diff options
Diffstat (limited to 'VRCSDK3Worlds/Assets/Editor/x64/Bakery/lmSSS.ptx')
| -rw-r--r-- | VRCSDK3Worlds/Assets/Editor/x64/Bakery/lmSSS.ptx | 1343 |
1 files changed, 1343 insertions, 0 deletions
diff --git a/VRCSDK3Worlds/Assets/Editor/x64/Bakery/lmSSS.ptx b/VRCSDK3Worlds/Assets/Editor/x64/Bakery/lmSSS.ptx new file mode 100644 index 00000000..6c1ca0ac --- /dev/null +++ b/VRCSDK3Worlds/Assets/Editor/x64/Bakery/lmSSS.ptx @@ -0,0 +1,1343 @@ +// +// 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 8 .b8 texCoords[8]; +.global .align 1 .b8 uvpos[1]; +.global .align 1 .b8 uvnormal[1]; +.global .align 1 .b8 rnd_seeds[1]; +.global .align 1 .b8 lightmapDirect[1]; +.global .align 4 .u32 samples; +.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_typeinfo9texCoordsE[8] = {82, 97, 121, 0, 8, 0, 0, 0}; +.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[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_typename9texCoordsE[7] = {102, 108, 111, 97, 116, 50, 0}; +.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[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_typeenum9texCoordsE = 4919; +.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 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 16 .b8 _ZN21rti_internal_semantic9texCoordsE[20] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 101, 120, 67, 111, 111, 114, 100, 115, 0}; +.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[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_annotation9texCoordsE[1]; +.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[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[40]; + .reg .b64 %SP; + .reg .b64 %SPL; + .reg .pred %p<95>; + .reg .b16 %rs<27>; + .reg .f32 %f<624>; + .reg .b32 %r<311>; + .reg .b64 %rd<105>; + + + mov.u64 %rd104, __local_depot0; + cvta.local.u64 %SP, %rd104; + ld.global.v2.u32 {%r97, %r98}, [pixelID]; + cvt.u64.u32 %rd18, %r97; + cvt.u64.u32 %rd19, %r98; + mov.u64 %rd22, uvnormal; + cvta.global.u64 %rd17, %rd22; + mov.u32 %r95, 2; + mov.u32 %r96, 4; + mov.u64 %rd21, 0; + // inline asm + call (%rd16), _rt_buffer_get_64, (%rd17, %r95, %r96, %rd18, %rd19, %rd21, %rd21); + // inline asm + ld.u32 %r1, [%rd16]; + shr.u32 %r101, %r1, 16; + cvt.u16.u32 %rs1, %r101; + and.b16 %rs2, %rs1, 255; + cvt.u16.u32 %rs3, %r1; + or.b16 %rs4, %rs3, %rs2; + setp.eq.s16 %p4, %rs4, 0; + mov.f32 %f582, 0f00000000; + mov.f32 %f583, %f582; + mov.f32 %f584, %f582; + @%p4 bra BB0_2; + + ld.u8 %rs5, [%rd16+1]; + and.b16 %rs7, %rs3, 255; + cvt.rn.f32.u16 %f135, %rs7; + div.rn.f32 %f136, %f135, 0f437F0000; + fma.rn.f32 %f137, %f136, 0f40000000, 0fBF800000; + cvt.rn.f32.u16 %f138, %rs5; + div.rn.f32 %f139, %f138, 0f437F0000; + fma.rn.f32 %f140, %f139, 0f40000000, 0fBF800000; + cvt.rn.f32.u16 %f141, %rs2; + div.rn.f32 %f142, %f141, 0f437F0000; + fma.rn.f32 %f143, %f142, 0f40000000, 0fBF800000; + mul.f32 %f144, %f140, %f140; + fma.rn.f32 %f145, %f137, %f137, %f144; + fma.rn.f32 %f146, %f143, %f143, %f145; + sqrt.rn.f32 %f147, %f146; + rcp.rn.f32 %f148, %f147; + mul.f32 %f582, %f137, %f148; + mul.f32 %f583, %f140, %f148; + mul.f32 %f584, %f143, %f148; + +BB0_2: + ld.global.v2.u32 {%r102, %r103}, [pixelID]; + ld.global.v2.u32 {%r105, %r106}, [tileInfo]; + add.s32 %r2, %r102, %r105; + add.s32 %r3, %r103, %r106; + setp.eq.f32 %p5, %f583, 0f00000000; + setp.eq.f32 %p6, %f582, 0f00000000; + and.pred %p7, %p6, %p5; + setp.eq.f32 %p8, %f584, 0f00000000; + and.pred %p9, %p7, %p8; + @%p9 bra BB0_97; + bra.uni BB0_3; + +BB0_97: + ld.global.u32 %r310, [imageEnabled]; + and.b32 %r269, %r310, 1; + setp.eq.b32 %p93, %r269, 1; + @!%p93 bra BB0_99; + bra.uni BB0_98; + +BB0_98: + cvt.u64.u32 %rd86, %r2; + cvt.u64.u32 %rd87, %r3; + mov.u64 %rd90, image; + cvta.global.u64 %rd85, %rd90; + mov.u64 %rd89, 0; + // inline asm + call (%rd84), _rt_buffer_get_64, (%rd85, %r95, %r96, %rd86, %rd87, %rd89, %rd89); + // inline asm + mov.u16 %rs24, 0; + st.v4.u8 [%rd84], {%rs24, %rs24, %rs24, %rs24}; + ld.global.u32 %r310, [imageEnabled]; + +BB0_99: + and.b32 %r272, %r310, 4; + setp.eq.s32 %p94, %r272, 0; + @%p94 bra BB0_101; + + cvt.u64.u32 %rd93, %r2; + cvt.u64.u32 %rd94, %r3; + mov.u64 %rd97, image_HDR; + cvta.global.u64 %rd92, %rd97; + mov.u32 %r274, 8; + mov.u64 %rd96, 0; + // inline asm + call (%rd91), _rt_buffer_get_64, (%rd92, %r95, %r274, %rd93, %rd94, %rd96, %rd96); + // inline asm + mov.f32 %f552, 0f00000000; + // inline asm + { cvt.rn.f16.f32 %rs25, %f552;} + + // inline asm + mov.u16 %rs26, 0; + st.v4.u16 [%rd91], {%rs25, %rs25, %rs25, %rs26}; + bra.uni BB0_101; + +BB0_3: + ld.global.v2.u32 {%r115, %r116}, [pixelID]; + cvt.u64.u32 %rd25, %r115; + cvt.u64.u32 %rd26, %r116; + mov.u64 %rd30, uvpos; + cvta.global.u64 %rd24, %rd30; + mov.u32 %r114, 12; + // inline asm + call (%rd23), _rt_buffer_get_64, (%rd24, %r95, %r114, %rd25, %rd26, %rd21, %rd21); + // inline asm + ld.f32 %f157, [%rd23+8]; + ld.f32 %f158, [%rd23+4]; + ld.f32 %f159, [%rd23]; + mul.f32 %f160, %f159, 0f3456BF95; + mul.f32 %f161, %f158, 0f3456BF95; + mul.f32 %f162, %f157, 0f3456BF95; + abs.f32 %f163, %f582; + div.rn.f32 %f164, %f160, %f163; + abs.f32 %f165, %f583; + div.rn.f32 %f166, %f161, %f165; + abs.f32 %f167, %f584; + div.rn.f32 %f168, %f162, %f167; + abs.f32 %f169, %f164; + abs.f32 %f170, %f166; + abs.f32 %f171, %f168; + mov.f32 %f172, 0f38D1B717; + max.f32 %f173, %f169, %f172; + max.f32 %f174, %f170, %f172; + max.f32 %f175, %f171, %f172; + fma.rn.f32 %f585, %f582, %f173, %f159; + fma.rn.f32 %f586, %f583, %f174, %f158; + fma.rn.f32 %f587, %f584, %f175, %f157; + add.u64 %rd29, %SP, 28; + cvta.to.local.u64 %rd31, %rd29; + mov.u32 %r113, 0; + st.local.u32 [%rd31+8], %r113; + st.local.u32 [%rd31+4], %r113; + st.local.u32 [%rd31], %r113; + ld.global.u32 %r112, [root]; + neg.f32 %f154, %f584; + neg.f32 %f153, %f583; + neg.f32 %f152, %f582; + mov.f32 %f591, 0f00000000; + mov.f32 %f156, 0f6C4ECB8F; + // inline asm + call _rt_trace_64, (%r112, %f585, %f586, %f587, %f152, %f153, %f154, %r113, %f591, %f156, %rd29, %r114); + // inline asm + ld.local.f32 %f13, [%rd31]; + abs.f32 %f14, %f152; + abs.f32 %f15, %f154; + setp.geu.f32 %p10, %f13, 0f00000000; + @%p10 bra BB0_5; + + fma.rn.f32 %f176, %f582, %f13, %f585; + fma.rn.f32 %f177, %f583, %f13, %f586; + fma.rn.f32 %f178, %f584, %f13, %f587; + mul.f32 %f179, %f176, 0f3456BF95; + mul.f32 %f180, %f177, 0f3456BF95; + mul.f32 %f181, %f178, 0f3456BF95; + div.rn.f32 %f182, %f179, %f14; + abs.f32 %f183, %f153; + div.rn.f32 %f184, %f180, %f183; + div.rn.f32 %f185, %f181, %f15; + abs.f32 %f186, %f182; + abs.f32 %f187, %f184; + abs.f32 %f188, %f185; + max.f32 %f190, %f186, %f172; + max.f32 %f191, %f187, %f172; + max.f32 %f192, %f188, %f172; + fma.rn.f32 %f585, %f190, %f152, %f176; + fma.rn.f32 %f586, %f191, %f153, %f177; + fma.rn.f32 %f587, %f192, %f154, %f178; + +BB0_5: + mov.u32 %r279, 4; + mov.u32 %r278, 2; + setp.gt.f32 %p11, %f14, %f15; + selp.f32 %f196, %f583, 0f00000000, %p11; + selp.f32 %f197, %f152, %f584, %p11; + selp.f32 %f198, 0f00000000, %f153, %p11; + mul.f32 %f199, %f197, %f197; + fma.rn.f32 %f200, %f196, %f196, %f199; + fma.rn.f32 %f201, %f198, %f198, %f200; + sqrt.rn.f32 %f202, %f201; + rcp.rn.f32 %f203, %f202; + mul.f32 %f22, %f196, %f203; + mul.f32 %f23, %f197, %f203; + mul.f32 %f24, %f198, %f203; + ld.global.v2.u32 {%r121, %r122}, [pixelID]; + cvt.u64.u32 %rd34, %r121; + cvt.u64.u32 %rd35, %r122; + mov.u64 %rd38, rnd_seeds; + cvta.global.u64 %rd33, %rd38; + // inline asm + call (%rd32), _rt_buffer_get_64, (%rd33, %r278, %r279, %rd34, %rd35, %rd21, %rd21); + // inline asm + ld.global.u32 %r281, [samples]; + setp.lt.s32 %p12, %r281, 1; + @%p12 bra BB0_6; + + mov.u32 %r282, 0; + cvt.rn.f32.s32 %f207, %r281; + rcp.rn.f32 %f25, %f207; + ld.u32 %r307, [%rd32]; + mul.f32 %f26, %f585, 0f3456BF95; + mul.f32 %f27, %f586, 0f3456BF95; + mul.f32 %f28, %f587, 0f3456BF95; + mul.f32 %f208, %f23, %f152; + mul.f32 %f209, %f22, %f153; + sub.f32 %f29, %f209, %f208; + mul.f32 %f210, %f22, %f154; + mul.f32 %f211, %f24, %f152; + sub.f32 %f30, %f211, %f210; + mul.f32 %f212, %f24, %f153; + mul.f32 %f213, %f23, %f154; + sub.f32 %f31, %f213, %f212; + mov.f32 %f591, 0f00000000; + abs.f32 %f214, %f27; + abs.f32 %f215, %f26; + max.f32 %f216, %f215, %f214; + abs.f32 %f217, %f28; + max.f32 %f218, %f216, %f217; + mov.f32 %f592, %f591; + mov.f32 %f593, %f591; + +BB0_8: + setp.lt.s32 %p13, %r281, 1; + @%p13 bra BB0_59; + + max.f32 %f36, %f218, %f172; + mov.u32 %r284, 0; + +BB0_10: + cvt.rn.f32.s32 %f578, %r282; + mad.lo.s32 %r127, %r307, 1664525, 1013904223; + and.b32 %r128, %r127, 16777215; + cvt.rn.f32.u32 %f220, %r128; + fma.rn.f32 %f221, %f220, 0f33800000, %f578; + mul.f32 %f222, %f25, %f221; + mad.lo.s32 %r307, %r127, 1664525, 1013904223; + and.b32 %r129, %r307, 16777215; + cvt.rn.f32.u32 %f223, %r129; + cvt.rn.f32.s32 %f224, %r284; + fma.rn.f32 %f225, %f223, 0f33800000, %f224; + mul.f32 %f226, %f25, %f225; + sqrt.rn.f32 %f40, %f222; + mul.f32 %f600, %f226, 0f40C90FDB; + abs.f32 %f42, %f600; + setp.neu.f32 %p14, %f42, 0f7F800000; + mov.f32 %f594, %f600; + @%p14 bra BB0_12; + + mov.f32 %f227, 0f00000000; + mul.rn.f32 %f594, %f600, %f227; + +BB0_12: + mul.f32 %f228, %f594, 0f3F22F983; + cvt.rni.s32.f32 %r295, %f228; + cvt.rn.f32.s32 %f229, %r295; + neg.f32 %f230, %f229; + mov.f32 %f231, 0f3FC90FDA; + fma.rn.f32 %f232, %f230, %f231, %f594; + mov.f32 %f233, 0f33A22168; + fma.rn.f32 %f234, %f230, %f233, %f232; + mov.f32 %f235, 0f27C234C5; + fma.rn.f32 %f595, %f230, %f235, %f234; + abs.f32 %f236, %f594; + setp.leu.f32 %p15, %f236, 0f47CE4780; + @%p15 bra BB0_23; + + add.u64 %rd40, %SP, 0; + cvta.to.local.u64 %rd3, %rd40; + mov.b32 %r13, %f594; + shr.u32 %r14, %r13, 23; + shl.b32 %r132, %r13, 8; + or.b32 %r15, %r132, -2147483648; + mov.u32 %r286, 0; + mov.u64 %rd101, 0; + mov.u64 %rd100, %rd3; + mov.u32 %r287, %r286; + +BB0_14: + .pragma "nounroll"; + shl.b64 %rd41, %rd101, 2; + mov.u64 %rd42, __cudart_i2opi_f; + add.s64 %rd43, %rd42, %rd41; + ld.const.u32 %r135, [%rd43]; + // inline asm + { + mad.lo.cc.u32 %r133, %r135, %r15, %r287; + madc.hi.u32 %r287, %r135, %r15, 0; + } + // inline asm + st.local.u32 [%rd100], %r133; + add.s32 %r286, %r286, 1; + cvt.s64.s32 %rd101, %r286; + mul.wide.s32 %rd46, %r286, 4; + add.s64 %rd100, %rd3, %rd46; + setp.ne.s32 %p16, %r286, 6; + @%p16 bra BB0_14; + + and.b32 %r138, %r14, 255; + add.s32 %r139, %r138, -128; + shr.u32 %r140, %r139, 5; + and.b32 %r20, %r13, -2147483648; + cvta.to.local.u64 %rd48, %rd40; + st.local.u32 [%rd48+24], %r287; + mov.u32 %r141, 6; + sub.s32 %r142, %r141, %r140; + mul.wide.s32 %rd49, %r142, 4; + add.s64 %rd8, %rd48, %rd49; + ld.local.u32 %r288, [%rd8]; + ld.local.u32 %r289, [%rd8+-4]; + and.b32 %r23, %r14, 31; + setp.eq.s32 %p17, %r23, 0; + @%p17 bra BB0_17; + + mov.u32 %r143, 32; + sub.s32 %r144, %r143, %r23; + shr.u32 %r145, %r289, %r144; + shl.b32 %r146, %r288, %r23; + add.s32 %r288, %r145, %r146; + ld.local.u32 %r147, [%rd8+-8]; + shr.u32 %r148, %r147, %r144; + shl.b32 %r149, %r289, %r23; + add.s32 %r289, %r148, %r149; + +BB0_17: + shr.u32 %r150, %r289, 30; + shl.b32 %r151, %r288, 2; + add.s32 %r290, %r150, %r151; + shl.b32 %r29, %r289, 2; + shr.u32 %r152, %r290, 31; + shr.u32 %r153, %r288, 30; + add.s32 %r30, %r152, %r153; + setp.eq.s32 %p18, %r152, 0; + @%p18 bra BB0_18; + bra.uni BB0_19; + +BB0_18: + mov.u32 %r291, %r20; + mov.u32 %r292, %r29; + bra.uni BB0_20; + +BB0_19: + not.b32 %r154, %r290; + neg.s32 %r292, %r29; + setp.eq.s32 %p19, %r29, 0; + selp.u32 %r155, 1, 0, %p19; + add.s32 %r290, %r155, %r154; + xor.b32 %r291, %r20, -2147483648; + +BB0_20: + clz.b32 %r294, %r290; + setp.eq.s32 %p20, %r294, 0; + shl.b32 %r156, %r290, %r294; + mov.u32 %r157, 32; + sub.s32 %r158, %r157, %r294; + shr.u32 %r159, %r292, %r158; + add.s32 %r160, %r159, %r156; + selp.b32 %r38, %r290, %r160, %p20; + mov.u32 %r161, -921707870; + mul.hi.u32 %r293, %r38, %r161; + setp.eq.s32 %p21, %r20, 0; + neg.s32 %r162, %r30; + selp.b32 %r295, %r30, %r162, %p21; + setp.lt.s32 %p22, %r293, 1; + @%p22 bra BB0_22; + + mul.lo.s32 %r163, %r38, -921707870; + shr.u32 %r164, %r163, 31; + shl.b32 %r165, %r293, 1; + add.s32 %r293, %r164, %r165; + add.s32 %r294, %r294, 1; + +BB0_22: + mov.u32 %r166, 126; + sub.s32 %r167, %r166, %r294; + shl.b32 %r168, %r167, 23; + add.s32 %r169, %r293, 1; + shr.u32 %r170, %r169, 7; + add.s32 %r171, %r170, 1; + shr.u32 %r172, %r171, 1; + add.s32 %r173, %r172, %r168; + or.b32 %r174, %r173, %r291; + mov.b32 %f595, %r174; + +BB0_23: + mul.rn.f32 %f48, %f595, %f595; + add.s32 %r46, %r295, 1; + and.b32 %r47, %r46, 1; + setp.eq.s32 %p23, %r47, 0; + @%p23 bra BB0_25; + bra.uni BB0_24; + +BB0_25: + mov.f32 %f239, 0f3C08839E; + mov.f32 %f240, 0fB94CA1F9; + fma.rn.f32 %f596, %f240, %f48, %f239; + bra.uni BB0_26; + +BB0_24: + mov.f32 %f237, 0fBAB6061A; + mov.f32 %f238, 0f37CCF5CE; + fma.rn.f32 %f596, %f238, %f48, %f237; + +BB0_26: + @%p23 bra BB0_28; + bra.uni BB0_27; + +BB0_28: + mov.f32 %f244, 0fBE2AAAA3; + fma.rn.f32 %f245, %f596, %f48, %f244; + mov.f32 %f246, 0f00000000; + fma.rn.f32 %f597, %f245, %f48, %f246; + bra.uni BB0_29; + +BB0_27: + mov.f32 %f241, 0f3D2AAAA5; + fma.rn.f32 %f242, %f596, %f48, %f241; + mov.f32 %f243, 0fBF000000; + fma.rn.f32 %f597, %f242, %f48, %f243; + +BB0_29: + fma.rn.f32 %f598, %f597, %f595, %f595; + @%p23 bra BB0_31; + + mov.f32 %f247, 0f3F800000; + fma.rn.f32 %f598, %f597, %f48, %f247; + +BB0_31: + and.b32 %r175, %r46, 2; + setp.eq.s32 %p26, %r175, 0; + @%p26 bra BB0_33; + + mov.f32 %f248, 0f00000000; + mov.f32 %f249, 0fBF800000; + fma.rn.f32 %f598, %f598, %f249, %f248; + +BB0_33: + @%p14 bra BB0_35; + + mov.f32 %f250, 0f00000000; + mul.rn.f32 %f600, %f600, %f250; + +BB0_35: + mul.f32 %f251, %f600, 0f3F22F983; + cvt.rni.s32.f32 %r305, %f251; + cvt.rn.f32.s32 %f252, %r305; + neg.f32 %f253, %f252; + fma.rn.f32 %f255, %f253, %f231, %f600; + fma.rn.f32 %f257, %f253, %f233, %f255; + fma.rn.f32 %f601, %f253, %f235, %f257; + abs.f32 %f259, %f600; + setp.leu.f32 %p28, %f259, 0f47CE4780; + @%p28 bra BB0_46; + + add.u64 %rd51, %SP, 0; + cvta.to.local.u64 %rd9, %rd51; + mov.b32 %r49, %f600; + shr.u32 %r50, %r49, 23; + shl.b32 %r178, %r49, 8; + or.b32 %r51, %r178, -2147483648; + mov.u32 %r296, 0; + mov.u64 %rd102, %rd9; + mov.u64 %rd103, %rd21; + mov.u32 %r297, %r296; + +BB0_37: + .pragma "nounroll"; + shl.b64 %rd52, %rd103, 2; + mov.u64 %rd53, __cudart_i2opi_f; + add.s64 %rd54, %rd53, %rd52; + ld.const.u32 %r181, [%rd54]; + // inline asm + { + mad.lo.cc.u32 %r179, %r181, %r51, %r297; + madc.hi.u32 %r297, %r181, %r51, 0; + } + // inline asm + st.local.u32 [%rd102], %r179; + add.s32 %r296, %r296, 1; + cvt.s64.s32 %rd103, %r296; + mul.wide.s32 %rd55, %r296, 4; + add.s64 %rd102, %rd9, %rd55; + setp.ne.s32 %p29, %r296, 6; + @%p29 bra BB0_37; + + and.b32 %r184, %r50, 255; + add.s32 %r185, %r184, -128; + shr.u32 %r186, %r185, 5; + and.b32 %r56, %r49, -2147483648; + cvta.to.local.u64 %rd57, %rd51; + st.local.u32 [%rd57+24], %r297; + mov.u32 %r187, 6; + sub.s32 %r188, %r187, %r186; + mul.wide.s32 %rd58, %r188, 4; + add.s64 %rd15, %rd57, %rd58; + ld.local.u32 %r298, [%rd15]; + ld.local.u32 %r299, [%rd15+-4]; + and.b32 %r59, %r50, 31; + setp.eq.s32 %p30, %r59, 0; + @%p30 bra BB0_40; + + mov.u32 %r189, 32; + sub.s32 %r190, %r189, %r59; + shr.u32 %r191, %r299, %r190; + shl.b32 %r192, %r298, %r59; + add.s32 %r298, %r191, %r192; + ld.local.u32 %r193, [%rd15+-8]; + shr.u32 %r194, %r193, %r190; + shl.b32 %r195, %r299, %r59; + add.s32 %r299, %r194, %r195; + +BB0_40: + shr.u32 %r196, %r299, 30; + shl.b32 %r197, %r298, 2; + add.s32 %r300, %r196, %r197; + shl.b32 %r65, %r299, 2; + shr.u32 %r198, %r300, 31; + shr.u32 %r199, %r298, 30; + add.s32 %r66, %r198, %r199; + setp.eq.s32 %p31, %r198, 0; + @%p31 bra BB0_41; + bra.uni BB0_42; + +BB0_41: + mov.u32 %r301, %r56; + mov.u32 %r302, %r65; + bra.uni BB0_43; + +BB0_42: + not.b32 %r200, %r300; + neg.s32 %r302, %r65; + setp.eq.s32 %p32, %r65, 0; + selp.u32 %r201, 1, 0, %p32; + add.s32 %r300, %r201, %r200; + xor.b32 %r301, %r56, -2147483648; + +BB0_43: + clz.b32 %r304, %r300; + setp.eq.s32 %p33, %r304, 0; + shl.b32 %r202, %r300, %r304; + mov.u32 %r203, 32; + sub.s32 %r204, %r203, %r304; + shr.u32 %r205, %r302, %r204; + add.s32 %r206, %r205, %r202; + selp.b32 %r74, %r300, %r206, %p33; + mov.u32 %r207, -921707870; + mul.hi.u32 %r303, %r74, %r207; + setp.eq.s32 %p34, %r56, 0; + neg.s32 %r208, %r66; + selp.b32 %r305, %r66, %r208, %p34; + setp.lt.s32 %p35, %r303, 1; + @%p35 bra BB0_45; + + mul.lo.s32 %r209, %r74, -921707870; + shr.u32 %r210, %r209, 31; + shl.b32 %r211, %r303, 1; + add.s32 %r303, %r210, %r211; + add.s32 %r304, %r304, 1; + +BB0_45: + mov.u32 %r212, 126; + sub.s32 %r213, %r212, %r304; + shl.b32 %r214, %r213, 23; + add.s32 %r215, %r303, 1; + shr.u32 %r216, %r215, 7; + add.s32 %r217, %r216, 1; + shr.u32 %r218, %r217, 1; + add.s32 %r219, %r218, %r214; + or.b32 %r220, %r219, %r301; + mov.b32 %f601, %r220; + +BB0_46: + mul.rn.f32 %f65, %f601, %f601; + and.b32 %r82, %r305, 1; + setp.eq.s32 %p36, %r82, 0; + @%p36 bra BB0_48; + bra.uni BB0_47; + +BB0_48: + mov.f32 %f262, 0f3C08839E; + mov.f32 %f263, 0fB94CA1F9; + fma.rn.f32 %f602, %f263, %f65, %f262; + bra.uni BB0_49; + +BB0_47: + mov.f32 %f260, 0fBAB6061A; + mov.f32 %f261, 0f37CCF5CE; + fma.rn.f32 %f602, %f261, %f65, %f260; + +BB0_49: + @%p36 bra BB0_51; + bra.uni BB0_50; + +BB0_51: + mov.f32 %f267, 0fBE2AAAA3; + fma.rn.f32 %f268, %f602, %f65, %f267; + mov.f32 %f269, 0f00000000; + fma.rn.f32 %f603, %f268, %f65, %f269; + bra.uni BB0_52; + +BB0_50: + mov.f32 %f264, 0f3D2AAAA5; + fma.rn.f32 %f265, %f602, %f65, %f264; + mov.f32 %f266, 0fBF000000; + fma.rn.f32 %f603, %f265, %f65, %f266; + +BB0_52: + fma.rn.f32 %f604, %f603, %f601, %f601; + @%p36 bra BB0_54; + + mov.f32 %f270, 0f3F800000; + fma.rn.f32 %f604, %f603, %f65, %f270; + +BB0_54: + and.b32 %r221, %r305, 2; + setp.eq.s32 %p39, %r221, 0; + @%p39 bra BB0_56; + + mov.f32 %f271, 0f00000000; + mov.f32 %f272, 0fBF800000; + fma.rn.f32 %f604, %f604, %f272, %f271; + +BB0_56: + mul.f32 %f281, %f40, %f598; + mul.f32 %f282, %f281, %f281; + mov.f32 %f283, 0f3F800000; + sub.f32 %f284, %f283, %f282; + mul.f32 %f285, %f40, %f604; + mul.f32 %f286, %f285, %f285; + sub.f32 %f287, %f284, %f286; + mov.f32 %f288, 0f00000000; + max.f32 %f289, %f288, %f287; + sqrt.rn.f32 %f290, %f289; + mul.f32 %f291, %f22, %f285; + mul.f32 %f292, %f23, %f285; + mul.f32 %f293, %f24, %f285; + fma.rn.f32 %f294, %f31, %f281, %f291; + fma.rn.f32 %f295, %f30, %f281, %f292; + fma.rn.f32 %f296, %f29, %f281, %f293; + fma.rn.f32 %f276, %f290, %f152, %f294; + fma.rn.f32 %f277, %f290, %f153, %f295; + fma.rn.f32 %f278, %f290, %f154, %f296; + mov.u32 %r225, -1082130432; + st.local.u32 [%rd31+8], %r225; + st.local.u32 [%rd31+4], %r225; + st.local.u32 [%rd31], %r225; + ld.global.u32 %r222, [root]; + mov.u32 %r223, 0; + // inline asm + call _rt_trace_64, (%r222, %f585, %f586, %f587, %f276, %f277, %f278, %r223, %f36, %f156, %rd29, %r114); + // inline asm + ld.local.f32 %f77, [%rd31]; + setp.lt.f32 %p40, %f77, 0f00000000; + @%p40 bra BB0_58; + + add.f32 %f593, %f593, %f77; + ld.local.f32 %f297, [%rd31+4]; + add.f32 %f592, %f592, %f297; + ld.local.f32 %f298, [%rd31+8]; + add.f32 %f591, %f591, %f298; + +BB0_58: + ld.global.u32 %r281, [samples]; + add.s32 %r284, %r284, 1; + setp.lt.s32 %p41, %r284, %r281; + @%p41 bra BB0_10; + +BB0_59: + add.s32 %r282, %r282, 1; + setp.lt.s32 %p42, %r282, %r281; + @%p42 bra BB0_8; + bra.uni BB0_60; + +BB0_6: + mov.f32 %f592, %f591; + mov.f32 %f593, %f591; + +BB0_60: + mul.lo.s32 %r226, %r281, %r281; + cvt.rn.f32.s32 %f299, %r226; + rcp.rn.f32 %f300, %f299; + mul.f32 %f90, %f593, %f300; + mul.f32 %f91, %f592, %f300; + mul.f32 %f92, %f591, %f300; + ld.global.u32 %r309, [imageEnabled]; + and.b32 %r227, %r309, 1; + setp.eq.b32 %p43, %r227, 1; + @!%p43 bra BB0_95; + bra.uni BB0_61; + +BB0_61: + mov.f32 %f303, 0f3E666666; + cvt.rzi.f32.f32 %f304, %f303; + fma.rn.f32 %f305, %f304, 0fC0000000, 0f3EE66666; + abs.f32 %f93, %f305; + abs.f32 %f94, %f90; + setp.lt.f32 %p44, %f94, 0f00800000; + mul.f32 %f306, %f94, 0f4B800000; + selp.f32 %f307, 0fC3170000, 0fC2FE0000, %p44; + selp.f32 %f308, %f306, %f94, %p44; + mov.b32 %r228, %f308; + and.b32 %r229, %r228, 8388607; + or.b32 %r230, %r229, 1065353216; + mov.b32 %f309, %r230; + shr.u32 %r231, %r228, 23; + cvt.rn.f32.u32 %f310, %r231; + add.f32 %f311, %f307, %f310; + setp.gt.f32 %p45, %f309, 0f3FB504F3; + mul.f32 %f312, %f309, 0f3F000000; + add.f32 %f313, %f311, 0f3F800000; + selp.f32 %f314, %f312, %f309, %p45; + selp.f32 %f315, %f313, %f311, %p45; + add.f32 %f316, %f314, 0fBF800000; + add.f32 %f302, %f314, 0f3F800000; + // inline asm + rcp.approx.ftz.f32 %f301,%f302; + // inline asm + add.f32 %f317, %f316, %f316; + mul.f32 %f318, %f301, %f317; + mul.f32 %f319, %f318, %f318; + mov.f32 %f320, 0f3C4CAF63; + mov.f32 %f321, 0f3B18F0FE; + fma.rn.f32 %f322, %f321, %f319, %f320; + mov.f32 %f323, 0f3DAAAABD; + fma.rn.f32 %f324, %f322, %f319, %f323; + mul.rn.f32 %f325, %f324, %f319; + mul.rn.f32 %f326, %f325, %f318; + sub.f32 %f327, %f316, %f318; + neg.f32 %f328, %f318; + add.f32 %f329, %f327, %f327; + fma.rn.f32 %f330, %f328, %f316, %f329; + mul.rn.f32 %f331, %f301, %f330; + add.f32 %f332, %f326, %f318; + sub.f32 %f333, %f318, %f332; + add.f32 %f334, %f326, %f333; + add.f32 %f335, %f331, %f334; + add.f32 %f336, %f332, %f335; + sub.f32 %f337, %f332, %f336; + add.f32 %f338, %f335, %f337; + mov.f32 %f339, 0f3F317200; + mul.rn.f32 %f340, %f315, %f339; + mov.f32 %f341, 0f35BFBE8E; + mul.rn.f32 %f342, %f315, %f341; + add.f32 %f343, %f340, %f336; + sub.f32 %f344, %f340, %f343; + add.f32 %f345, %f336, %f344; + add.f32 %f346, %f338, %f345; + add.f32 %f347, %f342, %f346; + add.f32 %f348, %f343, %f347; + sub.f32 %f349, %f343, %f348; + add.f32 %f350, %f347, %f349; + mov.f32 %f351, 0f3EE66666; + mul.rn.f32 %f352, %f351, %f348; + neg.f32 %f353, %f352; + fma.rn.f32 %f354, %f351, %f348, %f353; + fma.rn.f32 %f355, %f351, %f350, %f354; + mov.f32 %f356, 0f00000000; + fma.rn.f32 %f357, %f356, %f348, %f355; + add.rn.f32 %f358, %f352, %f357; + neg.f32 %f359, %f358; + add.rn.f32 %f360, %f352, %f359; + add.rn.f32 %f361, %f360, %f357; + mov.b32 %r232, %f358; + setp.eq.s32 %p46, %r232, 1118925336; + add.s32 %r233, %r232, -1; + mov.b32 %f362, %r233; + add.f32 %f363, %f361, 0f37000000; + selp.f32 %f364, %f362, %f358, %p46; + selp.f32 %f95, %f363, %f361, %p46; + mul.f32 %f365, %f364, 0f3FB8AA3B; + cvt.rzi.f32.f32 %f366, %f365; + mov.f32 %f367, 0fBF317200; + fma.rn.f32 %f368, %f366, %f367, %f364; + mov.f32 %f369, 0fB5BFBE8E; + fma.rn.f32 %f370, %f366, %f369, %f368; + mul.f32 %f371, %f370, 0f3FB8AA3B; + ex2.approx.ftz.f32 %f372, %f371; + add.f32 %f373, %f366, 0f00000000; + ex2.approx.f32 %f374, %f373; + mul.f32 %f375, %f372, %f374; + setp.lt.f32 %p47, %f364, 0fC2D20000; + selp.f32 %f376, 0f00000000, %f375, %p47; + setp.gt.f32 %p48, %f364, 0f42D20000; + selp.f32 %f615, 0f7F800000, %f376, %p48; + setp.eq.f32 %p49, %f615, 0f7F800000; + @%p49 bra BB0_63; + + fma.rn.f32 %f615, %f615, %f95, %f615; + +BB0_63: + setp.lt.f32 %p50, %f90, 0f00000000; + setp.eq.f32 %p51, %f93, 0f3F800000; + and.pred %p1, %p50, %p51; + mov.b32 %r234, %f615; + xor.b32 %r235, %r234, -2147483648; + mov.b32 %f377, %r235; + selp.f32 %f617, %f377, %f615, %p1; + setp.eq.f32 %p52, %f90, 0f00000000; + @%p52 bra BB0_66; + bra.uni BB0_64; + +BB0_66: + add.f32 %f380, %f90, %f90; + selp.f32 %f617, %f380, 0f00000000, %p51; + bra.uni BB0_67; + +BB0_64: + setp.geu.f32 %p53, %f90, 0f00000000; + @%p53 bra BB0_67; + + cvt.rzi.f32.f32 %f379, %f351; + setp.neu.f32 %p54, %f379, 0f3EE66666; + selp.f32 %f617, 0f7FFFFFFF, %f617, %p54; + +BB0_67: + abs.f32 %f553, %f90; + add.f32 %f381, %f553, 0f3EE66666; + mov.b32 %r236, %f381; + setp.lt.s32 %p56, %r236, 2139095040; + @%p56 bra BB0_72; + + abs.f32 %f570, %f90; + setp.gtu.f32 %p57, %f570, 0f7F800000; + @%p57 bra BB0_71; + bra.uni BB0_69; + +BB0_71: + add.f32 %f617, %f90, 0f3EE66666; + bra.uni BB0_72; + +BB0_69: + abs.f32 %f571, %f90; + setp.neu.f32 %p58, %f571, 0f7F800000; + @%p58 bra BB0_72; + + selp.f32 %f617, 0fFF800000, 0f7F800000, %p1; + +BB0_72: + mov.f32 %f561, 0fB5BFBE8E; + mov.f32 %f560, 0fBF317200; + mov.f32 %f559, 0f00000000; + mov.f32 %f558, 0f35BFBE8E; + mov.f32 %f557, 0f3F317200; + mov.f32 %f556, 0f3DAAAABD; + mov.f32 %f555, 0f3C4CAF63; + mov.f32 %f554, 0f3B18F0FE; + setp.eq.f32 %p59, %f90, 0f3F800000; + selp.f32 %f106, 0f3F800000, %f617, %p59; + abs.f32 %f107, %f91; + setp.lt.f32 %p60, %f107, 0f00800000; + mul.f32 %f384, %f107, 0f4B800000; + selp.f32 %f385, 0fC3170000, 0fC2FE0000, %p60; + selp.f32 %f386, %f384, %f107, %p60; + mov.b32 %r237, %f386; + and.b32 %r238, %r237, 8388607; + or.b32 %r239, %r238, 1065353216; + mov.b32 %f387, %r239; + shr.u32 %r240, %r237, 23; + cvt.rn.f32.u32 %f388, %r240; + add.f32 %f389, %f385, %f388; + setp.gt.f32 %p61, %f387, 0f3FB504F3; + mul.f32 %f390, %f387, 0f3F000000; + add.f32 %f391, %f389, 0f3F800000; + selp.f32 %f392, %f390, %f387, %p61; + selp.f32 %f393, %f391, %f389, %p61; + add.f32 %f394, %f392, 0fBF800000; + add.f32 %f383, %f392, 0f3F800000; + // inline asm + rcp.approx.ftz.f32 %f382,%f383; + // inline asm + add.f32 %f395, %f394, %f394; + mul.f32 %f396, %f382, %f395; + mul.f32 %f397, %f396, %f396; + fma.rn.f32 %f400, %f554, %f397, %f555; + fma.rn.f32 %f402, %f400, %f397, %f556; + mul.rn.f32 %f403, %f402, %f397; + mul.rn.f32 %f404, %f403, %f396; + sub.f32 %f405, %f394, %f396; + neg.f32 %f406, %f396; + add.f32 %f407, %f405, %f405; + fma.rn.f32 %f408, %f406, %f394, %f407; + mul.rn.f32 %f409, %f382, %f408; + add.f32 %f410, %f404, %f396; + sub.f32 %f411, %f396, %f410; + add.f32 %f412, %f404, %f411; + add.f32 %f413, %f409, %f412; + add.f32 %f414, %f410, %f413; + sub.f32 %f415, %f410, %f414; + add.f32 %f416, %f413, %f415; + mul.rn.f32 %f418, %f393, %f557; + mul.rn.f32 %f420, %f393, %f558; + add.f32 %f421, %f418, %f414; + sub.f32 %f422, %f418, %f421; + add.f32 %f423, %f414, %f422; + add.f32 %f424, %f416, %f423; + add.f32 %f425, %f420, %f424; + add.f32 %f426, %f421, %f425; + sub.f32 %f427, %f421, %f426; + add.f32 %f428, %f425, %f427; + mul.rn.f32 %f430, %f351, %f426; + neg.f32 %f431, %f430; + fma.rn.f32 %f432, %f351, %f426, %f431; + fma.rn.f32 %f433, %f351, %f428, %f432; + fma.rn.f32 %f435, %f559, %f426, %f433; + add.rn.f32 %f436, %f430, %f435; + neg.f32 %f437, %f436; + add.rn.f32 %f438, %f430, %f437; + add.rn.f32 %f439, %f438, %f435; + mov.b32 %r241, %f436; + setp.eq.s32 %p62, %r241, 1118925336; + add.s32 %r242, %r241, -1; + mov.b32 %f440, %r242; + add.f32 %f441, %f439, 0f37000000; + selp.f32 %f442, %f440, %f436, %p62; + selp.f32 %f108, %f441, %f439, %p62; + mul.f32 %f443, %f442, 0f3FB8AA3B; + cvt.rzi.f32.f32 %f444, %f443; + fma.rn.f32 %f446, %f444, %f560, %f442; + fma.rn.f32 %f448, %f444, %f561, %f446; + mul.f32 %f449, %f448, 0f3FB8AA3B; + ex2.approx.ftz.f32 %f450, %f449; + add.f32 %f451, %f444, 0f00000000; + ex2.approx.f32 %f452, %f451; + mul.f32 %f453, %f450, %f452; + setp.lt.f32 %p63, %f442, 0fC2D20000; + selp.f32 %f454, 0f00000000, %f453, %p63; + setp.gt.f32 %p64, %f442, 0f42D20000; + selp.f32 %f618, 0f7F800000, %f454, %p64; + setp.eq.f32 %p65, %f618, 0f7F800000; + @%p65 bra BB0_74; + + fma.rn.f32 %f618, %f618, %f108, %f618; + +BB0_74: + setp.lt.f32 %p66, %f91, 0f00000000; + and.pred %p2, %p66, %p51; + mov.b32 %r243, %f618; + xor.b32 %r244, %r243, -2147483648; + mov.b32 %f455, %r244; + selp.f32 %f620, %f455, %f618, %p2; + setp.eq.f32 %p68, %f91, 0f00000000; + @%p68 bra BB0_77; + bra.uni BB0_75; + +BB0_77: + add.f32 %f458, %f91, %f91; + selp.f32 %f620, %f458, 0f00000000, %p51; + bra.uni BB0_78; + +BB0_75: + setp.geu.f32 %p69, %f91, 0f00000000; + @%p69 bra BB0_78; + + mov.f32 %f577, 0f3EE66666; + cvt.rzi.f32.f32 %f457, %f577; + setp.neu.f32 %p70, %f457, 0f3EE66666; + selp.f32 %f620, 0f7FFFFFFF, %f620, %p70; + +BB0_78: + abs.f32 %f572, %f91; + add.f32 %f459, %f572, 0f3EE66666; + mov.b32 %r245, %f459; + setp.lt.s32 %p72, %r245, 2139095040; + @%p72 bra BB0_83; + + abs.f32 %f575, %f91; + setp.gtu.f32 %p73, %f575, 0f7F800000; + @%p73 bra BB0_82; + bra.uni BB0_80; + +BB0_82: + add.f32 %f620, %f91, 0f3EE66666; + bra.uni BB0_83; + +BB0_80: + abs.f32 %f576, %f91; + setp.neu.f32 %p74, %f576, 0f7F800000; + @%p74 bra BB0_83; + + selp.f32 %f620, 0fFF800000, 0f7F800000, %p2; + +BB0_83: + mov.f32 %f573, 0f3EE66666; + mov.f32 %f569, 0fB5BFBE8E; + mov.f32 %f568, 0fBF317200; + mov.f32 %f567, 0f00000000; + mov.f32 %f566, 0f35BFBE8E; + mov.f32 %f565, 0f3F317200; + mov.f32 %f564, 0f3DAAAABD; + mov.f32 %f563, 0f3C4CAF63; + mov.f32 %f562, 0f3B18F0FE; + setp.eq.f32 %p75, %f91, 0f3F800000; + selp.f32 %f119, 0f3F800000, %f620, %p75; + abs.f32 %f120, %f92; + setp.lt.f32 %p76, %f120, 0f00800000; + mul.f32 %f462, %f120, 0f4B800000; + selp.f32 %f463, 0fC3170000, 0fC2FE0000, %p76; + selp.f32 %f464, %f462, %f120, %p76; + mov.b32 %r246, %f464; + and.b32 %r247, %r246, 8388607; + or.b32 %r248, %r247, 1065353216; + mov.b32 %f465, %r248; + shr.u32 %r249, %r246, 23; + cvt.rn.f32.u32 %f466, %r249; + add.f32 %f467, %f463, %f466; + setp.gt.f32 %p77, %f465, 0f3FB504F3; + mul.f32 %f468, %f465, 0f3F000000; + add.f32 %f469, %f467, 0f3F800000; + selp.f32 %f470, %f468, %f465, %p77; + selp.f32 %f471, %f469, %f467, %p77; + add.f32 %f472, %f470, 0fBF800000; + add.f32 %f461, %f470, 0f3F800000; + // inline asm + rcp.approx.ftz.f32 %f460,%f461; + // inline asm + add.f32 %f473, %f472, %f472; + mul.f32 %f474, %f460, %f473; + mul.f32 %f475, %f474, %f474; + fma.rn.f32 %f478, %f562, %f475, %f563; + fma.rn.f32 %f480, %f478, %f475, %f564; + mul.rn.f32 %f481, %f480, %f475; + mul.rn.f32 %f482, %f481, %f474; + sub.f32 %f483, %f472, %f474; + neg.f32 %f484, %f474; + add.f32 %f485, %f483, %f483; + fma.rn.f32 %f486, %f484, %f472, %f485; + mul.rn.f32 %f487, %f460, %f486; + add.f32 %f488, %f482, %f474; + sub.f32 %f489, %f474, %f488; + add.f32 %f490, %f482, %f489; + add.f32 %f491, %f487, %f490; + add.f32 %f492, %f488, %f491; + sub.f32 %f493, %f488, %f492; + add.f32 %f494, %f491, %f493; + mul.rn.f32 %f496, %f471, %f565; + mul.rn.f32 %f498, %f471, %f566; + add.f32 %f499, %f496, %f492; + sub.f32 %f500, %f496, %f499; + add.f32 %f501, %f492, %f500; + add.f32 %f502, %f494, %f501; + add.f32 %f503, %f498, %f502; + add.f32 %f504, %f499, %f503; + sub.f32 %f505, %f499, %f504; + add.f32 %f506, %f503, %f505; + mul.rn.f32 %f508, %f573, %f504; + neg.f32 %f509, %f508; + fma.rn.f32 %f510, %f573, %f504, %f509; + fma.rn.f32 %f511, %f573, %f506, %f510; + fma.rn.f32 %f513, %f567, %f504, %f511; + add.rn.f32 %f514, %f508, %f513; + neg.f32 %f515, %f514; + add.rn.f32 %f516, %f508, %f515; + add.rn.f32 %f517, %f516, %f513; + mov.b32 %r250, %f514; + setp.eq.s32 %p78, %r250, 1118925336; + add.s32 %r251, %r250, -1; + mov.b32 %f518, %r251; + add.f32 %f519, %f517, 0f37000000; + selp.f32 %f520, %f518, %f514, %p78; + selp.f32 %f121, %f519, %f517, %p78; + mul.f32 %f521, %f520, 0f3FB8AA3B; + cvt.rzi.f32.f32 %f522, %f521; + fma.rn.f32 %f524, %f522, %f568, %f520; + fma.rn.f32 %f526, %f522, %f569, %f524; + mul.f32 %f527, %f526, 0f3FB8AA3B; + ex2.approx.ftz.f32 %f528, %f527; + add.f32 %f529, %f522, 0f00000000; + ex2.approx.f32 %f530, %f529; + mul.f32 %f531, %f528, %f530; + setp.lt.f32 %p79, %f520, 0fC2D20000; + selp.f32 %f532, 0f00000000, %f531, %p79; + setp.gt.f32 %p80, %f520, 0f42D20000; + selp.f32 %f621, 0f7F800000, %f532, %p80; + setp.eq.f32 %p81, %f621, 0f7F800000; + @%p81 bra BB0_85; + + fma.rn.f32 %f621, %f621, %f121, %f621; + +BB0_85: + setp.lt.f32 %p82, %f92, 0f00000000; + and.pred %p3, %p82, %p51; + mov.b32 %r252, %f621; + xor.b32 %r253, %r252, -2147483648; + mov.b32 %f533, %r253; + selp.f32 %f623, %f533, %f621, %p3; + setp.eq.f32 %p84, %f92, 0f00000000; + @%p84 bra BB0_88; + bra.uni BB0_86; + +BB0_88: + add.f32 %f536, %f92, %f92; + selp.f32 %f623, %f536, 0f00000000, %p51; + bra.uni BB0_89; + +BB0_86: + setp.geu.f32 %p85, %f92, 0f00000000; + @%p85 bra BB0_89; + + mov.f32 %f574, 0f3EE66666; + cvt.rzi.f32.f32 %f535, %f574; + setp.neu.f32 %p86, %f535, 0f3EE66666; + selp.f32 %f623, 0f7FFFFFFF, %f623, %p86; + +BB0_89: + abs.f32 %f579, %f92; + add.f32 %f537, %f579, 0f3EE66666; + mov.b32 %r254, %f537; + setp.lt.s32 %p88, %r254, 2139095040; + @%p88 bra BB0_94; + + abs.f32 %f580, %f92; + setp.gtu.f32 %p89, %f580, 0f7F800000; + @%p89 bra BB0_93; + bra.uni BB0_91; + +BB0_93: + add.f32 %f623, %f92, 0f3EE66666; + bra.uni BB0_94; + +BB0_91: + abs.f32 %f581, %f92; + setp.neu.f32 %p90, %f581, 0f7F800000; + @%p90 bra BB0_94; + + selp.f32 %f623, 0fFF800000, 0f7F800000, %p3; + +BB0_94: + mov.u32 %r276, 4; + mov.u64 %rd98, 0; + mov.u32 %r275, 2; + setp.eq.f32 %p91, %f92, 0f3F800000; + selp.f32 %f538, 0f3F800000, %f623, %p91; + cvt.u64.u32 %rd66, %r3; + cvt.u64.u32 %rd65, %r2; + mov.u64 %rd69, image; + cvta.global.u64 %rd64, %rd69; + // inline asm + call (%rd63), _rt_buffer_get_64, (%rd64, %r275, %r276, %rd65, %rd66, %rd98, %rd98); + // inline asm + cvt.sat.f32.f32 %f539, %f538; + mul.f32 %f540, %f539, 0f437FFD71; + cvt.rzi.u32.f32 %r257, %f540; + cvt.sat.f32.f32 %f541, %f119; + mul.f32 %f542, %f541, 0f437FFD71; + cvt.rzi.u32.f32 %r258, %f542; + cvt.sat.f32.f32 %f543, %f106; + mul.f32 %f544, %f543, 0f437FFD71; + cvt.rzi.u32.f32 %r259, %f544; + cvt.u16.u32 %rs9, %r257; + cvt.u16.u32 %rs10, %r259; + cvt.u16.u32 %rs11, %r258; + mov.u16 %rs12, 255; + st.v4.u8 [%rd63], {%rs9, %rs11, %rs10, %rs12}; + ld.global.u32 %r309, [imageEnabled]; + +BB0_95: + and.b32 %r260, %r309, 4; + setp.eq.s32 %p92, %r260, 0; + @%p92 bra BB0_101; + + mov.u64 %rd99, 0; + mov.u32 %r277, 2; + ld.global.v2.u32 {%r265, %r266}, [pixelID]; + cvt.u64.u32 %rd72, %r265; + cvt.u64.u32 %rd73, %r266; + mov.u64 %rd82, lightmapDirect; + cvta.global.u64 %rd71, %rd82; + mov.u32 %r264, 8; + // inline asm + call (%rd70), _rt_buffer_get_64, (%rd71, %r277, %r264, %rd72, %rd73, %rd99, %rd99); + // inline asm + ld.v4.u16 {%rs20, %rs21, %rs22, %rs23}, [%rd70]; + // inline asm + { cvt.f32.f16 %f545, %rs20;} + + // inline asm + // inline asm + { cvt.f32.f16 %f546, %rs21;} + + // inline asm + // inline asm + { cvt.f32.f16 %f547, %rs22;} + + // inline asm + add.f32 %f548, %f90, %f545; + add.f32 %f549, %f91, %f546; + add.f32 %f550, %f92, %f547; + cvt.u64.u32 %rd79, %r3; + cvt.u64.u32 %rd78, %r2; + mov.u64 %rd83, image_HDR; + cvta.global.u64 %rd77, %rd83; + // inline asm + call (%rd76), _rt_buffer_get_64, (%rd77, %r277, %r264, %rd78, %rd79, %rd99, %rd99); + // inline asm + // inline asm + { cvt.rn.f16.f32 %rs18, %f550;} + + // inline asm + // inline asm + { cvt.rn.f16.f32 %rs17, %f549;} + + // inline asm + // inline asm + { cvt.rn.f16.f32 %rs16, %f548;} + + // inline asm + mov.f32 %f551, 0f3F800000; + // inline asm + { cvt.rn.f16.f32 %rs19, %f551;} + + // inline asm + st.v4.u16 [%rd76], {%rs16, %rs17, %rs18, %rs19}; + +BB0_101: + ret; +} + + |