%rd6]; sub.f32 %f423, %f422, %f10; ld.global.f32 %f424, [%rd6+4]; sub.f32 %f425, %f424, %f11; ld.global.f32 %f426, [%rd6+8]; sub.f32 %f427, %f426, %f12; mul.f32 %f428, %f425, %f425; fma.rn.f32 %f429, %f423, %f423, %f428; fma.rn.f32 %f430, %f427, %f427, %f429; sqrt.rn.f32 %f37, %f430; rcp.rn.f32 %f431, %f37; mul.f32 %f38, %f423, %f431; mul.f32 %f39, %f425, %f431; mul.f32 %f40, %f427, %f431; mul.f32 %f432, %f37, %f37; div.rn.f32 %f41, %f16, %f432; ld.global.u8 %rs15, [%rd4]; cvt.rn.f32.u16 %f433, %rs15; div.rn.f32 %f434, %f433, 0fC37F0000; fma.rn.f32 %f435, %f434, 0f3F333333, 0f3F800000; setp.gt.f32 %p20, %f41, %f435; setp.ne.s32 %p21, %r5, 0; and.pred %p22, %p21, %p20; mov.pred %p197, -1; @%p22 bra $L__BB0_23; mul.f32 %f42, %f37, %f17; mov.f32 %f440, 0f40800000; abs.f32 %f44, %f42; setp.lt.f32 %p23, %f44, 0f00800000; mul.f32 %f442, %f44, 0f4B800000; selp.f32 %f443, %f442, %f44, %p23; selp.f32 %f444, 0fC3170000, 0fC2FE0000, %p23; mov.b32 %r83, %f443; and.b32 %r84, %r83, 8388607; or.b32 %r85, %r84, 1065353216; mov.b32 %f445, %r85; shr.u32 %r86, %r83, 23; cvt.rn.f32.u32 %f446, %r86; add.f32 %f447, %f444, %f446; setp.gt.f32 %p24, %f445, 0f3FB504F3; mul.f32 %f448, %f445, 0f3F000000; add.f32 %f449, %f447, 0f3F800000; selp.f32 %f450, %f449, %f447, %p24; selp.f32 %f451, %f448, %f445, %p24; add.f32 %f452, %f451, 0fBF800000; add.f32 %f453, %f451, 0f3F800000; rcp.approx.ftz.f32 %f454, %f453; add.f32 %f455, %f452, %f452; mul.f32 %f456, %f455, %f454; mul.f32 %f457, %f456, %f456; mov.f32 %f458, 0f3C4CAF63; mov.f32 %f459, 0f3B18F0FE; fma.rn.f32 %f460, %f459, %f457, %f458; mov.f32 %f461, 0f3DAAAABD; fma.rn.f32 %f462, %f460, %f457, %f461; mul.rn.f32 %f463, %f462, %f457; mul.rn.f32 %f464, %f463, %f456; sub.f32 %f465, %f452, %f456; add.f32 %f466, %f465, %f465; neg.f32 %f467, %f456; fma.rn.f32 %f468, %f467, %f452, %f466; mul.rn.f32 %f469, %f454, %f468; add.f32 %f470, %f464, %f456; sub.f32 %f471, %f456, %f470; add.f32 %f472, %f464, %f471; add.f32 %f473, %f469, %f472; add.f32 %f474, %f470, %f473; sub.f32 %f475, %f470, %f474; add.f32 %f476, %f473, %f475; mov.f32 %f477, 0f3F317200; mul.rn.f32 %f478, %f450, %f477; mov.f32 %f479, 0f35BFBE8E; mul.rn.f32 %f480, %f450, %f479; add.f32 %f481, %f478, %f474; sub.f32 %f482, %f478, %f481; add.f32 %f483, %f474, %f482; add.f32 %f484, %f476, %f483; add.f32 %f485, %f480, %f484; add.f32 %f486, %f481, %f485; sub.f32 %f487, %f481, %f486; add.f32 %f488, %f485, %f487; mul.rn.f32 %f489, %f440, %f486; neg.f32 %f490, %f489; fma.rn.f32 %f491, %f440, %f486, %f490; fma.rn.f32 %f492, %f440, %f488, %f491; mov.f32 %f493, 0f00000000; fma.rn.f32 %f494, %f493, %f486, %f492; add.rn.f32 %f495, %f489, %f494; neg.f32 %f496, %f495; add.rn.f32 %f497, %f489, %f496; add.rn.f32 %f498, %f497, %f494; mov.b32 %r87, %f495; setp.eq.s32 %p25, %r87, 1118925336; add.s32 %r88, %r87, -1; mov.b32 %f499, %r88; add.f32 %f500, %f498, 0f37000000; selp.f32 %f45, %f500, %f498, %p25; selp.f32 %f501, %f499, %f495, %p25; mov.f32 %f502, 0f3FB8AA3B; mul.rn.f32 %f503, %f501, %f502; cvt.rzi.f32.f32 %f504, %f503; abs.f32 %f505, %f504; setp.gt.f32 %p26, %f505, 0f42FC0000; mov.b32 %r89, %f504; and.b32 %r90, %r89, -2147483648; or.b32 %r91, %r90, 1123811328; mov.b32 %f506, %r91; selp.f32 %f507, %f506, %f504, %p26; mov.f32 %f508, 0fBF317218; fma.rn.f32 %f509, %f507, %f508, %f501; mov.f32 %f510, 0f3102E308; fma.rn.f32 %f511, %f507, %f510, %f509; mul.f32 %f512, %f511, 0f3FB8AA3B; add.f32 %f513, %f507, 0f4B40007F; mov.b32 %r92, %f513; shl.b32 %r93, %r92, 23; mov.b32 %f514, %r93; ex2.approx.ftz.f32 %f515, %f512; mul.f32 %f46, %f515, %f514; setp.eq.f32 %p27, %f46, 0f7F800000; mov.f32 %f1456, 0f7F800000; @%p27 bra $L__BB0_10; fma.rn.f32 %f1456, %f46, %f45, %f46; $L__BB0_10: mov.f32 %f1436, 0f40000000; cvt.rzi.f32.f32 %f1435, %f1436; add.f32 %f1434, %f1435, %f1435; mov.f32 %f1433, 0f40800000; sub.f32 %f1432, %f1433, %f1434; abs.f32 %f1431, %f1432; setp.lt.f32 %p28, %f42, 0f00000000; setp.eq.f32 %p29, %f1431, 0f3F800000; and.pred %p1, %p28, %p29; setp.eq.f32 %p30, %f42, 0f00000000; @%p30 bra $L__BB0_14; bra.uni $L__BB0_11; $L__BB0_14: add.f32 %f520, %f42, %f42; selp.f32 %f1458, %f520, 0f00000000, %p29; bra.uni $L__BB0_15; $L__BB0_11: mov.b32 %r94, %f1456; xor.b32 %r95, %r94, -2147483648; mov.b32 %f516, %r95; selp.f32 %f1458, %f516, %f1456, %p1; setp.geu.f32 %p31, %f42, 0f00000000; @%p31 bra $L__BB0_15; mov.f32 %f517, 0f40800000; cvt.rzi.f32.f32 %f518, %f517; setp.eq.f32 %p32, %f518, 0f40800000; @%p32 bra $L__BB0_15; mov.f32 %f1458, 0f7FFFFFFF; $L__BB0_15: add.f32 %f521, %f44, 0f40800000; mov.b32 %r96, %f521; setp.lt.s32 %p34, %r96, 2139095040; @%p34 bra $L__BB0_20; setp.gtu.f32 %p35, %f44, 0f7F800000; @%p35 bra $L__BB0_19; bra.uni $L__BB0_17; $L__BB0_19: add.f32 %f1458, %f42, 0f40800000; bra.uni $L__BB0_20; $L__BB0_17: setp.neu.f32 %p36, %f44, 0f7F800000; @%p36 bra $L__BB0_20; selp.f32 %f1458, 0fFF800000, 0f7F800000, %p1; $L__BB0_20: mov.f32 %f522, 0f3F800000; sub.f32 %f523, %f522, %f1458; setp.eq.f32 %p37, %f42, 0f3F800000; selp.f32 %f524, 0f00000000, %f523, %p37; cvt.sat.f32.f32 %f525, %f524; mul.f32 %f526, %f41, %f525; ld.global.f32 %f527, [%rd6+12]; mul.f32 %f528, %f38, %f527; ld.global.f32 %f529, [%rd6+16]; mul.f32 %f530, %f39, %f529; neg.f32 %f531, %f530; sub.f32 %f532, %f531, %f528; ld.global.f32 %f533, [%rd6+20]; mul.f32 %f534, %f40, %f533; sub.f32 %f535, %f532, %f534; cvt.sat.f32.f32 %f536, %f535; mul.f32 %f55, %f526, %f536; mul.f32 %f537, %f1438, %f39; fma.rn.f32 %f538, %f1437, %f38, %f537; fma.rn.f32 %f56, %f1439, %f40, %f538; setp.leu.f32 %p38, %f55, 0f3727C5AC; @%p38 bra $L__BB0_22; cvt.sat.f32.f32 %f548, %f56; mov.f32 %f554, 0f38D1B717; max.f32 %f545, %f553, %f554; sub.f32 %f546, %f37, %f545; mov.f32 %f547, 0f00000000; mov.u32 %r133, 2; mov.u32 %r134, 1; mov.u32 %r135, 3; mov.u32 %r138, 1065353216; mov.u32 %r167, 0; // begin inline asm call(%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,%r128),_optix_trace_typed_32,(%r167,%rd5,%f13,%f14,%f15,%f38,%f39,%f40,%f545,%f546,%f547,%r134,%r167,%r134,%r133,%r134,%r135,%r138,%r138,%r138,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167,%r167); // end inline asm mov.b32 %f555, %r97; mov.b32 %f556, %r98; mov.b32 %f557, %r99; ld.global.f32 %f558, [%rd6+24]; mul.f32 %f559, %f558, %f555; ld.global.f32 %f560, [%rd6+28]; mul.f32 %f561, %f560, %f556; ld.global.f32 %f562, [%rd6+32]; mul.f32 %f563, %f562, %f557; mul.f32 %f564, %f55, 0f3EA2F983; mul.f32 %f565, %f564, %f559; mul.f32 %f566, %f564, %f561; mul.f32 %f567, %f564, %f563; fma.rn.f32 %f1471, %f548, %f565, %f1471; fma.rn.f32 %f1472, %f548, %f566, %f1472; fma.rn.f32 %f1473, %f548, %f567, %f1473; fma.rn.f32 %f1465, %f38, %f565, %f1465; fma.rn.f32 %f1466, %f38, %f566, %f1466; fma.rn.f32 %f1467, %f38, %f567, %f1467; fma.rn.f32 %f1462, %f39, %f565, %f1462; fma.rn.f32 %f1463, %f39, %f566, %f1463; fma.rn.f32 %f1464, %f39, %f567, %f1464; fma.rn.f32 %f1459, %f40, %f565, %f1459; fma.rn.f32 %f1460, %f40, %f566, %f1460; fma.rn.f32 %f1461, %f40, %f567, %f1461; add.f32 %f1468, %f1468, %f565; add.f32 %f1469, %f1469, %f566; add.f32 %f1470, %f1470, %f567; add.f32 %f1474, %f1474, %f555; $L__BB0_22: add.s32 %r450, %r450, 1; setp.lt.s32 %p40, %r450, %r4; mov.pred %p197, %p17; @%p40 bra $L__BB0_7; $L__BB0_23: cvt.rn.f32.s32 %f568, %r450; mov.f32 %f569, 0f3F800000; max.f32 %f570, %f568, %f569; rcp.rn.f32 %f571, %f570; mul.f32 %f1554, %f1471, %f571; mul.f32 %f1555, %f1472, %f571; mul.f32 %f1556, %f1473, %f571; div.rn.f32 %f1557, %f1474, %f570; mul.f32 %f1551, %f1468, %f571; mul.f32 %f1552, %f1469, %f571; mul.f32 %f1553, %f1470, %f571; mul.f32 %f1548, %f1465, %f571; mul.f32 %f1549, %f1466, %f571; mul.f32 %f1550, %f1467, %f571; mul.f32 %f1545, %f1462, %f571; mul.f32 %f1546, %f1463, %f571; mul.f32 %f1547, %f1464, %f571; mul.f32 %f1542, %f1459, %f571; mul.f32 %f1543, %f1460, %f571; mul.f32 %f1544, %f1461, %f571; not.pred %p41, %p197; @%p41 bra $L__BB0_54; setp.lt.s32 %p42, %r5, 1; mov.f32 %f586, 0f00000000; mov.f32 %f1520, %f586; mov.f32 %f1519, %f586; mov.f32 %f1518, %f586; mov.f32 %f1517, %f586; mov.f32 %f1516, %f586; mov.f32 %f1515, %f586; mov.f32 %f1514, %f586; mov.f32 %f1513, %f586; mov.f32 %f1512, %f586; mov.f32 %f1511, %f586; mov.f32 %f1510, %f586; mov.f32 %f1509, %f586; mov.f32 %f1508, %f586; mov.f32 %f1507, %f586; mov.f32 %f1506, %f586; @%p42 bra $L__BB0_53; cvt.rn.f32.s32 %f602, %r6; rcp.rn.f32 %f121, %f602; mul.f32 %f122, %f13, 0f3456BF95; mul.f32 %f123, %f14, 0f3456BF95; mul.f32 %f124, %f15, 0f3456BF95; mov.u32 %r168, 0; ld.const.u64 %rd7, [params+96]; add.s64 %rd8, %rd1, 24; mov.u64 %rd51, __cudart_i2opi_f; abs.f32 %f660, %f123; abs.f32 %f661, %f122; max.f32 %f662, %f661, %f660; abs.f32 %f663, %f124; max.f32 %f664, %f662, %f663; mov.u32 %r451, %r168; $L__BB0_26: cvt.rn.f32.s32 %f140, %r451; mov.u32 %r453, %r168; $L__BB0_27: mad.lo.s32 %r170, %r454, 1664525, 1013904223; and.b32 %r171, %r170, 16777215; cvt.rn.f32.u32 %f603, %r171; fma.rn.f32 %f604, %f603, 0f33800000, %f140; mul.f32 %f605, %f121, %f604; mad.lo.s32 %r454, %r170, 1664525, 1013904223; and.b32 %r172, %r454, 16777215; cvt.rn.f32.u32 %f606, %r172; cvt.rn.f32.s32 %f607, %r453; fma.rn.f32 %f608, %f606, 0f33800000, %f607; mul.f32 %f609, %f121, %f608; fma.rn.f32 %f610, %f121, %f604, %f605; mov.f32 %f611, 0f3F800000; sub.f32 %f156, %f611, %f610; mul.f32 %f612, %f156, %f156; sub.f32 %f613, %f611, %f612; mov.f32 %f614, 0f00000000; max.f32 %f615, %f614, %f613; sqrt.rn.f32 %f157, %f615; mul.f32 %f158, %f609, 0f40C90FDB; mul.f32 %f616, %f158, 0f3F22F983; cvt.rni.s32.f32 %r461, %f616; cvt.rn.f32.s32 %f617, %r461; mov.f32 %f618, 0fBFC90FDA; fma.rn.f32 %f619, %f617, %f618, %f158; mov.f32 %f620, 0fB3A22168; fma.rn.f32 %f621, %f617, %f620, %f619; mov.f32 %f622, 0fA7C234C5; fma.rn.f32 %f1524, %f617, %f622, %f621; abs.f32 %f160, %f158; setp.ltu.f32 %p43, %f160, 0f47CE4780; mov.u32 %r458, %r461; mov.f32 %f1521, %f1524; @%p43 bra $L__BB0_35; setp.eq.f32 %p44, %f160, 0f7F800000; @%p44 bra $L__BB0_34; bra.uni $L__BB0_29; $L__BB0_34: mov.f32 %f625, 0f00000000; mul.rn.f32 %f1521, %f158, %f625; mov.u32 %r458, 0; bra.uni $L__BB0_35; $L__BB0_29: mov.b32 %r19, %f158; bfe.u32 %r174, %r19, 23, 8; add.s32 %r20, %r174, -128; shl.b32 %r175, %r19, 8; or.b32 %r21, %r175, -2147483648; shr.u32 %r22, %r20, 5; mov.u64 %rd122, 0; mov.u32 %r455, 0; mov.u64 %rd120, %rd1; mov.u64 %rd121, %rd51; $L__BB0_30: .pragma "nounroll"; ld.global.nc.u32 %r176, [%rd121]; mad.wide.u32 %rd53, %r176, %r21, %rd122; shr.u64 %rd122, %rd53, 32; st.local.u32 [%rd120], %rd53; add.s64 %rd121, %rd121, 4; add.s64 %rd120, %rd120, 4; add.s32 %r455, %r455, 1; setp.ne.s32 %p45, %r455, 6; @%p45 bra $L__BB0_30; st.local.u32 [%rd8], %rd122; mov.u32 %r177, 4; sub.s32 %r25, %r177, %r22; mov.u32 %r178, 6; sub.s32 %r179, %r178, %r22; mul.wide.s32 %rd54, %r179, 4; add.s64 %rd55, %rd1, %rd54; ld.local.u32 %r456, [%rd55]; ld.local.u32 %r457, [%rd55+-4]; and.b32 %r28, %r20, 31; setp.eq.s32 %p46, %r28, 0; @%p46 bra $L__BB0_33; mov.u32 %r180, 32; sub.s32 %r181, %r180, %r28; shr.u32 %r182, %r457, %r181; shl.b32 %r183, %r456, %r28; add.s32 %r456, %r182, %r183; mul.wide.s32 %rd56, %r25, 4; add.s64 %rd57, %rd1, %rd56; ld.local.u32 %r184, [%rd57]; shr.u32 %r185, %r184, %r181; shl.b32 %r186, %r457, %r28; add.s32 %r457, %r185, %r186; $L__BB0_33: and.b32 %r187, %r19, -2147483648; shr.u32 %r188, %r457, 30; shl.b32 %r189, %r456, 2; or.b32 %r190, %r188, %r189; shr.u32 %r191, %r190, 31; shr.u32 %r192, %r456, 30; add.s32 %r193, %r191, %r192; neg.s32 %r194, %r193; setp.eq.s32 %p47, %r187, 0; selp.b32 %r458, %r193, %r194, %p47; setp.ne.s32 %p48, %r191, 0; xor.b32 %r195, %r187, -2147483648; selp.b32 %r196, %r195, %r187, %p48; selp.b32 %r197, -1, 0, %p48; xor.b32 %r198, %r190, %r197; shl.b32 %r199, %r457, 2; xor.b32 %r200, %r199, %r197; cvt.u64.u32 %rd58, %r198; cvt.u64.u32 %rd59, %r200; bfi.b64 %rd60, %rd58, %rd59, 32, 32; cvt.rn.f64.s64 %fd1, %rd60; mul.f64 %fd2, %fd1, 0d3BF921FB54442D19; cvt.rn.f32.f64 %f623, %fd2; setp.eq.s32 %p49, %r196, 0; neg.f32 %f624, %f623; selp.f32 %f1521, %f623, %f624, %p49; $L__BB0_35: add.s32 %r35, %r458, 1; and.b32 %r36, %r35, 1; setp.eq.s32 %p50, %r36, 0; selp.f32 %f164, %f1521, 0f3F800000, %p50; mul.rn.f32 %f165, %f1521, %f1521; mov.f32 %f1522, 0fB94D4153; @%p50 bra $L__BB0_37; mov.f32 %f627, 0fBAB607ED; mov.f32 %f628, 0f37CBAC00; fma.rn.f32 %f1522, %f628, %f165, %f627; $L__BB0_37: selp.f32 %f629, 0f3C0885E4, 0f3D2AAABB, %p50; fma.rn.f32 %f630, %f1522, %f165, %f629; selp.f32 %f631, 0fBE2AAAA8, 0fBEFFFFFF, %p50; fma.rn.f32 %f632, %f630, %f165, %f631; mov.f32 %f633, 0f00000000; fma.rn.f32 %f634, %f165, %f164, %f633; fma.rn.f32 %f1523, %f632, %f634, %f164; and.b32 %r202, %r35, 2; setp.eq.s32 %p52, %r202, 0; @%p52 bra $L__BB0_39; mov.f32 %f636, 0fBF800000; fma.rn.f32 %f1523, %f1523, %f636, %f633; $L__BB0_39: @%p43 bra $L__BB0_47; setp.eq.f32 %p54, %f160, 0f7F800000; @%p54 bra $L__BB0_46; bra.uni $L__BB0_41; $L__BB0_46: mov.f32 %f639, 0f00000000; mul.rn.f32 %f1524, %f158, %f639; mov.u32 %r461, 0; bra.uni $L__BB0_47; $L__BB0_41: mov.b32 %r37, %f158; bfe.u32 %r203, %r37, 23, 8; add.s32 %r38, %r203, -128; shl.b32 %r204, %r37, 8; or.b32 %r39, %r204, -2147483648; shr.u32 %r40, %r38, 5; mov.u64 %rd123, 0; mov.u64 %rd124, %rd123; $L__BB0_42: .pragma "nounroll"; shl.b64 %rd63, %rd123, 2; mov.u64 %rd64, __cudart_i2opi_f; add.s64 %rd65, %rd64, %rd63; ld.global.nc.u32 %r205, [%rd65]; mad.wide.u32 %rd66, %r205, %r39, %rd124; shr.u64 %rd124, %rd66, 32; add.s64 %rd67, %rd1, %rd63; st.local.u32 [%rd67], %rd66; cvt.u32.u64 %r206, %rd123; add.s32 %r207, %r206, 1; cvt.s64.s32 %rd123, %r207; setp.ne.s32 %p55, %r207, 6; @%p55 bra $L__BB0_42; st.local.u32 [%rd8], %rd124; mov.u32 %r208, 4; sub.s32 %r41, %r208, %r40; mov.u32 %r209, 6; sub.s32 %r210, %r209, %r40; mul.wide.s32 %rd68, %r210, 4; add.s64 %rd69, %rd1, %rd68; ld.local.u32 %r459, [%rd69]; ld.local.u32 %r460, [%rd69+-4]; and.b32 %r44, %r38, 31; setp.eq.s32 %p56, %r44, 0; @%p56 bra $L__BB0_45; mov.u32 %r211, 32; sub.s32 %r212, %r211, %r44; shr.u32 %r213, %r460, %r212; shl.b32 %r214, %r459, %r44; add.s32 %r459, %r213, %r214; mul.wide.s32 %rd70, %r41, 4; add.s64 %rd71, %rd1, %rd70; ld.local.u32 %r215, [%rd71]; shr.u32 %r216, %r215, %r212; shl.b32 %r217, %r460, %r44; add.s32 %r460, %r216, %r217; $L__BB0_45: and.b32 %r218, %r37, -2147483648; shr.u32 %r219, %r460, 30; shl.b32 %r220, %r459, 2; or.b32 %r221, %r219, %r220; shr.u32 %r222, %r221, 31; shr.u32 %r223, %r459, 30; add.s32 %r224, %r222, %r223; neg.s32 %r225, %r224; setp.eq.s32 %p57, %r218, 0; selp.b32 %r461, %r224, %r225, %p57; setp.ne.s32 %p58, %r222, 0; xor.b32 %r226, %r218, -2147483648; selp.b32 %r227, %r226, %r218, %p58; selp.b32 %r228, -1, 0, %p58; xor.b32 %r229, %r221, %r228; shl.b32 %r230, %r460, 2; xor.b32 %r231, %r230, %r228; cvt.u64.u32 %rd72, %r229; cvt.u64.u32 %rd73, %r231; bfi.b64 %rd74, %rd72, %rd73, 32, 32; cvt.rn.f64.s64 %fd3, %rd74; mul.f64 %fd4, %fd3, 0d3BF921FB54442D19; cvt.rn.f32.f64 %f637, %fd4; setp.eq.s32 %p59, %r227, 0; neg.f32 %f638, %f637; selp.f32 %f1524, %f637, %f638, %p59; $L__BB0_47: mul.f32 %f174, %f157, %f1523; and.b32 %r51, %r461, 1; setp.eq.s32 %p60, %r51, 0; selp.f32 %f175, %f1524, 0f3F800000, %p60; mul.rn.f32 %f176, %f1524, %f1524; mov.f32 %f1525, 0fB94D4153; @%p60 bra $L__BB0_49; mov.f32 %f641, 0fBAB607ED; mov.f32 %f642, 0f37CBAC00; fma.rn.f32 %f1525, %f642, %f176, %f641; $L__BB0_49: selp.f32 %f643, 0f3C0885E4, 0f3D2AAABB, %p60; fma.rn.f32 %f644, %f1525, %f176, %f643; selp.f32 %f645, 0fBE2AAAA8, 0fBEFFFFFF, %p60; fma.rn.f32 %f646, %f644, %f176, %f645; mov.f32 %f647, 0f00000000; fma.rn.f32 %f648, %f176, %f175, %f647; fma.rn.f32 %f1526, %f646, %f648, %f175; and.b32 %r233, %r461, 2; setp.eq.s32 %p62, %r233, 0; @%p62 bra $L__BB0_51; mov.f32 %f650, 0fBF800000; fma.rn.f32 %f1526, %f1526, %f650, %f647; $L__BB0_51: mov.f32 %f665, 0f38D1B717; max.f32 %f657, %f664, %f665; mul.f32 %f655, %f157, %f1526; mov.f32 %f658, 0f6C4ECB8F; mov.u32 %r267, 1; mov.u32 %r270, 2; mov.u32 %r272, 4; mov.u32 %r275, 1065353216; mov.u32 %r276, 2139095039; mov.u32 %r304, 0; // begin inline asm call(%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,%r265),_optix_trace_typed_32,(%r304,%rd7,%f13,%f14,%f15,%f174,%f655,%f156,%f657,%f658,%f647,%r267,%r304,%r304,%r270,%r304,%r272,%r275,%r275,%r275,%r276,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304,%r304); // end inline asm mov.b32 %f666, %r234; mov.b32 %f667, %r235; mov.b32 %f668, %r236; mov.b32 %f669, %r237; max.f32 %f670, %f666, %f647; max.f32 %f671, %f667, %f647; max.f32 %f672, %f668, %f647; setp.ltu.f32 %p63, %f669, 0f00000000; selp.f32 %f673, %f672, 0f00000000, %p63; selp.f32 %f674, %f671, 0f00000000, %p63; selp.f32 %f675, %f670, 0f00000000, %p63; fma.rn.f32 %f1514, %f174, %f675, %f1514; fma.rn.f32 %f1513, %f174, %f674, %f1513; fma.rn.f32 %f1512, %f174, %f673, %f1512; fma.rn.f32 %f1517, %f655, %f675, %f1517; fma.rn.f32 %f1516, %f655, %f674, %f1516; fma.rn.f32 %f1515, %f655, %f673, %f1515; fma.rn.f32 %f1520, %f156, %f675, %f1520; fma.rn.f32 %f1519, %f156, %f674, %f1519; fma.rn.f32 %f1518, %f156, %f673, %f1518; add.f32 %f1511, %f1511, %f675; add.f32 %f1510, %f1510, %f674; add.f32 %f1509, %f1509, %f673; mul.f32 %f676, %f1438, %f655; fma.rn.f32 %f677, %f1437, %f174, %f676; fma.rn.f32 %f678, %f1439, %f156, %f677; cvt.sat.f32.f32 %f679, %f678; fma.rn.f32 %f1508, %f675, %f679, %f1508; fma.rn.f32 %f1507, %f674, %f679, %f1507; fma.rn.f32 %f1506, %f673, %f679, %f1506; add.s32 %r453, %r453, 1; setp.lt.s32 %p64, %r453, %r6; @%p64 bra $L__BB0_27; add.s32 %r451, %r451, 1; setp.lt.s32 %p65, %r451, %r6; @%p65 bra $L__BB0_26; $L__BB0_53: mul.lo.s32 %r305, %r6, %r6; cvt.rn.f32.s32 %f680, %r305; rcp.rn.f32 %f681, %f680; mul.f32 %f682, %f681, %f1508; mul.f32 %f683, %f681, %f1507; mul.f32 %f684, %f681, %f1506; div.rn.f32 %f1557, %f586, %f680; mul.f32 %f1551, %f681, %f1511; mul.f32 %f1552, %f681, %f1510; mul.f32 %f1553, %f681, %f1509; mul.f32 %f1548, %f681, %f1514; mul.f32 %f1549, %f681, %f1513; mul.f32 %f1550, %f681, %f1512; mul.f32 %f1545, %f681, %f1517; mul.f32 %f1546, %f681, %f1516; mul.f32 %f1547, %f681, %f1515; mul.f32 %f1542, %f681, %f1520; mul.f32 %f1543, %f681, %f1519; mul.f32 %f1544, %f681, %f1518; fma.rn.f32 %f1554, %f681, %f1508, %f682; fma.rn.f32 %f1555, %f681, %f1507, %f683; fma.rn.f32 %f1556, %f681, %f1506, %f684; $L__BB0_54: ld.const.u32 %r448, [params+616]; setp.lt.s32 %p66, %r448, 0; selp.f32 %f244, %f1554, %f1557, %p66; ld.const.u32 %r54, [params+104]; and.b32 %r306, %r54, 8; setp.eq.s32 %p67, %r306, 0; @%p67 bra $L__BB0_68; ld.const.u64 %rd76, [params+192]; cvta.to.global.u64 %rd19, %rd76; ld.const.u32 %r307, [params+184]; mad.lo.s32 %r308, %r307, %r8, %r7; cvt.u64.u32 %rd20, %r308; mov.f32 %f687, 0f3E68BA2E; cvt.rzi.f32.f32 %f688, %f687; add.f32 %f689, %f688, %f688; mov.f32 %f690, 0f3EE8BA2E; sub.f32 %f691, %f690, %f689; abs.f32 %f245, %f691; abs.f32 %f246, %f244; setp.lt.f32 %p68, %f246, 0f00800000; mul.f32 %f692, %f246, 0f4B800000; selp.f32 %f693, %f692, %f246, %p68; selp.f32 %f694, 0fC3170000, 0fC2FE0000, %p68; mov.b32 %r309, %f693; and.b32 %r310, %r309, 8388607; or.b32 %r311, %r310, 1065353216; mov.b32 %f695, %r311; shr.u32 %r312, %r309, 23; cvt.rn.f32.u32 %f696, %r312; add.f32 %f697, %f694, %f696; setp.gt.f32 %p69, %f695, 0f3FB504F3; mul.f32 %f698, %f695, 0f3F000000; add.f32 %f699, %f697, 0f3F800000; selp.f32 %f700, %f699, %f697, %p69; selp.f32 %f701, %f698, %f695, %p69; add.f32 %f702, %f701, 0fBF800000; add.f32 %f703, %f701, 0f3F800000; rcp.approx.ftz.f32 %f704, %f703; add.f32 %f705, %f702, %f702; mul.f32 %f706, %f705, %f704; mul.f32 %f707, %f706, %f706; mov.f32 %f708, 0f3C4CAF63; mov.f32 %f709, 0f3B18F0FE; fma.rn.f32 %f710, %f709, %f707, %f708; mov.f32 %f711, 0f3DAAAABD; fma.rn.f32 %f712, %f710, %f707, %f711; mul.rn.f32 %f713, %f712, %f707; mul.rn.f32 %f714, %f713, %f706; sub.f32 %f715, %f702, %f706; add.f32 %f716, %f715, %f715; neg.f32 %f717, %f706; fma.rn.f32 %f718, %f717, %f702, %f716; mul.rn.f32 %f719, %f704, %f718; add.f32 %f720, %f714, %f706; sub.f32 %f721, %f706, %f720; add.f32 %f722, %f714, %f721; add.f32 %f723, %f719, %f722; add.f32 %f724, %f720, %f723; sub.f32 %f725, %f720, %f724; add.f32 %f726, %f723, %f725; mov.f32 %f727, 0f3F317200; mul.rn.f32 %f728, %f700, %f727; mov.f32 %f729, 0f35BFBE8E; mul.rn.f32 %f730, %f700, %f729; add.f32 %f731, %f728, %f724; sub.f32 %f732, %f728, %f731; add.f32 %f733, %f724, %f732; add.f32 %f734, %f726, %f733; add.f32 %f735, %f730, %f734; add.f32 %f736, %f731, %f735; sub.f32 %f737, %f731, %f736; add.f32 %f738, %f735, %f737; mul.rn.f32 %f739, %f690, %f736; neg.f32 %f740, %f739; fma.rn.f32 %f741, %f690, %f736, %f740; fma.rn.f32 %f742, %f690, %f738, %f741; mov.f32 %f743, 0f00000000; fma.rn.f32 %f744, %f743, %f736, %f742; add.rn.f32 %f745, %f739, %f744; neg.f32 %f746, %f745; add.rn.f32 %f747, %f739, %f746; add.rn.f32 %f748, %f747, %f744; mov.b32 %r313, %f745; setp.eq.s32 %p70, %r313, 1118925336; add.s32 %r314, %r313, -1; mov.b32 %f749, %r314; add.f32 %f750, %f748, 0f37000000; selp.f32 %f247, %f750, %f748, %p70; selp.f32 %f751, %f749, %f745, %p70; mov.f32 %f752, 0f3FB8AA3B; mul.rn.f32 %f753, %f751, %f752; cvt.rzi.f32.f32 %f754, %f753; abs.f32 %f755, %f754; setp.gt.f32 %p71, %f755, 0f42FC0000; mov.b32 %r315, %f754; and.b32 %r316, %r315, -2147483648; or.b32 %r317, %r316, 1123811328; mov.b32 %f756, %r317; selp.f32 %f757, %f756, %f754, %p71; mov.f32 %f758, 0fBF317218; fma.rn.f32 %f759, %f757, %f758, %f751; mov.f32 %f760, 0f3102E308; fma.rn.f32 %f761, %f757, %f760, %f759; mul.f32 %f762, %f761, 0f3FB8AA3B; add.f32 %f763, %f757, 0f4B40007F; mov.b32 %r318, %f763; shl.b32 %r319, %r318, 23; mov.b32 %f764, %r319; ex2.approx.ftz.f32 %f765, %f762; mul.f32 %f248, %f765, %f764; setp.eq.f32 %p72, %f248, 0f7F800000; mov.f32 %f1558, 0f7F800000; @%p72 bra $L__BB0_57; fma.rn.f32 %f1558, %f248, %f247, %f248; $L__BB0_57: setp.lt.f32 %p73, %f244, 0f00000000; setp.eq.f32 %p74, %f245, 0f3F800000; and.pred %p3, %p73, %p74; setp.eq.f32 %p75, %f244, 0f00000000; @%p75 bra $L__BB0_61; bra.uni $L__BB0_58; $L__BB0_61: add.f32 %f770, %f244, %f244; selp.f32 %f1560, %f770, 0f00000000, %p74; bra.uni $L__BB0_62; $L__BB0_165: mov.f32 %f1394, 0f00000000; mov.u32 %r463, 0; // begin inline asm { cvt.rn.f16.f32 %rs113, %f1394;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs112, %f1394;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs111, %f1394;} // end inline asm mov.u16 %rs114, 0; st.global.v4.u16 [%rd28], {%rs111, %rs112, %rs113, %rs114}; $L__BB0_166: ld.const.u64 %rd108, [params+256]; cvta.to.global.u64 %rd109, %rd108; ld.const.u32 %r440, [params+248]; mad.lo.s32 %r441, %r440, %r8, %r7; mul.wide.u32 %rd110, %r441, 8; add.s64 %rd29, %rd109, %rd110; setp.eq.s32 %p193, %r463, 0; @%p193 bra $L__BB0_168; ld.global.v4.u16 {%rs121, %rs122, %rs123, %rs124}, [%rd29]; // begin inline asm { cvt.f32.f16 %f1395, %rs121;} // end inline asm // begin inline asm { cvt.f32.f16 %f1396, %rs122;} // end inline asm // begin inline asm { cvt.f32.f16 %f1397, %rs123;} // end inline asm add.f32 %f1398, %f1395, 0f00000000; add.f32 %f1399, %f1396, 0f00000000; add.f32 %f1400, %f1397, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs120, %f1400;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs119, %f1399;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs118, %f1398;} // end inline asm mov.u16 %rs125, 0; st.global.v4.u16 [%rd29], {%rs118, %rs119, %rs120, %rs125}; bra.uni $L__BB0_169; $L__BB0_168: mov.f32 %f1403, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs128, %f1403;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs127, %f1403;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs126, %f1403;} // end inline asm mov.u16 %rs129, 0; st.global.v4.u16 [%rd29], {%rs126, %rs127, %rs128, %rs129}; $L__BB0_169: ld.const.u64 %rd111, [params+272]; cvta.to.global.u64 %rd112, %rd111; ld.const.u32 %r442, [params+264]; mad.lo.s32 %r443, %r442, %r8, %r7; mul.wide.u32 %rd113, %r443, 8; add.s64 %rd30, %rd112, %rd113; @%p193 bra $L__BB0_171; ld.global.v4.u16 {%rs136, %rs137, %rs138, %rs139}, [%rd30]; // begin inline asm { cvt.f32.f16 %f1404, %rs136;} // end inline asm // begin inline asm { cvt.f32.f16 %f1405, %rs137;} // end inline asm // begin inline asm { cvt.f32.f16 %f1406, %rs138;} // end inline asm add.f32 %f1407, %f1404, 0f00000000; add.f32 %f1408, %f1405, 0f00000000; add.f32 %f1409, %f1406, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs135, %f1409;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs134, %f1408;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs133, %f1407;} // end inline asm mov.u16 %rs140, 0; st.global.v4.u16 [%rd30], {%rs133, %rs134, %rs135, %rs140}; bra.uni $L__BB0_172; $L__BB0_171: mov.f32 %f1412, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs143, %f1412;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs142, %f1412;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs141, %f1412;} // end inline asm mov.u16 %rs144, 0; st.global.v4.u16 [%rd30], {%rs141, %rs142, %rs143, %rs144}; $L__BB0_172: ld.const.u64 %rd114, [params+288]; cvta.to.global.u64 %rd115, %rd114; ld.const.u32 %r444, [params+280]; mad.lo.s32 %r445, %r444, %r8, %r7; mul.wide.u32 %rd116, %r445, 8; add.s64 %rd31, %rd115, %rd116; @%p193 bra $L__BB0_174; ld.global.v4.u16 {%rs151, %rs152, %rs153, %rs154}, [%rd31]; // begin inline asm { cvt.f32.f16 %f1413, %rs151;} // end inline asm // begin inline asm { cvt.f32.f16 %f1414, %rs152;} // end inline asm // begin inline asm { cvt.f32.f16 %f1415, %rs153;} // end inline asm add.f32 %f1416, %f1413, 0f00000000; add.f32 %f1417, %f1414, 0f00000000; add.f32 %f1418, %f1415, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs150, %f1418;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs149, %f1417;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs148, %f1416;} // end inline asm mov.u16 %rs155, 0; st.global.v4.u16 [%rd31], {%rs148, %rs149, %rs150, %rs155}; bra.uni $L__BB0_175; $L__BB0_174: mov.f32 %f1421, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs158, %f1421;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs157, %f1421;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs156, %f1421;} // end inline asm mov.u16 %rs159, 0; st.global.v4.u16 [%rd31], {%rs156, %rs157, %rs158, %rs159}; $L__BB0_175: ld.const.u64 %rd117, [params+304]; cvta.to.global.u64 %rd118, %rd117; ld.const.u32 %r446, [params+296]; mad.lo.s32 %r447, %r446, %r8, %r7; mul.wide.u32 %rd119, %r447, 8; add.s64 %rd32, %rd118, %rd119; @%p193 bra $L__BB0_177; ld.global.v4.u16 {%rs166, %rs167, %rs168, %rs169}, [%rd32]; // begin inline asm { cvt.f32.f16 %f1422, %rs166;} // end inline asm // begin inline asm { cvt.f32.f16 %f1423, %rs167;} // end inline asm // begin inline asm { cvt.f32.f16 %f1424, %rs168;} // end inline asm add.f32 %f1425, %f1422, 0f00000000; add.f32 %f1426, %f1423, 0f00000000; add.f32 %f1427, %f1424, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs165, %f1427;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs164, %f1426;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs163, %f1425;} // end inline asm mov.u16 %rs170, 0; st.global.v4.u16 [%rd32], {%rs163, %rs164, %rs165, %rs170}; bra.uni $L__BB0_178; $L__BB0_177: mov.f32 %f1430, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs173, %f1430;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs172, %f1430;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs171, %f1430;} // end inline asm mov.u16 %rs174, 0; st.global.v4.u16 [%rd32], {%rs171, %rs172, %rs173, %rs174}; bra.uni $L__BB0_178; $L__BB0_58: mov.b32 %r320, %f1558; xor.b32 %r321, %r320, -2147483648; mov.b32 %f766, %r321; selp.f32 %f1560, %f766, %f1558, %p3; setp.geu.f32 %p76, %f244, 0f00000000; @%p76 bra $L__BB0_62; mov.f32 %f767, 0f3EE8BA2E; cvt.rzi.f32.f32 %f768, %f767; setp.eq.f32 %p77, %f768, 0f3EE8BA2E; @%p77 bra $L__BB0_62; mov.f32 %f1560, 0f7FFFFFFF; $L__BB0_62: add.f32 %f771, %f246, 0f3EE8BA2E; mov.b32 %r322, %f771; setp.lt.s32 %p79, %r322, 2139095040; @%p79 bra $L__BB0_67; setp.gtu.f32 %p80, %f246, 0f7F800000; @%p80 bra $L__BB0_66; bra.uni $L__BB0_64; $L__BB0_66: add.f32 %f1560, %f244, 0f3EE8BA2E; bra.uni $L__BB0_67; $L__BB0_64: setp.neu.f32 %p81, %f246, 0f7F800000; @%p81 bra $L__BB0_67; selp.f32 %f1560, 0fFF800000, 0f7F800000, %p3; $L__BB0_67: mul.f32 %f772, %f1560, 0f437F0000; setp.eq.f32 %p82, %f244, 0f3F800000; selp.f32 %f773, 0f437F0000, %f772, %p82; cvt.rzi.u32.f32 %r323, %f773; shl.b64 %rd77, %rd20, 1; add.s64 %rd78, %rd19, %rd77; cvt.u16.u32 %rs16, %r323; mov.u16 %rs17, 255; st.global.v2.u8 [%rd78], {%rs16, %rs17}; $L__BB0_68: ld.const.v2.f32 {%f774, %f775}, [params+648]; mul.f32 %f259, %f1554, %f774; mul.f32 %f260, %f1555, %f775; ld.const.f32 %f261, [params+656]; mul.f32 %f262, %f1556, %f261; and.b32 %r324, %r54, 1; setp.eq.b32 %p83, %r324, 1; mov.pred %p84, 0; xor.pred %p85, %p83, %p84; not.pred %p86, %p85; @%p86 bra $L__BB0_142; mov.f32 %f777, 0f3E666666; cvt.rzi.f32.f32 %f778, %f777; add.f32 %f779, %f778, %f778; mov.f32 %f780, 0f3EE66666; sub.f32 %f781, %f780, %f779; abs.f32 %f263, %f781; abs.f32 %f264, %f259; setp.lt.f32 %p87, %f264, 0f00800000; mul.f32 %f782, %f264, 0f4B800000; selp.f32 %f783, %f782, %f264, %p87; selp.f32 %f784, 0fC3170000, 0fC2FE0000, %p87; mov.b32 %r325, %f783; and.b32 %r326, %r325, 8388607; or.b32 %r327, %r326, 1065353216; mov.b32 %f785, %r327; shr.u32 %r328, %r325, 23; cvt.rn.f32.u32 %f786, %r328; add.f32 %f787, %f784, %f786; setp.gt.f32 %p88, %f785, 0f3FB504F3; mul.f32 %f788, %f785, 0f3F000000; add.f32 %f789, %f787, 0f3F800000; selp.f32 %f790, %f789, %f787, %p88; selp.f32 %f791, %f788, %f785, %p88; add.f32 %f792, %f791, 0fBF800000; add.f32 %f793, %f791, 0f3F800000; rcp.approx.ftz.f32 %f794, %f793; add.f32 %f795, %f792, %f792; mul.f32 %f796, %f795, %f794; mul.f32 %f797, %f796, %f796; mov.f32 %f798, 0f3C4CAF63; mov.f32 %f799, 0f3B18F0FE; fma.rn.f32 %f800, %f799, %f797, %f798; mov.f32 %f801, 0f3DAAAABD; fma.rn.f32 %f802, %f800, %f797, %f801; mul.rn.f32 %f803, %f802, %f797; mul.rn.f32 %f804, %f803, %f796; sub.f32 %f805, %f792, %f796; add.f32 %f806, %f805, %f805; neg.f32 %f807, %f796; fma.rn.f32 %f808, %f807, %f792, %f806; mul.rn.f32 %f809, %f794, %f808; add.f32 %f810, %f804, %f796; sub.f32 %f811, %f796, %f810; add.f32 %f812, %f804, %f811; add.f32 %f813, %f809, %f812; add.f32 %f814, %f810, %f813; sub.f32 %f815, %f810, %f814; add.f32 %f816, %f813, %f815; mov.f32 %f817, 0f3F317200; mul.rn.f32 %f818, %f790, %f817; mov.f32 %f819, 0f35BFBE8E; mul.rn.f32 %f820, %f790, %f819; add.f32 %f821, %f818, %f814; sub.f32 %f822, %f818, %f821; add.f32 %f823, %f814, %f822; add.f32 %f824, %f816, %f823; add.f32 %f825, %f820, %f824; add.f32 %f826, %f821, %f825; sub.f32 %f827, %f821, %f826; add.f32 %f828, %f825, %f827; mul.rn.f32 %f829, %f780, %f826; neg.f32 %f830, %f829; fma.rn.f32 %f831, %f780, %f826, %f830; fma.rn.f32 %f832, %f780, %f828, %f831; mov.f32 %f833, 0f00000000; fma.rn.f32 %f834, %f833, %f826, %f832; add.rn.f32 %f835, %f829, %f834; neg.f32 %f836, %f835; add.rn.f32 %f837, %f829, %f836; add.rn.f32 %f838, %f837, %f834; mov.b32 %r329, %f835; setp.eq.s32 %p89, %r329, 1118925336; add.s32 %r330, %r329, -1; mov.b32 %f839, %r330; add.f32 %f840, %f838, 0f37000000; selp.f32 %f265, %f840, %f838, %p89; selp.f32 %f841, %f839, %f835, %p89; mov.f32 %f842, 0f3FB8AA3B; mul.rn.f32 %f843, %f841, %f842; cvt.rzi.f32.f32 %f844, %f843; abs.f32 %f845, %f844; setp.gt.f32 %p90, %f845, 0f42FC0000; mov.b32 %r331, %f844; and.b32 %r332, %r331, -2147483648; or.b32 %r333, %r332, 1123811328; mov.b32 %f846, %r333; selp.f32 %f847, %f846, %f844, %p90; mov.f32 %f848, 0fBF317218; fma.rn.f32 %f849, %f847, %f848, %f841; mov.f32 %f850, 0f3102E308; fma.rn.f32 %f851, %f847, %f850, %f849; mul.f32 %f852, %f851, 0f3FB8AA3B; add.f32 %f853, %f847, 0f4B40007F; mov.b32 %r334, %f853; shl.b32 %r335, %r334, 23; mov.b32 %f854, %r335; ex2.approx.ftz.f32 %f855, %f852; mul.f32 %f266, %f855, %f854; setp.eq.f32 %p91, %f266, 0f7F800000; mov.f32 %f1561, 0f7F800000; @%p91 bra $L__BB0_71; fma.rn.f32 %f1561, %f266, %f265, %f266; $L__BB0_71: setp.lt.f32 %p92, %f259, 0f00000000; setp.eq.f32 %p93, %f263, 0f3F800000; and.pred %p4, %p92, %p93; setp.eq.f32 %p94, %f259, 0f00000000; @%p94 bra $L__BB0_75; bra.uni $L__BB0_72; $L__BB0_75: add.f32 %f860, %f259, %f259; selp.f32 %f1563, %f860, 0f00000000, %p93; bra.uni $L__BB0_76; $L__BB0_72: mov.b32 %r336, %f1561; xor.b32 %r337, %r336, -2147483648; mov.b32 %f856, %r337; selp.f32 %f1563, %f856, %f1561, %p4; setp.geu.f32 %p95, %f259, 0f00000000; @%p95 bra $L__BB0_76; mov.f32 %f857, 0f3EE66666; cvt.rzi.f32.f32 %f858, %f857; setp.eq.f32 %p96, %f858, 0f3EE66666; @%p96 bra $L__BB0_76; mov.f32 %f1563, 0f7FFFFFFF; $L__BB0_76: add.f32 %f861, %f264, 0f3EE66666; mov.b32 %r338, %f861; setp.lt.s32 %p98, %r338, 2139095040; @%p98 bra $L__BB0_81; setp.gtu.f32 %p99, %f264, 0f7F800000; @%p99 bra $L__BB0_80; bra.uni $L__BB0_78; $L__BB0_80: add.f32 %f1563, %f259, 0f3EE66666; bra.uni $L__BB0_81; $L__BB0_78: setp.neu.f32 %p100, %f264, 0f7F800000; @%p100 bra $L__BB0_81; selp.f32 %f1563, 0fFF800000, 0f7F800000, %p4; $L__BB0_81: setp.eq.f32 %p101, %f259, 0f3F800000; selp.f32 %f275, 0f3F800000, %f1563, %p101; abs.f32 %f276, %f260; setp.lt.f32 %p102, %f276, 0f00800000; mul.f32 %f863, %f276, 0f4B800000; selp.f32 %f864, %f863, %f276, %p102; selp.f32 %f865, 0fC3170000, 0fC2FE0000, %p102; mov.b32 %r339, %f864; and.b32 %r340, %r339, 8388607; or.b32 %r341, %r340, 1065353216; mov.b32 %f866, %r341; shr.u32 %r342, %r339, 23; cvt.rn.f32.u32 %f867, %r342; add.f32 %f868, %f865, %f867; setp.gt.f32 %p103, %f866, 0f3FB504F3; mul.f32 %f869, %f866, 0f3F000000; add.f32 %f870, %f868, 0f3F800000; selp.f32 %f871, %f870, %f868, %p103; selp.f32 %f872, %f869, %f866, %p103; add.f32 %f873, %f872, 0fBF800000; add.f32 %f874, %f872, 0f3F800000; rcp.approx.ftz.f32 %f875, %f874; add.f32 %f876, %f873, %f873; mul.f32 %f877, %f876, %f875; mul.f32 %f878, %f877, %f877; mov.f32 %f879, 0f3C4CAF63; mov.f32 %f880, 0f3B18F0FE; fma.rn.f32 %f881, %f880, %f878, %f879; mov.f32 %f882, 0f3DAAAABD; fma.rn.f32 %f883, %f881, %f878, %f882; mul.rn.f32 %f884, %f883, %f878; mul.rn.f32 %f885, %f884, %f877; sub.f32 %f886, %f873, %f877; add.f32 %f887, %f886, %f886; neg.f32 %f888, %f877; fma.rn.f32 %f889, %f888, %f873, %f887; mul.rn.f32 %f890, %f875, %f889; add.f32 %f891, %f885, %f877; sub.f32 %f892, %f877, %f891; add.f32 %f893, %f885, %f892; add.f32 %f894, %f890, %f893; add.f32 %f895, %f891, %f894; sub.f32 %f896, %f891, %f895; add.f32 %f897, %f894, %f896; mov.f32 %f898, 0f3F317200; mul.rn.f32 %f899, %f871, %f898; mov.f32 %f900, 0f35BFBE8E; mul.rn.f32 %f901, %f871, %f900; add.f32 %f902, %f899, %f895; sub.f32 %f903, %f899, %f902; add.f32 %f904, %f895, %f903; add.f32 %f905, %f897, %f904; add.f32 %f906, %f901, %f905; add.f32 %f907, %f902, %f906; sub.f32 %f908, %f902, %f907; add.f32 %f909, %f906, %f908; mov.f32 %f910, 0f3EE66666; mul.rn.f32 %f911, %f910, %f907; neg.f32 %f912, %f911; fma.rn.f32 %f913, %f910, %f907, %f912; fma.rn.f32 %f914, %f910, %f909, %f913; mov.f32 %f915, 0f00000000; fma.rn.f32 %f916, %f915, %f907, %f914; add.rn.f32 %f917, %f911, %f916; neg.f32 %f918, %f917; add.rn.f32 %f919, %f911, %f918; add.rn.f32 %f920, %f919, %f916; mov.b32 %r343, %f917; setp.eq.s32 %p104, %r343, 1118925336; add.s32 %r344, %r343, -1; mov.b32 %f921, %r344; add.f32 %f922, %f920, 0f37000000; selp.f32 %f277, %f922, %f920, %p104; selp.f32 %f923, %f921, %f917, %p104; mov.f32 %f924, 0f3FB8AA3B; mul.rn.f32 %f925, %f923, %f924; cvt.rzi.f32.f32 %f926, %f925; abs.f32 %f927, %f926; setp.gt.f32 %p105, %f927, 0f42FC0000; mov.b32 %r345, %f926; and.b32 %r346, %r345, -2147483648; or.b32 %r347, %r346, 1123811328; mov.b32 %f928, %r347; selp.f32 %f929, %f928, %f926, %p105; mov.f32 %f930, 0fBF317218; fma.rn.f32 %f931, %f929, %f930, %f923; mov.f32 %f932, 0f3102E308; fma.rn.f32 %f933, %f929, %f932, %f931; mul.f32 %f934, %f933, 0f3FB8AA3B; add.f32 %f935, %f929, 0f4B40007F; mov.b32 %r348, %f935; shl.b32 %r349, %r348, 23; mov.b32 %f936, %r349; ex2.approx.ftz.f32 %f937, %f934; mul.f32 %f278, %f937, %f936; setp.eq.f32 %p106, %f278, 0f7F800000; mov.f32 %f1564, 0f7F800000; @%p106 bra $L__BB0_83; fma.rn.f32 %f1564, %f278, %f277, %f278; $L__BB0_83: setp.lt.f32 %p107, %f260, 0f00000000; and.pred %p5, %p107, %p93; setp.eq.f32 %p109, %f260, 0f00000000; @%p109 bra $L__BB0_87; bra.uni $L__BB0_84; $L__BB0_87: add.f32 %f942, %f260, %f260; selp.f32 %f1566, %f942, 0f00000000, %p93; bra.uni $L__BB0_88; $L__BB0_84: mov.b32 %r350, %f1564; xor.b32 %r351, %r350, -2147483648; mov.b32 %f938, %r351; selp.f32 %f1566, %f938, %f1564, %p5; setp.geu.f32 %p110, %f260, 0f00000000; @%p110 bra $L__BB0_88; mov.f32 %f939, 0f3EE66666; cvt.rzi.f32.f32 %f940, %f939; setp.eq.f32 %p111, %f940, 0f3EE66666; @%p111 bra $L__BB0_88; mov.f32 %f1566, 0f7FFFFFFF; $L__BB0_88: add.f32 %f943, %f276, 0f3EE66666; mov.b32 %r352, %f943; setp.lt.s32 %p113, %r352, 2139095040; @%p113 bra $L__BB0_93; setp.gtu.f32 %p114, %f276, 0f7F800000; @%p114 bra $L__BB0_92; bra.uni $L__BB0_90; $L__BB0_92: add.f32 %f1566, %f260, 0f3EE66666; bra.uni $L__BB0_93; $L__BB0_90: setp.neu.f32 %p115, %f276, 0f7F800000; @%p115 bra $L__BB0_93; selp.f32 %f1566, 0fFF800000, 0f7F800000, %p5; $L__BB0_93: setp.eq.f32 %p116, %f260, 0f3F800000; selp.f32 %f287, 0f3F800000, %f1566, %p116; abs.f32 %f288, %f262; setp.lt.f32 %p117, %f288, 0f00800000; mul.f32 %f945, %f288, 0f4B800000; selp.f32 %f946, %f945, %f288, %p117; selp.f32 %f947, 0fC3170000, 0fC2FE0000, %p117; mov.b32 %r353, %f946; and.b32 %r354, %r353, 8388607; or.b32 %r355, %r354, 1065353216; mov.b32 %f948, %r355; shr.u32 %r356, %r353, 23; cvt.rn.f32.u32 %f949, %r356; add.f32 %f950, %f947, %f949; setp.gt.f32 %p118, %f948, 0f3FB504F3; mul.f32 %f951, %f948, 0f3F000000; add.f32 %f952, %f950, 0f3F800000; selp.f32 %f953, %f952, %f950, %p118; selp.f32 %f954, %f951, %f948, %p118; add.f32 %f955, %f954, 0fBF800000; add.f32 %f956, %f954, 0f3F800000; rcp.approx.ftz.f32 %f957, %f956; add.f32 %f958, %f955, %f955; mul.f32 %f959, %f958, %f957; mul.f32 %f960, %f959, %f959; mov.f32 %f961, 0f3C4CAF63; mov.f32 %f962, 0f3B18F0FE; fma.rn.f32 %f963, %f962, %f960, %f961; mov.f32 %f964, 0f3DAAAABD; fma.rn.f32 %f965, %f963, %f960, %f964; mul.rn.f32 %f966, %f965, %f960; mul.rn.f32 %f967, %f966, %f959; sub.f32 %f968, %f955, %f959; add.f32 %f969, %f968, %f968; neg.f32 %f970, %f959; fma.rn.f32 %f971, %f970, %f955, %f969; mul.rn.f32 %f972, %f957, %f971; add.f32 %f973, %f967, %f959; sub.f32 %f974, %f959, %f973; add.f32 %f975, %f967, %f974; add.f32 %f976, %f972, %f975; add.f32 %f977, %f973, %f976; sub.f32 %f978, %f973, %f977; add.f32 %f979, %f976, %f978; mov.f32 %f980, 0f3F317200; mul.rn.f32 %f981, %f953, %f980; mov.f32 %f982, 0f35BFBE8E; mul.rn.f32 %f983, %f953, %f982; add.f32 %f984, %f981, %f977; sub.f32 %f985, %f981, %f984; add.f32 %f986, %f977, %f985; add.f32 %f987, %f979, %f986; add.f32 %f988, %f983, %f987; add.f32 %f989, %f984, %f988; sub.f32 %f990, %f984, %f989; add.f32 %f991, %f988, %f990; mov.f32 %f992, 0f3EE66666; mul.rn.f32 %f993, %f992, %f989; neg.f32 %f994, %f993; fma.rn.f32 %f995, %f992, %f989, %f994; fma.rn.f32 %f996, %f992, %f991, %f995; mov.f32 %f997, 0f00000000; fma.rn.f32 %f998, %f997, %f989, %f996; add.rn.f32 %f999, %f993, %f998; neg.f32 %f1000, %f999; add.rn.f32 %f1001, %f993, %f1000; add.rn.f32 %f1002, %f1001, %f998; mov.b32 %r357, %f999; setp.eq.s32 %p119, %r357, 1118925336; add.s32 %r358, %r357, -1; mov.b32 %f1003, %r358; add.f32 %f1004, %f1002, 0f37000000; selp.f32 %f289, %f1004, %f1002, %p119; selp.f32 %f1005, %f1003, %f999, %p119; mov.f32 %f1006, 0f3FB8AA3B; mul.rn.f32 %f1007, %f1005, %f1006; cvt.rzi.f32.f32 %f1008, %f1007; abs.f32 %f1009, %f1008; setp.gt.f32 %p120, %f1009, 0f42FC0000; mov.b32 %r359, %f1008; and.b32 %r360, %r359, -2147483648; or.b32 %r361, %r360, 1123811328; mov.b32 %f1010, %r361; selp.f32 %f1011, %f1010, %f1008, %p120; mov.f32 %f1012, 0fBF317218; fma.rn.f32 %f1013, %f1011, %f1012, %f1005; mov.f32 %f1014, 0f3102E308; fma.rn.f32 %f1015, %f1011, %f1014, %f1013; mul.f32 %f1016, %f1015, 0f3FB8AA3B; add.f32 %f1017, %f1011, 0f4B40007F; mov.b32 %r362, %f1017; shl.b32 %r363, %r362, 23; mov.b32 %f1018, %r363; ex2.approx.ftz.f32 %f1019, %f1016; mul.f32 %f290, %f1019, %f1018; setp.eq.f32 %p121, %f290, 0f7F800000; mov.f32 %f1567, 0f7F800000; @%p121 bra $L__BB0_95; fma.rn.f32 %f1567, %f290, %f289, %f290; $L__BB0_95: setp.lt.f32 %p122, %f262, 0f00000000; and.pred %p6, %p122, %p93; setp.eq.f32 %p124, %f262, 0f00000000; @%p124 bra $L__BB0_99; bra.uni $L__BB0_96; $L__BB0_99: add.f32 %f1024, %f262, %f262; selp.f32 %f1569, %f1024, 0f00000000, %p93; bra.uni $L__BB0_100; $L__BB0_96: mov.b32 %r364, %f1567; xor.b32 %r365, %r364, -2147483648; mov.b32 %f1020, %r365; selp.f32 %f1569, %f1020, %f1567, %p6; setp.geu.f32 %p125, %f262, 0f00000000; @%p125 bra $L__BB0_100; mov.f32 %f1021, 0f3EE66666; cvt.rzi.f32.f32 %f1022, %f1021; setp.eq.f32 %p126, %f1022, 0f3EE66666; @%p126 bra $L__BB0_100; mov.f32 %f1569, 0f7FFFFFFF; $L__BB0_100: add.f32 %f1025, %f288, 0f3EE66666; mov.b32 %r366, %f1025; setp.lt.s32 %p128, %r366, 2139095040; @%p128 bra $L__BB0_105; setp.gtu.f32 %p129, %f288, 0f7F800000; @%p129 bra $L__BB0_104; bra.uni $L__BB0_102; $L__BB0_104: add.f32 %f1569, %f262, 0f3EE66666; bra.uni $L__BB0_105; $L__BB0_102: setp.neu.f32 %p130, %f288, 0f7F800000; @%p130 bra $L__BB0_105; selp.f32 %f1569, 0fFF800000, 0f7F800000, %p6; $L__BB0_105: setp.eq.f32 %p131, %f262, 0f3F800000; mov.f32 %f1027, 0f3F800000; selp.f32 %f1028, 0f3F800000, %f1569, %p131; ld.const.u64 %rd79, [params+144]; cvta.to.global.u64 %rd21, %rd79; ld.const.u32 %r367, [params+136]; mad.lo.s32 %r368, %r367, %r8, %r7; cvt.u64.u32 %rd22, %r368; min.f32 %f1029, %f275, %f1027; mov.f32 %f1030, 0f00000000; max.f32 %f299, %f1030, %f1029; min.f32 %f1031, %f287, %f1027; max.f32 %f300, %f1030, %f1031; min.f32 %f1032, %f1028, %f1027; max.f32 %f301, %f1030, %f1032; mov.f32 %f1033, 0f3E555555; cvt.rzi.f32.f32 %f1034, %f1033; add.f32 %f1035, %f1034, %f1034; mov.f32 %f1036, 0f3ED55555; sub.f32 %f1037, %f1036, %f1035; abs.f32 %f302, %f1037; abs.f32 %f303, %f299; setp.lt.f32 %p132, %f303, 0f00800000; mul.f32 %f1038, %f303, 0f4B800000; selp.f32 %f1039, %f1038, %f303, %p132; selp.f32 %f1040, 0fC3170000, 0fC2FE0000, %p132; mov.b32 %r369, %f1039; and.b32 %r370, %r369, 8388607; or.b32 %r371, %r370, 1065353216; mov.b32 %f1041, %r371; shr.u32 %r372, %r369, 23; cvt.rn.f32.u32 %f1042, %r372; add.f32 %f1043, %f1040, %f1042; setp.gt.f32 %p133, %f1041, 0f3FB504F3; mul.f32 %f1044, %f1041, 0f3F000000; add.f32 %f1045, %f1043, 0f3F800000; selp.f32 %f1046, %f1045, %f1043, %p133; selp.f32 %f1047, %f1044, %f1041, %p133; add.f32 %f1048, %f1047, 0fBF800000; add.f32 %f1049, %f1047, 0f3F800000; rcp.approx.ftz.f32 %f1050, %f1049; add.f32 %f1051, %f1048, %f1048; mul.f32 %f1052, %f1051, %f1050; mul.f32 %f1053, %f1052, %f1052; mov.f32 %f1054, 0f3C4CAF63; mov.f32 %f1055, 0f3B18F0FE; fma.rn.f32 %f1056, %f1055, %f1053, %f1054; mov.f32 %f1057, 0f3DAAAABD; fma.rn.f32 %f1058, %f1056, %f1053, %f1057; mul.rn.f32 %f1059, %f1058, %f1053; mul.rn.f32 %f1060, %f1059, %f1052; sub.f32 %f1061, %f1048, %f1052; add.f32 %f1062, %f1061, %f1061; neg.f32 %f1063, %f1052; fma.rn.f32 %f1064, %f1063, %f1048, %f1062; mul.rn.f32 %f1065, %f1050, %f1064; add.f32 %f1066, %f1060, %f1052; sub.f32 %f1067, %f1052, %f1066; add.f32 %f1068, %f1060, %f1067; add.f32 %f1069, %f1065, %f1068; add.f32 %f1070, %f1066, %f1069; sub.f32 %f1071, %f1066, %f1070; add.f32 %f1072, %f1069, %f1071; mov.f32 %f1073, 0f3F317200; mul.rn.f32 %f1074, %f1046, %f1073; mov.f32 %f1075, 0f35BFBE8E; mul.rn.f32 %f1076, %f1046, %f1075; add.f32 %f1077, %f1074, %f1070; sub.f32 %f1078, %f1074, %f1077; add.f32 %f1079, %f1070, %f1078; add.f32 %f1080, %f1072, %f1079; add.f32 %f1081, %f1076, %f1080; add.f32 %f1082, %f1077, %f1081; sub.f32 %f1083, %f1077, %f1082; add.f32 %f1084, %f1081, %f1083; mul.rn.f32 %f1085, %f1036, %f1082; neg.f32 %f1086, %f1085; fma.rn.f32 %f1087, %f1036, %f1082, %f1086; fma.rn.f32 %f1088, %f1036, %f1084, %f1087; fma.rn.f32 %f1089, %f1030, %f1082, %f1088; add.rn.f32 %f1090, %f1085, %f1089; neg.f32 %f1091, %f1090; add.rn.f32 %f1092, %f1085, %f1091; add.rn.f32 %f1093, %f1092, %f1089; mov.b32 %r373, %f1090; setp.eq.s32 %p134, %r373, 1118925336; add.s32 %r374, %r373, -1; mov.b32 %f1094, %r374; add.f32 %f1095, %f1093, 0f37000000; selp.f32 %f304, %f1095, %f1093, %p134; selp.f32 %f1096, %f1094, %f1090, %p134; mov.f32 %f1097, 0f3FB8AA3B; mul.rn.f32 %f1098, %f1096, %f1097; cvt.rzi.f32.f32 %f1099, %f1098; abs.f32 %f1100, %f1099; setp.gt.f32 %p135, %f1100, 0f42FC0000; mov.b32 %r375, %f1099; and.b32 %r376, %r375, -2147483648; or.b32 %r377, %r376, 1123811328; mov.b32 %f1101, %r377; selp.f32 %f1102, %f1101, %f1099, %p135; mov.f32 %f1103, 0fBF317218; fma.rn.f32 %f1104, %f1102, %f1103, %f1096; mov.f32 %f1105, 0f3102E308; fma.rn.f32 %f1106, %f1102, %f1105, %f1104; mul.f32 %f1107, %f1106, 0f3FB8AA3B; add.f32 %f1108, %f1102, 0f4B40007F; mov.b32 %r378, %f1108; shl.b32 %r379, %r378, 23; mov.b32 %f1109, %r379; ex2.approx.ftz.f32 %f1110, %f1107; mul.f32 %f305, %f1110, %f1109; setp.eq.f32 %p136, %f305, 0f7F800000; mov.f32 %f1570, 0f7F800000; @%p136 bra $L__BB0_107; fma.rn.f32 %f1570, %f305, %f304, %f305; $L__BB0_107: setp.lt.f32 %p137, %f299, 0f00000000; setp.eq.f32 %p138, %f302, 0f3F800000; and.pred %p7, %p137, %p138; setp.eq.f32 %p139, %f299, 0f00000000; @%p139 bra $L__BB0_111; bra.uni $L__BB0_108; $L__BB0_111: add.f32 %f1115, %f299, %f299; selp.f32 %f1572, %f1115, 0f00000000, %p138; bra.uni $L__BB0_112; $L__BB0_108: mov.b32 %r380, %f1570; xor.b32 %r381, %r380, -2147483648; mov.b32 %f1111, %r381; selp.f32 %f1572, %f1111, %f1570, %p7; setp.geu.f32 %p140, %f299, 0f00000000; @%p140 bra $L__BB0_112; mov.f32 %f1112, 0f3ED55555; cvt.rzi.f32.f32 %f1113, %f1112; setp.eq.f32 %p141, %f1113, 0f3ED55555; @%p141 bra $L__BB0_112; mov.f32 %f1572, 0f7FFFFFFF; $L__BB0_112: add.f32 %f1116, %f303, 0f3ED55555; mov.b32 %r382, %f1116; setp.lt.s32 %p143, %r382, 2139095040; @%p143 bra $L__BB0_117; setp.gtu.f32 %p144, %f303, 0f7F800000; @%p144 bra $L__BB0_116; bra.uni $L__BB0_114; $L__BB0_116: add.f32 %f1572, %f299, 0f3ED55555; bra.uni $L__BB0_117; $L__BB0_114: setp.neu.f32 %p145, %f303, 0f7F800000; @%p145 bra $L__BB0_117; selp.f32 %f1572, 0fFF800000, 0f7F800000, %p7; $L__BB0_117: abs.f32 %f314, %f300; setp.lt.f32 %p146, %f314, 0f00800000; mul.f32 %f1118, %f314, 0f4B800000; selp.f32 %f1119, %f1118, %f314, %p146; selp.f32 %f1120, 0fC3170000, 0fC2FE0000, %p146; mov.b32 %r383, %f1119; and.b32 %r384, %r383, 8388607; or.b32 %r385, %r384, 1065353216; mov.b32 %f1121, %r385; shr.u32 %r386, %r383, 23; cvt.rn.f32.u32 %f1122, %r386; add.f32 %f1123, %f1120, %f1122; setp.gt.f32 %p147, %f1121, 0f3FB504F3; mul.f32 %f1124, %f1121, 0f3F000000; add.f32 %f1125, %f1123, 0f3F800000; selp.f32 %f1126, %f1125, %f1123, %p147; selp.f32 %f1127, %f1124, %f1121, %p147; add.f32 %f1128, %f1127, 0fBF800000; add.f32 %f1129, %f1127, 0f3F800000; rcp.approx.ftz.f32 %f1130, %f1129; add.f32 %f1131, %f1128, %f1128; mul.f32 %f1132, %f1131, %f1130; mul.f32 %f1133, %f1132, %f1132; mov.f32 %f1134, 0f3C4CAF63; mov.f32 %f1135, 0f3B18F0FE; fma.rn.f32 %f1136, %f1135, %f1133, %f1134; mov.f32 %f1137, 0f3DAAAABD; fma.rn.f32 %f1138, %f1136, %f1133, %f1137; mul.rn.f32 %f1139, %f1138, %f1133; mul.rn.f32 %f1140, %f1139, %f1132; sub.f32 %f1141, %f1128, %f1132; add.f32 %f1142, %f1141, %f1141; neg.f32 %f1143, %f1132; fma.rn.f32 %f1144, %f1143, %f1128, %f1142; mul.rn.f32 %f1145, %f1130, %f1144; add.f32 %f1146, %f1140, %f1132; sub.f32 %f1147, %f1132, %f1146; add.f32 %f1148, %f1140, %f1147; add.f32 %f1149, %f1145, %f1148; add.f32 %f1150, %f1146, %f1149; sub.f32 %f1151, %f1146, %f1150; add.f32 %f1152, %f1149, %f1151; mov.f32 %f1153, 0f3F317200; mul.rn.f32 %f1154, %f1126, %f1153; mov.f32 %f1155, 0f35BFBE8E; mul.rn.f32 %f1156, %f1126, %f1155; add.f32 %f1157, %f1154, %f1150; sub.f32 %f1158, %f1154, %f1157; add.f32 %f1159, %f1150, %f1158; add.f32 %f1160, %f1152, %f1159; add.f32 %f1161, %f1156, %f1160; add.f32 %f1162, %f1157, %f1161; sub.f32 %f1163, %f1157, %f1162; add.f32 %f1164, %f1161, %f1163; mov.f32 %f1165, 0f3ED55555; mul.rn.f32 %f1166, %f1165, %f1162; neg.f32 %f1167, %f1166; fma.rn.f32 %f1168, %f1165, %f1162, %f1167; fma.rn.f32 %f1169, %f1165, %f1164, %f1168; mov.f32 %f1170, 0f00000000; fma.rn.f32 %f1171, %f1170, %f1162, %f1169; add.rn.f32 %f1172, %f1166, %f1171; neg.f32 %f1173, %f1172; add.rn.f32 %f1174, %f1166, %f1173; add.rn.f32 %f1175, %f1174, %f1171; mov.b32 %r387, %f1172; setp.eq.s32 %p148, %r387, 1118925336; add.s32 %r388, %r387, -1; mov.b32 %f1176, %r388; add.f32 %f1177, %f1175, 0f37000000; selp.f32 %f315, %f1177, %f1175, %p148; selp.f32 %f1178, %f1176, %f1172, %p148; mov.f32 %f1179, 0f3FB8AA3B; mul.rn.f32 %f1180, %f1178, %f1179; cvt.rzi.f32.f32 %f1181, %f1180; abs.f32 %f1182, %f1181; setp.gt.f32 %p149, %f1182, 0f42FC0000; mov.b32 %r389, %f1181; and.b32 %r390, %r389, -2147483648; or.b32 %r391, %r390, 1123811328; mov.b32 %f1183, %r391; selp.f32 %f1184, %f1183, %f1181, %p149; mov.f32 %f1185, 0fBF317218; fma.rn.f32 %f1186, %f1184, %f1185, %f1178; mov.f32 %f1187, 0f3102E308; fma.rn.f32 %f1188, %f1184, %f1187, %f1186; mul.f32 %f1189, %f1188, 0f3FB8AA3B; add.f32 %f1190, %f1184, 0f4B40007F; mov.b32 %r392, %f1190; shl.b32 %r393, %r392, 23; mov.b32 %f1191, %r393; ex2.approx.ftz.f32 %f1192, %f1189; mul.f32 %f316, %f1192, %f1191; setp.eq.f32 %p150, %f316, 0f7F800000; mov.f32 %f1573, 0f7F800000; @%p150 bra $L__BB0_119; fma.rn.f32 %f1573, %f316, %f315, %f316; $L__BB0_119: setp.lt.f32 %p151, %f300, 0f00000000; and.pred %p8, %p151, %p138; setp.eq.f32 %p153, %f300, 0f00000000; @%p153 bra $L__BB0_123; bra.uni $L__BB0_120; $L__BB0_123: add.f32 %f1197, %f300, %f300; selp.f32 %f1575, %f1197, 0f00000000, %p138; bra.uni $L__BB0_124; $L__BB0_120: mov.b32 %r394, %f1573; xor.b32 %r395, %r394, -2147483648; mov.b32 %f1193, %r395; selp.f32 %f1575, %f1193, %f1573, %p8; setp.geu.f32 %p154, %f300, 0f00000000; @%p154 bra $L__BB0_124; mov.f32 %f1194, 0f3ED55555; cvt.rzi.f32.f32 %f1195, %f1194; setp.eq.f32 %p155, %f1195, 0f3ED55555; @%p155 bra $L__BB0_124; mov.f32 %f1575, 0f7FFFFFFF; $L__BB0_124: add.f32 %f1198, %f314, 0f3ED55555; mov.b32 %r396, %f1198; setp.lt.s32 %p157, %r396, 2139095040; @%p157 bra $L__BB0_129; setp.gtu.f32 %p158, %f314, 0f7F800000; @%p158 bra $L__BB0_128; bra.uni $L__BB0_126; $L__BB0_128: add.f32 %f1575, %f300, 0f3ED55555; bra.uni $L__BB0_129; $L__BB0_126: setp.neu.f32 %p159, %f314, 0f7F800000; @%p159 bra $L__BB0_129; selp.f32 %f1575, 0fFF800000, 0f7F800000, %p8; $L__BB0_129: abs.f32 %f325, %f301; setp.lt.f32 %p160, %f325, 0f00800000; mul.f32 %f1200, %f325, 0f4B800000; selp.f32 %f1201, %f1200, %f325, %p160; selp.f32 %f1202, 0fC3170000, 0fC2FE0000, %p160; mov.b32 %r397, %f1201; and.b32 %r398, %r397, 8388607; or.b32 %r399, %r398, 1065353216; mov.b32 %f1203, %r399; shr.u32 %r400, %r397, 23; cvt.rn.f32.u32 %f1204, %r400; add.f32 %f1205, %f1202, %f1204; setp.gt.f32 %p161, %f1203, 0f3FB504F3; mul.f32 %f1206, %f1203, 0f3F000000; add.f32 %f1207, %f1205, 0f3F800000; selp.f32 %f1208, %f1207, %f1205, %p161; selp.f32 %f1209, %f1206, %f1203, %p161; add.f32 %f1210, %f1209, 0fBF800000; add.f32 %f1211, %f1209, 0f3F800000; rcp.approx.ftz.f32 %f1212, %f1211; add.f32 %f1213, %f1210, %f1210; mul.f32 %f1214, %f1213, %f1212; mul.f32 %f1215, %f1214, %f1214; mov.f32 %f1216, 0f3C4CAF63; mov.f32 %f1217, 0f3B18F0FE; fma.rn.f32 %f1218, %f1217, %f1215, %f1216; mov.f32 %f1219, 0f3DAAAABD; fma.rn.f32 %f1220, %f1218, %f1215, %f1219; mul.rn.f32 %f1221, %f1220, %f1215; mul.rn.f32 %f1222, %f1221, %f1214; sub.f32 %f1223, %f1210, %f1214; add.f32 %f1224, %f1223, %f1223; neg.f32 %f1225, %f1214; fma.rn.f32 %f1226, %f1225, %f1210, %f1224; mul.rn.f32 %f1227, %f1212, %f1226; add.f32 %f1228, %f1222, %f1214; sub.f32 %f1229, %f1214, %f1228; add.f32 %f1230, %f1222, %f1229; add.f32 %f1231, %f1227, %f1230; add.f32 %f1232, %f1228, %f1231; sub.f32 %f1233, %f1228, %f1232; add.f32 %f1234, %f1231, %f1233; mov.f32 %f1235, 0f3F317200; mul.rn.f32 %f1236, %f1208, %f1235; mov.f32 %f1237, 0f35BFBE8E; mul.rn.f32 %f1238, %f1208, %f1237; add.f32 %f1239, %f1236, %f1232; sub.f32 %f1240, %f1236, %f1239; add.f32 %f1241, %f1232, %f1240; add.f32 %f1242, %f1234, %f1241; add.f32 %f1243, %f1238, %f1242; add.f32 %f1244, %f1239, %f1243; sub.f32 %f1245, %f1239, %f1244; add.f32 %f1246, %f1243, %f1245; mov.f32 %f1247, 0f3ED55555; mul.rn.f32 %f1248, %f1247, %f1244; neg.f32 %f1249, %f1248; fma.rn.f32 %f1250, %f1247, %f1244, %f1249; fma.rn.f32 %f1251, %f1247, %f1246, %f1250; mov.f32 %f1252, 0f00000000; fma.rn.f32 %f1253, %f1252, %f1244, %f1251; add.rn.f32 %f1254, %f1248, %f1253; neg.f32 %f1255, %f1254; add.rn.f32 %f1256, %f1248, %f1255; add.rn.f32 %f1257, %f1256, %f1253; mov.b32 %r401, %f1254; setp.eq.s32 %p162, %r401, 1118925336; add.s32 %r402, %r401, -1; mov.b32 %f1258, %r402; add.f32 %f1259, %f1257, 0f37000000; selp.f32 %f326, %f1259, %f1257, %p162; selp.f32 %f1260, %f1258, %f1254, %p162; mov.f32 %f1261, 0f3FB8AA3B; mul.rn.f32 %f1262, %f1260, %f1261; cvt.rzi.f32.f32 %f1263, %f1262; abs.f32 %f1264, %f1263; setp.gt.f32 %p163, %f1264, 0f42FC0000; mov.b32 %r403, %f1263; and.b32 %r404, %r403, -2147483648; or.b32 %r405, %r404, 1123811328; mov.b32 %f1265, %r405; selp.f32 %f1266, %f1265, %f1263, %p163; mov.f32 %f1267, 0fBF317218; fma.rn.f32 %f1268, %f1266, %f1267, %f1260; mov.f32 %f1269, 0f3102E308; fma.rn.f32 %f1270, %f1266, %f1269, %f1268; mul.f32 %f1271, %f1270, 0f3FB8AA3B; add.f32 %f1272, %f1266, 0f4B40007F; mov.b32 %r406, %f1272; shl.b32 %r407, %r406, 23; mov.b32 %f1273, %r407; ex2.approx.ftz.f32 %f1274, %f1271; mul.f32 %f327, %f1274, %f1273; setp.eq.f32 %p164, %f327, 0f7F800000; mov.f32 %f1576, 0f7F800000; @%p164 bra $L__BB0_131; fma.rn.f32 %f1576, %f327, %f326, %f327; $L__BB0_131: setp.lt.f32 %p165, %f301, 0f00000000; and.pred %p9, %p165, %p138; setp.eq.f32 %p167, %f301, 0f00000000; @%p167 bra $L__BB0_135; bra.uni $L__BB0_132; $L__BB0_135: add.f32 %f1279, %f301, %f301; selp.f32 %f1578, %f1279, 0f00000000, %p138; bra.uni $L__BB0_136; $L__BB0_132: mov.b32 %r408, %f1576; xor.b32 %r409, %r408, -2147483648; mov.b32 %f1275, %r409; selp.f32 %f1578, %f1275, %f1576, %p9; setp.geu.f32 %p168, %f301, 0f00000000; @%p168 bra $L__BB0_136; mov.f32 %f1276, 0f3ED55555; cvt.rzi.f32.f32 %f1277, %f1276; setp.eq.f32 %p169, %f1277, 0f3ED55555; @%p169 bra $L__BB0_136; mov.f32 %f1578, 0f7FFFFFFF; $L__BB0_136: add.f32 %f1280, %f325, 0f3ED55555; mov.b32 %r410, %f1280; setp.lt.s32 %p171, %r410, 2139095040; @%p171 bra $L__BB0_141; setp.gtu.f32 %p172, %f325, 0f7F800000; @%p172 bra $L__BB0_140; bra.uni $L__BB0_138; $L__BB0_140: add.f32 %f1578, %f301, 0f3ED55555; bra.uni $L__BB0_141; $L__BB0_138: setp.neu.f32 %p173, %f325, 0f7F800000; @%p173 bra $L__BB0_141; selp.f32 %f1578, 0fFF800000, 0f7F800000, %p9; $L__BB0_141: fma.rn.f32 %f1281, %f1572, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p174, %f299, 0f3F800000; mov.f32 %f1282, 0f3F800000; selp.f32 %f1283, 0f3F7FFFFF, %f1281, %p174; mul.f32 %f1284, %f299, 0f414EB852; setp.lt.f32 %p175, %f299, 0f3B4D2E1C; selp.f32 %f1285, %f1284, %f1283, %p175; fma.rn.f32 %f1286, %f1575, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p176, %f300, 0f3F800000; selp.f32 %f1287, 0f3F7FFFFF, %f1286, %p176; mul.f32 %f1288, %f300, 0f414EB852; setp.lt.f32 %p177, %f300, 0f3B4D2E1C; selp.f32 %f1289, %f1288, %f1287, %p177; fma.rn.f32 %f1290, %f1578, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p178, %f301, 0f3F800000; selp.f32 %f1291, 0f3F7FFFFF, %f1290, %p178; mul.f32 %f1292, %f301, 0f414EB852; setp.lt.f32 %p179, %f301, 0f3B4D2E1C; selp.f32 %f1293, %f1292, %f1291, %p179; min.f32 %f1294, %f1285, %f1282; mov.f32 %f1295, 0f00000000; max.f32 %f1296, %f1295, %f1294; mul.f32 %f1297, %f1296, 0f43800000; cvt.rzi.u32.f32 %r411, %f1297; min.u32 %r412, %r411, 255; min.f32 %f1298, %f1289, %f1282; max.f32 %f1299, %f1295, %f1298; mul.f32 %f1300, %f1299, 0f43800000; cvt.rzi.u32.f32 %r413, %f1300; min.u32 %r414, %r413, 255; min.f32 %f1301, %f1293, %f1282; max.f32 %f1302, %f1295, %f1301; mul.f32 %f1303, %f1302, 0f43800000; cvt.rzi.u32.f32 %r415, %f1303; min.u32 %r416, %r415, 255; shl.b64 %rd80, %rd22, 2; add.s64 %rd81, %rd21, %rd80; cvt.u16.u32 %rs18, %r416; cvt.u16.u32 %rs19, %r414; cvt.u16.u32 %rs20, %r412; mov.u16 %rs21, 255; st.global.v4.u8 [%rd81], {%rs20, %rs19, %rs18, %rs21}; $L__BB0_142: and.b32 %r417, %r54, 4; setp.eq.s32 %p180, %r417, 0; ld.const.u32 %r462, [params+108]; @%p180 bra $L__BB0_146; setp.eq.s32 %p181, %r462, 0; ld.const.u64 %rd82, [params+224]; cvta.to.global.u64 %rd83, %rd82; ld.const.u32 %r418, [params+216]; mad.lo.s32 %r419, %r418, %r8, %r7; mul.wide.u32 %rd84, %r419, 8; add.s64 %rd23, %rd83, %rd84; @%p181 bra $L__BB0_145; ld.global.v4.u16 {%rs29, %rs30, %rs31, %rs32}, [%rd23]; // begin inline asm { cvt.f32.f16 %f1304, %rs29;} // end inline asm // begin inline asm { cvt.f32.f16 %f1305, %rs30;} // end inline asm // begin inline asm { cvt.f32.f16 %f1306, %rs31;} // end inline asm add.f32 %f1307, %f259, %f1304; add.f32 %f1308, %f260, %f1305; add.f32 %f1309, %f262, %f1306; mov.f32 %f1310, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs27, %f1309;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs26, %f1308;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs25, %f1307;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs28, %f1310;} // end inline asm st.global.v4.u16 [%rd23], {%rs25, %rs26, %rs27, %rs28}; bra.uni $L__BB0_146; $L__BB0_145: mov.f32 %f1314, 0f3F800000; mov.u32 %r462, 0; // begin inline asm { cvt.rn.f16.f32 %rs36, %f1314;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs35, %f262;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs34, %f260;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs33, %f259;} // end inline asm st.global.v4.u16 [%rd23], {%rs33, %rs34, %rs35, %rs36}; $L__BB0_146: selp.f32 %f1315, 0f3F800000, 0f3E800000, %p197; mul.f32 %f1316, %f1315, %f1551; mul.f32 %f1317, %f1315, %f1552; mul.f32 %f1318, %f1315, %f1553; mul.f32 %f1319, %f1315, %f1548; mul.f32 %f1320, %f1315, %f1549; mul.f32 %f1321, %f1315, %f1550; mul.f32 %f1322, %f1315, %f1545; mul.f32 %f1323, %f1315, %f1546; mul.f32 %f1324, %f1315, %f1547; mul.f32 %f1325, %f1315, %f1542; mul.f32 %f1326, %f1315, %f1543; mul.f32 %f1327, %f1315, %f1544; mul.f32 %f336, %f1316, %f774; mul.f32 %f337, %f1317, %f775; mul.f32 %f338, %f1318, %f261; mul.f32 %f339, %f1319, %f774; mul.f32 %f340, %f1320, %f775; mul.f32 %f341, %f1321, %f261; mul.f32 %f342, %f1322, %f774; mul.f32 %f343, %f1323, %f775; mul.f32 %f344, %f1324, %f261; mul.f32 %f345, %f1325, %f774; mul.f32 %f346, %f1326, %f775; mul.f32 %f347, %f1327, %f261; ld.const.u64 %rd85, [params+256]; cvta.to.global.u64 %rd86, %rd85; ld.const.u32 %r421, [params+248]; mad.lo.s32 %r422, %r421, %r8, %r7; mul.wide.u32 %rd87, %r422, 8; add.s64 %rd24, %rd86, %rd87; setp.eq.s32 %p182, %r462, 0; @%p182 bra $L__BB0_148; ld.global.v4.u16 {%rs44, %rs45, %rs46, %rs47}, [%rd24]; // begin inline asm { cvt.f32.f16 %f1328, %rs44;} // end inline asm // begin inline asm { cvt.f32.f16 %f1329, %rs45;} // end inline asm // begin inline asm { cvt.f32.f16 %f1330, %rs46;} // end inline asm add.f32 %f1331, %f336, %f1328; add.f32 %f1332, %f337, %f1329; add.f32 %f1333, %f338, %f1330; mov.f32 %f1334, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs42, %f1333;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs41, %f1332;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs40, %f1331;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs43, %f1334;} // end inline asm st.global.v4.u16 [%rd24], {%rs40, %rs41, %rs42, %rs43}; bra.uni $L__BB0_149; $L__BB0_148: mov.f32 %f1338, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs51, %f1338;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs50, %f338;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs49, %f337;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs48, %f336;} // end inline asm st.global.v4.u16 [%rd24], {%rs48, %rs49, %rs50, %rs51}; $L__BB0_149: mov.f32 %f1339, 0f34000000; max.f32 %f1340, %f336, %f1339; div.rn.f32 %f1341, %f339, %f1340; max.f32 %f1342, %f337, %f1339; div.rn.f32 %f1343, %f340, %f1342; max.f32 %f1344, %f338, %f1339; div.rn.f32 %f1345, %f341, %f1344; fma.rn.f32 %f348, %f1341, 0f3F000000, 0f3F000000; fma.rn.f32 %f349, %f1343, 0f3F000000, 0f3F000000; fma.rn.f32 %f350, %f1345, 0f3F000000, 0f3F000000; div.rn.f32 %f1346, %f342, %f1340; div.rn.f32 %f1347, %f343, %f1342; div.rn.f32 %f1348, %f344, %f1344; fma.rn.f32 %f351, %f1346, 0f3F000000, 0f3F000000; fma.rn.f32 %f352, %f1347, 0f3F000000, 0f3F000000; fma.rn.f32 %f353, %f1348, 0f3F000000, 0f3F000000; div.rn.f32 %f1349, %f345, %f1340; div.rn.f32 %f1350, %f346, %f1342; div.rn.f32 %f1351, %f347, %f1344; fma.rn.f32 %f354, %f1349, 0f3F000000, 0f3F000000; fma.rn.f32 %f355, %f1350, 0f3F000000, 0f3F000000; fma.rn.f32 %f356, %f1351, 0f3F000000, 0f3F000000; ld.const.u64 %rd88, [params+272]; cvta.to.global.u64 %rd89, %rd88; ld.const.u32 %r423, [params+264]; mad.lo.s32 %r424, %r423, %r8, %r7; mul.wide.u32 %rd90, %r424, 8; add.s64 %rd25, %rd89, %rd90; @%p182 bra $L__BB0_151; ld.global.v4.u16 {%rs59, %rs60, %rs61, %rs62}, [%rd25]; // begin inline asm { cvt.f32.f16 %f1352, %rs59;} // end inline asm // begin inline asm { cvt.f32.f16 %f1353, %rs60;} // end inline asm // begin inline asm { cvt.f32.f16 %f1354, %rs61;} // end inline asm add.f32 %f1355, %f348, %f1352; add.f32 %f1356, %f349, %f1353; add.f32 %f1357, %f350, %f1354; mov.f32 %f1358, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs57, %f1357;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs56, %f1356;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs55, %f1355;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs58, %f1358;} // end inline asm st.global.v4.u16 [%rd25], {%rs55, %rs56, %rs57, %rs58}; bra.uni $L__BB0_152; $L__BB0_151: mov.f32 %f1362, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs66, %f1362;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs65, %f350;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs64, %f349;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs63, %f348;} // end inline asm st.global.v4.u16 [%rd25], {%rs63, %rs64, %rs65, %rs66}; $L__BB0_152: ld.const.u64 %rd91, [params+288]; cvta.to.global.u64 %rd92, %rd91; ld.const.u32 %r425, [params+280]; mad.lo.s32 %r426, %r425, %r8, %r7; mul.wide.u32 %rd93, %r426, 8; add.s64 %rd26, %rd92, %rd93; @%p182 bra $L__BB0_154; ld.global.v4.u16 {%rs74, %rs75, %rs76, %rs77}, [%rd26]; // begin inline asm { cvt.f32.f16 %f1363, %rs74;} // end inline asm // begin inline asm { cvt.f32.f16 %f1364, %rs75;} // end inline asm // begin inline asm { cvt.f32.f16 %f1365, %rs76;} // end inline asm add.f32 %f1366, %f351, %f1363; add.f32 %f1367, %f352, %f1364; add.f32 %f1368, %f353, %f1365; mov.f32 %f1369, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs72, %f1368;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs71, %f1367;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs70, %f1366;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs73, %f1369;} // end inline asm st.global.v4.u16 [%rd26], {%rs70, %rs71, %rs72, %rs73}; bra.uni $L__BB0_155; $L__BB0_154: mov.f32 %f1373, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs81, %f1373;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs80, %f353;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs79, %f352;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs78, %f351;} // end inline asm st.global.v4.u16 [%rd26], {%rs78, %rs79, %rs80, %rs81}; $L__BB0_155: ld.const.u64 %rd94, [params+304]; cvta.to.global.u64 %rd95, %rd94; ld.const.u32 %r427, [params+296]; mad.lo.s32 %r428, %r427, %r8, %r7; mul.wide.u32 %rd96, %r428, 8; add.s64 %rd27, %rd95, %rd96; @%p182 bra $L__BB0_157; ld.global.v4.u16 {%rs89, %rs90, %rs91, %rs92}, [%rd27]; // begin inline asm { cvt.f32.f16 %f1374, %rs89;} // end inline asm // begin inline asm { cvt.f32.f16 %f1375, %rs90;} // end inline asm // begin inline asm { cvt.f32.f16 %f1376, %rs91;} // end inline asm add.f32 %f1377, %f354, %f1374; add.f32 %f1378, %f355, %f1375; add.f32 %f1379, %f356, %f1376; mov.f32 %f1380, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs87, %f1379;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs86, %f1378;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs85, %f1377;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs88, %f1380;} // end inline asm st.global.v4.u16 [%rd27], {%rs85, %rs86, %rs87, %rs88}; bra.uni $L__BB0_178; $L__BB0_157: mov.f32 %f1384, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs96, %f1384;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs95, %f356;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs94, %f355;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs93, %f354;} // end inline asm st.global.v4.u16 [%rd27], {%rs93, %rs94, %rs95, %rs96}; $L__BB0_178: ret; }