// // 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<148>; .reg .b16 %rs<57>; .reg .f32 %f<920>; .reg .b32 %r<333>; .reg .f64 %fd<5>; .reg .b64 %rd<85>; mov.u64 %SPL, __local_depot0; add.u64 %rd1, %SPL, 0; // begin inline asm call (%r51), _optix_get_launch_index_x, (); // end inline asm // begin inline asm call (%r52), _optix_get_launch_index_y, (); // end inline asm ld.const.u32 %r3, [params+540]; ld.const.u64 %rd21, [params+400]; cvta.to.global.u64 %rd22, %rd21; ld.const.u32 %r57, [params+392]; mad.lo.s32 %r58, %r57, %r52, %r51; mul.wide.u32 %rd23, %r58, 4; add.s64 %rd2, %rd22, %rd23; ld.global.v2.u8 {%rs7, %rs56}, [%rd2]; or.b16 %rs9, %rs7, %rs56; and.b16 %rs10, %rs9, 255; setp.eq.s16 %p7, %rs10, 0; @%p7 bra $L__BB0_2; ld.global.u8 %rs55, [%rd2+2]; bra.uni $L__BB0_3; $L__BB0_2: ld.global.u8 %rs55, [%rd2+2]; setp.eq.s16 %p8, %rs55, 0; mov.f32 %f877, 0f00000000; mov.u16 %rs56, 0; mov.f32 %f878, %f877; mov.f32 %f879, %f877; @%p8 bra $L__BB0_4; $L__BB0_3: cvt.rn.f32.u16 %f152, %rs7; div.rn.f32 %f153, %f152, 0f437F0000; fma.rn.f32 %f154, %f153, 0f40000000, 0fBF800000; and.b16 %rs13, %rs56, 255; cvt.rn.f32.u16 %f155, %rs13; div.rn.f32 %f156, %f155, 0f437F0000; fma.rn.f32 %f157, %f156, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f158, %rs55; div.rn.f32 %f159, %f158, 0f437F0000; fma.rn.f32 %f160, %f159, 0f40000000, 0fBF800000; mul.f32 %f161, %f157, %f157; fma.rn.f32 %f162, %f154, %f154, %f161; fma.rn.f32 %f163, %f160, %f160, %f162; sqrt.rn.f32 %f164, %f163; rcp.rn.f32 %f165, %f164; mul.f32 %f879, %f165, %f160; mul.f32 %f878, %f165, %f157; mul.f32 %f877, %f154, %f165; $L__BB0_4: ld.const.v2.u32 {%r59, %r60}, [params]; add.s32 %r4, %r59, %r51; add.s32 %r5, %r60, %r52; setp.eq.f32 %p9, %f877, 0f00000000; setp.eq.f32 %p10, %f878, 0f00000000; and.pred %p11, %p9, %p10; setp.eq.f32 %p12, %f879, 0f00000000; and.pred %p13, %p12, %p11; @%p13 bra $L__BB0_116; bra.uni $L__BB0_5; $L__BB0_116: ld.const.u32 %r50, [params+104]; and.b32 %r308, %r50, 1; setp.eq.b32 %p141, %r308, 1; mov.pred %p142, 0; xor.pred %p143, %p141, %p142; not.pred %p144, %p143; @%p144 bra $L__BB0_118; ld.const.u64 %rd66, [params+144]; cvta.to.global.u64 %rd67, %rd66; ld.const.u32 %r309, [params+136]; mad.lo.s32 %r310, %r309, %r5, %r4; mul.wide.u32 %rd68, %r310, 4; add.s64 %rd69, %rd67, %rd68; mov.u16 %rs38, 0; st.global.v4.u8 [%rd69], {%rs38, %rs38, %rs38, %rs38}; $L__BB0_118: and.b32 %r311, %r50, 4; setp.eq.s32 %p145, %r311, 0; @%p145 bra $L__BB0_122; ld.const.u32 %r312, [params+108]; setp.eq.s32 %p146, %r312, 0; ld.const.u64 %rd70, [params+224]; cvta.to.global.u64 %rd71, %rd70; ld.const.u32 %r313, [params+216]; mad.lo.s32 %r314, %r313, %r5, %r4; mul.wide.u32 %rd72, %r314, 8; add.s64 %rd19, %rd71, %rd72; @%p146 bra $L__BB0_121; ld.global.v4.u16 {%rs45, %rs46, %rs47, %rs48}, [%rd19]; // begin inline asm { cvt.f32.f16 %f846, %rs45;} // end inline asm // begin inline asm { cvt.f32.f16 %f847, %rs46;} // end inline asm // begin inline asm { cvt.f32.f16 %f848, %rs47;} // end inline asm add.f32 %f849, %f846, 0f00000000; add.f32 %f850, %f847, 0f00000000; add.f32 %f851, %f848, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs44, %f851;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs43, %f850;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs42, %f849;} // end inline asm mov.u16 %rs49, 0; st.global.v4.u16 [%rd19], {%rs42, %rs43, %rs44, %rs49}; bra.uni $L__BB0_122; $L__BB0_5: ld.const.u64 %rd24, [params+432]; cvta.to.global.u64 %rd25, %rd24; ld.const.u32 %r63, [params+424]; mad.lo.s32 %r64, %r63, %r52, %r51; mul.wide.u32 %rd26, %r64, 12; add.s64 %rd27, %rd25, %rd26; ld.global.f32 %f170, [%rd27]; mul.f32 %f171, %f170, 0f3456BF95; ld.global.f32 %f172, [%rd27+4]; mul.f32 %f173, %f172, 0f3456BF95; ld.global.f32 %f174, [%rd27+8]; mul.f32 %f175, %f174, 0f3456BF95; abs.f32 %f176, %f877; div.rn.f32 %f177, %f171, %f176; abs.f32 %f178, %f878; div.rn.f32 %f179, %f173, %f178; abs.f32 %f180, %f879; div.rn.f32 %f181, %f175, %f180; abs.f32 %f182, %f177; abs.f32 %f183, %f179; abs.f32 %f184, %f181; mov.f32 %f185, 0f38D1B717; max.f32 %f186, %f182, %f185; max.f32 %f187, %f183, %f185; max.f32 %f188, %f184, %f185; fma.rn.f32 %f10, %f877, %f186, %f170; fma.rn.f32 %f11, %f878, %f187, %f172; fma.rn.f32 %f12, %f879, %f188, %f174; setp.gt.f32 %p14, %f176, %f180; neg.f32 %f189, %f878; selp.f32 %f190, %f189, 0f00000000, %p14; mov.f32 %f894, 0f00000000; neg.f32 %f191, %f879; selp.f32 %f192, %f877, %f191, %p14; selp.f32 %f193, 0f00000000, %f878, %p14; mul.f32 %f194, %f192, %f192; fma.rn.f32 %f195, %f190, %f190, %f194; fma.rn.f32 %f196, %f193, %f193, %f195; sqrt.rn.f32 %f197, %f196; rcp.rn.f32 %f198, %f197; mul.f32 %f13, %f190, %f198; mul.f32 %f14, %f192, %f198; mul.f32 %f15, %f193, %f198; ld.const.u64 %rd28, [params+128]; cvta.to.global.u64 %rd29, %rd28; ld.const.u32 %r65, [params+120]; mad.lo.s32 %r66, %r65, %r52, %r51; mul.wide.u32 %rd30, %r66, 4; add.s64 %rd3, %rd29, %rd30; setp.lt.s32 %p15, %r3, 1; mov.f32 %f895, %f894; mov.f32 %f896, %f894; mov.f32 %f897, %f894; @%p15 bra $L__BB0_36; ld.const.u32 %r6, [params+588]; cvt.rn.f32.s32 %f203, %r3; rcp.rn.f32 %f16, %f203; ld.global.u32 %r325, [%rd3]; mul.f32 %f17, %f10, 0f3456BF95; mul.f32 %f18, %f11, 0f3456BF95; mul.f32 %f19, %f12, 0f3456BF95; ld.const.u64 %rd4, [params+96]; mul.f32 %f204, %f877, %f14; mul.f32 %f205, %f878, %f13; sub.f32 %f20, %f205, %f204; mul.f32 %f206, %f879, %f13; mul.f32 %f207, %f877, %f15; sub.f32 %f21, %f207, %f206; mul.f32 %f208, %f878, %f15; mul.f32 %f209, %f879, %f14; sub.f32 %f22, %f209, %f208; mov.u32 %r67, 0; add.s64 %rd5, %rd1, 24; mov.u64 %rd31, __cudart_i2opi_f; abs.f32 %f272, %f17; abs.f32 %f273, %f18; max.f32 %f274, %f272, %f273; abs.f32 %f275, %f19; max.f32 %f276, %f274, %f275; mov.u32 %r322, %r67; $L__BB0_7: mov.u32 %r324, %r67; $L__BB0_8: cvt.rn.f32.s32 %f861, %r322; mad.lo.s32 %r69, %r325, 1664525, 1013904223; and.b32 %r70, %r69, 16777215; cvt.rn.f32.u32 %f210, %r70; fma.rn.f32 %f211, %f210, 0f33800000, %f861; mul.f32 %f32, %f16, %f211; mad.lo.s32 %r325, %r69, 1664525, 1013904223; and.b32 %r71, %r325, 16777215; cvt.rn.f32.u32 %f212, %r71; cvt.rn.f32.s32 %f213, %r324; fma.rn.f32 %f214, %f212, 0f33800000, %f213; mul.f32 %f215, %f16, %f214; mul.f32 %f216, %f32, %f32; mov.f32 %f217, 0f3F800000; sub.f32 %f218, %f217, %f216; mov.f32 %f219, 0f00000000; max.f32 %f220, %f219, %f218; sqrt.rn.f32 %f33, %f220; mul.f32 %f34, %f215, 0f40C90FDB; mul.f32 %f221, %f34, 0f3F22F983; cvt.rni.s32.f32 %r332, %f221; cvt.rn.f32.s32 %f222, %r332; mov.f32 %f223, 0fBFC90FDA; fma.rn.f32 %f224, %f222, %f223, %f34; mov.f32 %f225, 0fB3A22168; fma.rn.f32 %f226, %f222, %f225, %f224; mov.f32 %f227, 0fA7C234C5; fma.rn.f32 %f891, %f222, %f227, %f226; abs.f32 %f36, %f34; setp.ltu.f32 %p16, %f36, 0f47CE4780; mov.u32 %r329, %r332; mov.f32 %f888, %f891; @%p16 bra $L__BB0_16; setp.eq.f32 %p17, %f36, 0f7F800000; @%p17 bra $L__BB0_15; bra.uni $L__BB0_10; $L__BB0_15: mov.f32 %f230, 0f00000000; mul.rn.f32 %f888, %f34, %f230; mov.u32 %r329, 0; bra.uni $L__BB0_16; $L__BB0_10: mov.b32 %r14, %f34; bfe.u32 %r73, %r14, 23, 8; add.s32 %r15, %r73, -128; shl.b32 %r74, %r14, 8; or.b32 %r16, %r74, -2147483648; shr.u32 %r17, %r15, 5; mov.u64 %rd82, 0; mov.u32 %r326, 0; mov.u64 %rd80, %rd1; mov.u64 %rd81, %rd31; $L__BB0_11: .pragma "nounroll"; ld.global.nc.u32 %r75, [%rd81]; mad.wide.u32 %rd33, %r75, %r16, %rd82; shr.u64 %rd82, %rd33, 32; st.local.u32 [%rd80], %rd33; add.s64 %rd81, %rd81, 4; add.s64 %rd80, %rd80, 4; add.s32 %r326, %r326, 1; setp.ne.s32 %p18, %r326, 6; @%p18 bra $L__BB0_11; st.local.u32 [%rd5], %rd82; mov.u32 %r76, 4; sub.s32 %r20, %r76, %r17; mov.u32 %r77, 6; sub.s32 %r78, %r77, %r17; mul.wide.s32 %rd34, %r78, 4; add.s64 %rd35, %rd1, %rd34; ld.local.u32 %r327, [%rd35]; ld.local.u32 %r328, [%rd35+-4]; and.b32 %r23, %r15, 31; setp.eq.s32 %p19, %r23, 0; @%p19 bra $L__BB0_14; mov.u32 %r79, 32; sub.s32 %r80, %r79, %r23; shr.u32 %r81, %r328, %r80; shl.b32 %r82, %r327, %r23; add.s32 %r327, %r81, %r82; mul.wide.s32 %rd36, %r20, 4; add.s64 %rd37, %rd1, %rd36; ld.local.u32 %r83, [%rd37]; shr.u32 %r84, %r83, %r80; shl.b32 %r85, %r328, %r23; add.s32 %r328, %r84, %r85; $L__BB0_14: and.b32 %r86, %r14, -2147483648; shr.u32 %r87, %r328, 30; shl.b32 %r88, %r327, 2; or.b32 %r89, %r87, %r88; shr.u32 %r90, %r89, 31; shr.u32 %r91, %r327, 30; add.s32 %r92, %r90, %r91; neg.s32 %r93, %r92; setp.eq.s32 %p20, %r86, 0; selp.b32 %r329, %r92, %r93, %p20; setp.ne.s32 %p21, %r90, 0; xor.b32 %r94, %r86, -2147483648; selp.b32 %r95, %r94, %r86, %p21; selp.b32 %r96, -1, 0, %p21; xor.b32 %r97, %r89, %r96; shl.b32 %r98, %r328, 2; xor.b32 %r99, %r98, %r96; cvt.u64.u32 %rd38, %r97; cvt.u64.u32 %rd39, %r99; bfi.b64 %rd40, %rd38, %rd39, 32, 32; cvt.rn.f64.s64 %fd1, %rd40; mul.f64 %fd2, %fd1, 0d3BF921FB54442D19; cvt.rn.f32.f64 %f228, %fd2; setp.eq.s32 %p22, %r95, 0; neg.f32 %f229, %f228; selp.f32 %f888, %f228, %f229, %p22; $L__BB0_16: add.s32 %r30, %r329, 1; and.b32 %r31, %r30, 1; setp.eq.s32 %p23, %r31, 0; selp.f32 %f40, %f888, 0f3F800000, %p23; mul.rn.f32 %f41, %f888, %f888; mov.f32 %f889, 0fB94D4153; @%p23 bra $L__BB0_18; mov.f32 %f232, 0fBAB607ED; mov.f32 %f233, 0f37CBAC00; fma.rn.f32 %f889, %f233, %f41, %f232; $L__BB0_18: selp.f32 %f234, 0f3C0885E4, 0f3D2AAABB, %p23; fma.rn.f32 %f235, %f889, %f41, %f234; selp.f32 %f236, 0fBE2AAAA8, 0fBEFFFFFF, %p23; fma.rn.f32 %f237, %f235, %f41, %f236; mov.f32 %f238, 0f00000000; fma.rn.f32 %f239, %f41, %f40, %f238; fma.rn.f32 %f890, %f237, %f239, %f40; and.b32 %r101, %r30, 2; setp.eq.s32 %p25, %r101, 0; @%p25 bra $L__BB0_20; mov.f32 %f241, 0fBF800000; fma.rn.f32 %f890, %f890, %f241, %f238; $L__BB0_20: @%p16 bra $L__BB0_28; setp.eq.f32 %p27, %f36, 0f7F800000; @%p27 bra $L__BB0_27; bra.uni $L__BB0_22; $L__BB0_27: mov.f32 %f244, 0f00000000; mul.rn.f32 %f891, %f34, %f244; mov.u32 %r332, 0; bra.uni $L__BB0_28; $L__BB0_22: mov.b32 %r32, %f34; bfe.u32 %r102, %r32, 23, 8; add.s32 %r33, %r102, -128; shl.b32 %r103, %r32, 8; or.b32 %r34, %r103, -2147483648; shr.u32 %r35, %r33, 5; mov.u64 %rd83, 0; mov.u64 %rd84, %rd83; $L__BB0_23: .pragma "nounroll"; shl.b64 %rd43, %rd83, 2; mov.u64 %rd44, __cudart_i2opi_f; add.s64 %rd45, %rd44, %rd43; ld.global.nc.u32 %r104, [%rd45]; mad.wide.u32 %rd46, %r104, %r34, %rd84; shr.u64 %rd84, %rd46, 32; add.s64 %rd47, %rd1, %rd43; st.local.u32 [%rd47], %rd46; cvt.u32.u64 %r105, %rd83; add.s32 %r106, %r105, 1; cvt.s64.s32 %rd83, %r106; setp.ne.s32 %p28, %r106, 6; @%p28 bra $L__BB0_23; st.local.u32 [%rd5], %rd84; mov.u32 %r107, 4; sub.s32 %r36, %r107, %r35; mov.u32 %r108, 6; sub.s32 %r109, %r108, %r35; mul.wide.s32 %rd48, %r109, 4; add.s64 %rd49, %rd1, %rd48; ld.local.u32 %r330, [%rd49]; ld.local.u32 %r331, [%rd49+-4]; and.b32 %r39, %r33, 31; setp.eq.s32 %p29, %r39, 0; @%p29 bra $L__BB0_26; mov.u32 %r110, 32; sub.s32 %r111, %r110, %r39; shr.u32 %r112, %r331, %r111; shl.b32 %r113, %r330, %r39; add.s32 %r330, %r112, %r113; mul.wide.s32 %rd50, %r36, 4; add.s64 %rd51, %rd1, %rd50; ld.local.u32 %r114, [%rd51]; shr.u32 %r115, %r114, %r111; shl.b32 %r116, %r331, %r39; add.s32 %r331, %r115, %r116; $L__BB0_26: and.b32 %r117, %r32, -2147483648; shr.u32 %r118, %r331, 30; shl.b32 %r119, %r330, 2; or.b32 %r120, %r118, %r119; shr.u32 %r121, %r120, 31; shr.u32 %r122, %r330, 30; add.s32 %r123, %r121, %r122; neg.s32 %r124, %r123; setp.eq.s32 %p30, %r117, 0; selp.b32 %r332, %r123, %r124, %p30; setp.ne.s32 %p31, %r121, 0; xor.b32 %r125, %r117, -2147483648; selp.b32 %r126, %r125, %r117, %p31; selp.b32 %r127, -1, 0, %p31; xor.b32 %r128, %r120, %r127; shl.b32 %r129, %r331, 2; xor.b32 %r130, %r129, %r127; cvt.u64.u32 %rd52, %r128; cvt.u64.u32 %rd53, %r130; bfi.b64 %rd54, %rd52, %rd53, 32, 32; cvt.rn.f64.s64 %fd3, %rd54; mul.f64 %fd4, %fd3, 0d3BF921FB54442D19; cvt.rn.f32.f64 %f242, %fd4; setp.eq.s32 %p32, %r126, 0; neg.f32 %f243, %f242; selp.f32 %f891, %f242, %f243, %p32; $L__BB0_28: mul.f32 %f50, %f33, %f890; and.b32 %r46, %r332, 1; setp.eq.s32 %p33, %r46, 0; selp.f32 %f51, %f891, 0f3F800000, %p33; mul.rn.f32 %f52, %f891, %f891; mov.f32 %f892, 0fB94D4153; @%p33 bra $L__BB0_30; mov.f32 %f246, 0fBAB607ED; mov.f32 %f247, 0f37CBAC00; fma.rn.f32 %f892, %f247, %f52, %f246; $L__BB0_30: selp.f32 %f248, 0f3C0885E4, 0f3D2AAABB, %p33; fma.rn.f32 %f249, %f892, %f52, %f248; selp.f32 %f250, 0fBE2AAAA8, 0fBEFFFFFF, %p33; fma.rn.f32 %f251, %f249, %f52, %f250; mov.f32 %f252, 0f00000000; fma.rn.f32 %f253, %f52, %f51, %f252; fma.rn.f32 %f893, %f251, %f253, %f51; and.b32 %r132, %r332, 2; setp.eq.s32 %p35, %r132, 0; @%p35 bra $L__BB0_32; mov.f32 %f255, 0fBF800000; fma.rn.f32 %f893, %f893, %f255, %f252; $L__BB0_32: mul.f32 %f256, %f33, %f893; mul.f32 %f257, %f13, %f256; mul.f32 %f258, %f14, %f256; mul.f32 %f259, %f15, %f256; fma.rn.f32 %f260, %f22, %f50, %f257; fma.rn.f32 %f261, %f21, %f50, %f258; fma.rn.f32 %f262, %f20, %f50, %f259; fma.rn.f32 %f58, %f877, %f32, %f260; fma.rn.f32 %f59, %f878, %f32, %f261; fma.rn.f32 %f60, %f879, %f32, %f262; setp.leu.f32 %p36, %f59, 0f00000000; setp.ne.s32 %p37, %r6, 0; and.pred %p38, %p37, %p36; @%p38 bra $L__BB0_34; mov.f32 %f277, 0f38D1B717; max.f32 %f269, %f276, %f277; mov.f32 %f270, 0f6C4ECB8F; mov.f32 %f271, 0f00000000; mov.u32 %r169, 2; mov.u32 %r171, 1; mov.u32 %r172, 1065353216; mov.u32 %r203, 0; // begin inline asm call(%r133,%r134,%r135,%r136,%r137,%r138,%r139,%r140,%r141,%r142,%r143,%r144,%r145,%r146,%r147,%r148,%r149,%r150,%r151,%r152,%r153,%r154,%r155,%r156,%r157,%r158,%r159,%r160,%r161,%r162,%r163,%r164),_optix_trace_typed_32,(%r203,%rd4,%f10,%f11,%f12,%f58,%f59,%f60,%f269,%f270,%f271,%r171,%r203,%r171,%r169,%r171,%r171,%r172,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203,%r203); // end inline asm mov.b32 %f278, %r133; fma.rn.f32 %f894, %f58, %f278, %f894; fma.rn.f32 %f895, %f59, %f278, %f895; fma.rn.f32 %f896, %f60, %f278, %f896; mul.f32 %f279, %f878, %f59; fma.rn.f32 %f280, %f877, %f58, %f279; fma.rn.f32 %f281, %f879, %f60, %f280; cvt.sat.f32.f32 %f282, %f281; fma.rn.f32 %f897, %f282, %f278, %f897; $L__BB0_34: add.s32 %r324, %r324, 1; setp.lt.s32 %p39, %r324, %r3; @%p39 bra $L__BB0_8; add.s32 %r322, %r322, 1; setp.lt.s32 %p40, %r322, %r3; @%p40 bra $L__BB0_7; $L__BB0_36: mul.lo.s32 %r204, %r3, %r3; cvt.rn.f32.s32 %f283, %r204; div.rn.f32 %f284, %f897, %f283; add.f32 %f285, %f284, %f284; ld.const.v4.f32 {%f286, %f287, %f288, %f289}, [params+576]; mul.f32 %f73, %f285, %f286; mul.f32 %f74, %f285, %f287; mul.f32 %f75, %f285, %f288; ld.const.u32 %r49, [params+104]; and.b32 %r205, %r49, 1; setp.eq.b32 %p41, %r205, 1; mov.pred %p42, 0; xor.pred %p43, %p41, %p42; not.pred %p44, %p43; @%p44 bra $L__BB0_110; mov.f32 %f297, 0f3EE66666; abs.f32 %f77, %f73; setp.lt.f32 %p45, %f77, 0f00800000; mul.f32 %f299, %f77, 0f4B800000; selp.f32 %f300, %f299, %f77, %p45; selp.f32 %f301, 0fC3170000, 0fC2FE0000, %p45; mov.b32 %r206, %f300; and.b32 %r207, %r206, 8388607; or.b32 %r208, %r207, 1065353216; mov.b32 %f302, %r208; shr.u32 %r209, %r206, 23; cvt.rn.f32.u32 %f303, %r209; add.f32 %f304, %f301, %f303; setp.gt.f32 %p46, %f302, 0f3FB504F3; mul.f32 %f305, %f302, 0f3F000000; add.f32 %f306, %f304, 0f3F800000; selp.f32 %f307, %f306, %f304, %p46; selp.f32 %f308, %f305, %f302, %p46; add.f32 %f309, %f308, 0fBF800000; add.f32 %f310, %f308, 0f3F800000; rcp.approx.ftz.f32 %f311, %f310; add.f32 %f312, %f309, %f309; mul.f32 %f313, %f312, %f311; mul.f32 %f314, %f313, %f313; mov.f32 %f315, 0f3C4CAF63; mov.f32 %f316, 0f3B18F0FE; fma.rn.f32 %f317, %f316, %f314, %f315; mov.f32 %f318, 0f3DAAAABD; fma.rn.f32 %f319, %f317, %f314, %f318; mul.rn.f32 %f320, %f319, %f314; mul.rn.f32 %f321, %f320, %f313; sub.f32 %f322, %f309, %f313; add.f32 %f323, %f322, %f322; neg.f32 %f324, %f313; fma.rn.f32 %f325, %f324, %f309, %f323; mul.rn.f32 %f326, %f311, %f325; add.f32 %f327, %f321, %f313; sub.f32 %f328, %f313, %f327; add.f32 %f329, %f321, %f328; add.f32 %f330, %f326, %f329; add.f32 %f331, %f327, %f330; sub.f32 %f332, %f327, %f331; add.f32 %f333, %f330, %f332; mov.f32 %f334, 0f3F317200; mul.rn.f32 %f335, %f307, %f334; mov.f32 %f336, 0f35BFBE8E; mul.rn.f32 %f337, %f307, %f336; add.f32 %f338, %f335, %f331; sub.f32 %f339, %f335, %f338; add.f32 %f340, %f331, %f339; add.f32 %f341, %f333, %f340; add.f32 %f342, %f337, %f341; add.f32 %f343, %f338, %f342; sub.f32 %f344, %f338, %f343; add.f32 %f345, %f342, %f344; mul.rn.f32 %f346, %f297, %f343; neg.f32 %f347, %f346; fma.rn.f32 %f348, %f297, %f343, %f347; fma.rn.f32 %f349, %f297, %f345, %f348; mov.f32 %f350, 0f00000000; fma.rn.f32 %f351, %f350, %f343, %f349; add.rn.f32 %f352, %f346, %f351; neg.f32 %f353, %f352; add.rn.f32 %f354, %f346, %f353; add.rn.f32 %f355, %f354, %f351; mov.b32 %r210, %f352; setp.eq.s32 %p47, %r210, 1118925336; add.s32 %r211, %r210, -1; mov.b32 %f356, %r211; add.f32 %f357, %f355, 0f37000000; selp.f32 %f78, %f357, %f355, %p47; selp.f32 %f358, %f356, %f352, %p47; mov.f32 %f359, 0f3FB8AA3B; mul.rn.f32 %f360, %f358, %f359; cvt.rzi.f32.f32 %f361, %f360; abs.f32 %f362, %f361; setp.gt.f32 %p48, %f362, 0f42FC0000; mov.b32 %r212, %f361; and.b32 %r213, %r212, -2147483648; or.b32 %r214, %r213, 1123811328; mov.b32 %f363, %r214; selp.f32 %f364, %f363, %f361, %p48; mov.f32 %f365, 0fBF317218; fma.rn.f32 %f366, %f364, %f365, %f358; mov.f32 %f367, 0f3102E308; fma.rn.f32 %f368, %f364, %f367, %f366; mul.f32 %f369, %f368, 0f3FB8AA3B; add.f32 %f370, %f364, 0f4B40007F; mov.b32 %r215, %f370; shl.b32 %r216, %r215, 23; mov.b32 %f371, %r216; ex2.approx.ftz.f32 %f372, %f369; mul.f32 %f79, %f372, %f371; setp.eq.f32 %p49, %f79, 0f7F800000; mov.f32 %f902, 0f7F800000; @%p49 bra $L__BB0_39; fma.rn.f32 %f902, %f79, %f78, %f79; $L__BB0_39: mov.f32 %f867, 0f3E666666; cvt.rzi.f32.f32 %f866, %f867; add.f32 %f865, %f866, %f866; mov.f32 %f864, 0f3EE66666; sub.f32 %f863, %f864, %f865; abs.f32 %f862, %f863; setp.lt.f32 %p50, %f73, 0f00000000; setp.eq.f32 %p51, %f862, 0f3F800000; and.pred %p1, %p50, %p51; setp.eq.f32 %p52, %f73, 0f00000000; @%p52 bra $L__BB0_43; bra.uni $L__BB0_40; $L__BB0_43: add.f32 %f377, %f73, %f73; selp.f32 %f904, %f377, 0f00000000, %p51; bra.uni $L__BB0_44; $L__BB0_121: mov.f32 %f854, 0f00000000; // begin inline asm { cvt.rn.f16.f32 %rs52, %f854;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs51, %f854;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs50, %f854;} // end inline asm mov.u16 %rs53, 0; st.global.v4.u16 [%rd19], {%rs50, %rs51, %rs52, %rs53}; $L__BB0_122: and.b32 %r315, %r50, 64; setp.eq.s32 %p147, %r315, 0; @%p147 bra $L__BB0_124; ld.const.u64 %rd73, [params+208]; cvta.to.global.u64 %rd74, %rd73; ld.const.u32 %r316, [params+200]; mad.lo.s32 %r317, %r316, %r5, %r4; mul.wide.u32 %rd75, %r317, 4; add.s64 %rd76, %rd74, %rd75; mov.u16 %rs54, 0; st.global.v4.u8 [%rd76], {%rs54, %rs54, %rs54, %rs54}; bra.uni $L__BB0_124; $L__BB0_40: mov.b32 %r217, %f902; xor.b32 %r218, %r217, -2147483648; mov.b32 %f373, %r218; selp.f32 %f904, %f373, %f902, %p1; setp.geu.f32 %p53, %f73, 0f00000000; @%p53 bra $L__BB0_44; mov.f32 %f374, 0f3EE66666; cvt.rzi.f32.f32 %f375, %f374; setp.eq.f32 %p54, %f375, 0f3EE66666; @%p54 bra $L__BB0_44; mov.f32 %f904, 0f7FFFFFFF; $L__BB0_44: abs.f32 %f868, %f73; add.f32 %f378, %f868, 0f3EE66666; mov.b32 %r219, %f378; setp.lt.s32 %p56, %r219, 2139095040; @%p56 bra $L__BB0_49; abs.f32 %f869, %f73; setp.gtu.f32 %p57, %f869, 0f7F800000; @%p57 bra $L__BB0_48; bra.uni $L__BB0_46; $L__BB0_48: add.f32 %f904, %f73, 0f3EE66666; bra.uni $L__BB0_49; $L__BB0_46: abs.f32 %f870, %f73; setp.neu.f32 %p58, %f870, 0f7F800000; @%p58 bra $L__BB0_49; selp.f32 %f904, 0fFF800000, 0f7F800000, %p1; $L__BB0_49: setp.eq.f32 %p59, %f73, 0f3F800000; selp.f32 %f88, 0f3F800000, %f904, %p59; abs.f32 %f89, %f74; setp.lt.f32 %p60, %f89, 0f00800000; mul.f32 %f380, %f89, 0f4B800000; selp.f32 %f381, %f380, %f89, %p60; selp.f32 %f382, 0fC3170000, 0fC2FE0000, %p60; mov.b32 %r220, %f381; and.b32 %r221, %r220, 8388607; or.b32 %r222, %r221, 1065353216; mov.b32 %f383, %r222; shr.u32 %r223, %r220, 23; cvt.rn.f32.u32 %f384, %r223; add.f32 %f385, %f382, %f384; setp.gt.f32 %p61, %f383, 0f3FB504F3; mul.f32 %f386, %f383, 0f3F000000; add.f32 %f387, %f385, 0f3F800000; selp.f32 %f388, %f387, %f385, %p61; selp.f32 %f389, %f386, %f383, %p61; add.f32 %f390, %f389, 0fBF800000; add.f32 %f391, %f389, 0f3F800000; rcp.approx.ftz.f32 %f392, %f391; add.f32 %f393, %f390, %f390; mul.f32 %f394, %f393, %f392; mul.f32 %f395, %f394, %f394; mov.f32 %f396, 0f3C4CAF63; mov.f32 %f397, 0f3B18F0FE; fma.rn.f32 %f398, %f397, %f395, %f396; mov.f32 %f399, 0f3DAAAABD; fma.rn.f32 %f400, %f398, %f395, %f399; mul.rn.f32 %f401, %f400, %f395; mul.rn.f32 %f402, %f401, %f394; sub.f32 %f403, %f390, %f394; add.f32 %f404, %f403, %f403; neg.f32 %f405, %f394; fma.rn.f32 %f406, %f405, %f390, %f404; mul.rn.f32 %f407, %f392, %f406; add.f32 %f408, %f402, %f394; sub.f32 %f409, %f394, %f408; add.f32 %f410, %f402, %f409; add.f32 %f411, %f407, %f410; add.f32 %f412, %f408, %f411; sub.f32 %f413, %f408, %f412; add.f32 %f414, %f411, %f413; mov.f32 %f415, 0f3F317200; mul.rn.f32 %f416, %f388, %f415; mov.f32 %f417, 0f35BFBE8E; mul.rn.f32 %f418, %f388, %f417; add.f32 %f419, %f416, %f412; sub.f32 %f420, %f416, %f419; add.f32 %f421, %f412, %f420; add.f32 %f422, %f414, %f421; add.f32 %f423, %f418, %f422; add.f32 %f424, %f419, %f423; sub.f32 %f425, %f419, %f424; add.f32 %f426, %f423, %f425; mov.f32 %f427, 0f3EE66666; mul.rn.f32 %f428, %f427, %f424; neg.f32 %f429, %f428; fma.rn.f32 %f430, %f427, %f424, %f429; fma.rn.f32 %f431, %f427, %f426, %f430; mov.f32 %f432, 0f00000000; fma.rn.f32 %f433, %f432, %f424, %f431; add.rn.f32 %f434, %f428, %f433; neg.f32 %f435, %f434; add.rn.f32 %f436, %f428, %f435; add.rn.f32 %f437, %f436, %f433; mov.b32 %r224, %f434; setp.eq.s32 %p62, %r224, 1118925336; add.s32 %r225, %r224, -1; mov.b32 %f438, %r225; add.f32 %f439, %f437, 0f37000000; selp.f32 %f90, %f439, %f437, %p62; selp.f32 %f440, %f438, %f434, %p62; mov.f32 %f441, 0f3FB8AA3B; mul.rn.f32 %f442, %f440, %f441; cvt.rzi.f32.f32 %f443, %f442; abs.f32 %f444, %f443; setp.gt.f32 %p63, %f444, 0f42FC0000; mov.b32 %r226, %f443; and.b32 %r227, %r226, -2147483648; or.b32 %r228, %r227, 1123811328; mov.b32 %f445, %r228; selp.f32 %f446, %f445, %f443, %p63; mov.f32 %f447, 0fBF317218; fma.rn.f32 %f448, %f446, %f447, %f440; mov.f32 %f449, 0f3102E308; fma.rn.f32 %f450, %f446, %f449, %f448; mul.f32 %f451, %f450, 0f3FB8AA3B; add.f32 %f452, %f446, 0f4B40007F; mov.b32 %r229, %f452; shl.b32 %r230, %r229, 23; mov.b32 %f453, %r230; ex2.approx.ftz.f32 %f454, %f451; mul.f32 %f91, %f454, %f453; setp.eq.f32 %p64, %f91, 0f7F800000; mov.f32 %f905, 0f7F800000; @%p64 bra $L__BB0_51; fma.rn.f32 %f905, %f91, %f90, %f91; $L__BB0_51: setp.lt.f32 %p65, %f74, 0f00000000; and.pred %p2, %p65, %p51; setp.eq.f32 %p67, %f74, 0f00000000; @%p67 bra $L__BB0_55; bra.uni $L__BB0_52; $L__BB0_55: add.f32 %f459, %f74, %f74; selp.f32 %f907, %f459, 0f00000000, %p51; bra.uni $L__BB0_56; $L__BB0_52: mov.b32 %r231, %f905; xor.b32 %r232, %r231, -2147483648; mov.b32 %f455, %r232; selp.f32 %f907, %f455, %f905, %p2; setp.geu.f32 %p68, %f74, 0f00000000; @%p68 bra $L__BB0_56; mov.f32 %f456, 0f3EE66666; cvt.rzi.f32.f32 %f457, %f456; setp.eq.f32 %p69, %f457, 0f3EE66666; @%p69 bra $L__BB0_56; mov.f32 %f907, 0f7FFFFFFF; $L__BB0_56: abs.f32 %f871, %f74; add.f32 %f460, %f871, 0f3EE66666; mov.b32 %r233, %f460; setp.lt.s32 %p71, %r233, 2139095040; @%p71 bra $L__BB0_61; abs.f32 %f872, %f74; setp.gtu.f32 %p72, %f872, 0f7F800000; @%p72 bra $L__BB0_60; bra.uni $L__BB0_58; $L__BB0_60: add.f32 %f907, %f74, 0f3EE66666; bra.uni $L__BB0_61; $L__BB0_58: abs.f32 %f873, %f74; setp.neu.f32 %p73, %f873, 0f7F800000; @%p73 bra $L__BB0_61; selp.f32 %f907, 0fFF800000, 0f7F800000, %p2; $L__BB0_61: setp.eq.f32 %p74, %f74, 0f3F800000; selp.f32 %f100, 0f3F800000, %f907, %p74; abs.f32 %f101, %f75; setp.lt.f32 %p75, %f101, 0f00800000; mul.f32 %f462, %f101, 0f4B800000; selp.f32 %f463, %f462, %f101, %p75; selp.f32 %f464, 0fC3170000, 0fC2FE0000, %p75; mov.b32 %r234, %f463; and.b32 %r235, %r234, 8388607; or.b32 %r236, %r235, 1065353216; mov.b32 %f465, %r236; shr.u32 %r237, %r234, 23; cvt.rn.f32.u32 %f466, %r237; add.f32 %f467, %f464, %f466; setp.gt.f32 %p76, %f465, 0f3FB504F3; mul.f32 %f468, %f465, 0f3F000000; add.f32 %f469, %f467, 0f3F800000; selp.f32 %f470, %f469, %f467, %p76; selp.f32 %f471, %f468, %f465, %p76; add.f32 %f472, %f471, 0fBF800000; add.f32 %f473, %f471, 0f3F800000; rcp.approx.ftz.f32 %f474, %f473; add.f32 %f475, %f472, %f472; mul.f32 %f476, %f475, %f474; mul.f32 %f477, %f476, %f476; mov.f32 %f478, 0f3C4CAF63; mov.f32 %f479, 0f3B18F0FE; fma.rn.f32 %f480, %f479, %f477, %f478; mov.f32 %f481, 0f3DAAAABD; fma.rn.f32 %f482, %f480, %f477, %f481; mul.rn.f32 %f483, %f482, %f477; mul.rn.f32 %f484, %f483, %f476; sub.f32 %f485, %f472, %f476; add.f32 %f486, %f485, %f485; neg.f32 %f487, %f476; fma.rn.f32 %f488, %f487, %f472, %f486; mul.rn.f32 %f489, %f474, %f488; add.f32 %f490, %f484, %f476; sub.f32 %f491, %f476, %f490; add.f32 %f492, %f484, %f491; add.f32 %f493, %f489, %f492; add.f32 %f494, %f490, %f493; sub.f32 %f495, %f490, %f494; add.f32 %f496, %f493, %f495; mov.f32 %f497, 0f3F317200; mul.rn.f32 %f498, %f470, %f497; mov.f32 %f499, 0f35BFBE8E; mul.rn.f32 %f500, %f470, %f499; add.f32 %f501, %f498, %f494; sub.f32 %f502, %f498, %f501; add.f32 %f503, %f494, %f502; add.f32 %f504, %f496, %f503; add.f32 %f505, %f500, %f504; add.f32 %f506, %f501, %f505; sub.f32 %f507, %f501, %f506; add.f32 %f508, %f505, %f507; mov.f32 %f509, 0f3EE66666; mul.rn.f32 %f510, %f509, %f506; neg.f32 %f511, %f510; fma.rn.f32 %f512, %f509, %f506, %f511; fma.rn.f32 %f513, %f509, %f508, %f512; mov.f32 %f514, 0f00000000; fma.rn.f32 %f515, %f514, %f506, %f513; add.rn.f32 %f516, %f510, %f515; neg.f32 %f517, %f516; add.rn.f32 %f518, %f510, %f517; add.rn.f32 %f519, %f518, %f515; mov.b32 %r238, %f516; setp.eq.s32 %p77, %r238, 1118925336; add.s32 %r239, %r238, -1; mov.b32 %f520, %r239; add.f32 %f521, %f519, 0f37000000; selp.f32 %f102, %f521, %f519, %p77; selp.f32 %f522, %f520, %f516, %p77; mov.f32 %f523, 0f3FB8AA3B; mul.rn.f32 %f524, %f522, %f523; cvt.rzi.f32.f32 %f525, %f524; abs.f32 %f526, %f525; setp.gt.f32 %p78, %f526, 0f42FC0000; mov.b32 %r240, %f525; and.b32 %r241, %r240, -2147483648; or.b32 %r242, %r241, 1123811328; mov.b32 %f527, %r242; selp.f32 %f528, %f527, %f525, %p78; mov.f32 %f529, 0fBF317218; fma.rn.f32 %f530, %f528, %f529, %f522; mov.f32 %f531, 0f3102E308; fma.rn.f32 %f532, %f528, %f531, %f530; mul.f32 %f533, %f532, 0f3FB8AA3B; add.f32 %f534, %f528, 0f4B40007F; mov.b32 %r243, %f534; shl.b32 %r244, %r243, 23; mov.b32 %f535, %r244; ex2.approx.ftz.f32 %f536, %f533; mul.f32 %f103, %f536, %f535; setp.eq.f32 %p79, %f103, 0f7F800000; mov.f32 %f908, 0f7F800000; @%p79 bra $L__BB0_63; fma.rn.f32 %f908, %f103, %f102, %f103; $L__BB0_63: setp.lt.f32 %p80, %f75, 0f00000000; and.pred %p3, %p80, %p51; setp.eq.f32 %p82, %f75, 0f00000000; @%p82 bra $L__BB0_67; bra.uni $L__BB0_64; $L__BB0_67: add.f32 %f541, %f75, %f75; selp.f32 %f910, %f541, 0f00000000, %p51; bra.uni $L__BB0_68; $L__BB0_64: mov.b32 %r245, %f908; xor.b32 %r246, %r245, -2147483648; mov.b32 %f537, %r246; selp.f32 %f910, %f537, %f908, %p3; setp.geu.f32 %p83, %f75, 0f00000000; @%p83 bra $L__BB0_68; mov.f32 %f538, 0f3EE66666; cvt.rzi.f32.f32 %f539, %f538; setp.eq.f32 %p84, %f539, 0f3EE66666; @%p84 bra $L__BB0_68; mov.f32 %f910, 0f7FFFFFFF; $L__BB0_68: abs.f32 %f874, %f75; add.f32 %f542, %f874, 0f3EE66666; mov.b32 %r247, %f542; setp.lt.s32 %p86, %r247, 2139095040; @%p86 bra $L__BB0_73; abs.f32 %f875, %f75; setp.gtu.f32 %p87, %f875, 0f7F800000; @%p87 bra $L__BB0_72; bra.uni $L__BB0_70; $L__BB0_72: add.f32 %f910, %f75, 0f3EE66666; bra.uni $L__BB0_73; $L__BB0_70: abs.f32 %f876, %f75; setp.neu.f32 %p88, %f876, 0f7F800000; @%p88 bra $L__BB0_73; selp.f32 %f910, 0fFF800000, 0f7F800000, %p3; $L__BB0_73: setp.eq.f32 %p89, %f75, 0f3F800000; mov.f32 %f544, 0f3F800000; selp.f32 %f545, 0f3F800000, %f910, %p89; min.f32 %f546, %f88, %f544; mov.f32 %f547, 0f00000000; max.f32 %f112, %f547, %f546; min.f32 %f548, %f100, %f544; max.f32 %f113, %f547, %f548; min.f32 %f549, %f545, %f544; max.f32 %f114, %f547, %f549; mov.f32 %f553, 0f3ED55555; abs.f32 %f116, %f112; setp.lt.f32 %p90, %f116, 0f00800000; mul.f32 %f555, %f116, 0f4B800000; selp.f32 %f556, %f555, %f116, %p90; selp.f32 %f557, 0fC3170000, 0fC2FE0000, %p90; mov.b32 %r250, %f556; and.b32 %r251, %r250, 8388607; or.b32 %r252, %r251, 1065353216; mov.b32 %f558, %r252; shr.u32 %r253, %r250, 23; cvt.rn.f32.u32 %f559, %r253; add.f32 %f560, %f557, %f559; setp.gt.f32 %p91, %f558, 0f3FB504F3; mul.f32 %f561, %f558, 0f3F000000; add.f32 %f562, %f560, 0f3F800000; selp.f32 %f563, %f562, %f560, %p91; selp.f32 %f564, %f561, %f558, %p91; add.f32 %f565, %f564, 0fBF800000; add.f32 %f566, %f564, 0f3F800000; rcp.approx.ftz.f32 %f567, %f566; add.f32 %f568, %f565, %f565; mul.f32 %f569, %f568, %f567; mul.f32 %f570, %f569, %f569; mov.f32 %f571, 0f3C4CAF63; mov.f32 %f572, 0f3B18F0FE; fma.rn.f32 %f573, %f572, %f570, %f571; mov.f32 %f574, 0f3DAAAABD; fma.rn.f32 %f575, %f573, %f570, %f574; mul.rn.f32 %f576, %f575, %f570; mul.rn.f32 %f577, %f576, %f569; sub.f32 %f578, %f565, %f569; add.f32 %f579, %f578, %f578; neg.f32 %f580, %f569; fma.rn.f32 %f581, %f580, %f565, %f579; mul.rn.f32 %f582, %f567, %f581; add.f32 %f583, %f577, %f569; sub.f32 %f584, %f569, %f583; add.f32 %f585, %f577, %f584; add.f32 %f586, %f582, %f585; add.f32 %f587, %f583, %f586; sub.f32 %f588, %f583, %f587; add.f32 %f589, %f586, %f588; mov.f32 %f590, 0f3F317200; mul.rn.f32 %f591, %f563, %f590; mov.f32 %f592, 0f35BFBE8E; mul.rn.f32 %f593, %f563, %f592; add.f32 %f594, %f591, %f587; sub.f32 %f595, %f591, %f594; add.f32 %f596, %f587, %f595; add.f32 %f597, %f589, %f596; add.f32 %f598, %f593, %f597; add.f32 %f599, %f594, %f598; sub.f32 %f600, %f594, %f599; add.f32 %f601, %f598, %f600; mul.rn.f32 %f602, %f553, %f599; neg.f32 %f603, %f602; fma.rn.f32 %f604, %f553, %f599, %f603; fma.rn.f32 %f605, %f553, %f601, %f604; fma.rn.f32 %f606, %f547, %f599, %f605; add.rn.f32 %f607, %f602, %f606; neg.f32 %f608, %f607; add.rn.f32 %f609, %f602, %f608; add.rn.f32 %f610, %f609, %f606; mov.b32 %r254, %f607; setp.eq.s32 %p92, %r254, 1118925336; add.s32 %r255, %r254, -1; mov.b32 %f611, %r255; add.f32 %f612, %f610, 0f37000000; selp.f32 %f117, %f612, %f610, %p92; selp.f32 %f613, %f611, %f607, %p92; mov.f32 %f614, 0f3FB8AA3B; mul.rn.f32 %f615, %f613, %f614; cvt.rzi.f32.f32 %f616, %f615; abs.f32 %f617, %f616; setp.gt.f32 %p93, %f617, 0f42FC0000; mov.b32 %r256, %f616; and.b32 %r257, %r256, -2147483648; or.b32 %r258, %r257, 1123811328; mov.b32 %f618, %r258; selp.f32 %f619, %f618, %f616, %p93; mov.f32 %f620, 0fBF317218; fma.rn.f32 %f621, %f619, %f620, %f613; mov.f32 %f622, 0f3102E308; fma.rn.f32 %f623, %f619, %f622, %f621; mul.f32 %f624, %f623, 0f3FB8AA3B; add.f32 %f625, %f619, 0f4B40007F; mov.b32 %r259, %f625; shl.b32 %r260, %r259, 23; mov.b32 %f626, %r260; ex2.approx.ftz.f32 %f627, %f624; mul.f32 %f118, %f627, %f626; setp.eq.f32 %p94, %f118, 0f7F800000; mov.f32 %f911, 0f7F800000; @%p94 bra $L__BB0_75; fma.rn.f32 %f911, %f118, %f117, %f118; $L__BB0_75: mov.f32 %f860, 0f3E555555; cvt.rzi.f32.f32 %f859, %f860; add.f32 %f858, %f859, %f859; mov.f32 %f857, 0f3ED55555; sub.f32 %f856, %f857, %f858; abs.f32 %f855, %f856; setp.lt.f32 %p95, %f112, 0f00000000; setp.eq.f32 %p96, %f855, 0f3F800000; and.pred %p4, %p95, %p96; setp.eq.f32 %p97, %f112, 0f00000000; @%p97 bra $L__BB0_79; bra.uni $L__BB0_76; $L__BB0_79: add.f32 %f632, %f112, %f112; selp.f32 %f913, %f632, 0f00000000, %p96; bra.uni $L__BB0_80; $L__BB0_76: mov.b32 %r261, %f911; xor.b32 %r262, %r261, -2147483648; mov.b32 %f628, %r262; selp.f32 %f913, %f628, %f911, %p4; setp.geu.f32 %p98, %f112, 0f00000000; @%p98 bra $L__BB0_80; mov.f32 %f629, 0f3ED55555; cvt.rzi.f32.f32 %f630, %f629; setp.eq.f32 %p99, %f630, 0f3ED55555; @%p99 bra $L__BB0_80; mov.f32 %f913, 0f7FFFFFFF; $L__BB0_80: add.f32 %f633, %f116, 0f3ED55555; mov.b32 %r263, %f633; setp.lt.s32 %p101, %r263, 2139095040; @%p101 bra $L__BB0_85; setp.gtu.f32 %p102, %f116, 0f7F800000; @%p102 bra $L__BB0_84; bra.uni $L__BB0_82; $L__BB0_84: add.f32 %f913, %f112, 0f3ED55555; bra.uni $L__BB0_85; $L__BB0_82: setp.neu.f32 %p103, %f116, 0f7F800000; @%p103 bra $L__BB0_85; selp.f32 %f913, 0fFF800000, 0f7F800000, %p4; $L__BB0_85: abs.f32 %f127, %f113; setp.lt.f32 %p104, %f127, 0f00800000; mul.f32 %f635, %f127, 0f4B800000; selp.f32 %f636, %f635, %f127, %p104; selp.f32 %f637, 0fC3170000, 0fC2FE0000, %p104; mov.b32 %r264, %f636; and.b32 %r265, %r264, 8388607; or.b32 %r266, %r265, 1065353216; mov.b32 %f638, %r266; shr.u32 %r267, %r264, 23; cvt.rn.f32.u32 %f639, %r267; add.f32 %f640, %f637, %f639; setp.gt.f32 %p105, %f638, 0f3FB504F3; mul.f32 %f641, %f638, 0f3F000000; add.f32 %f642, %f640, 0f3F800000; selp.f32 %f643, %f642, %f640, %p105; selp.f32 %f644, %f641, %f638, %p105; add.f32 %f645, %f644, 0fBF800000; add.f32 %f646, %f644, 0f3F800000; rcp.approx.ftz.f32 %f647, %f646; add.f32 %f648, %f645, %f645; mul.f32 %f649, %f648, %f647; mul.f32 %f650, %f649, %f649; mov.f32 %f651, 0f3C4CAF63; mov.f32 %f652, 0f3B18F0FE; fma.rn.f32 %f653, %f652, %f650, %f651; mov.f32 %f654, 0f3DAAAABD; fma.rn.f32 %f655, %f653, %f650, %f654; mul.rn.f32 %f656, %f655, %f650; mul.rn.f32 %f657, %f656, %f649; sub.f32 %f658, %f645, %f649; add.f32 %f659, %f658, %f658; neg.f32 %f660, %f649; fma.rn.f32 %f661, %f660, %f645, %f659; mul.rn.f32 %f662, %f647, %f661; add.f32 %f663, %f657, %f649; sub.f32 %f664, %f649, %f663; add.f32 %f665, %f657, %f664; add.f32 %f666, %f662, %f665; add.f32 %f667, %f663, %f666; sub.f32 %f668, %f663, %f667; add.f32 %f669, %f666, %f668; mov.f32 %f670, 0f3F317200; mul.rn.f32 %f671, %f643, %f670; mov.f32 %f672, 0f35BFBE8E; mul.rn.f32 %f673, %f643, %f672; add.f32 %f674, %f671, %f667; sub.f32 %f675, %f671, %f674; add.f32 %f676, %f667, %f675; add.f32 %f677, %f669, %f676; add.f32 %f678, %f673, %f677; add.f32 %f679, %f674, %f678; sub.f32 %f680, %f674, %f679; add.f32 %f681, %f678, %f680; mov.f32 %f682, 0f3ED55555; mul.rn.f32 %f683, %f682, %f679; neg.f32 %f684, %f683; fma.rn.f32 %f685, %f682, %f679, %f684; fma.rn.f32 %f686, %f682, %f681, %f685; mov.f32 %f687, 0f00000000; fma.rn.f32 %f688, %f687, %f679, %f686; add.rn.f32 %f689, %f683, %f688; neg.f32 %f690, %f689; add.rn.f32 %f691, %f683, %f690; add.rn.f32 %f692, %f691, %f688; mov.b32 %r268, %f689; setp.eq.s32 %p106, %r268, 1118925336; add.s32 %r269, %r268, -1; mov.b32 %f693, %r269; add.f32 %f694, %f692, 0f37000000; selp.f32 %f128, %f694, %f692, %p106; selp.f32 %f695, %f693, %f689, %p106; mov.f32 %f696, 0f3FB8AA3B; mul.rn.f32 %f697, %f695, %f696; cvt.rzi.f32.f32 %f698, %f697; abs.f32 %f699, %f698; setp.gt.f32 %p107, %f699, 0f42FC0000; mov.b32 %r270, %f698; and.b32 %r271, %r270, -2147483648; or.b32 %r272, %r271, 1123811328; mov.b32 %f700, %r272; selp.f32 %f701, %f700, %f698, %p107; mov.f32 %f702, 0fBF317218; fma.rn.f32 %f703, %f701, %f702, %f695; mov.f32 %f704, 0f3102E308; fma.rn.f32 %f705, %f701, %f704, %f703; mul.f32 %f706, %f705, 0f3FB8AA3B; add.f32 %f707, %f701, 0f4B40007F; mov.b32 %r273, %f707; shl.b32 %r274, %r273, 23; mov.b32 %f708, %r274; ex2.approx.ftz.f32 %f709, %f706; mul.f32 %f129, %f709, %f708; setp.eq.f32 %p108, %f129, 0f7F800000; mov.f32 %f914, 0f7F800000; @%p108 bra $L__BB0_87; fma.rn.f32 %f914, %f129, %f128, %f129; $L__BB0_87: setp.lt.f32 %p109, %f113, 0f00000000; and.pred %p5, %p109, %p96; setp.eq.f32 %p111, %f113, 0f00000000; @%p111 bra $L__BB0_91; bra.uni $L__BB0_88; $L__BB0_91: add.f32 %f714, %f113, %f113; selp.f32 %f916, %f714, 0f00000000, %p96; bra.uni $L__BB0_92; $L__BB0_88: mov.b32 %r275, %f914; xor.b32 %r276, %r275, -2147483648; mov.b32 %f710, %r276; selp.f32 %f916, %f710, %f914, %p5; setp.geu.f32 %p112, %f113, 0f00000000; @%p112 bra $L__BB0_92; mov.f32 %f711, 0f3ED55555; cvt.rzi.f32.f32 %f712, %f711; setp.eq.f32 %p113, %f712, 0f3ED55555; @%p113 bra $L__BB0_92; mov.f32 %f916, 0f7FFFFFFF; $L__BB0_92: add.f32 %f715, %f127, 0f3ED55555; mov.b32 %r277, %f715; setp.lt.s32 %p115, %r277, 2139095040; @%p115 bra $L__BB0_97; setp.gtu.f32 %p116, %f127, 0f7F800000; @%p116 bra $L__BB0_96; bra.uni $L__BB0_94; $L__BB0_96: add.f32 %f916, %f113, 0f3ED55555; bra.uni $L__BB0_97; $L__BB0_94: setp.neu.f32 %p117, %f127, 0f7F800000; @%p117 bra $L__BB0_97; selp.f32 %f916, 0fFF800000, 0f7F800000, %p5; $L__BB0_97: abs.f32 %f138, %f114; setp.lt.f32 %p118, %f138, 0f00800000; mul.f32 %f717, %f138, 0f4B800000; selp.f32 %f718, %f717, %f138, %p118; selp.f32 %f719, 0fC3170000, 0fC2FE0000, %p118; mov.b32 %r278, %f718; and.b32 %r279, %r278, 8388607; or.b32 %r280, %r279, 1065353216; mov.b32 %f720, %r280; shr.u32 %r281, %r278, 23; cvt.rn.f32.u32 %f721, %r281; add.f32 %f722, %f719, %f721; setp.gt.f32 %p119, %f720, 0f3FB504F3; mul.f32 %f723, %f720, 0f3F000000; add.f32 %f724, %f722, 0f3F800000; selp.f32 %f725, %f724, %f722, %p119; selp.f32 %f726, %f723, %f720, %p119; add.f32 %f727, %f726, 0fBF800000; add.f32 %f728, %f726, 0f3F800000; rcp.approx.ftz.f32 %f729, %f728; add.f32 %f730, %f727, %f727; mul.f32 %f731, %f730, %f729; mul.f32 %f732, %f731, %f731; mov.f32 %f733, 0f3C4CAF63; mov.f32 %f734, 0f3B18F0FE; fma.rn.f32 %f735, %f734, %f732, %f733; mov.f32 %f736, 0f3DAAAABD; fma.rn.f32 %f737, %f735, %f732, %f736; mul.rn.f32 %f738, %f737, %f732; mul.rn.f32 %f739, %f738, %f731; sub.f32 %f740, %f727, %f731; add.f32 %f741, %f740, %f740; neg.f32 %f742, %f731; fma.rn.f32 %f743, %f742, %f727, %f741; mul.rn.f32 %f744, %f729, %f743; add.f32 %f745, %f739, %f731; sub.f32 %f746, %f731, %f745; add.f32 %f747, %f739, %f746; add.f32 %f748, %f744, %f747; add.f32 %f749, %f745, %f748; sub.f32 %f750, %f745, %f749; add.f32 %f751, %f748, %f750; mov.f32 %f752, 0f3F317200; mul.rn.f32 %f753, %f725, %f752; mov.f32 %f754, 0f35BFBE8E; mul.rn.f32 %f755, %f725, %f754; add.f32 %f756, %f753, %f749; sub.f32 %f757, %f753, %f756; add.f32 %f758, %f749, %f757; add.f32 %f759, %f751, %f758; add.f32 %f760, %f755, %f759; add.f32 %f761, %f756, %f760; sub.f32 %f762, %f756, %f761; add.f32 %f763, %f760, %f762; mov.f32 %f764, 0f3ED55555; mul.rn.f32 %f765, %f764, %f761; neg.f32 %f766, %f765; fma.rn.f32 %f767, %f764, %f761, %f766; fma.rn.f32 %f768, %f764, %f763, %f767; mov.f32 %f769, 0f00000000; fma.rn.f32 %f770, %f769, %f761, %f768; add.rn.f32 %f771, %f765, %f770; neg.f32 %f772, %f771; add.rn.f32 %f773, %f765, %f772; add.rn.f32 %f774, %f773, %f770; mov.b32 %r282, %f771; setp.eq.s32 %p120, %r282, 1118925336; add.s32 %r283, %r282, -1; mov.b32 %f775, %r283; add.f32 %f776, %f774, 0f37000000; selp.f32 %f139, %f776, %f774, %p120; selp.f32 %f777, %f775, %f771, %p120; mov.f32 %f778, 0f3FB8AA3B; mul.rn.f32 %f779, %f777, %f778; cvt.rzi.f32.f32 %f780, %f779; abs.f32 %f781, %f780; setp.gt.f32 %p121, %f781, 0f42FC0000; mov.b32 %r284, %f780; and.b32 %r285, %r284, -2147483648; or.b32 %r286, %r285, 1123811328; mov.b32 %f782, %r286; selp.f32 %f783, %f782, %f780, %p121; mov.f32 %f784, 0fBF317218; fma.rn.f32 %f785, %f783, %f784, %f777; mov.f32 %f786, 0f3102E308; fma.rn.f32 %f787, %f783, %f786, %f785; mul.f32 %f788, %f787, 0f3FB8AA3B; add.f32 %f789, %f783, 0f4B40007F; mov.b32 %r287, %f789; shl.b32 %r288, %r287, 23; mov.b32 %f790, %r288; ex2.approx.ftz.f32 %f791, %f788; mul.f32 %f140, %f791, %f790; setp.eq.f32 %p122, %f140, 0f7F800000; mov.f32 %f917, 0f7F800000; @%p122 bra $L__BB0_99; fma.rn.f32 %f917, %f140, %f139, %f140; $L__BB0_99: setp.lt.f32 %p123, %f114, 0f00000000; and.pred %p6, %p123, %p96; setp.eq.f32 %p125, %f114, 0f00000000; @%p125 bra $L__BB0_103; bra.uni $L__BB0_100; $L__BB0_103: add.f32 %f796, %f114, %f114; selp.f32 %f919, %f796, 0f00000000, %p96; bra.uni $L__BB0_104; $L__BB0_100: mov.b32 %r289, %f917; xor.b32 %r290, %r289, -2147483648; mov.b32 %f792, %r290; selp.f32 %f919, %f792, %f917, %p6; setp.geu.f32 %p126, %f114, 0f00000000; @%p126 bra $L__BB0_104; mov.f32 %f793, 0f3ED55555; cvt.rzi.f32.f32 %f794, %f793; setp.eq.f32 %p127, %f794, 0f3ED55555; @%p127 bra $L__BB0_104; mov.f32 %f919, 0f7FFFFFFF; $L__BB0_104: add.f32 %f797, %f138, 0f3ED55555; mov.b32 %r291, %f797; setp.lt.s32 %p129, %r291, 2139095040; @%p129 bra $L__BB0_109; setp.gtu.f32 %p130, %f138, 0f7F800000; @%p130 bra $L__BB0_108; bra.uni $L__BB0_106; $L__BB0_108: add.f32 %f919, %f114, 0f3ED55555; bra.uni $L__BB0_109; $L__BB0_106: setp.neu.f32 %p131, %f138, 0f7F800000; @%p131 bra $L__BB0_109; selp.f32 %f919, 0fFF800000, 0f7F800000, %p6; $L__BB0_109: ld.const.u64 %rd79, [params+144]; cvta.to.global.u64 %rd78, %rd79; ld.const.u32 %r319, [params+136]; mad.lo.s32 %r318, %r319, %r5, %r4; cvt.u64.u32 %rd77, %r318; fma.rn.f32 %f798, %f913, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p132, %f112, 0f3F800000; mov.f32 %f799, 0f3F800000; selp.f32 %f800, 0f3F7FFFFF, %f798, %p132; mul.f32 %f801, %f112, 0f414EB852; setp.lt.f32 %p133, %f112, 0f3B4D2E1C; selp.f32 %f802, %f801, %f800, %p133; fma.rn.f32 %f803, %f916, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p134, %f113, 0f3F800000; selp.f32 %f804, 0f3F7FFFFF, %f803, %p134; mul.f32 %f805, %f113, 0f414EB852; setp.lt.f32 %p135, %f113, 0f3B4D2E1C; selp.f32 %f806, %f805, %f804, %p135; fma.rn.f32 %f807, %f919, 0f3F870A3D, 0fBD6147AE; setp.eq.f32 %p136, %f114, 0f3F800000; selp.f32 %f808, 0f3F7FFFFF, %f807, %p136; mul.f32 %f809, %f114, 0f414EB852; setp.lt.f32 %p137, %f114, 0f3B4D2E1C; selp.f32 %f810, %f809, %f808, %p137; min.f32 %f811, %f802, %f799; mov.f32 %f812, 0f00000000; max.f32 %f813, %f812, %f811; mul.f32 %f814, %f813, 0f43800000; cvt.rzi.u32.f32 %r292, %f814; min.u32 %r293, %r292, 255; min.f32 %f815, %f806, %f799; max.f32 %f816, %f812, %f815; mul.f32 %f817, %f816, 0f43800000; cvt.rzi.u32.f32 %r294, %f817; min.u32 %r295, %r294, 255; min.f32 %f818, %f810, %f799; max.f32 %f819, %f812, %f818; mul.f32 %f820, %f819, 0f43800000; cvt.rzi.u32.f32 %r296, %f820; min.u32 %r297, %r296, 255; shl.b64 %rd57, %rd77, 2; add.s64 %rd58, %rd78, %rd57; cvt.u16.u32 %rs15, %r297; cvt.u16.u32 %rs16, %r295; cvt.u16.u32 %rs17, %r293; mov.u16 %rs18, 255; st.global.v4.u8 [%rd58], {%rs17, %rs16, %rs15, %rs18}; $L__BB0_110: ld.const.u32 %r320, [params+104]; and.b32 %r298, %r320, 4; setp.eq.s32 %p138, %r298, 0; @%p138 bra $L__BB0_114; ld.const.u32 %r299, [params+108]; setp.eq.s32 %p139, %r299, 0; ld.const.u64 %rd59, [params+224]; cvta.to.global.u64 %rd60, %rd59; ld.const.u32 %r300, [params+216]; mad.lo.s32 %r301, %r300, %r5, %r4; mul.wide.u32 %rd61, %r301, 8; add.s64 %rd18, %rd60, %rd61; @%p139 bra $L__BB0_113; ld.global.v4.u16 {%rs26, %rs27, %rs28, %rs29}, [%rd18]; // begin inline asm { cvt.f32.f16 %f821, %rs26;} // end inline asm // begin inline asm { cvt.f32.f16 %f822, %rs27;} // end inline asm // begin inline asm { cvt.f32.f16 %f823, %rs28;} // end inline asm add.f32 %f824, %f73, %f821; add.f32 %f825, %f74, %f822; add.f32 %f826, %f75, %f823; mov.f32 %f827, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs24, %f826;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs23, %f825;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs22, %f824;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs25, %f827;} // end inline asm st.global.v4.u16 [%rd18], {%rs22, %rs23, %rs24, %rs25}; bra.uni $L__BB0_114; $L__BB0_113: mov.f32 %f831, 0f3F800000; // begin inline asm { cvt.rn.f16.f32 %rs33, %f831;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs32, %f75;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs31, %f74;} // end inline asm // begin inline asm { cvt.rn.f16.f32 %rs30, %f73;} // end inline asm st.global.v4.u16 [%rd18], {%rs30, %rs31, %rs32, %rs33}; $L__BB0_114: ld.const.u32 %r321, [params+104]; and.b32 %r302, %r321, 64; setp.eq.s32 %p140, %r302, 0; @%p140 bra $L__BB0_124; mul.f32 %f832, %f895, %f895; fma.rn.f32 %f833, %f894, %f894, %f832; fma.rn.f32 %f834, %f896, %f896, %f833; sqrt.rn.f32 %f835, %f834; rcp.rn.f32 %f836, %f835; mul.f32 %f837, %f894, %f836; mul.f32 %f838, %f895, %f836; mul.f32 %f839, %f896, %f836; ld.const.u64 %rd62, [params+208]; cvta.to.global.u64 %rd63, %rd62; ld.const.u32 %r303, [params+200]; mad.lo.s32 %r304, %r303, %r5, %r4; fma.rn.f32 %f840, %f837, 0f3F000000, 0f3F000000; mul.f32 %f841, %f840, 0f437F0000; cvt.rzi.u32.f32 %r305, %f841; fma.rn.f32 %f842, %f838, 0f3F000000, 0f3F000000; mul.f32 %f843, %f842, 0f437F0000; cvt.rzi.u32.f32 %r306, %f843; fma.rn.f32 %f844, %f839, 0f3F000000, 0f3F000000; mul.f32 %f845, %f844, 0f437F0000; cvt.rzi.u32.f32 %r307, %f845; mul.wide.u32 %rd64, %r304, 4; add.s64 %rd65, %rd63, %rd64; cvt.u16.u32 %rs34, %r307; cvt.u16.u32 %rs35, %r306; cvt.u16.u32 %rs36, %r305; mov.u16 %rs37, 255; st.global.v4.u8 [%rd65], {%rs36, %rs35, %rs34, %rs37}; $L__BB0_124: ret; }