// // Generated by NVIDIA NVVM Compiler // // Compiler Build ID: CL-23083092 // Cuda compilation tools, release 9.1, V9.1.85 // Based on LLVM 3.4svn // .version 6.1 .target sm_30 .address_size 64 // .globl _Z6oxMainv .global .align 8 .b8 pixelID[8]; .global .align 8 .b8 resolution[8]; .global .align 4 .b8 normal[12]; .global .align 4 .b8 camPos[12]; .global .align 4 .b8 root[4]; .global .align 4 .u32 imageEnabled; .global .texref lightmap; .global .align 16 .b8 tileInfo[16]; .global .align 4 .u32 additive; .global .align 1 .b8 image[1]; .global .align 1 .b8 image_HDR[1]; .global .align 1 .b8 image_HDR2[1]; .global .align 1 .b8 image_RNM0[1]; .global .align 1 .b8 image_RNM1[1]; .global .align 1 .b8 image_RNM2[1]; .global .align 1 .b8 image_RNM3[1]; .global .align 1 .b8 image_RNM4[1]; .global .align 1 .b8 image_RNM5[1]; .global .align 1 .b8 image_RNM6[1]; .global .align 1 .b8 image_RNM7[1]; .global .align 1 .b8 image_RNM8[1]; .global .align 1 .b8 uvpos[1]; .global .align 1 .b8 uvnormal[1]; .global .align 1 .b8 rnd_seeds[1]; .global .texref sky; .global .align 4 .b8 skyColor[12]; .global .align 4 .u32 samples; .global .align 4 .u32 hemispherical; .global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0}; .global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0}; .global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0}; .global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0}; .global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0}; .global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0}; .global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0}; .global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0}; .global .align 4 .b8 _ZN21rti_internal_typeinfo8skyColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0}; .global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0}; .global .align 4 .b8 _ZN21rti_internal_typeinfo13hemisphericalE[8] = {82, 97, 121, 0, 4, 0, 0, 0}; .global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE; .global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E; .global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E; .global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E; .global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E; .global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E; .global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E; .global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E; .global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E; .global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E; .global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E; .global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E; .global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E; .global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E; .global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E; .global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E; .global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E; .global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E; .global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E; .global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E; .global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E; .global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE; .global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE; .global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE; .global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0}; .global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0}; .global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0}; .global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0}; .global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0}; .global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0}; .global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0}; .global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0}; .global .align 8 .b8 _ZN21rti_internal_typename8skyColorE[7] = {102, 108, 111, 97, 116, 51, 0}; .global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0}; .global .align 4 .b8 _ZN21rti_internal_typename13hemisphericalE[4] = {105, 110, 116, 0}; .global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919; .global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919; .global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919; .global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919; .global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919; .global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919; .global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919; .global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919; .global .align 4 .u32 _ZN21rti_internal_typeenum8skyColorE = 4919; .global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919; .global .align 4 .u32 _ZN21rti_internal_typeenum13hemisphericalE = 4919; .global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0}; .global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0}; .global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0}; .global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1]; .global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1]; .global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1]; .global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1]; .global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1]; .global .align 1 .b8 _ZN21rti_internal_semantic8skyColorE[1]; .global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1]; .global .align 1 .b8 _ZN21rti_internal_semantic13hemisphericalE[1]; .global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1]; .global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1]; .global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1]; .global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1]; .global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1]; .global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1]; .global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1]; .global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1]; .global .align 1 .b8 _ZN23rti_internal_annotation8skyColorE[1]; .global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1]; .global .align 1 .b8 _ZN23rti_internal_annotation13hemisphericalE[1]; .const .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 _Z6oxMainv( ) { .local .align 4 .b8 __local_depot0[32]; .reg .b64 %SP; .reg .b64 %SPL; .reg .pred %p<116>; .reg .b16 %rs<294>; .reg .f32 %f<977>; .reg .b32 %r<451>; .reg .b64 %rd<488>; mov.u64 %rd487, __local_depot0; cvta.local.u64 %SP, %rd487; ld.global.u32 %r1, [samples]; shl.b32 %r2, %r1, 1; ld.global.v2.u32 {%r94, %r95}, [pixelID]; cvt.u64.u32 %rd22, %r94; cvt.u64.u32 %rd23, %r95; mov.u64 %rd26, uvnormal; cvta.global.u64 %rd21, %rd26; mov.u32 %r92, 2; mov.u32 %r93, 4; mov.u64 %rd25, 0; // inline asm call (%rd20), _rt_buffer_get_64, (%rd21, %r92, %r93, %rd22, %rd23, %rd25, %rd25); // inline asm ld.u32 %r3, [%rd20]; shr.u32 %r98, %r3, 16; cvt.u16.u32 %rs1, %r98; and.b16 %rs12, %rs1, 255; cvt.u16.u32 %rs13, %r3; or.b16 %rs14, %rs13, %rs12; setp.eq.s16 %p4, %rs14, 0; mov.f32 %f913, 0f00000000; mov.f32 %f914, %f913; mov.f32 %f915, %f913; @%p4 bra BB0_2; ld.u8 %rs15, [%rd20+1]; and.b16 %rs17, %rs13, 255; cvt.rn.f32.u16 %f193, %rs17; div.rn.f32 %f194, %f193, 0f437F0000; fma.rn.f32 %f195, %f194, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f196, %rs15; div.rn.f32 %f197, %f196, 0f437F0000; fma.rn.f32 %f198, %f197, 0f40000000, 0fBF800000; cvt.rn.f32.u16 %f199, %rs12; div.rn.f32 %f200, %f199, 0f437F0000; fma.rn.f32 %f201, %f200, 0f40000000, 0fBF800000; mul.f32 %f202, %f198, %f198; fma.rn.f32 %f203, %f195, %f195, %f202; fma.rn.f32 %f204, %f201, %f201, %f203; sqrt.rn.f32 %f205, %f204; rcp.rn.f32 %f206, %f205; mul.f32 %f913, %f195, %f206; mul.f32 %f914, %f198, %f206; mul.f32 %f915, %f201, %f206; BB0_2: ld.global.v2.u32 {%r99, %r100}, [pixelID]; ld.global.v2.u32 {%r102, %r103}, [tileInfo]; add.s32 %r4, %r99, %r102; add.s32 %r5, %r100, %r103; setp.eq.f32 %p5, %f914, 0f00000000; setp.eq.f32 %p6, %f913, 0f00000000; and.pred %p7, %p6, %p5; setp.eq.f32 %p8, %f915, 0f00000000; and.pred %p9, %p7, %p8; @%p9 bra BB0_122; bra.uni BB0_3; BB0_122: ld.global.u32 %r450, [imageEnabled]; and.b32 %r325, %r450, 1; setp.eq.b32 %p103, %r325, 1; @!%p103 bra BB0_124; bra.uni BB0_123; BB0_123: cvt.u64.u32 %rd273, %r5; cvt.u64.u32 %rd272, %r4; mov.u64 %rd276, image; cvta.global.u64 %rd271, %rd276; mov.u64 %rd275, 0; // inline asm call (%rd270), _rt_buffer_get_64, (%rd271, %r92, %r93, %rd272, %rd273, %rd275, %rd275); // inline asm mov.u16 %rs163, 0; st.v4.u8 [%rd270], {%rs163, %rs163, %rs163, %rs163}; ld.global.u32 %r450, [imageEnabled]; BB0_124: cvt.u64.u32 %rd18, %r4; cvt.u64.u32 %rd19, %r5; and.b32 %r328, %r450, 4; setp.eq.s32 %p104, %r328, 0; @%p104 bra BB0_128; ld.global.u32 %r329, [additive]; setp.eq.s32 %p105, %r329, 0; @%p105 bra BB0_127; mov.u64 %rd289, image_HDR; cvta.global.u64 %rd278, %rd289; mov.u32 %r333, 8; mov.u64 %rd288, 0; // inline asm call (%rd277), _rt_buffer_get_64, (%rd278, %r92, %r333, %rd18, %rd19, %rd288, %rd288); // inline asm ld.v4.u16 {%rs170, %rs171, %rs172, %rs173}, [%rd277]; // inline asm { cvt.f32.f16 %f786, %rs170;} // inline asm // inline asm { cvt.f32.f16 %f787, %rs171;} // inline asm // inline asm { cvt.f32.f16 %f788, %rs172;} // inline asm // inline asm call (%rd283), _rt_buffer_get_64, (%rd278, %r92, %r333, %rd18, %rd19, %rd288, %rd288); // inline asm add.f32 %f789, %f786, 0f00000000; add.f32 %f790, %f787, 0f00000000; add.f32 %f791, %f788, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs169, %f791;} // inline asm // inline asm { cvt.rn.f16.f32 %rs168, %f790;} // inline asm // inline asm { cvt.rn.f16.f32 %rs167, %f789;} // inline asm mov.u16 %rs174, 0; st.v4.u16 [%rd283], {%rs167, %rs168, %rs169, %rs174}; bra.uni BB0_128; BB0_3: ld.global.v2.u32 {%r111, %r112}, [pixelID]; cvt.u64.u32 %rd29, %r111; cvt.u64.u32 %rd30, %r112; mov.u64 %rd39, uvpos; cvta.global.u64 %rd28, %rd39; mov.u32 %r108, 12; // inline asm call (%rd27), _rt_buffer_get_64, (%rd28, %r92, %r108, %rd29, %rd30, %rd25, %rd25); // inline asm ld.f32 %f217, [%rd27+8]; ld.f32 %f218, [%rd27+4]; ld.f32 %f219, [%rd27]; mul.f32 %f220, %f219, 0f3456BF95; mul.f32 %f221, %f218, 0f3456BF95; mul.f32 %f222, %f217, 0f3456BF95; abs.f32 %f223, %f913; div.rn.f32 %f224, %f220, %f223; abs.f32 %f225, %f914; div.rn.f32 %f226, %f221, %f225; abs.f32 %f227, %f915; div.rn.f32 %f228, %f222, %f227; abs.f32 %f229, %f224; abs.f32 %f230, %f226; abs.f32 %f231, %f228; mov.f32 %f232, 0f38D1B717; max.f32 %f233, %f229, %f232; max.f32 %f234, %f230, %f232; max.f32 %f235, %f231, %f232; fma.rn.f32 %f7, %f913, %f233, %f219; fma.rn.f32 %f8, %f914, %f234, %f218; fma.rn.f32 %f9, %f915, %f235, %f217; ld.global.u32 %r6, [hemispherical]; setp.gt.f32 %p10, %f223, %f227; neg.f32 %f236, %f914; selp.f32 %f237, %f236, 0f00000000, %p10; neg.f32 %f238, %f915; selp.f32 %f239, %f913, %f238, %p10; selp.f32 %f240, 0f00000000, %f914, %p10; mul.f32 %f241, %f239, %f239; fma.rn.f32 %f242, %f237, %f237, %f241; fma.rn.f32 %f243, %f240, %f240, %f242; sqrt.rn.f32 %f244, %f243; rcp.rn.f32 %f245, %f244; mul.f32 %f10, %f237, %f245; mul.f32 %f11, %f239, %f245; mul.f32 %f12, %f240, %f245; ld.global.v2.u32 {%r115, %r116}, [pixelID]; cvt.u64.u32 %rd35, %r115; cvt.u64.u32 %rd36, %r116; mov.u64 %rd40, rnd_seeds; cvta.global.u64 %rd34, %rd40; // inline asm call (%rd33), _rt_buffer_get_64, (%rd34, %r92, %r93, %rd35, %rd36, %rd25, %rd25); // inline asm mov.f32 %f957, 0f00000000; setp.lt.s32 %p11, %r1, 1; mov.f32 %f956, %f957; mov.f32 %f955, %f957; mov.f32 %f954, %f957; mov.f32 %f953, %f957; mov.f32 %f952, %f957; mov.f32 %f951, %f957; mov.f32 %f950, %f957; mov.f32 %f949, %f957; mov.f32 %f948, %f957; @%p11 bra BB0_56; cvt.rn.f32.s32 %f256, %r2; rcp.rn.f32 %f13, %f256; ld.u32 %r428, [%rd33]; mul.f32 %f14, %f7, 0f3456BF95; mul.f32 %f15, %f8, 0f3456BF95; mul.f32 %f16, %f9, 0f3456BF95; mul.f32 %f257, %f913, %f11; mul.f32 %f258, %f914, %f10; sub.f32 %f17, %f258, %f257; mul.f32 %f259, %f915, %f10; mul.f32 %f260, %f913, %f12; sub.f32 %f18, %f260, %f259; mul.f32 %f261, %f914, %f12; mul.f32 %f262, %f915, %f11; sub.f32 %f19, %f262, %f261; mov.f32 %f957, 0f00000000; mov.u32 %r119, 0; abs.f32 %f337, %f15; abs.f32 %f338, %f14; max.f32 %f339, %f338, %f337; abs.f32 %f340, %f16; max.f32 %f341, %f339, %f340; mov.u32 %r425, %r119; mov.f32 %f956, %f957; mov.f32 %f955, %f957; mov.f32 %f954, %f957; mov.f32 %f953, %f957; mov.f32 %f952, %f957; mov.f32 %f951, %f957; mov.f32 %f950, %f957; mov.f32 %f949, %f957; mov.f32 %f948, %f957; BB0_5: mov.u32 %r427, %r119; BB0_6: mov.u32 %r11, %r428; cvt.rn.f32.s32 %f895, %r425; mad.lo.s32 %r121, %r11, 1664525, 1013904223; and.b32 %r122, %r121, 16777215; cvt.rn.f32.u32 %f263, %r122; fma.rn.f32 %f264, %f263, 0f33800000, %f895; mul.f32 %f265, %f13, %f264; mad.lo.s32 %r12, %r121, 1664525, 1013904223; and.b32 %r123, %r12, 16777215; cvt.rn.f32.u32 %f266, %r123; cvt.rn.f32.s32 %f267, %r427; fma.rn.f32 %f268, %f266, 0f33800000, %f267; mul.f32 %f269, %f13, %f268; fma.rn.f32 %f41, %f265, 0fC0000000, 0f3F800000; mul.f32 %f270, %f41, %f41; mov.f32 %f271, 0f3F800000; sub.f32 %f272, %f271, %f270; mov.f32 %f273, 0f00000000; max.f32 %f274, %f273, %f272; sqrt.rn.f32 %f42, %f274; mul.f32 %f942, %f269, 0f40C90FDB; abs.f32 %f44, %f942; setp.neu.f32 %p12, %f44, 0f7F800000; mov.f32 %f936, %f942; @%p12 bra BB0_8; mov.f32 %f896, 0f00000000; mul.rn.f32 %f936, %f942, %f896; BB0_8: mul.f32 %f276, %f936, 0f3F22F983; cvt.rni.s32.f32 %r438, %f276; cvt.rn.f32.s32 %f277, %r438; neg.f32 %f278, %f277; mov.f32 %f279, 0f3FC90FDA; fma.rn.f32 %f280, %f278, %f279, %f936; mov.f32 %f281, 0f33A22168; fma.rn.f32 %f282, %f278, %f281, %f280; mov.f32 %f283, 0f27C234C5; fma.rn.f32 %f937, %f278, %f283, %f282; abs.f32 %f284, %f936; setp.leu.f32 %p13, %f284, 0f47CE4780; @%p13 bra BB0_19; add.u64 %rd42, %SP, 4; cvta.to.local.u64 %rd483, %rd42; mov.u32 %r429, 0; mov.u64 %rd484, 0; mov.u32 %r430, %r429; BB0_10: .pragma "nounroll"; mov.b32 %r406, %f936; shl.b32 %r405, %r406, 8; or.b32 %r404, %r405, -2147483648; add.u64 %rd479, %SP, 4; cvta.to.local.u64 %rd478, %rd479; shl.b64 %rd43, %rd484, 2; mov.u64 %rd44, __cudart_i2opi_f; add.s64 %rd45, %rd44, %rd43; ld.const.u32 %r129, [%rd45]; // inline asm { mad.lo.cc.u32 %r127, %r129, %r404, %r430; madc.hi.u32 %r430, %r129, %r404, 0; } // inline asm st.local.u32 [%rd483], %r127; add.s32 %r429, %r429, 1; cvt.s64.s32 %rd484, %r429; mul.wide.s32 %rd48, %r429, 4; add.s64 %rd483, %rd478, %rd48; setp.ne.s32 %p14, %r429, 6; @%p14 bra BB0_10; mov.b32 %r408, %f936; shr.u32 %r407, %r408, 23; add.u64 %rd477, %SP, 4; and.b32 %r132, %r407, 255; add.s32 %r133, %r132, -128; shr.u32 %r134, %r133, 5; cvta.to.local.u64 %rd50, %rd477; st.local.u32 [%rd50+24], %r430; mov.u32 %r135, 6; sub.s32 %r136, %r135, %r134; mul.wide.s32 %rd51, %r136, 4; add.s64 %rd8, %rd50, %rd51; ld.local.u32 %r431, [%rd8]; ld.local.u32 %r432, [%rd8+-4]; and.b32 %r24, %r407, 31; setp.eq.s32 %p15, %r24, 0; @%p15 bra BB0_13; mov.u32 %r137, 32; sub.s32 %r138, %r137, %r24; shr.u32 %r139, %r432, %r138; shl.b32 %r140, %r431, %r24; add.s32 %r431, %r139, %r140; ld.local.u32 %r141, [%rd8+-8]; shr.u32 %r142, %r141, %r138; shl.b32 %r143, %r432, %r24; add.s32 %r432, %r142, %r143; BB0_13: mov.b32 %r411, %f936; and.b32 %r434, %r411, -2147483648; shr.u32 %r144, %r432, 30; shl.b32 %r145, %r431, 2; add.s32 %r433, %r144, %r145; shl.b32 %r30, %r432, 2; shr.u32 %r146, %r433, 31; shr.u32 %r147, %r431, 30; add.s32 %r31, %r146, %r147; setp.eq.s32 %p16, %r146, 0; @%p16 bra BB0_14; bra.uni BB0_15; BB0_14: mov.u32 %r435, %r30; bra.uni BB0_16; BB0_15: mov.b32 %r413, %f936; and.b32 %r412, %r413, -2147483648; not.b32 %r148, %r433; neg.s32 %r435, %r30; setp.eq.s32 %p17, %r30, 0; selp.u32 %r149, 1, 0, %p17; add.s32 %r433, %r149, %r148; xor.b32 %r434, %r412, -2147483648; BB0_16: mov.b32 %r415, %f936; and.b32 %r414, %r415, -2147483648; clz.b32 %r437, %r433; setp.eq.s32 %p18, %r437, 0; shl.b32 %r150, %r433, %r437; mov.u32 %r151, 32; sub.s32 %r152, %r151, %r437; shr.u32 %r153, %r435, %r152; add.s32 %r154, %r153, %r150; selp.b32 %r39, %r433, %r154, %p18; mov.u32 %r155, -921707870; mul.hi.u32 %r436, %r39, %r155; setp.eq.s32 %p19, %r414, 0; neg.s32 %r156, %r31; selp.b32 %r438, %r31, %r156, %p19; setp.lt.s32 %p20, %r436, 1; @%p20 bra BB0_18; mul.lo.s32 %r157, %r39, -921707870; shr.u32 %r158, %r157, 31; shl.b32 %r159, %r436, 1; add.s32 %r436, %r158, %r159; add.s32 %r437, %r437, 1; BB0_18: mov.u32 %r160, 126; sub.s32 %r161, %r160, %r437; shl.b32 %r162, %r161, 23; add.s32 %r163, %r436, 1; shr.u32 %r164, %r163, 7; add.s32 %r165, %r164, 1; shr.u32 %r166, %r165, 1; add.s32 %r167, %r166, %r162; or.b32 %r168, %r167, %r434; mov.b32 %f937, %r168; BB0_19: add.s32 %r47, %r438, 1; and.b32 %r48, %r47, 1; setp.eq.s32 %p21, %r48, 0; @%p21 bra BB0_21; bra.uni BB0_20; BB0_21: mul.rn.f32 %f904, %f937, %f937; mov.f32 %f287, 0f3C08839E; mov.f32 %f288, 0fB94CA1F9; fma.rn.f32 %f938, %f288, %f904, %f287; bra.uni BB0_22; BB0_20: mul.rn.f32 %f900, %f937, %f937; mov.f32 %f285, 0fBAB6061A; mov.f32 %f286, 0f37CCF5CE; fma.rn.f32 %f938, %f286, %f900, %f285; BB0_22: @%p21 bra BB0_24; bra.uni BB0_23; BB0_24: mul.rn.f32 %f903, %f937, %f937; mov.f32 %f899, 0f00000000; mov.f32 %f292, 0fBE2AAAA3; fma.rn.f32 %f293, %f938, %f903, %f292; fma.rn.f32 %f939, %f293, %f903, %f899; bra.uni BB0_25; BB0_23: mul.rn.f32 %f901, %f937, %f937; mov.f32 %f289, 0f3D2AAAA5; fma.rn.f32 %f290, %f938, %f901, %f289; mov.f32 %f291, 0fBF000000; fma.rn.f32 %f939, %f290, %f901, %f291; BB0_25: fma.rn.f32 %f940, %f939, %f937, %f937; @%p21 bra BB0_27; mul.rn.f32 %f902, %f937, %f937; mov.f32 %f887, 0f3F800000; fma.rn.f32 %f940, %f939, %f902, %f887; BB0_27: and.b32 %r169, %r47, 2; setp.eq.s32 %p24, %r169, 0; @%p24 bra BB0_29; mov.f32 %f888, 0f00000000; mov.f32 %f297, 0fBF800000; fma.rn.f32 %f940, %f940, %f297, %f888; BB0_29: abs.f32 %f889, %f942; setp.neu.f32 %p115, %f889, 0f7F800000; @%p115 bra BB0_31; mov.f32 %f898, 0f00000000; mul.rn.f32 %f942, %f942, %f898; BB0_31: mov.f32 %f892, 0f27C234C5; mov.f32 %f891, 0f33A22168; mov.f32 %f890, 0f3FC90FDA; mul.f32 %f299, %f942, 0f3F22F983; cvt.rni.s32.f32 %r448, %f299; cvt.rn.f32.s32 %f300, %r448; neg.f32 %f301, %f300; fma.rn.f32 %f303, %f301, %f890, %f942; fma.rn.f32 %f305, %f301, %f891, %f303; fma.rn.f32 %f943, %f301, %f892, %f305; abs.f32 %f307, %f942; setp.leu.f32 %p26, %f307, 0f47CE4780; @%p26 bra BB0_42; add.u64 %rd53, %SP, 4; cvta.to.local.u64 %rd485, %rd53; mov.b32 %r50, %f942; shl.b32 %r172, %r50, 8; or.b32 %r52, %r172, -2147483648; mov.u32 %r439, 0; mov.u64 %rd486, %rd25; mov.u32 %r440, %r439; BB0_33: .pragma "nounroll"; add.u64 %rd482, %SP, 4; cvta.to.local.u64 %rd481, %rd482; shl.b64 %rd54, %rd486, 2; mov.u64 %rd55, __cudart_i2opi_f; add.s64 %rd56, %rd55, %rd54; ld.const.u32 %r175, [%rd56]; // inline asm { mad.lo.cc.u32 %r173, %r175, %r52, %r440; madc.hi.u32 %r440, %r175, %r52, 0; } // inline asm st.local.u32 [%rd485], %r173; add.s32 %r439, %r439, 1; cvt.s64.s32 %rd486, %r439; mul.wide.s32 %rd57, %r439, 4; add.s64 %rd485, %rd481, %rd57; setp.ne.s32 %p27, %r439, 6; @%p27 bra BB0_33; mov.b32 %r417, %f942; shr.u32 %r416, %r417, 23; add.u64 %rd480, %SP, 4; and.b32 %r178, %r416, 255; add.s32 %r179, %r178, -128; shr.u32 %r180, %r179, 5; cvta.to.local.u64 %rd59, %rd480; st.local.u32 [%rd59+24], %r440; mov.u32 %r181, 6; sub.s32 %r182, %r181, %r180; mul.wide.s32 %rd60, %r182, 4; add.s64 %rd15, %rd59, %rd60; ld.local.u32 %r441, [%rd15]; ld.local.u32 %r442, [%rd15+-4]; and.b32 %r60, %r416, 31; setp.eq.s32 %p28, %r60, 0; @%p28 bra BB0_36; mov.u32 %r183, 32; sub.s32 %r184, %r183, %r60; shr.u32 %r185, %r442, %r184; shl.b32 %r186, %r441, %r60; add.s32 %r441, %r185, %r186; ld.local.u32 %r187, [%rd15+-8]; shr.u32 %r188, %r187, %r184; shl.b32 %r189, %r442, %r60; add.s32 %r442, %r188, %r189; BB0_36: mov.b32 %r420, %f942; and.b32 %r444, %r420, -2147483648; shr.u32 %r190, %r442, 30; shl.b32 %r191, %r441, 2; add.s32 %r443, %r190, %r191; shl.b32 %r66, %r442, 2; shr.u32 %r192, %r443, 31; shr.u32 %r193, %r441, 30; add.s32 %r67, %r192, %r193; setp.eq.s32 %p29, %r192, 0; @%p29 bra BB0_37; bra.uni BB0_38; BB0_37: mov.u32 %r445, %r66; bra.uni BB0_39; BB0_38: mov.b32 %r422, %f942; and.b32 %r421, %r422, -2147483648; not.b32 %r194, %r443; neg.s32 %r445, %r66; setp.eq.s32 %p30, %r66, 0; selp.u32 %r195, 1, 0, %p30; add.s32 %r443, %r195, %r194; xor.b32 %r444, %r421, -2147483648; BB0_39: mov.b32 %r424, %f942; and.b32 %r423, %r424, -2147483648; clz.b32 %r447, %r443; setp.eq.s32 %p31, %r447, 0; shl.b32 %r196, %r443, %r447; mov.u32 %r197, 32; sub.s32 %r198, %r197, %r447; shr.u32 %r199, %r445, %r198; add.s32 %r200, %r199, %r196; selp.b32 %r75, %r443, %r200, %p31; mov.u32 %r201, -921707870; mul.hi.u32 %r446, %r75, %r201; setp.eq.s32 %p32, %r423, 0; neg.s32 %r202, %r67; selp.b32 %r448, %r67, %r202, %p32; setp.lt.s32 %p33, %r446, 1; @%p33 bra BB0_41; mul.lo.s32 %r203, %r75, -921707870; shr.u32 %r204, %r203, 31; shl.b32 %r205, %r446, 1; add.s32 %r446, %r204, %r205; add.s32 %r447, %r447, 1; BB0_41: mov.u32 %r206, 126; sub.s32 %r207, %r206, %r447; shl.b32 %r208, %r207, 23; add.s32 %r209, %r446, 1; shr.u32 %r210, %r209, 7; add.s32 %r211, %r210, 1; shr.u32 %r212, %r211, 1; add.s32 %r213, %r212, %r208; or.b32 %r214, %r213, %r444; mov.b32 %f943, %r214; BB0_42: and.b32 %r83, %r448, 1; setp.eq.s32 %p34, %r83, 0; @%p34 bra BB0_44; bra.uni BB0_43; BB0_44: mul.rn.f32 %f912, %f943, %f943; mov.f32 %f310, 0f3C08839E; mov.f32 %f311, 0fB94CA1F9; fma.rn.f32 %f944, %f311, %f912, %f310; bra.uni BB0_45; BB0_43: mul.rn.f32 %f909, %f943, %f943; mov.f32 %f308, 0fBAB6061A; mov.f32 %f309, 0f37CCF5CE; fma.rn.f32 %f944, %f309, %f909, %f308; BB0_45: @%p34 bra BB0_47; bra.uni BB0_46; BB0_47: mul.rn.f32 %f911, %f943, %f943; mov.f32 %f897, 0f00000000; mov.f32 %f315, 0fBE2AAAA3; fma.rn.f32 %f316, %f944, %f911, %f315; fma.rn.f32 %f945, %f316, %f911, %f897; bra.uni BB0_48; BB0_46: mul.rn.f32 %f910, %f943, %f943; mov.f32 %f312, 0f3D2AAAA5; fma.rn.f32 %f313, %f944, %f910, %f312; mov.f32 %f314, 0fBF000000; fma.rn.f32 %f945, %f313, %f910, %f314; BB0_48: fma.rn.f32 %f946, %f945, %f943, %f943; @%p34 bra BB0_50; mul.rn.f32 %f905, %f943, %f943; mov.f32 %f893, 0f3F800000; fma.rn.f32 %f946, %f945, %f905, %f893; BB0_50: and.b32 %r215, %r448, 2; setp.eq.s32 %p37, %r215, 0; @%p37 bra BB0_52; mov.f32 %f894, 0f00000000; mov.f32 %f320, 0fBF800000; fma.rn.f32 %f946, %f946, %f320, %f894; BB0_52: mul.f32 %f321, %f42, %f940; mul.f32 %f322, %f42, %f946; mul.f32 %f323, %f10, %f322; mul.f32 %f324, %f11, %f322; mul.f32 %f325, %f12, %f322; fma.rn.f32 %f326, %f19, %f321, %f323; fma.rn.f32 %f327, %f18, %f321, %f324; fma.rn.f32 %f328, %f17, %f321, %f325; fma.rn.f32 %f79, %f913, %f41, %f326; fma.rn.f32 %f80, %f914, %f41, %f327; fma.rn.f32 %f81, %f915, %f41, %f328; setp.gt.f32 %p38, %f80, 0f00000000; setp.eq.s32 %p39, %r6, 0; or.pred %p40, %p39, %p38; @!%p40 bra BB0_54; bra.uni BB0_53; BB0_53: add.u64 %rd61, %SP, 0; cvta.to.local.u64 %rd62, %rd61; max.f32 %f335, %f341, %f232; mov.u32 %r219, 1065353216; st.local.u32 [%rd62], %r219; ld.global.u32 %r216, [root]; mov.u32 %r217, 1; mov.f32 %f336, 0f6C4ECB8F; // inline asm call _rt_trace_64, (%r216, %f7, %f8, %f9, %f79, %f80, %f81, %r217, %f335, %f336, %rd61, %r93); // inline asm mul.f32 %f343, %f914, %f80; fma.rn.f32 %f344, %f913, %f79, %f343; fma.rn.f32 %f345, %f915, %f81, %f344; ld.local.f32 %f346, [%rd62]; add.f32 %f949, %f949, %f346; fma.rn.f32 %f950, %f80, %f346, %f950; fma.rn.f32 %f951, %f81, %f346, %f951; fma.rn.f32 %f952, %f79, %f346, %f952; mul.f32 %f347, %f79, %f80; fma.rn.f32 %f953, %f347, %f346, %f953; mul.f32 %f348, %f80, %f81; fma.rn.f32 %f954, %f348, %f346, %f954; fma.rn.f32 %f349, %f81, %f81, 0fBEAAAAAB; fma.rn.f32 %f955, %f349, %f346, %f955; mul.f32 %f350, %f79, %f81; fma.rn.f32 %f956, %f350, %f346, %f956; mul.f32 %f351, %f80, %f80; mul.f32 %f352, %f79, %f79; sub.f32 %f353, %f352, %f351; fma.rn.f32 %f957, %f353, %f346, %f957; cvt.sat.f32.f32 %f354, %f345; fma.rn.f32 %f948, %f354, %f346, %f948; BB0_54: mad.lo.s32 %r401, %r11, 1664525, 1013904223; mad.lo.s32 %r428, %r401, 1664525, 1013904223; add.s32 %r427, %r427, 1; setp.lt.s32 %p41, %r427, %r2; @%p41 bra BB0_6; mad.lo.s32 %r403, %r11, 1664525, 1013904223; mad.lo.s32 %r428, %r403, 1664525, 1013904223; add.s32 %r425, %r425, 1; setp.lt.s32 %p42, %r425, %r2; @%p42 bra BB0_5; BB0_56: mul.lo.s32 %r220, %r2, %r2; cvt.rn.f32.s32 %f355, %r220; div.rn.f32 %f356, %f948, %f355; div.rn.f32 %f112, %f949, %f355; div.rn.f32 %f113, %f950, %f355; div.rn.f32 %f114, %f951, %f355; div.rn.f32 %f115, %f952, %f355; div.rn.f32 %f116, %f953, %f355; div.rn.f32 %f117, %f954, %f355; div.rn.f32 %f118, %f955, %f355; div.rn.f32 %f119, %f956, %f355; div.rn.f32 %f120, %f957, %f355; add.f32 %f357, %f356, %f356; ld.global.f32 %f358, [skyColor]; mul.f32 %f121, %f357, %f358; ld.global.f32 %f359, [skyColor+4]; mul.f32 %f122, %f357, %f359; ld.global.f32 %f360, [skyColor+8]; mul.f32 %f123, %f357, %f360; ld.global.u32 %r449, [imageEnabled]; and.b32 %r221, %r449, 1; setp.eq.b32 %p43, %r221, 1; @!%p43 bra BB0_91; bra.uni BB0_57; BB0_57: abs.f32 %f125, %f121; setp.lt.f32 %p44, %f125, 0f00800000; mul.f32 %f366, %f125, 0f4B800000; selp.f32 %f367, 0fC3170000, 0fC2FE0000, %p44; selp.f32 %f368, %f366, %f125, %p44; mov.b32 %r222, %f368; and.b32 %r223, %r222, 8388607; or.b32 %r224, %r223, 1065353216; mov.b32 %f369, %r224; shr.u32 %r225, %r222, 23; cvt.rn.f32.u32 %f370, %r225; add.f32 %f371, %f367, %f370; setp.gt.f32 %p45, %f369, 0f3FB504F3; mul.f32 %f372, %f369, 0f3F000000; add.f32 %f373, %f371, 0f3F800000; selp.f32 %f374, %f372, %f369, %p45; selp.f32 %f375, %f373, %f371, %p45; add.f32 %f376, %f374, 0fBF800000; add.f32 %f362, %f374, 0f3F800000; // inline asm rcp.approx.ftz.f32 %f361,%f362; // inline asm add.f32 %f377, %f376, %f376; mul.f32 %f378, %f361, %f377; mul.f32 %f379, %f378, %f378; mov.f32 %f380, 0f3C4CAF63; mov.f32 %f381, 0f3B18F0FE; fma.rn.f32 %f382, %f381, %f379, %f380; mov.f32 %f383, 0f3DAAAABD; fma.rn.f32 %f384, %f382, %f379, %f383; mul.rn.f32 %f385, %f384, %f379; mul.rn.f32 %f386, %f385, %f378; sub.f32 %f387, %f376, %f378; neg.f32 %f388, %f378; add.f32 %f389, %f387, %f387; fma.rn.f32 %f390, %f388, %f376, %f389; mul.rn.f32 %f391, %f361, %f390; add.f32 %f392, %f386, %f378; sub.f32 %f393, %f378, %f392; add.f32 %f394, %f386, %f393; add.f32 %f395, %f391, %f394; add.f32 %f396, %f392, %f395; sub.f32 %f397, %f392, %f396; add.f32 %f398, %f395, %f397; mov.f32 %f399, 0f3F317200; mul.rn.f32 %f400, %f375, %f399; mov.f32 %f401, 0f35BFBE8E; mul.rn.f32 %f402, %f375, %f401; add.f32 %f403, %f400, %f396; sub.f32 %f404, %f400, %f403; add.f32 %f405, %f396, %f404; add.f32 %f406, %f398, %f405; add.f32 %f407, %f402, %f406; add.f32 %f408, %f403, %f407; sub.f32 %f409, %f403, %f408; add.f32 %f410, %f407, %f409; mov.f32 %f411, 0f3EE66666; mul.rn.f32 %f412, %f411, %f408; neg.f32 %f413, %f412; fma.rn.f32 %f414, %f411, %f408, %f413; fma.rn.f32 %f415, %f411, %f410, %f414; mov.f32 %f416, 0f00000000; fma.rn.f32 %f417, %f416, %f408, %f415; add.rn.f32 %f418, %f412, %f417; neg.f32 %f419, %f418; add.rn.f32 %f420, %f412, %f419; add.rn.f32 %f421, %f420, %f417; mov.b32 %r226, %f418; setp.eq.s32 %p46, %r226, 1118925336; add.s32 %r227, %r226, -1; mov.b32 %f422, %r227; add.f32 %f423, %f421, 0f37000000; selp.f32 %f424, %f422, %f418, %p46; selp.f32 %f126, %f423, %f421, %p46; mul.f32 %f425, %f424, 0f3FB8AA3B; cvt.rzi.f32.f32 %f426, %f425; mov.f32 %f427, 0fBF317200; fma.rn.f32 %f428, %f426, %f427, %f424; mov.f32 %f429, 0fB5BFBE8E; fma.rn.f32 %f430, %f426, %f429, %f428; mul.f32 %f431, %f430, 0f3FB8AA3B; ex2.approx.ftz.f32 %f432, %f431; add.f32 %f433, %f426, 0f00000000; ex2.approx.f32 %f434, %f433; mul.f32 %f435, %f432, %f434; setp.lt.f32 %p47, %f424, 0fC2D20000; selp.f32 %f436, 0f00000000, %f435, %p47; setp.gt.f32 %p48, %f424, 0f42D20000; selp.f32 %f968, 0f7F800000, %f436, %p48; setp.eq.f32 %p49, %f968, 0f7F800000; @%p49 bra BB0_59; fma.rn.f32 %f968, %f968, %f126, %f968; BB0_59: mov.f32 %f859, 0f3E666666; cvt.rzi.f32.f32 %f858, %f859; fma.rn.f32 %f857, %f858, 0fC0000000, 0f3EE66666; abs.f32 %f856, %f857; setp.lt.f32 %p50, %f121, 0f00000000; setp.eq.f32 %p51, %f856, 0f3F800000; and.pred %p1, %p50, %p51; mov.b32 %r228, %f968; xor.b32 %r229, %r228, -2147483648; mov.b32 %f437, %r229; selp.f32 %f970, %f437, %f968, %p1; setp.eq.f32 %p52, %f121, 0f00000000; @%p52 bra BB0_62; bra.uni BB0_60; BB0_62: add.f32 %f440, %f121, %f121; selp.f32 %f970, %f440, 0f00000000, %p51; bra.uni BB0_63; BB0_127: mov.u64 %rd296, image_HDR; cvta.global.u64 %rd291, %rd296; mov.u32 %r335, 8; mov.u64 %rd295, 0; // inline asm call (%rd290), _rt_buffer_get_64, (%rd291, %r92, %r335, %rd18, %rd19, %rd295, %rd295); // inline asm mov.f32 %f792, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs175, %f792;} // inline asm mov.u16 %rs176, 0; st.v4.u16 [%rd290], {%rs175, %rs175, %rs175, %rs176}; BB0_128: ld.global.u32 %r336, [additive]; setp.eq.s32 %p106, %r336, 0; @%p106 bra BB0_130; mov.u64 %rd309, image_RNM0; cvta.global.u64 %rd298, %rd309; mov.u32 %r340, 8; mov.u64 %rd308, 0; // inline asm call (%rd297), _rt_buffer_get_64, (%rd298, %r92, %r340, %rd18, %rd19, %rd308, %rd308); // inline asm ld.v4.u16 {%rs183, %rs184, %rs185, %rs186}, [%rd297]; // inline asm { cvt.f32.f16 %f793, %rs183;} // inline asm // inline asm { cvt.f32.f16 %f794, %rs184;} // inline asm // inline asm { cvt.f32.f16 %f795, %rs185;} // inline asm // inline asm call (%rd303), _rt_buffer_get_64, (%rd298, %r92, %r340, %rd18, %rd19, %rd308, %rd308); // inline asm add.f32 %f796, %f793, 0f00000000; add.f32 %f797, %f794, 0f00000000; add.f32 %f798, %f795, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs182, %f798;} // inline asm // inline asm { cvt.rn.f16.f32 %rs181, %f797;} // inline asm // inline asm { cvt.rn.f16.f32 %rs180, %f796;} // inline asm mov.u16 %rs187, 0; st.v4.u16 [%rd303], {%rs180, %rs181, %rs182, %rs187}; bra.uni BB0_131; BB0_130: mov.u64 %rd316, image_RNM0; cvta.global.u64 %rd311, %rd316; mov.u32 %r342, 8; mov.u64 %rd315, 0; // inline asm call (%rd310), _rt_buffer_get_64, (%rd311, %r92, %r342, %rd18, %rd19, %rd315, %rd315); // inline asm mov.f32 %f799, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs188, %f799;} // inline asm mov.u16 %rs189, 0; st.v4.u16 [%rd310], {%rs188, %rs188, %rs188, %rs189}; BB0_131: ld.global.u32 %r343, [additive]; setp.eq.s32 %p107, %r343, 0; @%p107 bra BB0_133; mov.u64 %rd329, image_RNM1; cvta.global.u64 %rd318, %rd329; mov.u32 %r347, 8; mov.u64 %rd328, 0; // inline asm call (%rd317), _rt_buffer_get_64, (%rd318, %r92, %r347, %rd18, %rd19, %rd328, %rd328); // inline asm ld.v4.u16 {%rs196, %rs197, %rs198, %rs199}, [%rd317]; // inline asm { cvt.f32.f16 %f800, %rs196;} // inline asm // inline asm { cvt.f32.f16 %f801, %rs197;} // inline asm // inline asm { cvt.f32.f16 %f802, %rs198;} // inline asm // inline asm call (%rd323), _rt_buffer_get_64, (%rd318, %r92, %r347, %rd18, %rd19, %rd328, %rd328); // inline asm add.f32 %f803, %f800, 0f00000000; add.f32 %f804, %f801, 0f00000000; add.f32 %f805, %f802, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs195, %f805;} // inline asm // inline asm { cvt.rn.f16.f32 %rs194, %f804;} // inline asm // inline asm { cvt.rn.f16.f32 %rs193, %f803;} // inline asm mov.u16 %rs200, 0; st.v4.u16 [%rd323], {%rs193, %rs194, %rs195, %rs200}; bra.uni BB0_134; BB0_133: mov.u64 %rd336, image_RNM1; cvta.global.u64 %rd331, %rd336; mov.u32 %r349, 8; mov.u64 %rd335, 0; // inline asm call (%rd330), _rt_buffer_get_64, (%rd331, %r92, %r349, %rd18, %rd19, %rd335, %rd335); // inline asm mov.f32 %f806, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs201, %f806;} // inline asm mov.u16 %rs202, 0; st.v4.u16 [%rd330], {%rs201, %rs201, %rs201, %rs202}; BB0_134: ld.global.u32 %r350, [additive]; setp.eq.s32 %p108, %r350, 0; @%p108 bra BB0_136; mov.u64 %rd349, image_RNM2; cvta.global.u64 %rd338, %rd349; mov.u32 %r354, 8; mov.u64 %rd348, 0; // inline asm call (%rd337), _rt_buffer_get_64, (%rd338, %r92, %r354, %rd18, %rd19, %rd348, %rd348); // inline asm ld.v4.u16 {%rs209, %rs210, %rs211, %rs212}, [%rd337]; // inline asm { cvt.f32.f16 %f807, %rs209;} // inline asm // inline asm { cvt.f32.f16 %f808, %rs210;} // inline asm // inline asm { cvt.f32.f16 %f809, %rs211;} // inline asm // inline asm call (%rd343), _rt_buffer_get_64, (%rd338, %r92, %r354, %rd18, %rd19, %rd348, %rd348); // inline asm add.f32 %f810, %f807, 0f00000000; add.f32 %f811, %f808, 0f00000000; add.f32 %f812, %f809, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs208, %f812;} // inline asm // inline asm { cvt.rn.f16.f32 %rs207, %f811;} // inline asm // inline asm { cvt.rn.f16.f32 %rs206, %f810;} // inline asm mov.u16 %rs213, 0; st.v4.u16 [%rd343], {%rs206, %rs207, %rs208, %rs213}; bra.uni BB0_137; BB0_136: mov.u64 %rd356, image_RNM2; cvta.global.u64 %rd351, %rd356; mov.u32 %r356, 8; mov.u64 %rd355, 0; // inline asm call (%rd350), _rt_buffer_get_64, (%rd351, %r92, %r356, %rd18, %rd19, %rd355, %rd355); // inline asm mov.f32 %f813, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs214, %f813;} // inline asm mov.u16 %rs215, 0; st.v4.u16 [%rd350], {%rs214, %rs214, %rs214, %rs215}; BB0_137: ld.global.u32 %r357, [additive]; setp.eq.s32 %p109, %r357, 0; @%p109 bra BB0_139; mov.u64 %rd369, image_RNM3; cvta.global.u64 %rd358, %rd369; mov.u32 %r361, 8; mov.u64 %rd368, 0; // inline asm call (%rd357), _rt_buffer_get_64, (%rd358, %r92, %r361, %rd18, %rd19, %rd368, %rd368); // inline asm ld.v4.u16 {%rs222, %rs223, %rs224, %rs225}, [%rd357]; // inline asm { cvt.f32.f16 %f814, %rs222;} // inline asm // inline asm { cvt.f32.f16 %f815, %rs223;} // inline asm // inline asm { cvt.f32.f16 %f816, %rs224;} // inline asm // inline asm call (%rd363), _rt_buffer_get_64, (%rd358, %r92, %r361, %rd18, %rd19, %rd368, %rd368); // inline asm add.f32 %f817, %f814, 0f00000000; add.f32 %f818, %f815, 0f00000000; add.f32 %f819, %f816, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs221, %f819;} // inline asm // inline asm { cvt.rn.f16.f32 %rs220, %f818;} // inline asm // inline asm { cvt.rn.f16.f32 %rs219, %f817;} // inline asm mov.u16 %rs226, 0; st.v4.u16 [%rd363], {%rs219, %rs220, %rs221, %rs226}; bra.uni BB0_140; BB0_139: mov.u64 %rd376, image_RNM3; cvta.global.u64 %rd371, %rd376; mov.u32 %r363, 8; mov.u64 %rd375, 0; // inline asm call (%rd370), _rt_buffer_get_64, (%rd371, %r92, %r363, %rd18, %rd19, %rd375, %rd375); // inline asm mov.f32 %f820, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs227, %f820;} // inline asm mov.u16 %rs228, 0; st.v4.u16 [%rd370], {%rs227, %rs227, %rs227, %rs228}; BB0_140: ld.global.u32 %r364, [additive]; setp.eq.s32 %p110, %r364, 0; @%p110 bra BB0_142; mov.u64 %rd389, image_RNM4; cvta.global.u64 %rd378, %rd389; mov.u32 %r368, 8; mov.u64 %rd388, 0; // inline asm call (%rd377), _rt_buffer_get_64, (%rd378, %r92, %r368, %rd18, %rd19, %rd388, %rd388); // inline asm ld.v4.u16 {%rs235, %rs236, %rs237, %rs238}, [%rd377]; // inline asm { cvt.f32.f16 %f821, %rs235;} // inline asm // inline asm { cvt.f32.f16 %f822, %rs236;} // inline asm // inline asm { cvt.f32.f16 %f823, %rs237;} // inline asm // inline asm call (%rd383), _rt_buffer_get_64, (%rd378, %r92, %r368, %rd18, %rd19, %rd388, %rd388); // inline asm add.f32 %f824, %f821, 0f00000000; add.f32 %f825, %f822, 0f00000000; add.f32 %f826, %f823, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs234, %f826;} // inline asm // inline asm { cvt.rn.f16.f32 %rs233, %f825;} // inline asm // inline asm { cvt.rn.f16.f32 %rs232, %f824;} // inline asm mov.u16 %rs239, 0; st.v4.u16 [%rd383], {%rs232, %rs233, %rs234, %rs239}; bra.uni BB0_143; BB0_142: mov.u64 %rd396, image_RNM4; cvta.global.u64 %rd391, %rd396; mov.u32 %r370, 8; mov.u64 %rd395, 0; // inline asm call (%rd390), _rt_buffer_get_64, (%rd391, %r92, %r370, %rd18, %rd19, %rd395, %rd395); // inline asm mov.f32 %f827, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs240, %f827;} // inline asm mov.u16 %rs241, 0; st.v4.u16 [%rd390], {%rs240, %rs240, %rs240, %rs241}; BB0_143: ld.global.u32 %r371, [additive]; setp.eq.s32 %p111, %r371, 0; @%p111 bra BB0_145; mov.u64 %rd409, image_RNM5; cvta.global.u64 %rd398, %rd409; mov.u32 %r375, 8; mov.u64 %rd408, 0; // inline asm call (%rd397), _rt_buffer_get_64, (%rd398, %r92, %r375, %rd18, %rd19, %rd408, %rd408); // inline asm ld.v4.u16 {%rs248, %rs249, %rs250, %rs251}, [%rd397]; // inline asm { cvt.f32.f16 %f828, %rs248;} // inline asm // inline asm { cvt.f32.f16 %f829, %rs249;} // inline asm // inline asm { cvt.f32.f16 %f830, %rs250;} // inline asm // inline asm call (%rd403), _rt_buffer_get_64, (%rd398, %r92, %r375, %rd18, %rd19, %rd408, %rd408); // inline asm add.f32 %f831, %f828, 0f00000000; add.f32 %f832, %f829, 0f00000000; add.f32 %f833, %f830, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs247, %f833;} // inline asm // inline asm { cvt.rn.f16.f32 %rs246, %f832;} // inline asm // inline asm { cvt.rn.f16.f32 %rs245, %f831;} // inline asm mov.u16 %rs252, 0; st.v4.u16 [%rd403], {%rs245, %rs246, %rs247, %rs252}; bra.uni BB0_146; BB0_145: mov.u64 %rd416, image_RNM5; cvta.global.u64 %rd411, %rd416; mov.u32 %r377, 8; mov.u64 %rd415, 0; // inline asm call (%rd410), _rt_buffer_get_64, (%rd411, %r92, %r377, %rd18, %rd19, %rd415, %rd415); // inline asm mov.f32 %f834, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs253, %f834;} // inline asm mov.u16 %rs254, 0; st.v4.u16 [%rd410], {%rs253, %rs253, %rs253, %rs254}; BB0_146: ld.global.u32 %r378, [additive]; setp.eq.s32 %p112, %r378, 0; @%p112 bra BB0_148; mov.u64 %rd429, image_RNM6; cvta.global.u64 %rd418, %rd429; mov.u32 %r382, 8; mov.u64 %rd428, 0; // inline asm call (%rd417), _rt_buffer_get_64, (%rd418, %r92, %r382, %rd18, %rd19, %rd428, %rd428); // inline asm ld.v4.u16 {%rs261, %rs262, %rs263, %rs264}, [%rd417]; // inline asm { cvt.f32.f16 %f835, %rs261;} // inline asm // inline asm { cvt.f32.f16 %f836, %rs262;} // inline asm // inline asm { cvt.f32.f16 %f837, %rs263;} // inline asm // inline asm call (%rd423), _rt_buffer_get_64, (%rd418, %r92, %r382, %rd18, %rd19, %rd428, %rd428); // inline asm add.f32 %f838, %f835, 0f00000000; add.f32 %f839, %f836, 0f00000000; add.f32 %f840, %f837, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs260, %f840;} // inline asm // inline asm { cvt.rn.f16.f32 %rs259, %f839;} // inline asm // inline asm { cvt.rn.f16.f32 %rs258, %f838;} // inline asm mov.u16 %rs265, 0; st.v4.u16 [%rd423], {%rs258, %rs259, %rs260, %rs265}; bra.uni BB0_149; BB0_148: mov.u64 %rd436, image_RNM6; cvta.global.u64 %rd431, %rd436; mov.u32 %r384, 8; mov.u64 %rd435, 0; // inline asm call (%rd430), _rt_buffer_get_64, (%rd431, %r92, %r384, %rd18, %rd19, %rd435, %rd435); // inline asm mov.f32 %f841, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs266, %f841;} // inline asm mov.u16 %rs267, 0; st.v4.u16 [%rd430], {%rs266, %rs266, %rs266, %rs267}; BB0_149: ld.global.u32 %r385, [additive]; setp.eq.s32 %p113, %r385, 0; @%p113 bra BB0_151; mov.u64 %rd449, image_RNM7; cvta.global.u64 %rd438, %rd449; mov.u32 %r389, 8; mov.u64 %rd448, 0; // inline asm call (%rd437), _rt_buffer_get_64, (%rd438, %r92, %r389, %rd18, %rd19, %rd448, %rd448); // inline asm ld.v4.u16 {%rs274, %rs275, %rs276, %rs277}, [%rd437]; // inline asm { cvt.f32.f16 %f842, %rs274;} // inline asm // inline asm { cvt.f32.f16 %f843, %rs275;} // inline asm // inline asm { cvt.f32.f16 %f844, %rs276;} // inline asm // inline asm call (%rd443), _rt_buffer_get_64, (%rd438, %r92, %r389, %rd18, %rd19, %rd448, %rd448); // inline asm add.f32 %f845, %f842, 0f00000000; add.f32 %f846, %f843, 0f00000000; add.f32 %f847, %f844, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs273, %f847;} // inline asm // inline asm { cvt.rn.f16.f32 %rs272, %f846;} // inline asm // inline asm { cvt.rn.f16.f32 %rs271, %f845;} // inline asm mov.u16 %rs278, 0; st.v4.u16 [%rd443], {%rs271, %rs272, %rs273, %rs278}; bra.uni BB0_152; BB0_151: mov.u64 %rd456, image_RNM7; cvta.global.u64 %rd451, %rd456; mov.u32 %r391, 8; mov.u64 %rd455, 0; // inline asm call (%rd450), _rt_buffer_get_64, (%rd451, %r92, %r391, %rd18, %rd19, %rd455, %rd455); // inline asm mov.f32 %f848, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs279, %f848;} // inline asm mov.u16 %rs280, 0; st.v4.u16 [%rd450], {%rs279, %rs279, %rs279, %rs280}; BB0_152: ld.global.u32 %r392, [additive]; setp.eq.s32 %p114, %r392, 0; @%p114 bra BB0_154; mov.u64 %rd469, image_RNM8; cvta.global.u64 %rd458, %rd469; mov.u32 %r396, 8; mov.u64 %rd468, 0; // inline asm call (%rd457), _rt_buffer_get_64, (%rd458, %r92, %r396, %rd18, %rd19, %rd468, %rd468); // inline asm ld.v4.u16 {%rs287, %rs288, %rs289, %rs290}, [%rd457]; // inline asm { cvt.f32.f16 %f849, %rs287;} // inline asm // inline asm { cvt.f32.f16 %f850, %rs288;} // inline asm // inline asm { cvt.f32.f16 %f851, %rs289;} // inline asm // inline asm call (%rd463), _rt_buffer_get_64, (%rd458, %r92, %r396, %rd18, %rd19, %rd468, %rd468); // inline asm add.f32 %f852, %f849, 0f00000000; add.f32 %f853, %f850, 0f00000000; add.f32 %f854, %f851, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs286, %f854;} // inline asm // inline asm { cvt.rn.f16.f32 %rs285, %f853;} // inline asm // inline asm { cvt.rn.f16.f32 %rs284, %f852;} // inline asm mov.u16 %rs291, 0; st.v4.u16 [%rd463], {%rs284, %rs285, %rs286, %rs291}; bra.uni BB0_155; BB0_154: mov.u64 %rd476, image_RNM8; cvta.global.u64 %rd471, %rd476; mov.u32 %r398, 8; mov.u64 %rd475, 0; // inline asm call (%rd470), _rt_buffer_get_64, (%rd471, %r92, %r398, %rd18, %rd19, %rd475, %rd475); // inline asm mov.f32 %f855, 0f00000000; // inline asm { cvt.rn.f16.f32 %rs292, %f855;} // inline asm mov.u16 %rs293, 0; st.v4.u16 [%rd470], {%rs292, %rs292, %rs292, %rs293}; bra.uni BB0_155; BB0_60: setp.geu.f32 %p53, %f121, 0f00000000; @%p53 bra BB0_63; mov.f32 %f883, 0f3EE66666; cvt.rzi.f32.f32 %f439, %f883; setp.neu.f32 %p54, %f439, 0f3EE66666; selp.f32 %f970, 0f7FFFFFFF, %f970, %p54; BB0_63: abs.f32 %f860, %f121; add.f32 %f441, %f860, 0f3EE66666; mov.b32 %r230, %f441; setp.lt.s32 %p56, %r230, 2139095040; @%p56 bra BB0_68; abs.f32 %f881, %f121; setp.gtu.f32 %p57, %f881, 0f7F800000; @%p57 bra BB0_67; bra.uni BB0_65; BB0_67: add.f32 %f970, %f121, 0f3EE66666; bra.uni BB0_68; BB0_65: abs.f32 %f882, %f121; setp.neu.f32 %p58, %f882, 0f7F800000; @%p58 bra BB0_68; selp.f32 %f970, 0fFF800000, 0f7F800000, %p1; BB0_68: mov.f32 %f869, 0fB5BFBE8E; mov.f32 %f868, 0fBF317200; mov.f32 %f867, 0f00000000; mov.f32 %f866, 0f35BFBE8E; mov.f32 %f865, 0f3F317200; mov.f32 %f864, 0f3DAAAABD; mov.f32 %f863, 0f3C4CAF63; mov.f32 %f862, 0f3B18F0FE; mov.f32 %f861, 0f3EE66666; setp.eq.f32 %p59, %f121, 0f3F800000; selp.f32 %f137, 0f3F800000, %f970, %p59; abs.f32 %f138, %f122; setp.lt.f32 %p60, %f138, 0f00800000; mul.f32 %f444, %f138, 0f4B800000; selp.f32 %f445, 0fC3170000, 0fC2FE0000, %p60; selp.f32 %f446, %f444, %f138, %p60; mov.b32 %r231, %f446; and.b32 %r232, %r231, 8388607; or.b32 %r233, %r232, 1065353216; mov.b32 %f447, %r233; shr.u32 %r234, %r231, 23; cvt.rn.f32.u32 %f448, %r234; add.f32 %f449, %f445, %f448; setp.gt.f32 %p61, %f447, 0f3FB504F3; mul.f32 %f450, %f447, 0f3F000000; add.f32 %f451, %f449, 0f3F800000; selp.f32 %f452, %f450, %f447, %p61; selp.f32 %f453, %f451, %f449, %p61; add.f32 %f454, %f452, 0fBF800000; add.f32 %f443, %f452, 0f3F800000; // inline asm rcp.approx.ftz.f32 %f442,%f443; // inline asm add.f32 %f455, %f454, %f454; mul.f32 %f456, %f442, %f455; mul.f32 %f457, %f456, %f456; fma.rn.f32 %f460, %f862, %f457, %f863; fma.rn.f32 %f462, %f460, %f457, %f864; mul.rn.f32 %f463, %f462, %f457; mul.rn.f32 %f464, %f463, %f456; sub.f32 %f465, %f454, %f456; neg.f32 %f466, %f456; add.f32 %f467, %f465, %f465; fma.rn.f32 %f468, %f466, %f454, %f467; mul.rn.f32 %f469, %f442, %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; mul.rn.f32 %f478, %f453, %f865; mul.rn.f32 %f480, %f453, %f866; 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 %f490, %f861, %f486; neg.f32 %f491, %f490; fma.rn.f32 %f492, %f861, %f486, %f491; fma.rn.f32 %f493, %f861, %f488, %f492; fma.rn.f32 %f495, %f867, %f486, %f493; add.rn.f32 %f496, %f490, %f495; neg.f32 %f497, %f496; add.rn.f32 %f498, %f490, %f497; add.rn.f32 %f499, %f498, %f495; mov.b32 %r235, %f496; setp.eq.s32 %p62, %r235, 1118925336; add.s32 %r236, %r235, -1; mov.b32 %f500, %r236; add.f32 %f501, %f499, 0f37000000; selp.f32 %f502, %f500, %f496, %p62; selp.f32 %f139, %f501, %f499, %p62; mul.f32 %f503, %f502, 0f3FB8AA3B; cvt.rzi.f32.f32 %f504, %f503; fma.rn.f32 %f506, %f504, %f868, %f502; fma.rn.f32 %f508, %f504, %f869, %f506; mul.f32 %f509, %f508, 0f3FB8AA3B; ex2.approx.ftz.f32 %f510, %f509; add.f32 %f511, %f504, 0f00000000; ex2.approx.f32 %f512, %f511; mul.f32 %f513, %f510, %f512; setp.lt.f32 %p63, %f502, 0fC2D20000; selp.f32 %f514, 0f00000000, %f513, %p63; setp.gt.f32 %p64, %f502, 0f42D20000; selp.f32 %f971, 0f7F800000, %f514, %p64; setp.eq.f32 %p65, %f971, 0f7F800000; @%p65 bra BB0_70; fma.rn.f32 %f971, %f971, %f139, %f971; BB0_70: setp.lt.f32 %p66, %f122, 0f00000000; and.pred %p2, %p66, %p51; mov.b32 %r237, %f971; xor.b32 %r238, %r237, -2147483648; mov.b32 %f515, %r238; selp.f32 %f973, %f515, %f971, %p2; setp.eq.f32 %p68, %f122, 0f00000000; @%p68 bra BB0_73; bra.uni BB0_71; BB0_73: add.f32 %f518, %f122, %f122; selp.f32 %f973, %f518, 0f00000000, %p51; bra.uni BB0_74; BB0_71: setp.geu.f32 %p69, %f122, 0f00000000; @%p69 bra BB0_74; mov.f32 %f880, 0f3EE66666; cvt.rzi.f32.f32 %f517, %f880; setp.neu.f32 %p70, %f517, 0f3EE66666; selp.f32 %f973, 0f7FFFFFFF, %f973, %p70; BB0_74: abs.f32 %f884, %f122; add.f32 %f519, %f884, 0f3EE66666; mov.b32 %r239, %f519; setp.lt.s32 %p72, %r239, 2139095040; @%p72 bra BB0_79; abs.f32 %f885, %f122; setp.gtu.f32 %p73, %f885, 0f7F800000; @%p73 bra BB0_78; bra.uni BB0_76; BB0_78: add.f32 %f973, %f122, 0f3EE66666; bra.uni BB0_79; BB0_76: abs.f32 %f886, %f122; setp.neu.f32 %p74, %f886, 0f7F800000; @%p74 bra BB0_79; selp.f32 %f973, 0fFF800000, 0f7F800000, %p2; BB0_79: mov.f32 %f878, 0fB5BFBE8E; mov.f32 %f877, 0fBF317200; mov.f32 %f876, 0f00000000; mov.f32 %f875, 0f35BFBE8E; mov.f32 %f874, 0f3F317200; mov.f32 %f873, 0f3DAAAABD; mov.f32 %f872, 0f3C4CAF63; mov.f32 %f871, 0f3B18F0FE; mov.f32 %f870, 0f3EE66666; setp.eq.f32 %p75, %f122, 0f3F800000; selp.f32 %f150, 0f3F800000, %f973, %p75; abs.f32 %f151, %f123; setp.lt.f32 %p76, %f151, 0f00800000; mul.f32 %f522, %f151, 0f4B800000; selp.f32 %f523, 0fC3170000, 0fC2FE0000, %p76; selp.f32 %f524, %f522, %f151, %p76; mov.b32 %r240, %f524; and.b32 %r241, %r240, 8388607; or.b32 %r242, %r241, 1065353216; mov.b32 %f525, %r242; shr.u32 %r243, %r240, 23; cvt.rn.f32.u32 %f526, %r243; add.f32 %f527, %f523, %f526; setp.gt.f32 %p77, %f525, 0f3FB504F3; mul.f32 %f528, %f525, 0f3F000000; add.f32 %f529, %f527, 0f3F800000; selp.f32 %f530, %f528, %f525, %p77; selp.f32 %f531, %f529, %f527, %p77; add.f32 %f532, %f530, 0fBF800000; add.f32 %f521, %f530, 0f3F800000; // inline asm rcp.approx.ftz.f32 %f520,%f521; // inline asm add.f32 %f533, %f532, %f532; mul.f32 %f534, %f520, %f533; mul.f32 %f535, %f534, %f534; fma.rn.f32 %f538, %f871, %f535, %f872; fma.rn.f32 %f540, %f538, %f535, %f873; mul.rn.f32 %f541, %f540, %f535; mul.rn.f32 %f542, %f541, %f534; sub.f32 %f543, %f532, %f534; neg.f32 %f544, %f534; add.f32 %f545, %f543, %f543; fma.rn.f32 %f546, %f544, %f532, %f545; mul.rn.f32 %f547, %f520, %f546; add.f32 %f548, %f542, %f534; sub.f32 %f549, %f534, %f548; add.f32 %f550, %f542, %f549; add.f32 %f551, %f547, %f550; add.f32 %f552, %f548, %f551; sub.f32 %f553, %f548, %f552; add.f32 %f554, %f551, %f553; mul.rn.f32 %f556, %f531, %f874; mul.rn.f32 %f558, %f531, %f875; add.f32 %f559, %f556, %f552; sub.f32 %f560, %f556, %f559; add.f32 %f561, %f552, %f560; add.f32 %f562, %f554, %f561; add.f32 %f563, %f558, %f562; add.f32 %f564, %f559, %f563; sub.f32 %f565, %f559, %f564; add.f32 %f566, %f563, %f565; mul.rn.f32 %f568, %f870, %f564; neg.f32 %f569, %f568; fma.rn.f32 %f570, %f870, %f564, %f569; fma.rn.f32 %f571, %f870, %f566, %f570; fma.rn.f32 %f573, %f876, %f564, %f571; add.rn.f32 %f574, %f568, %f573; neg.f32 %f575, %f574; add.rn.f32 %f576, %f568, %f575; add.rn.f32 %f577, %f576, %f573; mov.b32 %r244, %f574; setp.eq.s32 %p78, %r244, 1118925336; add.s32 %r245, %r244, -1; mov.b32 %f578, %r245; add.f32 %f579, %f577, 0f37000000; selp.f32 %f580, %f578, %f574, %p78; selp.f32 %f152, %f579, %f577, %p78; mul.f32 %f581, %f580, 0f3FB8AA3B; cvt.rzi.f32.f32 %f582, %f581; fma.rn.f32 %f584, %f582, %f877, %f580; fma.rn.f32 %f586, %f582, %f878, %f584; mul.f32 %f587, %f586, 0f3FB8AA3B; ex2.approx.ftz.f32 %f588, %f587; add.f32 %f589, %f582, 0f00000000; ex2.approx.f32 %f590, %f589; mul.f32 %f591, %f588, %f590; setp.lt.f32 %p79, %f580, 0fC2D20000; selp.f32 %f592, 0f00000000, %f591, %p79; setp.gt.f32 %p80, %f580, 0f42D20000; selp.f32 %f974, 0f7F800000, %f592, %p80; setp.eq.f32 %p81, %f974, 0f7F800000; @%p81 bra BB0_81; fma.rn.f32 %f974, %f974, %f152, %f974; BB0_81: setp.lt.f32 %p82, %f123, 0f00000000; and.pred %p3, %p82, %p51; mov.b32 %r246, %f974; xor.b32 %r247, %r246, -2147483648; mov.b32 %f593, %r247; selp.f32 %f976, %f593, %f974, %p3; setp.eq.f32 %p84, %f123, 0f00000000; @%p84 bra BB0_84; bra.uni BB0_82; BB0_84: add.f32 %f596, %f123, %f123; selp.f32 %f976, %f596, 0f00000000, %p51; bra.uni BB0_85; BB0_82: setp.geu.f32 %p85, %f123, 0f00000000; @%p85 bra BB0_85; mov.f32 %f879, 0f3EE66666; cvt.rzi.f32.f32 %f595, %f879; setp.neu.f32 %p86, %f595, 0f3EE66666; selp.f32 %f976, 0f7FFFFFFF, %f976, %p86; BB0_85: abs.f32 %f906, %f123; add.f32 %f597, %f906, 0f3EE66666; mov.b32 %r248, %f597; setp.lt.s32 %p88, %r248, 2139095040; @%p88 bra BB0_90; abs.f32 %f907, %f123; setp.gtu.f32 %p89, %f907, 0f7F800000; @%p89 bra BB0_89; bra.uni BB0_87; BB0_89: add.f32 %f976, %f123, 0f3EE66666; bra.uni BB0_90; BB0_87: abs.f32 %f908, %f123; setp.neu.f32 %p90, %f908, 0f7F800000; @%p90 bra BB0_90; selp.f32 %f976, 0fFF800000, 0f7F800000, %p3; BB0_90: mov.u32 %r399, 4; setp.eq.f32 %p91, %f123, 0f3F800000; selp.f32 %f598, 0f3F800000, %f976, %p91; cvt.u64.u32 %rd66, %r5; cvt.u64.u32 %rd65, %r4; mov.u64 %rd69, image; cvta.global.u64 %rd64, %rd69; // inline asm call (%rd63), _rt_buffer_get_64, (%rd64, %r92, %r399, %rd65, %rd66, %rd25, %rd25); // inline asm cvt.sat.f32.f32 %f599, %f598; mul.f32 %f600, %f599, 0f437FFD71; cvt.rzi.u32.f32 %r251, %f600; cvt.sat.f32.f32 %f601, %f150; mul.f32 %f602, %f601, 0f437FFD71; cvt.rzi.u32.f32 %r252, %f602; cvt.sat.f32.f32 %f603, %f137; mul.f32 %f604, %f603, 0f437FFD71; cvt.rzi.u32.f32 %r253, %f604; cvt.u16.u32 %rs19, %r251; cvt.u16.u32 %rs20, %r253; cvt.u16.u32 %rs21, %r252; mov.u16 %rs22, 255; st.v4.u8 [%rd63], {%rs19, %rs21, %rs20, %rs22}; ld.global.u32 %r449, [imageEnabled]; BB0_91: cvt.u64.u32 %rd16, %r4; cvt.u64.u32 %rd17, %r5; and.b32 %r254, %r449, 4; setp.eq.s32 %p92, %r254, 0; @%p92 bra BB0_95; ld.global.u32 %r255, [additive]; setp.eq.s32 %p93, %r255, 0; mov.f32 %f605, 0f3F800000; // inline asm { cvt.rn.f16.f32 %rs23, %f605;} // inline asm @%p93 bra BB0_94; mov.u64 %rd82, image_HDR; cvta.global.u64 %rd71, %rd82; mov.u32 %r259, 8; // inline asm call (%rd70), _rt_buffer_get_64, (%rd71, %r92, %r259, %rd16, %rd17, %rd25, %rd25); // inline asm ld.v4.u16 {%rs30, %rs31, %rs32, %rs33}, [%rd70]; // inline asm { cvt.f32.f16 %f606, %rs30;} // inline asm // inline asm { cvt.f32.f16 %f607, %rs31;} // inline asm // inline asm { cvt.f32.f16 %f608, %rs32;} // inline asm // inline asm call (%rd76), _rt_buffer_get_64, (%rd71, %r92, %r259, %rd16, %rd17, %rd25, %rd25); // inline asm add.f32 %f609, %f121, %f606; add.f32 %f610, %f122, %f607; add.f32 %f611, %f123, %f608; // inline asm { cvt.rn.f16.f32 %rs29, %f611;} // inline asm // inline asm { cvt.rn.f16.f32 %rs28, %f610;} // inline asm // inline asm { cvt.rn.f16.f32 %rs27, %f609;} // inline asm st.v4.u16 [%rd76], {%rs27, %rs28, %rs29, %rs23}; bra.uni BB0_95; BB0_94: mov.u64 %rd89, image_HDR; cvta.global.u64 %rd84, %rd89; mov.u32 %r261, 8; // inline asm call (%rd83), _rt_buffer_get_64, (%rd84, %r92, %r261, %rd16, %rd17, %rd25, %rd25); // inline asm // inline asm { cvt.rn.f16.f32 %rs36, %f123;} // inline asm // inline asm { cvt.rn.f16.f32 %rs35, %f122;} // inline asm // inline asm { cvt.rn.f16.f32 %rs34, %f121;} // inline asm st.v4.u16 [%rd83], {%rs34, %rs35, %rs36, %rs23}; BB0_95: ld.global.f32 %f616, [skyColor]; mul.f32 %f617, %f616, 0f40800000; ld.global.f32 %f618, [skyColor+4]; mul.f32 %f619, %f618, 0f40800000; ld.global.f32 %f620, [skyColor+8]; mul.f32 %f621, %f620, 0f40800000; mul.f32 %f622, %f112, %f617; mul.f32 %f623, %f112, %f619; mul.f32 %f624, %f112, %f621; mul.f32 %f163, %f622, 0f3F558750; mul.f32 %f164, %f623, 0f3F558750; mul.f32 %f165, %f624, 0f3F558750; ld.global.u32 %r262, [additive]; setp.eq.s32 %p94, %r262, 0; mov.f32 %f615, 0f3F800000; // inline asm { cvt.rn.f16.f32 %rs37, %f615;} // inline asm @%p94 bra BB0_97; mov.u64 %rd102, image_RNM0; cvta.global.u64 %rd91, %rd102; mov.u32 %r266, 8; // inline asm call (%rd90), _rt_buffer_get_64, (%rd91, %r92, %r266, %rd16, %rd17, %rd25, %rd25); // inline asm ld.v4.u16 {%rs44, %rs45, %rs46, %rs47}, [%rd90]; // inline asm { cvt.f32.f16 %f625, %rs44;} // inline asm // inline asm { cvt.f32.f16 %f626, %rs45;} // inline asm // inline asm { cvt.f32.f16 %f627, %rs46;} // inline asm // inline asm call (%rd96), _rt_buffer_get_64, (%rd91, %r92, %r266, %rd16, %rd17, %rd25, %rd25); // inline asm add.f32 %f628, %f163, %f625; add.f32 %f629, %f164, %f626; add.f32 %f630, %f165, %f627; // inline asm { cvt.rn.f16.f32 %rs43, %f630;} // inline asm // inline asm { cvt.rn.f16.f32 %rs42, %f629;} // inline asm // inline asm { cvt.rn.f16.f32 %rs41, %f628;} // inline asm st.v4.u16 [%rd96], {%rs41, %rs42, %rs43, %rs37}; bra.uni BB0_98; BB0_97: mov.u64 %rd109, image_RNM0; cvta.global.u64 %rd104, %rd109; mov.u32 %r268, 8; // inline asm call (%rd103), _rt_buffer_get_64, (%rd104, %r92, %r268, %rd16, %rd17, %rd25, %rd25); // inline asm // inline asm { cvt.rn.f16.f32 %rs50, %f165;} // inline asm // inline asm { cvt.rn.f16.f32 %rs49, %f164;} // inline asm // inline asm { cvt.rn.f16.f32 %rs48, %f163;} // inline asm st.v4.u16 [%rd103], {%rs48, %rs49, %rs50, %rs37}; BB0_98: ld.global.f32 %f635, [skyColor]; mul.f32 %f636, %f635, 0f40800000; ld.global.f32 %f637, [skyColor+4]; mul.f32 %f638, %f637, 0f40800000; ld.global.f32 %f639, [skyColor+8]; mul.f32 %f640, %f639, 0f40800000; mul.f32 %f641, %f113, %f636; mul.f32 %f642, %f113, %f638; mul.f32 %f643, %f113, %f640; mul.f32 %f166, %f641, 0f3FB8EBD1; mul.f32 %f167, %f642, 0f3FB8EBD1; mul.f32 %f168, %f643, 0f3FB8EBD1; ld.global.u32 %r269, [additive]; setp.eq.s32 %p95, %r269, 0; // inline asm { cvt.rn.f16.f32 %rs51, %f615;} // inline asm @%p95 bra BB0_100; mov.u64 %rd122, image_RNM1; cvta.global.u64 %rd111, %rd122; mov.u32 %r273, 8; // inline asm call (%rd110), _rt_buffer_get_64, (%rd111, %r92, %r273, %rd16, %rd17, %rd25, %rd25); // inline asm ld.v4.u16 {%rs58, %rs59, %rs60, %rs61}, [%rd110]; // inline asm { cvt.f32.f16 %f644, %rs58;} // inline asm // inline asm { cvt.f32.f16 %f645, %rs59;} // inline asm // inline asm { cvt.f32.f16 %f646, %rs60;} // inline asm // inline asm call (%rd116), _rt_buffer_get_64, (%rd111, %r92, %r273, %rd16, %rd17, %rd25, %rd25); // inline asm add.f32 %f647, %f166, %f644; add.f32 %f648, %f167, %f645; add.f32 %f649, %f168, %f646; // inline asm { cvt.rn.f16.f32 %rs57, %f649;} // inline asm // inline asm { cvt.rn.f16.f32 %rs56, %f648;} // inline asm // inline asm { cvt.rn.f16.f32 %rs55, %f647;} // inline asm st.v4.u16 [%rd116], {%rs55, %rs56, %rs57, %rs51}; bra.uni BB0_101; BB0_100: mov.u64 %rd129, image_RNM1; cvta.global.u64 %rd124, %rd129; mov.u32 %r275, 8; // inline asm call (%rd123), _rt_buffer_get_64, (%rd124, %r92, %r275, %rd16, %rd17, %rd25, %rd25); // inline asm // inline asm { cvt.rn.f16.f32 %rs64, %f168;} // inline asm // inline asm { cvt.rn.f16.f32 %rs63, %f167;} // inline asm // inline asm { cvt.rn.f16.f32 %rs62, %f166;} // inline asm st.v4.u16 [%rd123], {%rs62, %rs63, %rs64, %rs51}; BB0_101: ld.global.f32 %f654, [skyColor]; mul.f32 %f655, %f654, 0f40800000; ld.global.f32 %f656, [skyColor+4]; mul.f32 %f657, %f656, 0f40800000; ld.global.f32 %f658, [skyColor+8]; mul.f32 %f659, %f658, 0f40800000; mul.f32 %f660, %f114, %f655; mul.f32 %f661, %f114, %f657; mul.f32 %f662, %f114, %f659; mul.f32 %f169, %f660, 0f3FB8EBD1; mul.f32 %f170, %f661, 0f3FB8EBD1; mul.f32 %f171, %f662, 0f3FB8EBD1; ld.global.u32 %r276, [additive]; setp.eq.s32 %p96, %r276, 0; // inline asm { cvt.rn.f16.f32 %rs65, %f615;} // inline asm @%p96 bra BB0_103; mov.u64 %rd142, image_RNM2; cvta.global.u64 %rd131, %rd142; mov.u32 %r280, 8; // inline asm call (%rd130), _rt_buffer_get_64, (%rd131, %r92, %r280, %rd16, %rd17, %rd25, %rd25); // inline asm ld.v4.u16 {%rs72, %rs73, %rs74, %rs75}, [%rd130]; // inline asm { cvt.f32.f16 %f663, %rs72;} // inline asm // inline asm { cvt.f32.f16 %f664, %rs73;} // inline asm // inline asm { cvt.f32.f16 %f665, %rs74;} // inline asm // inline asm call (%rd136), _rt_buffer_get_64, (%rd131, %r92, %r280, %rd16, %rd17, %rd25, %rd25); // inline asm add.f32 %f666, %f169, %f663; add.f32 %f667, %f170, %f664; add.f32 %f668, %f171, %f665; // inline asm { cvt.rn.f16.f32 %rs71, %f668;} // inline asm // inline asm { cvt.rn.f16.f32 %rs70, %f667;} // inline asm // inline asm { cvt.rn.f16.f32 %rs69, %f666;} // inline asm st.v4.u16 [%rd136], {%rs69, %rs70, %rs71, %rs65}; bra.uni BB0_104; BB0_103: mov.u64 %rd149, image_RNM2; cvta.global.u64 %rd144, %rd149; mov.u32 %r282, 8; // inline asm call (%rd143), _rt_buffer_get_64, (%rd144, %r92, %r282, %rd16, %rd17, %rd25, %rd25); // inline asm // inline asm { cvt.rn.f16.f32 %rs78, %f171;} // inline asm // inline asm { cvt.rn.f16.f32 %rs77, %f170;} // inline asm // inline asm { cvt.rn.f16.f32 %rs76, %f169;} // inline asm st.v4.u16 [%rd143], {%rs76, %rs77, %rs78, %rs65}; BB0_104: ld.global.f32 %f673, [skyColor]; mul.f32 %f674, %f673, 0f40800000; ld.global.f32 %f675, [skyColor+4]; mul.f32 %f676, %f675, 0f40800000; ld.global.f32 %f677, [skyColor+8]; mul.f32 %f678, %f677, 0f40800000; mul.f32 %f679, %f115, %f674; mul.f32 %f680, %f115, %f676; mul.f32 %f681, %f115, %f678; mul.f32 %f172, %f679, 0f3FB8EBD1; mul.f32 %f173, %f680, 0f3FB8EBD1; mul.f32 %f174, %f681, 0f3FB8EBD1; ld.global.u32 %r283, [additive]; setp.eq.s32 %p97, %r283, 0; // inline asm { cvt.rn.f16.f32 %rs79, %f615;} // inline asm @%p97 bra BB0_106; mov.u64 %rd162, image_RNM3; cvta.global.u64 %rd151, %rd162; mov.u32 %r287, 8; // inline asm call (%rd150), _rt_buffer_get_64, (%rd151, %r92, %r287, %rd16, %rd17, %rd25, %rd25); // inline asm ld.v4.u16 {%rs86, %rs87, %rs88, %rs89}, [%rd150]; // inline asm { cvt.f32.f16 %f682, %rs86;} // inline asm // inline asm { cvt.f32.f16 %f683, %rs87;} // inline asm // inline asm { cvt.f32.f16 %f684, %rs88;} // inline asm // inline asm call (%rd156), _rt_buffer_get_64, (%rd151, %r92, %r287, %rd16, %rd17, %rd25, %rd25); // inline asm add.f32 %f685, %f172, %f682; add.f32 %f686, %f173, %f683; add.f32 %f687, %f174, %f684; // inline asm { cvt.rn.f16.f32 %rs85, %f687;} // inline asm // inline asm { cvt.rn.f16.f32 %rs84, %f686;} // inline asm // inline asm { cvt.rn.f16.f32 %rs83, %f685;} // inline asm st.v4.u16 [%rd156], {%rs83, %rs84, %rs85, %rs79}; bra.uni BB0_107; BB0_106: mov.u64 %rd169, image_RNM3; cvta.global.u64 %rd164, %rd169; mov.u32 %r289, 8; // inline asm call (%rd163), _rt_buffer_get_64, (%rd164, %r92, %r289, %rd16, %rd17, %rd25, %rd25); // inline asm // inline asm { cvt.rn.f16.f32 %rs92, %f174;} // inline asm // inline asm { cvt.rn.f16.f32 %rs91, %f173;} // inline asm // inline asm { cvt.rn.f16.f32 %rs90, %f172;} // inline asm st.v4.u16 [%rd163], {%rs90, %rs91, %rs92, %rs79}; BB0_107: ld.global.f32 %f692, [skyColor]; mul.f32 %f693, %f692, 0f40800000; ld.global.f32 %f694, [skyColor+4]; mul.f32 %f695, %f694, 0f40800000; ld.global.f32 %f696, [skyColor+8]; mul.f32 %f697, %f696, 0f40800000; mul.f32 %f698, %f116, %f693; mul.f32 %f699, %f116, %f695; mul.f32 %f700, %f116, %f697; mul.f32 %f175, %f698, 0f404EBF87; mul.f32 %f176, %f699, 0f404EBF87; mul.f32 %f177, %f700, 0f404EBF87; ld.global.u32 %r290, [additive]; setp.eq.s32 %p98, %r290, 0; // inline asm { cvt.rn.f16.f32 %rs93, %f615;} // inline asm @%p98 bra BB0_109; mov.u64 %rd182, image_RNM4; cvta.global.u64 %rd171, %rd182; mov.u32 %r294, 8; // inline asm call (%rd170), _rt_buffer_get_64, (%rd171, %r92, %r294, %rd16, %rd17, %rd25, %rd25); // inline asm ld.v4.u16 {%rs100, %rs101, %rs102, %rs103}, [%rd170]; // inline asm { cvt.f32.f16 %f701, %rs100;} // inline asm // inline asm { cvt.f32.f16 %f702, %rs101;} // inline asm // inline asm { cvt.f32.f16 %f703, %rs102;} // inline asm // inline asm call (%rd176), _rt_buffer_get_64, (%rd171, %r92, %r294, %rd16, %rd17, %rd25, %rd25); // inline asm add.f32 %f704, %f175, %f701; add.f32 %f705, %f176, %f702; add.f32 %f706, %f177, %f703; // inline asm { cvt.rn.f16.f32 %rs99, %f706;} // inline asm // inline asm { cvt.rn.f16.f32 %rs98, %f705;} // inline asm // inline asm { cvt.rn.f16.f32 %rs97, %f704;} // inline asm st.v4.u16 [%rd176], {%rs97, %rs98, %rs99, %rs93}; bra.uni BB0_110; BB0_109: mov.u64 %rd189, image_RNM4; cvta.global.u64 %rd184, %rd189; mov.u32 %r296, 8; // inline asm call (%rd183), _rt_buffer_get_64, (%rd184, %r92, %r296, %rd16, %rd17, %rd25, %rd25); // inline asm // inline asm { cvt.rn.f16.f32 %rs106, %f177;} // inline asm // inline asm { cvt.rn.f16.f32 %rs105, %f176;} // inline asm // inline asm { cvt.rn.f16.f32 %rs104, %f175;} // inline asm st.v4.u16 [%rd183], {%rs104, %rs105, %rs106, %rs93}; BB0_110: ld.global.f32 %f711, [skyColor]; mul.f32 %f712, %f711, 0f40800000; ld.global.f32 %f713, [skyColor+4]; mul.f32 %f714, %f713, 0f40800000; ld.global.f32 %f715, [skyColor+8]; mul.f32 %f716, %f715, 0f40800000; mul.f32 %f717, %f117, %f712; mul.f32 %f718, %f117, %f714; mul.f32 %f719, %f117, %f716; mul.f32 %f178, %f717, 0f404EBF87; mul.f32 %f179, %f718, 0f404EBF87; mul.f32 %f180, %f719, 0f404EBF87; ld.global.u32 %r297, [additive]; setp.eq.s32 %p99, %r297, 0; // inline asm { cvt.rn.f16.f32 %rs107, %f615;} // inline asm @%p99 bra BB0_112; mov.u64 %rd202, image_RNM5; cvta.global.u64 %rd191, %rd202; mov.u32 %r301, 8; // inline asm call (%rd190), _rt_buffer_get_64, (%rd191, %r92, %r301, %rd16, %rd17, %rd25, %rd25); // inline asm ld.v4.u16 {%rs114, %rs115, %rs116, %rs117}, [%rd190]; // inline asm { cvt.f32.f16 %f720, %rs114;} // inline asm // inline asm { cvt.f32.f16 %f721, %rs115;} // inline asm // inline asm { cvt.f32.f16 %f722, %rs116;} // inline asm // inline asm call (%rd196), _rt_buffer_get_64, (%rd191, %r92, %r301, %rd16, %rd17, %rd25, %rd25); // inline asm add.f32 %f723, %f178, %f720; add.f32 %f724, %f179, %f721; add.f32 %f725, %f180, %f722; // inline asm { cvt.rn.f16.f32 %rs113, %f725;} // inline asm // inline asm { cvt.rn.f16.f32 %rs112, %f724;} // inline asm // inline asm { cvt.rn.f16.f32 %rs111, %f723;} // inline asm st.v4.u16 [%rd196], {%rs111, %rs112, %rs113, %rs107}; bra.uni BB0_113; BB0_112: mov.u64 %rd209, image_RNM5; cvta.global.u64 %rd204, %rd209; mov.u32 %r303, 8; // inline asm call (%rd203), _rt_buffer_get_64, (%rd204, %r92, %r303, %rd16, %rd17, %rd25, %rd25); // inline asm // inline asm { cvt.rn.f16.f32 %rs120, %f180;} // inline asm // inline asm { cvt.rn.f16.f32 %rs119, %f179;} // inline asm // inline asm { cvt.rn.f16.f32 %rs118, %f178;} // inline asm st.v4.u16 [%rd203], {%rs118, %rs119, %rs120, %rs107}; BB0_113: ld.global.f32 %f730, [skyColor]; mul.f32 %f731, %f730, 0f40800000; ld.global.f32 %f732, [skyColor+4]; mul.f32 %f733, %f732, 0f40800000; ld.global.f32 %f734, [skyColor+8]; mul.f32 %f735, %f734, 0f40800000; mul.f32 %f736, %f118, %f731; mul.f32 %f737, %f118, %f733; mul.f32 %f738, %f118, %f735; mul.f32 %f181, %f736, 0f40330C94; mul.f32 %f182, %f737, 0f40330C94; mul.f32 %f183, %f738, 0f40330C94; ld.global.u32 %r304, [additive]; setp.eq.s32 %p100, %r304, 0; // inline asm { cvt.rn.f16.f32 %rs121, %f615;} // inline asm @%p100 bra BB0_115; mov.u64 %rd222, image_RNM6; cvta.global.u64 %rd211, %rd222; mov.u32 %r308, 8; // inline asm call (%rd210), _rt_buffer_get_64, (%rd211, %r92, %r308, %rd16, %rd17, %rd25, %rd25); // inline asm ld.v4.u16 {%rs128, %rs129, %rs130, %rs131}, [%rd210]; // inline asm { cvt.f32.f16 %f739, %rs128;} // inline asm // inline asm { cvt.f32.f16 %f740, %rs129;} // inline asm // inline asm { cvt.f32.f16 %f741, %rs130;} // inline asm // inline asm call (%rd216), _rt_buffer_get_64, (%rd211, %r92, %r308, %rd16, %rd17, %rd25, %rd25); // inline asm add.f32 %f742, %f181, %f739; add.f32 %f743, %f182, %f740; add.f32 %f744, %f183, %f741; // inline asm { cvt.rn.f16.f32 %rs127, %f744;} // inline asm // inline asm { cvt.rn.f16.f32 %rs126, %f743;} // inline asm // inline asm { cvt.rn.f16.f32 %rs125, %f742;} // inline asm st.v4.u16 [%rd216], {%rs125, %rs126, %rs127, %rs121}; bra.uni BB0_116; BB0_115: mov.u64 %rd229, image_RNM6; cvta.global.u64 %rd224, %rd229; mov.u32 %r310, 8; // inline asm call (%rd223), _rt_buffer_get_64, (%rd224, %r92, %r310, %rd16, %rd17, %rd25, %rd25); // inline asm // inline asm { cvt.rn.f16.f32 %rs134, %f183;} // inline asm // inline asm { cvt.rn.f16.f32 %rs133, %f182;} // inline asm // inline asm { cvt.rn.f16.f32 %rs132, %f181;} // inline asm st.v4.u16 [%rd223], {%rs132, %rs133, %rs134, %rs121}; BB0_116: ld.global.f32 %f749, [skyColor]; mul.f32 %f750, %f749, 0f40800000; ld.global.f32 %f751, [skyColor+4]; mul.f32 %f752, %f751, 0f40800000; ld.global.f32 %f753, [skyColor+8]; mul.f32 %f754, %f753, 0f40800000; mul.f32 %f755, %f119, %f750; mul.f32 %f756, %f119, %f752; mul.f32 %f757, %f119, %f754; mul.f32 %f184, %f755, 0f404EBF87; mul.f32 %f185, %f756, 0f404EBF87; mul.f32 %f186, %f757, 0f404EBF87; ld.global.u32 %r311, [additive]; setp.eq.s32 %p101, %r311, 0; // inline asm { cvt.rn.f16.f32 %rs135, %f615;} // inline asm @%p101 bra BB0_118; mov.u64 %rd242, image_RNM7; cvta.global.u64 %rd231, %rd242; mov.u32 %r315, 8; // inline asm call (%rd230), _rt_buffer_get_64, (%rd231, %r92, %r315, %rd16, %rd17, %rd25, %rd25); // inline asm ld.v4.u16 {%rs142, %rs143, %rs144, %rs145}, [%rd230]; // inline asm { cvt.f32.f16 %f758, %rs142;} // inline asm // inline asm { cvt.f32.f16 %f759, %rs143;} // inline asm // inline asm { cvt.f32.f16 %f760, %rs144;} // inline asm // inline asm call (%rd236), _rt_buffer_get_64, (%rd231, %r92, %r315, %rd16, %rd17, %rd25, %rd25); // inline asm add.f32 %f761, %f184, %f758; add.f32 %f762, %f185, %f759; add.f32 %f763, %f186, %f760; // inline asm { cvt.rn.f16.f32 %rs141, %f763;} // inline asm // inline asm { cvt.rn.f16.f32 %rs140, %f762;} // inline asm // inline asm { cvt.rn.f16.f32 %rs139, %f761;} // inline asm st.v4.u16 [%rd236], {%rs139, %rs140, %rs141, %rs135}; bra.uni BB0_119; BB0_118: mov.u64 %rd249, image_RNM7; cvta.global.u64 %rd244, %rd249; mov.u32 %r317, 8; // inline asm call (%rd243), _rt_buffer_get_64, (%rd244, %r92, %r317, %rd16, %rd17, %rd25, %rd25); // inline asm // inline asm { cvt.rn.f16.f32 %rs148, %f186;} // inline asm // inline asm { cvt.rn.f16.f32 %rs147, %f185;} // inline asm // inline asm { cvt.rn.f16.f32 %rs146, %f184;} // inline asm st.v4.u16 [%rd243], {%rs146, %rs147, %rs148, %rs135}; BB0_119: ld.global.f32 %f768, [skyColor]; mul.f32 %f769, %f768, 0f40800000; ld.global.f32 %f770, [skyColor+4]; mul.f32 %f771, %f770, 0f40800000; ld.global.f32 %f772, [skyColor+8]; mul.f32 %f773, %f772, 0f40800000; mul.f32 %f774, %f120, %f769; mul.f32 %f775, %f120, %f771; mul.f32 %f776, %f120, %f773; mul.f32 %f187, %f774, 0f3FCEBF87; mul.f32 %f188, %f775, 0f3FCEBF87; mul.f32 %f189, %f776, 0f3FCEBF87; ld.global.u32 %r318, [additive]; setp.eq.s32 %p102, %r318, 0; // inline asm { cvt.rn.f16.f32 %rs149, %f615;} // inline asm @%p102 bra BB0_121; mov.u64 %rd262, image_RNM8; cvta.global.u64 %rd251, %rd262; mov.u32 %r322, 8; // inline asm call (%rd250), _rt_buffer_get_64, (%rd251, %r92, %r322, %rd16, %rd17, %rd25, %rd25); // inline asm ld.v4.u16 {%rs156, %rs157, %rs158, %rs159}, [%rd250]; // inline asm { cvt.f32.f16 %f777, %rs156;} // inline asm // inline asm { cvt.f32.f16 %f778, %rs157;} // inline asm // inline asm { cvt.f32.f16 %f779, %rs158;} // inline asm // inline asm call (%rd256), _rt_buffer_get_64, (%rd251, %r92, %r322, %rd16, %rd17, %rd25, %rd25); // inline asm add.f32 %f780, %f187, %f777; add.f32 %f781, %f188, %f778; add.f32 %f782, %f189, %f779; // inline asm { cvt.rn.f16.f32 %rs155, %f782;} // inline asm // inline asm { cvt.rn.f16.f32 %rs154, %f781;} // inline asm // inline asm { cvt.rn.f16.f32 %rs153, %f780;} // inline asm st.v4.u16 [%rd256], {%rs153, %rs154, %rs155, %rs149}; bra.uni BB0_155; BB0_121: mov.u64 %rd269, image_RNM8; cvta.global.u64 %rd264, %rd269; mov.u32 %r324, 8; // inline asm call (%rd263), _rt_buffer_get_64, (%rd264, %r92, %r324, %rd16, %rd17, %rd25, %rd25); // inline asm // inline asm { cvt.rn.f16.f32 %rs162, %f189;} // inline asm // inline asm { cvt.rn.f16.f32 %rs161, %f188;} // inline asm // inline asm { cvt.rn.f16.f32 %rs160, %f187;} // inline asm st.v4.u16 [%rd263], {%rs160, %rs161, %rs162, %rs149}; BB0_155: ret; }